Datasets:

recursal

/

arXiv-CC0-v0.5

like 0

2 DAYOON PARK ofNoverZby anm-gonal form. As is well known, the problem of every positive integer Nwhich is represented by arbitrary given m- gonal form Fm(x) is not easy in general. In an effort to approach the problem, we suggest the reduced congruence equation (1.4) the a non-negative integer solution xNn 0) for every rZ, we say that Fm(x) locally represents N. The solvability of congruence for every prime pand natural number aNwould be enough to derive the local representability of NbyFm(x) in virtue of the Chinese Remainder Theorem. Note that for a prime p, the congruence equation Fm(x)N(modpa) has an integer solution xZnfor every natural number aNif and only if the equation Fm(x) =Nhas ap-adic integer solution xZn p. When the equation Fm(x) =N has ap-adic integer solution xZn p, we say that Fm(x)represents the (global) representability implies the local representa bility, but the converse does not hold in general. Especially, when the converse als o holds, i.e., an m-gonalform Fm(x) Nwhichis overZ(resp. over N0), we say that Fm(x) isregular over Z(resp. over N0). WhenFm(x) represents every positive integer Nwhich is locally but finitely many over Z(resp. over N0), we say that Fm(x) isalmost regular over Z(resp. over N0). The concept of regular form was firstly introduced by Dickson in [3]. In [12], Watson proved that there are only finitely many primitive positive defi nite Watson's L- transformation which has been widely applied in the study of regular form. In [5], Jagy, Kapla nsky, and Schiemann who used the Watson's L-transformation classified 9 13 candidates for primitive positive definite regular ternary quadratic forms. And they claimed the regularity of all of them but 22. Oh [9] showed the regularity of 8 forms of unconfirmed 22 forms. And Lemke Oliver [8] proved the regularity of the remaining 14 candidates under the assumption of generalized Riemann Hypoth esis. The finiteness of primitive regular ternary triangular forms (i.e., 3-g onal forms) was shown by Chan and Oh [1] and all the finitely many primitive regular t forms was determined by Kim and Oh [7]. As an improvement o f the result in [1], Chan and Ricci [2] proved the finiteness of primitive regula r polynomials with a fixed conductor in [2]. As another version 's improve- ment of the work, in [1], He and Kane showed the finiteness of primitive regular ternary polygonal forms (i.e., m-gonal forms for arbitrary m>=3). Which implies that there are only finitely many mwhich admits a regular m-gonal form of rank 3. In [6], Kim and the author showed that for each rank n>=4 too, there are only finitely many mwhich admits a regular m-gonal form of rank n. Moreover in [6], they completely determined the type of regular m-gonal forms of rank n>=4 for allm>=14 (mod 4) and m>=28 withm0 (mod 4) by which is well designed to track the regular m-gonal forms. Now let us turn to our main subject which is almost regular form in this a rticle. By Theorem 4.9 (1) in [1], every m-gonal form of rank n>=5 is almost regular over Z, i.e., every m-gonalform ofrank n>=5 representsall the sufficiently other word, all but finitely many integers) which are locally repres ented by the form over Z. But such the nice local-to-global principle is not guaranteed when the rank is less than 5 in general. There is non-almost regular m-gonal form ofNON-ALMOST REGULAR QUATERNARY m-GONAL FORMS 3 rank 4, and by the above argument, 4 should be the maximum rank wh ich admits a non-almost regular m-gonal form over Z. In this short article, we mainly look into the non-almost regular m-gonal forms of rank 4. The following two theorems are the main goal in this article. Theorem 1.1. given. (1)If the quaternary quadratic form Zpis isotropic for ev- ery prime p, then the quaternary m-gonal form mis almost regular over Zfor define T:={p|Zpis anisotropic }a non-empty set of primes. For any for (mod 4) if2T, them-gonal form mis almost regular over Z. For every sufficiently large mwhich is not satisfying (1.5), them-gonal formmis not almost regular over Z(so also over N0). Theorem 1.2. For any m>=3, there are infinitely many primitive forms which are not almost regular over Z(so also over N0). Remark 1.3. By using Theorem 1.1, for arbitrary given any sufficiently large m, we may directly determine whether the m-gonal formmis almost regular or not. Throughout this article, we adopt the arithmetic theory of quadra tic forms. Before we move on, we briefly set our languages and terminologies wh ich are used throughout this article. If a quadratic form (or m-gonal form) represents every p-adic integer in NZp overZp, then we say that the quadratic form (or m-gonal form) is especially, when NZp=Zp, we say that universal over Zp. If a quadratic form (or m-gonal form) is universal over Zpfor every prime p, then we say that the quadratic form (or m-gonal form) is locally universal . We say that a quadratic form Q(x) isisotropic if the equation Q(x) = 0 has a non-trivial solution otherwise. It is well known that any quadratic form Q(x) of rank n>=5 is isotropic over local ring Zp. This property is critical in the regularity of arbitrary quadratic form of rank n>=5. Any unexplained notation and terminology can be found in of all, we may get an easy but important observation and only allows transfer the issue of the representation by m-gonal form to the issue of the representation by diagonal quadratic form with a congruen ce condition. By using the following proposition, one may effectively determine the rep resentability of integer by an m-gonal form over non-archimedean local ring with help of known results in the quadratic form theory over non-archimedean local r ing.4 DAYOON PARK Proposition 2.1. LetFm(x) =a1Pm(x1) +***+anPm(xn)be a primitive (a1,***,an) = 1)m-gonal form. (1)Whenpis an odd prime with p|m-2,Fm(x)is universal over (mod 4) ,Fm(x)is universal over Z2. (3)Whenpis an odd prime with (p,m-2) = 1, an integer Nis represented by Fm(x)overZpif and only if the integer represented by the diagonal quadratic form (mod 4) , an integer Nis represented by only if the represented by the diagonal quadratic form Proposition 3.1 in [14]. /square Remark 2.2. Now, we may need understanding the behavior of representati on by a diagonal quadratic form which is anisotropic over a non-ar chimedean local ring. (1)First, we consider a quaternary quadratic form which is anis otropic over a non-dyadic (i.e., for an odd prime p) local ring Zp. A quadratic form over Zpis the form and-uand-u'are quadratic One may see that for a quadratic form which is anisotropic over Zp, some each i= 1,2,3,4. (2)A primitive quaternary anisotropic diagonal quadratic for m over dyadic Z2 is the form of 4) 4) 4) (mod 8) . One may get that for a quadratic form which is anisotropic over Z2, some each i= 1,2,3,4. (3)The maximal rank nwhich admits an anisotropic quadratic form over non- archimedean local ring is 4. And a diagonal anisotropic quadratic form over a non-archimedean local ring Zpof rankn(in effect, nshould be 1,2,3, or 4) would be a subform of a diagonal anisotropic quadratic form of rank4. So from the above (1) and (2), one may obtain that for a diagona l REGULAR QUATERNARY m-GONAL FORMS 5 which is anisotropic over a non-archimedean local ring Zp, 2.3. A diagonal quaternary quadratic form ispmax 1<=i<=4ordp(ai)Zp-universal over Zpfor every prime p. Proof.First, we assume that pis an odd prime. By 92:1b in [11], a ternary uni- modular (i.e., u1,u2,u3Zx p) quadratic form and a binary unimodular (i.e., u1,u2Zx p) quadratic form every p-adic unit Zx p. Which may yield the p= 2. uiZx 2, the quadratic over Z2. For two 2-adic units u,u'Zx 2, since uu'(mod 8) ==uu'(Zx 2)2, it would be enough to show the claim only for ui {1,3,5,7}. The tedious pro- cessing is omitted in this article. regular quaternary m-gonal forms Now we are ready to prove our main goals. At the end, in this last sect ion, we prove Theorem 1.1 and Theorem 1.2 by concretely constructing a form of rank 4. Proof of Theorem 1.1. Since the m-gonal form mis almost regular if and only if the m-gonal form almost regular for aN, throughout this proof, we assume that ( a1,a2,a3,a4) = 1. (1) Theorem 4.9 (2) in [1] directly yields the claim. (2) Since the representation over N0implies the representation over Z, it would be enough to show this theorem only over Z. For an odd prime p, if a primitive (i.e., satisfying ( a1,a2,a3,a4) = 1) quadratic formis anisotropicover Zp, then we have that a primitive quadratic form is anisotropicover Z2, we have Therefore, we may see that for for any pTandNN, that is to say, the order of integer 8( is absolutely bounded at every prime spot psuch that Zpis anisotropic. And so Theorem 4.9 (2) in [1] may yield the the second statement, define T':={p|Zpis not universal } (T'could be empty). We show that for any sufficiently large msuch DAYOON PARK ifmdoes not satisfy (1.5), then the m-gonal form mis not almost regular. If there is an odd prime in Twhich does not divide m-2, then we write the prime as qTand otherwise, then we put q= 2T(note that when q= 2, we have m0 (mod 4)). We may take a residue r(q) = For a prime them-gonal form mis universal over Zpby Proposition 2.1. For a prime pT', from our assumption that (a1,a2,a3,a4) = 1, we may take a residue r(p) which N is represented by moverZpfor anyNr(p) (modpordp(4p)). By combining the above arguments with the Chinese Remainder Theorem , we may take a residue that for any integer locally represented by m 8(m-2)N+(m-4)2(a1+a2+a3+a4)0 (mod So N0inthe interval[ that N0is locally represented by mand 8(m-2)N0+(m-4)2(a1+a2+a3+a4)0 (mod since the smallest m-gonal number is m-3 except 0 and 1, the integer N0 in the interval [ would not be (globally) represented bymoverZ(and so over N0too). And then we may get infinitely many positive integers Nnwhich are represented by mlocally, but not globally where 8( (mod 8( m-2)) when qis an odd prime qk1 (modm-2 2) when q= 2. For a contradiction, we suppose that Nnis represented by moverZ, i.e., 8( some ( x1,x2,x3,x4)Z4. Since 8(0 (mod q2kn+ max 1<=i<=4{ordq(64ai)}) andZqis anisotropic, we have (mod qkn+ordq(4)) by Remark 2.2. From our choice of kNsatisfying (3.1), we have REGULAR QUATERNARY m-GONAL FORMS 7 for some yiZ. And then we obtain a contradiction that N0=a1Pm(y1) This completes the 3.1. (1) = 1and a prime q, an m-gonal form a positive integer N, some (x1,***,xn)Z4for which divided by a high power of q, we have that would be divided by a high power of each i= 1,***,nby Remark 2.2. Note that the above situation occurs = 1 orm0 (mod 4) withq= 2 because ord anisotropic So we may getkNfor (mod 8( = 1 qk1 (modm-2 2)whenq= 2. When ord we may see that N'is represented by further a representation of N is derived from a representation of N'. Following the same processing as above, we may back track the representations of the positive integers N > N'> least the contraposition of the above argument, we may see that w hen a ord not represented by which the quadratic form anisotropic), the infinite series of positive integers Nnare not represented by (3.2). On the other hand, when mis large, a quaternary m-gonal form m would fail to represent a bunch of small positive integers si nce the smallest m-gonal number is m-3except0and1. As a specific observation, we suggest an example that a quaternary m-gonal form mwould represent at most 24-1 positive integers before m-3. Therefore there is a lot of chance to take a among the small positive integers) which is re presented by8 DAYOON PARK mlocally, but not globally and satisfy ord Very recently, it was shown that any m-gonal form of rank n>=5is almost regular over N0too (not only over Z) in[15]. The almost regularity of m-gonal rankn>=5is strongly bound to the property that quadratic rankn>=5is always isotropic over every non-archimedean local ring Zp. As the same stream, by Theorem 1.1 (1), when the positive defi nite qua form is isotropic over every non-archimedean local ring Zp, them-gonal form mis almost regular for every it is unknown whether such a principle in the case of quaternary form also holds over N0yet. The author suggests the interesting open problems that when a positive definite quaternary quadratic form is isotropic, to determine that whether the almost regularity of the m-gonal formsm are still preserved or not for m>=3overN0too and to determine that whether the first statement in Theorem 1.1 (2) holds over N0too or not. Proof of Theorem 1.2. Letpbe an odd prime with ( p,m-2) = 1. We may take an odd prime q1which is a quadratic residue modulo pand an odd prime for which-q2is a quadratic non-residue modulo p. Note that it would be possible to take infinitely many such odd primes p,q1andq2. Now we show that the not almost regular over Z(so also over N0). By Proposition 2.1, we may locally universal since the quaternary quadratic locally universal. But the m-gonal form (globally) represent the infinitely many positive (mod 8( m-2)). When n= 0, we may easily see that N0= 1 is not represented by >1. When n >0, for a contradiction, we suppose that Nnis represented by the other words, we suppose that for some ( x1,x2,x3,x4)Z4. Since the anisotropic over we have 1,2,3,4. Moreover, since pkn1 (mod 2( m-2)), we have some yiZ. So we may obtain that yielding that 1 is represented by Which is a contradiction. W. K. Chan, B.-K.Oh, Representations of integral quadratic polynomials , Contemp. Math. 587 (2013), 31-46.

2108.10845

Dayoon Park

Dayoon Park

Non-almost regular quaternary $m$-gonal forms

math.NT

http://creativecommons.org/publicdomain/zero/1.0/

The maximal $n$ which admits a non-almost regular $m$-gonal form of rank $n$ is $4$. In this article, we consider the $m$-gonal forms of rank $4$ which is not almost regular.

exposures. However, collecting the corresponding focal stack of an AiF image might be difficult because of the focus breathing phenomenon, where a camera's field of view changes as the lens moves. To address this problem, we use synthetic data without such effects during training, and apply a calibration process on real data during is trained supervisedly, our method outperforms methods in various comparisons, while our method also runs faster. *To our knowledge, the proposed method is the first that can learn depth estimation unsupervisedly from only AiF images and performs favorably against methods. *Domain gaps can be mitigated by our method with test- time optimization on real-world data, especially when ground truth depth data are not available. 2. Related Work 2.1. Depth from Focus Depth estimation is a fundamental computer vision task that aims to use different cues such as color, or the difference in image sequences to predict or fuse depth maps [ 5,40]. Most of the previous works assume that input images are all-in-focus, whereas in images are usually considered to be defocused in the background or with a shallow depth-of-field (DoF). Nev- ertheless, some approaches elaborate on depth estimation with defocused images. Conventional optimization based approaches [ 37,36] proposed to directly estimate depth from focal stacks, and a variant approach [ 26] tries to generate an index map in which every pixel is assigned to the focus position leading to the maximal sharpness. Chen et al. [6] found the relationship between relative blur and disparity, and make use of it to enhance the robustness of matching. Depth from focus sweep video [ 18] targets estimating depth from images with successive focus positions. Recently, deep learning based approaches [ 11,24] could model the blurri- ness more precisely and achieve much better depth quality. On the other hand, some works [ 2,4,35] use deep to remove the defocus blur for single images. 2.2. Multi-focus Image Fusion Although real-world images usually have defocus blur, most computer vision applications are supposed to coop- erate with all-in-focus (AiF) images. Therefore, more and more attention is given on shallow-DoF image deblurring and multi-focus image fusion. A multi-focus image fusion approach [ 41] applies Laplacian to different scales of images. Zhan et al. [43] proposed a guided filter to help edge preserv- ing during multi-focus image fusion. Nejati et al. [27] learna sparse representation of relative sharpness measurement and produce a pixel-level score map for decision through pooling. An unsupervised learning approach [ 32] was pro- posed to fuse either multi-exposure or multi-focus images to generate a high dynamic range (HDR) or an AiF image. Liuet al. [22] proposed to learn a focus map and a segmen- tation map through deep neural networks, and fuse images by integrating these maps. 2.3. Light Field A light field camera captures spatially distributed light rays. By re-rendering through digitized multi-view images, variable aperture pictures can be generated after capturing. With this characteristic, a post-capture refocusing can also be accomplished by properly manipulating those light field images [ 28]. As light field camera provides multi-view in- formation from different poses, it could help many compu- tational photography applications. As the number of views grows, more information could be retrieved. Light field images often suffer from low spatial resolu- tion as the sampling resources are limited. Cheng et al. [7] and Jin et al. [17] proposed to apply deep learning on con- structing super-resolution images from light field images to cover this situation. Depth estimation could also be achieved by leveraging the abundant information from light field im- ages [ Levoy et al. [20] proposed to use the epipolar plane images (EPIs), which contain both spatial and angular information, for depth esti- mation from light field images. Heber et al. [12] proposed to learn the end-to-end mapping between the 4D light field and its corresponding 4D depth field accompanied with a high- order regularization refinement. By following [ 12], they then designed an encoder-decoder architecture to extract the geometric information from light field images [ 13]. Some works [ 15,33] focus on providing light field datasets with ground truth depth or AiF images for further use. 2.4. Realistic Data Synthesis Deep learning significantly improves the quality of com- puter vision tasks with a large amount of training data. How- ever, collecting real-world data is often costly and time- consuming. Therefore, many works instead target synthe- sizing realistic data to provide sufficient data for training. Barron et al. [3] proposed to synthesize defocus images by a layer-wised rendering scheme with a z-buffer. Wadhwa et al. [39] combine a person segmentation mask with a depth map calculated from a dual-pixel camera to generated shal- low depth-of-field images. Gur et al. [10] use a spread function, namely the PSF convolution layer, to synthesize realistic defocus images and train a depth from de- focus network to generate depth maps. Herrmann et al. [14] proposed a learning-based approach for autofocus and re- cently provided a real dataset for training. However, thedataset was not open to the public at the submission time. 3. Method We describe the proposed method in this section. Sec. 3.1 gives an overview of our method. Then, a shared network used for both of depth estimation and AiF image reconstruc- tion is illustrated in Sec. 3.2. The attention mechanism that bridges the two tasks is introduced in Sec. 3.3. We then depict the core concept to turn AiF images into for unsupervised depth estimation in Sec. 3.4. Fi- nally, Sec. 3.5 shows the loss functions for training with depth supervision and AiF Overview Depth from focus aims to recover the depth from a focal stack through defocus cues. As shown in Fig. 1, given a with gradually varying focus positions P2RHW1F, our method produces several attention representations through a shared network. Then a depth map D2RHW1and an all-in-focus this scene can be generated with these attention maps. 3.2. Network [ 11] uses 2D ConvNets to address the problem, which is one of the main challenges in depth from focus. On the other hand, DefocusNet [ 24] ap- plies a global pooling layer as a communication tool between several weights-sharing 2D ConvNets. Another objective of this architecture is to allow focal stacks with arbitrary sizes. However, some important information across the stack dimension might not be effectively captured due to the limi- tation of 2D convolution and simple global pooling. For this reason, we adopt the Inception3D [ 1] as the backbone of our model. As shown in Fig. 1, our model is an encoder-decoder network consisting of 3D convolu- tions. With the 3D convolution, defocus cues could be better captured across frames and thus facilitate the tasks of depth estimation and AiF image reconstruction. Moreover, our model can also handle focal stacks with arbitrary sizes at- tributed to the nature of 3D Attention Mechanism The output of our network is an intermediate The underlying expectation of the inter- mediate attention Mis that it should reflect the probability of each focus position leading to the maximal sharpness. Then it can benefit both of depth estimation and AiF depth estimation, we propose to normalize the in- termediate attention Minto a depth attention Mdepthvia asoftplus (1 + exp ( (1 + exp ( Mi;j;1;n)): (2) The softplus function is a smooth version of ReLU. The depth attention Mdepthcan also be interpreted as the probability distribution because the softplus function ensures non-negativeness and normalizes Mdepthinto a distribution. Then the expected depth value of each pixel can be derived via: Di;j;1= (3) For AiF image reconstruction, we perform similar proce- dures except that the normalization function changes from softplus to softmax. That is, the AiF attention MAiFis ob- tained as: MAiF=(M); (Mi;j;1;n): (5) The AiF attention MAiFcan then be used for AiF The expected AiF image of each pixel is expressed as: Ii;j;k= (6) The reason why we adopt the softmax and softplus nor- malizations separately for depth and AiF image estimation is to tackle the problem of sparse focal stacks. By sparse, we mean that the stack size is small, and the focus positions inside the focal stack are not dense. For AiF image reconstruction, it is best to select the sharpest pixel along the stack dimension. Blending multiple pixels inside a sparse stack usually does not help. There- fore, we leverage the softmax function to pursue the to extract the clearest pixels. For depth estimation, the softmax normalization results in severe quantization for sparse focal stacks because it simply selects the nearest focal position of maximal sharpness. On the other hand, the softplus normalization leads to a so that more accurate depth can be predicted by interpolating among sparse focal positions. Fig. 2 shows the effects of this design stack SAll-in-focus IDepthD3D conv Figure 1: An overview of the proposed method. Given a stack of images with varying focus positions, i.e., focal stack, our model first produces an intermediate attention map M. The intermediate attention map can be shared between depth estimation and all-in-focus (AiF) image reconstruction. With different normalization functions, the attention map can be further manipulated to generate either depth or AiF 1 2 3 4 5 6 7 8 9 Stack Attention AiF Attention 0 1 2 3 4 5 6 7 8 9 Stack Seperate 0 1 2 3 4 5 6 7 8 9 Stack 0 1 2 3 4 5 6 7 8 9 Stack Figure 2: The effect of normalization on attention maps. The first row shows that if we adopt the same softmax nor- malization for depth and AiF attention, both of them tend to be a flatter distribution in supervised learning and face a peaking phenomenon in unsupervised learning (AiF su- pervision). The second row shows that with separate nor- malization as described in Sec. 3.3, each of depth and AiF attention becomes a proper distribution in both supervised and unsupervised learning. 3.4. Towards Unsupervised Depth to our method, DefocusNet [ 24] also combines the tasks of depth estimation and AiF image defocus maps . Their intermediate defocus maps necessitate supervisory signals, which can be derived by calculating the circle of confusions from the ground truth depth. Furthermore, they also propose to use a DepthNet fol- lowing the intermediate defocus maps to predict the output depth. Therefore, their method can only be trained supervis- edly when ground truth depth data are we propose to use the intermediate attention Mtobridge the two tasks, which does not necessitate signals from ground truth depth data. Moreover, there are no learnable parameters after M. We only use functions to generate the output depth and AiF images. Therefore, even when ground truth depth data are not available, we can still train the shared network to generate the intermediate attention Monly via supervisory signals from ground truth AiF images. That is, our method can be trained supervisedly or unsupvervisedly with or without ground truth depth data. 3.5. Training Loss For supervised depth estimation , our model is trained with a simple for the ground truth depth. For unsupervised depth estimation, our model can also be trained by a L1loss with AiF the ground truth AiF image. Furthermore, we also encourage our depth map to be locally smooth using an edge-aware weighting as in [ 9]. The smoothness loss is defined as: Lsmooth (9) where Wx= exp exp a hyper-parameter for the edge weighting based on the ground truth AiF image. The total loss of our unsuper- vised depth estimation is =LAiF+L smooth; the importance of the smoothness loss. 4. Evaluation In this section, we describe the datasets used in the ex- periments, and report ablation studies and comparisons with the state-of-the-art methods on depth from focus. In material, we present additional results for AiF image reconstruction, and visual comparisons on DDFF 12-Scene [ 11], Middlebury Stereo Datasets [ 34], and the DefocusNet dataset [24]. 4.1. Datasets Totally five datasets are used in the quantitative experi- ments and visual comparisons. Their descriptions and statis- tics are summarized in Table 1. For more details, please see the supplementary material. 4.2. Evaluation Metrics In this paper, we evaluate quantitative results with the fol- lowing metrics: mean-absolute error (MAE), (MSE), error (RMSE), log error (logRMS), error (Abs. rel.), relative-squared error (Sqr. rel.), bumpiness (Bump), accuracy with = 1:25, per-stack inference time (Secs.). 4.3. Implementation Details We implement our method in PyTorch [ 30]. Our model is trained from scratch using the Adam optimizer [ 19] (1= 0:9,2= 0:999), with a learning rate of 104. For DDFF 12-Scene [ 11], 4D Light Field Dataset [ 15], FlyingTh- ings3D [ 25], and the DefocusNet dataset [ 24], we use 10, 10, 15, and 5 as the input stack sizes, respectively. In un- supervised learning, the weight of smoothness loss is set to0:002in all experiments. We apply random spatial (flipping, cropping, rotation) and random color jittering (brightness, contrast, and gamma) for data augmen- tation during training. The training patch size is 256256 after random cropping except for DDFF 12-Scene [ 11]. Be- cause the image size of DDFF 12-Scene is 224224, which is larger than 256256, we do not apply random cropping on this dataset. All the experiments are conducted with a single NVIDIA GTX 1080 GPU. 4.4. Ablation Studies To understand the performance of each proposed com- ponents in our method, we conduct various ablation studies Stack Size of Test TimeDisparity 5 910 Same ArbitraryFigure 3: The effect of focal stack size. Each line indicates a stack size setting during training. Specifically, 10and Arbitrary respectively represent training with a fixed size of 10and training with arbitrary sizes, and Same refers to using the same setting as test stacks. on 4D Light Field Dataset [ 16]. All models are with disparity maps if not Table 2 shows the comparison between 2D and 3D convolutions and with/without the proposed atten- tion mechanism. The 3D convolution performs better than 2D convolution as the 3D convolution is able to capture and aggregate features across both spatial and stack dimensions. The design of attention bridges the tasks of depth estimation and AiF image reconstruction, and thus enables the possibil- ity of unsupervised depth estimation as described in Sec. 3.4 while maintaining roughly the same prediction quality. Focal stack size. Our network can handle arbitrary input stack sizes due to the nature of 3D convolution. Fig. 3 shows a comparison on different input stack sizes. We train the models in three different ways: 1) 10: training with a fixed input stack size of 10. 2) Arbitrary : training with arbitrary input stack sizes randomly sampled from 2 to 10. 3) Same : training with the same size as test stacks. The model trained with a fixed stack size performs poorly while testing with different stack sizes. As expected, the model trained with the same stack size as test data performs the best as the input setting for training and testing are consis- tent. However, the model trained with arbitrary input stack sizes performs favorably against the ones trained with the same stack size as test data. This demonstrates the robustness of our method across different input stack sizes. Fig. 4 shows the visual comparison of different input stack sizes, generated with the model trained with arbitrary input stack sizes. It is obvious that the quality of the map improves as the stack size learning using all-in-focus images. On the basis of the attention mechanism described in Sec. 3.3, our method is able to be trained with either ground truth dispar- ity maps (supervised) or ground truth AiF images Table 3 and Fig. 5 respectively show the quantitative and qualitative results for both of our supervised and unsu-Table 1: Summary of the evaluation Image source Cause of defocus Disparity/Depth GT AiF GT DefocusNet [24] Synthetic Blender rendering Depth DDFF 12-Scene [11] Real Light-field composition Depth 4D Light Field Dataset [16] Synthetic Light-field composition Disparity X Middlebury Stereo Datasets [34] Real Disparity rendering Disparity X Mobile Depth [37] Real Real Stitching by MRF Table 2: Ablation study on network architecture. The 3D convolution leads to better performance on disparity estima- tion because of its ability to capture features for both spatial and stack dimensions. Although the design of attention does not increase the performance explicitly, it bridges the tasks of depth estimation and AiF image reconstruction and thus enables the unsupervised learning for depth Attention MAE # MSE # RMSE #Bump. # 3DX 0.0788 0.0472 0.2014 1.5776 0.0851 0.0461 0.1984 2.3168 2DX 0.1070 0.0577 0.2259 2.1435 0.1179 0.0576 0.2291 2.3168 RGB stack size 3 stack size 5 stack size 9 GT Figure 4: Visual comparison on different focal stack sizes. The quality of the estimated disparity map improves as the stack size models. As shown in Table 3, with a large input stack size of 37and the smoothness loss, the performance of unsupervised learning is able to approximate that of super- vised learning with a smaller stack size of 10. However, the results of unsupervised learning with a smaller stack size suffer from the quantization problem as described in Sec. 3.4. As shown in Fig. 5, after adding the smoothness loss, the output disparity map becomes locally smooth and perform better Comparisons to the State-of-the-art Methods After ablation on different components of the proposed method, we conduct comparisons with the on various datasets in this section. DDFF 12-Scene. Table 4 shows the quantitative compari- son on DDFF 12-Scene [ 11]. All methods are on this dataset with ground truth depth. The pro- posed method performs favorably against the 3: Ablation study on supervision. The results of supervised learning perform better than the ones from unsu- pervised learning (AiF supervision). Unsupervised learning often generates disparity maps that suffer from the quan- tization effect and lead to poor results. After adding the smoothness loss, the output disparity maps become locally smooth and perform better Stack size Lsmooth MAE # MSE # RMSE # Yes 10 0.0788 0.0472 0.2014 No 10 0.2425 0.1174 0.3401 No 37 0.2099 0.1039 0.3202 No 10 X 0.1671 0.0746 0.2698 No 37 X 0.1116 0.0584 Yes No No No No Stack size 10 37 10 37 Lsmooth X X Figure 5: Visual comparison on different supervision set- tings. The results of supervised learning are better than the ones from unsupervised learning (AiF supervision). The out- put disparity maps from unsupervised learning suffer from the quantization effect. By adding the smoothness loss, the disparity results become locally smooth and perform favor- ably against the ones from supervised learning under all metrics. Furthermore, our method runs faster than another deep learning based method DDFF [ 11] . 4D Light Field Dataset. Compared to DDFF 12-Scene [ 11], 4D Light Field Dataset [ 16] provides the ability to simu- late shallower DoF images as its larger baseline leads to a larger synthetic aperture. Table 5 shows the with state-of-the-art methods. All of them are trained with this dataset. Our supervised model achieves the best MSE (0.0472), which outperforms the DefocusNet [ 24] (0.0593). Furthermore, our unsu- pervised model even performs better than most of the otherTable 4: Quantitative comparison on DDFF 12-Scene. Note that RMSE is not presented due to the mismatched results from the benchmark website and the paper of DDFF [ 11]. For DefocusNet [ 24], we only show the original metric reported in their paper, which is MSE. Red RedRed text indicates the best, and blue text indicates the second-best performing method. Method MSE # log RMS #Abs. rel. #Sqr. rel. #Bump. #= 1:25= 1:252= 1:253 Ours 0.25 0.250.25 0.01 0.010.01 0.63 68.33 68.3368.33 87.40 87.4087.40 93.96 DefocusNet [24] 9:1e4- - - - - - - DDFF [11] 9:7e40.32 0.29 0.01 0.010.01 0.59 61.95 85.14 92.98 PSPNet [44] 0.27 0.01 0.010.01 0.55 62.66 85.90 2:1e30.31 0.26 0.010.010.01 1.02 55.65 82.00 93.09 PSP-LF [44] 2:7e30.45 0.46 0.03 0.540.540.54 39.70 65.56 82.46 DFLF [11] 4:8e30.59 0.72 0.07 0.65 28.64 53.55 71.61 VDFF [26] 7:3e31.39 0.62 0.05 0.79 8.42 19.95 32.68 RGB VDFF [26] PSPNet [44] DDFF [11] DefocusNet [24] Ours (S) Ours (US) GT Figure 6: Visual comparison on 4D Light Field Dataset. S: supervised. US: unsupervised 5: Quantitative comparison on 4D Light Field Dataset. Our supervised model outperforms the method DefocusNet. Furthermore, our (with AiF supervision) even performs better than most of the other supervised methods. Please see Fig. 6 for the vi- sual comparison. (represents that the model is pre-trained on DDFF Supervised MSE #RMSE #Bump. # Ours Yes 0.2014 0.20140.2014 1.58 DefocusNet [24] Yes 0.0593 0.2355 2.69 *DDFF [11] Yes 0.19 0.42 1.92 *PSPNet [44] Yes 0.37 0.53 1.21 1.211.21 VDFF [26] Yes 1.3 1.15 1.58 Ours No 0.0746 0.2398 2.58 supervised methods, except for DefocusNet. Fig. 6 shows the qualitative results of 4D Light Field Dataset [ 16]. Our supervised model delivers sharper depth boundaries and less noise in textureless regions. Meanwhile, our unsupervised model also achieves visually with the supervised DefocusNet The DefocusNet dataset [ 24] is a syntheticTable 6: Quantitative comparison on the Our supervised model outperforms the method DefocusNet. Please refer to the for the full metric results. Method 0.0549 0.0127 [24] 0.0637 0.0175 0.1207 dataset using physically based rendering (PBR) shaders. Since only a subset of this dataset has been released and there are no AiF images, we can only conduct on the provided subset. The quantitative results of our method and DefocusNet [ 24] are shown in Table 6. Note that the results of DefocusNet are retrained on the provided subset with the released code and setting. Mobile Depth. The Mobile Depth dataset [ 37] is a real- world dataset of focal stacks captured with mobile phone and cameras. Due to the lack of training data in this dataset, all compared models are trained on FlyingThings3D [ 25], and the input focal stacks are synthesized by the used in Barron et al. [3]. As shown in Fig. 7, the output depth maps of our super-Table 7: Analysis on generalization ability across different datasets. We train our models and DefocusNet [ 24] [ 25] with synthesized focal stacks by the rendering technique used in Barron et al. [3], and test these models on Middlebury Stereo Datasets [ 34] and the DefocusNet dataset. The results show that both of our supervised and AiF supervision) models have better generalization ability than Supervised Test Dataset MAE # MSE # RMSE # absRel # sqrRel # Sec.# Ours Yes Middlebury 58.5698 58.569858.5698 5.9355 5.93555.9355 2.4776 2.47762.4776 45.5886 45.588645.5886 0.0287 Ours No Middlebury 5.4499 99.6029 8.2606 3.8954 80.7120 0.0283 [24] Yes Middlebury 7.4084 157.4397 9.0794 3.4698 63.6797 0.3798 Ours Yes DefocusNet 0.1827 0.079 5 0.2607 72.4664 40.4281 0.021 Ours No DefocusNet 0.0627 0.06270.0627 0.2380 0.23800.2380 59.4459 59.445959.4459 14.0227 14.022714.0227 0.01959 [24] Yes DefocusNet 0.3200 0.1478 0.3722 138.2917 70.0229 0.0527 vised model are smoother and exhibit less ambiguity than DefocusNet [ 24]. But the results of our unsupervised model (with AiF supervision) are worse than DefocusNet. Nonethe- less, since our method allows training without ground truth depth, we can perform test-time optimization on this dataset with only AiF images. After test-time optimization, our unsupervised model performs better than DefocusNet, and both of our supervised and unsupervised models against Mobile Depth. We also show the output AiF images in the supplementary ability analysis. Middlebury Stereo Datasets[ 34] is a real-world dataset of stereo images along with ground truth disparity maps. We use this dataset to analyze the generalization ability of our method. To this end, we firstly let our models and DefocusNet [ 24] be trained on FlyingThings3D [ 25] with synthesized focal stacks by the rendering technique used in Barron et al. [3], and then test these models on Middlebury Stereo Datasets as well as the DefocusNet dataset. The quantitative results are shown in Table 7. One can see that both of our supervised and unsupervised (with AiF supervision) models achieve than DefocusNet on these two time. As indicated in Table 7, our method is faster than the state-of-the-art method DefocusNet [ 24]. The main reason might be that our model does not have a bottleneck like the global pooling layer in DefocusNet. 5. Conclusion We have proposed a method to jointly estimate the depth map and the all-in-focus (AiF) image from an input focal stack with a shared network. By the design of the mechanism, the shared network can be trained either supervisedly, or unsupervisedly with AiF images. Our method can further mitigate domain gaps even in the absence of ground truth depth. Experimental results show that our method outperforms the state-of-the-art methods while also having higher efficiency in inference time. AiF GT DefocusNet [24] Ours (S) Ours (US) Mobile [37] Ours (S) * Ours (US) * AiF GT DefocusNet [24] Ours (S) Ours (US) Mobile [37] Ours (S) * Ours (US) * Figure 7: Visual comparison on the Mobile Depth dataset. With test-time optimization, our models perform better than DefocusNet [ 24] and favorably against Mobile Depth [ 37] qualitatively. (S: supervised. US: supervision). *: test-time Alayrac, Jo ~ao Carreira, and Andrew Zisserman. The visual centrifuge: Model-free layered video In IEEE Conference on Computer Vision and CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pages 2457-2466. Computer Vision Foundation / IEEE, 2019. 3 [2]Saeed Anwar, Zeeshan Hayder, and Fatih Porikli. Depth estimation and blur removal from a single out-of-focus image. InBritish Machine Vision Conference 2017, BMVC 2017, London, UK, September 4-7, 2017 . BMV A Press, 2017. 2 [3]Jonathan T. Barron, Andrew Adams, YiChang Shih, and Car- los Hern 'andez. Fast bilateral-space stereo for synthetic de- focus. In IEEE Conference on Computer Vision and CVPR 2015, Boston, MA, USA, June 7-12, 2015 , pages 4466-4474. IEEE Computer Society, 2015. 2, 7, 8 [4]Marcela Carvalho, Bertrand Le Saux, Pauline Trouv 'e-Peloux, Andr 'es Almansa, and Fr 'ed'eric Champagnat. On for deep depth estimation. In 2018 IEEE on Image Processing, ICIP 2018, Athens, Greece, October 7-10, 2018 , pages 2915-2919. IEEE, 2018. 2 [5]Yu-Lin Chang, Wei-Yin Chen, Jing-Ying Chang, Yi-Min Tsai, Chia-Lin Lee, and Liang-Gee Chen. Priority depth fusion for the 2d to 3d conversion system. In Brian D. Corner, Masaaki Mochimaru, and Robert Sitnik, editors, Proceedings of the Conference on Image Capture and Applications 2008, San Jose, CA, USA, January 28-29, 2008 , volume 6805 of SPIE Proceedings , page 680513. SPIE, Chen, Hui Zhou, and Timo Ahonen. estimation from defocus stereo images. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015 , pages 855-863. IEEE Computer Society, 2015. 2 [7]Zhen Cheng, Zhiwei Xiong, Chang Chen, and Dong Liu. Light field A benchmark. In IEEE Confer- ence on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2019, Long Beach, CA, USA, June 16-20, 2019 , pages 1804-1813. Computer Vision Foundation / IEEE, 2019. 2 [8]Donald G. Dansereau, Bernd Girod, and Gordon Wetzstein. Liff: Light field features in scale and depth. In IEEE Con- ference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pages 8042- 8051. Computer Vision Foundation / IEEE, 2019. 2 [9]Cl'ement Godard, Oisin Mac Aodha, and Gabriel J. Bros- tow. Unsupervised monocular depth estimation with In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pages 6602-6611. IEEE Computer Society, 2017. 1, 4 [10] Shir Gur and Lior Wolf. Single image depth estimation trained via depth from defocus cues. In IEEE Conference on Com- puter Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019 , pages 7683-7692. Com- puter Vision Foundation / IEEE, 2019. 2[11] Caner Hazirbas, Sebastian Georg Soyer, Maximilian Christian Staab, Laura Leal-Taix 'e, and Daniel Cremers. Deep depth from focus. In C. V . Jawahar, Hongdong Li, Greg Mori, and Konrad Schindler, editors, Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, 2-6, 2018, Revised Selected Papers, Part III , vol- ume 11363 of Lecture Notes in Computer Science , pages 525-541. Springer, 2018. 1, 2, 3, 5, 6, 7 [12] Stefan Heber and Thomas Pock. Convolutional networks for shape from light field. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016 , pages 3746-3754. IEEE 2016. 2 [13] Stefan Heber, Wei Yu, and Thomas Pock. Neural for shape from light field. In IEEE on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 , pages 2271-2279. IEEE 2017. 2 [14] Charles Herrmann, Richard Strong Bowen, Neal Wadhwa, Rahul Garg, Qiurui He, Jonathan T Barron, and Ramin Zabih. Learning to autofocus. In Proceedings of the on Computer Vision and Pattern Recognition , pages 2230-2239, 2020. 2 [15] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Shang-Hong Lai, Vincent Lepetit, Ko Nishino, and Yoichi Sato, editors, Com- puter Vision - ACCV 2016 - 13th Asian Conference on Com- puter Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part III , volume 10113 of Lecture Notes in Computer Science , pages 19-34. Springer, 2016. 2, 5 [16] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Asian Conference on Computer Vision , pages 19-34. Springer, 2016. 5, 6, 7 [17] Jing Jin, Junhui Hou, Jie Chen, and Sam Kwong. Light field spatial super-resolution via deep combinatorial geome- try embedding and structural consistency regularization. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pages 2257-2266. IEEE, 2020. 2 [18] Hyeongwoo Kim, Christian Richardt, and Christian Theobalt. Video In Fourth International Confer- ence on 3D Vision, 3DV 2016, Stanford, CA, USA, October 25-28, 2016 , pages 370-379. IEEE Computer Society, 2016. 2 [19] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015. 5 [20] Marc Levoy and Pat Hanrahan. Light field rendering. In John Fujii, editor, Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996 , pages 31-42. ACM, 1996. 2 [21] Haiting Lin, Can Chen, Sing Bing Kang, and Jingyi Yu. Depth recovery from light field using focal stack symmetry. In Pro-ceedings of the IEEE International Conference on Computer Vision , pages 3451-3459, 2015. 2 [22] Yu Liu, Xun Chen, Hu Peng, and Zengfu Wang. Multi- focus image fusion with a deep convolutional neural Fusion , 36:191-207, 2017. 2 [23] Maxim Maximov, Kevin Galim, and Laura Leal-Taix 'e. Fo- cus on defocus: bridging the synthetic to real domain gap for depth estimation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , pages 1071-1080, 2020. 1 [24] Maxim Maximov, Kevin Galim, and Laura Leal-Taix 'e. Focus on defocus: Bridging the synthetic to real domain gap for depth estimation. In 2020 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pages 1068-1077. IEEE, 2020. 2, 3, 4, 5, 6, 7, 8 [25] N. Mayer, E. Ilg, P. H ausser, P. Fischer, D. Cremers, A. Doso- vitskiy, and T. Brox. A large dataset to train for disparity, optical flow, and scene flow International Conference on Computer Vision and Pattern Recognition (CVPR) , 2016. 5, 7, 8 [26] Michael M oller, Martin Benning, Carola Sch onlieb, and Daniel Cremers. Variational depth from focus IEEE Trans. Image Process. , 2015. 1, 2, 7 [27] M. Nejati, S. Samavi, and S. Shirani. Multi-focus image fu- sion using dictionary-based sparse representation. Inf. Fusion , 25:72-84, 2015. 2 [28] Ren Ng, Marc Levoy, Mathieu Bredif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. In Stanford University Com- puter Science Tech Report CSTR 2005-02 . Stanford Univer- sity, 2005. 2 [29] In Kyu Park, Kyoung Mu Lee, et al. Robust light field depth estimation using occlusion-noise aware data costs. on pattern analysis and machine intelligence 2017. 2 [30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, An- dreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An im- perative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d 'Alch 'e-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Systems 32 , pages 8024-8035. Curran 2019. 5 [31] Jiayong Peng, Zhiwei Xiong, Yicheng Wang, Yueyi Zhang, and Dong Liu. Zero-shot depth estimation from light field using a convolutional neural network. IEEE Transactions on Computational Imaging , 6:682-696, 2020. 2 [32] K. Ram Prabhakar, V . Sai Srikar, and R. Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In IEEE on Computer Vision, ICCV 2017, Venice, Italy,October 22-29, 2017 , pages 4724-4732. IEEE 2017. 2 [33] Parikshit Sakurikar, Ishit Mehta, Vineeth N. P. J. Narayanan. Refocusgan: Scene refocusing using a single image. In Vittorio Ferrari, Martial Hebert, and Yair Weiss, editors, European Conference on Computer Vision ECCV 2016: Computer Vision - ECCV 2016 , pages 842-857. Springer, 2016. 2 [34] Daniel Scharstein, Heiko Hirschm uller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High- resolution stereo datasets with ground truth. In Xiaoyi Jiang, Joachim Hornegger, and Reinhard Koch, edi- tors, Pattern Recognition - 36th German Conference, GCPR 2014, M unster, Germany, September 2-5, 2014, Proceedings , volume 8753 of Lecture Notes in Computer Science , pages 31-42. Springer, 2014. 5, 6, 8 [35] Pratul P. Srinivasan, Rahul Garg, Neal Wadhwa, Ren Ng, and Jonathan T. Barron. Aperture supervision for monocular In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 , pages 6393-6401. IEEE 2018. 2 [36] Jaeheung Surh, Hae-Gon Jeon, Yunwon Park, Sunghoon Im, Hyowon Ha, and In So Kweon. Noise robust depth from focus using a ring difference filter. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Hon- olulu, HI, USA, July 21-26, 2017 , pages 2444-2453. IEEE Computer Society, 2017. 1, 2 [37] Supasorn Suwajanakorn, Carlos Hern 'andez, and Steven M. Seitz. Depth from focus with your mobile phone. In IEEE Conference on Computer Vision and Pattern 2015, Boston, MA, USA, June 7-12, 2015 , pages 3497- 3506. IEEE Computer Society, 2015. 1, 2, 6, 7, 8 [38] Yu-Ju Tsai, Yu-Lun Liu, Ming Ouhyoung, and Yung-Yu Chuang. Attention-based view selection networks for light- field disparity estimation. In The Thirty-Fourth AAAI Confer- ence on Artificial Intelligence, AAAI 2020, The Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Ad- vances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020 , pages 12095-12103. AAAI Press, 2020. 2 [39] Neal Wadhwa, Rahul Garg, David E. Jacobs, Bryan E. Feld- man, Nori Kanazawa, Robert Carroll, Yair T. Barron, Yael Pritch, and Marc Levoy. with a single-camera mobile phone. ACM Trans. Graph. , 2018. 2 [40] Ning-Hsu Wang, Bolivar Solarte, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. 360sd-net: 360 degstereo depth estimation with learnable cost volume. In 2020 IEEE on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020 , pages 582-588. IEEE, 2020. 2 [41] Wencheng Wang and Faliang Chang. A multi-focus im- age fusion method based on laplacian pyramid. J. Comput. 2011. 2 [42] Junyuan Xie, Ross B. Girshick, and Ali Farhadi. Deep3d: Fully automatic 2d-to-3d video conversion with deep convo-

2108.10843

Yu-Lun Liu

Ning-Hsu Wang, Ren Wang, Yu-Lun Liu, Yu-Hao Huang, Yu-Lin Chang, Chia-Ping Chen and Kevin Jou

Bridging Unsupervised and Supervised Depth from Focus via All-in-Focus Supervision

ICCV 2021. Project page: https://albert100121.github.io/AiFDepthNet/ Code: https://github.com/albert100121/AiFDepthNet

cs.CV

http://creativecommons.org/licenses/by/4.0/

Depth estimation is a long-lasting yet important task in computer vision. Most of the previous works try to estimate depth from input images and assume images are all-in-focus (AiF), which is less common in real-world applications. On the other hand, a few works take defocus blur into account and consider it as another cue for depth estimation. In this paper, we propose a method to estimate not only a depth map but an AiF image from a set of images with different focus positions (known as a focal stack). We design a to exploit the relationship between depth and AiF estimation. As a result, the proposed method can be trained either supervisedly with ground truth depth, or with AiF images as supervisory signals. We show in various experiments that our method outperforms the both quantitatively and qualitatively, and also has higher efficiency in inference time.

and geometric Brownian motion which enables pricing of financial derivatives [3]. This con- nection also plays a crucial role in the rigorous and constructive formulation of quantum field theories, which are essential for the description of nature at the most [4, 5]. On a more practical level, this formula is routinely utilized for the numer- ical evaluation of the quantum properties of atomic and subatomic systems in as well as in atomic and particle physics [6, 7, 8]. Recent research on the suitability of quantum computers for solving differential equa- tions has grown substantially due to the promise of efficiency and speed-up brought by solving these problems quantum mechanically. In [9, SSI], the link between the Feynman- Kac formula and stochastic processes is pointed out but not further developed. Similarly, on [10,SSII], a remark about a 'deep' connection between the heat equation and is made. In [11, SS2.1], Feynman-Kac formula is invoked to transform the price of a European call option from a conditional expectation value to the solution of the Wick- rotated Schrodinger equation through a particular Hamiltonian. Furthermore, in [12], the connection between trinomial trees for random walks, SDEs, and the Fokker-Planck equa- tion, is established but the full picture of generalizing to the Feynman-Kac formula is not developed. In this paper, we will explore the advantages of entirely transitioning from an SDE to a PDE by not relying on a random-walk tree or other kinds of discrete alternative approach to estimate properties of stochastic processes is by Amplitude Estimation (QAE) [13] and its variants [14, 15, 16], which can achieve a quadratic speed-up over classical Monte Carlo simulation. In contrast to classical Monte Carlo simulation, all possible paths of a (discretized) stochastic process are modelled in a single quantum state in superposition. In other words, a set of qubits is prepared such that measuring their state would result in sampling from the distribution of the of interest. Although preparing such states is in principle always possible for reasonable stochastic processes [17], efficient realization of this method demands a careful analysis [18, 19, 20, 21] and may not always result in a practical quantum advantage. Once such a state is prepared, QAE can be used to evaluate different (path-dependent) In [22], a similar connection between SDEs and PDEs is developed focused on Monte Carlo methods rather than variational methods associated to PDEs. In most of the existing literature on option pricing for equities using quantum com- puters, e.g. [18, 23], however, an SDE is tacitly solved, namely the so-called motion whereuandsare constants known as drift and volatility respectively, and Wtis the Brownian motion. Once this SDE is solved and shown to which produces a log-normal the pricing of a particular security begins by applying QAE. However, this SDE is one of the simplest and one of the few that can be solved analytically. For the most part SDEs rarely admit analytical solutions [24, 25] and are usually dealt with numerically with large amounts of classical resources. Here we bypass this problem by not solving the SDE and instead simulating an associated - and deterministic - PDE for the calculations of the conditional expectations of the stochastic process Xt. We now propose to synthesize these ideas into a unified framework and bring econ- omy of thought by describing how the heat, Schrodinger, Black-Scholes, equations can be thought of as different manifestations of the Feynman- Kac formulation. Once this framework is established, we then employ recent techniques from variational quantum time evolution [26, 27, 28] to introduce an algorithm for of SDEs on quantum computers by firstlinking them to PDEs and a differential operator, an infinitesimal generator, that can be evolved on a Accepted in click title to verify. Published under CC-BY 4.0. 2quantum computer. The paper is structured as follows. In Section 2, we introduce the Feynman-Kac for- mula. The algorithm along with its complexity and advantages is developed in Section 3 and numerically illustrated in Section 4. Future exploratory items are brought forward in Appendix H. 2 The Feynman-Kac formula In this section we first introduce the definition and existence of Brownian motion and immediately use its properties to derive a partial differential equation whose solution will be a conditional expectation of the Brownian motion. We formalize these results in one of second order PDEs is also provided. Moreover, we establish the that will be needed in Section 4 for our quantum and classical algorithms with periodic boundary conditions. 2.1 Brownian motion and differential equations We start by reminding the reader about Brownian motion and link it to the heat equation. We shall proceed axiomatically in order to provide all the tools needed throughout the paper. Suppose ohmis a sample space, oohmis an event of this space and N(u,s)is the normal distribution with mean uand variance s2, 2.1. Brownian motion a continuous that satisfies the following characteristics [29, SS5.3]: 1)W(0) = 0almost independent increments: W(t+u)-W(t)is independent of s(W(s) :s<=t) foru>=0, wheresdenotes the sigma algebra generated by W(s); 3)W(t)has Gaussian increments: W(t+u)-W(t)is normally distributed with mean 0and variance u, continuous paths: W(t)is a continuous function of t, in tfor alloohm. The probability measure under which Wtis a Brownian motion is denoted by Pand the filtered probability space will be denoted by (ohm,F,P). The existence of Brownian motion - often overlooked in the literature of quantum computing, e.g. [12, SSII.A], [30,SSII.A] or [23,SSII] - was established by Wiener by proving that indeed the first three items were compatible with the continuity of the paths of Wt. Theorem 2.1 motion exists. In Figure 1 we display some possible paths of Brownian motion and their distribution. A key result, which is best motivated informally, is due to Ito and it states the and [32, Appendix B], see also [33, 29, 34]. Theorem 2.2 (Ito's lemma) .For infinitesimal Brownian motion increments, one has that dW2 t=dt. (1) Accepted in click title to verify. Published under CC-BY 4.0. 30.2 0.4 0.6 0.8 1: Left : 10 realizations of Brownian motion. Right : 500 realizations of Brownian motion and their resulting result states that the variance of the Brownian increment is proportional to the time interval of the increment. In other words, Brownian motion accumulates at rate one per unit Brownian motion has some unusual properties. For instance while it is continuous everywhere it is (with probability one) nowhere differentiable and it scales like a fractal. Moreover, Brownian motion will eventually attain any and every value no matter how large or negative and once Brownian motion attains a value, it immediately attains it again infinitely often, and then after that again from time to time in the future, see [33,SS3.1] for other items in the 'bestiary' of Brownian motion. Globally, the law of large numbers states that Wt/t-0almost surely as t-, [5, Proposition 2.8]. Next, we move on to the backward heat equation [32, SS3.1.9] as an illustrative exam- ple. Suppose that (Wt)t>=0is a Brownian motion defined on the filtered probability space (ohm,F,P), and consider the simplest type of SDE dXt=dWt. (2) Letps:R-Rbe a Borel-measurable function such (3) for somea>0. Set the conditional (4) From the properties of Brownian motion we can write uas u(x,t) (5) for every (x,y,t )R2x[0,T]. Taking the partial derivative with respect to tand the second order partial derivative with respect to x, one can easily verify that usatisfies (6) with the terminal condition u(x,T) =ps(x)forxR. This result contains the essence of Feynman-Kac formulation as it expresses the solution to the parabolic PDE (6)as Accepted in click title to verify. Published under CC-BY 4.0. 4the expected value of a functional of Brownian motion (4)coming from an SDE (2) with a terminal condition that is subjected to (3). While (6)is indeed the heat equa- tion, some guesswork would be involved in generalizing this to a full SDE of the let alone to generalizations. In order to overcome this, it is worth investing in seeing how Ito's lemma plays a critical role in the derivation of the Feynman-Kac formula. This investment will also explain how (5)was arrived at. The details are in Appendices A and B. 2.2 One-dimensional Feynman-Kac formula As we have seen, the Feynman-Kac technique is a strategy for solving a certain type of partial differential equations by using Brownian motion and Monte Carlo methods. The converse can also be quite useful as it implies that expectations of Wiener processes such as(2)or more general, which are stochastic, can be translated to and computed by methods. Let us now formalize these a one-dimensional Brownian motion defined on a filtered (ohm,F,P). We shall consider the Wiener process Wtand the stochastic (7) where aandbare suitable analytic functions with a>0and with condition Xt=x, see [31,SS4]. Note how (7)now incorporates a drift term b(Xt,t), unlike(2). Let f:R-Rbe an analytic function. For a Borel-measurable function ps:R-R, we define the setting u(x,t) (8) where cis a suitable real function called the discount function. Moreover, suppose somed>0. The function uis defined for 0<T-t<1 2dandxR, and has derivatives of all orders. In particular, it belongs to the class C2,1(Rx(0,T)). The Feynman-Kac result is the fact that usatisfies the following partial differential = 0,fort<T u(x,T) =ps(x).(9) From this point, we shall drop the term involving fby setting it to zero as it will not be very useful in our formulation. In this case, equation =ps(x).(10) and its associated conditional expectation (8)reduces to u(x,t) (11) Accepted in click title to verify. Published under CC-BY 4.0. 5Shortly it will become clear that the termu xis not always helpful for our objectives and it is advantageous to remove it now by applying changes of variables and we set u(x,t) a function of aandbgiven(49)in Appendix C. Moreover, we switch to an initial value problem by setting t=T-tand then perform a Wick rotation x=-it, whereiis the imaginary unit. The last step requires This transforms the problem the new term wcan be written in terms of the other coefficients a,bandcof the PDE by the formula w(x,x) shall now write this in Dirac notation (12) where the wave function is written in Dirac notation as ps(x,t) it takes the role of the modified solution in the setting of an infinite dimensional Hilbert space. The resulting function not necessarily correspond to an wavefunction and ^His not necessarily Hermitian. The wave function not to be confused with the boundary condition psfrom(11). In this case, the Hamiltonian is given by the the identity operator. This Hamiltonian is time-dependent as xappears in both a andw. While performing a canonical quantization the momentum operator is taken to be ^px-~ i x. The differential operator xis skew-Hermitian, i.e. x+=- x, whereas the second order derivative is Therefore, when we performed the Wick rotation, the absence of the derivative of first order leaves the condition of the Hermicity of ^H solely on the term a(x,x)2 x2. In general this term will not be Hermitian unless However, for our case of interest in Section 4, we shall set ato be constant and this will result in our Hamiltonian being Hermitian. This means that the solution to equation xdependencyin first order derivative in (10)is associated to the drift term b(Xt,t)in(7). An SDE that has no dtterm is called a strict local martingale [35, SS2.3] or [34, 31]. an initial SDE into a martingale also helps in obtaining a Wick rotated equation without the first order derivative. In the real time evolution, the evolution operator e-i^Htmust be unitary. This operator is unitary if the Hamiltonian ^His Hermitian. In the imaginary time evolution, however, the evolution operator e-t^His not in general unitary and thus we do not need to be Accepted in click title to verify. Published under CC-BY 4.0. 6concerned about the Hermiticity of ^H. The situation for non-Hermitian Hamiltonians is worth discussing. In [11, SS2.1], a sequence of very judicious transformations and changes of variables transformed the Black-Scholes equation into the heat equation, see also [36, SS6.5.1] as well as the Lamperti transform. The advantage is that the Wick-rotated heat equation becomes a Schrodinger equation with a Hermitian Hamiltonian. In [9, SSII] (which does not employ variational methods), these changes were not applied and as a result the authors have to deal with a non-Hermitian Hamiltonian which gets embedded, by unitary dilation [37], into a larger space and then deal with probabilities of success when measuring the outcomes of quantum circuits. Moreover, in [10, SSII] (which does make use variational methods), these transformation were not employed either and the author has to utilize similarity transformations motivated from supersymmetry [38] in order to obtain a Hermitian Hamiltonian. These techniques are indeed useful for situations where no change of variables or transformations lead to Hermitian Hamiltonians, see also [39] for useful insights. 2.3 Special cases According to the values of the coefficients in (10), we can make the following formal : Suitable a,b,cwith f= : : The solution is an expectation (11)of a Wiener process motion. *Notes: Most general case before specializing the Heat :a= 1,b=c= : : The solution ucorresponds to the temperature field and the Lapla- cian indicates whether the material surrounding each point xis hotter or colder, on average, than the material at the point x. *Notes: Density and properties of materials can be taken into account by applying certain coefficients to the terms that appear in the equation. It is /lscript1-norm : The solution is the price of an option whose underlying follows a geometric Brownian motion. *Notes: Type of financial derivative depends on payoff boundary in click title to verify. Published under CC-BY 4.0. 72.3.4 :a=~2 2m,b= 0,c=V(x), withV(x)the : The solution is interpreted as the amplitude of wave function in quantum The resulting Hamiltonian ^His Hermitian and the equation is /lscript2-norm : HereDis the diffusion coefficient given by D(X,t) : The solution p(x,t)is the probability density function of the Ito process is Time evolution of the probability density function of a particle under the influence of drag forces and random forces. The equation is /lscript1-norm : See Fokker-Planck equation : These equations are used to characterize stochastic processes by de- scribing how they change over time. *Notes: 'Forward' equation is Fokker-Planck, and 'backward' equation is used to determine the probability distribution of a state at a later time s>t. The : Use ansatz ps(x,t) action on : See Schrodinger equation : Take ~-0for limiting semi-classical case and use where Lis the mechanical systems. It is Ornstein - : See Fokker-Planck above. In one dimension, Ornstein - : HereP(x,t)is the probability of finding the process in the state x and timetand diffusion D=s2 2. Accepted in click title to verify. Published under CC-BY 4.0. 8*Interpretation : The associated SDE is it is as a Langevin equation of the wheree(t)is white noise. Note that dWt/dtdoes not exist, as Brownian motion is not differentiable, so this is only The Ornstein - Uhlenbeck process is a stationary Gauss-Markov process and it is temporarily homogeneous. Wide applications in finance (interest rate systems (overdamped processes) and biology (modeling of kinds of formal identification in several dimensions also exist. 2.4 Feynman-Kac consider a D-dimensional stochastic process Xt= by an N-dimensional Brownian motion Wt= (14) whose time evolution is given by the following system of 1,2,***,D. (15) We may abbreviate this Xt,x0anduare inRD,Wtis inRN, and Sis theRDxNmatrix with The infinitesimal generator Gof the diffusion process Xtis given by the differ- ential operator (see e.g. [40, SS4.1] as well as [41, (16) We may further set The terms inGcorrespond to the terms that appear in the multidimensional Ito formula du(Xt,t) for the various stochastic terms involving the different dWj t. the deterministic, non-time dependent, part of (17). Now set u(x,t) (18) for some function pssatisfying appropriate Lipschitz and growth hypotheses [41, Theorem 6]. The function udefined by (18)satisfies the diffusion t> 0 (19) Accepted in click title to verify. Published under CC-BY 4.0. 9as well as the initial condition u(x,0) =ps(x). This problem could be mapped to Schrodinger-type ~G=G-rI, by introducing an appropriate Wick rotation, however ~Gmight not necessarily be is no need to impose correlations between the elements of the motion as we shall see in Section 4.1. Not only does this approach bypass the need to set individual correlations between Brownian motions, it also allows for 1,2,***,Nper dimension d= 1,2,***,Das in(14). We further note we present here, in addition to multi Brownian motions, also contains the discount function r. None of these features appear in the previous literature of VarQITE and stochastic differential equations [10, 11, 12]. 2.5 Finite difference approximation We briefly explain the discretization and the embedding schemes needed to solve the Feynman-Kac system using a Schrodinger equation solver, such as VarQITE which will be described in Section 3. Let nqbe the total number of qubits. We demonstrate procedure for a simple problem of a two dimensional Feynman-Kac genera- tor, as specified in (38)below which is a special case of (16), involving the second To solve the system using n= 4qubits (i.e. 24= basis states: we embed the solution uijin the quantum state as shown in the diagram below. This type of diagram in fact can be used to elucidate the embedding of a multidimensional differential operator on a spatial grid. The mesh is in general a hypercube in ddimensions, but if nqis not divisible by D, then the mesh is a D dimensional hypercube with the sum of its dimensions equal to that the size of the square mesh is 2nq D= 22= 4, wherenqwas the number of qubits and Dis the number of dimensions. Considering the above mesh, the second from (38)can be approximated by using the stencil method [42] in click title to verify. Published under CC-BY 4.0. xandyrefer to the lattice spacing in xandydirections respectively. We addi- tionally employ periodic boundary conditions on the grid in both directions. The second order derivatives appearing in the general case of Ddimensional Feynman- Kac operator (16)can be approximated Similarly if one intends to boundary conditions then appropriate modular operations should be taken into account. Using the approximations above we obtain the discretized form for the 2nq/D. The elements of the matrix are given Quantum algorithm for solving the Feynman-Kac formula In this section, we describe a quantum algorithm for solving the PDE associated to the SDE of interest as prescribed within the Feynman-Kac formulation. We first discuss vari- ational quantum imaginary time evolution (VarQITE), i.e., the fundamental algorithm used, then how to embed the problems into quantum states and the efficient decomposi- tion of the differential operators of the PDE into unitaries. After that, we discuss how to extract properties of the solution, and lastly how to (approximately) enforce the during the evolution. 3.1 VarQITE and state we have a time-independent differential operator Eacting onnqubits with as- sociated propagator exp(-iEt), evolving for real values of t. Provided the operator Eis Accepted in click title to verify. Published under CC-BY 4.0. 11Hermitian, the quantum real time evolution (QRTE) is unitary and can be by a suitable Trotter decomposition, solving the Schrodinger equation. If we apply the Wick rotation t=it, i.e., we translate the problem to the realm of quantum imaginary time evolution (QITE), the evolution operator exp(-Et)will not be a unitary operation anymore and the underlying equation is more complicated to solve on a quantum computer. If we are given an initial state then the normalized imaginary time evolution is defined by |ps(t)/angbracketright=a(t)e-Etwhere a(t) = the normalization factor with respect to the corresponding Wick-rotated Schrodinger equation the is an algorithm to approximate QITE, see [43, 12, 26, 11, 44, 27]. Instead of evolving quantum states in the complete exponential state space, a parameterized ansatz is used and the evolution is approximated by mapping it to the ansatz parameters via McLachlan's variational principle [45]. The main advantage is the possibility of imple- menting it using shallow quantum circuits, suitable for near-term quantum devices [44, p. 1]. More precisely, instead of considering a general state we encode the trial with th(t) = an integer depending on the ansatz circuit of our choice. The choice of the ansatz is crucial for the performance of the algorithm and we discuss this for our concrete example in Sec. for states that are possibly not [26]. In the situation of an by the (24) whereE(t)is now a linear time-dependent -not necessarily Hermitian- operator. The dynamical evolution of be simulated by introducing an ansatz the a parameter of the ansatz that scales the quantum state to the desired scale and G(th(t)) the product of Nparametric unitaries Gi, each composed of one para- metric rotation gates eithkGkwith G+ k=Gk. To produce the VarQITE evolution we need to replace mapping the dynamics of the quantum state to the dynamics of the ansatz. McLach- lan's variational 0 (25) yields the of equations eachk= 0,1,***N. (26) Here the elements of the matrix Mk,jare given by Mk,j= 0, Accepted in click title to verify. Published under CC-BY 4.0. 12as well >0, andM0,0= 1. This matrix is fully dependent on the ansatz circuit. The vector Vk, however, depends on the operator Eas follows Vk=a(t) the very first term gets simplified to V0= one change. These terms are explained in [27, 28]. Since implemented unitaries, the terms Mk,jandVkcan be computed parametrically by using the quantum circuit shown in Figure X GN*** Gk Gk Gk-1*** ohm U Figure 2: This quantum circuit evaluates the matrix elements Mk,jand the vector elements Vkde- pending on ohmandU. The extraction of the matrix elements Mk,jand the vector elements Vkis as follows. *One takes the sequence of gates Gk-1toohmto be Gk-1toGjandU =Gj, if we are interested in Mk,j, *or one sets the sequence of Gk-1toohmto be Gk-1toG1and then U =Uj, where Unare unitary operators such as multi-qubit Pauli operators as represented in (27)below, if we are interested in Vk. Further details on the respective implementation are given in [27, 44, 28]. From(26)and using the above formulas for Mk,jandVk, we can now obtain the evolution of the parameters .thkusing the forward Euler +.thdt=th(t) is, McLachlan's principle defines an ODE which can be numerically solved with an arbitrary ODE solver with NT=t/dttime steps. We would like to point out that the choice of the time step size dtis crucial to achieve reasonable results. Although, it is generally preferable to choose rather smaller time steps one has to consider the trade off between accuracy and computational cost. The use of an ODE solver that implements an adaptive step size scheme such as Runge-Kutta can help to find reasonable time steps. In certain cases the matrix Mwill not be well-conditioned and thus we will need to invoke inverse or use least squares, see e.g. [11, 44]. Notably, the construction cost of the terms MandVscales quadratically in the number of circuit parameters and Furthermore, O(N3). standard Moore-Penrose inverse calculations and solving systems of linear in click title to verify. Published under CC-BY 4.0. 13One has to decompose the operator E(t)as a sum of tensor products of Pauli matrices E(t) an easily implementable unitary operator made of Pauli operators. This requires an evaluation of O(N2) +O(NH)different quantum circuits. The will each require an ancilla qubit in addition to the qubits needed for the dynamics. must be feasible with depths of the unitaries no more than polynomial in n, i.e. the depth of the tensor products of Pauli matrices must be O(poly(n)). This will be explained in Section 3.2. Different manifestations of the Feynman-Kac equation will result in situations where the solution to the PDE might correspond to a probability distribution. To enable this we need to map the probabilities to amplitudes. Since quantum states have to be normalized with respect to the /lscript2norm, while probabilities are normalized with respect to the then need to re-scale the state to handle this mismatch. In other words, the have the form |~ps(t)/angbracketright=a(t)|~ps(t)/angbracketright=2n-1/summationdisplay i=0pi(t)|i/angbracketrightwithpi(t) =u(xi,t) =P[X(t) the random variable whose probability distribution function is given by function u, anda(t)is given by the means that we embed the probability density into a state-vector which is re-scaled with an additional scalar parameter. This will result in a bijection. Moreover, the be provided by the parameterized quantum circuit. It must be noted that indeed this differs from the traditional which is of the of this difference of embedding, extraction of quantities involving pi, such as the moments, need to be re-derived in this new context, see Section 3.3 and Appendix G for further details. Suppose we are confronted with solving (19). Obtaining the full distribution u(xi,t) = pi(t)as its solution is the most complete information, but in general quantum provide this type of result unless full-state tomography is employed, which is an expensive process [46]. However, there could be circumstances when knowing the first and second moments (say) of the solution, with quadratic speedup, might be This also raises the question of approximating a probability distribution from (a few of) its moments. Some of the instances where the moments of the solution are useful are listed below: *The type of heat equation in (19)describes the diffusion of point particles. In for example in one spatial dimension with u: ohm-R, in click title to verify. Published under CC-BY 4.0. 14the density of particles at a certain point xat timet. The first proportional to the total number of particles. If we consider a closed system, which is realized by imposing Neumann boundary conditions, then the conservation of energy means conservation of mass. If Neumann boundary conditions are not imposed, then the total mass is not relevant and it is sensible to instead define HereFis known as the energy functional for the heat equation and it measures the distance, in /lscript2-norm, from the equilibrium solution u= 0. Moreover, Fis related to the Dirichlet energy E[u(*,t)] see [47,SS2.2.5 andSS2.3.4], as follows. For solutions of the heat equation by parts = (29) When we integrate over t, the functional the accumulated loss of Dirichlet energy for t >0and equals our scaling factor a2(t). The time derivative can be approximated by employing finite differences as we have the time evolution of the proxy a(t). Dirichlet energy plays an important role in soft mapping [48], machine learning [49] and discrete differential geometry [50, 51]. *Suppose that qandpare the coordinate and momenta respectively of a wave function and thats(x) the standard deviation of the observable ^x, wherev(x) u(x)isthemeanof ^x. principle [52], the uncertainty length is defined as d[x] := wherep(x)is the probability distribution of the stochastic variable ^x. This is also known as the Susmann measure and it can be interpreted as the total length of all the intervals that produce rectangles of = 1. Here the height hxrepresents the mean value of the as its own weight factor. Its usefulness rests on the fact that other quantities related to width, like variance, might not necessarily exist for (e.g. the Cauchy distribution) whereas the Sussmann measure will. The inverse of the Sussmann measure can be thought of as the average height of quantum mechanical observables corresponding to Hermitian operators ^Oacting on the wave function via solving Schrodinger equation (12) or the Wick-rotated Schrodinger equation (23), are expectation values of the Thus, all such properties are addressable via the algorithm described in this work and the references mentioned herein. *Let us suppose we have an educated belief that the resulting distribution is normal, or a function thereof, e.g. lognormal. One plausible reason for this belief is that the Central Limit Theorem gives the normal distribution a special status and it is characterised by uands. Furthermore, as will be shown in Section 3.3 the evolution of the Feynman-Kac system needs to be evaluated only once to extract the final circuit parameters. Once these circuit parameters are obtained, additional higher order moments can be evaluated in parallel using several quantum devices, in those cases where skewness and kurtosis may play significant role in the description of the solution. As the dimension of the problem increases, so will the overhead. Lastly, for Accepted in click title to verify. Published under CC-BY 4.0. 15time-critical applications one might want to know the first moments using a quantum device while waiting for the full distribution to be computed classically. 3.2 Decomposition of the generator In order to implement VarQITE, we need to employ the matrix Mand shall start with only one dimension. First, we need to decompose the operator Ginto unitaries. The following procedure is borrowed from [12][ SSIII.C] and then adapted to our operator of interest Ggiven by (38)below. Start by defining the operators V+(n) (30) If we make the boundary identifications then the operators V+-can be manufactured from the n-qubit operators Cyc+-(n) per [53], these operators can be implemented as a product of O(n)Toffoli, CNOT, and X gates withO(n)ancilla qubits. The explicit construction of V+-from Cyc+-is achieved by defining an n-qubit-control gate According to [54], then implementation of Cn-1Zcan be achieved as a product of O(n2)Toffolli, CNOT and other single-qubit gates. Using the one can prove that the V's from(30)and the Cyc's are related by V+(n) = we can decompose V+-into a sum of two unitaries made of O(n2)few-qubit gates. Lastly, we bring in an operator composed of weighted projectors D(n) (31) whereZiis aZgate acting on the ith qubit. This implies that D(n)is now a sum ofO(n) unitaries composed of a single-qubit gate. Utilizing these operators we end up (D(n))m=2n-1/summationdisplay i=1im|i-1/angbracketright/angbracketlefti|. For example, let us suppose we have an operator Fwhose decomposition is given by F(t) : all other matrix elements (F(t))i,kbeing zero, which is a typical situation from the stencil method we are implementing in our formulation. Then we decompose the functions fjinto Taylor in click title to verify. Published under CC-BY 4.0. the mth derivative. Using xi= x*iallows us to write F(t) operators all made of the sum Each of these unitaries is made of O(n2)few-qubit gates. Our next step is to now generalize this to multiple dimensions (x1,x2,***,xD)T. builds up naturally from the one dimensional situation we just described above. In this case, we use multivariate Taylor expansions and the operators V(k) +-(n) operators will then (32) as well for different combinations of V's and different projectors can be obtained by similar methods. With (32)and(33)we may now decompose the generator ^Ggiven easily implementable unitaries composed of few-qubit gates and thereby generate a feasible VarQITE. Specifically, the operators the sums eachcomposedof procedures scales polynomially in the number of qubits ndue to the truncation of the Taylor approximation at a fixed order m. 3.3 Readout and efficiency In this section we discuss how to readout results of the solution of a PDE associated to an SDE as well as the resulting total algorithmic complexity. Further, we put this into context and identify the situations where this can lead to a potential quantum we go into the details of the presented algorithm, we discuss the complexity of classical methods. Suppose we want to solve an SDE in Ddimensions and time t[0,T] or the corresponding PDE. Starting with the SDE, we can run a Monte Carlo simulation to generate different scenarios and estimate the solution to the SDE. To achieve a target accuracye >0this will scale as O(T/e2). For small eand complicated SDEs this can Accepted in click title to verify. Published under CC-BY 4.0. 17become prohibitive. An alternative to Monte Carlo simulation is considering the corre- sponding PDE, discretizing the state space and using a classical PDE solver. We assume here that a discretization of the state space using 2ngrid points, i.e., with 2n/Dgrid points per dimension, is sufficient to reach the target accuracy. Then we assume that the total complexity would scale as O(T2n)to access the final state and evaluate properties of in- terest of that state, where we assumed the generator Gcorresponds to a sparse matrix such that matrix-vector products scale linearly in the dimension and that the number of iterations needed by the ODE solver scales linearly in in the following we compare a potential advantage between those two. The the classical and quantum PDE solvers is not completely fair: the classical solver yields the complete state, while the quantum solver only gives access to some properties of the solution. The classical algorithm allows us to control and evaluate the accuracy, while the quantum algorithm is a heuristic unless the ansatz is universal, which would defeat the purpose as this would scale exponentially in the number of qubits. We ignore these differences in the remainder of this section, and thus, the following can be seen as necessary - but not necessarily sufficient - conditions for a quantum advantage. While obtaining the full solution to the PDE generated by the presented algorithm is not feasible, we may try to extract certain properties of interest. Here we need to distinguish between two cases: when we solve our problem forward in time, as, e.g., for the heat equation, or backward in time, as for pricing financial options. The two situations imply differences in how the initial states are prepared and how the results are read out. In the following, we focus on the readout and its efficiency when solving the problem forward in time. The case of solving a PDE backward in time using the proposed method is discussed in Appendix E. Suppose the solution to the PDE is given by a scaled n-qubit quantum state i.e., the scaled amplitudes correspond to the elements of the dis- cretized solution to the considered PDE. We ignore in the following that we only approxi- mate this state. In [12] the authors discuss how to estimate properties of the (34) wherefis some function defined on a grid xi. If thepicorrespond to a of a random variable X(and for simplicity assuming f(x)>=0), then this results in an expected value E[f(X)]. This can be achieved if we can efficiently construct an operator Sfsuch by a weighted sum of such operators [12]. Note that this imposes an on the function f, 1.Iffis not, as we discuss later in this section. The operator Sfcan then be used to in click title to verify. Published under CC-BY 4.0. 18we can estimate by estimating the expected value of Cn-1Zfor the state As the authors of [12] point out, this can be achieved, for instance, up to accuracy e>0with a Hadamard test, or with Quantum Phase Estimation (QPE), using but with a O(1/e)-times longer circuit. Note that we evaluate the square of the expected value using the quantum computer and then multiply the result by a2. Thus, to achieve an overall accuracy eforE[f(X)], we need to estimate the expected value with In the case of the pibeing probabilities, it follows from the that a[1/ 2n,1], i.e.,amay have a positive impact on the estimation error. Although this looks promising, there are a couple of potential pitfalls that can diminish or at least significantly reduce the potential quantum advantage: 1. The operator Sfis often difficult to construct and in general requires a number of terms exponential in the represented dimension that usually scales linearly in the number of qubits. 2. The function fneeds to be that is we estimate a non- normalized function. When going back to the original scale, the resulting estimation error will Thus, the expected value needs to be estimated more accurately to compensate that scaling. This factor is maximal for constant functions, where it scales With the Hadamard test, this implies O(2n)samples depending on probability distributions) it may help to partially compen- sate the error amplification introduced by the normalization of the presented technique, and are often ignored in the related literature. For solving PDEs backwards in time, similar phenomena appear and we discuss them in more depth in Appendix E. For illustration, we assume the solution represents the probability distribution pof a D-dimensional random variable Xand we will focus on how to estimate its expected value E[X], for more details, see Appendix G. To this extent we assume that the scaled represents the final solution, i.e., we represent each dimension we assume an n/D-qubit we set f(x) =x, the required normalization constant. There are different ways to construct Sx, e.g., via a piece-wise constant approximation as suggested in [12] or following the approach in click title to verify. Published under CC-BY 4.0. 19in [55], which may be simplified due to the simple structure of the function considered. In the following, we focus on how to estimate a single component E[Xi] of the expected value. This needs to be repeated Dtimes to estimate the result. Without loss of generality we now assume i= 1and is easy to see that this corresponds to the objective we can apply the above approach to estimate To achieve a target accuracye >0forE[X1]using QPE requires an estimation accuracy of results in a longer circuits (ignoring a). To summarize, the proposed quantum algorithm can achieve an exponential speed- up for the (approximate) state evolution, but only up to a quadratic speed-up for the extraction of solution properties, which results in an overall complexity of O(Tpoly(n) + 2n)since the readout can be done using the final parameters without rerunning the evolution. Thus, the proposed algorithm might be particularly suitable for problems with a large time horizon. However, it could also accumulate a higher approximation error for those, and understanding where these two effects balance is subject to further research. 3.4 Enforcing PDEs describe the evolution of probability distributions, i.e., the components of a discretized solution should be in [0,1]and sum up to one. Although we allow the overcome the of the quantum state the in general does neither preserve the /lscript1-norm that the elements should be within [0,1]. Since it minimizes a geometric distance, the closest might not be In the following we describe how the procedure could be adjusted to (approximately) enforce while sacrificing a bit of the approach introduced in [56]. As shown in Section 3.3, we can estimate properties of the function fand amplitudes piof the current state to an accuracy e>0. If we now set f(xi) alli, then we can easily construct Sf=Hn, i.e., just a single layer of Hadamard gates, which will result in an estimation we estimate this to an accuracy e/ 2nusing QPE, this can be used as an approximation to of the state and allows to adjust the atsuch that the scaled quantum state would be (approximately) Note that the accuracy requirements on the estimation of the /lscript1-norm likely limit the potential advantage to a quadratic speed-up O(T 2n). 4.3. Appendix procedure described in this in click title to verify. Published under CC-BY 4.0. 203.5 Summarizing the algorithm The algorithm we have described is now summarized in Algorithm 1. For clarity, we use one spatial dimension as well as the forward Euler ODE solver. However, in practice, solvers such as for instance Runge-Kutta methods are usually more 1: Variational quantum algorithm for the Feynman-Kac 1,NT>0,dt > 0and SDE of the form Normalized state kets for solution 0. (36) 1Apply appropriate Feynman-Kac formulae to switch SDE to PDE where u(x,t) the appropriate normalization ( of (36). 3Extract Gfor periodic boundary wherelhCandUhare implementable composed of tensor products of Pauli ansatz circuit with N+ 1parameterized gates. 7Define initial condition initial parameters th0= =a(t), see ansatz-dependent matrix elements Mk,j= >0andM0,0= 1. 12ifM-1is singular then 13 Use least squares or Moore-Penrose time-dependent then 15 Update generator G(t)and its associated Pauli vector elements Vk=a(t) values of thwith th(t+dt)th(t) Numerical experiments We now provide numerical evidence of the algorithm we have proposed. To do so, we first need to establish the exact set of SDEs that we wish to study, in our case it will be a simple Accepted in click title to verify. Published under CC-BY 4.0. 21set of two correlated Brownian motions. From there we will write down the PDE for the expectation of these Brownian motions. This PDE is an anistropic heat equation that mixes the space variables xandydue to the correlation between the Brownian Laplace transforms. The ansatz circuit will be explained and we will also illustrate the behavior of the during the evolution. The VarQITE algorithm outputs the basis kets in one dimension, and since we shall use six and eight qubits, we will obtain 26and28basis kets. Although the problem is two dimensional, it is instructive to take amplitude vector in the computational basis and the comparison against the of forward Euler as well as Monte Carlo with periodic boundary conditions. 4.1 Two-dimensional anistropic heat equation We shall now specialize the number of dimensions, the number of Brownian motions and the coefficients in the SDE to illustrate the Feynman-Kac formula on a quantum we specialize to u1=u2= 0as well as s11=~s1, s 12= 0, s 21=~s2r, s a real in the interval [-1,1]which now is interpreted as the correlation between the two Brownian motions W1 tandW2 t. The resulting matrix infinitesimal generator ^Greduces this particular case. We shall also relabel x1=xandx2=y. Moreover for simplicity our discount function will be r(x,y,t ) = 0for allt, and we shall choose s1ands2to be 1. In summary, we are working with the differential operator and conditional ) associated Feynman-Kac PDE for the expectation of the system of SDEs t> 0, (39) with initial value condition psgiven by u(x,y,0) =d(x-x0)d(y-y0), (40) Accepted in click title to verify. Published under CC-BY 4.0. 22wheredis the Dirac delta function. The operator Gis an anisotropic Hamiltonian that includes the Laplacian operator 2=2 x2+2 y2,as well as the term mixing both first 2 xy.Equation (39)admits an exact heat kernel solution, which we derive - in a more general setting involving arbitrary constant coefficients - in Appendix F, u(x,y,t ) while the solution to (39)without the mixed term2 xyand with (40)is well-known, the general solution (41)was not immediately available in the literature [57, 5, 58], see Mehler's formula. The results of a large number of Monte Carlo simulations and the plot of exact solution are shown in Figure 3. Figure 3: Left : probability distribution of 1 million paths governed by the system of SDEs from (37). Right: Corresponding analytical solution to (39)given by Choice of ansatz, expressivity, barren plateaus and initialization of parameters The choice of ansatz will be what is called RealAmplitudes from Qiskit[59] and it can be found in Figure 4. It is a circuit that consists of six qubits with several layers of single qubit rotations as well as CNOT gates with circular entanglement. The circuit always returns Other ansatze are indeed possible for example using RxandRzgates. How- ever, this comes at the cost of taking the real part of the resulting ket at the end of the evolution, see e.g. [27]. Moreover, the non-parametric controlled gates could be replaced by parameter gates (such as controlled rotations) for more generality, see for instance the ansatz proposed in [11]. In order to get reliable simulation results, we need to ensure that the model ansatz is able to represent the problem in state space. However, we also need to consider that an ansatz which is sufficiently expressive, such that it corresponds to an [60, 61], is prone to result in barren plateaus [62]. More explicitly, it is known that for an ansatz that corresponds to an (approximate) t-design the gradient exponentially small in the system size which in turn may cause problems in the model simulation. It is thus vital to trade off the expressivity of a chosen ansatz with potential propagation issues. One particularly promising approach relies on the use of additional information to design problem-specific in click title to verify. Published under CC-BY 4.0. 23Repeat Ry(thn,6) Figure 4: Parametric quantum circuit used in the numerical experiments in this work. It is composed of a layer ofRyrotation gates on every qubit, a block of nonparametric CNOT gates connecting and another layer of Ryrotation gates on every qubit. This parametric quantum circuit is also referred to as RealAmplitudes ansatz due to the real-valued nature of the amplitudes in the resulting state our case, we are solving a PDE, (39), that admits only real solutions, which are in fact probability distributions. Therefore, it is convenient to keep the ansatz from amplitudes effectively avoiding dealing with the imaginary part of these ampli- tudes. Consequently, for the problem at hand it is reasonable to use the put forward in this section. The parameters th0= initialized classically using th0= (42) In [11,SS4], the authors make the point that care is needed when solving (42)due to the fact that the functions involved might not be convex and certain optimization techniques could be trapped in local minima. Moreover, the design of the ansatz should also go hand in hand with the optimization (42). Note that in general this scales poorly with the number of qubits. As with many quantum algorithms, one needs to be careful to ensure that the general state preparation techniques do not add costly overhead. Note that step might also be performed with a quantum computer using SWAP test when the dimension of the problem starts to be too large for classical computers. 4.3 VarQITE vs forward Euler vs Monte Carlo in computational basis Although we are dealing with the problem in (39), it is instructive to plot the amplitudes in the computational basis. The parameters are set to dx=dy= 1 as well asdt= 0.001. The number of evolutions is NT= 1000. We first use six qubits in the ansatz with twelve parameterized gates, i.e. n= 1, anda(t) =th13(t)as the norm encoder variable. With a circuit of six qubits and n= 3layers, we have a(t) in click title to verify. Published under CC-BY 4.0. 24norm encoder. Then we use eight qubits with n= 5layers anda(t) =th41(t)as the norm enconder enforcement of the evolution discussed in 3.4 can be seen in these plots of 1000 evolutions on the bottom left of Figures 6, 8 and 10. Earlier in the evolution period, the first kets came out with a flipped sign due to a phase. Moreover, periodic and the nature of the forward Euler method also yielded a norm slightly larger than one for the classical methods. Small deviations in the /lscript1-norm from VarQITE are coming from the approximate nature of the /lscript1-norm enforcement. Another useful metric to keep in mind is the difference between the /lscript2norms of forward Euler and VarQITE as well as Monte Carlo and VarQITE as shown on the bottom right of Figures 6, 8 and 10. We note the following analysis of the experiments: (1) One can see in Figure 5 that VarQITE was significantly off around the states 32 to 38 with six qubits and only n= 1layer. This situation is due to the choice of ansatz and its associated To see the effect of increasing the expressivity of the ansatz, we added more layers by increasing nfrom 1 to 3 while keeping the number of qubits constant at six. The results are plotted on Figure 7. Lastly, we then chose an eight qubit ansatz and added two additional layers of CNOTs and Ryrotations, i.e. n= 5. The results are in Figure 9. (3) In both instances (2) and (3), the match of the plots of VarQITE and forward Euler no longer show the discrepancy in the computational bits 32 to 38 for six 170 to 180 for eight qubits) while still maintaining a tight To perform the simulations with additional layers and additional qubits, to match more tightly with symbolic or circuit-based differentiation and this will be studied in future research. (5) We do not include finite-sampling noise or hardware noise. All numerical simulations have been tested and produced using Pythonas well as the NumPylibrary. The effect of noise on the results is an interesting topic and will be left for future research. (6) Although this methodology works with a high degree of generality, for high dimen- sional cases, the expressivity of the ansatz employed in VarQITE will limit the accu- racy of the results. 5 Conclusion We have proposed a quantum algorithm to transition from stochastic differential Wiener processes X1,X2,***,XDto partial differential the expectation of the Wiener processes. The resulting PDEs have been variational quantum imaginary time evolution (VarQITE). We have seen that well-known PDEs such as Schrodinger, heat and Fokker-Planck equations can be recast through this type of methodology and thereby can be unified in the single framework of the Feynman-Kac formula. Moreover, we have maintained a high degree of generality when dealing with dimensional cases, the expressivity of the ansatz will limit the Accepted in click title to verify. Published under CC-BY 4.0. 250 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad EulerFigure 5: Plots of uVarQITE computa- tional basis states with an ansatz consisting of six qubits and n= 1layer. The horizontal axis represents 26= 64computational basis and vertical axis represents u(x,t)withdt= 0.001and 1000 evolutions. 0 200 400 600 800 of Carlo Forward Euler 0 200 400 600 800 of the differences in u ||uFD uMC||2 Figure 6: Left :/lscript1norms for the first 1000 evolutions are plotted using three different Monte Carlo, and forward Euler method. Note that the scale on the y-axis has been narrowed down to demonstrate the approximate preservation of the /lscript1norms for all three methods. The time-steps and yaxis represents /lscript1-norm. Right : This error plot shows the /lscript2norms of the differences between the solutions uobtained via VarQITE and forward Euler (blue) and between VarQITE and Monte Carlo method (yellow). The x-axis represents time and yaxis represents the differences in u. accuracy of the results. Therefore, as the dimension of the problem increases one has to improve or enhance the ansatz accordingly. Some of the resulting PDEs do not preserve in click title to verify. Published under CC-BY 4.0. 260 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad Euler 0 10 20 30 40 50 60 Computational Carlo Fordwad EulerFigure 7: Plots of uVarQITE compu- tational basis states with an ansatz consisting of six qubits and n= 3layers. The horizontal axis represents 26= 64computational basis and vertical axis represents u(x,t)withdt= 0.001and 200 400 600 800 of Carlo Forward Euler 0 200 400 600 800 of the differences in u ||uFD uMC||2 Figure 8: Left :/lscript1norms for the first 1000 evolutions are plotted using three different Monte Carlo, and forward Euler method. Note that the scale on the y-axis has been narrowed down to demonstrate the approximate preservation of the /lscript1norms for all three methods. The time-steps and yaxis represents /lscript1-norm. Right : This error plot shows the /lscript2norms of the differences between the solutions uobtained via VarQITE and forward Euler (blue) and between VarQITE and Monte Carlo method (yellow). The x-axis represents time and yaxis represents the differences in u. normalized throughout the quantum evolution by introducing a proxy to the algorithm has shown substantial agreement in the solution of the PDE through VarQITE, classical Monte Carlo and forward Euler in a sufficiently complex and in click title to verify. Published under CC-BY 4.0. 270 50 100 150 200 Carlo Fordwad Euler 0 50 100 150 200 Carlo Fordwad Euler 0 50 100 150 200 Carlo Fordwad Euler 0 50 100 150 200 Carlo Fordwad Euler 0 50 100 150 200 Carlo Fordwad Euler 0 50 100 150 200 Carlo Fordwad EulerFigure 9: Plots of uVarQITE compu- tational basis states with an ansatz consisting of eight qubits and n= 5layers. The horizontal axis represents 28= 256computational basis and vertical axis represents u(x,t)withdt= 0.001and 200 400 600 800 of Carlo Forward Euler 0 200 400 600 800 of the differences in u ||uFD uMC||2 Figure 10: Left :/lscript1norms for the first 1000 evolutions are plotted using three different Monte Carlo, and forward Euler method. Note that the scale on the y-axis has been narrowed down to demonstrate the approximate preservation of the /lscript1norms for all three methods. The time-steps and yaxis represents /lscript1-norm. Right : This error plot shows the /lscript2norms of the differences between the solutions uobtained via VarQITE and forward Euler (blue) and between VarQITE and Monte Carlo method (yellow). The x-axis represents time and yaxis represents the differences in u. problem involving an anisotropic Fokker-Planck equation in (2 + 1)dimensions. the full solution to the PDE is an expensive problem, we need to instead the multidimensional moments of the solution through quantum in click title to verify. Published under CC-BY 4.0. 28In order to do so, we have decomposed the second order differential operator Gfrom the PDE into a feasible set of unitaries composed of few-qubit gates. For a the moments be computed by approximating fwith its multidimensional Taylor series to a fixed order and this approximation can be improved by increasing the number of qubits to allow for finer resolution. These moments can be computed through QPE or Hadamard tests with potential advantages depending on the concrete situation. The methodology provided here supplements the existing quantum-enhanced Monte Carlo algorithms and both techniques can coexist and complement each other. The scope of the topics we have addressed and their mapping to quantum computers allows us to open the possibilities for other interesting items for future work such as option pricing with stochastic volatility, see [36], as well as stochastic control problems such as American options which require the equation, see [63]. These topics might have to be explored in the context of real, as opposed to imaginary, time evolution. Lastly, the heat equation is not the only candidate for this algorithm. Ito's lemma can be used [40, SS4.1.1] to transform a conditional expectation arising from a vectorized SDE into the solution of the PDE. Another venue would be to consider the wave or Klein-Gordon equations (or in general, other type of hyperbolic differential equations). In this case, the time derivative is now of second order, unlike in the heat equation, which was of first order. For these situations, other types of probability distributions and Brownian motions need to be added, specially by considering jumps and Poisson processes, see [64]. 6 authors would like to thank Peter Tysowski and Vaibhaw Kumar for constructive dis- cussions regarding these topics. Antoine Jacquier was kind enough to read the CZ acknowledges support from the National Centre of Competence in Research Quantum Science and Technology (QSIT). The authors would like to acknowledge the ref- erees for the very helpful comments that have greatly increased the accuracy, precision, and readability of the manuscript. IBM, the IBM logo, and ibm.com are trademarks of International Business Machines Corp., registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. The current list of IBM trademarks is . 7 Code upon reasonable R. P. Feynman. The principle of least action in quantum mechanics. In New Approach To Quantum Theory , pages 1-69. World Scientific, 2005. Available online: . Accepted in click title to verify. Published under CC-BY 4.0. 29[2] M. Kac. On Distributions of Certain Wiener Functionals. Trans- actions of the American Mathematical Society , 65(1):1-13, 1949. DOI: . [3] F. Black and M. Scholes. The pricing of options and corporate liabilities. Journal of Political Economy , 81(3):637-654, 1973. Available online: . [4] J. Glimm and A. Jaffe. Quantum Physics: A Functional Integral Point of View New York, 2012. DOI: . [5] J. Lorinczi, F. Hiroshima, and V. Betz. Feynman-Kac-Type Theorems and Gibbs Measures on Path Space: With Applications to Rigorous Quantum Field Theory . De Gruyter, 2011. DOI: . [6] A. Korzeniowski, J. Fry, D. Orr, and N. Fazleev. Feynman-Kac path-integral calcu- lation of the ground-state energies of atoms. Phys Rev Lett , 10(69):893-896, 1992. DOI: . [7] M. Caffarel and P. Claverie. Development of a pure diffusion quantum Monte Carlo method using a full generalized Feynman-Kac formula. I. Formalism. The Journal of Chemical Physics , 88(2):1088-1099, 1988. DOI: 10.1063/1.454227 . [8] P. A. Faria da Veiga, M. O'Carroll, and R. Schor. Existence of baryons, baryon spectrum and mass splitting in strong coupling lattice QCD. Commun. Math. Phys. , 245:383-405, 2004. DOI: . [9] J. Gonzalez-Conde, A. Rodriguez-Rozas, E. Solano, and M. Sanz. Pricing Fi- nancial Derivatives with Exponential Quantum Speedup. arxiv:2101.04023 , 2021. DOI: . [10] S. K. Radha. Quantum option pricing using Wick rotated imaginary time , 2021. DOI: . [11] F. Fontanela, A. Jacquier, and M. Oumgari. A Quantum algorithm for linear PDEs arising in Finance. SSRN Electronic Journal , 2019. DOI: . [12] K. Kubo, Y. O. Nakagawa, S. Endo, and S. Nagayama. Variational of stochastic differential equations. Phys. Rev. A , 103:052425, 2021. DOI: . [13] G. Brassard, P. Hoyer, M. Mosca, and A. Tapp. Quantum Amplitude Amplifica- tion and Estimation. Quantum Computation and Quantum Information, Samuel J. Lomonaco, Jr. (editor), AMS Contemporary Mathematics , 305:53-74, 2002. DOI: 10.1090/conm/305 . [14] A. Montanaro. Quantum speedup of Monte Carlo methods. Proceedings of the Royal Society A , 471(2181), 2015. DOI: . [15] Y. Suzuki, S. Uno, R. Raymond, T. Tanaka, T. Onodera, and N. Yamamoto. Am- plitude Estimation without Phase Estimation. Quantum Inf Process , 19(75), 2020. DOI: . [16] D. Grinko, J. Gacon, C. Zoufal, and S. Woerner. Iterative quantum amplitude esti- mation.npj Quantum Information , 7:52, 2021. DOI: . [17] A. Carrera Vazquez and S. Woerner. Efficient state preparation for quantum amplitude estimation. Phys. Rev. Applied , 15:034027, 2021. DOI: . Accepted in click title to verify. Published under CC-BY 4.0. 30[18] N. Stamatopolous, D. J. Egger, Y. Sun, C. Zoufal, R. Iten, N. Shen, and S. Wo- erner. Option Pricing using Quantum Computers. Quantum , 4(291), 2020. DOI: . [19] S. Woerner and D. J. Egger. Quantum Risk Analysis. npj Quantum Information , 5(15), 2019. DOI: . [20] D. J. Egger, R. Garcia Gutierrez, J. Cahue Mestre, and S. Woerner. Credit risk analysis using quantum computers. IEEE Transactions on Computers , pages 1-1, 2020. DOI: . [21] S. Chakrabarti, R. Krishnakumar, G. Mazzola, N. Stamatopoulos, S. Woerner, and W. J. Zeng. A Threshold for Quantum Advantage in Derivative Pricing. Quantum , 5:463, 2021. DOI: . [22] D. An, N. Linden, J.-P. Liu, A. Montanaro, C. Shao, and J. Wang. multilevel Monte Carlo methods for stochastic differential equations in mathematical finance. Quantum , 5:481, 2021. DOI: . [23] P. Rebentrost, B. Gupt, and T. R. Bromley. Quantum computational finance: Monte Carlo pricing of financial derivatives. Phys. Rev. A , 98:022321, 2018. DOI: . [24] B. Oksendal. Stochastic Differential Equations . Springer, 2013. DOI: . [25] ocw.mit.edu. Lecture 21: Stochastic Differential online: . [26] S. Endo, J. Sun, Y. Li, S. C. Benjamin, and X. Yuan. Variational quan- tum simulation of general processes. Phys. Rev. Lett. , 125:010501, 2020. DOI: . [27] X. Yuan, S. Endo, Q. Zhao, Y. Li, and S. C. Benjamin. Theory of variational Quantum , 3(191), 2019. DOI: . [28] C. Zoufal, D. Sutter, and S. Woerner. Error Bounds for Variational Quantum Time Evolution. arXiv:2108.00022 , 2021. DOI: . [29] N. H. Bingham and R. Kiesel. Risk-Neutral Valuation: Pricing and Hedging Financial Derivatives . Springer-Verlag Berlin Heidelberg, 2004. DOI: . [30] S. Ramos-Calderer, A. Perez-Salinas, D. Garcia-Martin, C. Bravo-Prieto, J. Cortada, J. Planaguma, and J. I. Latorre. Quantum unary approach to option pricing. Phys. Rev. A, 103:032414, 2021. DOI: . [31] I. Karatzas and S. Shreve. Brownian Motion and Stochastic Calculus . Graduate Texts in Mathematics, Springer, 1984. DOI: . [32] M. Musiela and M. Rutkowski. Martingale Methods in Financial Modelling . Springer- Verlag Berlin Heidelberg, 2005. DOI: 10.1007/b137866 . [33] M. Baxter and A. Rennie. Financial Calculus: An introduction to derivative pricing . Cambridge University Press, 1996. DOI: . [34] E. Chin, D. Nel, and S. Olafsson. Problems and Solutions in Mathematical Calculus . Wiley Finance Series, 2014. DOI: . Accepted in click title to verify. Published under CC-BY 4.0. 31[35] A. Papanicolaou. Introduction to Stochastic Differential Equations (SDEs) for , 2015. DOI: . [36] P. Wilmott. Paul Wilmott On Quantitative Finance . Wiley, 2006. [37] T. Constantinescu. Schur parameters, Dilation and Factorization Problems . Birkhauser Verlag 82, 1996. DOI: . [38] F. Bagarello. Susy for non-Hermitian Hamiltonians, with a view to coherent states. Math. Phys. Anal. Geom. , 23(3):28, 2020. DOI: . [39] S. Dogra, A. A. Melnikov, and G. S. Paraoanu. Quantum simulation of breaking with a superconducting quantum processor. 4(26), 2021. DOI: . [40] J. N. D. of Mathematics at Stanford . equations and diffusion processes. Available online: http://math. stanford.edu/ . [41] S. L. D. of Statistics at University of . Stochastic Differen- tial Equations, Diffusion Processes and the Feynman-Kac formula. Available online: http://www.stat.uchicago.edu/ . [42] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. 3rd Edition: The Art of Scientific Computing . Cambridge University Press, 2007. DOI: . [43] Y. Li and S. C. Benjamin. Efficient variational quantum simulator in- corporating active error minimization. Phys. Rev. X , 7:021050, 2017. DOI: . [44] S. McArdle, T. Jones, S. Endo, Y. Li, S. Benjamin, and X. Yuan. Variational ansatz- based quantum simulation of imaginary time evolution. npj Quantum Information , 5(75), 2019. DOI: . [45] A. D. McLachlan. A variational solution of the time-dependent Schrodinger Physics , 8(1), 1964. DOI: . [46] G. M. D'Ariano, M. G. Paris, and M. F. Sacchi. Quantum tomog- raphy. Advances in Imaging and Electron Physics , 128:206-309, 2003. DOI: . [47] L. C. Evans. Partial Differential Equations . American Mathematical Society, 1998. DOI: . [48] J. Solomon, L. Guibas, and A. Butscher. Dirichlet energy for analysis and synthesis of soft maps. In Computer Graphics Forum , volume 32, pages 197-206. Wiley Online Library, 2013. DOI: . [49] K. Zhou, X. Huang, D. Zha, R. Chen, L. Li, S.-H. Choi, and X. Hu. Dirichlet learning for deep graph neural networks. Advances in Neural Systems , 34, 2021. DOI: . [50] O. Stein, M. Wardetzky, A. Jacobson, and E. Grinspun. A simple discretization of the vector dirichlet energy. In Computer Graphics Forum , volume 39, pages 81-92. Wiley Online Library, 2020. DOI: . Accepted in click title to verify. Published under CC-BY 4.0. 32[51] O. Stein, A. Jacobson, M. Wardetzky, and E. Grinspun. A smoothness energy without boundary distortion for curved surfaces. ACM Transactions on Graphics (TOG) , 39(3):1-17, 2020. DOI: 10.1145/3377406 . [52] G. Sussmann. Uncertainty Relation: From Inequality to Equality. Zeitschrift A , 1997. DOI: . [53] X.Li, G.Yang, C.M.Torres, D.Zheng, andK.L.Wang. Gates for Quantum Circuit Synthesis. International Journal of Modern Physics B , 28(1), 2014. DOI: . [54] M. A. Nielsen and I. L. Chuang. Quantum Computation and Quantum Information . Cambridge University Press, 2012. DOI: . [55] L. Grover and T. Rudolph. Creating superpositions that correspond to efficiently integrable probability distributions. arXiv:0208112 , 2002. DOI: . [56] A. Carrera Vazquez, R. Hiptmair, and S. Woerner. Enhancing the Quantum Lin- ear Systems Algorithm using Richardson Extrapolation. arXiv:2009.04484 , 2020. DOI: 10.1063/1.454227 . [57] N. Berline, E. Getzler, and M. Vergne. Heat Kernels and Dirac Operators . Springer- Verlag Berlin Heidelberg, 2004. [58] P. DuChateau and D. W. Zachmann. Schaum's outlines: partial differential equations . McGraw-Hill, 1986. DOI: . [59] 2021. Available online: qiskit.org . [60] A. Hayashi, T. Hashimoto, and M. Horibe. Reexamination of optimal quantum state estimation of pure states. Physical Review A , 72(3), 2005. DOI: . [61] A. Ambainis and J. Emerson. Quantum t-designs: t-wise Independence in the Quan- tumWorld. In Twenty-Second Annual IEEE Conference on Computational , 2007. DOI: . [62] J. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven. Barren plateaus in quantum neural network training landscapes. Nature Communications , 9, 2018. DOI: . [63] J. Vecer. A New PDE Approach for Pricing Arithmetic Average Asian Options. Jour- nal of Computational Finance , 4(4), 2001. DOI: . [64] R. C. Dalang, C. Mueller, and R. Tribe. A Feynman-Kac Type For- mula for the deterministic and stochastic wave equations and other of the American Mathematical Society , 2008. DOI: . [65] R. Brummelhuis. Mathematical Methods - MSc Financial Engineering - Birbeck College, 2004. Available online: . [66] V. V. Shende, S. S. Bullock, and I. L. Markov. Synthesis of quantum-logic circuits. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , 25(6), 2006. DOI: . Accepted in click title to verify. Published under CC-BY 4.0. 33[67] C. Zoufal, A. Lucchi, and S. Woerner. Quantum Generative Adversarial Networks for learning and loading random distributions. npj Quantum Information , 5(103), 2019. DOI: . [68] F. Oberhettinger. Tables of Mellin Transforms . Springer-Verlag Berlin DOI: . [69] E. Knill, G. Ortiz, and R. D. Somma. Optimal quantum measure- ments of expectation values of observables. Phys. Rev. A , 75:012328, 2007. DOI: . [70] D. Wang, O. Higgott, and S. Brierley. Accelerated variational quantum Rev. Lett. , 122:140504, 2019. DOI: . A Details on Feynman-Kac, Ito's lemma and the heat equation Suppose that a solution to the 'backward' 0,fort<T u(x,T) =ps(x).(43) Now let (Wt)t>=0be Brownian motion and consider u(Wt,t). Performing a small =u(Wt,t) Ito's lemma (1)we have the rule of 1, dt, ifk= (dt)k= 0fork>=2. Thus, we will ignore the last terms in (44). Passing the left-hand side we see that see that the term involving dtis zero, and therefore we are left with du(Wt,t) = u wdWt. Integrate this from ttoTso =u(WT,T)-u(Wt,t) after re-arrangement, leads to u(Wt,t) (45) Accepted in click title to verify. Published under CC-BY 4.0. 34where we used the final boundary condition of (43). The key part is that we must now conditioned on Wt=x, i.e.u(x,t) =E[ps(WT)|Wt=x], so that picking up from(45)we get u(x,t) =E[ps(WT)|Wt=x] (46) Let us deal with the second term first. Using tools from analysis, such as a Fubini's theorem, we can swap the expectation and the integral and the properties of Brownian motion we know that dWsis independent of Wsor, for that matter, any other function of Ws, hence by independence we can separate the expectation of the product into the product of the expectations again the independence of Brownian increments we get E[dWs|Wt=x] = E[dWs] = 0, since the Brownian motion is associated with a normal distribution with mean zero. Thus the expectation of the integral in (46)is simply zero and we are left with u(x,t) =E[ps(WT)|Wt=x]. The task at hand now is to compute this expectation and to this end, we use WT=Wt+ (WT-Wt)inside the expectation so =E[ps(Wt+ (WT-Wt))]. This is because the expressions WT-WtandWtare independent, therefore irrelevant. The conclusion is that if u(x,t)is a solution to (43), then u(x,t) =E[ps(w+WT-Wt)]. Now we need to remove the expectation. For that, normally distributed and therefore using the probability distribution function of the Gaussian distribution, we see that u(x,t) =E[ps(x+ T-tZ)] is to be compared to (5). This is the solution to the heat equation that would have been obtained through Fourier transform methods or through similarity reductions. B The full Feynman-Kac formula in one dimension This idea can be generalized with very little work to the following situation. Suppose =ps(x). We can associate to this PDE a stochastic process these coefficients as the SDE Then in click title to verify. Published under CC-BY 4.0. 35is the result that u(x,t) All issues of existence and uniqueness can be found in Shreve and Karatzas [31, SS4]. The sketch of the argument leading to Feynman-Kac formula is to first set Yt:= to then employ Ito to show the derivatives of uevaluated at (Xt,t). This allows us to establish that if udoes solve the PDE, then Ytabove has zero drift (i.e. it is a martingale) and then write E[Yt|Xt=x] =u(x,t). The details can be found in [65]. C Linking the Feynman-Kac PDE to the Schrodinger equation Let us make the substitution u(x,t) =eg(x,t)v(x,t), for some analytic functions the newly transformed derivatives appearing into (10)and re-grouping terms, we the coefficient of the unwanted term to be zero = (48) We may integrate this to g(x,t) (49) wherehis an arbitrary function of tonly. Using (48)on the coefficient of the linear term vappearing in (47)we get the : an arbitrary function we may take it to be h(t) = 0so thatdh dt= 0as this implies that g(x,t) the integration that the result of the integration with respect to y is a function of xwithout a constant of integration. This implies that w(x,t) in click title to verify. Published under CC-BY 4.0. 36In a substantial number of interesting cases the coefficients a,bandcare functions of x only, (50). Therefore, (10)becomes 0 are two additional substitutions we need to make. First, we need to turn this into an initial value problem by setting t=T-t, in which case we end up (51) Lastly, the Wick rotation x=-ittransforms that some straightforward modifications to the time variable need to be applied in these last two transformations to the functions uandvto keep consistency. D Additional norm plots To demonstrate the efficacy of the /lscript1-norm enforcement procedure as explained in Sec 3.4, we provide additional plots in Figure 11 comparing the VarQITE method against Monte Carlo and forward Euler methods. 0 200 400 600 800 without alpha(t) and without 1 Carlo Forward Euler 0 200 400 600 800 with alpha(t) and without 1 Carlo Forward Euler Figure 11: Plots of /lscript1norms for the first 100 evolutions with six qubits and n= 1layers of ansatz, left: withouta(t)and without the /lscript1-norm enforcement procedure and right : witha(t)but without enforcement procedure. E Backward propagation and option formula is required: one considers propagation backwards in time instead of forward, i.e., the final payoff is used as initial condition and we propagate back in time to the starting point of the process. Every element of the result then corresponds to the fair option price for the corresponding initial state. For instance, one may wish to solve the heat equation by specializing a(x,x) =1 2 andw(x,x) = 0as well as the payoff function ps(XT) = T. equation can be written as the heat equation using suitable into the domain of path-dependent options, another popular financial in click title to verify. Published under CC-BY 4.0. 37is the Asian option, see [63]. To price it, we specialize the coefficients of the = 0, whereq(t) 0andq(t) = 1-t Tif r= 0. The payoff is given by the average of process Xt, i.e. ps(XT,T) = (52) Inthiscase, Kandmaturity T. Weremarkthat the payoff function psnow depends on T. These cases are studied in [11, 10, 23, 18], not always under the optic of variational quantum imaginary time evolution. The European case is showcased in [9, 30] by amplitude estimation and without applying s. Much like in the case of forward propagation, cf. Section 3.3, the performance of the algorithm when going backward in time needs to be carefully analyzed. First, the payoff function, which is now considered as the initial condition, needs to be loaded into a quantum state, i.e., it needs to be normalized and we need to prepare a andK= (Xl+Xu)/2. If we discretize XTusingnqubits, i.e. with 2ngrid points and grid size h=B/2n, then the normalization factor of the quantum state As in forward propagation, this limits the potential advantage already to a quadratic speed-up, since we need to re-scale any result -including the estimation error- by a factor of size th(B 2n)to get back to the original scale. In addition to the a circuit to prepare arbitrary quantum states can be prohibitive [66]. For the simple payoffs considered efficient circuits should exist, and for more generic approach as suggested in [11] or [67] may be used. Suppose the initial state has been loaded, the algorithm has been applied to estimate a prepared quantum state that corresponds to the solution and one is interested in the result. The impact of the a-factor might be negligible here, since we already start in a quantum state with rather uniform amplitudes. However, the option price for a particular target initial price corresponds now to the amplitude of a single computational basis state. The readout of this can be achieved via QAE up to accuracy e>0at costO(1/e). Since the amplitude of interest likely will be of order O(1/ 2n), the required ewill be of equal scale. Combining the requirements coming from the normalization likely renders this for option pricing. F Analytical solution The equation we need to solve ), and initial condition u(x,y,0) all real constants such that 4BC-A2>0. The idea is to apply a Laplace transform on the time variable and a Mellin transform on each of the space variables. This will allow us to change the PDE into an algebraic equation which we can solve trivially. Then we proceed to invert the Laplace transforms as well as the to recover the solution of (53). Let us start (54) Accepted in click title to verify. Published under CC-BY 4.0. 38withDandEreal constants. Make the changes X= logx, Y = logy. (55) Then from (55)we see (D-B)u X+ (E-C)u Y-ru. have the same structure as (58) minding that x,yare related to X,Yby(55). Thus, we need to solve the conditions given by (58). The solution to is given by u(x,y,t ) = 4BC-A2>0. UsingX0= logx0andY0= logy0in addition to (55)(that is, reversing the change of variables), then the solution to is given by u(x,y,t ) start by taking the Laplace transform of uwith respect to tso that ^u(x,y,s ) =L{u(x,y,t )}(s) ). Next take a Mellin transform of ^uwith respect to x. This gives us ~^u(w1,y,s) =M{^u(x,y,s )}(w1) )dx Accepted in click title to verify. Published under CC-BY 4.0. )}(w1) )}(w1). (59) We shall need to employ =w(1 +w)M{f(x)}(w) to evaluate the terms appearing in (59). Using these Mellin transforms leads us to ~^u(w1,y,s) +B sw1(1 Mellin transform of ~^u(w1,y,s), this time with respect to y, +B sw1(1 we solve the above shall now bring in (58). Strictly speaking it is not necessary to use them and there is some loss of generality. However, it makes the problem much easier and using more generality in the coefficients is straightforward. Thus using (58)we see that (61) The task, now that we have terms of w1,w2ands, is to revert back all the transforms and translate the answer to the space of solutions of (53). Apply the inverse Laplace transform L-1{1 s-a}(t) =eatto(61)so that we end up =xw1-1 0yw2-1 0 in click title to verify. Published under CC-BY 4.0. 40wherea=Atw 2,andb=Bt. To invert the Mellin transform with respect to xwe use the following G(s) the inverses of G1andG2are given by [68] g1(x) =M-1{eas}(x) =d(a-logx) as well as g2(x) =M-1{ebs2}(x) the inverse of Gcan be found by computing the convolution g(x) Moving on from ~~u(w1,w2,t)we invertw1back intoxand see that ~u(x,w 2,t) =e-rtyw2-1 0 we have recycled aandbto now Lastly, we Mellin invert w2back intoyand we arrive at u(x,y,t ) setting = 4BC-A2, and doing some mild algebraic re-arrangements we end up with u(x,y,t ) is the solution we were seeking. G Moments While obtaining the full solution uto the PDF (19)generated by the Feynman-Kac al- gorithm is an expensive problem, familiar within the field of quantum metrology, we may nevertheless extract information about the moments of uwith a quadratic advantage using Accepted in click title to verify. Published under CC-BY 4.0. 41QPE which are deep circuits, see Section 3.3. Alternatively we could use Hadamard tests which are not as efficient but they consist of shallower circuits. This section generalizes to several dimensions some of the concepts introduced in [12], which then get applied to the system of SDEs introduced in (15). Along the way, we take this chance to amend some in- accuracies from [12]. To avoid confusion with exponentiation, the labeling i= 1,2,***,D of the processes Xi twill now be subscripted and the time variable twill be that nm= 2nq/D, wherenqis the total number of qubits and Dthe number of dimensions, represents the number of grid points allocated to each dimension. Suppose we have a function f:RD-Rand consider the =xi1,X2(T) =xi2,***,XD(T) the joint probability distribution func- tion of the random variables X1,X2,***,XDin situations where we have haveDdimensions of equal length [0,xi,max] = [0,xmax]and in each one of them we split the interval spaced by the same lattice spacing xi= x=hfori= 1,2,***,Dand the total number of qubits gets evenly allocated to each of the Ddimensions. The splitting of U(i), (63) that is, we have dintervals in each dimension D, yielding a total of dDintervals in all dimensions altogether. For an integer nwe havexn=nx. We further set the notation ai,k:= us recall Taylor's expansion in several dimensions. In order to do so, we need to introduce some notation first. If a= an n-tuple of then x= The the order, or the degree, of a. Moreover =a1!a2!***an!. Thus the order of ais the same as the order of xa as a monomial or the order or aas a partial derivative. In the interval ai,k, the function f(x) assumed to be well ap- proximated by its Lth order Taylor series ( ilabels dimension and klabels interval in (65) A natural assumption that we shall use towards the end is that Li,kis the same for of each dimension and hence Li,k=L. The error in the expectation (62) due to this approximation will be discussed shortly. Due to the finiteness of x, we can find an appropriate shift to have the range of fbe positive and this will ensure that the expectation (62)will also be the unnormalized state that we are working with is given =xi1,X2(t) =xi2,***,XD(t) in click title to verify. Published under CC-BY 4.0. 42where To compute the expectation (62)we first consider the non- unitary operator (67) This operator can be decomposed into easily implementable unitaries Q, (68) wherexkdC. a calculation shows =xa1,***,XD(T) =xc1,***,XD(T) =xa1,X2(T) =xa2,***,XD(T) =xc1,X2(T) =xc2,***,XD(T) =xcD] = (69) sincefwas a real function. We may now use some of the results presented in 3.2. a sum of easily implementable unitaries, and also be implementable as such a decomposition by the use can write the first and last terms of the left-hand side of is the sum of tensor products of unitaries. Thus in (69)we will only have expressions of the techniques will yield the quantities in (71). One could either use the Hadamard test circuits depicted in this section or use QPE [69, 70]. The circuit for the Hadamard test is given by Fig 12. Accepted in click title to verify. Published under CC-BY 4.0. U Figure 12: This circuit evaluates the Hadamard tests for (62), the real and imaginary parts. Here theU's are gates such that shall get to the pros and cons of each technique towards the end of this it is noteworthy that the circuit has the advantage of being shallow but not efficient in the terms of the number of measurements needed. The next step is to define the (72) where the interval indicator function khis defined by kh[a,b](x) must now connect Sfcoming from Note (73) In order to perform the sums over xij's we use a multidimensional Taylor in click title to verify. Published under CC-BY 4.0. (74) The operator Sfis extracted from this and seen to given by (31). We now note that if D= 1, then(74)becomes Sf|0/angbracketright=d-1/summationdisplay k1=0nm-1/summationdisplay i1=0f1,kj(xi1)Skha1,k1|0/angbracketright=d-1/summationdisplay k=02n-1/summationdisplay i=0fk(xi)Skhak| (76) since in this case there is only one dimension so that f1,kcan simply be written as fk,a1,k isakand all qubits are assigned to the same dimension. Moreover for one dimension (75) reduces we are now dealing an ordinary Taylor series of Lth order in one dimension fk(x) This means that the sum and each Qiis composed of at most O(n4)gates because the sum respectively, each composed of respectively. Equations inaccurately given in [12, Eqs. (39), (40) and (B24)]. Let us now return to the error term of the expectation (62)resulting from the Taylor expansion of the function f. We now assume, for simplicity of notation, that all the L's in each dimension are the same. The Lagrange form of the error term of the Lth Taylor series around given by RL,k(x,a) (78) where c= (c1,c2,***,cD)is such that ci[xi,(ki+ 1)h]. For every xixand ev- eryhkiwith same index iwe have that Hence we see that RL,k(x,a) =O(hD(L+1)). LetELbe the expectation that would result from approximating fwith theLth order Taylor series, in click title to verify. Published under CC-BY 4.0. the continuous joint probability distribution function. In order to secure the error below a desired threshold e, we need to take d>x final issue we need to resolve is the construction of the operator Sfand its com- plexity. This construction is explained in [12] for one dimension, and we now give the Ddimensional generalization needed for the rest of the bounds on the complexity of the algorithm. For any given i= 1,2,***,Dwe have that ai[0,xi,max]and hence there exists an integer kaiwith 0<kai<=n/Dsuch that The binary expansion of Next, we define the matrix of elements 1in ascending order for each i, and we define the collection of member of the matrix L. These intervals allow us to divide a dimension disjoint intervals as [0,ai/x] = (82) The effect of the operator the interval Saikhi,jis defined the operator Xtakes two different values according 1, I,ifs= 0. Accepted in click title to verify. Published under CC-BY 4.0. 46For the last interval on the right, the action of the operator is defined (84) Thus the interval Skh[0,ai]is constructed by summing over Lastly, which is a sum of at most O(n)unitaries last part is the bound on the computational complexity in several dimensions. In order to upper bound the error ecoming from shall upper bound the error e/primeof each of the terms in (70)and then find the number of measurements and gate complexity needed to achieve this error. Following [12, SSA] we assume be written as a linear combination of NUjunitaries for each dimension j= 1,2,***,D, unitary operators . (85) The normalized state =xj1,X2(t) =xj2,***,XD(t) (86) The errore/primeis defined as the error of the expectation value of each term in a state implies that the estimated expectation value of each term error in the linear combination of expectation values is determined using (88) If we write ~Efor the estimation of the expectation E, then by the use of (89) Accepted in click title to verify. Published under CC-BY 4.0. 47Therefore, in order to obtain an upper bound for ein the expectation E, the such (90) In terms of the Hadamard test circuit above, one requires to bound the error in the expectation value to e/prime. Up to the preparation of the state, the depth of the unitaries in the circuit is O(1). The trade-off with QPE is that the same accuracy in the expectation is O(log 1/e/prime)but the depth is now O(1/e/prime). With the Lth order Taylor approximation, one has that 1,2,***,D, whered was the number of intervals. Moreover, the total number of measurements is the number NUj. total number of measurements by Hadamard test is whereas using QPE, one would need log(1/ge)), but QPE needs gfrom(90). we see that Next, using (68)we (91) Moreover, from (31)we see that =O(x|b| max). For an interval in a Skha /lscript, such that 0<kai<=n/D. This implies that the largest |x/lscriptj|is as large Therefore, we O/parenleftbiggD/productdisplay j=1[max kjmax |b||fj,kj,|b||x|b| max]2d2n2L+2xmax/parenrightbigg , (92) which is the last bound we needed for g. The operator be the sum of and each Qk will be composed of O(n4)gates. The number of gates will grow exponentially with the dimensions and this is a universal problem in many algorithms, such as the use of QAE when computing moments when the arguments of the function are In such cases one needs exponentially many resources to build a multidimensional smaller while d2DnD(2L+2)grows larger. This means that the number of unitaries is not monotonic with respect to Land thus there In[12], of those unitaries could be computed in parallel with a second quantum device. H Future work As discussed in Section 5, some of the following areas of research might have to be pursued using a combination of real and imaginary time in click title to verify. Published under CC-BY 4.0. 48H.1 Stochastic volatility and stochastic interest rates A key weakness in the Black-Scholes is that the volatility term sis constant. One can go around this problem by defining two SDEs: geometric Brownian motion for the (93) wheredW1is standard Brownian motion, as well as another SDE for s dst=p(S,s,t )dt+q(S,s,t )dW2,t, LetV(S,s,t )be the value of an option on the underlying Sat timetand with s. the partial differential )is a universal function that is used in portfolio replication to ensure that options are only dependent on the variables S,sandt. The function lis called the market price of volatility risk. Comparing (95)with 0, (96) we see the appearance of the additional linear and second order partial derivatives with respect toscoming from (94)as well as the presence of the mixed term in Sandsdue to the correlation rof the two Brownian motions dW1anddW2. Note that if the volatility were to be constant, then (95)would reduce to (96). Different choices of pandqlead to different named models, e.g. *Hull and White: d(s2) model / Heston: d(s2) = model: d(s2) = d(logs2) = (a-blogs2)dt+cdW 2. see [36, p. 861]. Similar equations also exist for stochastic interest rates [36]. H.2 American and options The American option is also an involved one to price. Unlike in the European set-up, the holder of an American option may exercise at any time prior and up to expiration. The details of this set up can be found in [63, SS2]. We are dealing with our usual motion (93)as well as the additional Xq 0=X0, (97) whereqtis the number of shares held at time t. The strategy of qtis subject to the contractual constraint qt[at,bt], whereat<=bt. Financial arguments lead us to the conditional = t[0,T]. Accepted in click title to verify. Published under CC-BY 4.0. 49

2108.10846

Nicolas Robles

Hedayat Alghassi, Amol Deshmukh, Noelle Ibrahim, Nicolas Robles, Stefan Woerner, Christa Zoufal

A variational quantum algorithm for the Feynman-Kac formula

Quantum 6, 730 (2022)

10.22331/q-2022-06-07-730

quant-ph

http://creativecommons.org/licenses/by/4.0/

We propose an algorithm based on variational quantum imaginary time evolution for solving the Feynman-Kac partial differential equation resulting from system of stochastic differential equations. We utilize between the Feynman-Kac partial differential equation (PDE) and the Wick-rotated Schr{o}dinger equation for this purpose. The results for a $(2+1)$ dimensional Feynman-Kac system obtained through the variational are then compared against classical ODE solvers and Monte We see a remarkable agreement between the classical methods and the quantum variational method for an illustrative example on six and eight qubits. In the non-trivial case of PDEs which are preserving probability distributions -- rather than preserving the $ell_2$-norm -- we introduce a proxy norm which is efficient in keeping the solution approximately normalized throughout the evolution. The algorithmic complexity and costs associated to this methodology, in particular for the extraction of properties of the solution, Future research topics in the areas of quantitative finance and other types of PDEs are also discussed.

1 collisions at the Relativistic Heavy Ion Collider (RHIC) and the Large Hadron Collider (LHC) have provided strong evidence indicating the formation of the strongly coupled quark-gluon plasma (QGP), a deconned state of hadronic matter whose dynamics after the collision is dominated by non-perturbative eects [1{8]. While the perturbative QCD works only in the weak coupling regime, the string/gauge duality can be used as the powerful tool fo study the strongly coupled plasma [9{13] . However, there is no known gravity dual to QCD, the duality between the N= 4SU(Nc) super-Yang-Mills theory and type IIB string theory on AdS 5S5is the most studied example in the context of AdS/CFT correspondence [7, 14, 15] which results. For example, the ratio of shear viscosity over entropy density obtained from this duality, =s= 1=4[5, 16, 17] is consistent with the experimental data [18]. Lattice data suggests that the QGP formed at high-energy heavy-ion collisions is not a fully conformal uid, and bulk viscosity, which is a purely non-conformal eect, is needed for the precise extraction of the shear viscosity of the QGP [19]. Hydrodynamics eects successfully describes the smaller system such as p-Pb [20] and p-p [21, 22] collisions [23]. The original duality can be extended to the theories more similar to QCD or QGP using the well-known top-down [24{26] or a bottom-up [27, 28] approaches. One example of the latter approach is a ve-dimensional gravity model coupled to a scalar eld with a non-trivial potential [29]. In this model, the conformal invariance breaks even at zero temperature by coupling a scalar eld at pure gravity in AdS, which duals to a CFT deformed by a source  for a dimension-three operator. This source breaks the scale invariance of the resulting four-dimensional strongly coupled gauge theory explicitly and triggers a non-trivial renormalization Group (RG) ow from an ultraviolet (UV) xed point to an infrared (IR) xed point. The UV xed point assured that we are in the regime where the holographic duality is best understood and the bulk metric is asymptotically AdS. The IR xed point is needed to guarantees that the solutions are regular in the interior and solution is smooth in the deep IR. This simple model has been used to study dierent properties such as thermodynamics and relaxation channels [29], jet energy lost [30], and entanglement entropy [31]. One of the most important quantities of QGP is the energy loss of a quark moving through the plasma. An external quark was introduced in the context of AdS/CFT cor- respondence by adding a fundamental string attached to a avor brane in the AdS space [32{34]. The string endpoint species the heavy quark, while the string itself can be con- sidered as a gluonic cloud around the quark. It is straightforward to show that the mass of quark is proportional to the inverse distance of the string endpoint from the boundary [32]. The light quark or massless quark is mapped into a string attached to a avor brane which is extended from the boundary to the horizon, and its dynamic in dierent have been studied in [30, 35{37]. On the other hand, a string attached to the boundary of AdS space is dual to an innite mass quark. Since the rst attempts to study the holographic heavy quark in N= 4 super-Yang-Mills theory [32, 33], the heavy quark dynamics have been studied in various gauge theories with gravitational duals [38{61]. 1This paper is organised as follows: In section 2 we brie y review the introduced in [29]. In section 3 we discuss the string solution dual to a heavy quark in this non-conformal background. The numerical results of drag force and world- sheet temperature are presented in 3.1 and 3.2. Section 4 is devoted to a summary. 2 Non-conformal holographic model The action of 5-dimensional Einstein gravity coupled to a scalar eld is given (2.1) where5is the ve-dimensional Newton constant and V() is a potential encoding the details of the dual gauge theory. In [29], the following potential has been (2.2) which is characterized by a single parameter, M. This non-trivial relatively simple po- tential has a maximum at = 0 and a minimum at to two AdS solutions at UV and IR xed points. The radii of these two solutions are related as LIR=1 1 +1 62 ML: the number of degrees of freedom is smaller at IR. By parametrizing the vacuum metric (2.4) the vacuum solution can be analytically obtained for arbitrary (2.5) (r) =Ler=L q 1 (2.6) It is shown that the arbitrary parameter  is responsible for breaking the conformal in- variance explicitly [29]. The black brane solution of action 2.1 can be calculated numerically by the following ansatz in the being the radial coordinate such that the boundary and the horizon are located at = 0 and=H, respectively. In fact, the value of the scalar eld at the horizon, Hcharacterizes the black-brane solution and its thermodynamic p/L/() (b) Figure 1 : (Color online) (a) The ratio of non-conformal to the conformal entropy,(b) the ratio of speed of sound square to its conformal limit as a function of temperature for dierent values of M. explicitly. The dual gauge theory is conformal both at the UV and at the IR. Therefore the high and low temperature behavior of thermodynamical variables must match with the relativistic conformal theory. The entropy ratio to its conformal value, s=scon, and the ratio of speed of sound for dierent values of Mis plotted in Fig. 1a and Fig. 1b, entropy ratio approaches one at high temperatures and ( LIR=L)3at low speed of sound reaches its conformal value both at high and low temperatures while deviating from its conformal value at intermediate temperatures, which can be interpreted as a measure of the non-conformality of the gauge theory. The detail of the is presented in [29]. 3 Classical trailing string This section studied a heavy quark moving through an innite volume of with a xed velocity vat a nite temperature T. The heavy quark feels a drag force and consequently losses energy. In the dual theory, the dynamic of this quark can be described by an open string whose endpoint is attached to the UV boundary of AdS space and moves at the constant speed of vwith a string tail into the AdS bulk [32, 33]. The dynamics of the string is described by the Nambu-Goto the square of string's fundamental length and is the induced metric of the world-sheet dened the metric component of Eq. (2.7) in the Poincare =e2A();G() =L2 h()e2B(): we assume L= 1 for simplicity. By considering the following ansatz for the string embedding function in the static gauge ( =t;=) XM(;) = (3.4) the induced metric on the world-sheet becomes =e2A (3.5) leads to the corresponding (3.6) Since the Lagrangian does not depend on the explicitly, the string equation of motion simplies (3.7) whereis the constant of motion. From above, the equation for 0can be obtained as 0() (3.8) The numerator and denominator inside the square root are positive near the boundary and negative around the horizon. Requiring the real values for the string prole, 0means that the numerator and the denominator have to change their sign at the same point 0: (3.9) From the above equations, the constant of the equation of motion can be determined the into Eq. (3.8) yields to 0() (3.11) The square root is now well-dened for the region outside the horizon (0 <<H), and this equation can be solved numerically to obtain the string prole. 43.1 Drag force The drag force, which indicates the dissipation of quark momentum into the plasma, is given by Fdrag=1 x; the momentum that is lost by owing from the string to the horizon. The canonical momentum densities associated with the string can be obtained from varying the action with respect to the derivatives of the embedding functions (3.13) the 0 tand0 xalong the string gives us the total energy and the total momentum in the direction of motion of the string, respectively. While, 1 tand1 xare the energy and momentum ow along the string and similar to the case of N= 4 SYM plasma, 1 t=v1 x. By substituting the string solution, Eq. (3.11) in the Eq. (3.14), it is evident that the 1 x equals the constant of the equation of motion, . This means that if we pull the quark with a constant velocity, the fraction of energy ows at a given point along the string, 1 t is constant. Using the Eq. (3.10), and Eq. (3.12) the drag force (3.15) which is the momentum ow along the string at point s. If we use the metric of AdS-Sch, the drag force for N= 4 SYM plasma can be obtained as (3.16) where we have used the relation L4=02. In the case of non-conformal background, the rst equation of Eq. (3.9) should be solved numerically to obtain the sand then the drag force can be calculated using Eq. (3.15). Our numerical results for the drag force as a function of temperature for two values of quark velocity are plotted in Fig. 2. The drag force in this non-conformal plasma is smaller than the drag force in N= 4 SYM plasma with the same temperature and decreases by increasing the degree of M. Also, the drag force increases by increasing the quark velocity. In Fig. 3a, the drag force ratio to its conformal limit is plotted for dierent values of Mas a function of temperature. The solid and dashed lines represent the drag force for v= 0:9, andv= 0:5 respectively. As is evident from this gure, at where the background reaches its conformal limit, the ratio of drag force tends to one. In addition, the ratio of drag force to its conformal value vs velocity is presented for dierent values of Min Fig. 3b to explore the velocity dependency. The temperature in 5= = ph= ph= ph= ----- p/L(a) = = ph= ph= ph= ---- p/L (b) Figure 2 : (Color online) Drag force as a function of temperature for dierent values of Mcompared with the drag force in N= 4 / (b) Figure 3 : (Color Online) The ratio of drag force to its conformal value for dierent values ofMas a function of (a) temperature, (b) velocity. The solid and dashed lines in (a) representv= 0:9 andv= 0:5, respectively. The temperature of both plasma in (b) sets to 350 MeV. bothN= 4 SYM plasma and non-conformal plasma was set to 350 MeV. We nd that for all values of Mthe ratio is less than one at the xed temperature but increases slowly by increasing the velocity of quark. 3.2 World-sheet temperature One can reparametrize the world-sheet coordinate, Eq. (3.4) as +(): the background induced metric, Eq. (3.5) becomes =e2A (hv2) (3.18) 6By choosing a particular ansatz of the form of K0() =v0 hv2; (3.19) and using Eq. (3.10), and Eq. (3.11), the induced metric diagonalizes as =e2A() (h()v2) (3.20) which describe the metric of a two-dimensional world-sheet blackhole with a horizon radius ofs. The Hawking temperature of the world-sheet blackhole indicated as +h0(s)): (3.21) In the conformal limit, metric, the world-sheet temperature depends on the blackhole temperature and the velocity of quark (3.22) In Fig. 4, the ratio of world-sheet temperature to its conformal value is plotted in terms of the plasma temperature for dierent values of M. The world-sheet tempera- ture decreases by increasing the value of M. The deviation from conformal behaviour is dominant at intermediate temperature. At high temperatures, the ratio tends to one, as expected. 4 Summary This paper studied the dynamics of a heavy quark moving through a strongly coupled plasma with broken conformal symmetry using the gauge/gravity duality. We considered a holographic ve-dimensional model consisting of Einstein gravity coupled to a scalar eld with a non-trivial potential corresponding to a dual four-dimensional non-conformal gauge theory that exhibits a renormalization group ow between two dierent types of xed points (located at UV and IR) at zero temperature [29]. The parameter Mindicates the deviation from conformality, as shown in Fig. 1. According to the AdS/CFT dictionary, the heavy quark is associated with a string attached to the boundary of the AdS space. The string equation of motion was solved numerically for dierent values of parameter M, and then the drag force was computed by obtaining the corresponding Our results indicate that the drag force in this non-conformal plasma is smaller thanN= 4 super-Yang-Mills plasma. By increasing the level of non-conformality (the value of parameter M), the drag force decreases, see Fig. 2. The ratio of drag force to its conformal limit reaches one at high temperatures. At nite temperature, this ratio is smaller than one for dierent values of Mbut increases slowly by increasing the velocity of quark, see Fig. 3. 7ph= ph= ph= p/L/Figure 4 : (Color Online) The ratio of world-sheet temperature to its conformal value in terms of the plasma temperature for dierent values of M. The world-sheet horizon was also calculated, and it was shown that the drag force is equivalent to the energy ow at the world-sheet horizon. The corresponding world- sheet temperature has been calculated and compared with its conformal limit. Results conrm that the deviation from the conformal limit is dominated at the authors would like to thank Dr. Can Koz caz for his useful discussion and comments. R. Morad and T. Domurcukgul acknowledge the Scientic and Technological Research Council of Turkey (T ubitak), the Department of Science Fellowships and Grant Programs (BIDEB) for the fellowship program for visiting scientists (2221) and ARDEB 1001 { Support Program for Scientic and Technological Research Projects, J. Adams et al. [STAR], Nucl. Phys. A 757(2005), [nucl-ex]]. [2] K. Adcox et al. [PHENIX], Nucl. Phys. A 757(2005), [nucl-ex]]. [3] I. Arsene et al. [BRAHMS], Nucl. Phys. A 757(2005), [nucl-ex]]. 8[4] B. B. Back et al. [PHOBOS], Nucl. Phys. A 757(2005), [nucl-ex]]. [5] G. Policastro, D. T. Son and A. O. Starinets, Phys. Rev. Lett. 87(2001), [hep-th]]. [6] P. Kovtun, D. T. Son and A. O. Starinets, Phys. Rev. Lett. 94(2005), [hep-th]]. [7] R. Baier, Y. L. Dokshitzer, A. H. Mueller, S. Peigne and D. Schi, Nucl. Phys. B [hep-ph]]. [8] K. J. Eskola, H. Honkanen, C. A. Salgado and U. A. Wiedemann, Nucl. Phys. A [hep-ph]]. [9] J. M. Maldacena, Adv. Theor. Math. Phys. 2(1998), 231-252 [hep-th]]. [10] E. Witten, Adv. Theor. Math. Phys. 2(1998), 253-291 [hep-th]]. [11] S. S. Gubser, I. R. Klebanov and A. M. Polyakov, Phys. Lett. B 428(1998), [hep-th]]. [12] O. Aharony, S. S. Gubser, J. M. Maldacena, H. Ooguri and Y. Oz, Phys. Rept. [hep-th]]. [13] J. H. Liu, D. Mateos, K. Rajagopal and U. A. [arXiv:1101.0618 [hep-th]]. [14] E. Shuryak, Prog. Part. Nucl. Phys. 62(2009), 48-101 [hep-ph]]. [15] E. V. Shuryak, Nucl. Phys. A 750(2005), 64-83 [hep-ph]]. [16] P. Kovtun, D. T. Son and A. O. Starinets, JHEP 10(2003), [hep-th]]. [17] A. Buchel and J. T. Liu, Phys. Rev. Lett. 93(2004), [hep-th]]. [18] D. Teaney, Phys. Rev. C 68(2003), 034913 [nucl-th]]. [19] S. Ryu, J. F. Paquet, C. Shen, G. S. Denicol, B. Schenke, S. Jeon and C. Gale, Phys. Rev. Lett. 115(2015) no.13, 132301 P. Bozek, Phys. Rev. C 85(2012), 014911 B. Schenke and R. Venugopalan, Phys. Rev. Lett. 113(2014), [arXiv:1405.3605 [nucl-th]]. [22] M. Habich, G. A. Miller, P. Romatschke and W. Xiang, Eur. Phys. J. C 76(2016) no.7, [nucl-th]]. [23] S. Jeon and U. Heinz, Int. J. Mod. Phys. E 24(2015) no.10, [hep-ph]]. 9[24] J. Polchinski and M. J. Strassler, [hep-th]]. [25] A. Karch and E. Katz, JHEP 06(2002), 043 [hep-th]]. [26] T. Sakai and S. Sugimoto, Prog. Theor. Phys. 113(2005), 843-882 [hep-th]]. [27] U. Gursoy and E. Kiritsis, JHEP 02(2008), 032 [hep-th]]. [28] B. Galow, E. Megias, J. Nian and H. J. Pirner, Nucl. Phys. B 834(2010), [arXiv:0911.0627 [hep-ph]]. [29] M. Attems, J. D. Mateos, I. Papadimitriou, D. Santos-Oliv an, C. F. Sopuerta, M. Triana and M. Zilh~ ao, JHEP 10(2016), [hep-th]]. [30] S. Heshmatian, and R. Morad, JHEP 03(2019), 045 [hep-th]]. [31] M. Rahimi, M. Ali-Akbari and M. Lezgi, Phys. Lett. B 771(2017), [hep-th]]. [32] C. P. Herzog, A. Karch, P. Kovtun, C. Kozcaz and L. G. Yae, JHEP 07(2006), [hep-th]]. [33] S. S. Gubser, Phys. Rev. D 74(2006), 126005 [hep-th]]. [34] J. and D. Teaney, Phys. Rev. D 74(2006), [hep-ph]]. [35] P. M. Chesler, K. Jensen, A. Karch and L. G. Yae, Phys. Rev. D 79(2009), [arXiv:0810.1985 [hep-th]]. [36] R. Morad and W. A. Horowitz, JHEP 11(2014), 017 [hep-th]]. [37] K. Bitaghsir Fadafan and R. Morad, Eur. Phys. J. C 78(2018) no.1, [hep-th]]. [38] J. Reiten and A. V. Sadofyev, JHEP 07(2020), 146 [hep-th]]. [39] U. Gursoy, E. Kiritsis, G. Michalogiorgakis and F. Nitti, JHEP 12(2009), [arXiv:0906.1890 [hep-ph]]. [40] W. A. Horowitz and M. Gyulassy, Phys. Lett. B 666(2008), [arXiv:0706.2336 [nucl-th]]. [41] W. A. Horowitz, Phys. Rev. D 91(2015) no.8, 085019 [hep-ph]]. [42] S. Chakraborty and N. Haque, JHEP 12(2014), 175 [hep-th]]. [43] M. Lekaveckas and K. Rajagopal, JHEP 02(2014), 068 [hep-th]]. 10

2108.10853

Razieh Morad

Tolga Domurcukgul and Razieh Morad

Holographic drag force in non-conformal plasma

10.1140/epjc/s10052-022-10252-w

hep-th hep-ph

http://creativecommons.org/licenses/by/4.0/

In this study, the gauge/string duality is used to investigate the dynamics of a moving heavy quark in a strongly coupled, non-conformal plasma. The drag force in this non-conformal model is smaller than that of $mathcal{N} = Yang-Mills (SYM) plasma and decreases as the level is increased. At intermediate temperatures, the which is derived numerically by calculating the deviates from its conformal value, but at high temperatures, it tends to its conformal value.

that new dev elopments in Artificial Intelligence algorithms designed for quantum computers - also known as Quantum Artificial Intelligence (QAI) - may provide the key breakthroughs necessary to furthering the science of climate change. The resultant improvements in we ather and climate forecasts are expected to cascade to numerous societal benefits. Keywords: Quantum Computing, Artificial Intelligence , Climate Change , Quantum Artificial Intelligence , Climate Science 1. Introduction The Earth's mean temperature has ris en steeply over the last few decades precipitating a broad spectrum of global -scale impacts such as glacier melt, sea - level rise and an increasing frequency of weather extremes. These changes have resulted from the rising atmospheric carbon pollution duri ng the industrial era because of the use of fossil fuels. The Earth mean temperature today is about 1 degree Celsius (C) higher than pre -industrial times. Recent scientific advances increasingly suggest that exceeding 1.5 degrees C may cause the Earth syst em to lurch through a cascading set of 'tipping points' - states of no return - driving an irreversible shift to a hotter world. Climate change is global yet its manifestations and impacts will differ across the planet. Therefore, quantifying future chang es at regional and local scales is critical for informed policy formulation. This, however, remains a significant challenge. We begin with a discussion on state -of-the-art science and technology on these questions and their current limitations. With this c ontext, we come to the main theme of this article which is the potential of the emerging paradigm of 'quantum computing', and in particular Quantum Artificial Intelligence (QAI) , in providing some of the breakthroughs necessary in climate science. The rest of the chapter is organized as follows. In Sect. 2, we discuss science of climate change and the role of artificial intelligence . Section 3 discusses quantum artificial intelligence for the science of climate change . Section 4 concludes the chapter and highlights possible future directions. 2. Science of climate change and the role of artificial intelligence Climate models have become indispensable to studying changes in the Earth's climate , including its future re sponse to anthropogenic forcing. Climate modelling involves solving sets of coupled partial differential equations over the globe. Physical components of the Earth system - the atmosphere, ocean, land, cryosphere and biosphere - and the interactions betwee n them are represented in these models and executed on high performance supercomputers running at speeds of petaflops and beyond. Models operate by dividing the globe into grids of a specified size, defined by the model resolution. The dynamical equations are then solved to obtain output fields averaged over the size of the grid. Therefore, only physical processes operating at spatial scales larger than the grid size are explicitly resolved by the models based on partial differential equations ; processes that operate at finer scales, such as clouds and deep convection, are represented by approximate empirical relationships called This presents at least two significant challenges: 1. While climate models have become increasingly comprehensive, grid sizes of even state -of-the-art models are no smaller than about 25 km, placing limits on their utility towards regional climate projections and thereby for targeted policymaking. 2. Physical p rocesses organizing at sub -grid scales often critically shape regional climate. Therefore, errors in their are known to be the source of significant uncertainties and biases in climate models. Additionally, numerous biophysical processes are not yet well understood due to the complex and nonlinear nature of the interactions between the oceans, atmosphere and land. Therefore, rapid advances are necessary to 'downscale' climate model projections to higher resolutions, improving parameteriza tions of sub -grid scale processes and quantifying as yet poorly understood non -linear feedbacks in the climate system. A significant bottleneck in improving model resolution is the rapid increase in the necessary computational infrastructure such as memory , processing power and storage. For perspective, an atmosphere -only weather model with deep convection explicitly resolved was recently run in an experimental mode with a 1km grid size. The simulation used 960 compute nodes on SUMMIT, one of the fastest supercomputers in the world with a peak performance of nearly 150 petaflops, yet achieved a throughput of only one simulated week per day in simulating a four month period. A full -scale climate model, including coupled ocean, land, biosphere and cryosphere modules, must cumulatively simulate 1000s of years to perform comprehensive climate change studies. Towards surmounting the challenge of this massive scaling up in computing power, there have recently been calls for a push towards exascale computing (co mputing at exaflop speeds) in climate research. While the technology may be within reach, practical problems abound in terms of how many centres will be able to afford the necessary hardware and the nearly GW scale power requirements of exascale computing that will require dedicated power plants to enable it. Similar bottlenecks exist for improving of sub -grid scale processes. Satellite and ground based measurements have produced a deluge of observational data on key climate variables over the past few decades. However, these datasets are subject to several uncertainties such as data gaps, and errors arising during data acquisition, storage and transmission. The emerging challenge is to process and distil helpful information from this vast data deluge. Towards overcoming these challenges to improving climate projections, we discuss recent advances in Artificial Intelligence that have enabled new insights into climate system processes. These techniques, however, are also subject to their own limitations. It is in this context that we discuss how Quantum Artificial Intelligence may help overcome those limitations and advance both higher resolution climate model projections and reduce their biases. When machines learn decision -making or patterns from the data, they gain what is known as artificial intelligence (AI). Climate science has seen an explosion of datasets in the past three decades, particularly observational and simulation datasets. Artificial Intelligence (AI) has seen tremendous devel opments in the past decade and it is anticipated that its application to climate science will help improve the accuracy of future climate projections. Recent research has shown that the combination of computer vision and time -series models effectively mode ls the dynamics of the Earth system ([10]). It is anticipated that advances in this direction would enable artificial intelligence to simulate the physics of clouds and rainfall processes and reduce uncertainties in the present systems [27]. In addition to helping augment the representation of natural systems in climate models by using the now available high quality data, AI has also been proposed for climate change mitigation applications (6]) Other areas where AI is playing a leading role are the technolo gies of carbon capture, building information systems, improved transportation systems, and the efficient management of waste, to name a few [6]. There are however limitations to the present deep learning models, for example, their inability to differentiate between causation and correlation. Moreover , Moore's law is expected to end by about 2025 as it bumps up against fundamental physical limits such as quantum tunneling. With the increasing demands of deep learning and other software paradigm s, alternate hardware advancements are becoming necessary [15]. 3. Quantum Artificial Intelligence for the science of climate change Artificial intelligence algorithms suffer from two main problems: one is the availability of good quality data and the oth er is computational resources for processing big data at the scale of planet Earth. The impediments to the growth of AI based modelling can be understood from the way language models have developed in the past decade. In the early days of their success, de velopments were limited to computer vision, while natural language processing (NLP) lagged behind. Many researchers tried to use different algorithms for NLP problems but the only solution that broke ice was increasing the depth of the neural networks. Pre sent day GPT, BERT and T5 models are evolved versions from that era. Maximizing gains from the rapid advances in artificial intelligence algorithms requires that they be complemented by hardware developments; quantum computing is an emerging field in this regard. Quantum computers (QC) represent a conceptually different paradigm of information processing based on the laws of quantum physics [28]. The fundamental unit of information for a conventional / classical computer is the bit, which can exist in one of two states, usually denoted as 0 and 1. The fundamental unit of information for a quantum computer on the other hand is the qubit, a two -level quantum system that can exist as a superposition of the 0 and 1 states, interpreted as being simultaneously in both states although with different probabilities. What distinguishes quantum from classical information processing is that multiple qubits can be prepared in states sharing strong non - classical interactions called entanglement that simultaneously sample a much wider informational space than the same number of bits, thereby enabling, in principle, massively parallel computation. This makes quantum computers far more efficiently scalable than their classical counterparts for certain classes of proble ms. One trend of quantum computing is the race to demonstrate at least one problem that remains intractable to classical computers, but which can be practically solved by a quantum computer. Google coined this feat quantum supremacy, and claimed, not wi thout controversy, to achieve it with its 540qubit Sycamore chip [18]. A research team in China introduced Jiuzhang, a new light - based special -purpose quantum computer prototype, to demonstrate quantum advantage in 2020 [19]. The University of Science and Technology of China has successfully designed a 66 -qubit programmable superconducting quantum processor, named ZuChongzhi [20]. IBM plans to have a practical quantum chip containing in excess of one thousand qubits by 2023 [21]. Artificial intelligence on quantum computers is known as quantum artificial intelligence and holds the promise of providing major breakthroughs in furthering the achievements of deep learning. NASA has Quantum Artificial Intelligence Laboratory (QuAIL) which aims to explore the oppo rtunities where quantum computing and algorithms address machine learning problems arising in NASA's missions [24]. The JD AI research center announced that they have a 15-year research plan for quantum machine learning. Baidu's open -source machine learnin g framework Paddle has a subproject called paddle quantum, which provides libraries for building quantum neural networks [25]. However, for practical purposes, the integration of AI and quantum computing is still in its infant stage. The use of quantum neu ral networks is developing at a fast pace in the research labs, however, pragmatically useful integration is in its infant stages [22, 23]. The current challenges to industrial -scale QAI include how to prepare quantum datasets, how to design quantum machin e learning algorithms, how to combine quantum and classical computations and identifying potential quantum advantage in learning tasks [26]. In the past 5 years, algorithms using quantum computing for neural networks have been developed ([3], [4]). Just as the open -source TensorFlow, PyTorch and other deep learning libraries stimulated the use of deep learning for various applications, we may anticipate that software such as TensorFlowQ (TFQ), QuantumFlow and others, already in development will stimulate ad vances in QAI 3.1 Complex problems in Earth system science: Potential for QAI Quantum artificial intelligence can be used to learn intelligent models of earth system science bringing new insights into the science of climate change. Quantum AI (QAI) can pl ay an essential role in designing climate change strategies based on improved, high -resolution scientific knowledge powered by QAI. Recent studies (for example, [13]) have attempted to develop physics schemes based on deep learning. However, these are larg ely proof -of-principle studies in nascent stages. Challenges such as the spherical nature of the data over Earth, complex and non -linear spatio -temporal dynamics and others exist in AI for improved climate models. Various techniques such as cubed spheres and tangent planes have been proposed to address the spatial errors arising out of sphericity. QAI can further develop advanced physical schemes using AI by incorporating high -resolution datasets, more extended training, and hyper parameter optimization. A necessary condition for quantum speedup of classical AI is that the task in question can be parallelized for training. Present libraries such as TensorFlow and PyTorch offer both data and model parallelism capabilities. They have also been released for qua ntum computers and need to be further developed for industrial scale quantum computers of the future. 3.2 Technological Solutions for implementing QAI Figure 1 shows the technological solutions for implementing Quantum Artificial Intelligence for the climate science. The main technologies to empower quantum artificial intelligence for climate science includes the resources in the two GitHub repositories on Awesome Quantum Machine Learning (QML) [29,30], QML [31], Pennylane [32], QEML (Quantum Enhanced Machine Learning) [33], TensorFlowQ [34] and NetKet [35]. Figure 1: Technologies on Quantum artificial intelligence for the science of climate change 3.3 Case study on the use of QAI for climate science We demonstrate an example of the application of Quantum Artificial Intelligence for land -use land cover classification on UC Merced dataset. The dataset is first transformed to Quantum data using Pennylan e library and the training is performed. The code can be found on the GitHub repository for this article at . Figure 2: A case study on using Quantum artificial intelligence for land -use land - cover classification Figure 2 shows the working of the case study. Initially, satellite data is transformed as Quantum data using Pennylane. Then, we apply the deep convolutional classification algorithm for land -use land -cover classification. The output classes consist of forests, agricultural fields etc 4. Conclusions and Future Work Simulation studies are u sed to understand the science of climate change and are computationally expensive tools to understand the role of various forcings on the climate system. For example, recently, a study in the journal Science Advances showed how volcanic eruptions could for ce the coupling of El Nino Southern Oscillation and the South Asian Monsoon systems. Works of such kind are critical in advancing the understanding of the climate system and its response to various forcings. However, they are computationally demanding and are time -consuming to complete. Large ensemble climate simulations is an area that requires further work using Quantum computing and Quantum Artificial Intelligence. Quantum machine learning can play an important role, especially in pattern recognition for weather and climate science with problems that will benefit the most from quantum speedup being those that are inherently parallelizable. However, various challenges present themselves in designing and operating useful quantum computers. State -of-the-art implementations of quantum computers today can control and manipulate on the order of 100 qubits whereas it is estimated that any real -world applications where quantum computers can reliably outperform classical computers would require on the order of a million qubits. This presents a formidable technological challenge. Additionally, entanglement - the heart of quantum computing - is a fragile resource prone to being destroyed with even the slightest disturbance (called decoherence). Therefore, op erational quantum computers may be several years into the future. Yet, given their potential to affect a genuine paradigm shift, much effort has been invested in exploring its application to various fields and the focus of the present article concerns its possible climate science applications. In summary, quantum artificial intelligence is projected to be a powerful technology of the future. Developments in the field of quantum artificial intelligence include the development of both computer vision and seq uence algorithms capable of being implemented on large quantum computers. All these advancements would be driven by one factor, i.e. the development of high performance quantum computing hardware. Software Availability We have released the code for demons trating the application of QAI on land- use land cover classification dataset from UC Merced dataset at: References [1] Wichert, Andreas Miroslaus. Principles of Quantum Artificial Intelligence: Quantum Problem Solving and Machine Learning . World scientific, 2020. [2] O'Brien, Karen L. Climate change and social transformations: is it time for a quantum leap?. Wiley Reviews: Climate Change 7, no. 5 (2016): 618 -626. [3] Huntingford, Chris, Elizabeth S. Jeffers, Michael B. Bonsall, Hannah M. Christensen, Thomas Lees, and Hui Yang. Machine learning and artificial intelligence to aid climate change resear ch and preparedness. Environmental Research Letters 14, no. 12 (2019): 124007. [4] Jiang, W., Xiong, J. and Shi, Y., 2021, January. When Machine Learning Meets Quantum Computers: A Case Study. In 2021 26th Asia and South Pacific Design Automation Conference ( ASP-DAC) (pp. 593 - 598). IEEE. [5] Jiang, W., Xiong, J. and Shi, Y., 2021. A co -design framework of neural networks and quantum circuits towards quantum advantage. Nature communications, 12(1), pp.1 -13. [6] Rolnick, D., Donti, P.L., Kaack, L.H., Kochanski, K., Laco ste, A., Sankaran, K., Ross, A.S., Milojevic -Dupont, N., Jaques, N., Waldman - Brown, A. and Luccioni, A., 2019. Tackling climate change with machine learning. arXiv preprint [7] Reichstein, M., Camps -Valls, G., Stevens, B., Jung, M., Denzler, J. and Carvalhais, N., 2019. Deep learning and process understanding for data-driven Earth system science. Nature, 566(7743), pp.195 -204. [8] Rasheed, A., San, O. and Kvamsdal, T., 2020. Digital twin: Values, challenges and enablers from a modeling perspective . IEEE Access, 8, pp.21980 -22012. [9] [10] Weyn, J.A., Durran, D.R. and Caruana, R., 2020. Improving Data - Driven Global Weather Prediction Using Deep Convolutional Neural Networks on a Cubed Sphere. Journal of Advances in Modeling Earth Systems, 12(9), p.e2020MS002109. [11] Schiermeier, Q., 2010. The real holes in climate science. Nature News, 463(7279), pp.284 -287. [12] Britt, K.A. and Humble, TS, 2017. High -performance computing with quantum processing units. ACM Journal on Emerging Technologies in Computing Systems (JETC), 13(3), pp.1 -13. [13] Rasp, S., Pritchard, M.S. and Gentine, P., 2018. Deep learning to represent subgrid pro cesses in climate models. Proceedings of the National Academy of Sciences, 115(39), pp.9684 -9689. [14] Singh, M., Krishnan, R., Goswami, B., Choudhury, A.D., Swapna, P., Vellore, R., Prajeesh, A.G., Sandeep, N., Venkataraman, C., Donner, R.V. and Marwan, N., 20 20. Fingerprint of volcanic forcing on the ENSO -Indian monsoon coupling. Science advances, 6(38), p.eaba8164. [15] Marcus, G., 2018. Deep learning: A critical appraisal. arXiv preprint [16] [Eeckhout, L., 2017. Is moore's law slowing down? what's next?. IEEE Micro, 37(04), pp.4 -5. [17] Singh, M., Kumar, B., Niyogi, D., Rao, S., Gill, S.S., Chattopadhyay, R. and Nanjundiah, R.S., 2021. Deep learning for improved global precipitation in numerical weather prediction systems. arXiv preprint arXiv:210 6.12045. [18] Arute, F., Arya, K., Babbush, R. et al. Quantum supremacy using a Programmable superconducting processor. Nature 574, 505 -510 (2019). -019-1666 -5 [19] Quantum computational advantage using photons, Science, 2020: Vol. 370, Issue 6523, pp. 1460 -1463 DOI: [20] Yulin Wu, Wan -Su Bao, Sirui Cao et al. Strong quantum computational advantage using a superconducting quantum processor, ArXiv, [21] IBM's roadmap for scaling quantum te chnology, -quantum -roadmap/ [22] Beer, K., Bondarenko, D., Farrelly, T.et al. Training deep quantum neural networks. Nature Communic ation 11,808 (2020). -020-14454 -2 [23] Fanjie Kong, Xiao -Yang Liu, Ricardo Henao, Quantum Tensor Network in Machine Learning: An Application to Tiny Object Classification. NIPS workshop 2020 [24] NASA Quantum Artificial Intelligence Lab oratory. [25] Paddle Quantum. [26] Huang, HY., Broughton, M., Mohseni, M. et al. Power of data in quantum machine learn ing. Nature Communications 12, 2631 (2021). -021-22539 -9 [27] Manmeet Singh, Shreshth Tuli, Rachel Jane Butcher, Rupinder Kaur, and Sukhpal Singh Gill, Dynamic Shift from Cloud Computing to Industry 4.0: Eco -Friendly Choice or Climate Change Threat, P. Krause and F. Xhafa (eds.), IoT -based Intelligent Modelling for Environmental and Ecological Engineering, Lecture Notes on Data Engineering and Communications Technologies 6 7 , Springer, pp. , 2021. [28] Sukhpal Singh Gill, Adarsh Kumar, Harvinder Singh, Manmeet Singh, Kamalpreet Kaur, Muhammad Usman, and Rajkumar Buyya, Quantum Computing: A Taxonomy, Systematic Review and Future Directions, Software: Practice and Experience (SPE) , Volume 52, Issue 1, Pages: 66-114, January 2022. [29] -quantum -machine - learning [30] -quantum -ml [31] [32] Bergholm, V., Izaac, J., Schuld, M., Gogolin, C., Alam, M.S., Ahmed, S., Arrazola, J.M., Blank, C., Delgado, A., Jahangiri, S. and McKiernan, K., 2018. Pennylane: Automatic differentiation of hybrid quantum - classical computations. arXiv preprint [33] Sharma, S., 2020. Qeml (quantum enhanced machine learning): Using quantum computing to enhance ml classifiers and feature spaces. arXiv preprint [34] Broughton, M., Verdon, G., McCourt, T., Martinez, A.J., Yoo, J.H., Isakov, S.V., Massey, P., Halavati, R., Niu, M.Y., Zlokapa, A. and Peters, E., 2020. Tensorflow quan tum: A software framework for quantum machine learning. arXiv preprint

2108.10855

Manmeet Singh

Manmeet Singh, Chirag Dhara, Adarsh Kumar, Sukhpal Singh Gill and Steve Uhlig

Quantum Artificial Intelligence for the Science of Climate Change

cs.AI physics.ao-ph quant-ph

http://creativecommons.org/licenses/by/4.0/

Climate change has become one of the biggest global problems the Earth's habitability. Recent developments such as heat waves in California & Canada, and the devastating floods in Germany point to the role of climate change in the ever-increasing frequency of extreme weather. Numerical modelling of the weather and climate have seen tremendous improvements in the last five decades, yet stringent to be overcome. Spatially and temporally localized forecasting is the need of the hour for effective adaptation measures towards minimizing the loss of life and property. Artificial methods are results in improving predictions, but are still limited by the availability of requisite hardware and software required to process the vast deluge of data at a scale of the planet Earth. Quantum computing is an that has found potential applicability in several fields. In this opinion piece, we argue that new developments in Artificial designed for quantum computers - also known as Quantum (QAI) - may provide the key breakthroughs necessary to furthering the science of climate change. The resultant improvements in weather and climate forecasts are expected to cascade to numerous societal benefits.

2 Figure 1: Chemical and magnetic unit cells of bulk CoTiO 3. (a) The rhombohedral and hexagonal cells that reproduce the crystalline chemical periodicity. (b) Conventional and primitive magnetic cells exemplied using found G-AFM magnetic conguration. Ferromagnetic hexagonal abplanes of cobalt coupled in the (doubled) hexagonal c-axis. This gure was prepared using the VESTA software [10]. II. THEORETICAL APPROACH A. Computational delocalization within DFT may lead to an incorrect description of the magnetic properties. In particular, for systems with localized electrons such as d-electrons of transition metals acting as dopants or constituting a component of transi- tion metal oxides, Coulomb interaction eects may lead to qualitatively dierent results [11{13]. The DFT+U method is one approach that aims to correct the ten- dency of DFT towards itineracy by explicitly correcting the Coulomb interaction with a Hubbard-like interaction for a subset of states in the system[9]. By including the on-site Coulomb interaction Uand exchange the non-integer or double occupation of these states is penalized, thus localizing them in the DFT calculations were performed using the Vi- enna Ab-initio Software Package (VASP)[14, 15] using the projector augmented wave method (PAW). We em- ployed the GGA for exchange using (PBE) approach. Extra interactions are taken into account with the GGA+U approach implemented in the code. We em- ployed the simplied (rotationally invariant) approach by Dudarev et al. [16], which includes the UandJterms as an eective For brevity, and unless stated otherwise, we refer to this the rest of the paper. The electrons Co(3 p, 3d,4s), Ti(3p, 3d, 4s) and O(2s, 2p) were treated as va- lence states. Tests using all-electron calculations were conducted to check that the number of valence electrons per element were properly considered, as described in Appendix A. For most of the calculations presented in this paper, the total energy of the system was converged with re- spect to the plane-wave cuto energy and reciprocal space samplings. The convergence criterion was less than 1 meV/atom, and we found that a plane-wave cuto of 800 eV, and a -centered 8x8x8 Monkhorst-Pack k- point mesh to yield results within the stated precision. In the spin-orbit calculations, where the energy dier- ences are on the order of 101meV, additional conver- gence tests for the anisotropy ener- gies (MAE) with respect to the reciprocal space sampling were performed to ensure numerically precise results (see Appendix E). B. Chemical and magnetic structures We rst discuss the dierence between the chemical and magnetic structures of bulk cobalt titanate. The compound CoTiO 3is reported to have an ilmenite with trigonal space group R3, which consists of alternating layers of corner sharing CoO 6and TiO 6 octahedra, stacked along the c-axis in the hexagonal set- ting, as shown in Fig. 1(a)[5, 7]. The experimental are a= 5.48 A and= 55in the rhombohe- dral setting, and a=b= 5.06 A andc= 13.91 A in the3 hexagonal setting. The Co, Ti and O atoms are located at the Wycko positions (0,0, 0.355), (0, 0, 0.146) and (0.316, 0.021, 0.246), respectively [6]. The CoTiO 3magnetic conguration is reported as G- type ordering below the N eel tem- perature of 38 K [5{8, 17]. This conguration consists of coupled hexagonal ab-planes, anti- coupled along the c-axis, as shown in Fig.1b. It should be noted that in order to reproduce the periodicity of this magnetic cell in the implemented in the ab initio codes, building a cell larger than the chemical rhombohedral or hexagonal cells is needed. Although the c-axis doubled hexagonal cell, which con- sists of 60 atoms, is a straightforward candidate, there is a primitive magnetic cell of just 20 atoms that still sat- ises this periodicity[18]. It can be dened by means of the 2 2 21 2 11 23 52 4a b c3 5; (1) where a,bandcare the hexagonal lattice vectors. We refer to this cell as the primitive magnetic cell, and unless stated otherwise, all the calculations in this paper are performed in this RESULTS AND DISCUSSION A. Structural properties Using the primitive magnetic cell in the G-AFM cong- uration, the lattice parameters, cell volume, and atomic positions were fully relaxed for a range of dierent tita- nium and cobalt Uvalues. The stability of the structure was conrmed by additional phonon calculations that can be found in Appendix B. In Fig.2(a,b), the parameter and the cell volume are plotted Panels (c) and (d) display the values of the Co-Co and Co-Ti distances. The lattice parameter cand volume Vare presented in the hexagonal setting in order to facilitate the results indicate that the GGA+U approach con- sistently overestimates the experimental lattice parame- ters, which is manifested in the volume expansion of the unit-cell with increasing Uparameters. This expansion ranges from 1.2% of the experimental cell in the bare GGA case, to a 6.44% value for the GGA+U case with (UTi=6, U Co=5). This trend is in agreement with performed for TiO 2in the framework of the DFT+U theory[19]. The volume increase is linked to the expansion of which is in turn closely related to the Co-Co and Co-Ti interatomic distances. However, the Co-Co distance decreases for small UCo, before stabiliz- ing, while the Co-Ti distance increases for all UCo. In both cases, the eect of the titanium parameter UTiis to Figure 2: Change of the structural parameters with respect to the chosen values: (a) hexagonal lattice parameter c, (b) cell volume, (c) Co-Co distance, and (d) Co-Ti 3: Band gap as a function of cobalt UCoandUTi, given in eV. decrease the distance with decreasing UTi. Note that the trends of these distances with Uare opposite in cobalt and titanium, as they are respectively above and of the d-shell. The Ti-O bonds are the key ones determining the expansion in volume.4 B. Electronic properties In order to investigate the electronic properties of the material, the band structure of CoTiO 3was calculated. As in previous sections, we focused on the eect of the parameters UCoandUTion the electronic states of the system. Figure 3 shows the change in the band gap with respect to these parameters. Figure 4 displays band structures in the GGA+U ap- proach. The density of states (DOS) is also shown in the GGA+U case, as well as a sketch of the valence and conduction bands of cobalt titanate. For comparison the band structures using GGA are included in Appendix C. In Fig.3, we observe how the cobalt UCoparameter in- creases the band gap, yielding values that range from an almost metallic state in the GGA case, to an with band gaps of the order of 3 eV. The ef- fect of the titanium parameter UTiis, in contrast, much more moderate, making appreciable dierences only in the cases with a UColarger than 3.0 eV. For Uvalues in the range'3-4 eV, the bandgap takes values slightly below 3 eV. In the GGA approach, the band gap of the system is given by cobalt d-bands close to the Fermi energy, with oxygen and titanium bands occupying the valence and conduction states, respectively. Hybridization seems neg- ligible in the highest occupied valence band and conduction bands, and cobalt can be thought of as a recombination center, as shown in Appendix C. As theUparameters increase, these cobalt bands around the Fermi energy become more localized, enhancing the energetic separation between them and increasing the in- sulating behavior of the system, as illustrated in Fig.3. In the GGA+U approach, shown in Fig. 4, these fea- tures are clearly observed. Cobalt titanate has a bandgap of 2.9 eV, and cobalt bands are integrated into the bulk continuum, mixing with oxygen and titanium in the va- lence and conduction bands, respectively. It should be noted that this mixing does not occur in a fully symmet- ric way, as the top of the valence band is characterized by Co-O states, while the bottom of the conduction band is mostly titanium-based. This explains the eect of UTiin the electronic structure, which becomes notable when the band gap is given by the titanium bands in which this pa- rameter acts. This nding would also have implications in the optical processes, as one would expect electrons to localize in the Ti-O layers, with holes concentrating in the Co-O layers, potentially leading to interesting exci- tonic behavior between the hexagonal ab-planes. In the right panel of Fig. 4, we plot the projected den- sity of states of a single cobalt atom along the densities of titanium and oxygen. We nd how the in the conduction band, while the in-plane are hybridized over a wide energy range. This spatial distribution of the electronic states gives a picture of the bonding in the crystal, with the cobalt in the Ti-Co interaction, and the rest of hybridizing with oxygen. C. Magnetic properties 1. Spin conguration To analyze the magnetic structure of cobalt titanate, we perform total energy calculations for various the G-AFM structure, the and ull- (F-AFM) structures shown in Fig.5. We nd that for all considered ( UTi,UCo) pairs, the energy ordering of the three structures is the same: the G-AFM conguration is the ground-state of the system, followed by the ferromagnetic FM state, with the full- F-AFM structure presenting a considerable higher energy. We refer to the between the G-AFM and the FM structures as
E1, and label the dierence between the G-AFM and F-AFM states as
E2. In the G-AFM state, all cobalt atoms have a local mag- netic moment ofjCoj, wherejCojranges from 2.5 B (GGA) to 2.8 B(UTi=6,UCo=5). This change in the lo- cal magnetic moment is also the cause of the due to the Uparameters, which concentrates the electronic density around the cobalt atoms as the Upa- rameters increase. The calculated magnetic moments are close to the expected S=3/2 value derived from Hund's rules, and the slight dierence can be attributed to the fact that the local magnetization is numerically com- puted by integrating in the spherical region given by the Wigner-Seitz radius, which can lead to an of the measured magnetization. However, it should not be forgotten that due to the hybridization mentioned in the previous section, cobalt presents a that modies the ionic Co2+picture. The previously dened energy dierences
E1and
E2can be linked to the thermal energy needed to in- vert the spin ordering of their respective coupling, which causes a phase transition. A critical temperature can be associated with each of these transitions, e.g. in the form ofkBTi=
Ei=NCo. The inter-layer superexchange J1 and intra-layer direct exchange J2couplings can also be calculated from these energies (see Appendix D). We get approximate values of J1= 1.33 meV and J2= -1.25 meV in theUCo=UTi=4.0 eV case. These values have not to be confused with the ones in Refs.[5, 6], which are cal- culated for dierent model Hamiltonians and other computed critical temperatures are shown in Fig.6, where a phase diagram of the system behavior is presented. For temperatures lower than the N eel tem- perature (T N), the system will exhibit the G-AFM state, which consists of ferromagnetic hexagonal abplanes an- ti coupled along the c-axis. When the temperature ranges between T Nand T C, the inter-layer ordering will be broken, but the intra-layer ferromagnetic ordering will still be present.5 Figure 4: Element- and band-structure of cobalt titanate. The size of the marks represents the fractional contribution of each element following a color code. Titanium( d) and oxygen( s+p) are shown in green and blue, respectively. In red, the contribution of the d-orbitals of a cobalt atom in a spin-up G-AFM layer is shown. Panel (c) shows the density of states (DOS) of a single cobalt atom in conjunction to the titanium and oxygen densities, as well as a sketch of the splitting of cobalt UCo=UTi= 4.0 eV were employed in this 5: (a) G-type , (b) ferromagnetic and (c) full- congurations of ilmenite CoTiO 3.J1andJ2are the inter-layer and intra-layer magnetic couplings, respectively, and
E1and
E2denote the energetic dierences between the congurations. In the (a) and (b) settings, the intra-layer coupling with (a), or ferromagnetic (b), inter-layer coupling. In the conguration (c), both couplings are . Lastly, T Cindicates the beginning of the fully param- agnetic behavior, where the thermal energy overcomes the in-layer coupling, breaking the ferromagnetic order- ing of the layers. Note that the Co ions in the paramag- netic state still present disordered local being fully spin compensated. These ndings suggest that individual layers can be ferromagnetic in the T N< T<TCrange above the N eel temperature, an interest- ing result regarding applications that might merit work.6 Figure 6: Magnetic phases of cobalt titanate Energy dierences
E1and to temperature units. T Nis the N eel temperature, which is related to the breaking of the inter-layer ordering, and is shown along the experimental value of T N=38K indicated by a dashed line. T Cis related to the breaking of the intra-layer ferromagnetic ordering, and represents the starting point of paramagnetic behavior which exists for higher temperatures. 2. Magnetic anisotropy We next consider the magnetic anisotropy due to cobalt layers in CoTiO 3bulk. In order to determine whether cobalt titanate presents an in-plane or out-of-plane magnetic anisotropy, we rst perform to- tal energy calculations including the spin-orbit term as implemented in VASP for a number of spin respect to the ferromagnetic cobalt layers. The anisotropy energy (MAE) is dened as the energetic dierence between the lowest energy and the conguration under analysis, and is given by MAE( ) =E() -Ez. Here,is the polar angle in the hexagonal ac(cartesianxz) plane. We found that the eect of the in-plane orientation was negligible, only varying the MAE in the order of eV for dierent values of the azimuthal angle within the hexagonal ab plane. In the G-AFM setting of the primitive magnetic cell, we calculated the MAE in the GGA and with UTi=3.9 eV and UCo=4.5 eV. The MAE values using GGA are larger than those for the GGA+U cases because the GGA structure is slightly compressed. In fact, the role of the structural parameters seems key Figure 7: Magnetic anisotropy of bulk cobalt titanate. (a) Angular dependence of the anisotropy, where  changes from the out-of-plane c-axis to the In red, the MAE is shown as calculated using DFT. In dark red, the Bruno model tting obtained with the calculated Lorbital moments is shown; in black, the tting to Eq.(2); in blue, the due to the dipole-dipole interaction is shown. (b) Magnetic anisotropy at the abplane (= 90) with respect to the variation of the electron number in the unit cell (N). The dipole-dipole term is shown in dashed lines. For an electron excess of around 0.2, the MAE term becomes smaller than the dipole-dipole term. as the MAE for the experimental lattice is even larger. Some comments on the eect of U in the anisotropy can be found in Appendix F. We then focus on the MAE per atom in the GGA+U case, as shown in Fig.7(a). We observe that the anisotropy is minimum in the out-of-plane hexagonal c-axis, and in-7 creases as spins align with the hexagonal abplane. This tendency is observed in both the GGA and GGA+U ap- proaches, suggesting the easy-axis character of the The MAE was also calculated for the fer- romagnetic conguration and found to be nding conrms that the ab layers have a character. In order to understand the an- gular dependence of the MAE, we t our results to =K1sin2() +K2sin4(); the Using the total energies per unit cell, our tting yields values of K1= 0.29 (0.52) meV and K2= 0.068 (0.025) meV for GGA+U (GGA) cases. The K1 value is much larger than the K2one, but not negligible for GGA+U. This indicates the strong uniaxial character of the anisotropy. The element dependence of the anisotropy can also be analyzed by tting the MAE to the Bruno model[21] given by MAE() (3) where'50 meV is the spin-orbit constant, and GS L andL() are the orbital magnetic moments of cobalt atoms in the ground-state conguration and in the axis under analysis, respectively. Our tting to Eq.(3) yielded a spin-orbit constant of '48 meV ('60 meV) in the GGA+U (GGA) approach, close to the aforemen- tioned value. These MAE values calculated from the or- bital magnetic momenta Lusing the Bruno model are in great agreement with the directly calculated DFT+U values. This agreement suggests that the MAE could be directly correlated to the angular dependence of the density around cobalt ions in CoTiO 3. In the GGA+U approach, we get values of Lbetween 0.16 and 0.19 B, in good agreement with the only-GGA values in Ref.[22]. By being non-negligible, these Lvalues are pointing to the relevance of spin-orbit coupling in these are noncollinear with Sones when the eld is not exactly aligned with the easy axis or the hard plane (see Appendix G). Previous reports point to an in-plane anisotropy in cobalt titanate [5, 7, 8, 18], which is in contrast to In order to understand this discrepancy, we also calculated the anisotropy due to the interaction[23]. This interaction is given by the miandmjare the local magnetic moments around the interacting cobalt ions pairs, and rijis the vec- tor joining the two cobalt atoms. We computed thisterm from the atomic positions and local magnetic mo- ments derived from the DFT calculations in which the spin-orbit interaction was included. Our results for the GGA+U structure are shown along the MAE in Fig. 7(a). In contrast to the spin-orbit term, the dipole term fa- vors in-plane spin orientation, and competes with the MAE term in magnitude. Nevertheless, the total mag- netic anisotropy still favors an out-of-plane orientation in our calculations. This eect presumably increases with growing temperature, as the dipole-dipole term decays faster than the MAE term (/M(T)) with the spontaneous This could lead to potential out-of-plane ferromagnetic layers in the T range. To reconcile our results with experiments, we ana- lyze the eect of doping in the system, see Fig. 7(b). This is performed by the addition and the substraction of electrons in the unit cell. Including defects in this compound explicitly implies a dierent set of calcula- tions beyond the scope of the actual paper. We nd that removing electrons (p-doping) leads to an increase of the MAE, while adding electrons (n-doping) lowers the MAE even past the dipole-dipole term. This later mechanism could be a consequence of the presence of Ti atoms at some cobalt sites in the sample, as suggested in the experimental literature[7, 25]. Our results indicate that adding 0.2 electrons (which roughly corresponds to 2.5% of cobalt sites being occupated by titanium) could be enough to turn the out-of-plane anisotropy to an in- plane anisotropy, consistent with experiment. We further remark that the eect of mesoscopic domains, suggested in the literature[5, 18], may result in domains with in- plane anisotropy. However, it should be noted that do- mains with an out-of-plane component could also lead to the compensation of the MAE, yielding an as already shown in magnetic alloys [26, 27]. In summary, we nd that crystalline bulk CoTiO 3 presents a strong out-of-plane anisotropy, due to the spin-orbit coupling of cobalt atoms. The value is larger in magnitude to that of pure hcp cobalt[28, 29], a fact that is interesting because cobalt can be seen in this compound as a Co2+ion instead of being metallic. Furthermore, the is also estimated to be signicant in this material due to cobalt ferromagnetic coupling in layers. Summing the two contributions, we observed that the presence of cobalt-titanium anti-site disorder could be responsible of the experimentally observed of the bulk CoTiO 3. IV. CONCLUSIONS In this paper we analyzed the structural, electronic and magnetic properties of ilmenite CoTiO 3in the DFT+U framework. We observed that while the addition of the U correction terms slightly expands the unit cell of the sys-8 tem, it greatly improves the description of the by partially correcting the electron and thus enhancing the semiconducting character of the the magnetism of cobalt titanate, we found that the G-AFM structure is the ground state of the sys- tem, and that there are two critical temperatures which correspond to the transition between the G-AFM structure, and to the beginning of the paramagnetic phase. The existence of at temperatures above T N, could potentially lead to interesting magnetic calculations including spin-orbit coupling indi- cate that the anisotropy would be out-of-plane, a nd- ing in contrast with experiments. However, we found that the presence of 0.2 electrons in the unit cell (which roughly corresponds to 2.5% of cobalt sites being occu- pated by titanium) could be enough to turn the out- of-plane anisotropy to an in-plane anisotropy, consistent with experiment. We believe that further such as high pressure experiments, could further deepen our understanding of the magnetic anisotropy in this material. On the theoretical front, slab and single- layer calculations seem of great interest for future inves- tigations of intriguing thin-lm systems. V. gratefully acknowledge primary funding from the National Science Foundation through the Center for Dy- namics and Control of Materials:an NSF MRSEC Agreement No. DMR-1720595, with addi- tional support from NSF DMR-1949701 and NSF DMR- 2114825. This work was performed in part at the As- pen Center for Physics, which is supported by National Science Foundation grant PHY-1607611. M. R-V. was supported by LANL LDRD Program and by the U.S. Department of Energy, Oce of Science, Basic Energy Sciences, Materials Sciences and Engineering Matter Theory Program. This work has been supported by the Spanish Ministry of Science and Innovation with PCI2019-103657. We acknowledge nancial support by the European Commission from the (GA 870114). The Basque Government sup- ported this work through Project No. IT-1246-19. M.A was supported by the Spanish Ministry of Science and Innovation through the FPI PhD Fellowship A: Tests with an all-electron method In order to check the validity of the chosen number of valence electrons to be included per element, we compare the VASP calculation with a more precise all-electron cal- culation performed with the Elk code[30]. We nd that Figure 8: F-AFM band structure in the cell, in the GGA (panel (a)) and (b)) approaches. both band structures are in great qualitative agreement, which conrms the validity of the chosen number of va- lence electrons per element in our B: Phonon calculations were performed using the supercell method implemented in the Phonopy software package[31], with VASP used as the 2nd order force-constant calculator. Calculations of the phonon supercell size were carried out on 2 x2x2 ex- pansions of the primitive-cell. For the DFT force calcula- tions, we employ spin-collinear formalism with an energy cut-o of 700 eV, a 6 x 6 x 6 k-point Monkhorst pack mesh and the DFT-D3 Grimme[32] van der Waals cor- rection method. The U parameters where chosen to be: UTi= 3.9 eV and U Co= 4.5 eV. Appendix C: Electronic structure using mentioned in the main text, the eect of the cobalt U parameter is to split the cobalt bands localized around the Fermi energy, gradually increasing the band gap and9 Figure 9: Phonon dispersion of the primitive magnetic cell in the G-AFM the hybridization with titanium and 10: (a) GGA and (b) band structures of cobalt titanate. U parameters U Co= U Ti= 4.0 eV were chosen. Color indicates the fractional character of each element in the bands, with cobalt given in red, titanium in green and oxygen in blue. These band structures were calculated using the sumo D: Calculation of the magnetic the magnetic congurations under analysis, the Heisenberg Hamiltonian H =X ijJijSi
Sj; (D1) Figure 11: MAE of the hexagonal abplane against the Brillouin Zone sampling per unit cell. Regular grids were used. yields the following energies per primitive magnetic cell: EG-AFM = (2J1+ 6J2)~S2; (D2) EFM= (2J1+ 6J2)~S2; (D3) EF-AFM = (2J16J2)~S2: (D4) Here, ~Sis the pseudospin 3/2, and J1andJ2are the inter-layer and intra-layer magnetic couplings (given in meV). From the energy dierences
E 1= E FMEG-AFM and
E 2= E F-AFMEG-AFM , we get the following ex- pressions for the (D6) Appendix E: Convergence of the MAE The MAE is a small magnitude, in our case in the order of 104eV. To ensure that our results are we calculate the MAE amplitude with respect to the Brillouin Zone sampling nk(Fig. 11). We show that the MAE has a fast convergence in this system, and that the 8x8x8 Monkhorst-Pack grid used in our a well converged anisotropy energy. Appendix F: Eect of U in the anisotropy energy As commented in the main text, we nd that the MAE values obtained in the GGA approach are larger than those using GGA+U, as shown in Fig. 12. The 12: Comparison between the MAE in the GGA and GGA+U approaches. The MAE values are given per cobalt atom. expansion induced by the U parameter plays a leading role in this trend, as the GGA structure is to the experimental cell. The GGA+U anisotropy with the experimental lattice parameters is slightly larger (MAE ab=NCo'0.145 meV), a fact that points to the structural expansion as the main responsible for a de- creasing MAE. Appendix G: Orbital moment and show the calculated orbital magnetic moment val- uesLin the GGA+U approach, as well as the an- gular dierence
between the spin and orbital mag- netic moments that arises when the spin-orbit coupling is included. Note the overall non-collinearity between spin and orbital moments unless the values are just = 0;=2 and. Appendix H: Charge density with doping The addition (substraction) of electrons in the unit cell creates an excess (decit) of charge. The charge den- sity dierences between the non doped CoTiO 3and the n-doped (panel (a)) and p-doped (panel (b)) structures are shown in Fig.14. On the one hand, electrons local-ize around titanium atoms and the d2 zorbitals of cobalt atoms. On the other hand, holes are localized around the rest of d orbitals in cobalt and the p orbitals in oxygen atoms. This trend is in good agreement with the elec- tronic band structure of CoTiO 3shown in Fig. 4, where the bottom of the conduction band consists of titanium and cobalt d2 zorbitals, while the top of the valence band is a mixture of the rest of the d orbitals of cobalt and the Figure 13: (a) Orbital moment Lof cobalt in the GGA+U approximation. (b)
between the spin and orbital orbitals of oxygen. [1] R. J. D. Tilley, Perovskites: Relation- ships (John Wiley & Sons).[2] J. H. Hwang, E. N. Son, R. Lee, S. H. Kim, J. I. Baek, H. J. Ryu, K. T. Lee, and J. M. Sohn, A study of CoTiO 3as oxygen carrier for 14: Charge density dierence for the (a) n-doped and (b) p-doped cobalt titanate with respect to the non doped case. The density is plotted in units of e=a03, whereeis the electron charge and a0is the Bohr radius. The isosurface level is set to 0.00161 e=a03in panel (a) and to 0.00118 e=a03in panel (b). Yellow and cyan denote excess and decit charge density Charge densities were plotted using the VESTA combustion, Catalysis Today 303, 13 (2018). [3] Tien-Sheng Chao, Wei-Ming Ku, Hong-Chin Lin, D. Landheer, Yu-Yang Wang, and Y. Mori, CoTiO on HSG for DRAM applications, IEEE Trans. Electron Devices 51, 2200 (2004). [4] M. Siemons and U. Simon, Gas sensing properties of volume-doped CoTiO 3synthesized via polyol method, Sens. Actuators, B 126, 595 (2007). [5] B. Yuan, I. Khait, G.-J. Shu, F. C. Chou, M. B. Stone, J. P. Clancy, A. Paramekanti, and Y.-J. Kim, Dirac Magnons in a Honeycomb Lattice Quantum XY Mag- net CoTiO 3, Phys. Rev. X 10, 011062 (2020). [6] R. Dubrovin, N. Siverin, M. Prosnikov, V. Chernyshev, N. Novikova, P. Christianen, A. Balbashov, and R. Pis- arev, Lattice dynamics and spontaneous eect in ilmenite CoTiO 3, J. Alloys Compd. , R. E. Newnham, J. H. Fang, and R. P. Santoro, Crys- tal structure and magnetic properties of CoTiO 3, 17, 240 (1964). [8] H. Watanabe, H. Yamauchi, and H. Takei, in MTiO 3(M = Co, Ni), J. Magn. Magn. Mater. 15-18 , 549 (1980). [9] S. A. Tolba, K. M. Gameel, B. A. Ali, H. A. Almossalami, and N. K. Allam, The DFT+U: Approaches, Accuracy,and Applications, in Density Functional Calculations , edited by G. Yang (IntechOpen, Rijeka, 2018) Chap. 1. [10] K. Momma and F. Izumi, VESTA3 for of crystal, volumetric and morphology data, J. Appl. Crystallogr. 44, 1272 (2011). [11] H. Raebiger, S. Bae, C. Echeverr a-Arrondo, and A. Ayuela, Control of hole localization in by axial strain, Phys. Rev. Materials 2, 024402 (2018). [12] H. J. Kulik, Perspective: Treating electron with the DFT+U method, J. Chem. Phys. 142, 240901 (2015). [13] F. Aguilera-Granja and A. Ayuela, Magnetism and dis- tortions in two-dimensional transition-metal dioxides: on the quest for intrinsic magnetic semiconductor layers, J. Phys. Chem. C 124, 2634 (2019). [14] G. Kresse and J. Furthm uller, Ecient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Phys. Rev. B 54, 11169 (1996). [15] G. Kresse and D. Joubert, From ultrasoft to the projector augmented-wave method, Phys. Rev. B 59, 1758 (1999). [16] S. L. Dudarev, G. A. Botton, S. Y. Savrasov, C. J. Humphreys, and A. P. Sutton, spec- tra and the structural stability of nickel oxide: An LSDA+U study, Phys. Rev. B 57, 1505 (1998). [17] F. Schoofs, M. Egilmez, T. Fix, J. L. M. G. Blamire, Structural and magnetic properties of CoTiO 3thin lms on SrTiO 3(001), J. Magn. Magn. Mater. 332, 67 (2013). [18] M. Elliot, P. A. McClarty, D. Prabhakaran, R. D. John- son, H. C. Walker, P. Manuel, and R. Coldea, from bond-dependent exchange and of nodal quasiparticles in a honeycomb cobal- tate, Nat. Commun. 12, 3936 (2021). [19] M. E. Arroyo-de Dompablo, A. Morales-Garc a, and M. Taravillo, DFT+U calculations of crystal lattice, elec- tronic structure, and phase stability under pressure of TiO 2polymorphs, J. Chem. Phys. 135, 054503 (2011). [20] T. Wakiyama, Magnetic anisotropy and in Physics and Engineering Applications of Mag- netism , edited by Y. Ishikawa and N. Miura (Springer Berlin Heidelberg, Berlin, Heidelberg, 1991) pp. 133{158. [21] P. Bruno, Tight-binding approach to the orbital magnetic moment and anisotropy of monolayers, Phys. Rev. B 39, 865 (1989). [22] S. Das, S. Voleti, T. Saha-Dasgupta, and A. Paramekanti, Can we realize kitaev quantum spin liquids in the J e= 1=2 honeycomb cobaltates? (2021), P. Bruno, Physical Origins and Theoretical Models of Magnetic Anisotropy (1993) Chap. 24. [24] C. Kittel, Introduction to Solid State Physics , 8th ed. (Wiley, 2004) pp. 326{328. [25] G. Shirane, S. J. Pickart, and Y. Ishikawa, Neu- tron diraction study of MnTiO 3and NiTiO 3, J. Phys. Soc. Jpn. 14, 1352 (1959). [26] J. Enkovaara, A. Ayuela, L. Nordstr om, and R. M. Niem- inen, Magnetic anisotropy in Ni 2MnGa, Phys. Rev. B 65, 134422 (2002). [27] J. Enkovaara, A. Ayuela, L. Nordstr om, and R. M. Niem- inen, Structural, thermal, and magnetic properties of Ni2MnGa, J. Appl. Phys. 91, 7798 (2002).

2108.10847

Mikel Arruabarrena

M. Arruabarrena, A. Leonardo, M. Rodriguez-Vega, Gregory A. Fiete, A. Ayuela

Out-of-plane magnetic anisotropy in bulk ilmenite $ ext{CoTiO}_3$

Phys. Rev. B 105, 144425 - Published 20 April 2022

10.1103/PhysRevB.105.144425

cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

Structural, electronic and magnetic properties of bulk ilmenite CoTiO$_3$ are analyzed in the framework of density functional theory (DFT), using the generalized gradient approximation (GGA) and approaches. We find that the G-type (G-AFM) structure, which consists of ally coupled ferromagnetic $ab$ planes, is the ground-state of the system, in agreement with experiments. Furthermore, cobalt two critical temperatures related to the breaking of the inter- and intra-layer magnetic ordering. This would result in the individual planes remaining ferromagnetic even at temperatures above the N'{e}el spin-orbit coupling is included in our calculations, we find an out-of-plane magnetic anisotropy, which can be converted to an with a small doping of electrons corresponding to about 2.5% Ti substitution for Co, consistent with experimental expectations. We thus present a study of the magnetic anisotropy in bulk will determine its magnon properties, including topological aspects.

Figure 1: Overview of our framework. a A quadtree is constructed from a given input image, image regions are extracted from each node of the quadtree and down sampled to be of size 244244pixels and stored as a bag of patches. bPatches are passed through a feature network. We used a ResNet18 [ 7] pre-trained on ImageNet [ 8] for all our experiments to encode the feature vectors. These feature vectors are then passed through a MIL framework to generate the final output of using a neural network to encode adjacent regions of a histopathology image, we propose to use quadtrees built in a top down fashion from the entire image [ 9]. Quadtrees are tree data structures, historically used in [ 10]. They are constructed by recursively partitioning a two-dimensional space into four sub-regions of equal size and storing the information of each region in a node within a tree. Each node contains information about its corresponding region and has either exactly four children or none in the case of leaf nodes. As such, the tree data structure can be used to represent the information contained in the 2D space in a more compressed, data efficient way. By traversing the tree, the information from the original space can be extracted. The depth of the quadtree can depend on the size of the input image and the distribution of information within the space that is being decomposed. If there is sufficient relevant information within a given region, the algorithm will continue adding nodes to the tree and extending its depth until this no longer holds. The maximum depth of the tree can be limited by truncating the growth at a user defined depth. Quadtrees have not been used extensively within machine learning. The analogue, octrees, have been explored for tasks such as shape and scene completion with voxel arrays [ 11,12]. Wang et al. [11] use an Neural Network (O-CNN) for 3D shape analysis. By building an octree representation of 3D shapes and restricting the computation to the leaf nodes they are able to more efficiently store the information and performance with existing methods while using less memory. Jayaraman et al. [13] applied this idea with Quadtree Convolutional Neural Network (QCNN) to sparse two-dimensional handwriting data sets resulting in more efficient memory usage and computation time compared to a standard CNN. 2However, Wang et al. 's [11,12] and Jayaraman et al. 's [13] work solely uses binary data. Wang et al's O-CNN works with positional data while Jayaraman et al's QCNN uses greyscale image data. Histopathology data is 2D RGB colour data. In this work, we propose to use the quadtree itself as an image representation while taking advantage of the compression benefits of quadtrees by applying them to RGB data. To our knowledge this is the first time quadtrees have been used for this purpose. 1.2 Quadtree image propose a new image representation built with computational pathology in mind using a type of tree data structure called a quadtree. Unlike existing digital pathology representation methods our framework uses an image is not necessarily a feature representation. We use this quadtree representation to extract patches at varied resolutions in different regions of the images. This leads to greater performance in downstream tasks compared to existing patch extraction techniques all while using less data due to the quadtree construction. We show that our method is able to identify significant regions within an image relevant to clinical diagnosis and generate an interpretable, tree contributions can be summarised as follows: 1.We propose an image representation for computational pathology images and a pipeline able to predict image level labels using a single GPU. 2.We compare several methods for constructing the quadtree representation using different colour spaces and information quantification functions. 3.We evaluate our pipeline on a histopathology colorectal adenocarcinoma (CRA) data set used to train the system to classify an image as cancerous or generate attention heatmaps to discover which regions of the image the model found significant in predicting the image level label. The remainder of the paper is organised as follows: Section 2 explains our method in depth; Section 3 details our experimental results; our discussion and conclusions from our results are stated in Sections 4 and 5 respectively. 2 The Quadtree Framework In this section, we present our method for constructing quadtrees from histopathology images and how we use them for downstream analysis. Our pipeline method consists of two main stages. For a given image, a quadtree is constructed and patches extracted from the tree's nodes. These patches are then treated as a bag of instances and used in a multiple instance learning (MIL) paradigm to generate the image level Building quadtrees A quadtree represents a partition of a two-dimensional space obtained by recursively decomposing the region into four equal quadrants and sub-quadrants where each node in the tree contains information corresponding to a given specific partition of the original space [ 9]. In our case the two-dimensional space is that of an RGB image. The works on the intuition that if an image or a sub-region within an image, represented by a node in the sufficient interesting information it should be divided further into four equally sized sub regions. If this occurs, the tree is expanded by adding four child nodes to the original node being evaluated. However, if a region does not contain sufficient interesting information, then no child nodes are added and the node is made a leaf node. When all regions have been decomposed down to leaf nodes, i.e.they do not need to be split any further, then the quadtree is complete. An example of the quadtree representation is shown in Figure 1a including an example of a decomposed image and its quadtree for use with RGB images we represent this idea with the following components: 1. A function to quantify the information within a given region which we refer to as the criterion ; 2.A threshold value which determines the amount of information a region needs to contain to justify splitting the region 1: Quadtree construction algorithm QuadTree Input : Image:i2Rmn3, threshold:t, tree depth: d, criterion:c, : quadtreeQ add nodentoQat depthd ifnhas a parent then connect parent to 4 subquadrants of equal size [ components allow us to compute whether the splitting process should occur in a region given the criterion and splitting threshold. For example, let us say we have a given region i2Rmn3, a splitting threshold tand a criterion function c. We can compute c(i)and then compare it with the threshold t. Ifc(i)>tthen this means the given region has to justify splitting the region further and expanding the overall quadtree. Conversely, if c(i)tthen the region does not have sufficient information to warrant expanding the quadtree and the given node of the tree becomes a leaf node. 2.1.1 Splitting criteria We explored several different splitting criterion functions. We used the entropy [ 15] and mean pixel value of an image region as a measure of the amount of significant information in a region. Furthermore, we used these criteria with the images in the native RGB space and two additional colour spaces. All images we used were stained with Haematoxylin and Eosin (H&E). The haematoxylin dye stains basophilic tissue structures blue, in particular nuclei are heavily stained while cytoplasmic tissue regions are lightly stained. We hypothesise that by focusing on the bluer regions of the image, this will guide the quadtree splitting method to focus on the nuclei within the images which are usually the main feature of interest for most downstream tasks. As such we explored converting the images into blue ratio space [ 16] to highlight the areas dense in nuclei. We also used colour deconvolution [ 17] to separate out the different stains as colour channels and used the Haematoxylin channel as a proxy for the number of cells. 2.1.2 Determining the splitting threshold The splitting threshold tused in the quadtree algorithm is a hyperparameter which we determined from the data set. For a given criterion function and colour space, such as entropy in RGB space, we computed the value of this criterion for each image in the data set. We then calculated the mean ()and standard deviation ()of this distribution and used these values to determine the splitting threshold. We performed an ablation study for this threshold and how it affects both qualitative and downstream performance, as described in Section 3.1. 2.1.3 Learning from quadtrees In each node of the quadtree we stored a down-sampled copy of the original image region which corresponds to that node's position within the quadtree decomposition. We used image interpolation to down-sample the image regions to all be of size 244244for later processing by a pre-trained CNN for feature extraction. This allows us to store patches at different scales from the original image. This incorporates the wider FoV present in the patches corresponding to shallow nodes in the tree while the deeper nodes contain the fine grained detail at high MIL framework Having generated a collection of down-sampled image regions from the original image using our quadtree method we treated this collection as a bag of instances or patches. This lends itself to the Multiple Instance Learning (MIL) paradigm, a weakly supervised learning approach, where the bag of instances has a single label but instance level labels 4Figure 2: Example of quadtree decompositions in different colour spaces .aImage decomposed using entropy in RGB space. bImage decomposed using mean pixel value in blue ratio space. cImage decomposed using mean pixel value of haematoxylin channel. The splitting threshold was set to the mean ( ) minus one standard deviation ( ) for each case. are not available [ 14]. The label of the bag is positive (Y= 1) if at least one instance within the bag is positive and conversely the overall bag label is negative (Y= 0) if no instance within the bag is for a given image X2Rmn3with associated label Y2f0;1gwe create a bag Bof The instance level labels ynare hidden to the learner. To classify this bag of images, we can use the following function (X) =g((f(x1);:::;f (xk))); (2) wherefis a transformation of instances xkto a embedding, is a and gis a transformation to generate the overall class probabilities for the bag. There are two main approaches to : Herefclassifies each instance individually, assigning a class label for each instance and then combines these predictions together to generate an overall prediction for the bag using operations such as mean, maximum etc [18]. Finally, gis the identity function in this : Instead of classifying each instance individually, here fmaps each instance to a embedding, then obtains a bag representation independent of the number of instances in the these bag representations to obtain the overall prediction [19]. We initially tested the instance-based method but found the classifier was difficult to train to a sufficient accuracy causing poor performance. The embedding-based approach has less bias than the instance-based one and was found to perform better, hence was used for our experiments. For a bag of image patches we used a ResNet-18 pre-trained on ImageNet to encode each image patch as a feature vector turning the bag of image patches into a bag of feature vectors. 2.2.1 MIL methods We used two different MIL methods based on an attention mechanism [ 20]. First, we used Attention-based MIL (AMIL) [21]. Attention-based MIL modifies the embedding-based MIL approach by changing the MIL pooling operation, , where for a bag B=fh1;:::;h the overall prediction is obtained by z=KX k=1akhk; hT jg; parameters. For full details we refer the reader to Ilse et al. 's original paper [21]. We also explored a recent expansion of Ilse et al. 's Attention-based MIL. Lu et al. 's Instance Learning (CLAM) was developed with computational pathology in mind [ 22]. In CLAM, the predicts a set of attention scores for each class in the classification problem. Lu et al. use the same in the first two layers of the network as Ilse et al. but then split the network into nparallel attention branches in ann-class classification problem along with an additional instance-level clustering layer for each class to obtain the overall prediction. Again, we refer the reader to Lu et al. 's original paper for full details of the method [22]. 5Method Accuracy AUROC All patches AMIL 87:837:97 0:930:06 Blue ratio AMIL 77:016:77 0:940:06 RGB AMIL 90:728:82 AMIL 94:253:31 0:980:02 All patches CLAM 88:558:60 0:940:06 Blue ratio CLAM 80:545:88 0:930:04 RGB CLAM 92:068:25 CLAM 97.131.21 0.980.02 Table 1: Comparison of Attention MIL (AMIL) and CLAM models trained on extracted patches from the 3 different colour spaces at a threshold of and with all possible non-overlapping patches from the image. Results deviation , the variation in performance occurs due to differences in cross validation performance 3 Experimental Results To evaluate our proposed method we used the CRAG data set from Awan et al. [23]. This contains 139 extracted from 38 digitised WSIs of colorectal adenocarcinoma (CRA) at 20 magnification. The images varied slightly but on average were around 45007500 pixels in size. Of the 139 images, 71 were classified as normal tissue, 33 as low grade and 35 as high grade. We merged the low grade and high grade classes into a cancerous class to create a binary classification problem suitable for the MIL paradigm. In total, we had 71 images in the non-cancerous or negative class and 68 cancerous or positive images creating a fairly balanced classification problem. Awan et al. also provided the split they used for 3-fold cross validation so a fair comparison between our results and theirs could be drawn. Given the size of the original images, we set the maximum depth dof our quadtrees to be 4. This was done because if we allowed the trees to create nodes at a deeper level then the original image regions each node represents would be smaller in size than the 244244patch size that are fed into pre-trained networks. Upsampling would have been required to fix this and it would have been inconsistent with the process used on the nodes at every other level within the tree. We kept the trees at depth 4for all of our training, we normalised each patch using the mean and standard deviation of ImageNet, augmented each extracted patch using random horizontal and vertical flips with probability 0.5 as well as small, random adjustments to the brightness, contrast, saturation and hue of the input images. We used the Adam optimiser [ 24] with a loss rate of 5e4, betas of 0:99and0:999and weight decay of 1e4. All experiments were performed on a single Nvidia Quadro RTX 5000 GPU. 3.1 Quadtree evaluate the different proposed splitting methods for the quadtrees described in Section 2.1.1 we performed a qualitative analysis of all 139 images' quadtree decompositions under each method. For each method, we tested a variety of thresholds but kept them consistent in our comparisons. For example, we would only compare images that had been split where the threshold was set to be the same e.g. mean ( ) minus 1 standard deviation ( ) for each given colour space and method with images split at the same threshold, not with decompositions where the threshold had been set to1:25or any other level. We found that the mean pixel value in the Haematoxylin channel gave the best qualitative decomposition of the input images. Figure 2 shows an example of this. The method consistently split the image into finer grained detail in the presence of tissue while ignoring the background or less informative regions such as empty space or large areas of adipose tissue. Furthermore, it would aggressively split regions in the presence of cancerous tissue while with non cancerous tissue the method would sometimes only split down to a depth of 2 or 3 in more homogeneous areas such as large regions of connective tissue. This was not found to be the case in large regions of cancerous tissue where the method would consistently split the image down to the maximum allowed depth. Additionally, the threshold of  was empirically found to give the best qualitative splits by consistently following and splitting the tissue regions within the images. We trained the Attention MIL [ 21] and CLAM [ 22] models for 20 epochs using the patches extracted with each As a baseline control comparison, we also divided every image into non-overlapping patches of size 244244 denoted as the all patches baseline. The results in Table 1 show that our qualitative findings that the method produced the best decompositions of the images also corresponded to the best downstream model 6Figure 3: Accuracy of a CLAM model trained on data sets extracted at different thresholds ranging from 2to + 2in the Haematoxylin channel using the mean pixel value. Grey area indicates one standard deviation in the model's performance across 3 fold cross with both Attention MIL and CLAM when compared to the other methods with all other variables kept constant. Furthermore, two of the three colour spaces, RGB and Haematoxylin channel, outperformed our baseline of using every possible patch in an image as a bag of instances. The third colour space, blue ratio space, poorly as it almost always predicted the positive class giving a very high sensitivity but leading to overall worse performance than the other methods. Additionally, it was found that in each respective colour space and with the all patches baseline the CLAM model outperformed the Attention MIL model in terms of average performance with a smaller standard deviation for all cases except the all patches baseline. 3.2 Ablation study We performed an ablation study of the splitting thresholds used with the best performing approach, using the mean pixel value in the Haematoxylin channel. We varied the threshold from 2to+ 2in steps of 0:25and trained a CLAM model with 3 fold cross validation at each of the 17 thresholds, all other hyperparameters were kept constant. We found the model had the greatest performance using the best qualitative threshold of as it had the highest average performance and the smallest standard deviation across cross validation folds as shown in Figure 3. 3.3 Comparisons with other methods We performed an additional comparison using the standard practice for extracting patches in tissue regions in computa- tional pathology. To select only regions with tissue we thresholded an image's intensity to separate the tissue from Method Accuracy AUROC Average Training Time (s) % of pixel data used in training BAM 1 [23] 95:702:10 100 BAM 2 [23] 97:121:27 100 All patches 88:558:60 0:940:06 2110 100 Segmented patches 97:153:24 1.000.01 2097 69 Leaf nodes 95:652:11 0:990:02 1041 32 All nodes 97.131.21 0:980:02 1368 43 Table 2: Performance of CLAM model trained on 4 different sets of extracted patches compared to state-of-the-art from original CRAG paper on two class problem. Haematoxylin channel patches were extracted using the threshold . Reported metric values are meanstandard deviation on the validation set, the variation is due to the 3-fold Training times are averaged across the 3 cross validation folds for 20 epochs with all hyperparameters 4: a, b, c AUROC curve plots of cross validation average AUROC for CLAM models trained on all quadtree nodes, the leaf nodes from the quadtree and all patches respectively. The quadtree used was the best channel method extracted at a threshold of . Grey shaded area indicates one standard deviation across the three folds. dAverage AUROC curve for all three methods the background and created a tissue mask with Otsu's method [ 25]. We obtained a set of locations within the tissue area to extract patches from with the super-pixel algorithm such that the locations covered the entire tissue region [ 26]. To explore the significance of the non-leaf nodes with the quadtree structure in downstream tasks we also using just the leaf nodes patches extracted by the Haematoxylin channel method. Table 2 shows that the patches extracted by our Haematoxylin channel quadtree method yield as good performance as the segmented tissue patches with a smaller standard deviation between cross validation folds. In total the extracted 52,029 patches, 374 per image, while the Haematoxylin channel quadtree method extracted 32,099, 231 per image, a 38.31% reduction in data which still yielded as good if not better final model performance with all other training hyperparameters the same. 4 3 shows that our qualitative finding of setting the splitting threshold at yielded the best results also held true when compared to a wide range of possible thresholds in downstream performance. When the splitting threshold was set too high e.g. t>(+)we found that frequently the non-cancerous class images would not be split at all resulting in a quadtree with one root node only and a corresponding bag containing a single instance. This occurred because the non-cancerous images had more empty space on average than the cancerous images and contained less regions heavily stained with Haematoxylin. This in turn led to poor performance as the model was operating on a single instance which was heavily downsampled and did not have enough data to learn from and perform well. As this threshold was decreased the images were split more effectively by the algorithm resulting in improved performance. The performance peaked at the qualitatively best performing threshold of with the smallest standard 5: Attention heatmap visualisations For a non-cancerous (top) and a cancerous (bottom) image respectively the attention heatmap from the CLAM model trained using the Haematoxylin channel extracted patches at  overlaid on the quadtree decomposition of the original image (right). We smoothed the attention heatmaps for visibility by using 50% overlap between the patches and averaged the attention weights. The most highly attended regions are denoted in red while the less attended regions are denoted in blue. Regions have been re-scaled back to their original shape and size for display cross validation folds. The performance then decreased beyond this point with a much higher variation across cross validation folds. Inspecting the individual fold performance showed this was driven by one of the three folds performing much poorer than the others; however, this was not always the same fold in each case so this cannot have been occurring due to a difference in the data distribution between folds. We found that when t<the algorithm started to split regions that contained a low percentage of tissue where it was not before. This resulted in one patch containing tissue being generated but potentially two or three or all blank space. We believe this is why performance becomes more varied, more instances are being generated that are not conducive to making an accurate prediction. The threshold of appears to be a good middle ground between decomposing the tissue regions accurately and not over decomposing the images. While it is logical that such a middle ground should exist theoretically, we do not yet know why this is found at . We plan to explore this approach with other data sets in future to see if this finding holds true in different data found that our quadtree image representation is able to achieve better performance than the existing state of the art method for patch extraction in computational pathology. The CLAM model trained with the patches extracted from our quadtree method was able to achieve as good average cross validation performance with a smaller standard to a model trained using patches extracted using a tissue segmentation mask while using 38.31% less patches. When we excluded the non-leaf nodes' patches, the final model performance did degrade slightly but only by 1% compared to when all patches from the quadtree were included. We hypothesise this indicates that the higher level nodes in the tree (at lower magnifications) only contribute marginally to the final prediction. To verify this we overlaid the attention weights from the trained CLAM model in a heatmap over the original images. If the higher level non-leaf nodes were weighted highly we would expect to see large regions of high attention overlaid over regions that have been split further. However, the example in Figure 5 does not show this. It is representative of many other instances in our results where the leaf nodes are clearly the most highly weighted. The attention heatmap in Figure 5 also shows that the CLAM model trained using our quadtree strategy is able to assign higher attention values to clinically relevant regions. In the non-cancerous example, almost all of the highly weighted patches are located in a region of healthy glands and tissue, highlighted in red and orange, while the stromal region has 9been coloured a dark blue indicating very low attention has been assigned in this region by the model. This means the model has correctly identified that the presence of healthy glands indicates a non-cancerous region. 5 Conclusion We have shown that quadtrees can be used in computational pathology as an efficient image representation which can be used for fast and highly accurate downstream performance. Our quadtree method is able to decompose by identifying regions significant to clinical diagnosis while ignoring less significant regions such as empty space or connective tissue. We show that this image representation is able to be leveraged in a MIL setting to achieve better performance using 38% less data than the currently widely adopted thresholding based tissue mask approach used in the field while also providing an interpretable visualisation of which regions within the image are important for the algorithm in generating its prediction. In future, we plan to explore this approach with other computational pathology data sets to further verify our findings and explore how the framework performs for other tasks and tissue types as well as how sensitive it is to other staining methods and visual artifacts. If our results here hold true, the method should be very helpful in WSIs to reduce the significant data requirement currently present in processing these large A. Pryalukhin, W. Hulla, A. Bychkov, J. Fukuoka, A. Madabhushi, V . Achter, L. Nieroda, R. Buttner, A. Quaas, and Y . Tolkach, Quality control stress test for deep learning-based diagnostic model in digital pathology, Modern Pathology , pp. 1 - 11, 2021. [2]Y . LeCun, Y . Bengio, and G. Hinton, Deep learning, Nature , vol. 521, no. 7553, pp. 436-444, May 2015. [Online]. Available: Graham, D. Epstein, and N. Rajpoot, Dense steerable filter cnns for exploiting rotational symmetry in IEEE Transactions on Medical Imaging , vol. 39, pp. 4124-4136, 2020. [4]D. J. Ho, D. V . K. Yarlagadda, T. D'Alfonso, M. Hanna, A. Grabenstetter, P. Ntiamoah, E. Brogi, L. Tan, and T. J. Fuchs, Deep networks for multi-class breast cancer image segmentation, imaging and graphics : the official journal of the Computerized Medical Imaging Society , vol. 88, p. 101866, 2021. [5]G. Litjens, T. Kooi, B. E. Bejnordi, A. Setio, F. Ciompi, M. Ghafoorian, J. V . D. Laak, B. Ginneken, and C. Sanchez, A survey on deep learning in medical image analysis, Medical image analysis , vol. 42, pp. Tellez, G. Litjens, J. A. van der Laak, and F. Ciompi, Neural image compression for gigapixel analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 43, pp. 567-578, 2021. [7]K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 770-778, 2016. [8]J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, Imagenet: A large-scale hierarchical image IEEE Conference on Computer Vision and Pattern Recognition , 2009, pp. 248-255. [9]R. Finkel and J. Bentley, Quad trees a data structure for retrieval on composite keys, Acta Informatica , vol. 4, pp. 1-9, 1974. [10] S. Tanimoto and T. Pavlidis, A hierarchical data structure for picture processing, Computer Graphics and Image Processing , vol. 4, pp. 104-119, 1975. [11] P.-S. Wang, Y . Liu, Y .-X. Guo, C.-Y . Sun, and X. Tong, O-cnn, ACM Transactions on Graphics (TOG) , vol. 36, pp. 1 - 11, 2017. [12] P.-S. Wang, C.-Y . Sun, Y . Liu, and X. Tong, Adaptive o-cnn: A patch-based deep representation of 3d shapes, arXiv: Computer Vision and Pattern Recognition , 2018. [13] P. Jayaraman, J. Mei, J. Cai, and J. Zheng, Quadtree convolutional neural networks, in ECCV , 2018. [14] T. G. Dietterich, R. Lathrop, and T. Lozano-Perez, Solving the multiple instance problem with Artif. Intell. , vol. 89, pp. 31-71, 1997. [15] C. Shannon, A mathematical theory of communication, Bell Syst. Tech. J. , vol. 27, pp. 379-423, 1948. 10

2108.10873

Rob Jewsbury

Rob Jewsbury, Abhir Bhalerao, Nasir Rajpoot

A QuadTree Image Representation for Computational Pathology

11 pages, 5 figures, accepted to CDPath ICCV 2021

cs.CV cs.LG

http://creativecommons.org/licenses/by/4.0/

The field of computational pathology presents many challenges for computer vision algorithms due to the sheer size of pathology images. are large and need to be split up into image tiles or patches so neural networks (CNNs) can process them. In this work, we present a method to generate an interpretable image representation of images using quadtrees and a pipeline to use these highly accurate downstream classification. To the best of our knowledge, this is the first attempt to use quadtrees for pathology image data. We show it is highly accurate, able to achieve as good results as the currently widely adopted tissue mask patch extraction methods all while using over 38% less data.

2 JESSE THORNER AND ASIF Ifb1exists and q<=(logx)A, then Theorem 1.1 holds with th=7 12, but at a cost: c1will depend ineffectively on A. We justify this in Section 4 using Siegel's theorem [2, Ch. 21]: For all e >0, there exists a constant be>0 such that (1.2) 1 -b1>=beq-e. The constant beis effectively computable when e>=1 2[13, Thm 3]. Remark1.3.If there exists a constant c2>0, independent of q, such that b1<=1-c2 logq, then one may take l= 1 and th=7 12in Theorem 1.1. The implied constant and c1will on c2. We justify these assertions in Section 4. Ifb1does not exist, then l= 1 andh ph(q)is the main term in Theorem 1.1. If b1exists, then the notion of a main term and an error term in Theorem 1.1 d epends on how one views the role of b1. We suggest the perspective of the sum of a primary main term and a secondary main term, lea an error term. A similar perspective is implicit in Linnik's work on the least prime pa(modq) [9]. In light of our piecewise definition of l, this perspective accommodates all ranges of qandhunder our consideration, regardless of whether then there exists x(x-h,x) such that (1.5) l= if there exists A>=1 such that q<=(logx)A, then (1.2) with e=1 3Ashows that (1.6) l= in (1.6) is readily absorbed by (1.4). When qis larger, we allow the (1.3) to function as a secondary main term rather than attemp t to push the contribution into any sort of error term. This permits Theorem 1 .1 to effectively count primes in arithmetic progressions modulo qas long as the expected number of primes con- gruent to a(modq) in the interval ( x-h,x], namely lh/ph(q), exceeds xth+e. If, in addition, one has (log x)/(logq)- , then the expected number of primes becomes an effective as- ymptotic count, even if b1exists. The point is that (1.3) and (1.4) are each scaled by l, so the existence of b1proportionally affects both quantities. This differs from of the prime number theorem for arithmetic progre ssions (e.g., [2]) where the contribution from b1is viewed as an error term instead of a secondary main term. Even ifb1exists, we can always effectively estimate lusing the effective upper (1.2) as well as the l <2, which weprove inSection 4. Therefore, even case, forall0 < e <1-th, there exists an effectively computable constant c3=c3(e)>0 such that the Theorem 1.1 is satisfied when x>=c3and (1.8) h>=x1-d1, q<=xd2, d TO THE PRIME NUMBER THEOREM FOR ARITHMETIC PROG RESSIONS 3 Ifb1does not exist, then the3 2can be replaced by Theorem 1.1, we arrive at the following result using (1.8). Corollary 1.4. There exists a constant c4>0such that if q>=2andaare coprime as in Theorem 1.1, implied constant and c4are absolute and effectively then l(x,q,a,x) = with classical results. To display the uniformity of Theorem 1.1, we deduce the Siegel-Walfisz theorem with the best known error ter m, Hoheisel's asymp- totic prime number theorem for short intervals with the best known error term, a Brun- Titchmarsh type bound for short intervals, and Linnik's bound on th e least prime in an arithmetic progression as corollaries of Theorem 1.1, the ineffective bound (1.6), and the effective bound (1.7). While all of these results can be proved individu ally, the point here is that Theorem 1.1 provides a single asymptotic result that is uniform enough to deduce all of them. These deductions, which require no information on b1except for (1.2) and the effective computability of b1/2, showcase the flexibility of Theorem 1.1 in handling any combination of large, medium, or small moduli with short, medium, or lo ng andq<=(logx)A, then (1.6) and Theorem 1.1 with h=ximply that there exists an ineffective constant cA>0 such also Koukoulopoulos [8, Thm 1.1].) If e >0 and one adds the condition that then(1.6), Theorem1.1, thestrongest k nownversion of Hoheisel's short interval prime number theorem [4] for arithmetic progressions to There exists an ineffective constant cA,e>0 such now the situation where qis large. Applying Theorem 1.1 with e=3 110, we find that (logx)/(logq)- , then lsatisfies (1.7). Consequently, for all fixed d >0, there exists an constant c5=c5(d)>0 such that if x>=qc5, JESSE THORNER AND ASIF ZAMAN Whenb1does not exist, the (1 -b1)logxfactor is deleted, and 2+ dbecomes 1+ d. The upper bound is a weak form of the Brun-Titchmarsh theorem for sh ort intervals (cf. [12, Theorem 1.2]). Since b1<1, the lower bound implies that there exists a prime as Linnik proved In Section 2, we state a new version of the prime number theorem fo r proof of Theorem 2.3 uses a certain log-free zero density estim ate (Theorem 2.1) and zero-free region for Dirichlet L-functions. In Section 3, we use The- orem 2.3 and a zero-free region due to Iwaniec to refine Theorem 1.1 whenqis sufficiently powerful and provide an application to primes with pres cribed digits. We also contrast our results with those of Gallagher [3]. We prove Th eorem 2.3 in Section We thank Kevin Ford, Andrew Granville, Kannan Joni Ter av ainen for their encouragement and helpful conv ersations. We especially thank Roger Heath-Brown for his comments regarding the proof of Theo rem 2.1. 2.Proof of Theorem 1.1 Letq>=2,T>=1 and 0<=s<=1. For all Dirichlet characters kh(modq), we define Nkh(s,T) :=|{r=b+ig:b,gR, b>=s,|g| <=T, L(r,kh) = 0}|. Ifb1exists, then we :=|{r=b+ig:b,gR, b>=s,|g| <=T, = 0}|. Let Nq(s,T) then let N* q(s,T) with the standard theory of Dirichlet L-functions [2], we use a log-free zero density estimate that encodes the phenomenon that if b1exists, then b1repels all other zeros 2.1. Lete >0. LetT>=1,q>=2, and0<=s<=1. We have the and n(U) := then (2.2) 2.2.Bombieri's original proof in [1, Ch. 6] of the existence of an absolute a nd effectively computable constant c6>0 such that In an earlier version of this manuscript, we had pushed Bombieri's strategy to its limit, proving that one can take c6to be slightly smaller than75 4. The idea behind the simple proof below, which has the added benefit of producing strong er numerical exponents, was later communicated to us by TO THE PRIME NUMBER THEOREM FOR ARITHMETIC PROG RESSIONS 5 Proof.It is well-known that (2.1) follows from the work of Huxley [5] and Jutila [7]. To deduce (2.2), let 0 < e <1 and let nean effectively computable and sufficiently small positive constant depending only on e. Ifn(qT)> ne, then the conclusion follows from (2.1) by inflating the implied constant if necessary. Therefore, we assum e thatn(qT)<=ne, in which case our effective lower bound for b1implies that qTis sufficiently large (depending only one). The version of the phenomenon proved by Jut ila [7, Thm 2] implies that if s >5 6and (2.3) = 0. Hence, (2.2) holds under these assumptions. For the remainin g cases, our starting point is a trivial consequence from (2.1), we can rescale eaccordingly, it therefore suffices to show the proof of (2.2). If s<=9 10, then our effective lower bound for b1implies as required. Otherwise, if s >9 10and (2.3) does not hold, required. This proves (2.2). /square Using Theorem 2.1, we prove a highly uniform version of the prime numb er theorem for arithmetic progressions that depends only on a zero-free region f or Dirichlet 2.3. Letq>=2andabe coprime integers, and let 4<=h<=x. Letlandthbe as in Theorem 1.1. Let 0< e <1-th, and let d: [1,)-[0,1]be a function such that for all T>=e, the not vanish in the region {sC: <=T} except possibly for b1, if it exists. If lh/ph(q)>=xth+e, implied constant is effectively computable and depends a t most on (resp. does not exist) and the in Theorem 2.1 is improved to some constant c7>0, then Theorem 2.3 holds with th= 1-1 c7. Proof of Theorem 1.1. We use the zero-free region for Dirichlet L- functions [11, Ch. 9] in Theorem 2.3. In particular, there exists an a bsolute and constant c8>0 such that for t>=e, (2.4) >0, it follows JESSE THORNER AND ASIF ZAMAN fort>=e. We maximize these two exponentials over t[ex h,). The first exponential is maximized at the leftmost endpoint t=ex/h. For the second exponential, the same order of magnitude when log the second exponential is maximized on the interval [ex h,) for a value of t=t0such that either t0=ex/h, or This implies >0. Therefore, by the conditions on t0, there exists an effectively computable constant c1=c1(e)>0, depending at most on e, such this estimate in the conclusion of Theorem 2.3 proves Theo rem 1.1. /square The quality of the zero-free region does not impose constraints on the range of qor the length of theinterval ( x-h,x]. These ranges aredetermined only by the log-freezero density estimate in Theorem 2.1. The zero-free region only influences quality of regions other than (2.4) can be used with Theorem 2.3. I n some situations, can lead to improvements over Theorem 1.1. We conside r two such moduli. Letq>=2, and let d=d(q) the squarefree part of q. It follows from work of Iwaniec [6] that in Theorem 2.3, we may take (3.1) (3.1) into Theorem 2.3, we immediately obtain another highly uniform version of the prime number theorem for arithmetic 3.1. Letq>=2andabe coprime integers and let 3<=h < x. the squarefree part of q, letlandthbe as in Theorem 1.1. For all e >0, there exists a constant c9=c9(e)>0such that if implied constant and c9are effectively TO THE PRIME NUMBER THEOREM FOR ARITHMETIC PROG RESSIONS 7 Corollary 3.1 is a substantial improvement over Theorem 1.1 when qis sufficiently large relative to d, in which case qis powerful. If b1exists, then by [15, Lem 6.2], there exists an absolute and effectively computable constant c10>0 such that ifq thenb1does not exist, l= 1, and if there exists an effectively computable congmstant c11=c11(d)>0 such can be used to study primes with prescribed digits. Namely, one c an primes while simultaneously specifying 41.6% of the first and last 3.2. integers with A+B < N. LetPl(N)be the set of primes with Ndigits in their base lexpansions. For a prime p Pl(N), write its base l expansion Fixd0,d1,...,d N-1 0andgcd(d1,l) = 1. For all 0< e <5 12, there exists a constant that if A+B<=(5 12-e)N, then #{p Pl(N):dj(p) =djfor 0<=j<=A-1 implied constant and c12are effectively computable. The proof is nearly immediate from (3.2). The arithmetic progression condition significant digits, while the short interval condition prescribes theBmost signif- icant digits. Corollary 3.2 complements work of Swaenepoel [15] whe rein a small proportion of the digits can be prescribed without any restriction on how signific ant the digits with work of Gallagher. As far as the authors know, the result closest to Theorem 1.1 in the existing literature follows from the seminal work of Gallagher [3, Thm 7], which we now describe using our notation. Recall the definition of lin Theorem 1.1. With the effective bounds in (1.7), Gallagher's work implies that there e xist absolute and effectively computable constants c13,c14>0 such that if (3.3) proof of (3.4) uses a log-free bound on Nq(s,T) as in Theorem 2.1, the standard zero- free region when b1exists, the zero-free region (2.3) derived from the zero repulsion phenomenon of Deuring and Heilbronn [3, (27 )]. This contrasts with our use of N* q(s,T) whenb1exists and the zero-free region notexist and q>=exp(logx), Theorem 1.1achieves the asymptotic (3.4) in a stronger range than (3.3). When b1exists, an exact comparison of (3.4) with Theorem 1.1 depends on how large b1is as an explicit function of q. We observe here that we can estimate8 JESSE THORNER AND ASIF ZAMAN (2.5) using both the zero-free region and (2 .3). This observation leads to a mutual refinement of Theorem 1.1 and (3.4) in a range that impro ves upon (3.3). Corollary 3.3. Letq>=2andabe coprime integers and 4<=h<=x. Ifb1exists and lis as in Theorem 1.1, then there exist c16,c17(0,1)such that if lh/ph(q)>=xc16, implied constant, c16, andc17are absolute and effectively defined as in Theorem 1.1 and let 0 < e < 1-th. Without loss of generality, we may assume that that qis sufficiently large. The zero-free region (2.3) implies that there exists a n absolute and effec- tively computable constant c18>0 such that fore<=t<=x q. Consequently, if e<=t<=x q, maximum value of this exponential over t[ex h,) occurs at a value t=t1such or Therefore, there exists an effectively computable constant c19=c19(e)>0 such that if lh/ph(q)>=xth+e, estimate, in conjunction with (2.5), leads to the best bound tha t Theorem 2.3 and (2.3) allow. We slightly weaken this bound for easy comparison with (3.4 ). Let 0 < e < 1-th-e. Henceforth, assume that lh/ph(q)>=x1-e2, in which case (1.7) (1 -b1)logqq-1/2in (1.2), we obtain the these estimates, we deduce that there exists effectively computable constant c20= c20(e)>0 such that (3.5) is The last equality follows upon fixing e=c20 2. Taking the minimum of this estimate for (3.5) and bound (2.5), the result follows since min {u,v} <=u1/2v1/2foru,v >0. absorbed by the exponential when c17is chosen to be suitably small. /square 4.A flexible prime number theorem: Proof of Theorem 2.3 We provide the details in the most complicated case where b1exists using N* q(s,T) and th=32 37. All implied constants here are effectively computable. We begin by pr oving (1.7). Lemma 4.1. Ifb1exists and 4<=x1/2<=h<=x, l <2.REFINEMENTS TO THE PRIME NUMBER THEOREM FOR ARITHMETIC PROG RESSIONS then the aforementioned work of McCurley is primitive, then L(s,kh) has no real nontrivial zeros [14]. Therefore, if b1exists, then99 100< b1<1. Ifx(x-h,x) satisfies (1.5), then we have the explicit equality The boundsx 3< x < x follow from our Sincex>=4, we have x >1 and log x<=5logx, which we will use repeatedly. To bound lfrom above, we have l= 1-kh1(a)xb1-1<1 + (x/3)b1-1<2. For lower bounds, we use casework. If kh1(a) =-1, then min {1,(1-b1)logx} that n(U) = so 0< n(x)<=1. Ifkh1(a) = 1 and n(x) = 1, <=5n(x) = = 5l. Ifkh1(a) = 1 and 0 < n(x)<1, thenl= 1-e-n(x), and the from the Taylor expansion oft 1-e-tatt= 0. These cases exhaust all possibilities. /square Proof of Theorem 2.3. It suffices to consider h<=x-1. Fix 0< e <1-th, and assume that (4.1) explicit formula [2, Ch. 17-20] implies that if q,T,x>=2, a sum over characters kh(modq) a sum over non-trivial zeros The contribution from pk<=xwithk>=2 is trivially bounded by xusing Chebyshev's leading to the apply (4.2) twice with with p<=xand once with p<=x-h). Our assumption (4.1) then that h<=x-1. Using the JESSE THORNER AND ASIF ZAMAN as well as Taylor's theorem <=1 logx, we find 0< b <1 4and|g| consider the cases where h<=x 2andx 2< h<=x-1.) FortR, there zeros of L(s,kh) such that |t-g| <=1 [2,SS16]. Once we use this to the contribution from the it follows from (4.1) that (4.3) E dyadically decompose [ intoO(logx) subintervals, so that (4.3) leads to (4.4) E claim that if this claim, we sum over k, applying (4.4) and (4.5), to obtain E that Lemma 4.1 and (4.1). Also, since h<=x, it follows that logqelogx, Overall, this shows that establishes Theorem 2.3 subject to our claim, which we now prove By partial summation and Theorem 2.1, we find logqelogx, the definition of nimplies that bounds for lin (1.7) and the inequality q1-e2< qthen imply

2108.10878

Jesse Thorner

Jesse Thorner and Asif Zaman

Refinements to the prime number theorem for arithmetic progressions

11 pages. Theorems 1.1 and 2.1 improved

math.NT

http://creativecommons.org/licenses/by/4.0/

We prove a version of the prime number theorem for arithmetic is uniform enough to deduce the Siegel-Walfisz theorem, for intervals of length $x^{1-delta}$, a Brun-Titchmarsh bound, and Linnik's bound on the least prime in an arithmetic progression as corollaries. Our proof uses the zero-free region, a log-free zero and the zero repulsion phenomenon. when the modulus is sufficiently powerful.

2 contrast, Ossowski (2021) disfavored the globular clus- ter channel in favor of isolated binary evolution. Fran- ciolini et al. (2021) investigated four formation binaries, globular clusters, nuclear star clusters, and primordial BHs) and found a high likelihood that pri- mordial BH-BH mergers are part of LIGO/Virgo of these studies perform proper model comparison to infer which model or mixture of models is model comparison requires precisely specied models, i.e. models with precisely dened physics. Here we emphasize that the physics uncertainties (Schoote- meijer et al. 2019) in at least one of those models, of isolated binary evolution, are suciently large that (to put it in Bayesian terms) the prior dominates the con- clusion. That is, dierent assumptions can lead to very dierent outcomes, which means that population stud- ies are not yet at the stage allowing strong and to be drawn. We demonstrate these model weaknesses using var- ious, dierent assumptions in performing aim is to determine the fate of the most mas- sive known binary system. Despite the high binary fre- quency of massive stars (Sana et al. 2012), eclipsing sys- tems with primary masses 50 M are Within the Milky Way the most extreme double- lined systems are located in young, rich star clusters: A1 within NGC 3603 (Schnurr et al. 2008), F2 within the Arches (Lohr et al. 2018) and WR20a in Westerlund 2 (Bonanos et al. 2004). All three are short period (days), low eccentricity systems with main sequence WNh) primaries. The most massive double-lined eclipsing system in the Large Magellanic Cloud (LMC) is also a short period sys- tem, Hunter 38 in the Tarantula Nebula with a O-type primary, whose mass is 57 M (Massey et al. 2002). The most extreme LMC binaries are non eclipsing sys- tems, also within the Tarantula Nebula. They have min- imum dynamical primary masses in excess of 50 M , albeit with orbits whose periods are an order of mag- nitude longer and high eccentricities. From comparison with Bonn stellar evolutionary models at LMC metallic- ity (Brott et al. 2011), R139 ( Porb= 154 d,e= 0:38) has a primary O supergiant mass of 80 M (Mahy et al. 2020), R144 ( Porb= 74 d,e= 0:51) has a primary WNh mass of110 M (Shenar et al. 2021) and Mel- nick 34 (Mk 34, Porb= 155 d,e= 0:68) { the - has component WNh+WNh masses 12717 M (Tehrani et al. 2019) with initial masses of 144 M and 131 M . Pollock et al. (2018) rst established that Mk 34 is a colliding wind binary from analysis of X-ray time series Tehrani et al. (2019) noted that the potential fate of Mk 34 involves a double stellar mass black hole predict the fate of Mk 34, using dierent applying two rapid population synthesis codes ( StarTrack and COMPAS ) and the detailed code ( MESA ). We nd a wide variety of pos- sible outcomes (see Tab. 1) in terms of the black hole masses and orbital separations, and even in terms of whether black holes will form at all. We therefore urge caution in drawing important and credible conclusionsTABLE 1 Fate of Mk 34 model MBH1MBH2tdelay fate [ M] [ M ] [Myr] StarTrack1 22.5 22.1 47.5 close 35.7 33.3 10,035 close 36.0 32.5 11,663 close BH-BH?a COMPAS1 19.8 20.2 >thub wide BH-BH COMPAS2 31.8 31.7 >thub wide BH-BH COMPAS3 31.8 31.7 >thub wide BH-BH MESA1 21.9 51.6b| Thorne- _Zytkow MESA2 35.2 80.9b| Thorne- _Zytkow MESA3 35.3 85.4b| Thorne- 21.9 22.1 >thub wide 35.2 33.3 >thub wide | | 3.3 stellar | | | 6060 M>thub wide 3030 M>thub wide 2020 M>thub wide BH-BH a: optimistic (non-standard) StarTrack models are used to get this result b: for MESA models we list CE donor mass in column MBH2c: merger of post-MS star and MS star: formation of very massive single star, fate: PSN or a single BH d: pair-instability supernovae disrupting binary components about the LVC BH-BH population based on models evolution containing, by necessity, un- certain initial properties of Mk 34 we select: Ma= 144 M ,Mb= 131 M ,e= 0:68,a= 760 R cho- sen to result in an orbital period of Porb= 155 d after 0:6 Myr (current age, Tehrani et al. 2019) of system evo- lution with the StarTrack code. We adopt the LMC metallicity of Z= 0:006 (Rolleston et al. 2002). 2.1. StarTrack calculations We use the population synthesis code StarTrack (Bel- czynski et al. 2020a), which employs analytic ts to evo- lutionary tracks of non-rotating stellar models (Hurley et al. 2000). We adopt standard wind losses for mas- sive stars from Vink et al. (2001) and LBV winds as (dM=dt 1:5 from Belczynski et al. (2010a). For stars that overll their Roche lobes, we initiate mass transfer between binary components and associated (if any) mass loss from binary systems. If the binary is not circularized by tidal interactions we circularize it (enew= 0) to periastron distance ( anew=a(1e)) in one timestep and only then start Roche lobe over ow (RLOF). For mass transfer (NTMT) and mass transfer (TTMT) we use the standard formalism, while we use a diagnostic mass ratio diagram as a criterion for common envelope (CE) devel- opment (see Sec.5 of Belczynski et al. (2008)). During the TTMT/NTMT the fraction of mass lost by the donor star that is accumulated by non-degenerate companion stars is set to fa= 0:5, while the rest is lost with specic3 angular momentum (expressed in units of 2 a2=Porb) of jloss= 1:0 (see eq. 33 of Belczynski et al. (2008)). The accumulation of mass on compact objects (e.g., NS/BH) is limited by the (Eddington) critical accretion rate and mass is lost with the specic angular momentum of the compact accretor (King et al. 2001; Mondal et al. 2020). We employ the delayed core-collapse supernova (SN) engine in NS/BH mass calculation (Fryer et al. 2012) which allows for populating the lower mass gap between NSs and BHs (Belczynski et al. 2012; Zevin et al. 2020). We employ weak pair-instability pulsation su- pernova (PPSN) mass-loss and pair-instability supernova (PSN) model that results in upper mass gap: no BHs with massMBH&55 M (Belczynski et al. 2020a). We allow for the fallback decreased NS/BH natal kicks with = 265 km s1and no natal kicks for direct BH for- mation. This is our standard input physics marked as StarTrack1 model in Table 1. The development of the CE phase is a big issue astrophysics (Ivanova et al. 2013b; Olejak et al. 2021). We are agnostic about which systems should be sent to a CE and which should evolve through stable RLOF. In StarTrack models we allow for the most op- timistic scenario (see Sec. 3.1), and we send nearly all systems through the CE to form (potentially) BH-BH mergers. We do not do this on regular basis. According to our standard input physics, donors with radiative en- velopes (e.g., in the Hertzsprung gap) do not enter the CE phase. Since it is not fully understood how exactly a CE develops, we test various assumptions to show how these in uence the future fate of binary systems such as Mk 34. Contrasting models are being presented as well. During CE events the entire envelope of the donor is assumed to be lost from the binary, with the exception of compact object companions that are allowed to ac- crete a small fraction of donor's envelope at 5% of the Bondi rate (MacLeod et al. 2017a). The CE orbital de- cay is calculated with the standard energy-balance for- malism (Webbink 1984) in which we adopt a 100% e- ciency of the orbital energy transfer ( = 1:0) into the envelope, while the binding energy is parameterized by detailed stellar models ( scaling: Xu & Li (2010); Do- minik et al. (2012)). In StarTrack2 we decrease wind mass{loss rates for LBV stars to flbv= 0:48 and we increase the He core mass at the end of the main sequence by a factor of fcore= 1:5 with respect to the original Hurley et al. (2000) models. This model approximately reproduces the basic properties of the 131 M and 144 M models at terminal-age main sequence (TAMS) obtained in our MESA computations (see Sec. 2.3). In StarTrack3 we circularize massive binaries with angular momentum conservation ( anew=a(1e2)), we setjloss= 0:1, andfa= 0:25 while keeping the rest of the input physics as in StarTrack2. This model aims to test the survival of the CE phase in an Mk 34-like future evolution (see Sec. 3). In practice, such set-up allows the secondary star of Mk 34 to have a large radius (wide binary orbit) during the RLOF and therefore potentially to develop a convective envelope but to survive the CE phase. 2.2. COMPAS calculationsWe use the population synthesis code COMPAS (Steven- son et al. 2017; Vigna-G omez et al. 2018; Chattopadhyay et al. 2021), which incorporates stellar (Hurley et al. 2000) and binary evolution (Hurley et al. 2002) with updated wind prescriptions for massive stars (Vink et al. 2001; Belczynski et al. 2010a). The LBV wind losses are as in StarTrack (see Sec. 2.1). In its default set-up, the mass accretion during RLOF onto degenerate stars (i.e. black holes, neutron stars and white dwarfs) is Eddington limited in COMPAS . For non- degenerate stars, this accretion limit is set by the ratio of the rate at which mass is being donated to the rate of which mass can be incorporated into the companion star. These two rates, in turn, are determined by the timescales of the donor and the com- panion respectively (Kalogera & Webbink 1996). The thermal timescale is an explicit function of the mass, ra- dius and luminosity of the star. For a star of total mass M, envelope mass Menv, radiusRand luminosity L, the thermal timescale KHis given by KH=GMM env=RL, whereGis the universal gravitational constant. In bi- naries, since these stellar parameters are determined by the orbital period at which the donor over ows its Roche lobe, the thermal timescale of the donor becomes an im- plicit function of the orbital properties of the et al. 2015). Thus the accretion eciency in binaries is primarily determined by their orbital pe- riod. The mass transfer eciency acc(ratio of the mass gained by the donor to the mass lost from the compan- ion)1in a non-degenerate star can be thus expressed as acc= min(1;10KH ;acc KH ;don), thermal timescales of the accretor and donor respec- tively (Hurley et al. 2002; Schneider et al. 2015; Chat- topadhyay et al. 2021). The stability of the mass loss is determined by the mass-radius exponent for development of CE) in COMPAS using ts from the Ge et al. (2015) sim- ulations as described in Vigna-G omez et al. (2018). In nearly equal-mass, close binaries the thermal timescales of the donor and accretor being very similar, the mass transfer is usually conservative and remains stable. In close systems with more extreme mass ratio, the of the donor being much longer than the accre- tor, the mass transfer becomes non-conservative leading to a CE phase. The Ge et al. (2015) criteria renders the mass transfer from evolved (non-main sequence, non- degenerate) massive stars as predominantly stable (Nei- jssel et al. 2019), which is very similar to the Pavlovskii et al. (2017) model (discussed in section. 2.4). We assume an isotropic re-emission model for angular momentum loss during non-conservative stable RLOF (Pols et in StarTrack , the binary is not before RLOF (Vigna-G omez et al. 2018). But bi- naries that survive the CE events are always circular- ized (Vigna-G omez et al. 2018). All other default RLOF and CE mass transfer specications in COMPAS are iden- tical to StarTrack as described in Sec. 2.1. Our standard model COMPAS1 utilizes the Fryer et al. (2012) delayed' supernovae prescription core mass to post-supernova of the parameter fainStarTrack described in Sec. 2.1.4 mass mapping. The (pulsational) pair-instability super- novae modeling is implemented in COMPAS (Steven- son et al. 2019) with polynomial tting from the models by Marchant et al. (2018) as the default input. The natal kick distributions (including fallback) for BHs and NSs are identical to StarTack1 model. In the model COMPAS2 we reduce the LBV wind mass loss rate to flbv= 0:48 (fromflbv= 1:5 in COMPAS1). We also increase the He core mass of the terminal main sequence stars by a factor of fcore= 1:5 multiplied to the tting formula from Hurley et al. (2000) (eq. 30). COMPAS2, like StarTrack2 is adjusted to the total mass and the core mass at TAMS of 131 M and 144 M models computed with MESA . The model COMPAS3 is identical to COMPAS2 but here we allow the binary to circularize (while conserving its angular momentum) right before the onset of RLOF as detailed for StarTrack in Sec. 2.1. 2.3. MESA Calibration of TAMS core masses The tting formula by Hurley et al. (2000) to the evo- lutionary tracks from Pols et al. (1998), which are the basis for StarTrack and COMPAS population synthesis codes, are based on stellar models computed for stars with masses up to 40 M . The treatment of more mas- sive stars in StarTrack and COMPAS relies on extrap- olation. In the mass range considered in this study (130145 M ), this can lead to a signicant devia- tion in basic stellar properties from what detailed stel- lar models produce (or what is inferred from observa- tions). One property that is particularly inaccurate due to the extrapolation of Hurley et al. (2000) formulae (and which is crucial for considerations of the nal fate of the Mk 34 system) is the ratio of the helium core mass to the total star mass at TAMS ( Mcore;TAMS=MTAMS ). For very massive stars this ratio is close to unity (Yusof et al. 2013; K ohler et al. 2015). In contrast, a 144 M star atZ= 0:006 metallicity evolved with , even though signicantly stripped through MS winds ( MTAMS = 77:4 M) is far away from be- ing a helium star with the helium core mass of only Mcore;TAMS = 34:1 M. To correct for this and calibrate the properties of the StarTrack and COMPAS models at TAMS, we use the MESA 1D code (Paxton et al. 2011, 2013, 2015, 2018, 2019)2. We compute single models of 131 M and 144 M stars atZ= 0:006 metallicity. The rela- tive initial abundances of metals follow Grevesse et al. (1996). We model convection by using (B ohm-Vitense 1958) with a mixing-length pa- rameter= 2:0, and we adopt the Schwarzschild crite- rion for convection. We used the Dutch wind setup in MESA , which combines dierent prescriptions depending on the eective temperature Teand the fractional sur- face hydrogen abundance Hsur. As shown in Table 2, for Te<104K the code uses the mass-loss rates from de Jager et al. (1988), regardless of the hydrogen For Te104K,MESA adopts either the Nugis & Lamers (2000) prescriptions (if Hsur<0:4), or mass-loss rates from Vink et al. (2001) (if Hsur0:4). 2MESA version r15140, 2 Dutch Stellar Winds in MESAa Te<104K Te104K { de Jager et al. (1988) { Hsur<0:4 { Nugis & Lamers (2000) Hsur0:4 { Vink et al. (2001) a: based onTewinds either depend or do not depend on mass loss rates in MESA can be increased or decreased by changing a specic scaling factor fwind. The standard prescription of Vink et al. (2001) is known to underestimate the empirical mass-loss rates of very mas- sive MS stars, which increase dramatically as they ap- proach the Eddington limit, e(Vink et al. 2011; Besten- lehner et al. 2014; Bestenlehner 2020). Indeed, mass-loss rates of the components of Mk 34 from Tehrani et al. (2019) exceed Vink et al. (2001) pre- scriptions by factors of 2{3. We account for convective overshooting above core by applying the step overshoot- ing formalism, which extends the convective core by a fractionovof the local pressure scale height. We initialize our models with the initial rotation of Vi= 250 km s1(guided by the analysis of Tehrani et al. 2019). For rotational mixing, we include the eects of circulation, secular shear the instability, with an ef- ciency factor fc= 1=30 (Heger et al. 2000; Brott et al. 2011). We avoid using the MLT++ option in MESA (Paxton et al. 2013). As a result, models that reach the red super- giant stage encounter numerical diculties in their su- peradiabatic outer envelope layers (Pavlovskii & Ivanova 2015; Klencki et al. 2020), which prohibits us from fol- lowing their evolution to the point of maximum radial ex- pansion. However, for the purpose for the current study, we are only interested in the properties of MESA models at TAMS as well as whether or not the models expand suf- ciently to lead to a RLOF in the Mk 34 binary system. We thus stop the MESA computations when the radius of 2000 R is reached. Otherwise, we stop the simulation after 10,000 MESA steps. Such stopping conditions are sucient for our purposes in all the considered scenarios for the Mk 34 system. For calibration of population synthesis models at TAMS we calculate a MESA stellar model with = 144 M , metallicity Z= 0:006, and Vi= 250 km s1. We apply the standard Dutch winds (fwind = 1:0, though see above) and step 0:12, to maintain consistency with the over- shooting in models by Pols et al. (1998) and the Hurley et al. (2000) ts. At the end of the MS, this model has a mass ofMTAMS = 94:0 Mwith a He core mass of Mcore;TAMS = 66:6 M(see Table 3). Post-MS expan- sion leads to a maximum radius of Rmax= 1968 R at the end of our simulation, at which point the star is still expanding as a red supergiant. A MESA stellar model with MZAMS = 131 M ,Z= 0:006, andVi= 250 km s1 results inMMS= 86:6 MandMcore= 58:5 Mand expands beyond 2000 R (see Table 3). TheMZAMS = 144 M (Z= 0:006) orCOMPAS model produces MTAMS = 77:4 MandMcore;TAMS = 34:1 M. We decrease5 the winds during MS, keeping the original Vink et al. (2001) prescriptions, but decreasing the LBV winds to flbv= 0:48 to get a model with MTAMS = 94:2 M andMcore;TAMS = 44:5 M. Next, we increase the core size byfcore= 1:5 to get the target values: MTAMS = 94:2 MandMcore;TAMS = 66:7 Min the popula- tion synthesis codes. Note that such a massive star is already luminous ( L3106L) and cold K) to be beyond the (Humphreys & Davidson 1994) and subject to LBV winds on the MS. Applying the same calibration to Mzams = 131 M (Z= 0:006) we obtain in codes: MMS= 89:5 MandMcore= 62:3 M. We apply this calibration for all metallicities. However, one should note that some observations may be in con- tradiction of LBV winds (Gilkis et al. 2021). TABLE 3 Properties of MESA models used to calibrate the population synthesis simulations (the two top rows) as well as a few models with increased wind mass-loss or core overshooting (see Sec. 2.3.2). We show the maximum radius reached in our simulation as well as the total mass and the He core mass at the end of the MS phase. model R] [ M ] [ M ] 144 M (ov= 0:12,fwind = 1)>1968 94 67 131 M (ov= 0:12,fwind = 1)>2000 87 58 144 M (ov= 0:33,fwind = 1) 46 62 61 131 M (ov= 0:33,fwind = 1) 41 59 58 144 M (ov= 0:12,fwind = 1:5) 40 57 55 131 M (ov= 0:12,fwind = 1:5) 40 63 53 2.3.2. Calibration of post-MS expansion When considering the future fate of the Mk 34 system, a key question is whether its very massive components will expand after the end of the MS and initiate a mass- transfer interaction or whether they will lose their hy- drogen envelopes already during the MS and to become compact helium stars, avoiding any RLOF. The two crucial aspects that aect the degree of the post-MS expansion of very massive stars is the amount of core overshooting and the strength of stellar winds. Here, we explore this by computing a small grid ofMESA models with dierent overshooting and wind the calibration of TAMS properties of 131 M and 144 M models, we assumed a modest core 0:12, following the calibration to low-mass stars in open clusters by Pols et al. (1998). More recently, Choi et al. (2016) found the best agreement with the properties of the Sun for a MESA model with ov= 0:16. However, there is an increasing amount of evidence that core overshooting could be signicantly larger in the case of massive stars (e.g., Brott et al. 2011; Castro et al. 2014; Claret & Torres 2018; Scott et al. 2021). In particular, the calibration by Brott et al. (2011) to match the ob- served drop in rotational velocities of post-MS B stars (although see Vink et al. 2010) resulted in ov= 0:33, a value that has become widely used to compute stellar models of massive stars in the recent years. On the otherhand, there is no observational calibration of core over- shooting in the case of very massive stars of masses above 100 M . As such, we explore six dierent ovvalues from a wide range between 0 :12 and 0:5. The Dutch wind scheme in MESA incorporates the Vink et al. (2001) prescription for optically thin of hot MS stars. However, as noted above, very massive stars possess a suciently high luminosity to mass ratio on their MS that they approach their Edding- ton limit, leading to high mass-loss rates (Gr afener & Hamann 2008; Vink et al. 2011; Bestenlehner et al. 2020). Here, we attempt to correct for this by simply increasing the wind scaling factor fwind from 1:0 to 1:5 or to 2:0. We compute a grid of 131 M and 144 M models with the above variations in overshooting and winds (and all the other assumptions same as in our calibration models in Sec. 2.3.1). All the results are shown in Appendix A and a few selected examples in Table 3. In short, we nd that any model with overshooting of ov= 0:33 or higher, or a wind multiplication factor fwind1:5 evolves to become a helium WR star already by the end of MS, avoiding radial expansion beyond 100 R and any RLOF interaction in the Mk 34 binary. This result is at the basis of the quasi-single evolutionary scenario for Mk 34, see Sec. 2.5. 2.3.3. Calibration of envelope binding energies Recent studies by Klencki et al. (2021) and Marchant et al. (2021) have shown that the envelope binding en- ergies used in StarTrack andCOMPAS (i.e.scaling fol- lowing Xu & Li (2010); Dominik et al. (2012)) may be severely underestimated in the case of massive stars with outer radiative envelopes. Note that population synthe- sis codes do not perform CE evolution for Hertzsprung gap stars (radiative outer envelope) under standard as- sumptions on input physics, but typically the CE is ap- plied for core-helium burning stars even for those with outer radiative envelopes. To explore the eect of revised binding energies on the future fate of Mk 34, whenever our StarTrack bi- nary evolution calculation predicts a CE phase to oc- cur, we use MESA to compute a detailed stellar model of the donor star. We then follow the method outlined in Klencki et al. (2021) to integrate through the envelope of theMESA model and compute its binding energy. When matching the properties of a MESA model with those from StarTrack , we ensure that the CE donor has the same helium core mass and the same radius but allow for a lower envelope mass in the MESA model, so that we may be under- but never over-estimating the envelope bind- ing energy. This allows for conservative statements on inability of a binary to eject the donor's envelope and the CE survival. 2.4. Pavlovskii presented below are obtained with . In particular, we use more restrictive criteria for the CE development (Pavlovskii et al. 2017), and al- low more binaries to evolve through stable mass transfer instead (Olejak et al. 2021). The new criteria are post MS donor stars, (ii)for initial masses larger than 18 M ,(iii) when the mass ratio (compan- ion to donor mass at CE onset) fulls the condition6 qCE< qcritfor CE to develop, where qcrit= on donor mass and metallicity, and (iv)when the donor's radius at the onset of CE fulls specic cri- teria (shown at Fig. 2 and 3 of Olejak et al. (2021)) for CE to develop. These new criteria lead to the emergence of BH-BH formation channels without CE in (Olejak et al. 2021). This is the same chan- nel that was proposed by van den Heuvel et al. (2017) and that is also found in COMPAS simulations (Stevenson et al. 2019; Neijssel et al. 2019). Models labeled Pavlovskii1 and Pavlovskii2 cor- respond to models StarTrack1 and StarTrack2 but with modied CE development criteria, respectively. In model Pavlovskii3 we test dierent formulae for the loss of the angular momentum during mass trans- fer through the L2 point given by MacLeod & of our stan- dardjloss= 1:0 (Podsiadlowski et al. 1992). We expect much higher loss of angular momentum with this modi- cation, which may result in another potential fate for the future evolution of Mk 34 (see Sec. 3.4). The rest of input physics of Pavlovskii3 corresponds to Pavlovskii2. A change of the circularization scheme would only increase binary orbital separation. 2.5. Quasi single star calculations In this part we approximate the evolution of non- expanding components of Mk 34. This is driven by MESA models with large overshooting or increased stel- lar winds that do not show signicant post-MS ex- pansion (see Tab. A). In particular, MESA models with ov= 0:33 and standard winds ( fwind = 1:0) reach a maximum radius of 46 R forMZAMS = 144 M and 41 RforMZAMS = 131 M . Additionally, important for the development of the PPSN/PSN, stellar models that do not expand may have very dierent TAMS he- lium core masses. For example, the MZAMS = 144 M model produces Mcore;TAMS = 61 M forov= 0:33 and fwind = 1:0 orMcore;TAMS = 30 M forov= 0:4 and fwind = 1:5. The former model is possibly subject to a PPSN/PSN while the latter is not (Woosley 2017; Farmer et al. 2020). In Sec. 3.5 we explain our choice of models, show- ing how uncertainties can aect the future evolution of Mk 34. 3.EXAMPLES OF MK 34 FUTURE predicted models of Mk 34 future evolution are illustrated in Figure 1, summarized in Table 1, and described below. 3.1. StarTrack models A binary star resembling Mk 34 is evolved with the StarTrack2 model (Sec. 2.1). Star A (initially more mas- sive) expands as it evolves and nally periodically over- lls its Roche lobe ( Ra= 139 R ) at periastron pas- sages. At this point the orbit expanded from the initial a= 760 R to 1125 R due to wind mass loss from both binary components ( Ma= 94:2 M,Mb= 91:9 M) while the eccentricity remained much unchanged ( e= 0:68). The tidal circularization force is the strongest at periastron and we assume that the orbital motionis circularized to periastron distance and leads to nor- mal stable RLOF on new ( a= 360 R ) circular or- bit. RLOF leads rst to TTMT which to a NTMT. Star A is stripped almost en- tirely of its H-rich envelope ( Ma= 67:4 M,Ma;core= 67:3 M) while star B accreted half of that lost enve- lope (Mb= 105 M ) while the rest of the mass has been lost from the binary. In response the orbit increased in size (a= 439 R ). Star A is a Wolf-Rayet with heavy wind mass{loss and at the end of its nuclear evolution its mass decreases to Ma= 35:6 M(a= 1370 R , Mb= 99:1 M). Star A collapses directly to a BH with a mass ofMa= 35:2 M(1% neutrino mass loss, no baryonic mass loss, no natal kick). Then star B evolves and expands to ll its Roche lobe in a circular orbit witha= 1515 R (Mb= 88:5 M,Rb= 700 R , Tb;e= 8690 K). This time in our standard approach, due to relatively high mass ratio ( q= 88:5=35:2 = 2:5) RLOF is evaluated to lead to a CE phase. We estimate the binding energy of star's A envelope ( = 0:103) to be low enough to be ejected at the cost of orbital en- ergy. After envelope ejection star B becomes a massive stripped He core ( Mb= 62:0 M) and the orbit decays toa= 34:2 R. During CE the rst-formed BH accretes 0:5 M(Ma= 35:7 M). Star B, after a Wolf-Rayet wind mass loss ( Mb= 33:7 M,a= 47:8 R) to a BH ( Mb= 33:3 M). Aftertevol= 4:1Myr of binary evolution a close BH-BH binary is formed with a coalescence time of tcoal= 10:0Gyr. There are caveats in this scenario. Star B at the time of the RLOF onset has just nished core H-fusion and is a Hertzsprung-gap star in the transition to become a core-He burning giant. It was argued that such stars do not have a clear core-envelope structure and that the CE phase should always lead to merging of the donor star with its companion (Belczynski et al. 2007). This nds some support in observations as the predicted BH-BH merger rates that allow for such a scenario as presented above are too high to match the empirical (see submodels A in Table 4 of Belczynski et al. (2020a)). In addition, star B has an outer radiative en- velope: at the time of RLOF and with surface properties Rb= 700 R ,Tb;e= 8690K!L= 2:5106L, it is well above the eective temperature threshold below which stars at LMC metallicity have convective envelopes (T <3900K; see Fig.6 of Klencki et al. (2020)). Klencki et al. (2021) argued that massive gi- ants have binding energies that are too high to allow for a successful CE ejection in BH-BH merger progenitor bi- naries. In StarTrack models, the estimate of the binding energy of star B with the formalism is only an approx- imation that is needed for the use in rapid models. It should be also stressed that in the standard input physics of StarTrack models we would allow no Hertzsprung-gap star to survive a CE phase no matter what is our estimate of the star binding energy () or assumed eciency of orbital energy transfer to the envelope (). This is why we call the StarTrack models used here as optimistic scenarios. We test this with MESA in Sec. 3.3. In StarTrack1 model a similar scenario develops. How- ever, since the stars and their cores are less massive the BH masses are10 M smaller than in model Star- Track2 (see Tab. 1).7 Fig. 1.| Future evolution of Melnick 34. The fate of this massive binary is subject to a number of stellar and binary Depending on the adopted evolutionary model, Mk 34 may form a close or wide BH-BH system, a Thorne- _Zytkow (TZ { Thorne & Zytkow 1977) object or end its life in two pair-instability supernovae (PSN). RLOF: Roche lobe over ow, CE: common envelope, TTMT: mass transfer, ZAMS: Zero Age Main Sequence, BH: black hole, GRB: gamma-ray burst, PPSN: track gives somewhat lower mass BH (formed out of an initially more massive star) than other binary scenarios. In StarTrack3 model we adjust evolutionary parame- ters in such a way that CE survival is less caveated than in model StarTrack2. We alter circularization process and we change RLOF parameters setting mass trans- fer/loss to obtain wider binary than in StarTrack2 model. This allows star B to expand more before it initiates CE phase:Rb= 1439 R ,a= 3141 R ,Mb= 85:9 M, Mb;core= 60:1 M,Tb;e= 5986 K, = 0:050. This star has almost convective envelope, but not quite so. If we perform CE energy balance with the above param- eters this system survives CE and forms close BH-BH binary attevol= 3:9 Myr and with tcoal= 11:7 Gyr. Yet, the same caveats remain as for StarTrack2 model. 3.2. COMPAS models With the initial masses, orbital period and eccentric- ity of Mk 34 (see Sec. 2) we evolve COMPAS1, COM- PAS2 and COMPAS3 models with the individual vari- ations specied in Sec. 2.2. As in Sec. 3.1, we will always refer to the originally more massive star (with MZAMS = 144 M ) as star A. In COMPAS2 model, star A ends core-H burning witha total mass of 94 :2 M, while the stellar winds increase the separation to a= 1165:3 R. As star A leaves the MS, its core mass is calculated to be Ma;core= 66:8 M. Star A overlls its Roche lobe soon during post-MS evo- lution losing its envelope (27 :4 M) in a stable TTMT RLOF phase. This is the outcome of the binary being fairly wide and the donor being an evolved (post-MS) star, as commented on in Sec. 2.2 (see also Schneider et al. (2015)). Eccentricity remains un- changed (e= 0:68) during this phase and the orbit ex- pands toa= 1374:9 R. Star B, which had a total mass of 91:8 Mright before the mass transfer, becomes a 119:2 MMS star. Being stripped o its H-rich en- velope, star A enters the naked helium-star (WR) phase with mass Ma= 66:8 M. Evolution continues while both stars are losing mass in winds. Star A with mass Ma= 35:4 Mundergoes direct core-collapse, forming a BH of mass Ma= 31:8 M, while star B with Mb= 115:3 Mis still on the MS. The orbital separation, right after the formation of BH A, becomes a= 1712:1 R. Star B leaves the MS with mass Mb= 93:7 Mand core mass Mb;core= 66:3 Mwhen the orbital sepa-8 ration isa= 2006:6 R. Shortly thereafter star B lls its Roche lobe, loses most of its H-rich envelope (27:3 M) and becomes a naked helium star in sta- ble TTMT RLOF. The post-RLOF orbital separation is decreased to a= 918:1 R. Eddington limited accretion allows BH A to gain only 3:7105M. The binary at the onset of this RLOF, though has slightly than the previous RLOF phase, is still fairly wide and results in a stable TTMT RLOF despite rather high mass ratio ( q= 93:7=31:8 = 2:9). It is noted that both RLOF phases noted in COMPAS2 model are fairly similar to Pavlovskii2 model (see Sec. 3.4). However, the binary remains eccentric through both RLOF phases (e= 0:673). Star B, with a mass Mb= 35:2 M, un- dergoes direct core-collapse and forms a BH of 31 :7 M. COMPAS2 model creates a BH-BH system, which at the second BH formation has a separation a= 1372:6 R and an eccentricity of e= 0:658. This wide BH-BH sys- tem does not merge in a Hubble time. The evolution within COMPAS1 model is fairly analo- gous to COMPAS2 model. However, higher wind mass- loss and less internal mixing leads to formation of a much less massive BH-BH system. BHs in this model are 20 M while in COMPAS2 they are 32 M . The orbital separation at BH-BH formation is a= 740:6 R while the eccentricity remains virtually unchanged ( e= 0:68) resulting in a coalescence time longer than the Hub- ble time. Results of evolution in the COMPAS3 model are also broadly similar to those of the COMPAS2 model (see Tab. 1). The additional condition of pre-RLOF however, changes a few key points. The rst RLOF (stable, TTMT, star A to star B) decreases the orbital separation from 11165 :6 Rto 739:2 Rand the eccentricity becomes e= 0:0. As star B evolves, stellar winds increase this orbital separation to a= 1097:3 Rright before the second RLOF (stable, TTMT, star B to BH A). At this point, the eccentricity is e= 0:023, increased from the previous the formation of the rst BH. The second RLOF de- creases the separation to a= 501:9 R, and again the orbit is circularized. The binary orbital separation at the time of formation of the BH-BH system (with 32 M BHs) becomes a= 779:6 Rwhile the orbital eccen- tricity is negligible ( e= 0:054). Though COMPAS3 model evolution decreases the binary orbital we note that this change is not to create a close double BH system that merges in a Hubble time. We note that COMPAS3 creates the closest BH-BH binary obtained in the three COMPAS models. We note that to create a similarly circularized double BH system with the same masses, that merges within a Hubble time, the orbital separation at BH-BH formation can at most be about 46 :7 R(Peters 1964). A highly eccentric orbit can also decrease the merger time. However, for a BH- BH system with same masses and orbital separation as in the COMPAS3 model, this cut-o eccentricity should be at least e= the structure of the BH X-ray binary Cyg X-1 (though a less massive system than Mk42) was used in StarTrack (Wiktorowicz et al. 2014) and COM- PAS (Neijssel et al. 2021) calculations to argue that the future evolution of Cyg X-1 (Miller-Jones et al. 2021)may also lead to a wide BH-BH system which will not merge in a Hubble time. 3.3. MESA models In this section we use the MESA code to check the outcome of the CE phase encountered in the three StarTrack models from Sec. 3.1 (see Sec. 2.3.3 and Klencki et al. 2021, for the method). In the model MESA2 we evolve a star with Mzams = 133 M , overshooting ov= 0:2 and Dutch winds with fwind = 1:5. This model produces at some point of its post-MS evolution a star with a mass of M= 80:9 M and a He-core mass of Mcore = 62:2 M, radius of R= 700 R Te9000K (an outer radiative envelope; the envelope would not become convective until least Te.4500K Klencki et al. (2020)). At this point we cal- culate the envelope binding energy (obtained from inte- gration of the mass distribution over the entire for the internal energy of the envelope, then we subtract the BH accretion luminosity that binding energy, to obtain Ebind= translates to MESA = 0:012. This model resem- bles the star B at the onset of CE in the of evolution. The orbital energy at the onset of CE and the is a= 6:2 R(corresponding to a This was obtained under the assumption of a 100% eciency of the orbital en- ergy transfer to unbind the envelope ( = 1:0). The radius of the exposed core of star B is 2 :84 R( Hurley et al. (2000) formulae) while its new Roche lobe is only 2:64 Rand we assume a CE merger in such case. In the StarTrack2 model the binding energy ( = 0:103) as it scales with /1=by factor of9 as compared with detailed MESA estimate ( = 0:012). Note that since the BH accretes part of the envelope during the CE inspiral (for the numerical treatment of this process see Appendix in Belczynski et al. (2002)), the binary does not need to balance the entire binding energy of the envelope ( Ebind) with the orbital However, this has no in uence on our conclusion above. Accretion onto the BH is estimated at the level of0:5 M(see Sec. 2.1), while this MESA model underestimates the mass of the stellar envelope found in StarTrack2 simulation by 7:4 M(Menv= 26:1 Min StarTrack2 simulation, and Menv= 18:7 M in the above MESA simulation). The other two population synthesis models from Star- Track also produce CE mergers (Thorne- _Zytkow objects) if MESA binding energy estimate is used. In StarTrack1 model the CE donor was estimated to have = 0:050, while the MESA calculation gives = 0:007. In Star- Track3 model = 0:050 as contrasted with MESA esti- mate of= 0:008. 3.4. Pavlovskii models A binary star resembling Mk 34 is evolved with the Pavlovskii2 model (Sec. 2.4). Evolution to RLOF initi- ated by star B is the same as in the StarTrack2 model (.1% dierences in binary parameters are here the binary undergoes stable a RLOF: TTMT instead of CE. During mass exchange mass/loss the orbital separation changes from a= 1507!928 R 9 and star B is stripped from its H-rich envelope ( Mb= 87:7!62:3 M) becoming a massive Wolf-Rayet star. Accretion onto the BH is negligible as the mass transfer was highly supper-Eddington ( Ma= 35:2!35:2 M) and the loss of angular momentum (given the mass ra- tio at the onset of RLOF: q= 87:7=35:2 = 2:5) causes orbit to decrease in size by a factor of 1 :6; this may be compared with the orbital decrease by a factor of 63 dur- ing CE in the StarTrack2 model. After Wolf-Rayet wind mass loss (Mb= 33:9 M,a= 1427 R ) star B to a BH ( Mb= 33:3 M). Aftertevol= 4:16Myr of binary evolution a wide BH-BH binary is formed with a coalescence time of tcoal= 6:0106Gyr. In the Pavlovskii1 model, the binary follows a but a lower-mass wide BH-BH binary forms (21:9 + 22:1 M) due to stronger winds and lower core masses adopted in this model (see Tab. 1). In Pavlovskii3 model, due to strongly increased loss of the orbital angular momentum during TTMT, the system does not survive the rst TTMT. It ends with a stellar merger of the Hertzsprung gap star donor (Ma= 80:2 M) with its main-sequence star compan- ion (Mb= 98:9 M). The actual mass and the fate of the stellar-merger product is uncertain. Both observa- tions and simulations of stellar mergers are usually re- lated to low-mass stars, which are not BH progenitors (J. C. Lombardi et al. 2002; Tylenda & Kami nski 2016), or are calculated for dynamical collisions in dense stellar clusters (Glebbeek et al. 2013). It seems that a rather low mass fraction is lost during stellar mergers (J. C. Lom- bardi et al. 2002; Lombardi et al. 2006; Glebbeek et al. 2013). Assuming that the merger product in our simula- tion will become a Hertzsprung-gap star with the mass of 163 M (similarly to the scheme used in Olejak et al. (2020) with 20% of the less massive star being ejected during the merger) the single star will end its evolution either as (i)PSN leaving no remnant if classical PSN models are used (Woosley 2017; Leung et al. 2019), or (ii)as a single3040 M BH, if non-standard PSN models are used (see Fig.1 of Belczynski 2020, and ref- erences therein). The amount of angular-momentum loss through the L2 Lagrangian point adopted in the Pavlovskii3 considered to be an upper limit whereas the standard StarTrack jloss= 1:0 used in models Pavlovskii1 and Pavlovskii2 is instead close to the lower limit as indicated by MacLeod et al. (2018) and MacLeod & Loeb (2020). The allows to avoid a stellar merger dur- ing the rst TTMT and would lead to the formation of a wide BH-BH binary from Mk 34 is ofjL2of MacLeod & Loeb (2020)). This demon- strates that even with increased angular momentum losses, it is possible to form either a wide BH-BH bi- nary (tdelay> thub) with minimal separation of about a= 609 R (tdelay = 2:6105Gyr> thub) or a stellar merger but not a close BH-BH system. 3.5. Quasi single star evolution models 3This range corresponds to changing mass of donor and accretor during RLOF.Based on MESA models (see Sec. A), current literature and simple estimates, we follow the future evolution of the Mk 34 binary with non-expanding stars. We put the two stars on an eccentric ( e= 0:68) and wide orbit ( a= 780 R ). These stars lose &100 M during their MS life in stellar winds expanding the orbital separation ( a > 1000 R ; see Sec. A). At the post-MS closest encounter of these two stars (periastron), Roche lobe radii of both components are Rlobe>100 R . The radii of both stars areR< 100 R for many MESA models. There is no mass exchange between the stars. Depending on (i)the mass and core mass of non- expanding stellar models and (ii)the mass (in reality central temperature and density) and the range allowed for the onset of a PPSN/PSN, we can envision several dierent outcomes of Mk 34's future evolution. If both stars have core masses as high as 61 M and 58 M at TAMS (see MESA models with ov= 0:33 and fwind = 1:0 in Sec. A), these cores will reach 65 M at the time of oxygen burning, which will then become ex- plosive leading to PSN (Woosley 2017). Each star gets disrupted, leaving no compact object remnant but pro- ducing luminous PSN supernova (model: Tab. 1). There is a signicant caveat to the above to recent studies (Woosley 2017; Limongi & Chie 2018; Farmer et al. 2020; Costa et al. 2021; Far- rell et al. 2021) very low metallicity stars can produce BHs with mass as high as 8090 M . But there are also detailed MESA stellar evolutionary models that allow for the formation of BHs with 70 M avoiding a PPSN/PSN (Belczynski et al. 2020b) at high metal- licity. If this scenario is adopted then it is expected that 60 + 60 M wide BH-BH binary would form lower mass stars/cores at TAMS, we expect avoid- ing a PPSN/PSN and we predict the formation of a wide BH-BH binary. For example, if we take MESA models with ov= 0:4 andfwind = 1:5 they will produce stars with M= 32 M (Mcore;TAMS = 30 M ) andM= 32 M (Mcore;TAMS = 30 M ). These stars are not subject to PPSN/PSN and depending on the post-MS stellar wind mass loss will form 30 M BHs (model: formation of a wide BH-BH binary, with coalescence time exceeding the Hubble time, is predicted. If we push MESA models even further to higher over- shooting and stronger winds ( ov= 0:5 andfwind= 2:0) we produce stars with M= 21 M (Mcore;TAMS = 19 M ) andM= 21 M (Mcore;TAMS = 19 M ) at TAMS. This will also lead to the formation of a wide BH-BH binary but with 20 M BHs at most have investigated the future evolutionary tracks and fate of the most massive known binary system Mk 34. Several interesting possibilities seem to exist (see Tab. 1). However, it is impossible to decide with certainty (due to various stellar and binary physics uncertainties) which predicted fate is the correct one (if any). If very massive stars at LMC metallicity and with moderate rotation expand during their post-MS evolu- tion (expected for low overshooting) then we predict the10 following evolution sequence for Mk 34: RLOF A!BHA!RLOF B!TZA+B=BH B(1) where indices and mark the more- and less- massive component of Mk 34 respectively, BHA=Bde- notes the BH formation from a given component, and TZA+B means the formation of a Thorne- _Zytkow ob- ject from both binary components in the second RLOF. The rst RLOF (initiated by star A) is always found to be stable (TTMT/NTMT), while the second RLOF (donor: star B) can be either stable or dynamically un- stable (CE). Additionally, the binary system may not survive the rst RLOF while both stars merge forming a single star that will be either subject to PSN (no rem- nant) or will form a single BH: RLOF A!single star!PSN=BH (2) If such massive stars do not expand (for example, be- cause of signicant overshooting) the future proceeds without any binary A=PSN A!BHB=PPSN B=PSN B(3) Under very optimistic conditions (development and survival of CE initiated by a massive star with a ra- diative envelope and with comparable mass companion) Mk 34 may form a heavy BH-BH merger that would be a source of Depending on our assump- tions on mass loss and mixing in stellar interiors we nd formation of a20 + 20 M close BH-BH system that resembles LIGO/Virgo detection of GW190408 18:4+3:3 3:6M: Abbott et al. (2021)) or  30 + 30 M BH-BH system that would look similar to GW150914 (35 Abbott et al. (2016)) or to GW190828 063405 Abbott et al. (2021)). If this does not work, the formation of such LIGO/Virgo BH-BH mergers can still be obtained with more realistic CE input physics in the isolated bi- nary evolution (Belczynski et al. 2016; Spera et al. 2019; Patton et al. 2021). The detailed evolutionary estimates of a very massive star envelope binding energy do not allow for a CE sur- vival of the Mk 34 descendant binary even under assumptions (100% orbital energy used to eject the envelope with the help of internal energy of the gas and accretion luminosity from the If this is taken into account, then instead of form- ing a close BH-BH binary in CE scenario, we encounter the formation of a Thorne- _Zytkow object (BH sinks into the center of the post-MS massive star). This single ob- ject would rst appear as a post-MS massive star (most likely a classical Wolf-Rayet star), that starts to expand, cooling o and getting redder as the envelope pus up in response to the BH inspiralling in a H-rich envelope. Once the BH sinks into the star's core (the majority of star mass at this point) the accretion of helium is ex- tremely rapid ( &1 Mmin1, as in a collapsar engine; Fryer & Woosley 1998) and the core disappears and then the rest of the star is accreted as well and the entirely from sky. Such a transient should be visible in optical/infrared (initial expansion of the enve- lope before it collapses onto the BH), although at this moment there are no available calculations of the lightcurve or spectra for such heavy mergers ( 100 M post- MS star and2030 M BH). Observationally, various red novae/transients were proposed to be the outcome of CE mergers (Tylenda et al. 2005; Ivanova et al. 2013a; Kami nski et al. 2015; MacLeod et al. 2017b). It may be even possible that such a merger would lead to a gamma- ray burst (GRB). If there is enough angular momentum in the He core and BH then such a conguration may lead to formation of jets powering a GRB (Zhang & Fryer 2001). The angular momentum transport in stellar inte- riors of massive stars is not fully constrained, although low eective spins of LIGO/Virgo BH-BH mergers seem to indicate ecient angular momentum transport in mas- sive stars (Spruit 2002; Fuller et al. 2019; Bavera et al. 2020; Belczynski et al. 2020a). In such a case a slowly spinning He core{BH system would have only a small chance of producing a GRB. Another obstacle in produc- ing a GRB in this case is 30 M of H-rich envelope for jets to punch through. Yet, there are signs that such jets do form in massive stars in CE mergers and they try to breakout from stellar interiors (Th one et al. 2015). Another possibility for the future evolution of Mk 34 is to avoid the CE phase entirely, even if both stars in this binary expand. Since the orbital separation is not very large for this system, any RLOF encountered in the evo- lution is bound to happen when the donor star is not too large (a radiative envelope). Additionally, since the stars in Mk 34 are of similar mass, any RLOF is not bound to happen at extreme mass ratio. Taking this into account, this binary may evolve through two episodes of stable RLOF (rst initiated by initially more massive star, and then by the other star). Although such RLOF episodes may decrease the orbital separation, such orbit shrink- ing will be not large enough to lead to the formation of a close BH-BH binary (i.e., with a merger time smaller than the Hubble time) for the initial binary congura- tion of Mk 34. Instead, a wide BH-BH binary forms with separation as large as &1000 R . Such a descen- dant of Mk 34 cannot be a LIGO/Virgo source, but could possibly be detected by microlensing observations. The magnication of the source in such a microlensing event would last months and would include bumps typical of a binary lens and would be potentially detectable in LMC. Finally, it is also possible that neither of the stars in Mk 34 will experience any signicant expansion in their post-MS evolution. In such a case, Mk 34 expands due to wind mass loss from both stars and forms a wide sys- tem when both components end their nuclear of1000 R ). Depending on the highly uncertain mixing physics and not fully constrained nu- clear reaction rates (overshooting, rotation, that set in stellar cores, massive components of Mk 34 may or may not be a subject to signicant mass loss associated with PPSN during oxygen burning. We predict the formation of a wide BH-BH system with comparable-mass BHs in the mass range4090 M . Alternatively, both stars in Mk 34 may be subject to a full- edged get disrupted in luminous PSNe's (Higgins et al. 2021). To summarize, we conclude that we cannot yet pre- dict the fate of a massive binary such as Mk 34. The involved stellar and binary physics uncertainties are still too overwhelming. However, our study oers several con-11 ditional statements that shed light on the future evolu- tion of this massive binary. If the stars in Mk 34 expand in their post-MS evolution then they are bound to ini- tiate two RLOF interactions (one by each the rst interaction is always stable and does not threaten the survival of this system, the second one is more problematic. It is not at all clear if the second inter- action will be dynamically stable. If it is not, then a CE phase develops and the most likely fate of the system is then the merger of the two binary components, with a red nova or a GRB. If the second inter- action is stable then the RLOF will lead to the formation of a wide and massive BH-BH system with a merger time much larger than the Hubble time. Such a system is not a potential LIGO/Virgo source of high-frequency gravi- tational waves, but it may produce a microlensing event. If the stars in Mk 34 do not expand, which is also al- lowed by the current detailed evolutionary models, then we predict either the formation of a wide and potentially very massive BH-BH system, or the spectacular death of both stars in luminous pair-instability supernovae that leave no BHs any given origin scenario (i)there is a large number of input physics uncertainties or even unknowns (ii)the implementation of the physical pro- cesses that involve these uncertainties in the numerical codes are far from being based on rst-principle physics, so even probing the full range of a given parameter might not get the right answer, and (iii) there are more pa- rameters and thus uncertainties than are commonly re- alized (for example, there are at least 30 parameters in the isolated binary evolution, even though it is usually thought that a small subset of them are the most im- portant in determining binary outcomes). Thus are unjustied at this time. As an example of how biases could enter the suppose that we use the rapid supernova engine model of Fryer et al. (2012), which naturally pro- duces a25 Mmass gap between neutron stars and black holes. Then the discovery of the 2 :6 Mob- ject in GW190814 would rule strongly against isolated binary evolution and in favor of another channel, such as primordial black holes or mergers in dense stellar systems (e.g., two neutron stars could merge to make a2:6 Mblack hole). But perhaps the delayed supernova engine model of Fryer et al. (2012) is a better description; in this model the compact objects with mass 2 :6 Mare produced naturally and the iso- lated binary formation channel is perfectly viable. Or perhaps some other supernova model is selected by Na- ture, which would change the Bayes factor between the models that are considered. Similar considerations apply to the high-mass merger GW190521, which is consistent with binary stellar evolution given the substantial uncer-tainties (Farmer et al. 2020; Belczynski 2020; Kinugawa et al. 2020; Vink et al. 2021; Costa et al. 2021; Mehta et al. 2021). We have focused on the specic system Mk 34, but our caveat extends to analyses of the full population. For example, Olejak et al. (2021) explored the eects on the BH-BH population that stem from dierent treat- ments of the common envelopes. Over the range of mod- els they studied, the population characteristics For example, the BH-BH merger rate varied from 18 with the current LVC estimate of 15 :338:8 Gpc3yr1: The LIGO et al. (2021)) to 88 Gpc3yr1. The BH mass distribution can be consistent with the LVC esti- mate M masses), or very inconsistent with the < 15 M and/M3:3for heavier black holes). The mass ratio distribution can have one peak or two peaks. To reiterate, even this broad range of predicted population characteristics does not include many other possible variations of aspects of stellar and binary evolution. What must be done to reach a stage in which we can draw rm conclusions about the BH-BH system origins? More and better data will obviously help: for exam- ple, if multiple events point to a compact object in the 25 Mrange then this tells us that the lower mass gap is not a major feature of the mass distribution. Rare individual events, if they have denitively can point to particular origins. For exam- ple, an event with many cycles that is clearly highly ec- centric would favor a dynamical origin, and a compact object with a mass <0:5 Mwould signify a primordial black hole. But we emphasize that a detailed work on the physics of BH formation in each case is essential: sta- tistical analyses must be grounded in both thorough and accurate physics and astrophysics. We thank Lukasz Wyrzykowski, Tomasz Kami nski, Chris Fryer and Daniel Holz for useful comments on the manuscript. KB, AR, AO acknowledge support from the Polish National Science Center grant DC and SS acknowledge the support of the Australian Research Council Centre of Excellence for Gravitational Wave Discovery project number CE170100004. This work made use of the OzSTAR high performance computer which is funded by Swinburne University of Technology and the National Collaborative Research Infrastructure Strat- egy (NCRIS). MCM acknowledges support from NASA ADAP grant 80NSSC21K0649. He performed part of his work on this paper at the Aspen Center for Physics, which is supported by National Science Foundation JPL was supported in part by a grant from the French Space Agency B. P., et al. 2016, Phys. Rev. Lett., 116, 061102 Abbott, R., et al. 2020a, Phys. Rev. Lett., 125, 101102 |. 2020b, ApJ, 896, L44 |. 2021, Physical Review X, 11, 021053 Bavera, S. S., et al. 2020, A&A, 635, A97 Belczynski, K. 2020, ApJ, 905, L15Belczynski, K., Bulik, T., Fryer, C. L., Ruiter, A., Valsecchi, F., Vink, J. S., & Hurley, J. R. 2010a, ApJ, 714, 1217 Belczynski, K., Dominik, M., Bulik, T., O'Shaughnessy, R., Fryer, C. L., & Holz, D. E. 2010b, ApJ, 715, L138 Belczynski, K., Holz, D. E., Bulik, T., & O'Shaughnessy, R. 2016, Nature, 534, 512 Belczynski, K., Kalogera, V., & Bulik, T. 2002, ApJ, 572, K., Kalogera, V., Rasio, F. A., Taam, R. E., Zezas, A., Bulik, T., Maccarone, T. J., & Ivanova, N. 2008, ApJS, K., Taam, R. E., Kalogera, V., Rasio, F. A., & Bulik, T. 2007, ApJ, 662, 504 Belczynski, K., Wiktorowicz, G., Fryer, C. L., Holz, D. E., & Kalogera, V. 2012, ApJ, 757, 91 Belczynski, K., et al. 2020a, A&A, 636, A104 |. 2020b, ApJ, 890, J. M. 2020, MNRAS, 493, J. M., et al. 2014, A&A, 570, A38 B ohm-Vitense, E. 1958, ZAp, 46, 108 Bonanos, A. Z., et al. 2004, ApJ, 611, L33 Brott, I., et al. 2011, A&A, 530, A115 Castro, N., Fossati, L., Langer, N., Sim on-D az, S., Schneider, F. R. N., & Izzard, R. G. 2014, A&A, 570, D., Stevenson, S., Hurley, J. R., Bailes, M., & Broekgaarden, F. 2021, MNRAS, 504, 3682 Choi, J., Dotter, A., Conroy, C., Cantiello, M., Paxton, B., & Johnson, B. D. 2016, ApJ, 823, 102 Claret, A., & Torres, G. 2018, ApJ, 859, 100 Costa, G., Bressan, A., Mapelli, M., Marigo, P., Iorio, G., & Spera, M. 2021, MNRAS, 501, 4514 de Jager, C., Nieuwenhuijzen, H., & van der Hucht, K. A. 1988, A&AS, 72, 259 Dominik, M., Belczynski, K., Fryer, C., Holz, D., Berti, B., Bulik, T., Mandel, I., & O'Shaughnessy, R. 2012, ApJ, 759, 52 Farmer, R., Renzo, M., de Mink, S., Fishbach, M., & Justham, S. 2020, arXiv e-prints, B., Holz, D. E., & Farr, W. M. 2018, ApJ, 854, L9 Farr, W. M., Stevenson, S., Miller, M. C., Mandel, I., Farr, B., & Vecchio, A. 2017, Nature, 548, 426 Farrell, E., Groh, J. H., Hirschi, R., Murphy, L., Kaiser, E., Ekstr om, S., Georgy, C., & Meynet, G. 2021, MNRAS, 502, L40 Fishbach, M., & Holz, D. E. 2020, arXiv G., et al. 2021, arXiv e-prints, C. L., Belczynski, K., Wiktorowicz, G., Dominik, M., Kalogera, V., & Holz, D. E. 2012, ApJ, 749, 91 Fryer, C. L., & Woosley, S. E. 1998, ApJ, 502, L9 Fuller, J., & Ma, L. 2019, ApJ, 881, L1 Fuller, J., Piro, A. L., & Jermyn, A. S. 2019, MNRAS Ge, H., Webbink, R. F., Chen, X., & Han, Z. 2015, ApJ, 812, 40 Gilkis, A., Shenar, T., Ramachandran, V., Jermyn, A. S., Mahy, L., Oskinova, L. M., Arcavi, I., & Sana, H. 2021, MNRAS, E., Gaburov, E., Portegies Zwart, S., & Pols, O. R. 2013, Monthly Notices of the Royal Astronomical Society, afener, G., & Hamann, W. R. 2008, A&A, 482, 945 Grevesse, N., Noels, A., & Sauval, A. J. 1996, in of the Pacic Conference Series, Vol. 99, ed. S. S. Holt & G. Sonneborn, 117 Heger, A., Langer, N., & Woosley, S. E. 2000, ApJ, 528, 368 Higgins, E. R., Sander, A. A. C., Vink, J. S., & Hirschi, R. 2021, MNRAS, 505, 4874 Humphreys, R. M., & Davidson, K. 1994, PASP, 106, 1025 Hurley, J. R., Pols, O. R., & Tout, C. A. 2000, MNRAS, 315, 543 Hurley, J. R., Tout, C. A., & Pols, O. R. 2002, MNRAS, 329, 897 Ivanova, N., Justham, S., Avendano Nandez, J. L., & Lombardi, J. C. 2013a, Science, 339, 433 Ivanova, N., et al. 2013b, A&A Rev., 21, 59 J. C. Lombardi, J., Warren, J. S., Rasio, F. A., Sills, A., & Warren, A. R. 2002, The Astrophysical Journal, 568, 939 Kalogera, V., & Webbink, R. F. 1996, Astrophys. J., 458, 301 Kami nski, T., Mason, E., Tylenda, R., & Schmidt, M. R. 2015, A&A, 580, A34 King, A. R., Davies, M. B., Ward, M. J., Fabbiano, G., & Elvis, M. 2001, ApJ, 552, L109 Kinugawa, T., Nakamura, T., & Nakano, H. 2020, arXiv J., Nelemans, G., Istrate, A. G., & Chruslinska, M. 2021, A&A, 645, A54 Klencki, J., Nelemans, G., Istrate, A. G., & Pols, O. 2020, A&A, 638, A55 K ohler, K., et al. 2015, A&A, 573, A71 Leung, S.-C., Nomoto, K., & Blinnikov, S. 2019, arXiv M., & Chie, A. 2018, ApJS, 237, 13 Lohr, M. E., Clark, J. S., Najarro, F., Patrick, L. R., Crowther, P. A., & Evans, C. J. 2018, A&A, 617, A66 Lombardi, J. C., J., Proulx, Z. F., Dooley, K. L., Theriault, E. M., Ivanova, N., & Rasio, F. A. 2006, ApJ, 640, 441 MacLeod, M., Antoni, A., A., Macias, P., & Ramirez-Ruiz, E. 2017a, ApJ, 838, 56 MacLeod, M., & Loeb, A. 2020, ApJ, 893, 106 MacLeod, M., Macias, P., Ramirez-Ruiz, E., Grindlay, J., Batta, A., & Montes, G. 2017b, ApJ, 835, 282MacLeod, M., Ostriker, E. C., & Stone, J. M. 2018, Journal, 863, 5 Mahy, L., et al. 2020, A&A, 634, A118 Mandel, I., & Broekgaarden, F. S. 2021, arXiv P., Pappas, K. M. W., Gallegos-Garcia, M., Berry, C. P. L., Taam, R. E., Kalogera, V., & Podsiadlowski, P. 2021, A&A, 650, A107 Marchant, P., Renzo, M., Farmer, R., Pappas, K. M. W., Taam, R. E., de Mink, S., & Kalogera, V. 2018, arXiv e-prints Massey, P., Penny, L. R., & Vukovich, J. 2002, ApJ, 565, 982 Mehta, A. K., Buonanno, A., Gair, J., Miller, M. C., Farag, E., deBoer, R. J., Wiescher, M., & Timmes, F. X. 2021, arXiv e-prints, J. C. A., et al. 2021, Science, 371, 1046 Mondal, S., Belczy nski, K., Wiktorowicz, G., Lasota, J.-P., & King, A. R. 2020, MNRAS, 491, 2747 Neijssel, C. J., Vinciguerra, S., Vigna-G omez, A., Hirai, J. C. A., Bahramian, A., Maccarone, T. J., & Mandel, I. 2021, ApJ, 908, 118 Neijssel, C. J., et al. 2019, MNRAS, 490, 3740 Nitz, A. H., & Capano, C. D. 2021, ApJ, 907, L9 Nugis, T., & Lamers, H. J. G. L. M. 2000, A&A, 360, 227 Olejak, A., Belczynski, K., Bulik, T., & Sobolewska, M. 2020, A&A, 638, A94 Olejak, A., Belczynski, K., & Ivanova, N. 2021, arXiv M. 2021, A&A, 649, A57 Patton, R. A., Sukhbold, T., & Eldridge, J. J. 2021, arXiv e-prints, K., & Ivanova, N. 2015, MNRAS, 449, 4415 Pavlovskii, K., Ivanova, N., Belczynski, K., & Van, K. X. 2017, MNRAS, 465, 2092 Paxton, B., Bildsten, L., Dotter, A., Herwig, F., Lesare, P., & Timmes, F. 2011, ApJS, 192, 3 Paxton, B., et al. 2013, ApJS, 208, 4 |. 2015, ApJS, 220, 15 |. 2018, ApJS, 234, 34 |. 2019, arXiv e-prints Peters, P. C. 1964, Physical Review, 136, P., Joss, P. C., & Hsu, J. J. L. 1992, ApJ, 391, 246 Pollock, A. M. T., Crowther, P. A., Tehrani, K., Broos, P. S., & Townsley, L. K. 2018, MNRAS, 474, 3228 Pols, O. R., Schr oder, K.-P., Hurley, J. R., Tout, C. A., & Eggleton, P. P. 1998, MNRAS, 298, 525 Rolleston, W. R. J., Trundle, C., & Dufton, P. L. 2002, A&A, 396, 53 Sana, H., et al. 2012, Science, 337, 444 Schneider, F. R. N., Izzard, R. G., Langer, N., & de Mink, S. E. 2015, ApJ, 805, 20 Schnurr, O., Casoli, J., Chen e, A. N., Moat, A. F. J., & St-Louis, N. 2008, MNRAS, 389, A., Langer, N., Grin, N. J., & Wang, C. 2019, A&A, 625, A132 Scott, L. J. A., Hirschi, R., Georgy, C., Arnett, W. D., Meakin, C., Kaiser, E. A., Ekstr om, S., & Yusof, N. 2021, MNRAS, T., et al. 2021, A&A, 650, A147 Spera, M., Mapelli, M., Giacobbo, N., Trani, A. A., Bressan, A., & Costa, G. 2019, MNRAS, 485, 889 Spruit, H. C. 2002, A&A, 381, 923 Stevenson, S., Sampson, M., Powell, J., Vigna-G omez, A., Neijssel, C. J., Sz ecsi, D., & Mandel, I. 2019, arXiv S., Vigna-G omez, A., Mandel, I., Barrett, J. W., Neijssel, C. J., Perkins, D., & de Mink, S. E. 2017, 8, 14906, [Nature K. A., Crowther, P. A., Bestenlehner, J. M., Littlefair, S. P., Pollock, A. M. T., Parker, R. J., & Schnurr, O. 2019, MNRAS, 484, 2692 The LIGO Scientic Collaboration et al. 2021, ApJ, 913, L7 Th one, C. C., de Ugarte Postigo, A., Fryer, C. L., & Kann, D. A. 2015, in Extragalactic Jets from Every Angle, ed. F. Massaro, C. C. Cheung, E. Lopez, & A. Siemiginowska, Vol. 313, 396{397 Thorne, K. S., & Zytkow, A. N. 1977, ApJ, 212, 832 Tylenda, R., & Kami nski, T. 2016, A&A, 592, A134 Tylenda, R., Soker, N., & Szczerba, R. 2005, A&A, 441, 1099 van den Heuvel, E. P. J., Portegies Zwart, S. F., & de Mink, S. E. 2017, MNRAS, 471, 4256 Vigna-G omez, A., et al. 2018, MNRAS, 481, 4009 Vink, J. S., Brott, I., Gr afener, G., Langer, N., de Koter, A., & Lennon, D. J. 2010, A&A, 512, L7 Vink, J. S., de Koter, A., & Lamers, H. J. G. L. M. 2001, A&A, 369, 574 Vink, J. S., Higgins, E. R., Sander, A. A. C., & Sabhahit, G. N. 2021, MNRAS, 504, 146 Vink, J. S., Muijres, L. E., Anthonisse, B., de Koter, A., Gr afener, G., & Langer, N. 2011, A&A, 531, A132

2108.10885

Krzysztof Belczynski

K. Belczynski, A. Romagnolo, A. Olejak, J. Klencki, D. Chattopadhyay, S. Stevenson, M. Coleman Miller, J.-P. Lasota, P.A. Crowther

The Uncertain Future of Massive Binaries Obscures the Origin of LIGO/Virgo Sources

comments welcomed

10.3847/1538-4357/ac375a

astro-ph.HE gr-qc

http://creativecommons.org/publicdomain/zero/1.0/

The LIGO/Virgo observatories have detected 50 This sample is large enough to have allowed several recent studies to draw conclusions about the branching ratios between versus dense stellar clusters as the origin of double BHs. It has also led to the exciting suggestion that the population is highly likely to black holes. Here we demonstrate that such conclusions cannot yet be robust, because of the large current uncertainties in several key aspects of binary stellar evolution. These include the development and survival of a common envelope, the mass and angular momentum loss during binary in stellar interiors, pair-instability mass loss and Using standard tools such as the population synthesis codes StarTrack and COMPAS and the detailed stellar evolution code MESA, we examine as a case study the possible future evolution of Melnick 34, the most massive known binary star system. We show that, despite its well-known various assumptions regarding stellar and binary physics predict a wide variety of outcomes: from a close BH-BH binary (which would lead to a potentially detectable coalescence), through a wide BH-BH binary (which might be seen in microlensing observations), or a Thorne-Zytkow object, to a of both objects by pair-instability supernovae. Thus since the future of massive binaries is inherently uncertain, sound predictions about the properties of BH-BH systems are highly challenging at this time. conclusions about the formation channels for the LIGO/Virgo BH-BH merger population is premature.

2 1.1 mas per exposure, this would create incredibly maps of the 108stars that the EML survey will monitor. In this paper, we forward model the constraints that the Nancy Grace Roman Space Telescope's EML Sur- vey would place on DM subhalos in the Milky Way. In Section II we summarize the lensing signature we expect. In Section III we discuss our forward modeling approach. In Section IV we present our results, discuss possible sys- tematics, and identify avenues for improved constraints. The code used in this paper is publicly available . II. THEORY In this section, we provide a summary of the expected by DM substructure. For a more detailed discussion of these signatures, see Ref. [27]. The apparent change in a star's position caused by one lens (1) wherecis the speed of light, bis the impact parameter on the sky,Dlis the distance to the lens, Dsis the distance to the star, and Mlis the lensing mass within a cylinder of radiusband height Ds. From now on, we take limit: DsDl. Then, the apparent velocity change is [27, 28]: ~ 
_~= v)^b~ v M0 l b(^b
~ v)^b ;(2) where _d dt,~ v_~b, and0d db. The apparent accelera- tion due to a DM lens is then: ~ 
~= 4G c2Ml b3A(b) +M0 l b2B(b) =8(^b
~ v)2^b+ 2b(^b
_~ v)^b+ 2v2^b +4(^b
~ v)~ vb_~ v (4) B(b) = 5( ^b
~ v)2^b2(^b
~ v)~ vb(^b
_~ =(^b
~ v)2^b: (6) For the rest of this work, we set _~ v= 0. We also ignore any changes to the direction of ~b. Note the very strong dependence on the impact pa- rameter here, ~ /1=b3. This will be important to our statistical of the lensing signature for two dierent mass proles (NFW and point-source) are given in Fig. 1. The dierences in the signatures are due to the depen- dence on the mass prole derivatives in Equation 3. For an NFW prole, all of the mass-impact parameter ra- tios (Ml=b3,M0 l=b2, andM00 l=b) are approximately thesame order of magnitude, with M00 l=bhaving the oppo- site sign as the other terms (for b=rs>1). The give a dipole. For the point source 0. This allows the quadrupole moment to dominate. These behaviors are easiest to see when using a vector spherical harmonic decomposition [see Ref. 28, for some discussion of this topic]. III. METHODS In general, we expect there to be some population of lenses that change the apparent accelerations of stars. Each star can be lensed by multiple lenses and the proba- bility of being lensed increases with the number of lenses. Thus, the lens signal for any one star will rely on Ml,vl, andb, as well as the total number of lenses, Nl. Here we forward model the expected distribution of induced by DM weak lensing. We then perform a Markov Chain Monte Carlo (MCMC) over dierent lens parameters to nd the constraints the Roman Space Tele- scope would achieve. In the following, we rst give the details of our statistical model and then describe our sam- pling eorts. A. Statistical Formalism Upon a rst attempt at this problem, we could try to fully forward model the system. This would involve placing the lenses at random with each sample of the MCMC. However, this approach quickly becomes in- tractable without considerable computational resources. It requires many samples to fully probe the impact pa- rameter distribution at any xed Nl, nevermind when also sampling Ml, which sets Nl. In addition, this for- ward modeling would be very sensitive to the number density of simulated stars in the eld of view (FOV). To achieve the correct result, a run with the same num- ber as the true FOV would be required. Then the dis- tances from these 108stars to the Nllenses would need to be computed at each sample. In this paper, we in- stead opt for a semi-analytic approach, which we nd reasonably accurate and more computationally ecient. Rather than sampling lens positions, we nd the distri- bution of expected impact parameters given the number of lenses in the eld. For each star in the eld, we then sample from this distribution to nd its impact parame- ter to the nearest lens. In the following, we explain this process in more detail. Equation 3 gives the apparent acceleration of 1 star from 1 lens. We rst consider the likelihood of 1 star's given apparent acceleration due to a population of lenses. We will then generalize to N?stars. Throughout, we assume that all lenses have the same mass and that their mass proles are set solely by this mass. First, let us consider the apparent acceleration of 1 star from a population of Nllenses with the same mass,3 FIG. 1: Example lensing-induced apparent accelerations assuming dierent mass proles. The lens is located at the black `X' at the center of each plot and it is assumed to be moving in the positive y-direction. Each blue arrow shows the direction of the induced acceleration and its color shows the magnitude. The black circles demarcate the NFW scale radius. Left: NFW lens. Right: Point-source lens. The dierences in the signals are due to the dependence on derivatives of the mass prole with respect to the impact parameter (see text below). Ml. We assume that we are in the weak lensing regime. Then the total acceleration from these Nllenses is just the vector sum of the signal from each lens: ~ l=NlX i=1~ l;i(Ml;~ vl;i;~bi); (7) where~ l;iis given by Equation 3. In general, the closest lens will dominate (since Equa- tion 3 has such a strong dependence on b). Thus, we now only consider the lensing signature from the closest lens to each star. The impact parameter, b, is then taken to always be the minimum impact parameter. Then the above equation simplies to ~ l~ l(Ml;~ vl;~bmin). We assume that the number of lenses is uniquely deter- mined by the mass. Specically, we assume that all of the dark matter is contained in halos of one mass, Ml. Then, to conserve the local dark matter density, the number of lenses for a given lens mass (8) whereDMis the local density of dark matter, AFOV is the area covered by the FOV, and dmaxis the maximum distance of the lenses. This assumes that the volume probed by the telescope is a square pyramid, which is a ne assumption for small FOVs (in other words, we take the at sky approximation). We always assume at least 1 lens within the volume. Note that we do expect lenses within the FOV for the very largest masses here; however, quantifying this involves an assumed subhalo mass distribution. We leave this to future work.The number of lenses, Nl, then sets the impact pa- rameter distribution as follows. First, we assume that the distribution of lenses is isotropic. This allows us to split~binto an angular portion, b, which will just have a at prior on [0 ;2), and a magnitude portion, bk~bk. We now also split binto an angular distance on the sky, , and a distance along the line of sight, dl. Speci- cally, we assume the small angle approximation and de- neb=dl. Given the FOV of the Roman Space Telescope (0.28 deg2), the small angle approximation is valid. We assume that the distribution of is indepen- dent of the distribution of dl. The distances to the lenses are given by simply assuming that they follow the dark matter density prole of the Milky Way. In particular, we assume the lenses' radial positions follow a Navarro, Frenk & White [NFW; 36] prole: (r) =4s r rs(1 +r rs)2; (9) whereris the distance from the MW center, rs= 18 kpc is the scale radius of the MW halo, and the density at that radius [i.e., we as- sume the same parameters as Ref. 37]. We also assume that all lenses are within 1 kpc of the Sun. Ref. [28] found that most of the power from DM lenses is con- strained within this radius. This constraint also allows us to continue assuming the distant source can be found using some geometry and a little bit of numerical integration (see Appendix A for a full derivation). In essence, the calculation is one of nding the probability of the minimum distance between randomly placed points on a plane. It is nd the probability for the distance between two ran- domly placed points [38]. This can then be generalized to4 Npoints. Then the minimum can be found using logic. In general, it is a function with a long tail toward low values that peaks near (2pNl)1lFOV, wherelFOV is the side-length of the FOV. Fig. 2a shows the form of the function for various numbers of lenses. In summary: p(~b) is the Dirac Delta function, dxeis the and Nl(Ml) is given by Equation 8. The velocity distribution of the lenses is taken to be independent of Ml,Nl, orb. It is set by the potential of the Milky Way, which we set from observations. As with the impact parameter, we assume that the lenses have isotropic velocities, which allows us to set a uniform prior on the angle of the lens velocity on the sky, p(v) = U[0;2). For the magnitude, vl, we assume the same velocity distribution that Ref. [28] assumes: p(vl) =N(~ v;~ v); (11) wherevl=k~ vlk,v= 0 km=s andv= 220 km=s. In addition, we do not allow the lenses to have velocity greater than the escape velocity of the Milky Way, v < 550 km=s. Finally, the probability of a star having a acceleration, given a lens mass Ml, lens velocity vl, and impact parameter ~b, is then: p(~ ljMl;~ vl;~b) (12) where the factor in front, 1 =42is given by the priors the measurement errors on the acceleration are Gaussian, it is then straightforward to nd the like- lihood of a single stellar acceleration measurement given a DM lens with mass Ml: L(~ jMl) =Z p(~ ljMl)N(~ ~ l;2 )d~ l; (13) where~ is the observed acceleration, N(;2) is a nor- mal distribution with mean, , and variance, 2,is the error of the observation, and p(~ ljMl) is given by marginalizing Equation 12 over ~ vl,,Nl, the full likelihood is then: L(~ jMl) =N?Y i=1Z d~ l;iN(~ i~ l;ijMl;~ vl;~b)d~ is straightforward to generalize this to lenses that have parameters beyond Ml.Note that the above likelihood does not explicitly in- clude the possibility of multiple stars being lensed by the same lens. However, some of this signal is preserved { see Section IV for further discussion of this point. A summary of our likelihood terms is given in Table I. B. Details of the Sampling In this work, we focus on three types of mass proles for our halos: point source, Gaussian, and NFW. In all cases, we assume a log-uniform prior on the mass sample in log10Ml. The point source case is most relevant for MACHOs; however, as we describe in Section IV, we expect mi- crolensing to be a better model for these lenses. Thus, it mostly serves as a test case in this work. The point source case has M(b) =Mlfor 0. Unlike the other two cases, the Gaussian prole is not a realistic prole for any DM models. However, it is a commonly used prole due to its analytical nature. We compute the results here in order to compare to previous work. The prole is dened as: (r) (15) whereM0is the total lens mass and R0is a Here we set M0=Mland we sample in log10Ml and log10R0. We use a log-uniform prior of [104;104] pc for the scale radius. The NFW prole, given in Equation 9, is the expected prole for DM. We choose to model the prole using the mass,Ml, and concentration parameter, c200. Thus (16) wherecis the critical density of the universe today, and x= 1 +c200. The scale radius, rs, is then given (17) whereM200is dened as the mass of a halo with density 200c. Here we set M200=Mland sample in log10Ml. We also sample in log10c200with a log-uniform prior the EML survey itself, we assume the following: the EML survey will measure the accelerations of 108 stars with sensitivity 0 :1as=yr2[35]. We also assume a square FOV2with area equal to the true FOV area, AFOV= 0:28 deg2. 2The true FOV will be a more complicated shape { see Fig. 2 of Ref. [32].5 (a) (b) FIG. 2: Probability Distribution Functions for Theory Likelihoods. a)Probability distribution functions for the magnitude of the angular size of the impact parameter, , of 1 star given a xed number of lenses, Nl. The black, dashed line gives the eective length of one Roman Space Telescope Field of View (FOV; 0.53 deg = 1908 a square area with the same area as the true FOV (0.28 deg2). The functions are arbitrarily normalized to t on this plot. See Appendix A for the derivation of these functions. b)Probability of the magnitude of an apparent acceleration for 1 star produced by DM lensing given a specic mass and other lens 12 marginalized over velocity and impact parameter). The dierent colors correspond to dierent assumed lens types. The black, dashed line gives the Roman Space Telescope's assumed acceleration sensitivity for 1 star in the EML survey ( = Model Used Lens Mass, p(Ml) Log-Uniform Prior on of DM in Halos, p(f) Uniform Prior on [0 ;1] Scale Radius for the Gaussian Prole, p(R0) Log-Uniform Prior on [104;104] pc Concentration Parameter for the NFW Prole, p(c200) Log-Uniform Prior on [100;108] Number of Lenses, p(Nl) Delta Function Set by the Mass of the Lens (Eqn. 8) Impact Parameter (total), p(~b) Set by Eqn. 10 and the relevant probabilities below Impact Parameter (on-sky angle), p(b) Uniform prior on [0 ;2) Impact Parameter (distance to lens), p(dl) Follows Milky Way DM density, see Eqn. 9 Impact Parameter (on-sky angular size), p() See Fig. 2a and Appendix A Lens Velocity (magnitude), p(vl) Normal distribution, see Eqn. 11 Lens Velocity (on-sky angle), p(v) Uniform prior on [0 ;2) TABLE I: Summary of Likelihood Terms and Assumed Models We use the emcee package to perform our MCMC. We only explicitly sample in the mass prole parameters de- scribed above. However, each sample of Equation 8. This then sets the probability distribu- tion for the impact parameter using Equation 10. We sampleN?times from this distribution. We then sample N?times from the vldistribution. The apparent accel- eration from lensing is calculated using these for each star via Equation 3. The likelihood is then calculated via Equation 14. IV. RESULTS & DISCUSSION Using the forward modeling procedure described above, we forecast the expected limits on DM substruc-ture that the Nancy Grace Roman Space Telescope's EML survey would place from lensing-induced acceler- ations on stars. Below we describe our results in detail and discuss possible limitations and future work. Our results are summarized in Table II. A. Results For Dierent Lens Proles Here we discuss our results for 3 dierent types of lenses: point sources, Gaussian proles, and NFW pro- les. For each of these, we consider a case where these halos make up all of the DM and a case where the frac- tion of DM contained in these halos is treated as a a test case, we rst consider point source lenses.6 TABLE II: Forecasted 95% Limits on DM Substructure Using the Roman Space Telescope EML Survey Mass Prole Type Mass Limits [log10M] Radius Limits [log10pc] Concentration Limits [log10c200] Fraction Limits Point Source >0:9 &<3:4 { { { Point Source + Fraction >0:3 &<3:2 { { <0:89 Gaussian <7:42 >0:61 { { Gaussian + Fraction <7:45 >0:61 { <0:95 NFW <7:49 { <2:50 { NFW + Fraction <7:45 { <2:56 <0:95 FIG. 3: Forecasted Constraints on Point Source Lenses by the Roman Space Telescope's EML Survey. The blue lines/areas show the results when the fraction of DM in these lenses is also sampled. The orange lines give the results when we assume that the fraction of DM in these halos is equal to unity. The dashed, vertical lines give the 95% credible intervals. The 2D plot shows the 68% and 95% contours. Note that the eects from these lenses is probably best described with microlensing. However, at large they would produce the weak lensing signa- tures described in this paper. To account for this, we place limits on the maximum apparent acceleration in- duced by lensing. We limited the maximum acceleration to a signal-to-noise ratio (SNR) of 1 for any single star. In practice, any stars with acceleration signatures at high SNRs would be cut from the data sample anyway. For both the case where we allow the fraction of the DM in these halos to vary and in the nominal case where all of the dark matter is in these subhalos, we nd that the Roman Space Telescope will be able to place con- straints at both high and low masses. These limits arecomparable to previous MACHO constraints [18, 19, 39]. The previous MACHO constraints prohibit MACHOs from making up all of the dark matter at any mass. How- ever, the constraints are weakest in the 10 100 M range. Those constraints allow for MACHOs to make up to 40% of the DM at 10 M . In the very we show here, the Roman Space Telescope will place similar constraints using these weak lensing signatures. Note that the point-source case appears to feature a detection, despite the use of fake data with no injected signal (i.e., Gaussian acceleration signals). This detec- tion is just the weakest point between two limiting cases. At the low-mass end, there are large numbers of lenses, so the probability of a small impact parameter is very large. At the high-mass end, each lens has a large impact on the surrounding stars, since the acceleration signal scales with mass (see Equation 3). The peak in the posterior is the maximum mass at which we expect more than 1 lens within the survey volume. As we describe above, these signals are probably better described via eorts to nd planets via microlensing with the Ro- man Space Telescope [34, 40] will also be able to on MACHOs and other compact objects. Next, we consider the Gaussian prole case, which is analogous to the compact lens case that Refs. [28, 37] consider. Fig. 4 shows the results for the EML survey. In Fig. 5 we show the results if we assume the same survey parameters as Ref. [28]'s WFIRST-Like survey (similar to the EML, but with a wider eld of view). We nd somewhat tighter constraints than the results quoted in Ref. [28]. This is most likely due to the use of forward modeling rather than the power spectrum ap- proach. The acceleration signature is not Gaussian, so the power spectrum approach misses part of the signal. Our result suggests that an alternative approach using higher order statistics, such as a bispectrum or trispec- trum analysis, would be we consider the NFW prole case (see Fig. 6). The mass constraints for this case are the least constrain- ing, as is expected given the diuse nature of these ha- los. However, we nd that the Roman Space Telescope would place interesting constraints on the concentration of these halos. Namely, it would be very sensitive to any halos with c200>104. It would even be able to probe down toc200100 at a 2level forM108M. For typical NFW halos in CDM, we expect c200values of around 1020 for Milky Way-sized objects. However,7 FIG. 4: Forecasted Constraints on Gaussian Lenses by the Roman Space Telescope. The blue lines/areas show the results when the fraction of DM in these lenses is also sampled. The orange lines give the results when we assume that all the DM is in these halos. The 2D plot shows the 68% and 95% contours. FIG. 5: Forecasted Constraints on Gaussian Lenses by a survey like the Roman Space Telescope's EML Survey, but with a larger eld of view ( fsky= 0:05). The blue lines give the 68% and 95% contours. The dashed black line gives the 95% upper limit from Ref. [28].for subhalos very close to their host galaxies, simulations have found that c200could extend upwards to 70 [41]. Thus, with only a modest improvement in sensitivity (ei- ther through analysis methods or survey strategy), the Roman Space Telescope's EML survey could be sensitive to CDM subhalos within the galaxy. B. Discussion & Future Work We now discuss the caveats of this study. First and foremost, the EML survey parameters have not been - nalized and are subject to change. As mentioned above, we also make several simplications (e.g., shape of the FOV) that could marginally eect the results. We dis- cuss the most impactful assumptions below: star systematics, the use of mono-mass popula- tions of lenses, and our modeling of multiple star lensing by single lenses. One possible systematic is the acceleration from binary star systems. These systems are ubiquitous and could be a serious systematic since they would produce large ac- celeration signatures. For low-mass binary systems, the most likely semi-major axis is a= 5:3 AU [for a review of binary star systems, see Ref. 42]. This, with M= 0:1M and a distance of 8 kpc from the observer gives an of 35as=yr2, which is many orders of magnitude greater than the expected lensing signal from a population of lenses. However, careful modeling could be used to distinguish the signals. In most cases, the pe- riodic motion of the stellar positions would allow us to discard these sources [see 34, for a similar discussion of how to distinguish binaries from planet signals]. We only consider mono-mass populations of lenses in our work. This allows us to consider the at dierent masses more easily, but it is not a realistic scenario. The expected subhalo mass function is largely uncertain, due to tidal eects, baryonic feedback, and DM physics. In addition, the radial dependence and expected proles of the subhalos can also heavily depend on these factors and vary from the assumptions made in this paper [see, e.g., 43{45]. Future work could look into the constraints on the mass function, as well as consider other radial dependencies and mass proles. Constraints on these properties would let us learn about both DM physics and baryonic as mentioned in Section III, we do not explic- itly consider the eects of a single lens aecting multiple stars. This would add extra correlations to the acceler- ation signal expected in the stars. In some sense, our p(jNl) likelihood does account for the particular distri- bution of impact parameters assuming a xed number of lenses in the eld. The magnitude of the signal should be preserved by our formalism, even if the angular dis- tribution is not. Our analysis here can be taken as a conservative estimate of the true signal. There are also several ways of improving these con- straints. We discuss 2 avenues for further 6: Forecasted Constraints on NFW lenses by the Roman Space Telescope. The blue lines/areas show the results when the fraction of DM in these lenses is also sampled. The orange lines give the results when we assume that all the DM is in these halos. The 2D plot shows the 68% and 95% contours. here: 1) using time domain signatures; 2) using observa- tions of other systems to improve paper solely considered the overall of these stars at the end of the Roman Space Telescope's nominal mission. However, we could also consider the information contained in the time of these sources. For example, consider a lens 1 kpc from us moving at 300 km =s in front of a background of 108stars distributed across the 2 deg2 EML survey area. If this lens was directly in front of onestar, it would take 60 years to be in the line of sight of a dierent star. Over 5 years, we still would expect to see a large dierence in the lensing pattern from this lens, since the signal scales so strongly with impact parameter. A detailed exploration of this topic is left to future work. One other way of improving constraints is to use com- plementary observations to the EML survey. A guest ob- server program for the Nancy Grace Roman Space Tele- scope that targets nearby galaxies could yield even bet- ter limits, provided the number of stars observed at accuracy is sucient. For example, an M31 survey combined with the Hubble Space Telescope PHAT survey [46] could be incredibly powerful. This would give a longer baseline to the observations and allow us to lever- age the time domain signatures we discuss above. How- ever, the EML survey's timing resolution will likely allow it to remain the best survey for observing these lenses. In this paper, we considered the constraints the Nancy Grace Roman Space Telescope's Exoplanet Survey would place on dark matter their lensing eects on the apparent accelera- tions of stars in the Galaxy. We forward modeled the ef- fects of various lens proles using a semi-analytic frame- work. While Roman will not be able to place on the fraction of DM in these lenses, we found that Roman's EML survey will place on the masses and other prole parameters of point source, Gaussian, and NFW lenses. In particular, the NFW bounds will be complementary to those placed by satellite counts by probing lower masses than is pos- sible with MW satellites [see, for example, 47]. Future work on the timing signatures of these lenses could lead to further improvement on these authors would like to thank Siddharth Mishra- Sharma for his very helpful comments. This work was done at the Jet Propulsion Laboratory, California Insti- tute of Technology, under a contract with the and Space Ad-ministration. This work was supported by NASA grant 15-WFIRST15-0008 Cosmol- ogy with the High Latitude Survey Roman Science In- vestigation Team (SIT). (c)2021. California Institute of Technology. Government sponsorship astropy [48], emcee [49], matplotlib [50], numpy [51], scipy [52] Appendix A: Derivation of the Impact would like to nd the distribution of impact param- eters given a random set of DM lenses and stars. First, note that the number of stars does not matter for this dis- tribution { the populations of each are totally separate, and we can consider each star independently. star is a random draw of the impact parameter this appendix, we keep the lenses at a xed distance from us and only consider their on-sky distribu- tion. The distance and on-sky distributions should be totally independent and the distance can be sampled consider a toy problem: a 1D line with length a where we place uniformly place our lenses. For now, we place a single lens. Then place a star on this line. It toowill be randomly and uniformly placed. The probability of the squared distance between the star and the lens, x, is given by [38]: f(x) =1 apx1 a2;for 0xa2; (A1) and 0 Let the random variable Xifor i2[1;Nl] be a random variable describing the squared distance of lens ifrom our one star. As Eq. 3 shows, the acceleration signal depends very strongly on the impact parameter ( /1=b3). The signal is then mostly determined by the lens closest to the star. Thus, we want to nd the probability that the closest lens has a squared distance, x. In other words, we =x). Let us start by working with the cumulative distribu- tion function (CDF), Note that this is equivalent to the probability of any one lens having squared distance less than x. The probability of this is then the converse of the probability that all of the lenses have squared distances greater than x. The prob- ability that any one lens has a squared distance = 1F(x). For all nlenses this is: = the CDF we are looking = obtain the probability distribution function (PDF) in this case, we simply take the derivative of Eqn. A2. This =x) consider the 2D case. To be totally general, we can consider a rectangle with sides a,b, andab. For just 1 star and 1 lens, the probability of a given squared distance is [38]: g(x) if 0 elsewhere. The generalization to the minimum case is the same as in the 1D case. We numerically inte- grate Eqn. A4 to obtain it's CDF, G(x). We then apply Eqn. A2 and take the derivative to obtain the PDF.10 [1] F. Zwicky, Die Rotverschiebung von Helvetica Physica Acta 6, 110 (1933). [2] V. C. Rubin, J. Ford, W. K., and N. Thonnard, Ex- tended rotation curves of high-luminosity spiral galaxies. IV. Systematic dynamical properties, Sa -&gt; Sc., ApJL 225, L107 (1978). [3] V. C. Rubin, J. Ford, W. K., and N. Thonnard, Rota- tional properties of 21 SC galaxies with a large range of luminosities and radii, from NGC 4605 (R=4kpc) to UGC 2885 (R=122kpc)., ApJ 238, 471 (1980). [4] E. Aprile etal. (Xenon Collaboration), Dark Mat- ter Search Results from a One Ton-Year Exposure of XENON1T, Phys. Rev. Lett. 121, 111302 D. S. Akerib etal. (LUX Collaboration), Results from a search for dark matter in the complete lux exposure, Phys. Rev. Lett. 118, 021303 (2017). [6] X. Cui etal. (PandaX-II Collaboration), Dark matter re- sults from 54-ton-day exposure of pandax-ii Rev. Lett. 119, 181302 (2017). [7] Y. Meng, Z. Wang, Y. Tao, A. Abdukerim, Z. Bo, W. Chen, X. Chen, Y. Chen, C. Cheng, Y. Cheng, X. Cui, Y. Fan, D. Fang, C. Fu, M. Fu, L. Geng, K. Giboni, L. Gu, X. Guo, K. Han, C. He, J. He, D. Huang, Y. Huang, Z. Huang, R. Hou, X. Ji, Y. Ju, C. Li, M. Li, S. Li, S. Li, Q. Lin, J. Liu, X. Lu, L. Luo, W. Ma, Y. Ma, Y. Mao, N. Shaheed, X. Ning, N. Qi, Z. Qian, X. Ren, C. Shang, G. Shen, L. Si, W. Sun, A. Tan, A. Wang, M. Wang, Q. Wang, S. Wang, S. Wang, W. Wang, X. Wang, M. Wu, W. Wu, J. Xia, M. Xiao, X. Xiao, P. Xie, B. Yan, X. Yan, J. Yang, Y. Yang, C. Yu, J. Yuan, Y. Yuan, D. Zhang, M. Zhang, P. Zhang, T. Zhang, L. Zhao, Q. Zheng, J. Zhou, N. Zhou, X. Zhou, and Y. Zhou, Dark Matter Search Results from the PandaX-4T Commissioning Run, arXiv e-prints (2021), arXiv:2107.13438 [hep-ex]. [8] S. Ghigna, B. Moore, F. Governato, G. Lake, T. Quinn, and J. Stadel, Dark matter haloes within clusters, MN- RAS300, 146 (1998), [astro-ph]. [9] A. Klypin, A. V. Kravtsov, O. Valenzuela, and F. Prada, Where Are the Missing Galactic Satellites?, ApJ 522, 82 (1999), [astro-ph]. [10] A. M. Green, S. Hofmann, and D. J. Schwarz, The rst WIMPy halos, J. Cosmology Astropart. Phys. 2005 , 003 (2005), [astro-ph]. [11] T. Bringmann, Particle models and the small-scale struc- ture of dark matter, New Journal of Physics 11, 105027 (2009), arXiv:0903.0189 J. S. Bullock and M. Boylan-Kolchin, Small-Scale Chal- lenges to the CDM Paradigm, ARA&A 55, 343 G. Efstathiou, Suppressing the formation of dwarf galax- ies via photoionization, MNRAS 256, 43P (1992). [14] M. Hoeft, G. Yepes, S. Gottl ober, and V. Springel, Dwarf galaxies in voids: suppressing star formation MNRAS 371, 401 (2006), [astro-ph]. [15] T. Okamoto, L. Gao, and T. Theuns, Mass loss of galax- ies due to an ultraviolet background, MNRAS 390, 920 (2008), arXiv:0806.0378 [astro-ph]. [16] Y. Noh and M. McQuinn, A physical understandingof how reionization suppresses accretion on to dwarf haloes, MNRAS 444, 503 (2014), arXiv:1401.0737 B. Paczynski, Gravitational Microlensing by the Galactic Halo, ApJ 304, 1 (1986). [18] C. Alcock, R. A. Allsman, D. R. Alves, T. S. Axelrod, A. C. Becker, D. P. Bennett, K. H. Cook, N. Dalal, A. J. Drake, K. C. Freeman, M. Geha, K. Griest, M. J. Lehner, S. L. Marshall, D. Minniti, C. A. Nelson, B. A. Peterson, P. Popowski, M. R. Pratt, P. J. Quinn, C. W. Stubbs, W. Sutherland, A. B. Tomaney, T. Vandehei, and D. L. Welch, MACHO Project Limits on Black Hole Dark Mat- ter in the 1-30 M solar Range, ApJL 550, L169 [astro-ph]. [19] P. Tisserand, L. Le Guillou, C. Afonso, J. N. Al- bert, J. Andersen, R. Ansari, E. Aubourg, P. Bareyre, J. P. Beaulieu, X. Charlot, C. Coutures, R. Ferlet, P. Fouqu e, J. F. Glicenstein, B. Goldman, A. Gould, D. Gra, M. Gros, J. Haissinski, C. Hamadache, J. de Kat, T. Lasserre, E. Lesquoy, C. Loup, C. Magneville, J. B. Marquette, E. Maurice, A. Maury, A. Milsztajn, M. Moniez, N. O. Perdereau, Y. R. Rahal, J. Rich, M. Spiro, A. Vidal-Madjar, L. Vi- groux, S. Zylberajch, and EROS-2 Collaboration, Lim- its on the Macho content of the Galactic Halo from the EROS-2 Survey of the Magellanic Clouds, A&A 469, 387 (2007), [astro-ph]. [20] A. Bonaca, D. W. Hogg, A. M. Price-Whelan, and C. Conroy, The Spur and the Gap in GD-1: for a Dark Substructure in the Milky Way Halo, ApJ880, 38 (2019), arXiv:1811.03631 A. M. Nierenberg, T. Treu, G. Brammer, A. H. G. Pe- ter, C. D. Fassnacht, C. R. Keeton, C. S. Kochanek, K. B. Schmidt, D. Sluse, and S. A. Wright, Probing dark matter substructure in the gravitational lens HE 0435- 1223 with the WFC3 grism, MNRAS 471, 2224 Y. D. Hezaveh, N. Dalal, D. P. Marrone, Y.-Y. Mao, W. Morningstar, D. Wen, R. D. Blandford, J. E. Carl- strom, C. D. Fassnacht, G. P. Holder, A. Kemball, P. J. Marshall, N. Murray, L. Perreault Levasseur, J. D. Vieira, and R. H. Wechsler, Detection of Lensing Sub- structure Using ALMA Observations of the Dusty Galaxy SDP.81, ApJ 823, 37 (2016), arXiv:1601.01388 B. Ostdiek, A. Diaz Rivero, and C. Dvorkin, Ex- tracting the Subhalo Mass Function from Strong Lens Images with Image Segmentation, arXiv e-prints , arXiv:2009.06639 (2020), arXiv:2009.06639 S. Alexander, S. Gleyzer, H. Parul, P. Reddy, M. W. Toomey, E. Usai, and R. Von Klar, Decoding Dark Matter Substructure without Supervision, arXiv e-prints , arXiv:2008.12731 (2020), arXiv:2008.12731 A. L. Erickcek and N. M. Law, Astrometric Microlensing by Local Dark Matter Subhalos, ApJ 729, 49 F. Li, A. L. Erickcek, and N. M. Law, A new probe of the small-scale primordial power spectrum: by ultracompact minihalos, Phys. Rev. D11 86, 043519 (2012), arXiv:1202.1284 K. Van Tilburg, A.-M. Taki, and N. Weiner, Halometry from astrometry, J. Cosmology Astropart. Phys. 2018 , 041 (2018), arXiv:1804.01991 S. Mishra-Sharma, K. Van Tilburg, and N. Weiner, The Power of Halometry, arXiv e-prints , arXiv:2003.02264 K. Vattis, M. W. Toomey, and S. M. learning the astrometric signature of dark arXiv e-prints , arXiv:2008.11577 S. Mishra-Sharma, Inferring dark matter substruc- ture with astrometric lensing beyond the power spectrum, arXiv e-prints , arXiv:2110.01620 D. Spergel, N. Gehrels, J. Breckinridge, M. Donahue, A. Dressler, B. S. Gaudi, T. Greene, O. Guyon, C. Hi- rata, J. Kalirai, N. J. Kasdin, W. Moos, S. Perlmut- ter, M. Postman, B. Rauscher, J. Rhodes, Y. Wang, D. Weinberg, J. Centrella, W. Traub, C. Baltay, J. Col- bert, D. Bennett, A. Kiessling, B. Macintosh, J. Merten, M. Mortonson, M. Penny, E. Rozo, D. Savransky, K. Stapelfeldt, Y. Zu, C. Baker, E. Cheng, D. Content, J. Dooley, M. Foote, R. Goullioud, K. Grady, C. Jack- son, J. Kruk, M. Levine, M. Melton, C. Peddie, J. Rua, and S. Shaklan, WFIRST-2.4: What Every Know, arXiv e-prints , arXiv:1305.5425 R. Akeson, L. Armus, E. Bachelet, V. Bailey, L. Bar- tusek, A. Bellini, D. Benford, D. Bennett, A. Bhat- tacharya, R. Bohlin, M. Boyer, V. Bozza, G. Bryden, S. Calchi Novati, K. Carpenter, S. Casertano, A. Choi, D. Content, P. Dayal, A. Dressler, O. Dor e, S. M. Fall, X. Fan, X. Fang, A. Filippenko, S. Finkelstein, R. Foley, S. Furlanetto, J. Kalirai, B. S. Gaudi, K. Gilbert, J. Gi- rard, K. Grady, J. Greene, P. Guhathakurta, C. Hein- rich, S. Hemmati, D. Hendel, C. Henderson, T. Hen- ning, C. Hirata, S. Ho, E. Hu, A. Hutter, R. Jansen, S. Jha, S. Johnson, D. Jones, J. Kasdin, P. Kelly, R. Kir- shner, A. Koekemoer, J. Kruk, N. Lewis, B. Macin- tosh, P. Madau, S. Malhotra, K. Mandel, E. Massara, D. Masters, J. McEnery, K. McQuinn, P. Melchior, M. Melton, B. Mennesson, M. Peeples, M. Penny, S. Perl- mutter, A. Pisani, A. Plazas, R. Poleski, M. Postman, C. Ranc, B. Rauscher, A. Rest, A. Roberge, B. Robert- son, S. Rodney, J. Rhoads, J. Rhodes, J. Ryan, Rus- sell, K. Sahu, D. Sand, D. Scolnic, A. Seth, Y. Shvartz- vald, K. Siellez, A. Smith, D. Spergel, K. Stassun, R. Street, L.-G. Strolger, A. Szalay, J. Trauger, M. A. Troxel, M. Turnbull, R. van der Marel, A. von der Lin- den, Y. Wang, D. Weinberg, B. Williams, R. Wind- horst, E. Wollack, H.-Y. Wu, J. Yee, and N. Zimmerman, The Wide Field Infrared Survey Telescope: 100 Hubbles for the 2020s, arXiv e-prints , arXiv:1902.05569 B. S. Gaudi, R. Akeson, J. Anderson, E. Bachelet, D. P. Bennett, A. Bhattacharya, V. Bozza, S. Calchi Novati, C. B. Henderson, S. A. Johnson, J. Kruk, J. R. Lu, S. Mao, B. T. Montet, D. M. Nataf, M. T. Penny, R. Poleski, C. Ranc, K. Sahu, Y. Shvartzvald, D. N. Spergel, D. Suzuki, K. G. Stassun, and R. A. Science with the WFIRST Microlensing Sur- vey, BAAS 51, 211 (2019), arXiv:1903.08986 M. T. Penny, B. S. Gaudi, E. Kerins, N. J. Rattenbury, S. Mao, A. C. Robin, and S. Calchi Novati, Predictions of the WFIRST Microlensing Survey. I. Bound Planet Detection Rates, ApJS 241, 3 (2019), WFIRST Astrometry Working Group, R. E. Sander- son, A. Bellini, S. Casertano, J. R. Lu, P. Melchior, M. Libralato, D. Bennett, M. Shao, J. Rhodes, S. T. Sohn, S. Malhotra, S. Gaudi, S. M. Fall, E. Nelan, P. Guhathakurta, J. Anderson, and S. Ho, Astrometry with the Wide-Field Infrared Space Telescope, Journal of Astronomical Telescopes, Instruments, and Systems 5, 044005 (2019). [36] J. F. Navarro, C. S. Frenk, and S. D. M. White, A Uni- versal Density Prole from Hierarchical Clustering, ApJ 490, 493 (1997), [astro-ph]. [37] C. Mondino, A.-M. Taki, K. Van Tilburg, and N. Weiner, First Results on Dark Matter Substructure from As- trometric Weak Lensing, Phys. Rev. Lett. 125, 111101 (2020), arXiv:2002.01938 J. Philip, The Probability Distribution of the Distance Between Two Points in a Box, TRITA MAT 7(2007). [39] T. D. Brandt, Constraints on MACHO Dark Matter from Compact Stellar Systems in Ultra-faint Dwarf Galaxies, ApJL 824, L31 (2016), arXiv:1605.03665 S. A. Johnson, M. Penny, B. S. Gaudi, E. Kerins, N. J. Rattenbury, A. C. Robin, S. Calchi Novati, and C. B. Henderson, Predictions of the Nancy Grace Ro- man Space Telescope Galactic Exoplanet Survey. II. Free- oating Planet Detection Rates, AJ 160, 123 A. Molin e, M. A. S anchez-Conde, S. Palomares-Ruiz, and F. Prada, Characterization of subhalo structural prop- erties and implications for dark matter annihilation sig- nals, MNRAS 466, 4974 (2017), arXiv:1603.04057 G. Duch^ ene and A. Kraus, Stellar Multiplicity, ARA&A 51, 269 (2013), arXiv:1303.3028 S. Garrison-Kimmel, A. Wetzel, J. S. Bullock, P. F. Hopkins, M. Boylan-Kolchin, C.-A. Faucher-Gigu ere, D. Kere s, E. Quataert, R. E. Sanderson, A. S. Graus, and T. Kelley, Not so lumpy after all: modelling the deple- tion of dark matter subhaloes by Milky Way-like galax- ies, MNRAS 471, 1709 (2017), arXiv:1701.03792 S. B. Green, F. C. van den Bosch, and F. Jiang, The tidal evolution of dark matter substructure - II. The impact of articial disruption on subhalo mass functions and radial proles, MNRAS 503, 4075 (2021), E. O. Nadler, A. Banerjee, S. Adhikari, Y.-Y. Mao, and R. H. Wechsler, The Eects of Dark Matter and Baryonic Physics on the Milky Way Subhalo Population in the Presence of the Large Magellanic Cloud, ApJL 920, L11 (2021), arXiv:2109.12120 J. J. Dalcanton, B. F. Williams, D. Lang, T. R. Lauer, J. S. Kalirai, A. C. Seth, A. Dolphin, P. Roseneld, D. R. Weisz, E. F. Bell, L. C. Bianchi, M. L. Boyer, N. Cald- well, H. Dong, C. E. Dorman, K. M. Gilbert, L. Girardi, S. M. Gogarten, K. D. Gordon, P. Guhathakurta, P. W. Hodge, J. A. Holtzman, L. C. Johnson, S. S. Larsen, A. Lewis, J. L. Melbourne, K. A. G. Olsen, H.-W. Rix, K. Rosema, A. Saha, A. Sarajedini, E. D. Skillman,

2108.10886

Kris Pardo

Kris Pardo and Olivier Dor\'e

Detecting dark matter subhalos with the Nancy Grace Roman Space Telescope

11 pages, 6 figures. v2: small text changes. accepted for publication in PRD

10.1103/PhysRevD.104.103531

astro-ph.CO

http://creativecommons.org/licenses/by/4.0/

The dark matter subhalo mass function is a promising way of dark matter models. While cold dark matter predicts halos down to Earth-sized masses, other dark matter models typically predict a cut-off in the subhalo mass function. Thus a definitive detection or limits on the existence of subhalos at small masses can give us insight into the nature of dark matter. If these subhalos exist in the Milky Way, they would produce weak such as modified apparent positions, on background stars. These signatures would generate correlations in the apparent velocities of these stars, which could be observable given resolution and cadence. The Nancy Grace Roman Space Microlensing Survey will be perfectly suited to measuring the acceleration signatures of these halos. Here we forward model signatures and explore the Roman Space Telescope's on lens profiles and populations. We find that the Roman Space Telescope could place competitive bounds on point source, Gaussian, (NFW) profile lenses that are complementary to other proposed methods. In particular, it could place 95% upper limits on the $c_{200} < 10^{2.5}$. We discuss possible systematic effects that could hinder these efforts, but argue they should not prevent the Roman Space Telescope from placing strong limits. We also discuss further analysis methods for improving these constraints.

Contents 1 Introduction 2 2 Dynamics of primordial tensors and gauge elds during in ation 5 3 of in ationary tensor perturbation with primordial gauge elds 8 3.1 A complete calculation using in-in formalism . . . . . . . . . . . . . . . . . . . . . . . 8 3.1.1h AAi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 3.1.2h BBi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.1.3h EEi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 3.2 The magnetic and electric non-linearity parameters . . . . . . . . . . . . . . . . . . . . 14 3.3 The squeezed limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 4 Semi-classical derivation of the consistency relations 17 5 A novel correlation of tensor and curvature perturbations 20 6 Conclusions and discussions 24 A Useful integrals 25 B Validity of the semi-classical approach 27 C The in-in formalism and the dynamical correction term 28 1 Introduction In eective eld theories with extra dimensions, as derived from string theory, it is natural to expect that gauge elds are non-minimally coupled in the early universe through a dynamical dilaton eld or the moduli of the internal dimensions. This breaks the conformal invariance of the gauge eld action and enables the amplication of the gauge elds during in ation. If gauge elds are enhanced in the early universe, during the in ationary epoch, it opens the possibility that such a primordial gauge eld could have left a detectable imprint on the observable universe today. It is well known that gravitational waves may also be created during in ation by the same causal mechanism, and we therefore expect a non-vanishing non-Gaussian between primordial and primordial gauge elds, which we will calculate here. Focussing for simplicity on abelian gauge elds, we will consider a model where conformal invariance is broken by a non-minimal coupling between an evolving scalar eld and the kinetic term of the gauge eld [1,2], which takes the form Lgauge =1 4()FF: (1) Such models are often referred to as the kinetic coupling models. There could also be a with the parity violating term of the gauge eld. However, we shall not consider those models in this paper. In all these scenarios of broken conformal symmetry, one also assumes that the conformal invariance is restored by the end of in ation (or by the end of the reheating such a model only breaks the conformal invariance, and not the gauge invariance. 2A further motivation for studying this type of models stems from the fact that astrophysical obser- vations indicate the presence of coherent magnetic elds of femto-Gauss strength on large (Mpc scales or larger) exceeding the scales of galaxies and galaxy clusters [3{5]. It is unclear if any astrophysical process can produce magnetic elds with coherence lengths of cosmological sizes, and it has therefore been speculated that such elds must have an in ationary origin. After all, in ation is known to be the perfect mechanism for producing the very large, even of primordial uctuations, needed to explain the observed large scale correlations in the cosmic microwave background (CMB) in ationary magnetogenesis to work, it requires a breaking of the conformal invariance during in ation, and therefore imply a departure from standard in eq. (1) [1,2,6]. But even then, back-reaction arising due to the generated electromagnetic elds and issues with entering regimes of strong coupling severely restricts the possibility for elds to be generated during in ation [7]. Only in some very special cases, it is possible for the observed magnetic elds to have been generated during in ation [8{13]. It would therefore be very important to understand if magnetic elds observed on cosmological scales really have an or not. One possibility is to look at the correlation of the cosmological magnetic elds with the or equivalently the gauge-invariant curvature perturbation. Such correlations have ear- lier been discussed in [14{18], and also in [19,20] in a dierent set-up. The idea being that magnetic elds, if they are indeed generated during in ation, will be correlated with the primordial On the other hand, if they are generated after in ation, no such correlation will exist. Assuming an exact model for generating the magnetic elds during in ation, one can compute the expected correlation. Since one of the few models, which can indeed generate signicant elds during in ation, takes the general form in eq. (1), much attention was focussed on this model in the previous work, and in [16, 17] the non-Gaussian correlation function of the with the magnetic (2) was calculated. It was shown that such a non-Gaussian correlation function satises a new in the squeezed limit and its magnitude becomes quite large in the attened conguration in the Fourier space. Later, some consistency relations for the soft limit of the higher order magnetic elds and matter over-densities were proposed and it was pointed out that any violation of such consistency relations would point towards an in ationary origin of elds [21]. Here we will consider the natural extension of that work, and look at the non-trivial cross corre- lations of primordial gauge elds and primordial gravitational waves. Our main objective is therefore to compute the primordial non-Gaussian correlations (3) where is the tensor mode, Ais our abelian gauge boson (vector potential) and B,Eare the associated magnetic and electric elds, respectively. Since these are equal time correlation functions, we will calculate them using the full in-in formalism which is widely used to calculate such higher order correlators during in ation [22]. As a way of checking our nal results, we propose new semi-classical consistency relations in the limit wherein the wavelength of the graviton goes to innity (the squeezed limit). consistency relation for correlation functions only involving the curvature perturbation, , or the tensor mode, , was rst proposed by Maldacena in [34, 35] (and also subsequently by others [36, 37]), hence they are also sometimes called the Maldacena consistency relations. A relation for gauge elds was rst discussed in [16], wherein a magnetic consistency relation for the correlation function hB
Biwas proposed. Here we will consider the natural extension to the correlation functions above in eq. (3), and show that the full in-in correlation function in fact satises the semi-classical consistency relation in the appropriate limits5. One of the problems with generating large magnetic elds during in ation, is that the universe becomes an almost perfect conductor after reheating, which very rapidly erases any magnetic eld which would have been generated during in ation. Models of successful magnetogenesis require a very low scale of in ation [11,12], or at least a very low reheating temperature [13]. This problem is absent if the magnetic eld belongs to a dark gauge group. Therefore we can more easily imagine that large dark magnetic elds could have been generated during in ation. While such dark magnetic elds are not directly observable in the same way as the standard magnetic elds , they do leave indirect imprints in the curvature perturbation of the universe. The total curvature perturbations is then the sum of the usual contribution, 0generated independently of the magnetic eld, and the contribution, B, imprinted by magnetic elds ()'0() +B(); (4) where the curvature perturbation induced by any magnetic eld is [40] B() =Z 0dln (5) This expression is valid for both dark and ordinary magnetic elds. In the remainder of this paper, it can refer to both cases. The magnetic elds generated during in ation follow a Gaussian statistics to the leading order in perturbations which implies that the induced curvature perturbation Bis generally a non-Gaussian eld. Various imprints of this induced component such as anisotropies in the power spectrum and bispectrum have also been greatly studied [41{48]. One may naively expect that due to the induced curvature perturbations, there might exist a correlation of the primordial tensor mode with the primordial curvature perturbation of the (6) due to the presence of the non-vanishing correlator h B
Bi. A similar contribution will arise from the corresponding electric elds as well. As we shall show in section 5, while h Bivanishes in the 4Since the semi-classical relations relate a higher order correlation function involving an infrared (soft mode) to a lower order correlation function, the semi-classical relations are also sometimes called the soft theorems with a reference to the Weinberg's soft photon and graviton theorems [23], which can be related to the asymptotic symmetries and memory of at space [24,25]. For a discussion of the relation between semi-classical relations/soft symmetries and gravitational memory in in ation, see [26{28] and also [29,29{33]. 5The semi-classical relations serve an important role as a cross check of the correctness of the full in-in (not only Maldacena's three-point function). As was rst shown in [38], it can also be used to calculate the contribution to the four-point function in the limit where the momentum of the exchanged graviton goes to zero, the so called counter collinear limit, by simply taking the square of the three-point function in the squeezed limit. In [39], it was further realized that one can also extract the infrared divergent part of higher order loop diagrams using limit6, such a two point correlation function can actually be non-vanishing in the created by the long wavelength gauge elds, and may even receive contributions of quantum gravity. This paper is organized as follows. In the next section we will quickly review the dynamics of during in ation and the kinetic coupling model as a mechanism for the production of large scale primordial gauge elds during in ation. In section 3, We calculate the full with the cross correlations of the in ationary tensor perturbation with gauge elds using the very general in-in formalism and discuss the extent of non-linearities using the Fourier space shape functions. In section 4, we derive the consistency relations for these non-Gaussian cross a semi-classical approach. In section 5, we calculate a direct correlation of the mode and the primordial curvature perturbation induced by the gauge elds in the of long wavelength modes. Finally, in section 6, we summarise our results and conclude with a discussion. In appendix A, we have listed some useful integrals appearing in the in-in results of various cross correlations. In appendix B, we discuss the validity of the semi-classical approach used to derive our consistency relations and in appendix C, we discuss the presence of the terms arising in the cross correlation involving electric elds in the in-in this paper, we work in natural units with ~=c= 1, and the Planck mass M2 Pl= 1=8G is set to unity. Our metric convention is ( ;+;+;+). 2 Dynamics of primordial tensors and gauge elds during in ation In this section, we shall brie y discuss the dynamical evolution of primordial tensor modes and abelian gauge elds during in ation. The homogeneous and isotropic background during the is described by the spatially at FLRW (7) whereis the conformal time, dened by d=dt=a anda() is the scale factor. The perturbed FLRW metric in the presence of tensor modes is given (8) where ijis the metric tensor perturbation which is transverse and traceless i.e. ij;j= 0 and ii= 0. These conditions lead to only two independent radiative degrees of freedom of tensor correspond to the two polarizations of gravitational waves (GW). To the linear order in there is typically no active source of GWs during in ation. However, quantum uc- tuations of ijare parametrically amplied by the expansion of the universe. In order to describe this phenomenon, one needs to quantize the canonical degrees of freedom associated with the tensor perturbations. Following the usual quantization formalism, the corresponding mode expansion of tensor perturbation is dened as ij(x;) (9) 6Some scaling arguments of such a mixed correlator based on the special conformal transformations have been dis- cussed in [49] wherein there is an explicit parity violating term in the the polarization tensor corresponding to the helicity s, with the normalization s ijs0 ij= 2ss0 and the creation and annihilation operators do satisfy the condition, [ bs k;bs0y k0] = tensor helicity mode s(k;) can be dened through the following (10) We can now dene the two point function in Fourier space (11) where the power spectrum of tensor modes is given by P (k;) =j k()j2. It is often useful to dene a dimensionless tensor power spectrum P as P (k;) =k3 22P (k;) =k3 22j k()j2; (12) which is usually evaluated in the super-horizon limit jkj! 0. For a nearly de-Sitter whereHis the Hubble parameter, the Bunch-Davies normalized solution for the is given by k() =Hp k3(1 +ik)eik; (13) which, in the super-horizon limit, leads to P (H=2)2. The tensor power spectrum thus measures the energy scale during in ation and is exactly scale invariant in a de-Sitter background (and nearly scale invariant in a slow roll in ation). As mentioned already in the introduction, it is well known that gauge elds described by the canonical Maxwell's action are not amplied in the FLRW background due to the conformal symmetry and thus, a necessary condition for amplication of gauge elds during in ation is to break this conformal invariance. One of the simplest possibility which has been explored to a large extent in the literature is the so called kinetic coupling model which breaks the conformal invariance gauge invariance) by explicitly coupling gauge elds to the in aton as (14) where the gauge eld tensor is dened as F=@A@Aand the second term represents the interaction where jis the four current density. For simplicity, we shall neglect the interaction term and assume that there are no free charges. In the Coulomb gauge wherein A0= 0 and@iAi= 0, the quadratic action for the gauge eld (15) Since we are interested in the dynamical amplication of the quantum gauge eld originating from the vacuum, we promote the classical gauge eld Aito an operator and impose the canonical Following this, we dene the usual mode expansion as Ai(x;) (16) 6and impose the standard commutation = (17) HereAk() is the mode function in Fourier space and the polarization vectors will satisfy the usual relations: k
(^k) = =0, can now dene a canonically normalized gauge eld =(()). In this case, the quadratic action for the gauge eld takes the canonical form (18) and the equation of motion for the rescaled mode function, vk=S()Ak, takes the form of a with a time dependent mass term 0: (19) In order to compute the two point and higher order correlators, one needs to dene the For instance, the dilatonic coupling corresponds to ()exp(2) [2, 50]. Since  must be dynamical, it turns out to be more convenient to parametrize the time dependence of the coupling function as (()) the conformal time at the end of in ation, and Iis the coupling evaluated at the end of in ation =I. For a power law coupling () =I(=I)2n, one nds that S00=S= 1=2) and the mode solution can be written in terms of the Hankel functions. In the the mode solution must be normalized to the Bunch-Davies vacuum vk() =1p 2keik; (20) which leads to the full solution for vkas vk() (21) With this solution, the mode function Ak() is expressed as Ak() is a Hankel function of the rst kind. The two point correlation function of gauge elds Aican be obtained as (23) With respect to an observer characterised by the 4-velocity vector u, one can covariantly dene the electric eld Eand magnetic eld Bas [51]. E=Fu; B=Fu; the dual of the electromagnetic eld tensor Fwhich is dened by the (25) with being the totally antisymmetric permutation tensor of space-time with 0123=pg. In this work, the covariant magnetic and electric elds are dened with respect to a comoving observer having 4-velocity u= (1 a;0;0;0). The power spectra of the magnetic eld and electric eld are (27) Note that, both PBandPEare dimension-full. We can now easily express PBandPEin terms of the mode function Ak() as PB(k;) = can be readily calculated using the obtained mode solution (22) for the case of a power law coupling [14,17,50,51]. For the cases of n= 2 andn=3, the magnetic eld spectrum turns out to be scale invariant i.e. PB/k3while the electric eld spectrum behave as In both these cases, there arise various issues such as the strong coupling and the back reaction problem which have been discussed to a great extent in the literature of in ationary tensor perturbation with primor- dial gauge elds In this section, we shall discuss our calculations of the cosmological s of perturbation with gauge elds using the in-in formalism while in the next section, we shall discuss the semiclassical approach to arrive at these correlators. As expected, both these approaches lead to the same results for these s in the squeezed limit. 3.1 A complete calculation using in-in formalism In order to compute the correlation function during in ation, we adopt a very powerful tool of the in-in formalism which can be used to calculate various correlation functions at equal times. A of the in-in formalism can be found in appendix C. In this formalism, the expectation value of an operatorOat timeIis given int j0i (30) whereHintis the interaction Hamiltonian and T, T are the time ordering and anti-time respectively. The quantum state j0irefers to the physical vacuum state of the theory which we take to be the standard Bunch-Davies vacuum state. We shall rst apply the in-in formalism to calculate the s of the tensor mode with gauge elds. 8Figure 1: The leading order cosmological diagram contributing to all equal time Is of our interest. In addition to the contribution from this diagram, there will also be additional corrections for each correlator, as discussed in the text. For there is a term. But for the case of h EEi, there is not only a kinematical correction term but also a dynamical correction us rst compute the of a tensor mode with two gauge eld modes, i.e., a correlator of the formh AAi. We rst nd (31) where all the repeated indices in the above and following expressions are summed over (irrespective of up or down). By using the master formula of the in-in formalism (30), we can calculate which, to the leading order, is given the cubic order interaction Hamiltonian which can be calculated from equation (14), Hint() + (33) By substituting (31) in (32), we observe that the rst term on the right hand side (RHS) of equation (32) gives a non-zero contribution. It can be easily calculated from the Fourier mode functions (13) and (22) is the power spectrum of the tensor modes, as dened in (11) and PA(k;) is the power spectrum of the gauge elds dened (35) Thus, the vacuum expectation value of AAdoes contribute to this non-Gaussian correlator at the leading order. Since this contribution is originating purely from the kinematical using the dynamics of gauge elds, we refer to such contributions as the kinematical correction terms . As we shall see later, similar kinematical terms will also appear for the cases of the second term in the RHS of equation (32) is calculated using the cosmological diagram- matic rules [22]. The leading order diagram contributing to this correlator is shown in gure 1. Substituting the interaction Hamiltonian Hintfrom equation (33) in equation (32) and after we obtain the total contribution (36) where the integrals I1andI2can be expressed in terms of the mode functions of tensor mode and gauge eld and their time derivatives as I1= 2 (37) I2= 2 (38) and the explicit time dependence, I, of the operators and the power spectra in the equation (36) is omitted for notational convenience. It is useful to rewrite these integrals in (37) and (38) in terms of another set of auxiliary integrals, denoted by as andA(0) k(I) are the asymptotic super-horizon expansion of the mode functions (13) and (22), given the explicit forms of the integrals to both n >1 2andn <1 2are given in Appendix A. One can substitute (37) and (38) into (36) and after considerable nal result for the correlator h AAican be written (45) Here the subscript ` A' inMAandWArefers to the gauge eld A. Note that, in the nal result (43), the rst contribution arises from the interaction term corresponding to the leading order in gure 1 while the second term is the pure kinematical correction term which is independent of the coupling function this section, we will calculate the correlator h BBi. The covariant magnetic eld Bis dened with respect to a comoving observer as in (24). Then BBcan be expressed in terms of gauge eld tensor as (46) One can evaluate the correlator by substituting (46) and (33) into discussed in the case of the AAcorrelator, the rst term in the RHS of (47) leads to a non zero contribution to the vacuum expectation value of BB. This kinematical correction term is calculated using (46) to yield the following (48) where ~k2=k2+1 2k1. Once again, we notice that the vacuum expectation value of the leading order to the non-Gaussian correlator. However, we would like to emphasise that correction terms were not explored in earlier works [17,52]. 11The leading order contribution of the second term in the RHS of (47) is calculated using the cosmological diagrammatic rules [22]. Then one would obtain the total contribution the same integrals, as given in (37) and (38). By inserting (39) and (40) into (49), the nal result for the correlation function (52) These contributions borrow similar structure as derived earlier in the case of the correlator hB
Bi in [16, 17, 52]. Note the presence of the two non-trivial integrals which can be exactly calculated for various values of n. However, the second term in (50) is the pure kinematical correction term which was not pointed out in earlier works now turn to the calculation of the correlator h EEi, i.e., non-Gaussian cross correlation of tensor mode with the electric elds which is a bit more subtle than the other correlators. Using equation (24), one can express the covariant electric eld in terms of gauge elds as E(x;) =d dA(x;) +i[Hint;A]; (53) where the extra commutator i[Hint;A] arises due to the fact that the operators are all to be in the interaction picture, as discussed in Appendix C. With this denition, expressed as + (54) Again, using the master formula of the in-in formalism (30), we nd discussed earlier, the rst term in the RHS of the above equation produces a non zero contribution to the vacuum expectation value of EEdue to the second and the third term in the RHS of equation (54). The contributions arising from the second term can be referred as the dynamical correction term and the third term as the kinematical correction term, respectively. In the super-horizon limit, they can be calculated as (56) Unlike previous sections, the above vacuum expectation value is not only contributed by a term but also a dynamical correction term. Finally, the leading order contribution arising from the second term in the RHS of (55) is calculated using the cosmological diagrammatic rules [22] and the nal result (57) wherein the two integrals J1andJ2are dened in terms of the mode functions of tensor mode and gauge eld and their time derivatives as J1= 2 (58) J2= 2 (59) Again, it becomes useful to dene the two auxiliary integrals ~J(1) nand withJ1and J2, in a similar manner as earlier, andA0(0) k(I) are the asymptotic super-horizon values of the mode functions of the tensor mode and the time derivative of mode functions of the gauge eld respectively. In the to the end of in ation jkIj!0, we nd that ~J(2) n'~I(2) n, as shown in appendix A. Note that, a few more terms appear in the result for the correlator h EEias the covariant electric eld is dened in terms of the time derivative of the gauge 13eld while the covariant magnetic eld only contains spatial derivatives of the gauge eld. The nal result can now be written in a compact form =j k1(I)j2 (66) Note that, the above result has been obtained in the super-horizon limit. This correlator has a ash BBi, albeit an extra contribution which arises precisely due to the denition of the covariant electric eld in equation (53). We refer to this contribution as the dynamical correction term to the correlator. Moreover, the last term in equation (62) is similar to the kinematical correction term as appeared earlier for h BBiin equation (50). As a result, the associated non-linearity constant in the squeezed limit, as we shall discuss in the following section. 3.2 The magnetic and electric non-linearity the approach of our previous papers [16,17] for the case of a of the perturbation with primordial gauge elds, we can similarly dene convenient and useful non- linearity parameters to characterise the s of the in ationary tensor perturbation with primordial gauge elds. Let us rst dene the bispectra associated with   (68) where the bispectra only depend on the magnitude of the wavevectors due to spatial isotropy of the background expansion. The strength of the magnetic and electric eld bispectra can be characterised by dening the non-linearity parameters b NLande NLas +PB(k3) ; +PE(k3) ; the power spectra of the tensor perturbations (11), magnetic eld (28) and electric eld (29), respectively and the RHS of these denitions are written in a symmetric manner. 14If the two non-linearity parameters b NLande NLare momentum independent, they correspond to a local shape of the bispectra in the Fourier space which can be obtained from the relations (72) where BG,EGand Gare the Gaussian random elds. In the absence of any parity violating inter- actions, as is the case in our scenario, the Gaussian tensor mode Grepresents both the of freedom associated with the graviton. Similar denitions also allow one to compute the higher order cross correlations of curvature perturbations with gauge elds [40]. For the case of the we can now simply compare the equations (67) and (69) with (50) and read o the bispectrum which leads to the following result for the non-linearity 1=2>0 ~k12n 2 k12n 2 +k12n 3!# ; (73) where, once again, the rst term in the above result is due to the interaction Hamiltonian and the second term is the pure kinematical correction term. Similarly, on comparing (68) and (70) with (57), yields the result for the non-linearity parameter e NLforn1=2>0 2E(k1;k2;k3) +k2 ~k22 ~k32n 2 k32n 2 general expressions for b NLande NLare one of the main results of this paper and they depend only on the wave numbers k1,k2andk3. The presence of (3)(k1+k2+k3) in various s constrains the wave vectors to form a closed triangle in the Fourier space. In order to understand the extent of these non-linearities, it is useful to study b NLande NLfor dierent shapes of the Conventionally, such shape functions are studied by dening the dimensionless set at an arbitrary scale. For an isotropic background, the shape function only depends on any of the two momentum ratios. This allows us to rst calculate the following two expressions as ~k2 k2!2 =1 2 1 +x2 3x2 1 2 ; (76) In order to calculate b NLande NL, we also need to evaluate the integrals dierent values ofnwhich has already been done in [17]. For the most interesting case of n= 2 which leads to a scale invariant spectrum of magnetic elds, we present the results for Appendix A and use them to plot the extent of these non-linearity parameters as contour plots in the momentum space in gure 2. There are a few interesting points to note about the plots. First, both small in the squeezed limit which is characterised by the top left corner, i.e., x1!0;x3!1. 150.0 0.5 1.0 1.5 0.5 1.0 1.5 75 125 175 225 20 40 60 80 100 120Figure 2: The extent of the non-linearity parameters b NL(left) ande NL(right) corresponding to dierent triangular conguration are plotted for the case of n= 2. The colour legends representing the magnitude are also shown below each plot. While the top left corner of these plots corresponds to the squeezed limit, i.e., ( x1;x3) = (0;1), the top right corner represents the attened shape, i.e., (x1;x3) = (2;1). In the squeezed limit, b NLis relatively small while it is large in the nearly attened shape. However, e NLremains small both in the exact squeezed and attened limit but it becomes large asx3becomes smaller. Close to the attened limit dened by large but e NLbecomes very small in the exact attened limit. However, e NLbecomes large in a somewhat dierent limit x1!1;x3!0 in small. Note that, the allowed region in the space of x1x3plane is constrained by the triangle inequality in the Fourier space. 3.3 The squeezed limit Let us now consider the squeezed limit wherein the wavelength of the graviton mode is much larger than the corresponding wavelengths of the electromagnetic elds. Thus, a squeezed shape a correlation of very large scale with very small scales. Using our general results of the from the previous sections, we can easily write down the results for all the correlators in the squeezed limit. However, as we shall discuss later, one can also obtain the squeezed limit non- Gaussianity by using semi-classical methods. In this limit, we have k1!0 and k2!k3k. The 16two integrals of interest the Bessel functions of the rst and second kind, respectively. These integrals to evaluate and for integer values of n, one can show the following analytical =n k2; (79) and evidently, In the squeezed limit, k1!0, so using this result in all the correlators that we have obtained earlier, and after some simplication, we nd the following above results comprise of a set of novel consistency relations for the cross correlations involving the tensor perturbation and gauge elds. Note that, in the absence of other source terms, there exists an electromagnetic duality which states that the equations of motion and the power spectra of the electric and magnetic elds can be obtained from each other under a simultaneous To our knowledge, it has not been known if this duality is also maintained at the level of higher order correlation functions beyond the power spectra. In our case, we observe that this duality is indeed preserved in the squeezed limit of the cross the electric and magnetic elds. In particular, the results in equation (81) can be obtained from equation (82) using the electromagnetic duality and vice versa. The results in both the regimes of can be correctly obtained from the corresponding results for the under the electromagnetic duality transformations. In any case, it will be interesting to examine whether the electromagnetic duality is maintained for the higher order correlation functions away from the squeezed limit. 4 Semi-classical derivation of the consistency relations In this section, we want to apply semi-classical techniques to derive the three-point correlation function of the the squeezed limit k1k2k3. 7Note that, this result also holds true for real non-integer values of nwhich can be veried would use the similar approach developed earlier to compute loop corrections to the primordial power spectrum during a de-Sitter/slow roll background [39,54]. To start with, the correlator h YYi in real space can be written (83) In the limit where one of the momenta is much smaller than the other two i.e. k1k2k3, the wavelength corresponding to the graviton is very large and is frozen to a constant amplitude B outside the horizon. As a result, the eect due to a long wavelength graviton mode Bon the of gauge elds can be understood by rescaling the background as Thus we can compute the two point correlation function hYYiin the modied background using the similar arguments of [39], by Taylor expanding around the in real space ; (84) and in the Fourier space x
~ q1ei~x
~ (85) Since the long wavelength mode Bis almost constant over the Hubble horizon, and in particular, over the scales of variation of the short wavelength modes, we can choose x0=x. We can now express the modied coordinates due to the long wavelength graviton ; (86) and using that momentum transforms inversely, ; (87) such that ~ x
~ q=x
q, (88) It is useful to note that det[exp( B)] = 1, so that the Jacobian of the transformation ~ q!qis =(3)(q1+q2), so we can write hY(~ q1)Y(~ q2)iB= (89) 18In the squeezed limit, due to the rescaled background by the long wavelength graviton mode, one can write a three point correlation function in terms of the modied two point function (90) Let us now justify that the arguments used to obtain the consistency relations might only be trusted for specic regions of n. To illustrate this, let's write the evolution equation for the gauge eld, i.e., equation (19) again 0: (91) This equation is identical to the evolution equation of a massless scalar eld in the FLRW the only dierence is that the gauge eld experiences an eective scale factor Swhich we refer to as the pump eld. For the massless scalar eld in de-Sitter with a1=, there exists an eective event horizon and therefore, the long and the short wavelength modes decouple on the This allows one to capture all the physical eects of the long wavelength mode by rescaling the background for the short wavelength modes to calculate the three point s involving long and short wavelength modes of the scalar eld uctuations or the curvature perturbations. can be applied for any other correlator in the squeezed limit as far as the are xed by the values of the background quantities at horizon crossing. If we inspect the mode function equation (91) above in the long wavelength limit k2S00=S, the solution can be found by simple integration to be given by vk()C1S() integration constants. It is clear that if S=p n, then for !0, the rst term will dominate for n >1=2, which means that the solution for Ak=vk=S will freeze to a constant in the long wavelength limit. A trivial rescaling of the eld is of course not important, what is important is that the solution is in a super-horizon attractor regime. If the solution is dominated by the attractor, then the rescaling of the momentum by a long mode will enter trivially as a rescaling of the initial normalization through C1. However, if the non-attractor is important, the rescaling of momentum will also enter through the lower limit of the time integral multiplying C2and thus modify the dynamics non-trivially. The existence of an attractor solution of the long wavelength modes leads to the constraint n>1=2. Therefore, the semi-classical derivation for h AAican only be trusted for n >1=2. A similar condition can also be obtained for the corresponding the graviton and magnetic elds. However, we nd that the properly rescaled electric eld experiences an eective scale factor,1p , which reverses the condition on nfor electric eld. As a result, we conclude that the semi-classical derivation for the graviton electric elds only be trusted for n<1=2. By substituting equation (89) into equation (90) and after further simplications, we will obtain the desired correlators in the squeezed limit. For n>1=2, we get the following consistency for n<1=2, we the squeezed limit, these consistency relations are quite general as they have been derived only by using the semi-classical techniques. These relations do exactly match with the squeezed limit results of the full in-in correlates, as obtained in (80), (81) and (82). From these correlators, one can now read o the non-linearity parameters n>1=2 n< 1=2: (97) It is important to mention here that these are indeed the local non-linearity parameters as the momentum independent in the squeezed limit, as evident from equation (76). Moreover, both the parameters in the squeezed limit are proportional to nand are of order unity as the discussion of our earlier paper [16]. This behaviour is also quite evident from the contour plots in gure 2 wherein the top left corner indicates the extent of these non-linearity parameters in the squeezed limit. 5 A novel correlation of tensor and curvature perturbations As mentioned in the introduction, an interesting consequence of these three point correlators involving the primordial tensor mode and gauge elds is the possibility of a novel two point correlation of the primordial tensor mode with the primordial curvature perturbation. As we shall show later, such a two point correlation function can actually be non-vanishing in the anisotropic background created by the long wavelength gauge elds, and may even receive contributions of the form h B
Bifrom higher order quantum gravity corrections. The presence of primordial gauge elds can leave indirect imprints on the spectrum of during in ation. This can be illustrated by splitting the total curvature the sual contribution, 0generated independently of the gauge eld, and the by magnetic and electric elds, () =0() +B() +E(); induced by the corresponding magnetic and electric elds derived from the gauge elds only. Now, we can express the two-point correlation function of the tensor mode with the curvature perturbation as (99) 20To this end, we are going to generically denote BandEbyY. Following [40], they can be written in the Fourier space as Y(k;I) (100) whereis the rst slow roll parameter, His the Hubble parameter, and Y=fE;Bg. We can now express the two-point correlator h Yiin terms of the three point function (101) To computeh i, we need to evaluate this three dimensional integral in Fourier space. However, note that, these three point correlators appearing in the integrand of above equation are not equal time correlation functions [40]. Thus, we rst have to convert them in terms of an equal time  is the Heaviside step function and B=( 42n; n>1=2 6 + 2n; n> 1=2 4 + 2n; n< 1=2(104) Here, we have assumed that only super-horizon modes of gauge elds are contributing non trivially to these integrals with functions stressing this fact and k0a0H0is the horizon crossing scale at the time 0which we assume to be the onset of the generation of magnetic elds. In general, the momentum integral appearing in equation (102) is somewhat non-trivial to evaluate. However, without loss of generality, one can analytically calculate these two-point correlators by rst expressing these three-point correlators (105) and using it in equation (102) to arrive can now choose a convenient coordinate system to evaluate the momentum integral. We choose a spherical polar coordinate system in such a way that its polar axis is along the direction of k2. This 21allows us to construct a triplet of orthonormal basis vectors with (s) ij(^k2). Thus, any arbitrary unit vector ^qin such a spherical polar coordinate system can be expressed as follows ^q= sincos^e1+ sinsin^e2+ cos^k2: (107) Using this expression and through a straightforward calculation, we can now obtain the (108) SinceFis independent of the angle and so are all the  functions, we can evaluate the q-integral and obtain (for brevity, we are not writing all the  function terms this integral identically vanishes because of the integral and as a result, the direct two-point correlation of the tensor mode with the induced curvature perturbation also vanishes 0: one can also show the above result by performing the angular integral using the the spin-weighted spherical harmonics and using the transverse and traceless condition of the polarization tensor, along the lines of the discussion presented in [52]. This does, however, not mean that we should necessarily expect the h icorrelator to vanish. It is worth noting that isotropy is broken by the long wavelength modes of the vector eld. As long wavelength modes are stretched to super-horizon scales, they act as a background for the shorter scales. Thus, if in ation lasts for a total number of e-folds, Ntot, the modes which left the horizon during the rst (NtotN) e-folds, they contribute as a classical background for the modes leaving the remain. These long wavelength modes make up a random vector pointing in some random direction, which locally breaks isotropy. The typical amplitude in a given local realization is given by the square root of the variance of the long wavelength uctuations [44,55]. Let us, that in the case of in ationary magnetogenesis, a background magnetic eld is turned on in the x-direction. This is not completely general, as we already chose our coordinate system to be aligned withk, but it serves to show the point we want to make. In any case, we will Ei=Ei: (111) However, the discussion proceeds very similar for the case of the electric eld being amplied on super-horizon scales. We will turn our attention to two interesting contributions to the function. This rst comes from the induced direct coupling between and, while the second is a contribution arising from a quantum gravity induced higher dimensional operator such that the eective action is written (112) withMbeing the scale of quantum gravity. The rst contribution comes from expanding the coupling () in perturbations of the scalar eld, using the comoving gauge ; (113) 22so we nd the interaction (114) where we have dened () =@ @0() H: (115) Using the cosmological diagrammatic rules [22], we arrive at h (k1)(k2)i= (116) where we used xx= 1=p 2. This integral will not vanish in general, but its detailed value will depend on the time dependence of andBc, as previously discussed in [57,58]. The second contribution coming from a higher order quantum gravity term of the to an interaction (117) to linear order in . Expanding this to the quadratic order in the electromagnetic eld anisotropic part of the interaction Hamiltonian reads H(2) int, + 4 ixBiBx : (118) Apart from the changed polarization sum, this interaction Hamiltonian is similar to equation (33), if we take the eective coupling (119) Upon using the same techniques as earlier, we now = 2( (121) and the integrals I1andI2are dened in the same way as earlier but with the eective coupling ~ replacing. Note that, the structure of this result is quite similar to the previous result in equation (50), however, there are crucial dierences. Since the background magnetic eld is turned on in the x-direction, upon using equation (120) in equation (102), the angular integrals can be trivially carried out which will lead to a non-vanishing result for the h icorrelator. Therefore, such an imprint induced in this case can be considered as a novel consequence of such scenarios with higher order quantum gravity corrections. 236 Conclusions and order cosmological correlations generally provide new insights into the non-trivial interactions in the early universe. This paper has calculated a new set of such cosmological correlation functions and their corresponding semi-classical consistency relations. More precisely, we have computed the non-Gaussian correlation functions of primordial gravitons produced during in ation, with gauge elds that are non-minimally coupled in the early universe. Our correlation functions are derived, both in terms of the vector potential, corresponding to the correlation of gravitons with a vector boson, and in terms of the associated electric and magnetic elds for the U(1) gauge boson. We have obtained the full general results for these s for the rst time in the literature by taking into account the correct time evolution operator in the interaction picture of the in-in formalism. Moreover, we have also shown the presence of the leading order correction terms in all the correlators that had not been noticed earlier. In the squeezed limit, we found that the non-linearity parameters remain small but can be large for the other triangular shapes of the bispectra in Fourier space. In order to check our results, we have derived new semi-classical relations analogous to known semi- classical relations in the literature (also sometimes called the consistency relations or soft theorems, etc.). Some well-known examples are the Maldacena consistency relations [34] for the functions, the SSV relations [38] for the four-point functions, and the GS relations for the infrared part of loop diagrams [39]. These relations are generally valid for curvature and graviton modes, but they are not directly translatable to the present case. In the case, the gauge eld sector is invariant under the conformal transformations and therefore does only feel the expansion of spacetime through trivial conformal rescaling of the at Thus, it is evident that the non-trivial contribution to the correlation functions arises due to the time dependence of the non-minimal coupling that breaks the conformal invariance. Therefore all the correlation functions computed here are proportional to the time derivative of the function. However, a suitable redenition of the pump eld in terms of the non-minimal cou- pling function and a gauge eld redenition allows us to write the action for the polarization modes of the gauge elds in the identical form to a scalar eld in a ctitious expanding universe, with an expansion given by the time dependence of the non-minimal coupling. For a non-minimal to a ctitious in ationary universe, we show that our new semi-classical relations cor- rectly recover the squeezed limit of the full in-in calculations. We checked the agreement with our new consistency relations both for the vector eld itself, and also in terms of the corresponding electric and magnetic elds. In particular, it is worthwhile keeping in mind, that even if the gauge eld is not directly as in the case of dark sector gauge elds, it may be indirectly observable through its imprints on the curvature perturbations. Finally, we have calculated a direct correlation between one graviton mode and a curvature perturbation mode, induced by the three-point correlation function of one graviton mode and two gauge eld modes, and showed that it vanishes in the isotropic limit. However, as we brie y discussed in the previous section, this two-point correlator is actually non-vanishing in general due to the anisotropic background of long wavelength gauge elds and also receives non- trivial contributions in non-linear theories of electrodynamics arising in quantum gravity. It would further be interesting to study the observational consequences of such correlators in terms of non-zero s involving temperature and polarization anisotropies in the CMB and understand the prospects of their detectability with both ground and space based upcoming CMB would like to acknowledge nancial support from the new faculty seed start-up grant of IISc, the Core Research Grant CRG/2018/002200 from the Science and Engineering Research Board, De- partment of Science and Technology, Government of India and the Infosys Foundation, the Infosys Young Investigator award. MSS is supported by Villum Fonden grant 13384 and Independent Research Fund Denmark grant 0135-00378B. A Useful the integrals be explicitly written (123) Similarly for (125) 25In the limit corresponding to the end of in ation jkIj!0, one could show that To achieve this result, let's rewrite the integrals (58) and (59) as follows, J1= 2 (126) J2= 2 (127) Using equation (22), the time derivative of mode function can be (128) By denition, i.e. from (39), (40), (60) and (61), we can see (130) Now, let us take the limit jkIj!0 and consider the equations (124), (128) and (129), we will (131) Similarly by considering the equations (125), (128) and (132) One can evaluate these integrals for dierent values of nwhich are listed in the appendix of [17]. For completeness, we present the result below for the most interesting case of n= 2 which has been used to plot the non-linearity parameters in Sec. 3.2, + 3k3 1k2 t( + the Euler gamma constant. B Validity of the semi-classical approach Here, we shall quickly review the validity of the semi-classical approach for the three-point tensor perturbation. For the semi-classical derivation of single eld consistency relations, one argues that the eect of the long wavelength metric perturbation can be absorbed into a local coordinate transformation. Then the two point functions in a modied background can be calculated by rescaling the coordinates. And use them to calculate the squeezed limit three-point functions. We can see that the same arguments will be valid also in our scenario. So let's try to establish this fact at the action level. To start with, the gauge eld action of our interest is given (135) Now we can calculate the leading order action by introducing tensor (136) where ijis the tensor perturbation and S(0)denotes the free part of the action. So the second term in RHS of equation (136) corresponds to interacting part of the action. One can express this interacting part in the Fourier space and study its squeezed limit. In the squeezed limit, wherein the to ijis much smaller than the gauge elds, we can (137) Since ijis time independent in the squeezed limit, we can use the on-shell conditions to (137) as (138) where we have used the equation of motion of Aiin the last line of the above equation. (138) into (137) and (139) S(0)can be rewritten (140) 27Then the total action (141) where ~k2=k2kikj ij. One can identify this as the rescaled momentum in the rescaled So this establishes the fact that the eect of the long wavelength graviton mode can be absorbed into a local coordinate The in-in formalism and the dynamical correction term The electric eld is dened through the time-derivative of the vector potential. When promoting these to operators and taking expectation values, we need to be careful and distinguish between of operators in the Heisenberg picture and the interaction picture. The relation between the electric eld and the time-derivative of the vector potential is simplest in the Heisenberg picture where only the operators depend on time. We will therefore start by considering the time derivative of an operator in the Heisenberg picture, and then translate that into the corresponding operator in the interaction picture where also the state depends on time. In this appendix we will carefully interaction picture and Heisenberg picture operators, but in the main text all discussion is in the interaction picture and all operators are implicitly understood to be in the interaction picture unless explicitly dened be an observable of our interest which is constructed out of eld operators and The time evolution of this observable is given by the Heisenberg equation of motion d dOH() (142) One can use the standard procedure in quantum eld theory and express the solution in terms of a unitary operator Uas OH() the initial time and Uobeys the following =Htot()U(;i); (144) and satises the following = 1; U1(;i)U(;i) = 1; U (3;2)U(2;1) =U(3;1): (145) Thus,Ucan be interpreted as a time evolution operator which evolves the operator Ofrom an initial timeito a later time . In this work, we are interested in the expectation value of O(), the initial quantum state of the system. Since we are working in the the states are independent of time. Now, our main goal is to calculate equation (146). In order to do so, one has to split the total Hamiltonian as follows Htot() =H0() +Hint(); (147) whereH0() andHint() denote the free and the interacting parts of the Hamiltonian, we know the evolution of the system under free Hamiltonian, i.e. without any interaction, we can make use of it to understand the time evolution under full Hamiltonian. This can be easily illustrated by introducing the interaction picture. To this end, let us dene U0(;i) as i@ @U0(;i) =H0()U0(;i); (148) so one can rewrite the expectation value of O() as in equation (146) as (150) Then, we can dene the operator Oin the interaction picture in the following manner as OI() (151) Using this denition, we can now express the expectation value of O() (153) and the superscript Irefers to the interaction picture. The solution to the above equation can be written in terms of HI intas F(;i) (154) whereTis the time ordering operator. Using this equation, one can calculate the expectation value of an operatorOat (155) This result is the master formula of the in-in formalism for calculating equal time correlation functions. It was rst developed by Schwinger [59] and others [60{63] and then later applied to cosmology [64,65]. In this paper, we are interested in the correlators The construction of interaction picture of AAand BBis trivial but EErequires more attention, as it involves 29the time derivative of the vector potential. Let us therefore compare the time-derivative of the operator in the Heisenberg picture and in the interaction picture. The time derivative of the eld has a clear meaning as only the operators evolve in time in the Heisenberg (156) Now, identifying QH()(d=d )OH() and using (151), we have QI() equation (142), we get QI() (157) We conclude, that if we have an operator, which in the Heisenberg picture would take the form of a time derivative of an operator, i.e. QH()(d=d )OH(), then the interaction picture version of that operator is the time derivative of the interaction picture operator plus and extra piece, i.e. QI() = (d=d )OI() this observation to the electric eld in the interaction picture EI , we nd EI i() (158) Since all the operators in the above equation are in the interaction picture, let's suppress the corre- sponding superscript Ifor brevity. Then one can express the observable EE, to the leading order, in the interaction picture as + (159) We refer to the second term as the dynamical correction term and the third as the M. S. Turner and L. M. Widrow, In ation Produced, Large Scale Magnetic Fields ,Phys. Rev. D37 (1988) 2743. [2] B. Ratra, Cosmological 'seed' magnetic eld from in ation ,Astrophys. J. 391(1992) L1{L4. [3] F. Tavecchio et al. ,The intergalactic magnetic eld constrained by Fermi/LAT observations of the TeV blazar 1ES 0229+200 ,Mon. Not. Roy. Astron. Soc. 406(2010) ]. [4] A. Neronov and I. Vovk, Evidence for strong extragalactic magnetic elds from of TeV blazars ,Science 328(2010) 73{75, [ arXiv:1006.3504 ]. 30[5] W. Chen, J. H. Buckley, and F. Ferrer, Search for GeV -Ray Pair Halos Around Low Redshift Blazars ,Phys. Rev. Lett. 115(2015) 211103, [ arXiv:1410.7717 ]. [6] L. M. Widrow, Origin of Galactic and Extragalactic Magnetic Fields ,Rev. Mod. Phys. 74 (2002) 775{823, [ astro-ph/0207240 ]. [7] V. Demozzi, V. Mukhanov, and H. Rubinstein, Magnetic elds from in ation? ,JCAP 0908 (2009) 025, [ arXiv:0907.1030 ]. [8] R. Durrer, L. Hollenstein, and R. K. Jain, Can slow roll in ation induce relevant helical magnetic elds? ,JCAP 1103 (2011) 037, [ arXiv:1005.5322 ]. [9] C. T. Byrnes, L. Hollenstein, R. K. Jain, and F. R. Urban, Resonant magnetic elds from in ation ,JCAP 1203 (2012) 009, [ arXiv:1111.2030 ]. [10] R. K. Jain, R. Durrer, and L. Hollenstein, Generation of helical magnetic elds from in ation . [11] R. J. Ferreira, R. K. Jain, and M. S. Sloth, In ationary magnetogenesis without the strong coupling problem ,JCAP 10(2013) 004, [ arXiv:1305.7151 ]. [12] R. J. Ferreira, R. K. Jain, and M. S. Sloth, In ationary Magnetogenesis without the Strong Coupling Problem II: Constraints from CMB anisotropies and B-modes ,JCAP 06(2014) ]. [13] T. Kobayashi and M. S. Sloth, Early Cosmological Evolution of Primordial ,Phys. Rev. D 100(2019), no. 2 023524, [ arXiv:1903.02561 ]. [14] R. R. Caldwell, L. Motta, and M. Kamionkowski, Correlation of magnetic elds with scalar uctuations ,Phys. Rev. D84 (2011) 123525, [ arXiv:1109.4415 ]. [15] L. Motta and R. R. Caldwell, Non-Gaussian features of primordial magnetic elds in ,Phys. Rev. D85 (2012) 103532, [ arXiv:1203.1033 ]. [16] R. K. Jain and M. S. Sloth, Consistency relation for cosmic magnetic elds ,Phys. Rev. D 86 (2012) 123528, [ arXiv:1207.4187 ]. [17] R. K. Jain and M. S. Sloth, On the non-Gaussian correlation of the primordial with vector elds ,JCAP 02(2013) 003, [ arXiv:1210.3461 ]. [18] M. Biagetti, A. Kehagias, E. Morgante, H. Perrier, and A. Riotto, Symmetries of during the de Sitter Epoch ,JCAP 07(2013) 030, [ arXiv:1304.7785 ]. [19] D. Chowdhury, L. Sriramkumar, and M. Kamionkowski, s between and magnetic elds in bouncing universes ,JCAP 01(2019) ]. [20] D. Chowdhury, L. Sriramkumar, and M. Kamionkowski, Enhancing the s between magnetic elds and scalar perturbations through parity violation ,JCAP 10(2018) ]. 31[21] P. Berger, A. Kehagias, and A. Riotto, Testing the Origin of Cosmological Magnetic Fields through the Large-Scale Structure Consistency Relations ,JCAP 05(2014) ]. [22] S. B. Giddings and M. S. Sloth, Cosmological diagrammatic rules ,JCAP 1007 (2010) ]. [23] S. Weinberg, Infrared photons and gravitons ,Phys. Rev. 140(1965) B516{B524. [24] A. Strominger and A. Zhiboedov, Gravitational Memory, BMS and Soft Theorems ,JHEP 01(2016) 086, [ arXiv:1411.5745 ]. [25] A. Strominger, Lectures on the Infrared Structure of Gravity and Gauge Theory . [26] R. Z. Ferreira, M. Sandora, and M. S. Sloth, Asymptotic Symmetries in de Sitter and In ationary Spacetimes ,JCAP 04(2017) 033, [ arXiv:1609.06318 ]. [27] R. Z. Ferreira, M. Sandora, and M. S. Sloth, Patient Observers and Non-perturbative in In ation ,JCAP 02(2018) 055, [ arXiv:1703.10162 ]. [28] R. Z. Ferreira, M. Sandora, and M. S. Sloth, Is Patience a Virtue? Cosmic Censorship of Infrared Eects in de Sitter ,Int. J. Mod. Phys. D 26(2017), no. 12 ]. [29] L. Hui, A. Joyce, and S. S. C. Wong, In ationary soft theorems revisited: A relation ,JCAP 02(2019) 060, [ arXiv:1811.05951 ]. [30] A. Kehagias and A. Riotto, In ation and Conformal Invariance: The Perspective from ,Fortsch. Phys. 65(2017), no. 5 1700023, [ arXiv:1701.05462 ]. [31] K. Hinterbichler, A. Joyce, and J. Khoury, In ation in Flatland ,JCAP 01(2017) ]. [32] K. Hinterbichler, L. Hui, and J. Khoury, Conformal Symmetries of Adiabatic Modes in Cosmology ,JCAP 08(2012) 017, [ arXiv:1203.6351 ]. [33] K. Hinterbichler, L. Hui, and J. Khoury, An Innite Set of Ward Identities for Adiabatic Modes in Cosmology ,JCAP 01(2014) 039, [ arXiv:1304.5527 ]. [34] J. M. Maldacena, Non-Gaussian features of primordial uctuations in single eld ,JHEP 05(2003) 013, [ astro-ph/0210603 ]. [35] J. M. Maldacena and G. L. Pimentel, On graviton during in ation ,JHEP 09 (2011) 045, [ arXiv:1104.2846 ]. [36] P. Creminelli and M. Zaldarriaga, Single eld consistency relation for the 3-point function , JCAP 0410 (2004) 006, [ astro-ph/0407059 ]. [37] C. Cheung, A. L. Fitzpatrick, J. Kaplan, and L. Senatore, On the consistency relation of the 3-point function in single eld in ation ,JCAP 02(2008) 021, [ arXiv:0709.0295 ]. 32[38] D. Seery, M. S. Sloth, and F. Vernizzi, In ationary trispectrum from graviton exchange ,JCAP 03(2009) 018, [ arXiv:0811.3934 ]. [39] S. B. Giddings and M. S. Sloth, Semiclassical relations and IR eects in de Sitter and ,JCAP 1101 (2011) 023, [ arXiv:1005.1056 ]. [40] S. Nurmi and M. S. Sloth, Constraints on Gauge Field Production during In ation ,JCAP 07 (2014) 012, [ arXiv:1312.4946 ]. [41] S. Yokoyama and J. Soda, Primordial statistical anisotropy generated at the end of in ation , JCAP 0808 (2008) 005, [ arXiv:0805.4265 ]. [42] N. Barnaby, R. Namba, and M. Peloso, Observable non-gaussianity from gauge eld production in slow roll in ation, and a challenging connection with magnetogenesis ,Phys. Rev. D85 (2012) 123523, [ arXiv:1202.1469 ]. [43] J. Soda, Statistical Anisotropy from Anisotropic In ation ,Class. Quant. Grav. 29(2012) 083001, [ arXiv:1201.6434 ]. [44] N. Bartolo, S. Matarrese, M. Peloso, and A. Ricciardone, The anisotropic power spectrum and bispectrum in the f()F2mechanism ,Phys.Rev. D87 (2013) 023504, [ arXiv:1210.3257 ]. [45] D. H. Lyth and M. Karciauskas, The statistically anisotropic curvature perturbation 05(2013) 011, [ arXiv:1302.7304 ]. [46] A. A. Abolhasani, R. Emami, J. T. Firouzjaee, and H. Firouzjahi, Nformalism in and large anisotropic bispectrum and trispectrum ,JCAP 08(2013) ]. [47] M. Shiraishi, E. Komatsu, M. Peloso, and N. Barnaby, Signatures of anisotropic sources in bispectrum of the cosmic microwave background ,JCAP 05(2013) ]. [48] T. Fujita and S. Yokoyama, Higher order statistics of curvature perturbations in IFF model and its Planck constraints ,JCAP 09(2013) 009, [ arXiv:1306.2992 ]. [49] J. P. Beltr an Almeida, J. Motoa-Manzano, and C. A. Correlation functions of sourced gravitational waves in in ationary scalar vector models. A symmetry based approach , JHEP 09(2019) 118, [ arXiv:1905.00900 ]. [50] J. Martin and J. Yokoyama, Generation of Large-Scale Magnetic Fields in ,JCAP 0801 (2008) 025, [ arXiv:0711.4307 ]. [51] K. Subramanian, Magnetic elds in the early universe ,Astron. Nachr. 331(2010) ]. [52] M. Shiraishi, S. Saga, and S. Yokoyama, CMB power spectra induced by between metric perturbations and vector elds ,JCAP 1211 (2012) ]. [53] M. Giovannini, duality and the conditions of in ationary magnetogenesis , JCAP 04(2010) 003, [ arXiv:0911.0896 ]. 33

2108.10887

P. Jishnu Sai

Rajeev Kumar Jain, P. Jishnu Sai, Martin S. Sloth

On the primordial correlation of gravitons with gauge fields

v1: 33 pages, 2 figures; v2: 34 pages, 2 figures, discussion in section 5 extended, references added, version accepted in JCAP

JCAP 03 (2022) 054

10.1088/1475-7516/2022/03/054

hep-th astro-ph.CO gr-qc hep-ph

http://creativecommons.org/licenses/by/4.0/

We calculate the primordial correlation of gravitons with an abelian gauge field non-minimally coupled through a dynamical dilaton field or a volume moduli during inflation in the early universe. In particular, we compute of a tensor mode with two gauge field modes and correlation functions for the associated magnetic and electric fields using the in-in formalism. Moreover, using semi-classical methods, we show that the three-point functions satisfy new (soft theorems) in the squeezed limit. Our findings exhibit a complete agreement of the full in-in results with the new . An interesting consequence of our scenario is the possibility of a novel correlation of the primordial tensor mode with the primordial induced by higher order quantum gravity corrections. The anisotropic background created by long wavelength gauge field modes makes this correlation function non-vanishing. Finally, we discuss how these functions are imprinted on cosmological observables today and the applications to scenarios of inflationary magnetogenesis.

Sensors 2021 ,21, 5590 2 of 16 The Unihedron Sky Quality Meter (hereafter SQM), a low-cost and pocket-size NSB pho- tometer, has given the general public the possibility of quantifying the NSB at any place and time [12]. It is a silicon photodiode with a TSL237 converter and a Hoya CM-500 filter that limits its effective bandpass to 400-650 nm. Based on the same TSL237 photodiode SQM detector, the European funded project a new Telescope Encoder and Sky Sensor ( TESS [13] ), with a more extended spectral response in the red (400-800 nm) to include the emission lines of High Pressure Sodium (HPS) converter sensor technologies used in SQM and TESS expanded the range of inexpensive and easy-to-use NSB photometers, with- out sacrificing the precision and detail of professional instruments [ 14]. However, although the cost of photometers has dropped, their deployment in remote locations (e.g., areas) remains a major drawback for many projects. SG-WAS, a new, low-cost NSB photometer based on the TSL237 sensor, is presented in this article, which is organized as follows. In Section 2, a description of the SG-WAS photometer is given, together with a detailed explanation of WAS technology and field tests. The new Sky Integrating Sphere (SIS) method and its results are presented in Section 3, along with a brief discussion of the implications of the results and uncertainties. Finally, the main conclusions are provided in Section 4. 2. SG-WAS Photometer The SkyGlow Wireless Autonomous Sensor (SG-WAS) is a new NSB photometer partially funded by the EU project EELabs, and designed and manufactured by the Canarian R&D company Sieltec Canarias S.L., under the scientific coordination of the Instituto de Astrofisica de Canarias (IAC, Tenerife, Spain). SG-WAS is basically composed of the following. 1. A TSL237 converter sensor. 2. An optical dichroic that determines the spectral response of the device. 3. A concentrator optic that focuses the light and determines the effective field of view (FOV). 4. An infrared (IR) sensor that measures the sky temperature. 5. Two microcontrollers that convert the frequencies measured into an average magnitude and uncertainty , store and send the information, and control the power-saving strategy . 6. A communication unit (LoRa, WiFi, LTE-M) customizable to the place where it is installed. The photometer's dimensions are similar to those of other photometers of its class (10 cm 10 cm 4 cm) and its weight is approximately 100 g. It is designed to be on a flat surface (optionally provided by the manufacturer), so its 20 and the IR sky temperature sensor face the zenith (see Figure 1). On one side of its casing, there is a color-coded manual switch-off button that may be enabled in A photovoltaic cell on the upper part allows it to be recharged within a few hours with direct sunlight (see Section 2.2). The SG-WAS is the first fully autonomous and wireless NSB photometer. It is waterproof and resistant to adverse weather conditions. 2.1. Wireless In recent decades, there has been a growing interest in installing sensor networks to monitor different variables related to environmental protection [ 15,16], ecology [ 17,18], and volcanology [ 19], among others, in remote or locations with no These wireless sensor networks (WSN) are organized as a series of dispersed nodes with sensing capabilities that collect and send data to a centralized hub. All SG-WAS photometers are independent NSB sensors conceived to be used as one of these nodes. The 2021 ,21, 5590 3 of 16 Figure 1. SG-WAS photometer viewed from above ( a) and two sides ( b). Its dimensions are 10 cm  10 cm 4 cm and it weighs about 100 g. The main elements are sky brightness optics (1), (2), solar panel (3), fixing screws (4), and power button (5). and communication process is shown in Figure 2. There are two microcontrollers M2 is connected to the sensor, takes measurements, and stores them temporarily. * M1 handles timing, storage, data processing, and communication. It is connected to M2 and the central hub (GateWay in LoRa version or server in WiFi/LTE-M version). To reduce power consumption as much as possible, two independent processes are carried out: 1. Every 5 min, M1 calculates the average and instrumental uncertainty of the ten measure- ments taken by M2 in the previous slot, stores them, and commands M2 to take a new set of measures. 2. In the corresponding slot, depending on the device version, M1 encrypts, and sends the measurements stored in its memory to the central hub. These include the average and standard deviation of the ten continuous NSB measurements, ambient and sky temperature, battery charge, and communication signal strength. During the rest of the time, both microcontrollers remain in a deep sleep state, so the battery consumption is extremely low. Different versions of SG-WAS depend on the com- munication protocol (WiFi, LoRa, or LTE-M), the connection to the central hub and how the sending slots are assigned. 2.1.1. LoRa Version LoRa [ 20] is a patented wireless communication technology commercialized by since 2012. Signals are modulated in sub-GHz license-free Industrial, Scientific, and Medical (ISM) radio bands, making it affordable, deployable worldwide, and interopera- ble. With its ability to reach long distances of up to 15 km in rural areas, provide data rates in the range of kilobits per second, and more inexpensive base stations (GateWay), LoRa is a very attractive technology for WSN communication in remote there are some ALOHA-based communication protocols with random [ 21] is one of most widely known--that support networks with thou- sands of sensors connected to the same GateWay, networks with densities of less than oneSensors 2021 ,21, 5590 4 of 16 Figure 2. Measurement and communication diagram of the SG-WAS photometer. Each device has two M2 takes ten NSB measurements and stores them temporarily; M1 calculates the average and uncertainty of the data set taken in the previous slot, and stores them until they are transmitted to the central hub. This process is repeated every 5 min. In the LoRa version, the GateWay assigns a 6 s slot every 5 min to each photometer to send its measurements and synchronize its timestamp for the next cycle. The GateWay sends the measurements of all the devices connected to it to the server. In the WiFi and LTE-M version, the measurements are directly sent to the server once every hour and M1 is continuously synchronized with the network. During the rest of the time, both in a deep sleep state, so the battery consumption is extremely low. photometer per square kilometer are not expected [ 9]. Therefore, each GateWay is limited to several tens of photometers. In addition, random access channels require data re-sending or active listening strategies to ensure the reception of all the packets, which is a power- consuming process. Consequently, a deterministic communication protocol of Time Access (TDMA [22]) has been developed. When a new photometer is added to the network, a connection with the GateWay is established and a 6 s slot in the 5 min cycle is assigned. Within these 6 s, data collected by the device are sent to the GateWay, which returns a confirmation message and synchronizes the time to the next slot. The photometer will not connect again until then and enters a deep sleep state. All data collected by the devices in the network are sent to the server by the GateWay via WiFi-MQTT a 10% error in the slot time (a total sending time of 6.6 s), each GateWay can accept up to 45 devices using the protocol developed in this work. A test network of 20 SG-WAS LoRa has been successfully established around the Teide Observatory (OT, Canary Islands, Spain). 2.1.2. WiFi and LTE-M Version The WiFi version incorporates a communication module via the MQTT protocol. The LTE- M version has a 2G multiband card and may incorporate an additional antenna. The slottedSensors 2021 ,21, 5590 5 of 16 (a) 3. (a) Normal charge -discharge cycle for three photometers SG-WAS LoRa (green), LTE-M (brown), and WiFi (blue) versions located at the Teide Observatory (OT, Canary Islands, Spain). The measured by the Stella telescopes weather station, is shown at the bottom; (b) discharge curve of an SG-WAS WiFi when not exposed to direct sunlight. It takes more than 20 days to stop uninterruptedly until it reaches the hibernation state; ( c) charge curve of an SG-WAS once it receives direct sunlight from the hibernation state; ( d) voltage curve of the device 1 53 placed in the Valley of Tejeda (Gran Canaria, Spain), which is no longer exposed to direct sunlight in in hibernation for 4 months and is charged in March when it begins to receive sunlight algorithm described above is not used with these versions because devices are directly connected to the internet. Data are stored in each device (M1, see Figure 2 ) and sent to the server every hour. 2.2. Autonomous The measurement and communication algorithm described above has been developed to optimize the power consumption of the device. It incorporates a solar panel at the top that recharges the internal Li-Ion battery under direct sunlight (nominal 5.5 V and 35 mA with solar irradiance of 455 W/m2). Many battery life tests have been performed with the different SG-WAS versions to study their autonomy. Figure 3a shows the charge and discharge curve of three devices located at the OT, under a daytime solar irradiance ~1000 W/m2. In all versions, the charge lost during the night is recovered in the first 4 h of the day even in the presence of thin clouds (see a less smooth curve on 18 June), at a rate of 15-20 mV/hr. Once the voltage peak is reached, the battery begins to discharge slowly at a rate of 2.5 mV/h. During the self-charging and discharging phases, measuring and communication is not interrupted at any time. A complete discharge curve is shown in Figure 3b. The 1 101 (WiFi) photometer was placed inside a telescope dome that was closed during daytime. It took more than 20 days for the device to reach its battery hibernation point. This checkpoint is introduced to prevent the battery from reaching the deep discharge limit and being damaged. Once 2021 ,21, 5590 6 of 16 to sunlight, the device begins to charge and send measurements again, reaching the peak voltage just a week later (Figure 3c). The hibernation phase can be maintained for months; see Figure 3d for the battery voltage of photometer 1 53 located in the Valley of Tejeda (Gran Canaria, Spain). It is located on a rocky point that does not receive sunlight during the winter and part of the autumn, so the photometer started to discharge and went into hibernation in mid-November 2020. Four months later, in mid-March 2021, it started receiving direct sunlight, sending measurements and recharging again, returning to its normal operating state. The battery stability and autonomy of the device has been extensively tested Sensor TSL237 is a silicon photodiode combined with an electrical con- verter on a single monolithic CMOS integrated circuit. It was designed by Texas Solutions [ 23] and is the most widely used sensor in low-cost of its low price, high sensitivity, and accuracy. TSL237 covers a spectral range between 300 nm and 1050 nm with maximum sensitivity at 700 nm, which lies between the visible and near infrared. According to the manufacturer, it can work between 25oC and 70 oC, and is thermally compensated between 320 nm and 700 nm. The output signal is a square wave with a frequency that is linearly proportional to the intensity of the light (irradiance) incident on the photodiode in the range between 1 Hz and 1 MHz. A 1.8 mm diameter lens is integrated in the optical center and 0.07 mm above the to the manufacturer [ 23], the sensor linearity is fulfilled in the range between 1 Hz and 1 MHz when illuminated with a spectral lamp at 524 nm at ambient tempera- ture. This has been verified before with several SQM devices down to 19 measuring in darker places requires studying the literality down to magnitude to sub-Hz frequencies. A TSL237, an (A V ANTES) fiber spectrometer and a HALOSTAR 50W 12V GY6.35 (OSRAM) lamp were placed inside a black box. The intensity and position of the lamp were varied to cover the range of 0.01 to 10,000 W/m2received at the entry of the fiber and the sensor, which are placed together, while recording the frequency registered by the sensor. The collected measurements, including the correspondence to are shown in Figure 4(c). Linearity is maintained down to 0.01 Hz, equivalent to about 24 mag/arcsec2, with an R2very close to 1. The TSL237 sensor is therefore linear in the measurement of NSB in both light-polluted and natural dark Angular Response As for the SQM-L and TESS-W, a low-cost concentrator lens is located before the the optical path), allowing more light to be collected in a narrower FOV . To verify the angular response of the device in its principal axes, it has been placed on a a point light source. Positive values indicate a clockwise rotation, being negative otherwise. Measurements have been taken from 45to 45in 1steps with three different devices to avoid possible random errors in the mounting of one of them. The mean values and their deviation are shown in Figure 4(a). The FOV is approximately Gaussian in both axes, with a full width at half maximum (FWHM) of 18.20.3deg in the vertical and 19.50.3deg in the horizontal. The bumps that stand out from the fit are due to stray at very specific positions in the experimental set-up. This FWHM is close to that of SQM-L ( 20[12]) and TESS-W the FOV , the contribution of each integrated ring, defined from the with respect to the optical axis, to the total flux measured by the device can beSensors 2021 ,21, 5590 7 of 16 (a) (b) (c) Figure 4. (a)): FOV in the vertical (brown, defined in the IR direction) and horizontal (blue) axis, obtained in the laboratory by measuring received irradiance for different angles with respect to the optical axis. The FWHM obtained from the Gaussian fit in both axes is included. ( b): Weight function of each FOV ring over the total flux. ( c): Frequency measured by the TSL237 sensor against received irradiance; its approximate equivalence to magnitudes per square arcsecond is shown on the right axis. A fit is included to show the linearity of the detector to magnitudes greater than natural sky This is a weight function that should be considered when calculating the average radiance at the entrance of the detector plane--see, for example, Equation (7). The result on both axes is shown in Figure 4(b). Although the detector response decreases with separation, the integrated sky area is larger and the maximum is located the optical axis. This is particularly meaningful in the presence of by extended artificial (ALAN) or natural (twilight, moonlight, galactic plane, but less for point sources such as stars, whose fluxes are added to the rest of the sky in the integrated ring. 2.3.3. Transmittance A window and optical filter were placed in the entrance plane in the early versions of SG-WAS . This is a similar configuration to that of the SQM, with a Hoya CM-500, and TESS-W, with a dichroic filter. The empty gap between the two surfaces is exposed to condensation from any moisture that may remain inside the case. Manufacturers of the TESS-W included a heater to keep the internal temperature above 10 degC and avoid such condensation. Given the auton- omy requirements of the SG-WAS, this energy-intensive solution (more the measurement and communication processes) is not feasible. The dichroic +window assembly has been replaced, then, by a single dichroic, which limits the spectral response of the detector and ensures proper sealing of the device. This is also an opportunity to optimize the transmittance of the device, a critical con- sideration if an additional color filter is added to the SG-WAS optics (under window causes an approximately 8% loss in transmittance and the filter, 6%. If they were considered as two independent elements, the total transmittance loss would be Removing the window causes the optical system to become more sensitive, as shown in Figure 5. The spectral range of SG-WAS is fitted with BVR filters, from 400 to 720 nm . These filters cover the main optical airglow lines.Sensors 2021 ,21, 5590 8 of 16 300 400 500 600 700 800 900 Wavelength OH 5. Spectral response curve of the SG-WAS. The sensitivity of the TSL237 sensor is limited to the visible range by a dichroic filter, and the window has been removed to improve the transmittance of the optical system. The transmittance of the TESS-W and SQM-L photometers and Johnson- Cousins [25] BVRI filters has been included as a reference, as well as the night sky spectrum obtained using the SkyCalc tool [ 26], where the brightest airglow lines are labeled. Adapted with permission from [ 10] (c)AAS 2021. 3. Results and Discussion 3.1. SG-WAS Astronomical Magnitudes and standards in astronomy, brightness is usually expressed in units of magnitude per square second as log10(ffD), (1) where ZPis the zero point of the photometer using a reference light source, and fand fDare the signal and dark frequencies of the sensor, respectively, both measured in Hz. To avoid mistakes, it is more correct to consider the SQM, TESS, and SG-WAS responses as different photometric systems. The conversion factors between photometric systems can be obtained based on the kind of spectra of the observed object [14] . Hereafter, the brightness in in the SG-WAS passbands will be called m SGand Equation into mSG=ZP2.5 (2) At typical nighttime low operating temperatures (<20), the value of the dark frequency , fD SG, turns out to be negligible in most cases; therefore, the total error of the measurements of the SG- WAS photometer for output frequency fSG,d2mSG, is obtained and converted to units by propagating the expression (2) as (3)Sensors 2021 ,21, 5590 9 of Error (4) dmins SG=2.5 Error. (5) Both calibration and instrumental errors must be taken into account when determining the total uncertainty of the Sky Integrating Sphere (SIS) Calibration Method The determination of the ZPis the key issue in the reproducibility of measurements and consistency between different devices. Starting from the expression (2), and considering that the measurement of the reference device mSGrefis expressed in the photometric absolute (AB) magnitude system, the difference between this and a simultaneous measurement of another device will be given log10 the comparison calibration process, we are interested in establishing the ZPvalue from the magnitude difference. Under the same illumination conditions, two correctly cali- brated photometers, i.e., with the ZPwell determined, should measure the same magnitude within the uncertainty range, so the ratio between constant. The relationship between the frequency measured by the TSL237 sensor and the and FOV-averaged radiance Lat the entrance plane of the detector is given Ll(w)is the spectral radiance of the incident light field along the direction specified by the angular vector w,P(w)is the weight function describing the FOV of the such the spectral response of the device and Ka constant that provides the absolute link between the converter output frequency and [14]. Under angularly uniform radiance illumination is only if (ii) fD, which varies with temperature, is the same, i.e., fD SG=fD SGref, or negligible; (iii) there is no significant change for both devices in the spectral response T(l), bearing in mind that the incident radiance has the same i.e., LSG l=LSGref l). These three conditions must be fulfilled to obtain the ZPfrom the difference of simultaneous measurements of two photometers. A well-established procedure is to use a stable light source to illuminate a and take simultaneous measurements between a reference photometer and the one to be calibrated, thereby establishing the ZP from the differences. In both the SQM-L and an integrating sphere is used to ensure homogeneous irradiance. In the first case, a broadband spectral lamp is used, with increasing irradiance in the wavelength range of 350 to 500 nm that remains almost constant thereafter [ 27] and, in the second one, a LED with a spectral response centered at 596 nm and an FWHM of 14 nm [28] . This system presents several difficulties in ensuring reproducibility and usually requires a 1-by-1 2021 ,21, 5590 10 of 16 On the one hand, a dependence on the experimental setup is introduced that may lead to increased measurement uncertainty (as reported by [ 24]) due, for example, to or poor screening of the background stray light breaking the angular uniformity (i). Moreover, it requires careful control of the temperature at which the measurements are taken, as this may not only affect the emission spectrum of the lamp but also the dark frequency, which may not be negligible at working temperature (ii). On the other hand, it is known that the spectral response of different photometers is not the same, so condition (iii) is not always satisfied. In the extreme case, it is not correct to obtain the ZP by comparing simultaneous measurements of a TESS-W and an SQM-L because their spectral response is very different and hence also the frequency measured, even under the same illumination conditions. This effect also occurs, although to a lesser extent, in photometers of the same type. When calibrating with a narrow source, as in the case of the TESS-W, it is assumed that the difference in magnitudes obtained in the spectral range of the lamp can be extrapolated to the full spectral response of the device, thus introducing an undesired and unknown systematic error. To reduce this systematic error in night sky brightness measurements, a light source with a spectral distribution similar to that of the night sky may be much more appropriate. In this work, we have developed a novel calibration method based on of the night sky that can be used in any type of NSB device. The night sky is likely to behave as an integrating sphere if there is no light gradient in the photometer FOV . Therefore, it is necessary to avoid localized sources of both artificial (ALAN) and natural light emission (astronomical components such as the Moon or the Galaxy). All are calibrated on zenith observations taken from the OT, where light pollution is minimal [ 10] and condition (i) satisfied (further details on the calibration procedure are omitted because of a patent-pending process). When taking simultaneous nighttime only is the temperature the same in all devices, but also, at an altitude of 2500 m, it does not get high enough for the dark frequency to be significant. In addition, many up to 100 devices at the OT) can be calibrated at the same time as there is space available on the optical table. The most meaningful point of this method is that the calibration measurements are taken with the same spectrum of the night sky. This allows the systematic error to be very well constrained to the absolute calibration of the reference device. The difference between the measurements is thought to come mainly from the small variations in the spectral response of the photometer, resulting in a variation of the magnitude value with the changes that take place in the night sky spectrum over time. 3.3. SG-WAS Calibration Error The SIS method was used to obtain the ZP of 10 TESS-W and 21 SG-WAS photometers. The differences between simultaneous zenith measurements taken between pairs of photome- ters under dark conditions (following the definition described in [ 10]) for the period 2021 January 18-May 26 and for TESS-W and between 2021 April 14-June 10 for SG-WAS are shown in Figure 6(a). Both distributions are centered at 0, as expected, and have a of 0.01 mag/arcsec2in the case of SG-WAS and 0.02 mag/arcsec2for TESS-W. These values are greater than the combination of instrumental errors (more detail in the next section) and are due to the effect of variations in the night sky spectrum captured by a slightly response in the devices. The uncertainty in the determination of the ZP by the SIS method does not exceed 0.02 mag/arcsec2, as it is shown in Figure 6(b). This is a reduction of more than a factor 2 of the 0.044 as the calibration error in the TESS-W [14,28] and a factor of 5 of the SQM [24].Sensors 2021 ,21, 5590 11 of 16 0.10 0.05 0.00 0.05 0.10 0.15 NSB Difference CountsTESS-W (ori.) TESS-W (SIS) SG-WAS (SIS) 0.005 0.010 Error (SIS) TESS-W (SIS) Figure 6. (a): Distribution of differences in simultaneous NSB measurements between pairs of TESS-W (brown) and SG-WAS (blue) photometers after calibrating them using the Sky Integrating Sphere (SIS) method. The distribution of TESS-W photometer pairs with the original calibration (ori.) given by the manufacturer is also shown unfilled. ( b): Distribution of calibration errors for both with the SIS method. The original calibration of the TESS-W, which is made by the manufacturer in the Labo- ratory for Scientific Advanced Instrumentation (Laboratorio de Instrumentacion LICA) of Universidad Complutense de Madrid (UCM, Spain), presents a more scattered distribution in the differences, with a standard deviation of 0.042 unfilled distribution in Figure 6). Figure 7 shows the variation of the calibration error for a TESS-W and an SG-WAS as a function of the number of measurements considered, taken randomly from the whole set. The choice of these particular devices is arbitrary, and their reproducibility has been checked for all other available photometers. From 360 measurements onward, the ZP determination is considered to have reached its accuracy limit and the calibration error cannot be further reduced. Given the temporal resolution of the photometers, this is equivalent to 6 h of measurements in the TESS-W and 30 h in the SG-WAS. Although the time available to take such measurements depends on different astronomical and atmospheric components, it has been verified that this interval is easily achievable in one week at the OT, regardless of the time of year. The SIS method presented in this work not only reduce the calibration uncertainty but also allows a massive calibration to be performed -currently up to 100 devices - by measurements for no longer than one clear week. 3.4. SG-WAS Instrumental Error A statistical study of the data has been carried out to characterize the instrumental error of the system (sensor +optics). For this purpose, 5000 laboratory frequency taken for magnitude values between 17 and 22 mag/arcsec2(as expected in the field concerned) in steps of 1 mag/arcsec2. The optical setup is similar to that used in For each set of 5000 measurements, a histogram has been obtained and represented in Figure 8. Statistical parameters are shown in Table 1. The calculated standard deviation values for each distribution magnitude is lower than 0.004 mag/arcsec2(see Table 1). This experiment therefore shows that the instrumental error of the SG-WAS photometer is lower than a few Taking into account that this laboratory error calculation can still be affected by other factors such as the power supply error, the stability of the lamp used, the design of the experiment (background stray light),Sensors 2021 ,21, 5590 12 of 16 Figure 7. Calibration error obtained with the SIS method for a TESS-W photometer ( a) and an SG-WAS (b), as a function of the number of measurements taken randomly. It is considered that the lower limit is reached from 360 onward (dashed vertical lines), corresponding to 6 and 30 h of night sky 23.8 fSG countsmSG = 17.007 mins SG = 0.003 9.50 9.55 fSG 3.76 fSG 1.49 fSG 0.595 fSG 0.236 fSG 8. Distribution of 5000 laboratory frequency measurements with light intensities for between 17 and 22 mag/arcsec2range in steps of 1 mag/arcsec2. The standard deviation in magnitude units, corresponding to the instrumental uncertainty, is included on interference, etc., it should be considered an upper limit to the instrumen- tal error. The new measurement method developed for the SG-WAS allows us to obtain the average of ten continuous measurements of NSB and also its standard deviation. The time taken by the device to take a measurement is inversely proportional to the frequency. Following Table 1, it takes approximately 4.5 s to take one measurement at magnitude 22, 1.7 s at magnitude 21, and so on. In total, ten continuous measurements at magnitude 22 are taken in about 45 s. As the possible variations in sky brightness produced by airglow are not measurable on this time scale [ 10], the standard deviation of these measurements corresponds only to the instrumental error, which is provided along with the NSB every 5 min in the SG-WAS. 3.5. Night Sky Brightness than one-hundred SG-WAS photometers were installed at the OT during 2021 to test their resistance to extreme weather stress conditions (snow, ice, strong wind, high solar radiation, among others) and the reliability of their NSB measurements. Figure 9 shows the zenithal brightness curve on 2021 May 14 for five SG-WAS and two TESS-W placed together. In the inner plot, the dark and clear period (as defined by [ 10]) is shown enlarged. Note that the SG-WAS curves are smoother than those of TESS-W, which is a result of the averagingSensors 2021 ,21, 5590 13 of 16 Table 1. Frequency standard deviation values for each distribution. Following Equation (5), the magnitude, dmins SG, was also Magnitude (mag/arcsec2) (Hz) 17.989 0.0134 0.002 3.7564 19.003 0.0100 0.003 1.4833 20.012 0.0053 0.004 0.5918 21.009 0.0009 0.002 0.2350 22.013 0.0004 0.002 process. Variations associated with airglow are clearly visible in both devices, although some differences between them are appreciable. This is a consequence of the slight difference in the spectral response of both devices. The SG-WAS curves are very similar to each other, with their difference being less than one-hundredth of a magnitude. The lower plot shows the total uncertainty in the measurements, including the instrumen- tal error (calculated for each SG-WAS measurement and its upper bound of 0.004 mag/arcsec2 for TESS-W) and the calibration error (obtained by the SIS method for SG-WAS and TESS-W). The uncertainties of the SG-WAS measurements are below 0.02 throughout the night, except at twilight, when the rapid change in sky brightness increases the standard deviation of the ten measurements. These curves have been found to be very nearly the same for all other SG-WAS photometers, regardless of version and series. 3.6. Light Pollution Laboratory: IoT-EELab The measurements collected by the EELabs and STARS4ALL networks are openly avail- able in real-time at the IoT-EELab website2. It is an interactive dashboard that collects, controls, and analyses data from hundreds of NSB sensors and several all-sky cam- eras. There are more than 170 million entries available containing brightness data collected for more than 500 photometers (400 TESS and 100 includes a simple reading page with NSB data from a single with calculation of the lunar phase, mean and median values of NSB during the astro- nomical night, slopes of brightness curves in different time intervals, sunset and sunrise hours, and percentage of cloud cover. Furthermore, there are statistical tools to obtain diagrams, jellyfish plots, and monthly and annual NSB trends, among other features. All raw and filtered data can be downloaded in standard format. a unique opportunity for researchers to experiment with new and existing light pollution datasets. 4. Conclusions In this work, the characterization of the new SG-WAS, a low-cost NSB photometer based on the TSL237 sensor (similar to SQM and TESS), has been carried out. The main conclusions are summarized as follows: * SG-WAS is the first wireless NSB sensor to communicate via LoRa, WiFI, or LTE-M and be powered by solar energy. A measurement every 5 min is taken continuously and sent using a Time Division Multiple Access algorithm to avoid packet collision. * The device recharges to peak voltage in just 4 h with direct solar irradiance after a full night of operation. It can stay up to 20 days in darkness while taking 2021 ,21, 5590 14 of Sky 22:00 23:00 00:00 01:00 02:00 03:00 04:00 05:00 Time UTC 00:00 01:00 9. NSB curve of five SG-WAS (1 X) and two TESS-W (starsX) photometers taken on 2021 May 14 from Teide Observatory (Canary Islands, Spain). The vertical lines indicate the end ( left) and beginning (right ) of astronomical twilight (black), the leaving of the zodiacal light (blue) and the entrance of the Galaxy (green) in the FOV of the device. The inner graph shows the dark period [10] enlarged. The SG- WAS curves are smoother than those of TESS-W as a consequence of the 10-measurement The lower plot shows the total uncertainty in the measurements, including the for each SG-WAS measurement and its upper bound of 0.004 mag/arcsec2for TESS- W--and the calibration error--obtained by the SIS method for SG-WAS and 0.044 remain in hibernation for months before returning to operation once it is illuminated again. * Its optical design is very similar to the TESS-W, with a FOV approximately 19(FWHM) and a slightly less red spectral range, from 400 to 720 nm. The window has been removed to prevent condensation on the dichroic and increase transmittance. * A new Sky Integrating Sphere (SIS) calibration method (patent-pending) has been de- signed and demonstrated to achieve calibration errors in TESS-W and SG-WAS below 0.02 mag/arcsec2. * The robustness to adverse weather conditions and the stability of its measurements have been demonstrated in field tests. Taking the average and uncertainty of ten makes the NSB curves smoother and avoids spikes. Differences measurements of several SG-WAS photometers have been found to have a standard deviation of 0.01 mag/arcsec2, several times smaller than their Contributions: M.R.A., M.P .-S., and M.S.-R. conducted the design of the experiments, the inter- pretation of the results and the organization and writing of the manuscript. M.R.A., M.S.-R., and S.L.-P . developed the SIS calibration method and IoT-EELabs tool. The measurements taken in the laboratory of Sieltec Canarias SL were performed by M.P .-S., M.M., and C.L. All authors have read and agreed to the published version of the 2021 ,21, 5590 15 of 16 Funding: This research was supported by EELabs a project funded by the European Union INTERREG V-A MAC 2014-2020. This work has been supported by the STARS4ALL Foundation, which maintains the data infrastructure for the TESS photometer network. Sieltec Canarias S.L. used its own resources, in addition to a grant from the Ministry of Economy, Knowledge and Employment, Canary under PILA 157G0042, 85% co-financed by the ERDF within the 2014-2020 and PILA 207G0385 for investment projects of SMEs in the Canary Review Board Statement: Not Consent Statement: Not applicable. Data Availability Statement: Access and download of the data used is openly available at We are grateful to Ramon Negrillo Valdivia (Software Developer, Sieltec Canarias S.L.) and Patricia Hernandez Rodriguez (Electronics Engineer, Sieltec Canarias S.L.) for their contribu- tions to the development of the of Interest: MSR and SLP are members of the EU EELabs project SG-WAS development team. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the Holker, F.; Wolter, C.; Perkin, E.K.; Tockner, K. Light pollution as a biodiversity threat. Trends Ecol. Evol. 2010 ,25, 681-682. 2. Gaston, K.J.; Bennie, J.; Davies, T.W.; Hopkins, J. The ecological impacts of nighttime light pollution: A mechanistic appraisal. Biol. Rev. 2013 ,88, 912-927. 3. Bennie, J.; Davies, T.W.; Cruse, D.; Gaston, K.J. Ecological effects of artificial light at night on wild plants. J. Ecol. 2016 ,104, 611-620. 4. Owens, A.C.; Lewis, S.M. The impact of artificial light at night on nocturnal insects: A review and synthesis. Ecol. Evol. 2018 , 8, 11337-11358. 5. Falchi, F.; Cinzano, P .; Duriscoe, D.; Kyba, C.C.M.; Elvidge, C.D.; Baugh, K.; Portnov, B.A.; Rybnikova, N.A.; Furgoni, R. The new world atlas of artificial night sky brightness. Sci. Adv. 2016 ,2, e1600377, Stark, H.; Brown, S.S.; Wong, K.W.; Stutz, J.; Elvidge, C.D.; Pollack, I.B.; Ryerson, T.B.; Dube, W.P .; Wagner, N.L.; Parrish, D.D. City lights and urban air. Nat. Geosci. 2011 ,4, 730-731, de Miguel, A.S.; Kyba, C.C.M.; Zamorano, J.; Gallego, J.; Gaston, K.J. The nature of the diffuse light near cities detected in imagery. Sci. Rep. 2020 ,10, Jiang, W.; He, G.; Long, T.; Guo, H.; Yin, R.; Leng, W.; Liu, H.; Wang, G. Potentiality of Using Luojia 1-01 Nighttime Light Imagery to Investigate Artificial Light Pollution. Sensors 2018 ,18, 2900, Bara, S. Characterizing the zenithal night sky brightness in large territories: How many samples per square kilometre are needed? Mon. Not. R. Astron. Soc. 2018 ,473, 4164-4173, Alarcon, M.R.; Serra-Ricart, M.; Lemes-Perera, S.; Mallorquin, M. Natural Night Sky Brightness during Solar Minimum. Astron. J. 2021 ,162, 25, Aube, M.; Simoneau, A.; Munoz-Tunon, C.; Diaz-Castro, J.; Serra-Ricart, M. Restoring the night sky darkness at Observatorio del Teide: First application of the model Illumina version 2. Mon. Not. R. Astron. Soc. 2020 ,497, 2501-2516, Cinzano, P . Report on Sky Quality Meter, Version L .2007 . Available online, accessed on 28-06-2021. 13. Zamorano, J.; Garcia, C.; Tapia, C.; de Miguel, A.S.; Pascual, S.; Gallego, J. STARS4ALL Night Sky Brightness Photometer. Int. J. Sustain. Light. 2017 ,18, 49-54, Bara, S.; Tapia, C.; Zamorano, J. Absolute Radiometric Calibration of TESS-W and SQM Night Sky Brightness Sensors. Sensors 2019 , 19, 1336, Hart, J.K.; Martinez, K. Environmental sensor networks: A revolution in the earth system science? Earth Sci. Rev. 2006 ,78, 177-191. 16. Vairamani, K.; Mathivanan, N.; Venkatesh, K.A.; Kumar, U.D. Environmental parameter monitoring using wireless sensor Exp. Tech. 2013 ,56, 468-471. 17. Zhang, P .; Sadler, C.M.; Lyon, S.A.; Martonosi, M. Hardware design experiences in ZebraNet. In Proceedings of the 2nd on Embedded Networked Sensor Systems, Baltimore, MD, USA, 2004 ; pp. 227-238. 18. Vera-Amaro, R.; Angeles, M.E.R.; Luviano-Juarez, A. Design and analysis of wireless sensor networks for animal tracking in large monitoring polar regions using phase-type distributions and single sensor model. IEEE Access 2019 ,7, 45911-45929. 19. Awadallah, S.; Moure, D.; Torres-Gonzalez, P . An internet of things (IoT) application on volcano monitoring. Sensors 2019 ,19,

2108.10891

Miguel R. Alarcon

Miguel R. Alarcon, Marta Puig-Subir\`a, Miquel Serra-Ricart, Samuel Lemes-Perera, Manuel Mallorqu\'in, C\'esar L\'opez

SG-WAS: a new Wireless Autonomous Night Sky Brightness Sensor

15 pages, 9 figures

Sensors 2021, 21(16), 5590

10.3390/s21165590

astro-ph.IM

http://creativecommons.org/licenses/by/4.0/

The main features of SG-WAS (SkyGlow Wireless Autonomous Sensor), a low-cost device for measuring Night Sky Brightness (NSB), are presented. SG-WAS is based on the TSL237 sensor --like the Unihedron Sky Quality Meter (SQM) or the STARS4ALL Telescope Encoder and Sky Sensor (TESS)--, with (LoRa, WiFi, or LTE-M) and solar-powered rechargeable batteries. Field tests have been performed on its autonomy, proving that it can go up to 20 days without direct solar irradiance and remain hibernating after that for at least mbox{4 months}, returning to operation once re-illuminated. A new approach to the acquisition of average NSB measurements and their (of the order of thousandths of a magnitude) is presented. In addition, the results of a new Sky Integrating Sphere (SIS) method have shown the possibility of performing mass device calibration with uncertainties below 0.02 mag/arcsec$^2$. SG-WAS is the first fully autonomous and wireless low-cost NSB sensor to be used as an independent or networked device in remote any additional infrastructure.

2 R. Buttry et significantly to the 3km s1dispersions observed When multi-epoch spectroscopy is available, one can binary systems via their observed accelerations (e.g., Ol- szewski et al. 1996; Martinez et al. 2011; Koposov et al. 2011; Mi- nor et al. 2019); indeed modeling of multi-epoch spectroscopic of Galactic binaries that indicate relatively high at low metallicity (Badenes et al. 2018). In some cases, the removal of suspected binary stars has a significant impact Triangulum II (Tri II) ultra-faint dwarf galaxy provides spectroscopy of six member stars, measured a 51,40 14km of solar units, and an extremely high dark matter density of et al. 2015). An by Martin et al. (2016) obtained a spectroscopic sample of 13 member stars, finding complicated kinematics in which a v=44,28 20km metallicity dispersion, supporting the conclusion that Tri star follow-up spectroscopy soon provided a time From high-resolution spectra obtained primarily to an- alyze chemical abundances, Venn et al. (2017) measured a change in velocity for one star ( Star46) of25km s1with respect to the initial epoch measured by Martin et al. (2016). Kirby et al. (2017) added additional epochs for this star, independently confirming its velocity variability and finding that, when they excluded the likely binary from their analysis, the velocity dispersion was circumstances leave the case for a dominant dark matter halo in Tri II resting on the indirect argument provided by its (Venn et al. MMT. Second, we combine with the previously data in order to obtain a multi-epoch data set that then lets us model the orbital parameters of the likely binary star. Our orbital solution includes an inference for the binary motion, allowing us properly to include this star in our analysis of Tri II's stellar full orbital solution, based on 34 independent velocity measure- ments taken over a two-year baseline (Koch et al. 2014). Here, we develop methodology for inferring orbital solutions with as few as two velocity epochs. The problem of finding orbital parameters for a binary system given a small number of radial velocity (RV) mea- surements has been undertaken previously by Price-Whelan et al. (2017) to create the JOKER . Like the JOKER , the binary model we present in this paper takes the approach of performing with likelihood function marginalized over some However, our method has the added modifications suchas the marginalization over inclination rather than for the calculation of semi-major axis), the ability to take non-trivial priors over binary parameters, and using parameter sam- ples for hierarchical models of binary this analysis as well as the calculation of a zero-point the two instruments. In section 3 we present our Tri II binary system and the overall Tri II kinematics in a hierarchical model building off the methods used. 2 DATA 2.1 MMT at the 6.5-m MMT Observatory on Mt. Hopkins, deploys up to 240 optical fibers, each with aperture 15 arcsec, over a field of diameter 1. We observed using the 'RV31' filter, isolating the wavelength range of 51505300at resolution R 34000. We observed five different Hectochelle fiber configurations, each centered on the published center of Tri II, allowing us to observe up to500unique targets, with many stars included in stars, we rely on narrow-band, metallicity sensitive CaHK ob- servations of Tri II. The observations follow a similar strategy and goal as presented by Starkenburg et al. (2017) in the Pristine survey but correspond to a single CFHT MegaCam field centered on Tri II. The field covers the full extent of Tri II and the integration amount to 4,800 s in total from the CaHK images. Additional (Laevens et al. 2015). All were observed in service mode by the CFHT observing staff between 2015-07-18 and 2016-02-13. of the images with ELIXIR by CFHT and de-trending; Magnier & Cuillandre 2004), the is calibrated onto the Pan-STARRS1 gP1andiP1bands and we directly use the Pan-STARRS1 photometry at the bright end, where the CFHT observations saturate. Finally, the photometric catalog is de-reddened using the Schlegel, Finkbeiner and Davis dust maps. Similarly to what was done for the Pristine Inner Galaxy Survey (Arentsen et al. 2020) the large number of science fibers available on Hectochelle allowed us to perform a broad selection of targets in the part of the CaHK color-color diagram that contains metal- poor stars. This allowed us to bypass any potential calibration ones we relied on in Starkenburg et al. (2017). In particular, we selected stars based on their color-magnitude diagram survey, we loosely select stars from the metal-poor 000, 1-14 (2021)Triangulum II 3 Figure1. are shown as small grey dots and follow a stellar locus that is produced by Milky Way, metal-rich stars. In this space, metal-poor stars are above this locus. Large black symbols correspond to stars within 4' of the center of Tri II, with likely members shown in red. One of those stars is shown here to be metal-rich and unlikely to be a true members. Our selection for Hectoshell gives the highest priority to stars within the dashed polygon that provides a loose selection of stars away from the metal-rich foreground properties. Once fibers are assigned to these we fill the rest of the fibers with random CMD-selected of their CaHK information. We processed all raw Hectochelle spectra using the CfA al. (2015), we then analyzed each individual spectrum by fitting a model based on a library of synthetic template spectra that span a regular grid in effective temperature, surface gravity and In addition to the parameters, we fit for line-of-sight velocity as well as several free parameters that specify the continuum shape and correct for present purposes we update our estimation of systematic er- rors associated with line-of-sight velocity and metallicity. For this task we use our entire catalog of MMT/Hectochelle observations of dwarf galaxies and globular clusters, including observations span- ning the years 2005 to 2020. This sample includes 12517 indepen- dent observations of 7906 unique stars, including 2501 stars with up to 13 individual measurements. We model the pair-wise veloc- ity and metallicity differences as a mixture of a Gaussian with an outlier model (see Section 4.1 of Li et al. 2019 and Section 2.2 of Pace et al. 2021).The final uncertainty ( vcalib) is treated as a systematic error ( vsystematic) plus a scaling parameter ( kv) We the velocity metallicity systematic Keck 2016. They used the 1200G grating, which achieves a with a similar spectral configuration as Kirby et al. Kirby et al. (2015) selected stars using Keck/LRIS (Oke et al. 1995) photometry. They chose a generous selection region in the CMD around the red giant branch as defined by the ridgeline of the globular cluster M92. Martin et al. (2016) used a similar with photometry (Laevens et al. 2015) from the Large Binocular Camera. In general, the field of Tri II is sparse enough that most candidate member stars in the field of the slitmask (or stellar age or metallicity). Kirby et al. (2017) were in quantifying radial velocity variability, not in finding new members. As a result, they designed their slitmasks to al. were measured in slightly different ways. Kirby et al. (2015, 2017) reduced the spectra with custom modifications to the spec2d data reduction pipeline (Cooper et al. 2012). They spectral templates observed with DEIMOS to the Tri used their own custom pipeline (Ibata et al. 2011) to reduce the spectra. They determined radial velocities from the mean wave- lengths of Gaussian fits to the Ca imaging spectrographs can experience radial velocity zero- point shifts if the star is not perfectly centered in the slit. This frame. All of the studies used in this work performed such Zero-point between Keck and MMT. Although our complete set of Tri II Keck observations is a concatenation of the Kirby et al. (2017) and Martin et al. (2016) datasets, the zero-point offset between the two appears consistent with zero (see Section 3.2 of Kirby et al. 2017). For this analysis, we have defined the offset as the MMT Tri II galactic center with radial velocity measurements in both MMT and Keck catalogues. Measurements for a given object and instrument are combined into a single weighted mean value. We then calculate the offset between the sets of combined assuming Gaussian likelihood function, which is described in terms of model pre- diction, model error , given data in the form of the likelihood function is the product of the velocity 000, 1-14 (2021)4 R. Buttry et the ith star's MMT weighted mean velocity and its Keck weighted mean velocity, We assume that these velocity differences are consistent with the zero-point offset between instruments vand that there is no dispersion in v, errors. Thus, the total likelihood is the product of the likelihoods of individual some objects have a very large velocity difference binarity. A convenient way to account for these objects is to construct a mixture model to treat them as outliers, a similar model is used in Li et al. (2019). The new likelihood for a given object is the sum of the original Gaussian likelihood and outlier model like- lihood weighted by the probability the object is an outlier. The new likelihood is written the outlier fraction and vrepresents the uniform(10 km s1,10 km s1) outlier fraction : uniform(0, 0.5) outlier model standard deviation p: uniform( 4 km s1, 20 km s1) We define the posterior as the Gaussian likelihood multiplied error is taken as the width of this offset posterior Chain Monte Carlo (MCMC) by the via the emceepython package (Foreman-Mackey et datasets observed around the Tri II galactic center, we find an offset value of v=011102 km s1. While this value could be determined using only our Tri II dataset, the large error means that the offset is not resolved within 1 km s1and the introduction of a large offset error can make it more difficult to resolve the Tri II velocity dispersion. To improve the error on our zero-point offset between the in- struments, we also include Keck/DEIMOS (Pace et al. 2020) (Spencer et al. 2018) observations from an results in an offset correction of v=133033km s1. We opt to use the offset that is calculated while including for the offset. Thus, we bring all observations onto a 133033 km s1from each velocity measurement obtained with Keck/DEIMOS. We do not propagate the offset error as the total error is dominated by the combination of systematics and random errors. 2.4 Membership To determine Tri II membership, we sigma clip our dataset at step removes objects whose observations differ from the Tri II measurements measurement errors, = 2 Tri II,2 obs. We use the proper 10 5 0 5 10 15 20 * 31 star 46 (binary) Tri II0204060 0 501 0 1 21 012 Figure2. Tri II center from Gaia eDR3 after a parallax cut (blue). The green point marks the Tri II proper motion value that is used to determine membership. The black triangle and star are the two stars with uncertain membership in Kirby et al. (2017). The red dots are the remaining members that have Gaia proper motion members are consistent with zero parallax (i.e, 3817 km =058006 mas yr1= 011007 mas with Gaia EDR3 find similar results (McConnachie & Venn 2020; Battaglia et al. 2022). While the velocity dispersion is unresolved, we use Tri II=4 km s1(Figure 3). Our resulting sample consists of 16 member stars, one of which is the binary star (GAIA ID which we refer to as Star46or MIC2016-46 in this work). There is one and the remaining 13 are observed by Keck light radius after a star-galaxy separation cut observed by the Subaru/Hyper Suprime-Cam (HSC; Miyazaki et al. 2018) and Pan- STARRS1 (Flewelling et al. 2020). The isochrone parameters and the half-light radius ( rh=2.50) are taken from Carlin et al. (2017). In Figures 2, 3, and 4, we indicate the locations of the binary called Star31in their work). Its membership is brought into question due to it being (1) far from the galactic center, (2) the most metal rich star in the sample (to the point of driving Tri !1from the galaxy mean velocity. However, the now available a member. MNRAS 000, 1-14 (2021)Triangulum II 5 400 300 200 100 0 100 vlos 46(binary star) star 31 Tri II MMT obs Keck mem obs MMT mem obs02040 0 20 390 385 380 375 3. Mean metallicities and velocities of all stars observed in MMT around the TriII field (black). The greenarea represents the TheTri II mean velocity from Kirby et al. (2017) used in membership cuts. Our final sample of Tri II members is plotted with two colors/markers to distinguish between Keck (blue circles) and MMT (red squares) measurements. The binary mean of the velocity observations. 3 METHODS 3.1 Binary Stars When a star is a part of a binary system, the orbital motion can give rise to periodic variability in the line-of-sight velocity. If one star is much brighter than its companion (e.g., a red giant with a main sequence companion), then a single-epoch spectrum may not show any evidence of variability. Conventional methods of velocity if the system is sparsely observed and/or it results in an inflated error in the combined velocity. Thus, binary stars Though the true line-of-sight velocity will vary with time, it will oscillate around the center-of-mass (systemic) velocity. In terms of the true anomaly (angular position in the orbital plane from the periastron direction), argument of periapsis (angle in the orbital plane between the ascending node and the periastron) and systemic velocity v0, a member of a binary star system velocity that varies with time according terms of the binary parameters, period P, semi-major axis a, inclination sini, and eccentricity (4) This is the model that we use to fit the observed radial velocitycurves. It is worth mentioning that instead of using as a free pa- rameter, we sample in Ph0, representing the phase, which is related to a constant called the time of periastron passage ,T, by a factor of 2P. We can relate andPh0using the eccentricity anomaly (6) We furthermore include a jitter parameter that acts as secondary error to account for velocity variability not due to periodic jitter and surface gravity for red-giant branch stars, where stars with logg.1can have excess jitter 1 km s1. However, the majority of our members, including Star46, exist on the lower inflate the galaxy velocity allowing us to use the previously defined Gaussian equation 1. We generate samples of the posterior on the orbital binary parameters using rejection sampling. This sampling is carried out over the course of 105iterations, with each with 106points before rejection. After rejection, the surviv- ing samples across all iterations are combined. The choice to can have multiple possible orbital period solutions, which have corresponding total mass and systemic velocity values, repre- sented as peaks in the posterior. This specific scenario can lead to untouched areas of the posterior when using MCMC. The priors are quite uninformative, which ensures that we can fully sample all the possible binary parameters. Note that there is a pre-defined mass range, such that the prior in total mass is loguni- form from log101 Motolog110 Mo, which is used to bound orbits and a Beta distribution for eccentricities. The parame- ters for the Beta distribution are =15and=15, resulting in a near uniform between 0 and 1 for eccentricities, Beta11515o. explicitlystate thepriors oneach ofthe alongside the resulting Star46binary parameter posteriors in Table 1. We are able to make further assumptions about the system by modifying the surviving posterior samples. We remove corresponding to close binaries where the pericenter is less than the stellar radius or where the mass of the companion star is negative. These stellar radii and masses are calculated python package (Morton 2015) to determine for the binary star of age 13 Gyr, [Fe/H] of -2.2, and distance modulus of 17.24 using the MESA Isochrones and Stellar Tracks (MIST) grid (Dotter 2016; Choi et al. 2016). Figure 4 shows the relevant isochrone fit on a color-magnitude diagram with Tri II members plotted. For the binary star Star46, we determine a mass ofMbin=0777 Mand stellar radius of rbin=0027 AUfrom the other Tri II members by finding the point on the to each member star using the same Raghavan et al. (2010) looked at 259 confirmed Hipparcos 000, 1-14 (2021)6 R. Buttry et al. 0.0 0.2 0.4 0.6 0.8 1.0 1.2 g0 Age=13.0Gyr star 46(binary star) star 31 Keck members MMT 33.2 33.3 33.4 33.5 33.6 RA [deg]star 46 (binary star) star 31 Keck members MMT obs Figure 4. Left: Color-magnitude diagram of Tri II stars within 4half-light radii ( rh=2.50; Carlin et al. 2017) taken by Hyper Suprime-Cam (HSC) and Panstarrs with a MIST isochrone overlaid. The location of the Tri II binary star and Star31is indicated by a star and triangle, respectively. The isochrone is defined by the following values: mM=1744, [Fe/H]=22, and Age=13 Gyr. Right: Sky positions of members and stars observed with MMT. The dotted lines represents the 1 and 2 half-light radii of Triangulum II. In both plots, the members are plotted with either a red square or blue circle for MMT and Star Orbital Prior Short Per. Long Per. PeriodP linearly decreasing in log( P) from 1 day to 1010days 1481,28 14days Beta(1.5,1.5) velocity v0 Normal(0, 474 km s1) 3876,08 12km s1 PhasePh0 uniform(0,2) of periapsis uniform(0,2) Normal( 0 k ms12 k ms1o 13,14 09km s109,11 06km s1 Semi-major axis a loguniform, bounds determined from mass 071,026 015AU sini uniform(0,2) incosi the 0.6 years (219.2 days) boundary. MNRAS 000, 1-14 (2021)Triangulum II logP=503 and standard deviation logP=228, wherePis in days. This is a more informative prior on the period than our default period of this paper. We present this distribution with the minor caveat that most of the detectable member stars of dwarf galaxies are not solar-type (main-sequence) stars, but rather red giant branch stars. We are able to infer the posterior under this prior by to the default prior. All of these assumptions shift the posterior away from shorter orbital periods, but as we will see in section posterior. As such, we present the results both before and after the application of the Raghavan prior. 3.2 Non-Binary Stars In the case of non-binary stars, we expect line-of-sight velocity in each measurement proportional to the observational error. Our goal is to determine a star's true velocity by combining multi- ple epochs of radial velocity measurements into one stellar velocity. A common approach to this is calculating the weighted the measurements weighted by the inverse of the squared errors. This approach has the benefits of being analytic and thus intensive. However, it has the downside of the error when dealing with potentially variable systems. In this section, we offer another method to calculate the true stellar velocity by finding its posterior defined Gaussian likelihood function, equation 1, and performed with 24 walkers, 1000 steps each, and 100 step burn- in. For a set of line-of-sight velocity observations fvigwith 4 max1figo. We also use Jeffrey's prior for the standard deviation of a Gaussian velocity measurement is taken as one standard of a singular Gaussian to the resulting posterior for an example Tri II member star with multiple Gaussian Fit method has the advantage in that it does the RV curve, more than the weighted mean. For comparison, model methods (Figure 6). Applying the binary model to the, assumed non-binary, member stars takes up significantly Over short time-scales, the observations can be These methods are only applied to members with more than one velocity measurement as neither would be informative for of trying to fit a Gaussian distribution to a single 2016-01-14 2016-05-12 2016-09-07 Time [kms1] Keck 450 425 400 375 350 325 300 v0 Fit Binary KCS2015-76 Left:RVcurveof Star The velocity measurements are colored by instrument with the same labeling as Figure 6. posterior on the mean reduce the posterior scatter. For the binary model, there is not enough information in a single epoch to restrict the possible binary solutions. 3.3 Mean Velocity and Velocity Dispersion The Tri II velocity dispersion can be inferred from the posterior using the same Gaussian likelihood function as in previous 1). We define another Gaussian model in terms of the mean velocity -vand the velocity dispersion vof the galaxy. The total likelihood is the product of the likelihood of the model on the stellar velocity of each member. For stars with multiple the measured velocities are combined using one posterior, such as for Star46, we must fit a to the systemic velocity posterior and the likelihood for that object jthGaussian in the mixture model with nGaussians. The error on jis taken as the standard deviation of the Gaussian jand the number of Gaussians in the GMM is determined by manually inspecting the v0posterior. Figure 7 shows the GMM fits to the posterior before and after the application of the velocity dispersion prior is log-uniform from log101005 km s1otolog101100 km s1o, which is equiva- lent to Jeffrey's prior. The minimum value corresponds to the case where the galactic dynamics are fully determined by visible matter (see Appendix A). We have chosen to find the posterior of log10v rather thanvto better sample smaller velocity dispersion values. Once again, we use a uniform prior from min1fvigo 4 max1figoas our prior on mean velocity. 4 RESULTS 4.1 Binary Orbital of this object, we find a reduced chi-squared of this star to be part of a binary system. For comparison, the MNRAS 000, 1-14 (2021)8 R. Buttry et al. 2015-10-07 2016-01-27 2016-05-18 = -384.6, 1.0 B = -383.2, 5.8 = -384.7, 2.1 2016-01-14 2016-05-12 = -378.2, 2.0 B = -376.7, 5.0 = -378.3, 2.5 2015-10-17 2015-11-15 = -390.4, 6.8 B = -387.6, 25.4 = -392.0, 13.9 2015-11-01 2015-12-15 = -385.6, 15.8 B = -383.8, 31.1 = -387.6, 16.1 2015-11-01 2015-12-15 = -383.5, 4.7 B = -378.6, 26.9 = -382.2, 24.2 2016-01-14 2016-05-12 = -387.1, 3.5 B = -385.6, 6.4 = -387.2, 3.7 KCS2015-76 2016-01-14 2016-05-12 = -383.2, 3.6 B = -381.7, 6.0 = -383.3, 3.0 2016-01-14 2016-05-12 = -383.0, 2.6 B = -381.4, 4.6 = -382.9, 1.8 KCS2015-116 2016-01-14 2016-05-12 = -381.2, 4.0 B = -379.9, 5.4 = -381.9, 2.9 KCS2015-91 2015-11-01 2015-12-15 = -400.8, 11.1 B = -394.8, 31.7 = -398.3, 17.1 2016-10-30 2016-11-09 = -386.0, 1.3 B = -386.0, 7.8 = -385.5, 1.6 2016-02-08 2016-06-29 = -382.1, 0.5 B = -382.0, 3.1 = -382.2, 0.3 KCS2015-106 -v: weighted mean; B: fitting a Gaussian distribution to the binary model v0posterior;G: assuming the observations are taken from a centered on a true systemic velocity. For comparison, the standard deviation of the observations is also listed ( ). The stars are identified by their Gaia ID when applicable. All values listed and on the y-axis are in km s1. The dates are in the format MIC2016-27, its velocity variations are more consistent with model than Star46. In this analysis, we use the 4 velocity measurements of this star from MMT (Table D1) and 5 from Keck (Table D2). There are two additional velocity measurements made in Geminin/GRACES data (Venn et al. 2017; Ji et al. 2020) that we have opted to not use zero-point correction that is more difficult to quantify (See Appendix B). After an initial pass with our model on Star46, we find that and0.8 years. We are able to eliminate the law both limiting the allowed semi-major axis values. Because the period modes are distinct, we re-run the sampling with a of surviving posterior samples. The resulting binary presented in this section are derived from the posterior after removing samples that correspond to a negative companion mass but before the application of the Raghavan prior. Figure 8 is a corner plot of the posterior of select parameters and shows two distinct peaks corresponding to long and short period solutions. We separate the posterior samples at the 0.6 year boundary and present the orbital parameter values taken from the truncated posteriors in Table 1. Figure 9 shows the RV curve for this star, as well as the orbital solutions that correspond to the surviving samples. The posteriors of the semi-major axis for both solutions are con- sistent with values less than a=2AU, indicating that this is a from the prior to be informative, but all of the samples cor- respond to stellar masses potentially on the main sequence. For comparison, the only other binary star in an ultra-faint Her-3 in the Hercules dwarf, has a smaller eccentric- ity (eHer3=018) and period ( (Koch et al. 2014). Plugging into equation 4 we find semi-amplitudes, s1and MNRAS 000, 1-14 (2021)Triangulum II 9 400 395 390 385 380 375 PDFRaghavan prior: w=0.30 400 395 390 385 380 375 PDFRaghavan prior: w=1.00 Figure 7. The posterior on the systemic velocity of the binary star and removing non-physical 8. Orbital parameters' posterior for the binary star derived from our binary model. There are two distinct modes corresponding solution, KHer3=1448082 km longer periods, but not more than 02 km s1, much smaller than the width of the posterior peaks ( 2 km s1). Reweighting posterior in favor of the longer period solutions and the 2015-12-30 2016-04-18 2016-08-05 2016-11-23 Time 9. Possible fits to the radial velocity curve of the binary star, using our binary model. The orange and purple lines correspond to solutions with orbital periods of approximately 0.4 years and 0.8 The dates are in the format larger systemic velocities, removing the multi-modality of the posterior (Figure 7). Even before applying the Raghavan prior, there is a strong preference towards the long period solutions as entirely in favor of the long period solution as the ratio solution, but we also explore the effect using the Raghavan prior has on the velocity dispersion. This binary star is only 1 of the 16 stars in our sample. Moe et of[Fe/H]=-2.1 impliesa closebinary fraction suggests that there are more close binary members, but be observed Triangulum II Dynamics While we have briefly explored applying the binary model to Star46, we unsurprisingly find that the best variation is to treat confirmed the binary star under the binary model and the remaining These additional variations made it more difficult to minimum value. Thus, we focus our final results on the best variation. The galaxy mean velocity posterior is consistent with the previ- ously found mean velocity of Tri II, hvi=381711 km s1 (Kirby et al. 2017) within 2, whereis the error on the mean velocity from the posterior. We present a mean velocity of hvi= 382307 km s1from this analysis. However, even in this best variation, we were unable to completely resolve the velocity posterior has a small peak at log10v=02(which 000, 1-14 (2021)10 R. Buttry et al. 1 0 1 2 log10(v/kms1) prior: =0.319(84th) 1 0 1 2 log10(v/kms1) prior: 10. Velocity dispersion posterior with and without the use of the Raghavan prior on Star46. The 16th, 50th, and 84th percentiles are marked by the blue, orange and green lines 11. Complete posterior Tri II mean velocity and velocity dispersion when treating the binary star under the binary model and other under the Gaussian fit with the Raghavan prior. back to the minimum value. We see that applying the Raghavan prior only slightly reduces this tail (Figure 10). Figure 11 shows the complete posterior with the binary under the Raghavan prior. This remaining tail of probability means we were unable to resolve the velocity dispersion of Tri II, and we instead present upper limits on v. Before applying the Raghavan prior, we find a 90% confidence limit of 27 km s1and 95% confidence limit of 35 km s1. Af- ter applying the Raghavan prior, we find a 90% confidence limit of 26 km s1and 95% confidence limit of 34 km s1. Due to the differences between these values being so small, we again choose to present the values determined using the more informative period prior. 5 CONCLUSIONS We have found orbital parameters for the Tri II binary star an orbital solution after Her-3 in Hercules Dwarf (Koch et al.2014). In doing so, we have demonstrated a method of accounting for binary star systems in velocity dispersion calculations that does notdemand theremoval of suchsystems. Wealso seethat usingour binary model when unnecessary, such as with non-binary stars, velocity remains an open question. Though the use of this model has offered an improved analysis of the dwarf galaxy, we were unable to resolve a velocity dispersion that would confirm the existence of a dark matter halo in Tri II. However, Kirby et al. (2017) found 1>>FeH available metallicity measurements for the dataset. They make clear that metallicity dispersion relied on the membership two motions of both stars confirm that they are Tri 1>>FeH 1/4o calculation while including the new MMT metallicities. We resulting dispersion where the difference after inclusion was . 003 greatly strengthens the case for Tri II's classification as a dwarf galaxy, indicating the system is embedded in a dark matter halo (Willman & Strader 2012). The choice to apply our binary model to only one star means that we are operating with a hard binary fraction of fbin006. A calculated from marginalizing over the binary parameter pos- terior. These likelihoods would allow for the sampling of binary fraction as a parameter alongside with mean velocity and forming a hierarchical model that builds from the set velocities. However, this approach introduces the posterior. For instance, removing binary samples based on hierarchical model to be an interesting problem and important step to explored in future work is supported in part by NSF grants AST-1813881 and AST-1909584. E.N.K. is supported by NSF grant AST-1847909, and he gratefully acknowledges support from a Cottrell Scholar award administered by the Research Corporation for Science Ad- vancement. EO is partially supported by NSF grant AST-1815767. NC is supported by NSF grant AST-1812461. MM is supported CB and CMD are supported by NSF grant work has made use of data from the European Space Agency (ESA) mission 000, 1-14 (2021)Triangulum II (DPAC, in particular the institutions participating in the Agreement. The Pan-STARRS1 Surveys (PS1) and the PS1 public science archive have been made possible through contributions by the In- stitute for Astronomy, the University of Hawaii, the Office, the Max-Planck Society and its participating in- stitutes, the Max Planck Institute for Astronomy, Heidelberg and the Max Planck Institute for Extraterrestrial Physics, Garching, the Queen's University Belfast, the the Space Telescope Science Institute, the National Aeronau- tics and Space Administration under Grant No. NNX08AR22G is- sued through the Planetary Science Division of the NASA Science Mission Directorate, the National Science Foundation Grant No. AST-1238877, the University of Maryland, Eotvos Lorand Univer- sity (ELTE), the Los Alamos National Laboratory, and the Gordon and Betty Moore Foundation. This research has made use of NASA's Astrophysics Data Sys- tem Bibliographic Services. This paper made use of the Whole Sky Database (wsdb) created by Sergey Koposov and maintained at the Institute of Astronomy, Cambridge by Sergey Koposov, Vasily Be- lokurov and Wyn Evans with financial support from the Science & Technology Facilities Council (STFC) and the European Research Council reported here were obtained at the MMT cultural role and reverence that the summit of Maunakea has always had within the indigenous Hawaiian community. We mountain. For the purpose of open access, the author has applied a version arising from this submission. DATA AVAILABILITY The processed MMT/Hectochelle spectra for Triangulum II targets are publicly available at the Zenodo database, A., et al., 2020, MNRAS, 496, 4964 Badenes C., et al., 2018, ApJ, 854, 147 Battaglia G., Taibi S., Thomas G. F., Fritz T. K., 2022, A&A, 657, A54 Bullock J. S., Boylan-Kolchin M., 2017, ARA&A, 55, 343 Carlin J. L., et al., 2017, AJ, 154, 823, DEEP2 DEIMOS Spectral Pipeline A., 2016, ApJS, 222, 8 Errani R., Penarrubia J., Walker M. G., 2018, MNRAS, 481, Instrumentation Engineers (SPIE) Conference Series Vol. 4841,Instrument Design and Performance for Optical/Infrared pp 1657-1669, H. A., et al., 2020, ApJS, 251, 7 Foreman-Mackey D., Hogg D. W., Lang D., Goodman J., 2013, PASP, 125, 306 Gaia Collaboration et al., 2021, A&A, 649, A1 Hekker S., Snellen I. A. G., Aerts C., Quirrenbach A., Reffert S., Mitchell D. S., 2008, in Journal of Physics Conference Series. p. ), R., Sollima A., Nipoti C., Bellazzini M., Chapman S. C., Dalessandro E., 2011, ApJ, 738, 186 Illingworth G., 1976, ApJ, 204, 73 Irwin M., Lewis J., 2001, New Astron. Rev., 45, 105 Ji A. P., Simon J. D., Frebel A., Venn K. A., Hansen T. T., 2020, VizieR Online Data Catalog, p. E. N., Cohen J. G., Simon J. D., Guhathakurta P., 2015, ApJ, 814, L7 Kirby E. N., Cohen J. G., Simon J. D., Guhathakurta P., Thygesen A. O., Duggan G. E., 2017, ApJ, 838, 83 Koch A., Hansen T., Feltzing S., Wilkinson M. I., 2014, ApJ, 780, 91 Koposov S. E., et al., 2011, ApJ, 736, 146 Laevens B. P. M., et al., 2015, The Astrophysical Journal, 802, L18 Li T. S., et al., 2019, Monthly Notices of the Royal Astronomical Society, 490, 3508 Magnier E. A., Cuillandre J. C., 2004, PASP, 116, 449 Martin N. F., et al., 2016, The Astrophysical Journal, 818, 40 Martinez G. D., Minor Q. E., Bullock J., Kaplinghat M., Simon J. D., Geha M., 2011, ApJ, 738, 55 Mateo M. L., 1998, ARA&A, 36, 435 Mazzola C. N., et al., 2020, MNRAS, 499, 1607 McConnachie A. W., 2012, AJ, 144, 4 McConnachie A. W., Cote P., 2010, ApJ, 722, L209 McConnachie A. W., Venn K. A., 2020, Updated proper motions for Local Group dwarf galaxies using Gaia Early Data Release ) Minor Q. E., 2013, ApJ, 779, 1142 Minor Q. E., Pace A. B., Marshall J. L., Strigari L. E., 2019, MNRAS, S., et al., 2018, PASJ, 70, S1 Moe M., Kratter K. M., Badenes C., 2019, ApJ, 875, 61 Morton T. D., 2015, isochrones: Stellar model grid package C. D., Correia A. C. M., 2010, Keplerian Orbits and Dynamics of Exoplanets. pp 15-23 Oke J. B., et al., 1995, PASP, 107, 375 Olszewski E. W., Pryor C., Armandroff T. E., 1996, AJ, 111, 750 Pace A. B., et al., 2020, MNRAS, 495, 3022 Pace A. B., Walker M. G., Koposov S. E., Caldwell N., Mateo M., Ol- szewski E. W., Bailey John I. I., Wang M.-Y., 2021, arXiv e-prints, A. B., Erkal D., Li T. S., 2022, arXiv e-prints, p. F., et al., 2011, Journal of Machine Learning Research, 12, A. M., Hogg D. W., Foreman-Mackey D., Rix H.-W., 2017, ApJ, 837, 20 Raghavan D., et al., 2010, ApJS, 190, 1 Simon J. D., 2019, ARA&A, 57, 375 Sohn S. T., et al., 2007, ApJ, 663, 960 Spencer M. E., Mateo M., Walker M. G., Olszewski E. W., McConnachie A. W., Kirby E. N., Koch A., 2017, AJ, 153, 254 Spencer M. E., Mateo M., Olszewski E. W., Walker M. G., McConnachie A. W., Kirby E. N., 2018, AJ, 156, 257 Starkenburg E., et al., 2017, MNRAS, 471, A., et al., 2011, PASP, 123, 1188 Tolstoy E., Hill V., Tosi M., 2009, ARA&A, 47, 371 Venn K. A., Starkenburg E., Malo L., Martin N., Laevens B. P. M., 2017, MNRAS, 466, 3741 Walker M. G., Mateo M., Olszewski E. W., Penarrubia J., Evans N. W., Gilmore G., 2009, ApJ, 704, 1274 MNRAS 000, 1-14 (2021)12 R. Buttry et al. Source HJD vlos[km s1] Venn et al. 2017 2457372.5 397120 Ji et al. 2020 2457372.5 396532 Ji et al. 2020 2457638.5 381550 Table B1. Additional GRACES observations of the Tri II binary, Star46. Walker M. G., Olszewski E. W., Mateo M., 2015, MNRAS, 448, 2717 Willman B., Strader J., 2012, AJ, 144, J. D., Avedo F. F., 2010, MNRAS, 406, 1220 APPENDIX A: LOWER LIMIT ON log10VPRIOR From the virial theorem, we know that the mass contained within the half-light radius of a dwarf galaxy in virial equilibrium can be described by... M1 R the galaxy's velocity dispersion and Gis the gravi- tational constant. Cis a proportionality constant. Assuming that all of the mass is accounted for by the total stellar mass, the becomes true. M1 R these, we can solve for is estimated that Tri II has a total luminosity L400 Land a half-light radius R1234 pc(Laevens et al. 2015). Stellar mass to light ratios usually exist in the rage U05to30 ML1 and the proportionality constant is in the range C2to4. Taking the lower limit for Uand the upper limit for C, we find that vcan go as small as 005 km s1. This value will serve as the lower limit of our velocity dispersion prior and corresponds to a system where the dynamics are fully determined by visible matter. APPENDIX B: GRACES We briefly explored how the inclusion of the Venn et al. (2017) and Ji et al. (2020) would impact the binary orbital parameter posterior. For this, we do not apply an offset to the velocities measured in GRACES because of the lack of overlap in observation fields between GRACES and either MMT or in the main text (see Table B1 for the individual spectra taken in December 2015 are reduced using measured velocities for the same epoch (See Section 2 of Venn et al. 2017 and Section 2 of Ji et al. 2020). Figure B2 shows the inferred systemic velocity posterior when including GRACES observations. Under the default period prior, the posterior takes on the same bimodal shape as seen in Section 4.1 regardless of which set of GRACES observations is used. 2015-12-30 2016-04-18 2016-08-05 2016-11-23 Time et al. 2015 Ji et al. 2020Figure B1. Radial velocity curve of Star46including GRACES epochs. solutions with orbital periods of approximately 0.4 years and 0.8 withmodesat3876,08 12km modes are s1 when including either set of GRACES observations. Applying the period distribution found in Raghavan et al. (2010), there is between the resulting unimodal posteriors with the usingVenn et al. (2017)'s main text are well within may likely vanish after properly accounting for any of which of the two GRACES sets is preferable for this analysis. Rather than exploring this avenue that would likely yield only a marginal improvement, we opt to not include the of this star in this paper. APPENDIX C: METALLICITY affects the metalicity disperison of Tri II. From repeat mea- surements in the MMT catalog, we only consider an Fe/H error < 0.4 dex as being good-quality observations. At lower signal-to-noise, we see a bias for higher metallicities in dwarf galaxy members and applying this selection removes maximum likelihood estimation to fit a Gaussian distribution a Fe/H zero-point offset between the instruments which was de- termined using the non-outlier offset model (setting outlier fraction to zero in equation 2) with the two stars that are in the overlap dexto be added to the Kirby et al. (2017) metal- MNRAS 000, 1-14 (2021)Triangulum II 13 400 395 390 385 380 375 370 365 PDFRaghavan prior: False Incl Ji et al. 2020 Incl Venn et al. 2017 400 395 390 385 380 375 370 365 PDFRaghavan prior: True Incl Ji et al. 2020 Incl Venn et al. 2017 FigureB2. epochs. The posteriors from including Venn et al. (2017)and colored grey. 0 5 10 15 20 25 star46 star40 Figure C1. Fe/H measurements v. signal-to-noise of the Tri II members weighted mean Fe/H values for each star. licities. We note that a weighted mean is not the preferred way of combining metallicity measurements and improvements to shows Fe/H measurements for the stars observed in MMT. Kirby et al. (2017) found 1>>FeH Kirby et al. 2017). The effect of including the MMT observations can D: DATA TABLES 0.00 0.25 0.50 0.75 1.00 1.25 1.50 et al. 2017 Including MMTFigure C2. Metallicity dispersion likelihood distributions calculated when using only the Keck data (Kirby et al. 2017) for Tri II (black) and when also including our Tri II MMT measurements (red). The dotted lines mark the maximum likelihood values. MNRAS 000, 1-14 (2021)14 R. Buttry et al. This paper has been typeset from a T EX/LATEX file prepared by the author.Gaia ID KCS2015 MIC2016 Member RA (ICRS) Dec (ICRS) HJD (days) S/N vlos1km s1o[Fe/H] Teff True 33.4272917 36.0950281 2457683.72 8.67 -386.9 0.7 -3.00.3 4760226 True 33.4272917 36.0950281 2457688.77 10.68 -385.5 0.9 -3.20.4 4684255 True 33.4272917 36.0950281 2457711.69 6.07 -384.4 1.7 -3.20.6 4772453 True 33.4272917 36.0950281 2457685.74 4.74 -383.4 1.7 -3.00.7 4783495 1.70.9 65 46 True 33.33975 36.1659431 2457374.8 8.84 -400.4 0.7 -2.50.3 4823159 2.00.4 65 46 True 33.33975 36.1659431 2457683.72 6.78 -405.9 0.8 -2.90.3 4535215 0.90.4 65 46 True 33.33975 36.1659431 2457688.77 5.08 -405.1 1.0 -2.00.7 5416588 2.00.7 65 46 True 33.33975 36.1659431 2457711.69 2.88 -386.4 2.1 -0.80.9 6811855 106 40 True 33.3189583 36.17939 2457374.8 27.68 -382.8 0.6 -3.50.1 436787 106 40 True 33.3189583 36.17939 2457683.72 21.62 -381.8 0.5 -3.30.2 450795 106 40 True 33.3189583 36.17939 2457688.77 23.86 -381.9 0.5 -3.30.1 450582 106 40 True 33.3189583 36.17939 2457711.69 13.11 -381.9 0.7 -3.60.2 4253135 106 40 True 33.3189583 36.17939 2457685.74 12.85 -382.2 0.8 -3.20.3 4597195 1.10.5 Table D1: Individual observations of Tri II members taken with MMT Hectochelle. This table is a subset of the full MMT catalogue. MNRAS 000, 1-14 (2021)

2108.10867

Rachel Buttry

Rachel Buttry, Andrew B. Pace, Sergey E. Koposov, Matthew G. Walker, Nelson Caldwell, Evan N. Kirby, Nicolas F. Martin, Mario Mateo, Edward W. Olszewski, Else Starkenburg, Carles Badenes, Christine Mazzola Daher

Stellar kinematics of dwarf galaxies from multi-epoch spectroscopy: application to Triangulum II

10.1093/mnras/stac1441

astro-ph.GA

http://creativecommons.org/licenses/by/4.0/

We present new MMT/Hectochelle spectroscopic measurements for 257 stars observed along the line of sight to the ultra-faint dwarf galaxy Triangulum II. Combining with results from previous Keck/DEIMOS spectroscopy, we obtain a sample that includes 16 likely members of Triangulum II, with up to 10 independent redshift measurements per star. To this multi-epoch kinematic data set we apply methodology that we develop in order to infer binary from sparsely sampled radial velocity curves with as few as two epochs. For a (spatially unresolved) binary system in Tri~II, we infer an orbital solution with period m~ days$ , semi-major axis and a systemic velocity $ -380.0 1.7 m~km ~s^{-1}$ that we then use in the analysis of kinematics. Despite this improvement in the modeling of binary star systems, the current data remain insufficient to resolve the of Triangulum II. We instead find a 95% confidence upper limit of $sigma_{v} lesssim 3.4 m ~km~s^{-1}$.

Conformal Cyclic Cosmology, and the role of gravity in quantum state reduction. In this Introduction, we give a brief roadmap for this everything is a conviction that the conventional interpretation of the of Bell's inequality - that the laws of physics must either be indeterministic or violate local causality - lies at the heart of the seemingly insurmountable obstacles synthesising our theories of quantum and gravitational physics. Penrose has himself written about this on many that the violation of Bell's inequality seems at least in con ict with the spirit of even if the violation does not imply overt superluminal signalling. There is a that we haven't yet got to the bottom of this profound dilemma. In Section 2 we propose an alternative interpretation of the experimental violation of violation of the Statistical Independence (SI) assumption. This assumption has been thoroughly misunderstood by the quantum foundations community. Sometimes it is claimed that violation of SI would mean that sub-ensembles of particle pairs in a Bell experiment are not sta- tistically leading to apocalyptic claims about the end of objective Sometimes it is claimed that violation of SI would mean experimenters are not able to choose freely between measurement settings, that their brains have been subverted by the particles being measured! Both of these claims are incorrect. In Section 2 we note that SI can be violated (without any implications about the of sub-ensembles or experimenter free choice) if there exists a non-trivial measure on state space, such that on points where = 0, states are deemed inconsistent with the assumed laws of physics. An example where such a non-trivial measure occurs naturally is in dynamical systems where states are evolving on some dynamically invariant subset Iof state space. If this subset is fractal, so that the dynamical system is chaotic, the subset has non-integer dimension and is Hausdor. The fractal gaps in Iwhere= 0 correspond to specic counterfactual worlds which are inconsistent with the putative laws of physics. According to a theorem of Blum et al [3], many of the geometric properties of such fractal sets are algorithmically undecidable, i.e. we discuss, Penrose has often speculated that a quantum theory of gravity should be non- computable. In order to distinguish traditional approaches to the violation of SI, here a violation of SI based on a non-trivial measure is referred to as supermeasured [10]. Fig 1 provides a owchart for the logical progression of these ideas. In Fig 1, the word ealistic is assumed synonymous with Section 3 a theory of quantum physics is outlined based on the notion that the universe is evolving on such a cosmological invariant set IU- Invariant Set Theory (IST) [18], [19]. In this theory, state-space trajectories have a fractal structure homeomorphic to the set of p-adic Such a theory is very conceptually dierent from quantum mechanics - not least it implies that the quantum state vector has an ensemble intepretation. More dramatically, it implies that cosmology is not to be thought of as emergent from primitive laws on Planckian scales - the conventional approach where smaller is seen as inherently more fundamental. Rather, the essential properties of quantum physics would arise from the geometry of the invariant set of the undivided universe. This is what we mean in the title by a op-down [7] theory of evidence from cosmology is there for such a fractal invariant set IU? In Section 4, we discuss the Hawking Box, which Penrose discusses in a number of his books. Under we show that the asymptotic end state of an isolated conguration of matter is a measure-zero fractal invariant set. Moving to a realistic model of the universe, we discuss Conformal Cyclic Cosmology might similarly exhibit an fractal invari- 2Figure 1: A owchart for some of the ideas discussed in this paper ant set. On the other hand, at least under circumstances, we know that quantum theory is a supremely successful theory of physics. It is therefore important to understand the connection between IST and quantum theory. To do this, in Section 5 we reframe quantum states as Hilbert vectors, not over the continuum C, but over a set Cpof complex numbers Aei, rationals of the form m=p for integers mandp, wherep>> 1. Such a granular Hilbert state space can be shown to provide a symbolic description of ensembles of trajectories in ap-adic neighbourhood of IU, with possible links back to earlier combinatoric ideas of Penrose's on spin networks. Just as IUis gappy and lacks the property of counterfactual deniteness, Cpis gappy and lacks the arithmetic closure properties of a full eld. A small modication to the eld equations of general relativity is proposed to make it consistent with IST. This modication may potentially help explain the phenomenon of dark matter. In the Conclusions section, we discuss some future plans to develop IST. Roger Penrose was internal examiner for my PhD in general relativity; my supervisor was Roger's close colleague and friend Dennis Sciama. On the occasions when I was lucky enough to take part 3in small discussion groups in Roger's oce, I had a strong sense that the notion of was being violated - that I was sitting in a very privileged frame of reference close to the centre of the universe. 2 Bell's Theorem and the Statistical Independence Assump- tion Our failure to synthesise quantum and gravitational physics may ultimately be due to the fact that we still haven't resolved the tension that exists between relativity theory and quantum theory, posed by the experimental violation of Bell inequalities. Specically, general relativity is a deter- ministic locally causal theory, whilst, it is commonly asserted, deterministic locally causal theories of quantum physics must satisfy the (CHSH version of the) Bell +C(0;1) +C(1;0)C(1;1)2 (1) and thus be inconsistent with experiment. As usual, we imagine a source of entangled prepared so that the total angular momentum of any pair of particles is zero. The spins of the particles are measured by remote experimenters Alice and Bob who can each choose to orient their measuring apparatuses in one of two ways (relative to a reference direction): are referred to as X= 0,X= 1 (for Alice) and Y= 0,Y= 1 (for Bob). C(X;Y ) denotes the correlation in spin measurements for ensembles of particles, as a function of these detector settings. In the discussion below, X0= 1 ifX= 0 and vice versa , and similarly for Y. However, in order to show that a locally causal deterministic theory satises (1), then it is also necessary to assume that theory satises the Statistical Independence =() (2) Here we presume that each pair of entangled particles is described by a supplementary variable , often referred to as a hidden variable - though in the uncomputable type of theory proposed here, there is no need for to be hidden. For each in a conventional hidden-variable theory, a value of spin (here1) is dened for each of the four values of XandY.() is some probability density on the space of hidden variables. The assumption (2) is typically required to ensure that when the individual correlations in (1) are estimated from separate sub-ensembles of particle pairs (as necessarily happens in any real-world experimental test of the Bell inequality), then the hidden variables associated with are statistically equivalent to one another. In another interpretation, the assumption (2) is justied on the basis that experimenters' choices are not somehow subverted by the variables. A theory which violates (2) is referred to as [12]. As such, theories have typically been treated with contempt, and peremptorily both of these interpretations of (2) are incorrect - they are not implied by a violation of (2). The reason the interpretations are incorrect relates to the denition of the quantity (). As discussed in [10], may depend both of factors which are under the control of the factors which are not. To make this explicit, we replace (2) by Bell(jXY) =Bell() We now treat as a probability distribution under the control of the exper- imenter: for example, it depends on the nature of the prepared state and whether the experiment 4was performed on Monday or Tuesday, and in London or New York. , by contrast, is a on state space, determined by the laws of physics. Points where = 0 describe states which are inconsistent with the assumed laws of physics. The actions and choices of much how freedom they have to do as they please, must nevertheless be consistent with the laws of physics. With this notation, the Statistical Independence assumption is actually (3) not (2). Indeed, if (2) is respected, then it must be the case that (jXY) =() (4) is violated. A theory where (2) is violated (with the new denition of ) will continue to be called . By contrast, a theory where (4) is violated will be called supermeasured [10]. In the next Sections we describe a supermeasured theory of quantum physics. 3 Fractal Invariant Sets and the nonlinear dierential (5) wherer >0. Starting from any initial r2R+, it can easily be shown that rwill 1:r= 1 is a xed-point attractor of (5). We can add a second equation _= 1 (6) where 0 < 2, in which case the asymptotic evolution of the dynamical system (5) and (6), in the 2-dimensional state space described by the polar coordinate system ( r;), will be the circle r= 1. This is an example of a limit cycle. Such xed points and limit cycles are examples of invariant sets: if a point lies on an invariant set, then its future evolution will always lie on it, and its past evolution will have always lain on it. Conversely, for a point which does not lie on the invariant set, its future evolution will never will lie on it, and its past evolution has never has lain on it. The invariant measure associated with (5) is equal to 0 when r6= 1. A more general class of invariant set is the fractal (or 'strange') attractor - an invariant set of a class of chaotic nonlinear systems. Perhaps the most famous of these is the Lorenz attractor (see Fig 2) associated with the Lorenz equations (7) Such fractal attractors have a couple of important properties. Firstly they have zero volume in state space. That is to say, if an initial non-zero state-space volume is evolved in time using (7), the volume will asymptotically shrink to zero. This means that there are processes which velocity elds to be convergent. In the Lorenz equations, this process is described dissipation in the model equations, and is associated with one or more 2: The fractal invariant set of the Lorenz equations (7) Lyapunov exponents. As such, the invariant measure of the attractor is zero with respect to the Lebesgue measure of the full state space in which the invariant set is embedded. To create a fractal structure, a simple state-space volume must be stretched and folded as it shrinks. In a chaotic system, the stretching is associated with instabilities with positive In the Lorenz system the negative Lyapunov exponent associated with dissipation dom- inates in magnitude over the positive Lyapunov exponent associated with attractors provide another link to one of Penrose's key ideas. In [21] Penrose discusses a toy universe whose states correspond to pairs of nite sets of polyominoes. Dynamical evolution in this toy universe is a non-computable rule, drawing on the result that there is no procedure for deciding when a polyomino set will tile the plane. Penrose writes: What about the actual universe? Well, I have argued [elsewhere] that there is missing in our physics. Is there any reason from physics itself to think that there might be something non-computable in this missing physics? Well I think there is some such reason to believe this - that the true quantum gravity theory might his celebrated book The Emperor's New Mind [20], Penrose asks whether the fractal Mandel- brot Set is an example of a non-computable geometry. This possibility motivated the seminal book Complexity and Real Computation [3] (co-authored by Steve Smale one of the pioneers of chaos theory). Blum et al set about answering the question of whether membership of the Mandelbrot Set is algorithmically decidable. Their argumentation applies equally to the fractal attractor Aof a chaotic system. We consider a putative on the real numbers, which takes as 6input a point xin the state space of the chaotic system, and halts if x2A. Blum et al's Theorem implies that such an algorithm does not exist if Adoes not have The very denition of a fractal is one whose Hausdor dimension is not an integer. Hence we can conclude that indeed the fractal invariant sets of chaotic systems are notion was further developed by Dube [6], who showed that many of the classic of computing theory (e.g. the Post Correspondence Problem) can be recast in terms of geometric properties of fractal attractors. We now come to the crucial point of this paper. Instead of dening a dynamical system by its dierential equations (such as (5) or (7)), could we instead dene it by the invariant set geometry, or, equivalently, by the invariant measure ? In the case of (5), this would be trivial. In the case of (7) it is not trivial, precisely because the geometry is However, there are techniques to characterise such fractal invariant sets topologically through what is called [8]. Here one attaches symbolic labels, e.g. to the two lobes of the Lorenz attractor. The strings of symbolic labels associated with trajectory segments can be used to characterise the attractor topology. In this regard, it is worth recalling that Hilbert states in quantum mechanics are fundamentally symbolic in nature [24] (we do not need to know precisely what is the an alive and dead cat in order to dene a superposed state of dead/alive cats). equations (e.g. (7)) should merely be thought of as local representations of the global invariant set geometry. In particular, relative to this invariant set geometry, initial conditions and dynamical equations can no longer be considered independent of one another. In [19] a model of quantum physics, referred to as Invariant Set Theory (IST), was described, where state-space trajectories on an invariant set IUhave a fractal structure isomorphic to the p- adic integers. Here Udenotes the (undivided [4]) universe, and it is assumed Uis evolving precisely on its fractal invariant set IU. The fractal structure of trajectories on IUis illustrated in Fig 3. Using the notion of Section 2, it can be inferred from number-theoretic properties in IST, (see [19]) that, for 0 =)(jX0Y) =(jXY0) = 0 (8) That is to say, if the measurement settings for an actual Bell experiment on a particle pair with hidden variable areXandY, then by denition the world characterised by the triple ( ;X;Y ) lies onIUand is associated with 6= 0. In this situation, counterfactual worlds characterised by the triples ( ;X0;Y) and (;X;Y0) do not lie on IUand are associated with = 0. This means that it is possible to satisfy (jXY) =(jX0Y) =(jXY0) =(jX0;Y0) (9) while at the same time =Bell(jX0Y) =Bell(jXY0) =Bell(jX0;Y0) (10) Hence IST violates Statistical Independence (3) but not statistical independence (2). That is to say, the sub-ensembles of particle pairs which the experimenter chooses to perform his or her experiments are statically equivalent. Similarly, insofar as his or her choices do not fall foul of the laws of physics, experimenters are free to choose how to set their measuring apparatuses as they like: if they want to set their instruments based on the wavelength of light from distant quasars, so be it - there is nothing in IST stopping them from so doing. Because IST is non-computable, one cannot determine a priori which putative states lie on IUand which not: this can only be asserted 7Figure 3: A schematic of the local fractal state-space structure of the invariant set IUin Invariant Set Theory [19]. a) An ensemble of trajectories decoheres into two distinct clusters labelled aand
a. Under a second phase of decoherence, this trajectory, itself comprising a further ensemble, decoheres into two further distinct clusters labelled band
b. These clusters correspond to measurement eigen- states in quantum mechanics. b) Under magnication, a trajectory segment is found to comprise a helix ofptrajectories at the next fractal iterate. c) Top: a cross section through the helix of trajec- tories comprises p(herep= 16 ) disks coloured black or grey according to whether that to the acluster or the
acluster. Bottom: each of these pdisks itself comprises pfurther disks coloured black or grey according to whether each trajectory evolves to the clusters bor
b. The fractal set of disks is homeomorphic to the set of p-adic integers. Reprinted with permission from Palmer T. N., Proc. R. Soc. A. 476, 20190350 (2020), under a Creative Commons 4.0 posteriori . Of course, it is one thing to say that IST violates Bell's inequality, it is another thing to say it violates it as does quantum theory. We address this latter point in Section 5. Violations of Statistical Independence are sometimes criticised as being 'ne tuned' [25]. How- ever, by itself this criticism is meaningless: with respect to what measure or metric is the tuning ne? For example, the invariant measure (r) of the dynamical system dened by (5) is the From this point of view the precise value r= 1 is far from being ne tuned, it is generic! Similarly, relative to the Hausdor measure of IU, points which lie on IUare and not ne tuned. Indeed, as mentioned, IUis locally isomorphic to the set of p-adic integers. Relative to metric, a point which does not lie on IUis necessarily at least p-distant relative to a point which does lie on IU- even for pairs of points which, from a Euclidean extremely close. Hence, relative to measures and metrics which are naturally linked to the geometry of IU, violations of (3) are not ne tuned at all. And if the violation of (3) is not ne tuned, it cannot be called conspiratorial, by denition. It is worth examining these issues from the perspective of the example introduced by Bell himself [2]. Suppose the detector settings are determined by two pseudo-random number In turn suppose the output of a PRNG is determined by the parity of the millionth digit of the input number. Consider a situation (a given ) where the parity of both PRNG inputs was even. According to the discussion above, keeping xed, a putative world where one of the PRNG inputs was odd would not lie on IU. This may violate one's intuition that the parity of the PRNG input is (to quote Bell [1]) nlikely to be the vital piece [of information] for any distinctively dierent purpose, i.e. it is otherwise rather useless; that is to say, one's intuition may be that the rest of the universe couldn't care less about the parity of the PRNG inputs. Indeed, if these PRNG inputs weren't tied to a specic type of quantum experiment - a Bell experiment - then, perhaps the universe couldn't care less. However, when the PRNG inputs are linked to a quantum Bell experiment, then the universe does care and the counterfactual where the parity is perturbed, does not lie on IU. Even Bell (at the end of that his intuition in this matter may have been faulty. Is there any evidence that the universe may be evolving on a fractal invariant set in space? We address this point in the next two sections. 4 Hawking Boxes and Conformal Cyclic Cosmology We start with a description of a hypothetical system which Penrose calls the Hawking Box [11] [22], a vast (say galactic scale) box of matter whose walls are perfect mirrors, allowing no information to cross in or out. If we wait unimaginably long periods of time, then contents of the Box will vacillate irregularly between distributions of matter dominated by one or more black holes, and a eld of almost pure Hawking radiation with no black hole Penrose now considers the phase space/state space PHawking of such a system. This focus on state space is central to the ideas developed in the Sections above. A schematic of PHawking is shown in Fig 4. In the part of state space containing a black hole, Penrose assumes there will naturally be information loss, manifest as a convergence of the state-space velocity eld (associated with a con uence of state-space trajectories). By contrast, in the part of state space, quantum state reduction (in regions denoted R) will cause state-space velocities to be divergent. It might seem strange at rst sight to link the notion of state reduction to divergent ow. However, the construct merely illustrates the notion that, as normally envisaged, state 4: The Hawking Box. Loss of phase-space volume occurs when a black hole is present. This may be (partially) balanced against regions where phase-space volumes grow due to quantum state reduction - here shown by the symbol R. Adapted with permission from S. Hawking and R. Penrose, The Nature of Space and Time, Princeton Science Library, 2000. Copyright 2000 by Princeton University Press. occurs when the quantum state evolves into one of a number of macroscopic possibilities. The divergent ow describes these dierent macroscopic possibilities. In the Everettian imagines some kind of branching process where the quantum state literally splits into dierent worlds. Although Everettians like to think of this branching process the notion of a single state splitting into multiple states is the very epitome of inde- terminism in nonlinear dynamical systems theory. In the framework developed in this paper, which is very much motivated by nonlinear dynamical systems theory, the divergent process illustrated in Fig 4 is assumed to represent an exponential divergence of trajectories like that in the left-hand panel of Fig 3, where two trajectories never precisely lie on top of each other (contrasting with polynomial Everettian divergence). In Penrose's own writings, it is never quite clear which of these two paradigms he believes is most relevant in his collapse model. Here - consistent with - we will assume that trajectory segments never intersect or lie on top of each other and that both the divergent and convergent ow describe exponentially increasing or between neighbouring of these divergence and convergences, the phase-space ow is not Hamiltonian. It is strongly time asymmetric. Penrose postulates that the divergent and convergent processes are in balance so that ultimately there is phase-space volume preservation of the ow. However, here I we consider a possibility more consistent with chaos theory: that the convergence eect associated with black-hole dynamics (gravitational collapse more generally) wins out overall compared with the divergence eect associated with quantum state reduction. In this situation, the volume of ow will ultimately shrink to zero. This does not mean that the ow will shrink to a xed point or limit cycle. The asymptotic invariant set of the Hawing box could be a zero volume (measure zero) fractal. Penrose discusses the Hawking Box in the context of the Liouville theorem. For a the Liouville theorem demands that phase space volumes are conserved. In the asymptotic case where the phase-space volume of the Hawking Box is zero, this theorem will be obeyed. However, during the period of transient ow where volumes are shrinking, a generalised form of the Liouville theorem will be satised where probabilities, but not necessarily phase-space volumes, are conserved. This is an important point: loss of state-space volume does not necessarily imply a violation of conservation of probability. The convergence of the state-space ow during periods when the Hawking Box contains a black hole illustrates a feature of so-called nal space-time singularities. Such nal-time singularities are assumed to be dominated by Weyl curvature. The information in the degrees of freedom associated with the Weyl curvature is assumed lost when the black holes evaporate. Initial singularities, by contrast, can be expected to be free of Weyl curvature. As Penrose has explained in many of his books and papers, this asymmetry is needed to explain the Second Law of fundamental dierence between initial and nal singularities is encoded in Penrose's 'Weyl Curvature colleague Paul Tod, Penrose realised that it was not actually necessary to make such a hypothesis. If the asymptotic end state of an epoch of the universe is pure zero rest-mass radiation whose origin was Hawking evaporation, then the universe will be conformally equivalent to an (initial) space-time singularity with nite (potentially zero) Weyl curvature. This led Penrose to abandon the Weyl Curvature Hypothesis, and instead postulate that our universe has a conformal cyclic structure [23]. During CCC, the smooth matter elds at the beginning of a cosmological epoch clump gravitationally (and chaotically), ultimately forming black holes. These then evaporate leaving a pure radiation eld (some decay of mass is postulated to ensure this nal radiation eld has zero rest mass and therefore satises conformal nal radiation eld can be considered conformally equivalent to a new radiation dominated Big Bang. Let us consider the phase portrait of CCC. Like the Hawking Box, if gravitationally wins out over divergence due to quantum state reduction, then we can assume that the asymptotic volume of the cosmological state-space invariant set of CCC is zero. Again, like the Hawking Box, this does not mean we end up with a xed-point invariant set. Instead, we will assume that this asymptotic end state is a zero-volume fractal invariant set in cosmological state space. The key idea here is that the existence of a non-computable fractal invariant set in space ISTis consistent with multi-aeon cosmology, of which CCC is an example. 5 Complex Numbers and Invariant Set Theory Clearly quantum theory is an extraordinarily successful theory. Yet here we are suggesting it be replaced by a deterministic supermeasured theory based the assumed fractal geometry of the invariant set IUof the universe. Just as general relativity theory is a fundamentally dierent theory to Newtonian gravity, the predictions of general relativity theory are the same as in the limit of large c. What is the equivalent weak limit of Invariant Set Theory, from which quantum theory would be emergent? 11In quantum mechanics, states are represented by Hilbert vectors over the continuum eld Cof complex numbers, a mathematical object of great elegance, as Penrose has himself pointed out on many occasions (indeed Penrose's Twistor Theory is profoundly underpinned by the properties of complex holomorphic functions). However, is it possible the founding fathers of quantum mechanics were beguiled by such simplicity and elegance, ignoring the physical interpretational by complex Hilbert Space? I believe they were. The fact that Cis a eld means that all conceivable arithmetic operations on elements of Clead to elements which still lie in C. However, a basic feature of a fractal set is that it is fundamentally gappy: arbitrary operations acting on a state which lies on an invariant set do not automatically lead to a state which lies on the invariant set, and indeed, typically they do not. Based on its fractal structure, IST suggests an alternative to C; specically, it admits as physical only those wavefunctions =AeiwhereA2can be written in the form m=p andis of the form 2n=p wherem,nandpare integers. Let Cpdenote the set of such rational complex numbers. Since Cpis dense in Cin the limit p!1 , it will be hard, for large p, to detect between quantum mechanics and a theory based on Cpdirectly. However, clearly Cis a singular limit of Cp: just as fractals are inherently gappy, Cpis gappy relative to C, no matter how large is p. As discussed in [19], elements of Cphave constructive representations as operators acting on bit strings of nite length p. To give a simple example, let p= 4 and consider the bit (11) then it is easily shown that i2S=S. In this way i1=2is a multiplicative 4th root of unity and can be written ei=4.iitself is a square root of minus one. The formalism is readily generalised with bit strings of arbitrary length p, ifpis a power of 2. In this way, Hilbert vectors (and associated tensor products) over the set Cpcan be interpreted [19] as nite bit strings (and Cartesian products of bit strings). The bits themselves are interpreted as the symbolic labels of trajectories on IU, where the labels correspond to the clusters to evolve (see Fig 3). These clusters correspond to measurement eigenstates of the relevant quantum observable. In this way, a Hilbert state over Cprepresents an ensemble of trajectories (in ap-adic neighbourhood on IU), where the ensemble is represented by the symbolic ensembles of trajectory segments in IST can be described by complex Hilbert entanglement correlations in IST will be, for all practical purposes, identical to those in quantum theory (except in the limit where the number of entangled qubits is comparable with pwhence entanglement correlations will become classical). Hence, for small numbers of IST violates Bell's inequality exactly as does quantum theory. Consider the algebraic properties of Cp. to Cp=2, to Cp. But what about addition? From Euler's formula for complex numbers in Cwe have eia+eib= 2 (12) We now need to be mindful of Niven's Theorem [17, 13]: Theorem 1. Let=22Q. Then cos =2Qexcept when cos= ifaandbare rational fractions of 2 , then cos(ab) and hence irrational. That is to say, the sum of eia+eibdoes not belong to Cpeven if the summands do. Unlike say the Gaussian integers, Cpis not closed under both multiplication and addition. Because of this, Cpseems uninteresting from a mathematical point of view. However, from a physical point of view it is extremely interesting: strict superpositions of exponentials do not correspond to ensembles of trajectories on IU. Moreover, (12) a spatial grid where :::;k (x
x);kx;k (x+
x);:::are all rational. Then, again by Niven's theorem, the square of the amplitude term on the right hand side will typically not be that subtraction of this type is undened in Cp. Conversely, if we choose
xso that the complex number on the right hand side of (13) belongs to Cp, then the individual on the left hand side do not. However, in the limit
x!0, subtraction of this type denes the momentum form of the quantum wavefunction. Basing our theory of quantum physics on Cpwe have an immediate explanation of why the position and momentum of a particle are not One can put it like this. If the wavefunction describes an ensemble of trajectories where the position of a particle is measured, then a counterfactual ensemble of trajectories where the momentum of the same particle (same ) is being measured, does not lie on the invariant set IU. Is the loss of algebraic closure too heavy a price for such a simple physical might be if Hilbert vectors were fundamental. However, in a theory based on Cp, these vectors clearly are not fundamental - they merely represent ensembles of trajectories on IU. The fundamen- tal object here is the invariant set geometry itself. The Hilbert vectors merely describe of this geometry. Algebraic closure arises at this deeper geometric level. For example, it is possible to add and multiply two p-adic integers and the result is a p-adic integer [14]. It would be interesting to reformulate Penrose's Twistor Theory in terms of Cp- twistor space, is, after all, a form of state space with complex structure. This might allow incorporation of Penrose's ideas about state reduction and even CCC, into an approach to laws of physics which has been the focus of Penrose's research since he was a young man. 6 Towards a Top-Down Theory of Quantum wisdom suggests that a theory of quantum gravity is needed primarily to explain the structure of space-time on the Planck scale. This ts naturally with the philosophy of Method- ological Reductionism, where an attempt at providing an explanation at a deeper level a focus on smaller-scale entities. However, the discussion above suggests this philosophy may be profoundly we propose some possible building blocks for a theory of quantum gravity, whose fun- damental equations describe the phase-portrait geometry of the assumed fractal invariant set of a multi-aeon model of the universe, such as Conformal Cyclic Cosmology. Such a theory would provide a clear delineation between quantum processes on the one hand, and gravitational pro- cesses on the other. Specically, what we describe as quantum processes are those associated with the fractal structure of the invariant set, as illustrated in the middle and right-hand panels of Fig 3. These are linked to positive local Lyapunov exponents on the invariant set (i.e. locally diuent trajectories) as an otherwise isolated system interacts with its environment. By contrast, what we describe as 'gravitational' processes are associated with larger-scale of the invariant set, such as the clustering process illustrated in the left-hand panel of Fig 3. Such inhomogeneities link to negative Lyapunov exponents on IU(i.e. associated with Together, all the fundamental forces of nature, including gravity, will be describable in terms of the total (stable and unstable) geometry of IU. As yet, no clear picture of this global geometry has emerged. This delineation between quantum and gravity suggests there is no reason to think of gravity as some force subject to the rules of quantum eld theory. In IST there is no such thing as a graviton. Moreover, there is no reason for gravity to be a 'witness of quantum entanglement' [5] [16]. In IST, the convergence of trajectories into a nite number of clusters is, by our denition, an inherently gravitational process. This, therefore, is consistent with Penrose's notion of induced collapse of the wavefunction (R in Fig 4). The key dierence with is that in IST, there is no continuous 'wavefunction' from which collapse takes place. Instead, by describing quantum states as vectors over Cprather than C, then Hilbert states simply describe nite ensembles on IU. In this sense, the measurement problem is rather trivially solved in IST. IST suggests that one should generalise the equations of general relativity to G(M) space-time corresponding to our trajectory on IU, andM0denotes to other trajectories on IU.pis a measure on IUwhich 'smears out' the singular delta function (M0M ) of general relativity onto a p-adic neighbourhood of IU- Einstein always believed the right hand side of his eld equations was somewhat provisional. That is to say, consistent with the primacy of the geometry of IU, the curvature of space-time feels the energy- momentum of matter elds in space-times associated with neighbouring trajectories on IU, in addition to the energy-momentum of matter elds on the trajectory corresponding to our space- time. This extra source of space-time curvature would manifest itself as if there were some form of dark matter in our space-time (which, according to IST, there isn't). Additionally, ), the non-singular nature of p(M0;M) would likely prevent the blow-up of space- time curvature at space-time singularities in M- a requirement for any putative theory of indeed the quantum-matter elds on IUare associated with the (positive of) the geometry of IU. Then (14) describes a relationship between the of space-time and the p-adic geometry of IU. If this programme of research is successful, we will have reduced allof physics to geometry. An experimental test of IST [9] relates to the fact that the bit strings associated with complex Hilbert states have nite length p. This implies that only a nite number of qubits can be entangled and still show quantum entanglement statistics. This inherent limit of quantum coherence might be intimately related to Penrose's own ideas on gravitational decoherence, given the discussion about gravity and the heterogeneity of IU. 147 Penrose, more than anyone, has taught us not to be afraid to think out of the box. This has served me well in my professional career. Here I have returned to the topic I studied for my doctoral thesis under the guidance of both Roger Penrose and Dennis Sciama, and have suggested some out of the box ideas for synthesising quantum and gravitational physics. These ideas are motivated by ideas that Penrose himself has promoted: notably cosmic cyclic cosmology, and gravitationally induced quantum state reduction. This has led to a proposal for a future quantum theory of gravity which completely deviates from the usual philosophy reductionism, where smaller is more fundamental. The key is the of the invariant set IUof the undivided [4] universe. Relative to such a geometry, the basic concepts of determinism and causality which underpin general relativity, still hold, and yet Bell's inequalities can be violated as they are in quantum theory. Although we have formulated a small modication to the eld equations of general relativity, by far the larger change will be to the equations of quantum physics, where Hilbert states are no longer seen as fundamental, but are merely statistical descriptions of strings of symbolic labels given to state-space new framework for quantum physics is indeed as dierent from standard quantum mechanics as general relativity is from Newtonian gravity as Penrose wrote that it would have to be. In particular, quantum mechanics emerges as a singular limit of Invariant Set Theory IST. Of course, the ideas proposed here need considerable further development. So far, IST has only been developed to describe Hilbert States. A crucial question concerns the invariant set interpretations of continuous quantum wavefunctions (x;t) in physical space-time. This is needed to explore the concept of wave-particle duality, not discussed in this paper, from an invariant set perspective. The next paper on IST will develop this topic. A second paper will explore the extent to which the revision (14) of the Einstein eld equations can indeed explain the phenomenon of dark matter. It is an honour to write this paper to celebrate Roger's 90th Birthday. Data Availability Statement Data sharing is not applicable to this article as no new data were created or analysed in this J.S. Bell. Free variables and local causality. In `Speakable and unspeakable in quantum mechan- ics'. Cambridge University Press, 1993. [2] J.S. Bell. Speakable and unspeakable in quantum mechanics . Cambridge University Press, 1993. [3] L. Blum, F.Cucker, M.Shub, and S.Smale. Complexity and Real Computation . Springer, 1997. [4] D. Bohm and B.J.Hiley. The Undivided Universe . Routledge, 2003. [5] S. Bose, A.Mazumdar, G.W.Morley, H.Ulbricht, M.Toros, M. Paternostro, A.Geraci, P.Barker, M.Kim, and G.Milburn. A spin entanglement witness for quantum gravity. Phys.Rev.Lett , 119:240401, 2017. 15

2108.10902

Tim Palmer

T.N. Palmer

Bell's Theorem, and Conformal Cyclic Cosmology: A Top-Down Approach to Quantum Gravity

To appear in Special Edition of AVS Quantum Science, celebrating Roger Penrose's 90th Birthday

quant-ph gr-qc

http://creativecommons.org/licenses/by/4.0/

This paper draws on a number of Roger Penrose's ideas - including phase-space flow of the Hawking Box, Conformal and gravitationally induced quantum state reduction - in order to propose a radically unconventional approach to quantum gravity: Invariant Set Theory (IST). In IST, the fundamental laws of physics describe the geometry of the phase portrait of the universe as a whole: quantum process are associated with fine-scale fractal process with larger-scale heterogeneous geometry. With this, it becomes possible to explain the experimental violation of Bell having to abandon key ingredients of general relativity: determinism and local causality. Ensembles in IST can be described by complex Hilbert states over a finite set $mathbb C_p$ of complex numbers, where $p$ is a large finite integer. The quantum mechanics of Hilbert spaces is emergent as a singular limit when $p ightarrow infty$. A small modification to the field equations of general relativity is proposed to make it consistent with IST.

base for the standard model of cosmology, the LCDM model, in which Lrepresents the cosmo- logical constant and plays the role of dark energy, while CDM is the Cold Dark Matter, another sort of peculiar matter which is gravitating but non-baryonic. According to recent 95% of the total energy budget of the universe is composed of these dark sectors [6]. However, the cosmological constant approach, although promising, suffers a conceptual problem [7], [8], in addition to the famous coincidence problem [9]. Such problems have motivated the construction of alternative scenarios to the cosmological constant dark energy. In some of these scenarios, the dark energy is dynamic. In other scenarios the dark matter and dark energy are in- teracting [10], [11], while in some others they are non-interacting [12], [13]. There is also a class of models that considers unification between the two sectors. In these models, the two dark sec- tors are represented by a single dark fluid where the late transition from DM to DE can smoothly occur [14], [15]. Chaplygin gas model and its generalizations are examples of these models [16], [17]. There is also the new barotropic dark fluid model proposed by Elkhateeb [18]. As viscosity is a physical property of fluids, cosmological viscosity attracted attention. Dissi- pative processes in cosmology have been discussed since the work of Zel'dovich [19] and Misner [20] in 1967. In his work in 1971, Weinberg [21] claims that vanishing bulk viscosity for a fluid is just an exception rather than a rule. Many authors also discussed the viscous bulk pressure effects relating the cosmological particle production, e.g., [22]-[28]. Viscous pres- sure inducing cosmic expansion also brings attention. A study made by Giovannini [29] declines the possibility of the existence of an accelerated phase driven solely by viscosity, where the ten- sor to scalar ratio exceeds the observational limit. He concluded that the viscous modes are only tolerable as a subleading component of a dominant adiabatic mode from a different physical ori- gin. Some authors also discussed removing instabilities in the fluid power spectrum by as the physical origin for entropy perturbation, which in turn smoothes out anisotropies in the universe [30]-[32]. Cosmological viscosity also included studying the early universe and entropy generation via cosmic inflation [33]-[39], and bulk viscosity driven late universe accel- eration [40]-[44]. Recently, Brevik and Normann [45] inferred symmetry between the early-time and present-day viscous cosmology by using a bulk viscosity function proportional to the Hubble parameter H. Victor et al. [46] discussed the bulk viscosity effects and matter production as an alternative for the description of the late times behavior of the observable universe. Alleviation of the H0-tension between direct and indirect measurements is also a motivation for studying vis- cous models [47]-[49]. Many authors also studied The viscosity of the cosmic unified dark Other aspects concerning dissipative cosmology can be found in [56]-[61]. In this work, we aim to study the evolution of the universe under the effect of bulk viscosity using the latest cosmological observations. We adopt the new unified viscous dark fluid model introduced by Elkhateeb in a previous work [52]. The dark fluid introduced has an equation of state with two power-law asymptotes, consequently can interpolate between dust and dark energy at early and late times. The added dissipative effects are in the form of bulk viscosity with a constant viscosity coefficient. Shear was considered negligible due to the isotropy of the universe. We implement recent observational datasets including local H0measurements, Type Ia Supernovae (SNe Ia), observational Hubble data (OHD), Baryon Acoustic Oscillations (BAO), and Cosmic Microwave Background (CMB) to perform robust analysis. The paper is organized as follows: Sec. 2 represents the theory of viscous fluid-based cosmology. Sec. 3 constrains some of the model parameters based on the asymptotic behavior of the model. In sec. 4, recent datasets used to validate the analysis procedure are presented. In sec. 5, we compare models based on model selection statistics in the form of the Akaike Information Criterion (AIC) to select the favorable model according to the observational data used. In sec. 6, we study the due to our fluid model. Sec. 7 for estimation of the present-day viscosity coefficient of the cosmic fluid. We finally concluded in sec. 8. 22. Fluid Based Cosmological Model Einstein's equation for GR is given (1) where Rmnis the Ricci curvature tensor, Ris the scalar curvature, gmnis the space metric and Tmn is the energy-momentum tensor given (2) where ris the energy density of the cosmic fluid, pis its pressure, and hmn=UmUngmnis the projection tensor to the 3 space orthogonal to the fluid element, where in comoving coordinates the four-velocity units with c=1, the metric in the standard FRW cosmology is given (3) where a(t)represents the scale factor, kis the curvature constant for the space with the three distinct values1, 0, and 1, represent respectively a spatially open, flat, and a closed with qandfare the usual polar and azimuthal angles of spherical co- ordinates. The coordinates (t;r;q;f)are the comoving coordinates. Considering our universe is filled with several components: radiation, baryons, and dark fluid, each with pressure piand energy density ri, Einstein's equation (1) together with the metric (3) will lead to the (5) where we consider flat geometry and units with 8 pG=1. The conservation equation =0; (6) where Tm n(i)is the energy momentum tensors for the effective pressure of dark fluid component pfldis considered to consist of an adiabatic part paplus a viscous (7) where J=Um;m=3His the expansion rate of the fluid with H=.a ais the Hubble the coefficient of bulk viscosity that arises in the fluid and is restricted to be positive. For the adiabatic part, pa, we adopt a barotropic pressure that can asymptote between dust at the early time and DE at the late time [18]. It has the (8) where g,d,n, and mare free parameters. As mentioned in [18], this form of adiabatic pressure has the advantage that it enables interpolation between different powers for the density, which allows for smooth phase transitions during the universe evolution. It also has the advantage of having a general equation of state (EoS) for DE that enables the cosmological constant as special case. The bulk viscosity coefficient is considered to take the form z(rfld) =z0rn fld; (9) 3where z0andnare constants. In this work we adopt a constant bulk viscosity coefficient z(r) = z0, so that n=0. This assumption is attractive due to its mathematical simplicity while it is physically acceptable. Accordingly, our viscous dark fluid has the effective equation of (10) 3. The Model Parameters The dark sector of the universe is supposed to consist of viscous dark fluid with an EoS two power laws characterizing dust at early time and dark energy at late time. The asymptotic behavior of the model can shed light on the relationship between its parameters, which reduces the number of degrees of freedom of the model. In doing so, we define the (12) so Accordingly, the conservation equation for the fluid reduces (14) is also dimensionless, while [z0] = [H]. DUDF has to mimic dust for large ~rand DE for small ~r, which means that d~r=dtwill be in proportion to ~rfor large ~rand tends to zero for small ~r. Accordingly, the parameter nhas to be positive. Restricting malso to be positive, relation (13) tends asymptotically to the following form for large (15) For small~z0, the first term in the above equation dominates for (nm)1. Under this condition, Eq. (15) reduces (16) Comparing this with Friedmann equation for dust, (17) we directly get n=m+1;g=d: (18) As a (20) has the dimensions of H2;h ~di = [r] = H2 . 4In a previous work [52], we introduced the fundamental properties of the DUDF model and estimated its parameters using observations of different cosmological phenomena such as today's value of deceleration parameter, the redshift value of transition, and the age of the universe. In this work, we seek to use the available cosmological datasets to constrain the model parameters robustly. 4. Data Analysis 4.1 Observational Data We now aim to constrain the DUDF model using observational data. We relied on the datasets from local H0measurements, Type Ia Supernovae (SNe Ia), data (OHD), Baryon Acoustic Oscillations (BAO), and Cosmic Microwave our analysis, the cosmological model is specified by the total relative density, Wtot(z) (21) The radiation term Wrconsiders the standard thermal history into account with today's value Wr0 given by [62] Wr0(z) (22) where the current value of photon density, his the constant defined through H0=100h km s1Mpc1, and Neff=3:046 is the effective number of neutrino species [63]. 4.1.1 SNe Ia Type Ia supernovae (SNe Ia) have a known brightness whence can serve as standard candles and can be used for measuring cosmological distances. Pantheon Sample [64] is the most complete updated compilation of SNe Ia. Confirmed 276 SNe Ia from Pan-STARRS1 Medium Deep Survey at 0:03<z<0:65 are combined with previous available SNe Ia samples from other surveys to form a sample of 1048 SNe Ia spanning the redshift range 0 :01<z<2:3. Data for the m(z)is given instead of the distance modulus mas the absolute magnitude Mof SNe Ia degenerates with H0. Observations of the apparent magnitude of SNe Ia, m(z), are related to the distance modulus m(z)through the relation m(z) (23) where the absolute magnitude Mis considered constant for all supernovae of Type Ia due to the standard candle hypothesis, and DL(z)is the luminosity distance given by DL(z) (25) is the Hubble distance, and the function E(z)is given by E(z) =H(z) H0: (26) 54.1.2 Observational Hubble Data In addition to the SNe Ia observations, we also considered observational Hubble data (OHD) from CC and BAO measurements. CC techniques [65] allow direct information about Hubble function as they are probe for direct measure of the differential age,Dz Dt, of passively whence H(z) =1 1+zDz Dt. Accordingly, They allow for measurements of Hubble function. However, CC is a good tool for Hubble function measure at redshifts 2 [66]. In this work, we considered the updated list compiled by Maga ena et al. [67] which contains 31 points from CC measurements plus 20 points from clustering. However, we excluded the points of Alam et al. [68] as they are used in the BAO dataset. 4.1.3 Local we included in our analysis the local value of Hubble parameter H0provided by Riess et al. 2020 (R20) results [69]. 4.1.4 BAO Anisotropic BAO analysis can measure the Hubble function H(z)and comoving angular DM(z)due to the BAO shifts perpendicular and parallel to the line of sight. The comoving angular diameter distance DM(z)for a flat universe is defined to be DM(z) (27) As changes to the energy density alter the radius of the sound horizon, BAO measurements really constrain the combinations where rdis the radius of sound horizon at the drag zdis the redshift of the drag epoch, wb=Wbh2is the current physical density of baryons, and wg=Wgh2is the current physical density of photons. To calculate the redshift at the drag epoch we considered the recent formula fitted through the machine learning approach by Aizpuru et is accurate up to 0:001%, where wm=Wmh2is the current physical density of matter. We used the results of BAO measurements from Sloan Digital Sky Survey (SDSS-III) provided by Alam et al. [68] for redshifts z=0:38 , 0 :51, and 0 :61. The c2 BAOfor these data is constructed as c2 BAO= (30) where -Dobsis an abbreviation rfid d=147:78 Mpc. C1 SDSSis the inverse covariant matrix of the data. Observational data and covariant matrix are available publicly at the SDSS-III website [71]. 64.1.5 CMB CMB data provide information about the entire history of the universe up to the last In our analysis we use results of the compressed likelihood of the CMB power spectrum given in Chen et al. [72] based on the Planck-2018 TT,TE,EE+lowE data release. We use the results of R;lA;wbto constrain our model parameters. the CMB shift parameter, and lA= the acoustic scale at last scattering, with zis the redshift at the photon decoupling epoch calculated using the recent formula of Aizpuru et al (34) which is accurate up to 0:0005%, and rs(z) =ris the comoving size of the sound horizon at zgiven the CMB data is constructed the inverse covariant matrix of the CMB data, where (CCMB)i j=sisjDi j (38) with Dis the correlation matrix given by [72] D=0 BB@1:0 0 :460:66 0:46 1 1 :01 CCA(39) 4.2 Parameters Estimation The probability distribution function, posterior, p(jD)for the model parameters given dif- ferent observational datasets Dis given by Bayes' theorem: p(jD) (40) where p()is the parameters prior, and p(D), known as the evidence, is the probability of observing the data. The evidence p(D)is constant and can be set equal to 1 since, after all, 7we got the observations. The probability p(Dj)is the expected distribution of data given the parameters, likelihood, and is given (42) with d represents each data group. To constrain DUDF model we used the public Markov Chain Monte Carlo (MCMC) emcee code [73]. The parameter space were chosen to be the three basic cosmo- logical parameters: Hubble constant H0, today's baryon density Wbh2, and today's dark matter density Wch2, plus the three model parameters: m,ed, and z0, in addition to the nuisance parame- ter represents the absolute magnitude M, The posterior is constructed from the likelihood and a flat prior on all parameters. In table 1 we summarized the priors of our parameters space. Table 1: Priors of the model parameters. H0 Wbh2Wch2m 104ed z0 M [65, 80] [0.015, 0.028] [0.1, 0.13] [0.5, 2.5] [1.3, 2.7] [0, 6] [-20, -18] The previously mentioned five different datasets are used to constrain the parameters of our dissipative unified dark fluid, DUDF, in addition to the particular case of non-dissipative unified dark fluid, UDF. Similar analysis is also performed on the LCDM model for comparison. In Fig. 1 we plot the 68% and 95% two-dimensional contours for the three models. In the Fig. we also show the one-dimensional marginalized posterior distribution for the parameters of the models. From the Fig. we can notice the correlation between the different parameters. In is a correlation between H0and the dark fluid parameters mand~d. There is an the parameters z0and~dtoo. One can also notice that the viscosity parameter z0is by the data. The parameter ~din the DUDF model is poorly constrained, as well, due to its degeneracy with z0. In table 2 we list the 68% C.L. of the parameters of the three models under 2: 68% C.L. parameters of cosmological DUDF UDF LCDM H0 70:020:56 70:250:57 0:0223560:00015 0:12270:0013 2:0910:13 2:2990:063 z0 2:9+1:6 2:30:0 Model Comparison In table 3 we display the c2 min-values for models under consideration. We can see that between the three models, the DUDF model has the lower value of the c2 min. 868 70 0.0227 ob0.120 0.125 oc1.5 1: The marginalized 68% and 95% C.L. contours of the parameters of the three models. Shown also the one dimensional marginalized posterior distributions for the 3: Summary of the information criteria results DUDF UDF LCDM c2 min1059 :68 1059 :78 1065 :96 AIC 1073 :68 1071 :78 1073 :96 DAIC 1:9 0 2 enables the determination of the best-fit parameters of a model. are not sufficient for deciding which model is the best, especially if the models have a differ- ent number of parameters. Many model selection statistical methods are proposed in the context of cosmology in the literature. The most commonly used is the Akaike information criteria (AIC) [74]. It is more robust as it considers the number of parameters of each model. The AIC is defined for a given model as AIC=2lnL+2p; (43) 9where pis the number of free parameters of the model. The model which minimizes the AIC is considered the better model. As a result, the second term in the criteria imposes a penalty against the extra number of parameters. In general, the AIC criterion is inclined to select the model that better fits the data [75]. In the second row of table 3 we display the values of AIC for the three models at hand. We can note that between the three models, the UDF model has the lower value ofAIC. We can then conclude that the UDF model is the most favored by the data. However, the difference between AICs of different models is that making sense. According to [76], if Di=AIC iAIC min, then model iwhich has Dibetween 0 and 2 is considered as substantial on the level of empirical support, while models with Dibetween 4 and 7 has considerably less support. For Di>10, model imay be omitted in future the UDF model minimizes the AIC value, the Di's for other models are measured to this model. The results are included in the last row of table 3. We can see from the table that DAICfor the DUDF model is <2. Following [76], we conclude that the DUDF model shows a of empirical support and is not worth more than a bare mention of evidence against the UDF model. 6. Cosmological Evolution of the Model In this section, we study the cosmological evolution due to our fluid model and discuss the con- sistency of the model with The Hubble Parameter One of the most important physical parameters in cosmology is the Hubble parameter. It describes the expansion history of the universe. In the upper panel of Fig. 2 we plot the evolution of the Hubble parameter due to DUDF and UFD models against the OHD. We also displayed results from the LCDM model for comparison. In the lower panel of the figure, we present the relative ratio 1.5 2.0 2.5 3.0 3.5 1 +z1.001.05H/H LCDM Figure 2: Upper panel: evolution of the Hubble parameter. Lower panel: Ratio between Hubble parameter of the fluid models to that of the LCDM model. 10The graph shows that the model represents the observational data very well. From the lower panel, we can see that our fluid follows the history of the standard cosmology at early times. As time passes, its deviation from the LCDM model is in the range of 2%. 6.2 Supernovae SNe Ia measurements using SNe Ia lead to the discovery of the accelerating expansion of the universe. This was first discovered in 1998. SNe Ia are referred to as standard candles, and their different characteristics extend to their stellar age distribution. The upper panel of Fig. 3 represents the distance modulus calculated from the fluid models, DUDF and UDF, against obser- vations from Pantheon data. Results from the LCDM model are also displayed for comparison. The lower panel of the figure shows the ratio of the distance modulus from each of the fluid mod- els to that of the LCDM model. We can see from the Fig. that the fluid models excellently fit the SNe Ia data and highly follow the LCDM results, where the deviations are of the order of 0 0.5 1.0 1.5 2.0 LCDM Figure 3: Upper panel: distance modulus from different models vs Pantheon data. Lower panel: ratio between the distance modulus from the fluid models to that of the LCDM model. 6.3 BAO Observations BAO matter clustering provides a standard ruler for length scale and can be used to measure the expansion history of the universe. In Fig. 4 we compare theoretical calculations for Hubble function and comoving angular diameter distance from the three models at hand with the SDSS-III data. The Fig. shows that both DUDF and UDF models well represent the data. 6.4 CMB Observations The compressed likelihood of Planck data contains three quantities: the CMB shift parameter R, the angular scale of the sound horizon at last scattering lA, and the baryon density wb. In Fig. 5 we compare the theoretical results of lAfrom the three models at hand against the the Planck18 dataset. The Fig. shows that both DUDF and UDF models fit the DR12 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 4: A comparison between the observational BAO data and theoretical calculations from the three models at hand. 1080.0 1082.5 1085.0 1087.5 1090.0 1092.5 1095.0 1097.5 5: A comparison between the observational lAfrom Planck18 and the theoretical the three models at hand. 6.5 Effective Equation of State parameter In standard cosmology, the universe is isotropic and filled with a perfect fluid. The equation of state (EoS) describing this fluid is characterized by a constant parameter wthat is the ratio of fluid pressure to fluid density. For the LCDM model, today's universe is characterized mainly by two components: dark energy with w=1 and dark matter with w=0. On the other hand, the effective EoS parameter of DUDF model is calculated using relations 12(18) and (20) in (10) where we (45) In Fig. 6 we present the behavior of the EoS of DUDF and UDF models. The Fig shows the 1.0 10.0 2 3 4 5 6 789 1 UDF Figure 6: Behaviour of the EoS of fluid models. ability of the fluid for a smooth transition from dust at early times, where w(z) =0, to dark energy domination at late time. 6.6 Deceleration parameter is an important parameter in cosmology. Evolution of the expansion of the universe can be discussed through this parameter. Deceleration parameter is defined as q(z) =a(z)a .a2: (46) It was believed that the expansion of the universe is decelerating due to A study of the apparent magnitude of distant supernovae reveals that the universe is instead ac- celerating. However, cosmological observations indicated that this acceleration is a recent phe- nomenon, which was explained by the domination of dark energy in the late times. Fig. 7 displays the behavior of the deceleration parameter due to DUDF and UDF models. The Fig. also shows results from the LCDM scenario for comparison. One can see from the Fig. that the deceleration parameter due to the fluid model follows the standard evolution of the universe. The fluid can play the role of the cosmological constant in the late times, so able to switch the deceleration expansion to an accelerating one. The Fig also shows that viscosity slows the transition to the accelerating phase. The transition redshift increases from ztr0:78 for DUDF model to ztr0:8 for the UDF where the fluid is non-viscous. These values ofztrare consistent with the results of the observational analysis made in [77] and [78]. Similarly, the viscosity affects the current value of the deceleration parameter. Its value increases from q00:53 for DUDF model to q00:52 for the UDF model. These values of q0agree with the results of [79]. 131.0 10.0 2.0 3.0 4.0 7: Evolution of deceleration parameter. 6.7 Density of the indicate that the universe is flat with a mean energy density equals to its Friedmann equations enable the determination of the energy budget of the universe at any time. For any component xof the universe's fluid, the energy density is given by Wx(z) =rx(z) rcrit(z); (47) where rxis the density of the x-component, and rcritis the critical density given by rcrit(z) =3cH2(z) 8pG: (48) In Fig. 8 we displayed the density of different components of the fluid model vs the LCDM model. We can see from the Fig. that the density parameter of the unified fluid matches the dust den- sity due to the LCDM scenario at early times while gradually following the dark energy density to the future. 7. Estimation of the Viscosity of the Cosmic Fluid Different authors estimated the viscosity of the cosmic fluid to be z0107Pa.s. However, its value is highly model-dependent. As we mentioned above, the best model due to the Akaike criteria is the perfect fluid model, UDF, though the result of DAIC for the viscous model DUDF indicated that this model is substantial. We can consider the mean value of z0due to this model as an estimated value for the viscosity of the cosmic fluid. From table 2 we see that the mean value of z0is 2:9km s1Mpc1. Multiplying this by c2=8pG, we get an estimated value for the viscosity of the cosmic fluid as 5 :04106Pa.s. 8. Conclusions We considered a standard FRW flat cosmology with several energy source components. The two dark sectors of the cosmic fluid are replaced by a dissipative unified dark fluid. The EoS of this fluid can asymptote between two power laws. As a result, it enables fluid to smoothly UDF LCDMFigure 8: Evolution of density early and late times. The dissipative effects are described by a bulk viscosity with a constant coefficient. We have implemented data analysis for the DUDF model as well as its special perfect fluid case, UDF, using recent observational datasets from R20, SNe Ia, OHD, BAO, and CMB. Similar analysis is also performed on the LCDM model for comparison. We found that between the three analyzed models, the DUDF model has the lower c2 min-value. We then relied on model in the form of AIC to compare different models. We found that the UDF model has the minimum AIC value, with the conclusion that the data favored the perfect unified fluid model. On measuring Dito the lowest AIC model, the UDF model, we found that the DUDF model has a difference of less than 2, meaning that the DUDF model is substantial on the level of and is not worth more than a bare mention of evidence against the UDF model. Given this result, we estimated a present-day value for the viscosity of the cosmic fluid as 5 on, we studied the evolution of the cosmic fluid through the Hubble parameter, the distance modulus, BAO data, acoustic scale at last scattering lA, the effective equation of state of the dark fluid, the cosmic deceleration parameter, and the cosmic density parameter. Such a that the fluid model highly matches observations and follows the standard the evolution history. The evolution of the deceleration parameter proved to reflect the well-known behavior of such parameter, with a present-day value that agrees with the results obtained by many previous authors. The evolution of the EoS of the fluid manifests its ability for a smooth transition between dark sectors at early and late times. The density parameter indicates again that the fluid can smoothly evolve from one dark sector to the other, with a dust density that matches the LCDM model during the high redshifts, and a dark energy density dominates we conclude that our unified fluid model provides a viable framework for describing the authors would like to thank A. Awad, A. El-zant, and W. El Hanafy for valuable A.G. Riess, A.V . Filippenko, P. Challis, et al., AJ116(1998) 1009-1038. [2] S. Perlmutter, G. Aldering, G. Goldhaber, et al., APJ517(1999) 565-586. [3] D.N. Spergel, R. Bean and O. Dore, et al., ApJS 170(2007) 377. [4] A. Blanchard, M. Douspis, M. Rowan-Robinson and S. Sarkar, A&A 449(2006) 925. [5] S. Wienberg, Rev. of mod. phys. 61(1989) 1. [6] Planck Collaboration, Planck 2018 results VI, N. Aghanim et al., A&A 641(2020) A6. [7] M.F. Wondrak, The Cosmological Constant and Its Problems: A Review of Experimental Search for Quantum Gravity Hossenfelder S. (eds) FIAS Science Series. Springer, Cham (2018). [8] L. ombriser, Physics Letters B 797(2019) 134804. [9] S.M. Carroll, Living Rev. Rel. 4(2001) 1. [10] M. Khurshudyan and A.Z. Khurshudyan, Symmetry 10(2018) 577. [11] A. Paliathanasis, S. Pan and W. Yang, Int. J. mod. phys. D 2812 (2019) 1950161. [12] Gong-Bo Zhao, et al., Nature Astron. 1, 9 (2017) 627. [13] C. Escamilla-Rivera and A. Najera, arXiv:2103.02097 [gr-qc]. [14] K. Boshkayev, R. D'Agostino and O. Luongo, Eur. Phys. J. C 79(2019) 332. [15] W. Zimdahl, H.E.S. Velten and W.S. Hipolito Int. J. mod. phys. Conf. Ser. 03(2011) 312. [16] A. Kamenshchikab, U. Moschellac and V . Pasquier, Physics Letters B 511(2001) 265. [17] B.R. Dinda, S. Kumar and A.A. Sen, Phys. Rev. D 90(2014) 083515. [18] E. Elkhateeb, Astrophys. Space Sci. 363(2018) 7. [19] Ya B Zel'dovich Sov. Phys. Usp. 9(1967) 602. [20] C. W. Misner Phys. Rev. Lett. 19(1967) 533. [21] S. Weinberg Apj168(1971) 175. [22] V .A. Belinskil and I.M. Khalatnikov Zh. Eksp. Teor. Fiz. 69(1975) 401. [23] V .A. Belinskil and I.M. Khalatnikov Zh. Eksp. Teor. Fiz. 72(1977) 3. [24] J.A.S. Lima, A.S.M. Germano Phys. Lett. A 170(1992) 373. [25] W. Zimdahl and D. Pavon Gen. Relativ. Gravit. 26 (1994) 1259. [26] W. Zimdahl Phys. Rev. D 53(1996) 5483. [27] W. Zimdahl Mon. Not. R. Astron. Soc. 288665. [28] J.S. Farnes A & A 620(2018) A92. [29] M. Giovannini Physical Review D 93(2016) 083521. [30] W.S. H.E.S. Velten and W. Zimdahl, JCAP 0906 (2009) 016. [31] J.C. Fabris, P.L.C. de Oliveira and H.E.S. Velten, Eur. Phys. J. C 71(2011) 1773. [32] I. Brevik, o. Gron Journal of and Plasma Research 19, Iss. 1/2 (2014) 97. [33] S. Nojiri, S.D. Odintsov, V .K. Oikonomou and T. Paul Phys. rev. D 102 (2020) 023540. [34] I. Brevik and A.V . Timoshkin JETP 122(2016) 679. [35] K. Bamba and S.D. Odintsov Eur. Phys. J. C 76 (2016) 18. [36] S. Del Campo, R. Herrera and D. Pavon Phys. Rev. D 75 (2007) 083518. 16[37] o. Gron Astrophys. Space Sci. 173(1990) 191. [38] I. Waga, R.C. Falcao, and R. Chanda Phys. Rev. D 33(1986) 1839. [39] L. Diosi, B. Keszthelyi, B. Lukacs and G. Paal Phys. Lett. 157B (1985), 1, 23. [40] I. Brevik and A.V . Timoshkin (2020) N.D.J. Mohana, A. Sasidharanb and T. K Mathewc Eur. Phys. J. C 77(2017) 849. [42] I. Brevik, o. Gron, J. de Haro, S.D. Odintsov and E.N. Saridakis Int. J. Mod. Phys. D 26 (2017) 1730024. [43] I. Brevik Mod. Phys. Lett. A 31(2016) 8. [44] I. Brevik, E. Elizalde, S. Nojiri, and S.D. Odintsov Phys. Rev. D 84(2011) 103508. [45] I. Brevik and B.D. Normann Symmetry 12(2020) 1085. [46] V .H. C' ardenas, M. Cruz and S. Lepe Phys. Rev. D 102(2020) 123543. [47] E. Elizalde, M. Khurshudyan, S.D. Odintsov, and R. Myrzakulov Phys. Rev. D W.J.C da Silva and R. Silva Eur. Phys. J. C 81(2021) 403. [49] D. Wang, Y .J Yan, and X.H Meng Eur. Phys. J. C 77(2017) 660. [50] I. Brevik and A.V . Timoshkin International Journal of Geometric Methods in Modern Physics 17, 2 (2020) 2050023. [51] W. Yang, S. Pan, y. E. Di Valentino, z. A. Paliathanasis and J. Lu Phys. Rev. D E. Elkhateeb Int. J. mod. phys. D 28, no. 9 (2019) 1950110. [53] S.K. Tripathy, D. Behera and B. Mishra Eur. Phys. J. C 75(2015) 149. [54] V . Folomeev and V . Gurovich Phys. Lett. B 661Issues 2-3 (2008) 75. [55] J. Ren and X.H. Meng Int. J. Mod. Phys. D 16, 08 (2007) 1341. [56] N. Cruz1, E. Gonz' alez, S. Lepe and D. S' aez-Chill' on G'omezJCAP 12(2018) 017. [57] S. Anand, P. Chaubal, A.M. azumdar and S.M. ohanty JCAP 11(2017) 005. [58] M. Cruz, N. Cruz and S. Lepe Phys. Rev. D 96(2017) 124020. [59] N. Komatsu and S. Kimura Phys. Rev. D 90(2014) 123516. [60] A. Avelino and U. Nucamendi, JCAP 2010 , no. 08 (2010) 009. [61] J.C. Fabris, S.V .B. Goncalves and R. de Sa Ribeiro, Gen.Rel.Grav. 38(2006) 495. [62] Bing Xu et al., ApJ855(2018) 89. [63] P. F. de Salas and S. Pastor, J. Cosmology Astropart. Phys. 07(2016) 051. [64] D.M. Scolnic et al., ApJ859(2018) 101. [65] M. Moresco et al., JCAP 05(2016) 014. [66] R. Ji' menez and A. Loeb, Astrophys. J. 573(2002) 37. [67] J. Magaena, M.H. Amante, M.A. Garcia-Aspeitia, V . Motta, Mon. Not. Roy. Astron. Soc. 476 (2018) 1, 1036-1049. [68] S. Alam et al., Mon. Not. R. Astron. Soc. 470(2017) 2617. [69] A.G. Riess, S. Casertano, W. Yuan, J.B. Bowers, L. Macri, J.C. Zinn and D. Scolnic Astro- phys. J. Lett. 908(2021) L6. [70] Andoni Aizpuru, Ruben Arjona, and Savvas Nesseris Phys. Rev. D 104(2021) 043521. [71]

2108.10905

Esraa Elkhateeb

Esraa Ali Elkhateeb and Mahmoud Hashim

Dissipative Unified Dark Fluid: Observational Constraints

18 page, 8 figures

astro-ph.CO

http://creativecommons.org/licenses/by/4.0/

We adopt a standard FRW cosmology with a unified scenario, where the usual dark matter and dark energy sectors are replaced by a single dark fluid (DUDF). The equation of state of such fluid can two power laws. As a result, it enables fluid to have a from dust at early times to dark energy at late times. The dissipation is represented by a bulk viscosity with a constant shear viscosity is excluded due to the isotropy of the universe. We performed a likelihood analysis using recent observational datasets from local $H_0$ measurements, SNe Ia, observational Hubble data, BAO, and CMB to put cosmological constraints on the model. The special case of the non- dark fluid (UDF) is also studied, while a similar analysis is performed on the $Lambda$CDM model for comparison. We got an $H_0$ value of $70.02$ $km$ $s^{-1} Mpc^{-1}$ for DUDF and $70.25$ $km$ $s^{-1} Mpc^{-1}$ for the UDF model. Our analyses revealed that the DUDF model has the lower Based on model selection statistics in the form of the Criterion (AIC), we compare different models to select the favored one due to the observational data used. Our results revealed that the UDF model has the minimum AIC, with the conclusion that it is the most favorable model for the data. $Delta AIC$ value of other models are then measured to this model. This difference indicated that the DUDF model is a substantial model on the level of empirical support. Examining the evolution of the the effective equation of state parameter, and the density that our unified fluid doesn't deviate from the standard at early times, with the ability to play the role of the by accelerating the universe at the late times.

2 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB partial sums are closely related to invariance princ iples, which in turn play an important role in many areas of applications s uch as Fi- nance, Actuarial Sciences, Demography, etc. As an example, consider the claims problem for an insurer, whose clients subscribe to sp ecific products through determined policies. In vehicule insurance for ins tance, the pol- icy may include that at each accident, the client makes a clai mX, which depends on many factors as the severity of the crash for insta nce. For simplicity, we suppose that the claims are reported at discr ete times nt0, wheret0is a fixed period of time that may in days, weeks or months. At each time jt0, the claim is a random variables Xj. So, at time nt0, the total claim (referred as the total loss) up to time nt0is given by the equation (1.1) insurer should have a accurate estimation of Snto fix the premiums by clients should pay at the establishment of the policies, oth erwise the ruin would be highly probable. We remark that the the discrete tim e modeling of claims ( 1.1) can be extended to a continuous time one. In such a case, the number of reported claims up to time t, sayN(t), is a random variable and the total loss up to we suppose that the insurer has a capital uat the beginning, and that the premiums can be linearized, say as ct, the surplus process (measuring the financial balance of the insurer) at time tcan be given though the model uses continuous time, in practice , the time is discretized into multiples of a unit of time t0>0and the model n>=0.MOVING CENTRAL LIMIT THEOREMS 3 Finding the limiting law of the stochastic >0, for an appropriate sequence of normalization coefficients (cn)n>=1, to a stochastic process {Y(t),0<=t<=T}is the essence of the problem (or functional central limit theorem) pr oblem. For inde- pendent data, the most used limiting law is a Brownian motion . But, even in that case, the general solution is a L' evy process taken as a Poisson process. It is reasonable to expect that at least, the insurer should avoid incurring a ruin, say at a t imetn(r), An approximation of the probability ruin is given by (1.2) n>=0, and n(r) = accurately estimated, Equation ( 1.2) may help in pre-setting candu before contracting policies, to ensure profit and avoid ruin . To learn more on such modeling, the reader is directed to Klugman et al. (2004 ) page 252, Grandell (1991 ) and the references therein. For independent and square integrable data, the class of possible weak limits is exactly that of of infinitely decomposable laws and the associated invari ance princi- ple leads to Levy processes, the Brownian motion and the Pois son process being among them (see Lo` eve (1997 ),Applebaum (2004 ),Niang et al. (2021 ), etc.). The general problem of finding the weak law, in the random sche me as Sn(t) n>=1 and, in the non-random scheme as4 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Sn(t) n>=1, (usually for T= 1) is the core of the invariance principle problem. This problem is hard and quite general since we do not necessarily know the dependence type between the losses Xj's nor do we always have that N(*) is a Poisson process. The main results and achievements in th e literature are obtained for independent data and when Nis a classical Poisson pro- cess. The problem of moving partial sums arise in the important set ting of func- tional weak limits. In what follows, we provide some backgro und. Consider a sequence of centered random variables (Xn)n>=1defined on the same prob- ability space (ohm,A,P). Let(p(n))n>=1be an arbitrary sequence of We define (1.3) S' 0= 0, for all = 0 for alln>=1, we find ourselves in studying the usual the partial sums of variances s2 n=Var(Sn). In some situations, we may be concerned not with all the parti al sums Snbeginning by the first r.v.X1but with partial sums that can start at any part of the sequence (Xn)n>=1. In ( 1.3), the partial sums S' nbegin with the r.v.Xp(n)+1and is the sum of all the nobservations with index greater that p(n). A more appropriate notation should be Sn,p(n)that we denote as S' n, given that the sequence (p(n))n>=1is already defined. For ex- ample, when dealing with invariance principles, we need to h ave the limit of the sequence of stochastic processes {Yn(t),0<=t<=T, n>=1}to some stochastic process {Y(t),0<=t<=T}. The state of the art (see ),van der Vaart and Wellner (1996 ),Loet al. (2016 ), etc.) expresses that, in order that weak convergence holds, we need to have th e conver- gence in and that the sequence is unifor mly tight (as inMOVING CENTRAL LIMIT THEOREMS 5 Billinsgley (1968 )) or asymptotically tight (as in van der Vaart and Wellner (1996 )). For now, we focus on the weak convergence in finite distrib utions, i.e., of the 1),k>=2. It is clear that we have, for eachj-th with p(n) = the relation between snands' nplays a major role as we will see later. We refer to such partial sums as moving partial sums, since th e arbitrary. As we will see, the handling weak laws of moving partial sums ( MPS ) can be a lot easier for independent r.v.'s. Nevertheless, we have to make sure that all steps are rigorously taken into a ccount. But, for dependent data, the situation requires more attention a nd can get The problem is even more serious if the sequenc as expected in applications, especially in Actuaria l Sciences and Fi- nance . In this paper, we focus on associated data introduced in 3.1as associated data. We will assume knowledge of basic definitions and resul ts on assici- ated data, thus, we refer the reader to Sanghar' e and Lo (2016 ) for a quick review. In 3.1, we will review the most important facts on associated data. For much more details, Rao (2012 ) and Bulinski and Shashkin (2007 ) are more appropriate. Nevertheless, we will present a thorough review of some important results on independent data, in order to facilita te the passage to dependent data. Hence, the purpose of this paper is to extend Gaussian centra l limits for independent data to Gaussian limits for MPS and to compare the classi- cal conditions (Lyapounov and Lynderberg, Uniform Asympto tic Negligi- bility ( UAN ), Bounded Variance Hypothesis ( BVH ), Convergence ( CVH ), etc.) for full sums and moving partial sums. We will then study how such results can be used in invariance princip les to re- scaled Brownian motions. Since handling invariance princi ples for de- pendent data, here for associated data, requires similar as ymptotic weak6 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO laws for MPS and their applications, this will lead to more general resul ts in invariance principles for associated data in comparison with in particular in Oliveira (2012 ). The present analysis will open the door to studying asymptotic weak laws for MPS for other types of dependence and for other specific types limit laws, i.e., for any type limiting laws, in particular Poisson law s. Future re- search works will focus on generalizations of the results to random sums and their data-driven rest of the paper is organized as follows. In Section 2, we study the MPS in the independence situation under the n-+, wherea(t)is a non-decreasing function in t[0,1]. Next, we apply the re- sults to al invariance principles with weak convergence to re-scaled Brownian motions. We get moving versions for Ly and the theorem. We give totally de- tailed proofs that are postponed to the (appendix) section 6. In Section 3, we deal with dependent random variables, here associated s equences. We make profit of the moving versions to significantly extend t he central limit theorem of Sangar' e H. and Lo, G.S (2018 ) which in turn is an ex- tended version of Oliveira (2012 ). Getting weak limits of MPS 's requires the following more complicated =s1< s2< t1< t2<=T(T= 1), wherea(t)is a non-decreasing func- tion int[0,1]. Here again, we apply the results in the weak limits in the invariance principle for associated s equences. In that section, we used regrouped-data method as it is usual in weak laws on associated sequences. Therein, we can use both conditions o n regrouped data and no-regrouped data. We need a whole section, say Sect ion4, to give the links between these two type of conditions. We close the paper with concluding remarks (in 5).MOVING CENTRAL LIMIT THEOREMS 7 The obtained results will help in successfully addressing t he general set- ting of random invariance principles using random numbers N(nt)of data, t[0,1]in the innovative case of associated data, and beyond. Let us proceed to the study for each type of dependence mentio ned /f.sc/o.sc/r.sc this section, we are going to check that Lyapounov's Theor em and Theorem are unchanged in the moving frame. We exactly use the same proofs as in Lo` eve (1997 ) but we follow the detailed proofs in Lo (2018 ). We will show how to use them in establishing general invari ance principles for independent data at least for finite-distrib utions. Here, we adopt the following n>=1. Here are the moving versions of the two main central limit the orems for independent data. 2.1. Moving 1. (Lyapounov). Suppose that the Xk's are independent and for every d >0and (2.1) A' n(d) n-+. Then, AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Remark 1. As in the usual way, a version using arrays automatically written without any change in the proof. Theorem 2. Suppose that the Xk's are indepen- dent and have only finite second order moments. We have the equ ivalence between the assertions below: (2.2) whenn -+ and (2.3) o>0, gn(o) n-+. As promised, the proofs are direct and require checking all l ines in the mentioned proofs. They are given in the appendix. Let us focu s on the applications to invariance (R1) . A first remark is that if we the Lynderberg condition for the whole sequence implie s that it holds for the sequence {Xp(n)+j,j>=1}. Indeed, the condition above implies that for anye >0, there exists n0such that for any n>=n0, we have Hence, for CENTRAL LIMIT THEOREMS By letting n-+, we get the moving The same remark applies for the Lyapounov condit ion. 2.2. Application to finite distributions limit invariance princ iples. Let us set {Yn(t), invariance principle investigates whether such a seque nce of stochas- tic processes converges to a tight stochastic process, main ly to a motion. Here, a moving version of the central limit theorem is useful. It may be not very hard to proceed for independent dat a. But the way we use will serve as a basis for more complex dependent dat a as we will see in the second part of the paper. The limit of s[nt]/snplays an important role here. In the iidcase with EX2 1= s2, we have s2 n=Var(Sn) =ns2, which leads to the fact as n --+. In the general case, we do not have such a simple relation. We h ave to set assumptions, for as n --+, wherea(t)is a non-decreasing function of t(0,1). From this, we may use the moving version of the the orem to have the following result. Theorem 3. Suppose that the Xj's are independent, centered and Let us suppose that the Lynderberg condition ho lds. Then the sequence of stochastic processes converges in a Wiener stochastic process.10 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Proof . Let us set 0 =t0<t1<t2<...<t k+1= 1. We put : Zn(t1) =Yn(t1)-Yn(t0) =Yn(tj)-Yn(tj-1) =Yn(tk)-Yn(tk-1) each fixed j {1,...,k}, we have Zn(tj) the Lynderberg condition holds for the whole sequence , it holds for each sequence {1,...,k}[See Remark (R1), page 8]. So, we may apply the moving (CLT) for = [ntj-1]to get converges to the standard normal law N(0,1)and by n--+. So, for each j {1,***,k}, by Slutsky's lemma, we as eachj {1,...,k}. Now, for any u= CENTRAL LIMIT THEOREMS the vector Zn= converges to a random Gauss- ian vector Z, with independent components. For each variance of the continuous theorem mapping theorem, Yn=AZnwith A= 1 0...0 1 1...0 1...1 0 1... ...1 weakly converges to V=AZand its components we see we conclude that the finite margins of Yn(t)weakly converge to those ofW(a(t)), whereWis a Wienner process, since E(V2 j) =a(tj)andE(VjVh) =a(tj)a(th), for alljandhboth in /f.sc/o.sc/r.sc Easy introduction to Associated random variables. In fear of ren- dering this paper heavier, we refer the reader to Sanghar' e and Lo (2016 ) for a quick introduction on associated sequence of random va riables, and give the important Newman's inequality as follows. Readers are directed toRao (2012 ) and Bulinski and Shashkin (2007 ) for a more detailed in- troduction to associations. The above mentioned inequalit y is used below several times.12 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Lemma 1 (Newman and Wright (1981) Theorem, see Newman and Wright (1981 nbe associated, then we have for all t= us begin with the approximation result. 3.2. Moving version of the central limit theorem for associated d ata. As in Sangar' e H. and Lo, G.S (2018 ), we suppose that we have sequences of integer numbers that n=m(n)l(n) l(n)/n-0as n-+. We want to stress that the integers func- tions ofn>=1throughout the text even though we may and do drop the la- belnin many situations for simplicity's sake. Let us suppose tha t(p(n))n>=1 is an arbitrary sequence. The moving versions correspondin g to the hy- potheses in Sangar' e H. and Lo, G.S (2018 ) and alike papers as Oliveira (2012 ) are as follows. We denote S' 0= 0, n>=1. The hypotheses we will be using are the CENTRAL LIMIT THEOREMS the sequel, it may be handy to use the notation (3.2) us prove the moving version of Formula (4.12) in Sangar' e H. and Lo, G.S (2018 ). Proposition 1. Under Hypotheses (L), (H0), (Ha), (Hab) and (Hb), we have for anytR, . We follow the one in Sangar' e H. and Lo, G.S (2018 ) which we ap- propriately adapt. Let us have for (3.6)14 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO But for any |x|. Thus the second member of (3.6)is, by the Cauchy-Schwarz's tends to zero as n-+by(Hab) proves recall that Yj,l= for1<=j<=m. Observe to the Newman's inequality (see Lemma 1), we CENTRAL LIMIT THEOREMS (L3), we take the associativity into account to get 1 leads by (Ha) and (Hab) , we get, as n-+our conclusion, n-+.16 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Let us draw a first important application of our results regar ding the in- variance principle. 3.3. Gaussian Central limit theorem. Proposition 1says that under the hypotheses (L), (H0), (Ha), (Hab) and (Hb), S' n/s' nhas the same weak limit law for each independent components and for each the random variables in the left-hand of the equality a re already defined in Formula ( 3.2) and=dstands for the equality in distribution. A - Lyapounov central limit theorem . As a consequence a for independent data is enough to ensure the Gauss ian central limit theorem. Let us set for d n-+. By theCr-inequality ( real-valued numbers a,b andr>=1), have finite (2+d)-moments, d >0if theXj's do. So, we have Theorem 4. Letd >0. If theXj's have finite (2+d)-moments and (Hc) holds on top of the hypotheses (L), (H0), (Ha), (Hab) and (Hb), CENTRAL LIMIT THEOREMS 17 Proof Since the study transformed into that of some in independent and centered data, the Lyapounov condition is enough to get the c onclusion, but the involved variance is s'2 mlwhich is equivalent to s'2 nby - A central limit theorem . With the equivalence s'2 ml~s'2 nby Hypothesis (Ha), the Lynderberg condition on theTj,n's is equivalent to: denote Bn= (FLL) theorem is stated as fall ows. Theorem 5. We suppose that the Xj's have finite second order moments and that the hypotheses (L), (H0), (Ha), (Hab) and (Hb) hold. B and only if the Lynderberg condition (3.9)holds.18 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Here, we directly used the Lyapounov and the FLL conditions ( (Hc) and 3.9) for on the regrouped data (the Tj,n's). It may be more convenient to give sufficient conditions of the non-regrouped data for they hols. When we proceed to complete comparison between the situation for regrouped data and non-regrouped data as we will do it in Section , we arr ive at this final version. Theorem 6. Suppose that the random variables Xj,j>=1, are centered and square integrable and associated and that Hypotheses (L), ( H0), (Ha), (Hab) and (Hb) are satisfied. Then we have following results: (1) If theXj's for some d >0and the Suppose that the following uniform negligibility of the Then we and only if the following Lynderberg-type condition hold s: (3.11) for CENTRAL LIMIT THEOREMS 19 As said earlier, we will give a complete justification of that theorem in Sec- tion 4. Now, let us see the applications of the results to invariance principles. 3.4. Invariance principles of associated data. We already defined the sequence of stochastic page 9and we wish to find the weak law. Above, in the independent case , we used the hypothesis (H1) which controls how the variance s[nt]grows with respect to sn. Such a condition still works if the data are we may need the following more elaborated assump tion. (H1-NSA) . There exists a measurable function a(t)oft[0,1]such that this, we may use the moving version of the Levy-Feller-L or Lyapounov theorem, we have the following result. Theorem 7. Let us suppose that the Xj's are associated and (H1-NSA) holds. Let us suppose that the hypotheses (L), (H0), (Ha), (H ab) and (Hb) hold on top of the moving Lynderberg condition. Then the sequ ence of sto- chastic processes converges in finite distributions a Wienner stochastic process. Proof . Let us set 0 =t0<t1<t2<...<t k+1= 1. We put :20 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB each fixed j a moving partial sum with p(j,n) = [ntj-1],(j,n) = for each j have the central limit theorem for snZj,n, by Slutsky's lemma, we we denote a(tj) =a(tj)-a(tj-1),j {1,...,k}. Next, let us apply Wold Criterion (see Lo(2018 ), Chapter 1) and we haveMOVING CENTRAL LIMIT THEOREMS already proved that Rn(2)-0. It remains to prove that Rn(1)does. But, by Newman's inequality in Lemma 1, we for any both converge to a(tj1) + a(tj2)and hence there, the conclusion is the same as in the independence case since we move from the limit to the increments to the sequence the same way. /square22 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB /s.sc /o.sc/f.sc /o.sc/n.sc /d.sc/a.sc 4and 5used Lyapounov and FLL-type conditions on the re- grouped date. Using Proposition 1, we pointed out that under Hypotheses (H0), (Ha), (Hab) and (Hb), the weak limit law of exactly as that and the Tj,nare there, the theory on weak limits of independent case app lies. have the Bounded Variance Hypothesis (BVH) is 1. But, because of Hypothesis (Ha), we may change the sequence we study the (BVH) still holds for some c >0. Moreover, the Uniformly Asymp- totically Negligibility ( UAN ) condition in sums of independent random vari- ables theory : (4.1) U''(n) = >=et' n)-0, is controlled as follows:MOVING CENTRAL LIMIT THEOREMS 23 (4.2) e>0, in the first place, let us see the links of usual condition s related to theory of sums of independent random variables when they are expressed on theTj,n's and on the Xj's. A - Convergence conditions on the regrouped and simple data . (A1) The UAN condition . That condition is expressed on the It is controlled with B'' n. This itself is bounded through the simple data follows. By using the convexity of R+x-x2, we denoting (4.3) have (4.4) - Lyapounov condition . Ford >0, the Lyapounov condition is24 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO (4.5) A'' n(d) shorten the notation, we use the following notation by using the convexity of R+x-x2+d, we have A'' n(d) we get (4.6) A' n(d) - Lynderberg Condition . Fore >0, the Lynderberg Condition for regrouped data is (4.8) L'' n(e) CENTRAL LIMIT THEOREMS 25 Let us set, max Since we ohm, we are going to complete our comparison by using a moving Lynd erberg condition on the data Xj's:26 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO (4.9) L' n(e) by Hypothesis ( Ha) and ( Hab), we have (t' n)2/s'2 n-1, asn-+. So, fornlarge enough, (|Xj|> et' n)implies So, we have for n large enough, for any e>0, (4.10) we may justify Theorem 6using conditions of the real data Xj,j>=1, as announced, by summarizing the discussion above. (B) Weak limits of Moving Partial sums using no-regrouped da ta. Let us suppose that hypotheses (L), (H0), (Ha), (Hab) and (Hb ) hold. The BVH hypothesis for the regrouped data T' j,nsis satisfied. The UAN for re- grouped data is ensured by the uniform negligibility of the v ariances (See (4.2)) which itself is forced by a condition on the uniform neglig ibility of the variances for non-regrouped data (See ( 4.3)). So hypotheses (L), (H0), (Ha), (Hab) and (Hb) hold and if B' n-0, the UAN condition and the BVH are satisfied and hence, the FLL theorem for inde- pendent data Tj,napplies and Lynderberg condition for regrouped data is forced by the Lynderberg-type condition for the non-regrou ped data ( 3.11 ) (See ( 4.10 ) and ( 4.10 )). Applying the Feller-L' evy-Lynderberg (as Theorem 23 in Lo(2018 ), Chapter 7, section 4) to get Part (2) of Theorem 6 Also for data having 2 +dmoments, for d >0, the Lyapounov condition for regrouped data is forced by the same condition for non-re grouped data 3.10 (See ( 4.6) and ( 4.7)). Hence the Lyapounov theorem for the Tj,n's ap- plies under the Lyapounov-type condition ( A' n-0). We apply the Lya- pounov theorem (as in in Lo(2018 )) to get Part (1) of Theorem 22 6, Chap- ter 7, Section 4). CENTRAL LIMIT THEOREMS paper will be closed by Section 6where the full computations of weak limits to the Gaussian law for moving partial sums of indepen dent data are given. The paper, after offering a general handling moving pa rtial sums convergence for independent data opens the rich field of such kinds of as- ymptotic weak laws for dependent data and their application to s and beyond to modeling in Applied statistics. Th e next step is to remain of these two kinds of dependence but to characteriz e asymptotic of random sums with not necessarily the Poissonian hypothes is on the counting process N(*)in classical hypotheses.28 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB D. (2004). L' evy processes and stochastic Calculus . Press, P.(1968). Convergence of Probability measures . John A. and Shashkin A.(2007) Limit theorems for associated random fields and related systems . World Scientific Publishing. Singapore. Lo, G.S.(2016). Weak Convergence (IA). Sequences of random vectors. SPAS Books Series. Saint-Louis, Senegal - Calgary, Canada. Doi Arxiv : 1610.05415. ISBN : 978-2- 9559183- 1-9 Lo, G.S.(2018). Mathematical Foundation of Probability Theory . SPAS Books Series. Saint-Louis, Senegal - Calgary, Canada. Doi : J.(1991). Aspects of Risk Theory . S.A., Panjer H.H. and Willmot G.E. (2004) Loss model: from data to decisions . Wiley, New-Jersey. Lo` eve (1997). Probability Theory I . Springer Verlag. Fourth P.E.(2012). Asymptotics for Associated Random Variables. Springer-Verlag Berlin Heid elberg. Newman, C.M and Wright, A.L.(1981). An invariance principl e for sequences. Ann. probab. , Vol. 9 (4), 671-675. Niang A.B., Lo G.S. and Diallo M.(2004) Asymptotic laws of su mmands I: square integrable independent random variables. Imhotep I Center r-001.pdf). Rao P. B.L.S. (2012). Associated sequences, Demimartingales and Nonpara- metric Inference . Probability and its applications. Springer Basel Dore- drecht, London, e H. and Lo G.S. (2018). General Central Limit Theore ms for Associated Sequences. In A Collection of Papers in Mathematics and Related Sciences, a festschrift in honour of the late Galaye Dia(Editors : Seydi H., Lo G.S. and Diakhaby A.). Spas Editions, Euclid Ser ies Book, pp. . Doi : e, H. and Lo, G. S.(2016) A Review on asymptotic norma lity of sums of associated random variables. Afrika Statistika , , 11 ( 1), CENTRAL LIMIT THEOREMS 29 Doi : Arxiv 1405.4316. van der Vaart A. W. and Wellner J. A.(1996). Weak Convergence and Em- pirical Processes With Applications to Statistics . Springer, New-York.30 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB of Theorem 1. In the partial sums beginning by p(n) = 0 , by using Lemma 8 in Lo(2018 ), Theorem 1holds ford >1if it does for d= 1. So it is enough to prove the theorem for 0< d<=1.By Lemma 9 in Lo(2018 ), Assumption ( 2.1) impliessn' us use the expansion of the characteristic functions fk(u) the order two to get for each as given in Lemma 6 in Lo(2018 ), (6.1) fk(u/sn') = the characteristic function of Sn'/sn'is, use the uniform expansion of log(1+u)at the neighborhood at 0, that is (6.2) eachkin (6.1), we have (6.3) fk(u/sn') = 1-uknMOVING CENTRAL LIMIT THEOREMS 31 with the uniform bound |ukn| ( 6.2) to ( 6.3) to -u2/2. We get for u completes the proof.32 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO Proof of Theorem 2. The proof follows the lines of that of Lo` eve (1997 ). But they are extended by more details and adapted and changed in some parts. Much de tails were omitted. We get them back for making the proof understan dable for students who just finished the measure and probability cours e. Before we begin, let us establish an important property when (2.3) holds. Suppose that this latter holds. We want to show that there exi sts a that implying also that e-1 ngn(en) = o(en)-0and thatgn(en) =o(e2 n)-0).To this end, let ,we have enough. We will get what we want from an induction on this property. Fi xk= 1 and denote n1an integer such that we apply the same property on the sequence {gn(*),n1+ 1}withk= 2. We find a n2> n1such that we apply the same property on the sequence 3. We find a there exists an infinite sequence of integers n1< n2< ... < n k< nk+1< ... such that for each k>=1,one for each n>=1, for each k>=1, there exists k(n)such surely have is clear 1/k(n) we are going to use (6.4) CENTRAL LIMIT THEOREMS 33 Proof of : ( ( 2.3) holds.Thus there exists a positive numbers such that (( 6.4)) prevails. First, we see that, follows remains to prove that this end we are going to use this array of truncated randoms variables {Xnk, n>=1} defined as follows. For each fixed consider summands remark that for any e AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO and remark equivalent in probability. This implies that they have the same limit law or do not have a limit law together. So t o a limit law, we may prove that S' nn/s' nhas a limit law. Next by Slutsky lemma, it will suffice to establish the limit law of show that s' nn/s' n-1.We focus on this. Begin to remark that, since E(Xk) = 0 , we have the decomposition 0 =E(Xk) get also that E(Xnk) CENTRAL LIMIT THEOREMS 35 Combining all that leads help of this let us evaluate that for each fixed n>=1,the Xnkare still independent. The technic used in ( 6.5) may be summarized as follows : any measurable function g(*)such that g(0) = 0, (6.7) Eg(Xnk) putting these remarks together, we obtain36 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB leads use the simple inequality of real numbers conclude from the last inequality thatMOVING CENTRAL LIMIT THEOREMS (6.6) above, we arrive comes that the proof of this part will derive from the limit law of S' nn/s' nn.We center the Xnkat their expectations. To prove that the new summands T' nn/s' nn, to N(0,1), we check the Lyapounov's condition for d= 1, that .38 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB LO But we =|*|in (6.7) to see again also last formula Lyapounov's theorem, we CENTRAL LIMIT THEOREMS by ( conclude that to N(0,1). Proof of : ( 2.2)==(2.3). The convergence to N(0,1)implies that for any fixedtR,we are going to use uniform expansions of log(1+z). We implies (6.9) we use the = <=1 to get that (6.10) latter implies40 AKIM ADEKPEDJOU, ALADJI BABACAR NIANG, CH 'ERIF MAMADOU MOCTAR TRAOR 'E, AND GANE SAMB (6.10 ), we see that logfk(u/s' n) uniformly defined in ksuch that enough and ( 6.8) -u2/2 that using the uniform bound of <=1 and arrive CENTRAL LIMIT THEOREMS 41 If we take the real parts, we have for any fixed are going to use the following fact: for all this for d= 1to

2108.10906

Gane Samb Lo

Akim Adekpedjou, Aladji Babacar Niang, Ch\'erif Mamadou Moctar Traor\'e, Gane samb Lo

Central limit theorems for associated possibly moving partial sums and application to the non-stationary invariance principle

43 pages

math.PR

http://creativecommons.org/licenses/by/4.0/

General Central limit theorem deals with weak limits (in type) of sums of row-elements of array random variables. In some situations as in the problem, the sums may include only parts of the row-elements. For strictly stationary arrays (stationary for each row), there is no change to the asymptotic results. But for non-stationary data, especially for dependent data, asymptotic laws of partial sums moving in rows may require extra-conditions to exist. This paper deals with central limit theorems with Gaussian limits data. Our main focus is on dependent data, particularly on associated data. But the non-stationary independent data is also studied as a learning process. The results are applied to s for the types of data described above. In Moreover, results for associated sequences are interesting and innovative. Beyond their own interest, the results are expected to be applied for random sums of random variables and next in statistical modeling in many disciplines, in Actuarial sciences for example

gate array (FPGA) board, the proposed GGNB requires slices, LUTs, flip-flops, and DSP units, respectively, NN words: Controller area network, security, graph-theory. 1. Introduction The application of vehicular electronic control units (ECUs) depends on intra- and inter-vehicular communication networks' performance. are and Any cyber or physical attack on these communication networks has severe consequences, including private data leakage, financial loss, and jeopardizing human life. As a result, it is imperative to secure these networks of modern vehicular systems. The typical automotive data communication networks include twisted wire paired controller area network (CAN), single-wire local interconnect network (LIN), FlexRay, and optical media oriented (MOST). Figure 1 shows the vehicular network data and some of their example applications. Among all the vehicular data communication networks, CAN is the most popular and widely used in modern automobile systems [3]. However, the national highway traffic safety administration (NHTSA) and other researchers identified several security breaches on the vehicular CAN bus [4,5,6]. According to the NHTSA, cybersecurity plays a more important role in the vehicular systems than other computational and [ 4]. In addition to the NHTSA and industry, academic researchers have proposed various intrusion detection systems (IDS) for the CAN However, the existing research fails to provide a that can effectively tackle all kinds of attacks. We hypothesize that a graph-based Gaussian naiveBayes (GGNB) IDS for CAN bus to tackle this issue. In particular, the key contributions of this paper are: *This is the first introduction of graph-based model in naive Bayes algorithm for CAN intrusion including mixed attacks without modifying the area network (CAN) Ethernet Media oriented systems transport (MOST) FlexRay Local interconnect network (LIN) Figure 1: The intra-vehicular communication system uses CAN for collision MOST for audio-video bridging and time-sensitive ethernet networking, FlexRay for the brake-by-wire system, and LIN for multifunctional keyless system [7]. *It is the first algorithm that integrates common graph properties with PageRank (PR)-related features into a GNB the proposed methodology exhibits 98.57% detection accuracy, which is better or comparable with the proposed methodology yields many fold reduction in runtime com- pared to the existing classification gate array (FPGA) implementation of the proposed GGNB requires 5.7x,5.9x,5.1x, and 3.6xfewer slices, LUTs, flip-flops, and DSP units, respectively, than conventional NN architecture. We organized the rest of the paper as follows: Section 2 gives a brief overview of existing CAN bus attacks and IDSs. Section 3 proposes our GGNB algorithm. Section 4 compares our proposed IDS with existing Section 5 concludes the paper. 2. Background 2.1. Attacks on CAN Bus The ever increasing demand for autonomous vehicles and competition to produce fully autonomous cars by the original equipment has grown the CAN bus's security concern. Besides, several security vulnerabilities in the CAN bus [ 8,9]. This work studies viirtual attacks on the safety-critical systems using a compromised ECU through OBD-II port. The attacker is assumed to have full control of a wide range of functions, for example, disabling the brakes, stopping the engine, and controlling other vehicle functions using reverse Valasek and Miller demonstrated that it was possible to have real- world attacks on multiple vehicles using the CAN bus [ 10]. The attacked the brakes of a Jeep Cherokee while it was driving on a highway. Another interesting approach remotely attacked CAN and showed that it is possible to attack the vehicle's Bluetooth and [11]. 2.2. Existing Intrusion Detection Systems for CAN The importance of securing CAN bus led researchers to propose different IDSs [12]. Moore et al. used the regularity of CAN message frequency to detect the anomaly [ 12]. A similar detection method proposed by Gmiden et al. relied on the time intervals of CAN messages. Based on the regularity in the signal frequencies, they hypothesized that an accurate detection of signal injection attack is possible by monitoring the inter- signal wait times of CAN bus traffic. A highly accurate SVM-based IDS was proposed by Tanksale et al. [ 5]. However, it is only applicable to denial of service (DoS) attacks due to the characteristics of the IDS model. Another interesting IDS uses a deep convolutional neural network (DCNN) to secure the CAN [ 6]. This method used a frame builder that converts the CAN bus data into a grid-like structure fed into the DCNN. However, this method is not applicable to replay attacks due to the unique that attack. Zhou et al. integrate a deep neural network (DNN) with a triple loss network [ 13]. In this method, the researchers extract data features as a set of vectors and then compute the similarity between two CAN data sequences extracted in real-time. Then the integrated approach uses an additional calibrated data sequence to identify the malicious data. Verendel et al. proposed a honeypot security mechanism at the acting as a decoy in simulating the in-vehicle networks [ 14]. This methodology collected and analyzed the attacked information to update the later version of the security system. The major challenge in is to make it as realistic as possible and concealing it from et al. proposed a rule-based recurrent neural network (RNN) IDS for CAN [ 15]. However, this approach can handle only DoS, fuzzy, and replay attacks. On the other hand, Minawi et al. proposed a IDS considering only DoS, Fuzzy, RPM Impersonation, and Gear Impersonation attacks and exhibits good model accuracy [ 16]. et al. used only DoS, fuzzy, and spoofing attacks to build their private dataset [ 17]. However, unlike these existing methods, we consider mixed attack scenarios. As a result, we removed these methods [ 15,16,17] from our analysis and Existing Attack Prevention Methodologies To secure CAN bus, researchers proposed several techniques [ 18,19,20]. Wolf et al. [ 18] proposed a firewall signature-based architecture for communication gateways. The firewall only allows to exchange valid messages. However, this method can not fully shield the vehicle network because most modern vehicles have interfaces that enable access to the entire car system. Islam et al. [ 20] proposed the first dynamic arbitration ID refreshing algorithm for preventing CAN based attacks. They validated their methodology using real CAN and implementing virtual CAN to demonstrate their algorithm's efficiency. 2.4. Existing Graph-Based Anomaly Detection In recent years, we observed a significant increase in graph-based due to the ease of representation of complex data using used background knowledge of the evaluation metrics and biased the substructure discovery process towards discovering anomalous substruc- tures [21]. Background knowledge is added in the form of rule coverage, which conveys the percentage of the final graph covered by the The authors hypothesized that it is possible to discover by introducing negative weights to the rule coverage. This method can efficiently detect 100% of the attacks with no false positives rate using the KDD99 data set. Another relevant approach uses detection by analyzing the data for the suspicious employee activities at Kasios [ 22]. This approach provides a rich contextual and more of data due to its ability to discover patterns in databases which are not a trivial task to identify using traditional query or This method achieves a true positive rate and false positive rate of 100% and 65%, Types of Attacks on CAN The electronic fusion of automobiles brings many benefits, including driver comforts, performance, and vehicle efficiency. However, these benefits comes at a higher risk of cyber attacks and human safety due to on those devices. Researchers and OEMs already threats in CAN communication and agreed that it was not built with security in mind [ In this research, we consider both fabrication and masquerade CAN attacks. When an intruder using a compromised ECU with a forged ID is referred to as a fabrication attack. DoS, fuzzy, and spoofing attacks are also classified attacks [ 23,27,26,28]. The diagnostic attack consists of of CAN ID which values are in a particular range using a diagnostic tool [29,2]. Fuzzy attacks are mainly composed of fuzzing CAN ID and fuzzing payload attacks [ 1,2]. The fuzzing CAN ID consists of with a specific CAN ID. These CAN ID values are not a part of the legitimate values. Fuzzing payload attack involves modifying the payload of messages that are not used in legitimate traffic. Another weak attack, where hackers sniff in-vehicle networks and suspend message transmission of a certain ECU, is called suspension attack [ 30,2]. On the other hand, a masquerade attack requires two compromised ECUs to mount an attack. The masquerade attacks are considered strong attacks. A replay attack is known as a masquerade attack [24]. 3. Proposed Anomaly Detection Methodology In this paper, we propose a graph-based anomaly detection system to secure the CAN bus communication system. The proposed methodology is based on the PR analysis for detecting anomalies. PR is a used by Google Search Engine to rank their websites [ 31]. The PR algorithm can effectively suggest users for searching different topics in Google as their requirements. Before presenting the proposed anomaly we will discuss the PR algorithm. 3.1. PageRank Algorithm We will discuss the PR algorithm using a simple graph with four vertices A, B, C, and D, as shown in Figure 2. PR is initialized to the same value for all vertices, which assigns a probability distribution between 0 and 1. So, 6B ACDFigure 2: Example four-vertex directed graph for PR computation. the initial probability distribution of each C is1 n= 0.25, where n is the total number of vertices. Every vertex will contribute PR to other vertices with the outgoing edges. As a result, vertex D will contribute to A, B, and C vertices. Since D has three outlines, it will divide its PR equally into three parts. Here, we compute the PR of A as, PR(A) =PR(B) Outdegree (D)(1) where Outdegree(B) is the number of edges exiting from B. We PR of each vertex and assign the PR value. The iteration will stop when there will be no update of PR of vertices. We computed the PR values of vertices for both attacked and attack-free graphs. To identify the significance of PR values, we performed an analysis of many graphs. According to our observation, PR plays a vital role in detecting anomalies. Empirically, the distribution of PR values among attack-free vertices is significantly different from the attacked vertices. In an attack-free graph, the distribution of PR among different vertices is similar, whereas, in an attacked graph, the PR values of attacked vertices are significantly different from other vertices in the graph. As a result, we used PR-related features for our anomaly (b) Figure 3: (a) A simple graph where each vertex has an equal 0.25 PR value, and (b) when a DoS attack is introduced, the PR value of 0000 ID is significantly higher than the an intruder mounts a DoS attack on a network, it floods network with a certain ID with high priority. We depict a similar situation in Figure 3, where Figure 3(a) shows a simple graph with each vertex with equal 0.25 PR value. On the other hand, Figure 3(b) shows an example when vertex ID with 0000 has a PR value of 0.45, which is significantly higher than the other vertices of that graph. Algorithm 1 shows the PR and other PR-related feature We initialize all the necessary variables from Line 4 to Line 7. Then we iteratively compute the minimum, median, and maximum PR values of every graph giGin Line 8 to Line 13. 3.2. Proposed Graph-Based Gaussian Naive Bayes (GGNB) IDS The proposed graph-based IDS uses existing Gaussian naive Bayes (GNB) classifier [ 32]. The GNB follows Gaussian normal distribution and data. In general, naive Bayes uses the posterior probability of Bayesian rule to infer a classification label from input feature conditional posterior probability can be defined as the probability of an event (A) occurring given the probability of another event that has already occurred (B) and can be expressed as P(A|B) 1 PageRank-related features extraction collection of The list of minimum, median, and maximum PR of each graph in = []initializing the list of the PR of every vertex from each = []initializing the list of the median PR from each = []initializing the list of the maximum PR from each = []initializing the list of the minimum PR from each graph 8:for [i] =PR(gi) PR (gi)computes the PR of every vertex of [i] =median (graphList [i] =maximum (graphList [i] =minimum (graphList [gi]) 13:end for where P(A) and P(B) are distributions of events A and B, respectively. GNB classifiers assume that features are independent of each other and require a small training data to estimate the parameters needed for we assume that the likelihood of the features to be Gaussian. So, the likelihood probability is given the mean, and sis the standard deviation. Assume that we have an unknown binary outcome Yandncontinuous features of X. As we have a binary label, Yis Bernoulli, which can be defined by setting a positive probability. The likelihood of fX|Y(x|y)is the distribution of the observation for the known outcome. Our updated knowledge or posterior about the unknown after observation is pY|X(y|x). We can express the Bayes rule as pY|X(y|x) Equation 4 is true for the ncontinuous variables X=X0,X1,...,X n-1 9and can be expressed n-1) discussing the proposed algorithm, we give a simple example of an attack detection mechanism. Assume that in a certain number of graphs, we have the probability of attacked and attack free graphs are 0.5. Then we compute the normal distribution of each feature of those graphs. To identify a graph with Vivertices and Eiedges weather it is attacked or attack free, we utilized the likelihood (L) condition as sAtt =P(att)*L(Nodes =Vi|att)*L(Edges =P(attFr )*L(Nodes =Vi|attFr )* L(Edges =Ei|attFr )are probability of attack and attack free And If are underflow, a them can help us to restore a comparable value. The the to decide if the graph is attacked or attack free. If sAttis greater than sAttFrthen the graph is attacked. Table 1 shows the list of features we used in our proposed anomaly proposed anomaly detection methodology is shown in Algorithm 2. It accepts raw CAN data and window size as inputs and predicts whether the input is attacked or attack free. For the analysis, we used ~23msas window size considering 1Mbit/s CAN speed. The proposed algorithm will first create a list of graphs g1,g2,...,g nnamedGbased on the user-provided window size in Line 4. The CAN message IDs will act as vertices, and we construct an edge between two vertices depending on their sequence. We declared different list variables from Line 5 to Line 10. In Line 11, we declared a for loop. In each iteration of the for loop, we collect nodes, edges, maximum out-degree, minimum in-degree, minimum out-degree from each graphgfrom Line 10 to Line 18. Then we compute the PR-related features using (G)function which calls the Algorithm 1 in Line 19. The data processing and labeling are performed in Line 19. In 10Algorithm 2 Anomaly Detection CAN data, Prediction; Attacked if Prediction is correct else = []initializing the list of from each = []initializing the list of from each = []initializing the list of from each = []initializing the list of from each = []initializing the list of number of nodes from each = []initializing the list of number of edges from each graph 11:for [i] =numberOfNodes [i] =numberOfEdges [i] =maximumIndegree [i] [i] =minimumIndegree [i] (gi) 18:end (G)extract using Algorithm =trainTestSplit (data,labels )splitting the data and labels for training and testing 22:model =gaussianNb ) training =model.predict (test) making the 1: The proposed methodology uses graphs' vertices, edges, degrees, and features for Names Number of nodes in a graph Nodes or Vertex Number of edges in a graph Edges Maximum number of in-degree in a graph Maximum Indegree Maximum number of out-degree in a graph Maximum number of in-degree in a graph Minimum Indegree Minimum number of out-degree in a graph Minimum Outdegree Median of PageRanks among nodes in a graph Median PageRank Maximum of PageRanks among nodes in a graph Maximum PageRank Minimum of PageRanks among nodes in a graph Minimum PageRank between Line 21 and Line 22, we split the data and labels into training and test sets and fit the data into a GNB model. We perform attack prediction in Line 23. 4. Experiments For verifying the proposed methodology, we used real CAN data and performed analysis on an Intel Xeon(R) Silver 3.2GHz 20-core processor with 48GB RAM by using our proposed methods in Python language. In our research, we used raw CAN attacked and attack free data [ 1]. We refer to this data set as rawCAN. The attacked data includes DoS, fuzzy, spoofing, and replay attacks. Besides, we consider the mixed attacks that have all four kinds of attacks. From the raw CAN bus data, we build and~19K, and~80Kattack free, DoS, fuzzy, and mixed attacked graphs. To apply the proposed GGNB-based anomaly detection algorithm, we used standard graph features listed in Table 1. Besides, we identify s to help detect anomalies. It primarily helps removing the non- informative features from the model [ 33]. We computed feature importance for DoS, fuzzy, spoofing, replay, and mixed attacked data, as shown in Figure 4. We compute the mean differences of the sample features after which shows that distinctions are balanced around zero, as shown in Figure 5(a). However, the variance differences of the feature samples are entirely on the positive side. So, Figure 5(b) depicts that the positive 12Nodes Edges Maximum Indegree Minimum Indegree Maximum Outdegree Minimum OutdegreeMedian PageRank Maximum PageRank Minimum importance mean(a) Nodes Edges Maximum Indegree Minimum Indegree Maximum Outdegree Minimum OutdegreeMedian PageRank Maximum PageRank Minimum importance mean (b) Nodes Edges Maximum Indegree Minimum Indegree Maximum Outdegree Minimum OutdegreeMedian PageRank Maximum PageRank Minimum importance mean (c) Nodes Edges Maximum Indegree Minimum Indegree Maximum Outdegree Minimum OutdegreeMedian PageRank Maximum PageRank Minimum importance mean (d) Nodes Edges Maximum Indegree Minimum Indegree Maximum Outdegree Minimum OutdegreeMedian PageRank Maximum PageRank Minimum importance mean (e) Figure 4: The Maximum Indegree, Maximum Outdegree, and Maximum PageRank have the highest feature importance for DoS, spoofing, and mixed attacks, respectively and nodes and edges have the highest feature importance for fuzzy and replay distributions are more concentrated around their means than the negative ones. These indicate the discriminating power of the model. We can classify data if the mean difference is far away from the distributions' centers, or the variance difference is greater than the distributions' spread. After quantile transformation, we identified two features, Maximum Inde- gree and Maximum Outdegree, which distinguish the two classes very well. To detect positive or negative class, we observed a closeness between the trans- formed values of these two features for any graph. Hence, a using the mixed attacked data provides a straight line through the origin, as shown in Figure 6. The transformed features will maintain a range of values for most of the graphs of each class. So, after plotting altered Maximum Indegree and Maximum Outdegree features, we observe a good negative and positive class, as shown in Figure 6(a). Even without quantile transformation, the original features' values can separate the as demonstrated in Figure 6(b). Hence, using these features, the proposed methodology works well for detecting attacked and attack free graphs. The quantile transformation of the features helped us identify the classes. 131.5 1.0 0.5 0.0 0.5 1.0 1.5 Mean diffrences between attack and attack free 0.5 1.0 1.5 2.0 2.5 Variance differences between attack and attack free (b) Figure 5: (a) The mean differences between the two classes have symmetric spread in both positive and negative direction, and (b) the variance differences of the feature samples are entirely on the positive side. (a) (b) Figure 6: Separation of classes using Maximum Indegree and Maximum Outdegree (a) quantile transformed feature, and (b) original features values of 7: The quantile transformed features values in the range from u-to-+-3scan clearly separate attack and attack-free data. Then, we compute the transformed features' mean and standard deviations for both positive and negative categories. For the transformed Maximum Indegree and Maximum Outdegree, we select random values in the When we draw the resulting values, it can clearly distinguish two classes, as shown in Figure 7. 4.1. Comparison of Prediction Accuracy The proposed algorithm uses 67% and 33% data for training and The proposed algorithm efficiently detected DoS, fuzzy, spoof, and replay attacked graphs and detect any mixed attacks. The has 99.61%, 99.83%, 96.79%, and 93.35% detection DoS, fuzzy, spoofing, and replay attacks, respectively. Figure 8(a), Figure 8(b), Figure 8(c), Figure 8(d) show the confusion matrix of DoS, fuzzy, spoofing, and replay attacks. One of the major advantages of using is it can detect mixed attacks. According to our analysis, the proposed methodology has 96.20% detection accuracy for the mixed attacks, as shown in Figure 8(e). Precision indicates the proportion of correctly predicted positive cases, and a high precision indicates the low false-positive rate of an algorithm [ 34]. Recall tries to identify all actual positive cases, which is the ratio of positive observations to all the actual class observations. The F1 15Attack free Attacked Actual label Attack free label10290 22 26 free Attacked Actual label Attack free label10227 11 11 2212500010000 (b) Attack free Attacked Actual label Attack free label9875 406 36 3436 250050007500 (c) Attack free Attacked Actual label Attack free label6105 16 901 free Attacked Actual label Attack free label18536 41 994 7654 50001000015000 (e) Figure 8: The proposed methodology has a prediction accuracy of (a) 99.61% considering DoS, (b) 99.83% considering fuzzy, (c) 96.79% considering spoofing, (d) 93.35% and (e) 96.20% for the mixed attacks. Table 2: When considering DoS attacks, the proposed methodology has 53.79%, 10.75%, and 38.77% better precision, recall, and F1 scores, respectively, than the classifier, using rawCAN data set [1]. Types of attacksSVM [5] DCNN [6] CNB MNB GGNB PrReF1Pr Re F1 Pr Re F1 Pr Re F1 Pr Re F1 Dos 0.9995 0.9969 1.00.9985 0.9966 1.00.9983 0.9979 0.9975 0.9977 Fuzzy ---0.9995 0.9965 0.9980 0.9945 1.00.9972 0.9899 1.00.9949 0.9989 0.9989 0.9989 Spoofing ---0.9999 0.9992 0.9995 0.9599 0.9997 0.9794 0.9596 1.00.9794 0.9605 0.99634 0.9781 Replay ---- - -0.9960 0.8561 0.9208 0.9960 0.8576 0.9216 0.9974 0.8714 0.9301 Mixed ---- - -0.8889 0.9956 0.9392 0.8889 0.9992 0.9408 0.9978 0.9491 0.9728 Overall ---- - -0.9498 0.9807 0.9650 0.9488 0.9822 0.9652 0.9911 0.9655 0.9781 score is the harmonic mean of precision and recall. We analyze all three metrics for the DoS, fuzzy, spoofing, replay, and mixed attacks. We compare our method with a generic complement naive Bayes (CNB), multinomial naive Bayes (MNB), and the existing SVM-based [ 5] and DCNN-bassed [ 6] IDSs. The proposed GNB-based methodology has excellent 99.79% precision, which is better than all the competing methods. Using DoS attacks, the proposed method has 10.75% and 38.77% better recall and F1 scores, respectively, than the state-of-the-art SVM classifier. Overall, the proposed methodology has 99.11%, 96.55%, and 97.71% of precisions, recall, and F1 scores, respectively. In order to verify the efficiency of the proposed methodology, we used 16Attack free Attacked Actual label Attack free label1306 0 0 free Attacked Actual label Attack free label1328 1 1 25001000 (b) Attack free Attacked Actual label Attack free label1329 1 0 2 05001000 (c) Attack free Attacked Actual label Attack free label1328 0 0 free Attacked Actual label Attack free label1327 0 1 4 05001000 (e) Figure 9: The proposed methodology has excellent 100%, 99.85%, 99.92%, 100%, and 99.92% detection accuracy considering DoS, diagnostic, fuzzing CAN ID, fuzzing payload, and replay, CAN attacked and attack free OpelAstra data set [ 2]. The attacked data includes DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay, and suspension attacks. Besides, we consider the mixed attacks that have all six kinds of attacks. From the raw CAN bus data, we build and~17.5K, DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay, suspension, and mixed attacked graphs. The proposed algorithm efficiently detected DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay, suspension, and detect any mixed attacks. The proposed methodology has 100%, 99.85%, 99.92%, 100%, and 99.92% detection accuracy considering DoS, diagnostic, fuzzing CAN ID, fuzzing payload, and replay, respectively. Figure 9(a), Figure 9(b), Figure 9(c), Figure 9(d), and Figure 9(e) show the confusion matrix of DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay attacks. In addition, the proposed methodology has 97.75% and 99.57% detection accuracy for the suspension and mixed attacks, as shown in Figure 10(a) and Figure to the other data set [ 1], we analyze precision, recall, and F1 scores for the DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay, suspension, and mixed attacks. Table 3 shows the results of this analysis. The free Attacked Actual label Attack free label1235 0 30 free Attacked Actual label Attack free label12319 53 0 95 0500010000 (b) Figure 10: (a) The proposed methodology has 97.75% detection accuracy attacks, and (b) 99.57% detection accuracy considering mixed attacks, which includes all six types of CAN attacks. Table 3: When considering DoS attacks, the proposed methodology applying on the OpelAstra data set [ 2] has 54%, 11%, and 39% better precision, recall, and F1 than the state-of-the-art SVM of attacksSVM [5] DCN [6] CNB MNB GGNB PrReF1PrRe F1 Pr Re F1PrRe F1 Pr Re F1 DoS 0.9995 1.0 1.0 1.01.01.0 1.0 1.0 1.0 1.0 Diagnostic ----- -0.9932 0.9985 0.9959 1.00.9977 0.9989 0.9992 0.9992 0.9992 Fuzzing canid ----- -1.0 1.0 1.01.01.0 1.0 1.00.9992 0.9996 Fuzzing payload ----- -1.0 1.0 1.01.01.0 1.0 1.0 1.0 1.0 Replay ----- -1.0 1.0 1.01.00.9977 0.9989 0.9992 ----- -0.9960 0.9992 0.9976 1.00.9620 0.9806 0.9763 1.00.9880 Mixed ----- -1.00.9972 0.9986 1.00.9940 0.9970 1.00.9957 0.9979 Overall ----- -0.9993 0.9982 0.9987 1.00.9936 0.9968 0.9984 0.9973 0.9978 GNB-based methodology has excellent 100% precision, which is better than or comparable to all the competing methods. Using DoS attacks, the proposed method has 11% and 39% better recall and F1 scores, respectively, than SVM classifier. Overall, the proposed GGNB methodology has 99.84%, 99.73%, and 99.78% of precision, recall, and F1 scores, Sensitivity Analysis In order to measure the sensitivity of the proposed methodology for window size variation, we considered CAN communication window size from 11.5ms to 230ms. A wide range of window variation allows different frequency. According to our analysis, the proposed is fairly consistent with the variation of window size, as shown in Figure 11. The proposed method has a less than 1% accuracy variation in this extended window size variation considering DoS, spoofing, fuzzy, and 1850 100 150 200 Window size attack Fuzzy attackDos attack Mixed attackReplay attackFigure 11: The prediction accuracy varies by less than 1% for DoS, fuzzy, spoofing, and replay attacks for 11.5ms to 230ms window size variation, and it exhibits about 2% for the mixed attacks in the same window variation range using rawCAN [ 1] data set. replay attacks. The proposed method's accuracy dropped only from 94.42% to 92.54% for 11.5ms to 230ms window range for mixed attack. 4.3. Comparison of Time and Space Complexity For the GGNB algorithm, we compute the priors and the for each class in the training phase [ 35]. We compute the n records prior in linear time O(n). Considering k features, we calculate k likelihoods for each class. The likelihood probabilities will be two classes. For testing mcases, the proposed GNB algorithm time complexity isO(mk). For training, the space complexity only considers the likelihoods for two classes of GNB, resulting in O(k). We computed the likelihoods during the training phase, so the training phase only requires a loop up, resulting in testing space complexity of O(m). Generally, for a dataset with n records and kfeatures, the time isO(kn3). Fordssupport vectors, the test time is in the order ofO(kds). For training, the space complexity only needs to store As expected, the proposed GGNB-based methodology has training and test time, respectively, compared to the existing SVM-based anomaly detection system. 4.4. Features Reduction 4.4.1. Correlation Matrix To correlate the features with the label of the data, we used [36]. The correlation coefficient can be expressed r value in the range of -1 to +1. If two components x and y of the data are perfectly correlated, then r = +1 or -1, and r = 0 if they The positive value of r indicates the proportional correlation of the components. On the other hand, the negative value of r indicates the inverse proportional correlation of the reduction is the process of reducing the number of features in computation without losing important information and prediction accuracy [37]. To reduce number of feature variables, we used correlation matrix for our proposed features, as shown in Figure 12. Maximum Indegree and Maximum Outdegree are highly correlated with our desired output label Using only these two features the proposed algorithm exhibits 97.25 % accuracy, using mixed attack data. When we incorporate two more features Median PageRank and Maximum PageRank for detecting attacked label, we achieved up to 98.06 % accuracy. 4.4.2. Feature Reduction Impact on reduction leads to the use of fewer resources to complete com- putations due to the decrease in the number of variables [ 38]. reduction has a tremendous impact on the proposed as shown in the Table 4. All the results are normalized with the GGNB algorithm with all nine features. For both the full set of features and the reduced number of features, the GGNB methodology has the lowest training time than other naive Bayes algorithms. On the other hand, CNB has the most down testing time than the other naive Bayes algorithms. With reduced features, the proposed GGNB algorithm has 18% and 21% lower training time using rawCAN [ 1] and OpelAstra [ 2] data sets, GGNB has 37.5% lower testing time using only four features on IndegreeMinimum Indegree Maximum OutdegreeMinimum OutdegreeMedian PageRank Maximum PageRankMininum Indegree Minimum Indegree Maximum Outdegree Median PageRank Maximum PageRank Mininum 0.023 -0.53 0.016 -0.54 -0.81 -0.21 -0.76 0.43 0.92 10.25 -0.5 0.24 -0.49 -0.9-0.023 -0.77 0.58 0.023 0.25 1-0.051 0.98 -0.038 -0.51 0.94 0.12 0.83 -0.53 -0.5-0.051 1-0.041 0.28 0.45 0.079 0.49 -0.26 0.016 0.24 0.98 -0.041 1-0.04 -0.5 0.94 0.14 0.83 -0.54 -0.49 -0.038 0.28 -0.04 10.45 0.085 0.39 -0.26 -0.81 -0.9 -0.51 0.45 -0.5 0.45 1-0.3 0.59 -0.74 -0.21 -0.023 0.94 0.079 0.94 0.085 -0.3 10.37 0.69 -0.76 -0.77 0.12 0.49 0.14 0.39 0.59 0.37 1-0.18 0.43 0.58 0.83 -0.26 0.83 -0.26 -0.74 0.69 -0.18 12: The Maximum Indegree and Maximum Outdegree have the highest correlation with our desired output label or attack detection. Table 4: With the reduced number of features, the proposed GGNB algorithm has 18% and 21% lower training time using rawCAN [ 1] and OpelAstra [ 2] data sets, set type # of featuresCNB MNB GGNB Training time Test time Accuracy Training time Test time Accuracy Training time Test time Accuracy rawCAN [1] 9 1.1727 0.6337 0.9075 1.1338 0.6849 0.9098 1.00 1.00 0.9620 s 4 1.0555 0.5012 0.9647 1.0494 0.5733 . 0.7174 0.8196 0.6474 0.9618 OpelAstra [2] 9 0.9024 0.5417 0.9080 0.8780 0.5417 0.9090 1.00 1.00 0.9090 4 0.8415 0.4167 0.9090 0.8171 0.4583 . 0.9090 0.7927 0.6250 0.9087 the data sets with only a minor 0.03% maximum accuracy penalty, as shown in Table 4. 4.4.3. FPGA Implementation Results We have implemented the proposed naive Bayes-based algorithms in a Xilinx Zybo Z7 board as shown in Figure 13. Among different naive Bayes- based algorithms, CNB requires lower computation resources than the other two methodologies for attack prediction. For example, the proposed CNB 21requires 5.3x,4.7x,6.4x, and 6.5xfewer slices, LUTs, flip-flops, and DSP cores, respectively, than the GGNB method. Moreover, the proposed GGNB requires 5.7x,5.9x,5.1x, and 3.6xfewer slices, LUTs, flip-flops, and DSP units, respectively, compared to a conventional NN architecture with 4-hidden layers where each layer contains 500 neurons [13]. 5. Conclusion We have presented the first-ever GGNB algorithm by using common graph properties and PR-related features. When we apply the proposed algorithm on real rawCAN data set [ 1], the GGNB method has 99.61%, 99.83%, 96.79%, and 93.35% detection accuracy considering DoS, fuzzy, spoofing, and replay attacks, respectively. Moreover, this is the first-ever method with 96.20% detection accuracy for mixed attacks. Using OpelAsta data set [ 2], the proposed methodology has 100%, 99.85%, 99.92%, 100%, 99.92%, 97.75% and 99.57% detection accuracy considering DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay, suspension, and Zybo Z7 used for proposed IDS Figure 13: The proposed GGNB requires 5.7x,5.9x,5.1x, and 3.6xfewer slices, and DSP units, respectively, than conventional NN attacks, yet, the proposed GGNB-based methodology has about training and test time, respectively, compared to the existing SVM classifiers. Furthermore, the proposed correlation matrix-based method reduces training time by 18% and 21% using rawCAN [ 1] and OpelAstra [ 2] data sets, respectively. Using a reduced number of features, the proposed method shows 37.5% lower testing time on both the data sets with only a minor 0.03% maximum accuracy Lee, S. H. Jeong, H. K. Kim, Otids: A novel intrusion detection system for in-vehicle network by using remote frame, in: Annual Conference on Privacy, Security and Trust, Vol. 00, 2017, pp. 57-5709. [2]G. Dupont, A. Lekidis, J. den Hartog, S. Etalle, Automo- tive controller area network (CAN) bus intrusion dataset . [3] B. GmbH, CAN specification version 2.0. [4]N. H. T. S. Administration, Vehicle cybersecurity, . [5]V. Tanksale, Intrusion detection for controller area network using support vector machines, in: IEEE International Conference on Mobile Ad Hoc and Sensor Systems Workshops, IEEE, 2019, pp. 121-126. [6]H. M. Song, J. Woo, H. K. Kim, In-vehicle network intrusion detection using deep convolutional neural network, Vehicular Communications 21 (2020) 100198. [7]R. E. Corporation, In-vehicle networking solutions, . [8]K. Koscher, A. Czeskis, F. Roesner, S. Patel, T. Kohno, S. B.Kantor, D.Anderson, H.Shacham, S.Savage, analysis of a modern automobile, in: IEEE Symposium on Security and Privacy, 2010, pp. 447-462. [9]S. Jeong, B. Jeon, B. Chung, H. K. Kim, Convolutional neural network- based intrusion detection system for AVTP streams in networks, Vehicular Communications 29 (2021) . URL Miller, C. Valasek, Remote exploitation of an unaltered Black Hat USA 2015 (2015) 91. 24[11]S. Checkoway, D. McCoy, B. Kantor, D. Anderson, H. Shacham, S. Sav- age, K. Koscher, A. Czeskis, F. Roesner, T. Kohno, analyses of automotive attack surfaces, in: Proceedings of the USENIX Conference on Security, 2011, p. 6. [12]M. R. Moore, R. A. Bridges, F. L. Combs, M. S. Starr, S. J. inter-signal arrival times for accurate detection of CAN bus signal injection attacks: a data-driven approach to in-vehicle in: Proceedings of the Annual Conference on Cyber and Information Security Research, 2017, pp. 1-4. [13]A. Zhou, Z. Li, Y. Shen, Anomaly detection of CAN bus messages using a deep neural network for autonomous vehicles, Applied Sciences 9 (15). [14]V. Verendel, D. K. Nilsson, U. E. Larson, E. Jonsson, An approach to using honeypots in in-vehicle networks, in: IEEE Vehicular 2008, pp. 1-5. [15]S. Tariq, S. Lee, H. K. Kim, S. S. Woo, CAN-ADF: The controller area network attack detection framework, Computers & Security 94 (2020) 101857. . URL Minawi, J. Whelan, A. Almehmadi, K. El-Khatib, Machine learning- based intrusion detection system for controller area networks, in: Pro- ceedings of the ACM Symposium on Design and Analysis of Networks and Applications, DIVANet '20, Association for Computing Machinery, New York, NY, USA, 2020, p. 41?47. [17]M. D. Hossain, H. Inoue, H. Ochiai, D. Fall, Y. Kadobayashi, detection system for in-vehicle can bus communications, IEEE Access 8 (2020) 185489-185502. . [18]K. Lemke, C. Paar, M. Wolf, Embedded security in cars, Springer, 2006. [19]L. Wang, X. Liu, Secure cooperative communication scheme for networks, Vehicular Communications 11 (2018) . URL Islam, R. U. D. Refat, Improving CAN bus security by arbitration IDs, Journal of Transportation Security 13 (1) Velampalli, W. Eberle, Novel graph based anomaly detection using background knowledge, in: The International Flairs Conference, 2017. [22]S. Velampalli, L. Mookiah, W. Eberle, Discovering suspicious patterns using a graph based approach, in: The International Flairs Ueda, R. Kurachi, H. Takada, T. Mizutani, M. Inoue, S. authentication system for in-vehicle network, SEI technical review 81 (2015) 5-9. [24]I. Studnia, V. Nicomette, E. Alata, Y. Deswarte, M. Kaaniche, Y. Laarouchi, Survey on security threats and protection mechanisms in embedded automotive networks, in: Annual IEEE/IFIP Conference on Dependable Systems and Networks Workshop, 2013, pp. 1-12. [25]P. Carsten, T. R. Andel, M. Yampolskiy, J. T. McDonald, S. Russ, A system to recognize intruders in controller area network CAN, Symposium for ICS & SCADA Cyber Security Research, 2015, pp. 111-114. [26]A. Boudguiga, W. Klaudel, A. Boulanger, P. Chiron, A simple method for controller area network, in: IEEE on Communications, 2016, pp. 1-7. [27]J. Staggs, How to hack your mini cooper: reverse engineering CAN messages on passenger automobiles, Institute for Information Security. [28]T. Hoppe, S. Kiltz, J. Dittmann, Security threats to automotive examples and selected short-term Engineering & System Safety 96 (1) (2011) 11-25. [29]S. Woo, H. J. Jo, D. H. Lee, A practical wireless attack on the connected car and security protocol for in-vehicle CAN, IEEE Transactions on intelligent transportation systems 16 (2) (2014) 993-1006. 26

2108.10908

Riadul Islam

Riadul Islam, Maloy K. Devnath, Manar D. Samad, and Syed Md Jaffrey Al Kadry

GGNB: Graph-Based Gaussian Naive Bayes Intrusion Detection System for CAN Bus

cs.CR cs.LG

http://creativecommons.org/licenses/by/4.0/

The national highway traffic safety administration (NHTSA) of the automobile systems are more critical than the security of other information systems. Researchers already demonstrated remote attacks on critical vehicular electronic control units (ECUs) using controller area network (CAN). Besides, existing intrusion detection systems (IDSs) often propose to tackle a specific type of attack, which may leave a to numerous other types of attacks. A generalizable IDS that can identify a wide range of attacks within the shortest possible time has more practical value than attack-specific IDSs, which is not a trivial task to accomplish. In this paper we propose a novel { extbf g}raph-based { extbf n}aive { extbf B}ayes (GGNB) intrusion detection algorithm by leveraging graph properties and PageRank-related features. The GGNB on the real rawCAN data yields 99.61%, 99.83%, 96.79%, and 96.20% detection accuracy for denial of service (DoS), fuzzy, mixed attacks, respectively. Also, using OpelAstra the proposed methodology has 100%, 100%, 99.92%, 97.75% and 99.57% detection accuracy considering DoS, diagnostic, fuzzing CAN ID, fuzzing payload, replay, suspension, and mixed attacks, respectively. The GGNB-based methodology requires about $239 imes$ and $135 imes$ lower training and tests times, respectively, compared to the SVM classifier used in the same application. Using Xilinx Zybo gate array (FPGA) board, the proposed GGNB requires $5.7 imes$, $5.9 imes$, $5.1 imes$, and $3.6 imes$ fewer slices, and DSP units, respectively, than conventional NN architecture.

2 II. SUPERCONDUCTING INSTABILITIES IN DIRAC MATERIALS We investigate the pairing instabilities of a two- dimensional lattice with hexagonal symmetry hosting Dirac points at the Kpoints, which we describe via an extended Hubbard (1) wherecy r= (cy r;;cy r;#) are two component spinors and the spin inner product is implied in the single particle and interacting terms in the Hamiltonian (1). We choose lattice vectors a1= (a;0),a2= specialize to lattices possessing C6vsymmetry, which consists of twofold and threefold rotations in the plane and mirror re ections about the xandyaxes. We only consider models without spin-orbit interaction, thus we may choose a basis of Wannier orbitals so that T(r r0) are real. The Hamiltonian is invariant under SU(2) spin rotations and possesses a time reversal symmetry satisfying T2= 1, which is identical to complex conjugation in the coordinate representation. For the remainder of this paper we denote the time which leaves spin unaected by T. The resulting band structure features Dirac points which we distinguish by a valley index=. The eigenstates of the single particle Hamil- tonian may be classied by their eigenvalues rotations, where =. We shall refer to the degree of freedom as pseudospin . At the Dirac points there are four degenerate energy eigenstates into one another under time reversal, and mirror symmetry. We shall introduce the Pauli operators ;acting on the valley and of freedom. The degeneracy of the eigenstates at each valley is protected by either MxorRT. We obtain the following representation of crystal UMx=x; UR= cosIx+ sinIy;T=xxK (2) whereIis a phase which generally depends on the lattice and the chosen center of slightly above the valleys K=(4 3a;0), the Fermi surface consists of two circular pockets with Fermi momenta kFjKj, with the single particle energy eigenstates formed from linear com- binations of the pseudospin basis states (i.e. eigenstates of threefold rotations). By obtaining the band projecting the interactions onto states near the Dirac points, we obtain a quantum eld theory describ- ing interaction processes close to the Fermi level which accounts for the ,and spin (s) degrees of freedomand may be formulated using an 8-component local eld operator y s(r). The most general Hamiltonian with the symmetries (2) is of the form H= y(ivz
r) +1 2X a;bVab yJa
yJb
(3) where we have expanded the interaction in the adjoint basis consisting of products Ja=, andVabare con- stants which may be obtained by projecting the extended Hubbard interactions (1) onto the pseudospin and As we show in the Appendix, only operators which are even under time reversal symmetry are per- mitted in the interactions, which are Hamiltonian must transform as a scalar under the spatial symmetries (2) which implies that Vis a diagonal matrix. There are only ve constants g1;g2;g3;g4;g5, which we relate to the interaction matrix addition to a density-density interaction g1, the inter- actions involve ( J2=zz) and valley- mixing (J5=x;J6=y) mass operators. The remain- ing six operators are conserved SU(2) valley currents which can be expressed as products of the valley the electric current j= (jx;jy) relation between the coupling constants giappear- ing in the eld theory (3) and the extended Hubbard model parameters U(r1;r2;r3;r4) for the kagome and honeycomb lattices are given in Table I. We denote the three sublattices of the kagome lattice fA;B;Cg, and the two sublattices of the honeycomb lattice by fA;Bg. We list only interactions involving nearest neighbors, and denote the interactions in which the lattice on the In the kagome lattice, each site has two nearest neighbors in each of the other sublattices, and in the honeycomb lattice, each site has three in the other sublattice. The eigenstates of the Dirac Hamiltonian in the upper band are given by fy k;;s = ( y s(k) + eik y s(k))=p 2. Evaluating the Born amplitudes from (3) we nd that the scattering vertex 1234() in the Cooper channel is given by () = ++() = ++() (5)3 Kagome + 8UABAB + 8UAABA + 4UACBC +UAABB )1 2(UAAAA + 3UABAB + 4UAABA4UACBC + 2UAABB ) + 3UAABA3UACBC + 2UAABB ) +UABAB + 4UAABA + 8UACBC + 8UAABB )UAAAA TABLE I: The relations between coupling constants (3) and the extended Hubbard model parameters (1) for the Kagome and honeycomb lattice with couplings involving nearest neighbors. We denote where in sublattices at most by a nearest neighbor bond. whereis the scattering angle. BCS theory predicts that superconducting pairing occurs when the between Cooper pairs is negative, i.e. for attractive scattering. For pairing of electrons within the same valley this corresponds to ` <0, while pairing of electrons in opposite valleys requires the or critical temperature is `  for intravalley pairing or = intervalley pairing, and 0=kF=(2v) is the density of states per spin per valley at the Fermi level. In order to investigate the physical features of the condensate we introduce the mean eld h.c. (7) wherefy kis the creation operator for a Dirac fermion in the upper band and
( k) is the superconducting gap matrix (spin and valley indices are implied). Accounting for spin, valley and angular momentum structure, there are eight possible superconducting phases, which we list in Table II along with the corresponding scattering amplitude. The topological properties of the gaps are associated with their class [ 49{51], which clas- sies mean eld Hamiltonians based on time reversal, charge conjugation and chiral symmetries. For the spin singlet pairing phases,
/s0, accounting for SU(2) spin rotation symmetry we nd that the charge satises C2=1. In the case of spin triplet pairing, the spontaneous violation of SU(2) spin xes a direction d= (dx;dy;dz), with the gap
/ds. As we show in the Appendix, magnetization of the condensate is energetically disfavored, which implies thatdcan be chosen to be purely real. The is invariant under spin rotation about the d-direction. Accounting for this U(1) symmetry [88], we may decompose the mean eld Hamiltonian in two spin blocks, with each separately possessing a charge conjuga- tion symmetry satisfying C2= +1. We therefore conclude that the phases with spin singlet pairing are either in classCI or C, while those with spin triplet pairing are either in class BDI or D depending on whether the of the normal state Hamiltonian survives in the condensed phase. Consulting the periodic table of topo- logical invariants [ 51], we nd that rst order is possible for the phases in which time reversal symmetry is spontaneously broken, i.e. those in class C or D, and posses a Chern number which counts the number of chiral modes propagating along the bound- ary and whose parity is equal to the parity of Majorana modes localized at a vortex. However, both the spin triplet, and intervalley s-wave spin triplet states are in class BDI. For the intravalley spin triplet phase, the gap is p+ipin one valley and pipin the other, preserving time reversal symmetry, and we refer to this as the p+ipphase. In these cases lowest order topology is absent while second order topology is possible [14, 16, 20]. These phases are indicated by the existence of Majorana corner states of denite spin protected by crystalline symmetries. In addition to the spinless time re- versal symmetry operator satisfying T2= +1, our system possesses a spinful time reversal symmetry (T0)2=1 which interchanges opposite spin blocks; thus corner modes occur in Kramers pairs. In Section IV we present exact diagonalization results that the intravalley p+ipspin triplet phase indeed possesses a second order topology. We now apply our previous results to analyze the pos- sibility of superconducting instabilities in the kagome and honeycomb lattices. The leading Hubbard interac- tions do not involve tunneling between sites, thus we only account for interactions U(r1;r2;r3;r4) in (1) Denoting the parameter in (6) by 1for intervalley s-wave spin triplet pairing, 2for spin triplet pairing and 3for spin triplet pairing, we nd for the kagome 2(8) where is the area of the unit cell and 1= 0; 2=3 2=3UABAB 4: (9) for the honeycomb lattice.4 Gap structure
 AZ spin singlet, s-wave CI Intervalley, spin singlet, p-wave eik 1 4(g1g2+g4) C Intervalley, spin triplet, s-wave BDI Intervalley, spin triplet, p-wave D Intravalley, spin singlet, s-wave CI Intravalley, spin singlet, d-wave CI Intravalley, spin singlet, mixed C Intravalley, spin triplet, p-wave BDI TABLE II: Superconducting phases for 2D Dirac fermions in lattices with C6vsymmetry. First column : gap structure, second column : order parameter, third column : coupling constant which determines Tcvia Eq. (6), in terms of coupling constants in the Hamiltonian Eq. (3), fourth column : class. In the limit UABAB!0 the scattering amplitudes for all three spin triplet instabilities vanish for both that the system approaches a critical point for the onset of triplet including for nontrivial phases. The vanishing of the scattering amplitude in the a consequence of the Dirac nature of the and can be understood as follows: the components of the wavefunction in the pseudospin and valley basis have phase factors which are a manifestation of the Berry phase surrounding the Dirac points. Despite our model only containing local interac- tions, upon projecting the Dirac wavefunction onto the interactions in (3) we nd that the Berry phase gives rise top+ipscattering amplitudes which are essential to obtain a nontrivial topology. The pseudospin and of the operators Jaallows some gito con- tribute to the partial wave amplitudes with negative signs, as can be seen in Table II. This causes between the contributions to the with distinct pseudospin and valley structure { i.e. thegi. In the scattering channels which give rise to the topological phases, the contributions from dierent gi sum to zero, leading to a total cancellation of the on-site Coulomb repulsion. An additional negative contribution to the which can overcome the weak nearest neigh- bor repulsion would result in One interesting scenario would be a weak electron-phonon in- teraction, which in our case could drive the instability to an unconventional spin triplet state, rather than the spin singlet state which normally results. In the next section, however, we will show that even in the absence of additional attractive interactions, a negative contribution to the nearest neighbor repulsion arises as a result of over- screening. All three triplet amplitudes become negative if the bare nearest neighbor repulsion is suciently weak, and the density is suciently high.III. SCREENING EFFECTS We now consider how giare modied by screening pro- cesses. We show these eects here analytically in the Random Phase Approximation (RPA), equivalent to tak- ing the limit of large NwhereN= 4 is the degeneracy of the Dirac points [ 53{56]. This analysis was performed for the special case of an articial semiconductor honey- comb lattice in [ 41]; we extend this analysis to Dirac materials satisfying C6vand , which may be described by the Hamiltonian (3). The RPA results in replacing the interaction con- stantsVabin (3) with screened frequency and couplings, which are determined by the (10) where cd(!;q) is a generalized polarization operator. These represent the susceptibility of the system to a perturbation H=P aa(r;t)Ja, and are calculated in the for the frequency and momentum depen- dence of the screened scattering vertex, solution of the Eliashberg equations gives the critical temperature Tc= the RPA renormalised of Table II obtained from interactions between electrons on the Fermi surface and  is a frequency cuto determined by away from the Fermi surface; explicit in the Appendix) shows  EF. Scattering on the Fermi surface corresponds to the range of momentum and frequency transfer q<2kF,!= 0. In this range the polarization operators are simply given by the previous expressions (Table II) evaluated with screened couplings egiat!= 0;q= 0. We have studied the screening of the interactions g1 andg2in a previous paper [ 41], and found that g1is reduced while g2is enhanced as the chemical potential is5 increased. We can explain the qualitative eects of static and homogeneous screening on the interactions on general physical grounds. The coupling g1is screened by the long wavelength density-density response, which as increasing chemical potential. The intra and intervalley mass couplings g2andg4are screened via the polarization operator associated with the response of the system to an external perturbation gap the Dirac spectrum. This perturbation lowers the energy of negative energy states and increases the energy of the positive energy states. The response of the occupied negative energy states below the Dirac point is divergent and may be fully subtracted by regularization { the result of which is that g2andg4should be replaced by reduced interaction constants which are dierent from those provided in Table I; thus within the Dirac model the only physically signicant eect of screening is the increase in energy of the occupied positive energy states which grows stronger with increasing chemical potential. The interactions g3;g5are valley current interactions and the polarization operators responsible for their screening also contain a UV divergent part which may be subtracted by replacing the bare interaction constants with reduced values, and after this subtraction are unscreened due to a Ward identity. Explicitly, we nd in RPA eg1=g1 1 (12) wherei= 2;4 andj= 3;5. This results in replac- ing the nearest-neighbor coupling with a screened in the relations in Table I and (8), (9). The screened nearest neighbor interaction is given by eUABAB the kagome lattice and eUABAB the honeycomb lattice. For suciently large 0, we nd thateUABAB<0, resulting in and the dominant instability occurs in the intravalley, p-wave, spin triplet channel. This occurs for  > with the critical value = the kagome lattice and = the honeycomb lattice. The Dirac theory is only applicable up to a cuto be- yond which the electronic dispersion becomes nonlinear. Our analysis predicts a superconducting instability when the critical doping lies within the regime where Dirac theory is justied. In the limit g3!0 org2!0 respec- tively for the kagome and honeycomb lattices, the critical doping becomes innite. We therefore require dependent couplings, which are suppressed when the atomic orbitals are delocalized. So far, we have neglected the long ranged part of the Coulomb interaction, which provides weak couplings be- tween sites with large separation. This may be accounted for by replacing eg1with its screened value at q!0, eg1!1=N 0, and simply results in a shift of the critical doping to higher values. IV. TOPOLOGICAL EDGE AND CORNER MODES Our RPA results show that the dominant instability for both the honeycomb and Kagome lattice is the intravalley p{wave spin triplet phase, with gap structure
This gap possesses a number of unconventional features that distin- guish it from the superconducting phases more commonly studied in either high- or Dirac mate- rials. The gap exists at both Kpoints and has a p+ip wave angular structure, which points to the presence of a{Berry phase in two-particle scattering from pairing originates. However, the winding of the gap is opposite in the two valleys, with the order parameter at the two Kpoints being related to each other by complex conjugation. As a re- sult, the usual topological features associated with chiral p-wave superconductors are strongly modied. We may view the phase as consisting of four copies of a [ 10], one per spin and valley species, with the gaps at opposite valleys possessing of opposite sign. This results in a superconductor according to the classication. However, in the presence of a second order topology emerges which is characterized by anomalous corner modes. In the intravalley paired state, Cooper pairs carry a nite quasimomentum, and as a result the gap is forming a pair density wave which is commen- surate with the lattice with a periodicity of three unit cells [ 57{64]. The condensate possesses a U(1) order with the spatial position of the pair density wave and does not couple to the the system also exhibits Goldstone modes associated with uctuations of , which are physically the sliding modes of the pair density wave. The spatial modulation of the gap spontaneously breaks the crystalline symme- tries of the normal state except at special values =n 2, n2Z; at these values the gap may be classied according6 (a) (b) (c) 0.0 0.2 0.4 0.6 0.8 [t] (d) (e) (f) 0.0 0.2 0.4 0.6 0.8 [t] (g) (h) (i) 0.0 0.2 0.4 0.6 0.8 [t] FIG. 1: Results of the exact diagonalization of the Bogoliubov-de Gennes Hamiltonian within a single spin sector for hexagonal akes of a kagome lattice of three dierent geometries. The leftmost gures show the spatially modulated pairing between nearest neighbors for ==2. The color of the bonds between neighboring sites r;r0shows the value of the gap function
(r;r0) according to the color scale. The inset of the bottom left gure depicts a zoom into the bulk of the lattice with the unit cell boundary indicated by the thin lines. The central gures show the wavefunction densities of the three lowest The rightmost gures show the quasiparticle spectrum as a function of the pair density wave parameter . The three lowest quasiparticle modes are indicated in red. For the exact diagonalization we use the parameters = 0:4t;
0= 0:4t. to a second order topological invariant which predicts the number of Majorana corner modes of a nite sam- ple of the superconductor which preserves the study the topology of the p+ipintravalley spintriplet phase via the bulk boundary correspondence for higher order topological superconductors, and identify the anomalous boundary physics in hexagonal akes. By choosing a spin quantization axis which is aligned so that dk^y, we nd that pairing occurs between7 (a) (b) (c) 0.0 0.2 0.4 0.6 0.8 [t] (d) (e) (f) 0.0 0.2 0.4 0.6 0.8 [t] (g) (h) (i) 0.0 0.2 0.4 0.6 0.8 [t] FIG. 2: Results of the exact diagonalization of the Bogoliubov-de Gennes Hamiltonian within a single spin sector for hexagonal akes of a honeycomb lattice of three dierent geometries. The leftmost gures show the spatially modulated pairing between nearest neighbors for ==2. The color of the bonds between neighboring sites r;r0shows the value of the gap function
(r;r0) according to the color scale. The inset of the bottom left gure depicts a zoom into the bulk of the lattice with the unit cell boundary indicated by the thin lines. The central gures show the wavefunction densities of the three lowest The rightmost gures show the quasiparticle spectrum as a function of the pair density wave parameter . The three lowest quasiparticle modes are indicated in red. For the exact diagonalization we use the parameters = 0:4t;
0= 0:2t. electrons with both spins aligned either along the zorz axis. Thus the pairing term may be decomposed into a sum of two spin sectors. We derive the topological features of the superconducting state by analyzing the corresponding to pairing within a single spinsector. In order to make numerical diagonalization easier we employ a simplied form of the gap function in which we only include nearest neighbor pairing and hopping terms. In the Appendix, we derive the real space form of the gap from the p+ipmomentum space structure.8 The spinless mean eld lattice Hamiltonian which
(r0;r) =
(r;r0), and
(r;r0) =
0cos(K
(r+r0) the kagome lattice, and
(r;r0) =
0sin(K
(r+r0) the honeycomb lattice, and we distinguish the param- eter
0from the bulk gap
. We present exact diagonalization results for the kagome and honeycomb lattices in Figs. 1, 2 respectively. We have calculated the spectrum for all values of , however the topological classication must be performed at the values= 0; 2for which the gap obeys a twofold . We have performed calculations for a variety of edge geometries, and in each case we preserve the twofold rotational symmetry of the bulk. In a number of cases we observe a gapless edge spectrum for all values of. This occurs for the geometries shown in the last two rows of Fig. 1 and the last row of Fig. 2. These surfaces states are remnants of the chiral Majorana edge modes expected from the Read-Green model [ 10] which are the most easily identiable manifestions of the chiral p-wave nature of the gap. For other edge the edge spectrum is gapped, indicating that the gapless edge modes are not topologically protected, which is consistent with the fact that, since our gap lies in class BDI, it cannot possess a rst order For the other edge terminations, we nd that at= 2, the edge states are gapped, and six zero energy modes appear within each spin sector which are localized at each of the corners of the ake, forming the three lowest modes within the quasiparticle spectrum of a single spin sector highlighted in red in Figs. 1, 2c,f,i. This occurs for both the honeycomb and kagome lattices. At = 0, we either nd no Majorana zero modes when the edge spectrum is gapped or a gapless edge spectrum. This allows us to conclude that both the honeycomb and kagome lattices are in a second order topological trivial for = 0. We note that, since our results are for a spinless model obtained after performing a spin decomposition of the mean-eld Hamiltonian, the Majorana zero modes appear as Kramers pairs. Since pairing occurs between electrons with parallel spin, the Majorana modes individually possess denite spin, and are protected against hybridization by a spinful time reversal symmetry T0=eiSyT.Since the boundary modes in a higher order topolog- ical phase require crystalline symmetries, they are not protected against local perturbations. The robustness of the bulk-boundary correspondence in the presence of crystalline symmetry breaking therefore provides an im- portant, general question for systems exhibiting higher order topology. In our examples, for general values of the pair density wave order parameter , all the point group symmetries of the normal state are spontaneously broken by the pairing term except for mirror re ection about thexaxis, which is preserved due to the fact that +of the pair density wave does not depend on the ycoordinate. However, at discrete values =n 2, the superconductor recovers a twofold The parameter allows continuous tuning of the system between a second order topological phase at =(n+1 2) and a trivial phase at =n. We observe, despite the breaking of crystalline symmetries, that the anomalous corner modes persist and remain to zero energy in a nite region of values to a critical value at which the edge gap closes. The corner modes are protected by both the bulk and edge gap, and thus persist even when the point group symmetry is weakly broken [13, 14]. The anomalous corner modes coexist with a gapped edge spectrum. This originates from the p+ipnature of the gap, which provides a 1D boundary each copy of the Moore-Read model existing in each spin and valley species. The manifestation of the bulk second order topology is shadowed by the behavior of the boundary superconductor: for certain edge edge modes remain anomalous due to symmetries along the edge which forbid their hybridization. In partic- ular, for the bottom geometries of Fig. 1 and 2, the system possesses a mirror symmetry at =n 2that prohibits chiral Majorana modes from gapping out. In this case, the edge states are anomalous and associated with a bulk crystalline symmetry phase protected by mirror symmetry. For other edge terminations, mixing of the chiral modes be- tween opposite valleys along the same edge occurs, and the edge spectrum is gapped. Interestingly, the emergence of anomalous corner modes in this situation is not general but depends on the value of . We may make an interesting comparison to the p+ ipintervalley spin triplet phase, which is subleading in our analysis in the regime where the onsite over the nearest neighbor couplings, but may arise as the dominant instability for models with more complicated lattice interactions. In this case, pairing occurs between the electron pockets surrounding and the order parameter spontaneously chooses a winding number corresponding to either Time reversal symmetry is broken, which puts the superconductor in class D, rather than BDI as in the intravalley case. This phase possesses a Chern number which leads to a gapless, anomalous edge spectrum. In a forthcoming study, we show that this phase may become9 dominant when spin orbit coupling is present, and analyze the topology of this phase in detail. V. DISCUSSION Our results show that the rst order topological p+ip and second order topological emerge naturally in the presence of Dirac points near the chemical potential. While in a previous study we explored the pseudospin pairing mechanism giving rise to in a lattice within the RPA [ 41], in this work we have extended our analysis of the pairing mechanism to generic lattices satisfying C6vsymmetry and shown that this instability is not specic to any lattice structure but rather emerges due to the universal properties of the Dirac fermion excitations in these materials. We have illustrated that both the Berry phase and screening eects underlying the pairing mechanism are active in kagome lattices, and presented exact diagonalization results which demonstrate the striking similarity between the boundary physics in kagome and honeycomb lattices, which result from the fact that they possess identical topological properties. In addition to the p+ipphase, the intervalley p+ip ands-wave phase are possible, but are subleading in our analysis. Unlike the p-wave states, the s-wave state is protected against the eects of weak disorder by a generalized Anderson's theorem, so this state might posses a higherTcthan the topological phases in a system. In a forthcoming study we also suggest that this phase may arise due to spin-orbit coupling and study its properties in detail. We call attention to a number of relevant features of this mechanism: ( i) the mechanism can be applied for weak repulsive interactions, so strong correlations are not necessary, ( ii) appears beyond a critical doping beyond the Dirac points, at which the screening eects become strong enough to give rise to a net attraction, ( iii) the electronic origin of the pairing means the cuto which determines Tcis the Fermi energy rather than the Debye frequency, ( iv) is more pronounced in systems with localized has recently been discovered in quasi two dimensional vanadium-based kagome metals AV 3Sb3 (A=K,Rb,Cs) [ 42,45]. The agreement between DFT cal- culations of the bandstructure and ARPES suggests that these materials are weakly correlated, an supported by the results of DFT/DMFT The Fermi surface consists of two the Dirac points, a pockets surrounding the point and an additional Fermi contour [66]. exists alongside density wave order which sets in at a higher temperature and appears to compete with [ 67{73]. STM and Josephson STS mea- surements suggest a spatially modulated [ 47] and zero bias peaks inside magnetic cores are suggestive of Majorana bound states [ 48].Given that phonons appear unable to account for and yet the materials appear to be weakly correlated, we argue that our mechanism might provide an explanation for in kagome metals, in contrast to existing theoretical proposals which attribute to the eects of nesting or competing density wave order [ 75{77]. In this originates from the Dirac-like Fermi surfaces and is not directly related to the observed charge density wave order [ 67], which is potentially due to van Hove singularities or nested portions of the band struc- ture near the Fermi level. Future theoretical work the interplay between the pairing mechanism and the presence of density wave order, which symmetry [ 68]. Future experiments could look for evidence of the edge or corner modes present in [ 22,23], probe the spin structure of the gap through NMR, or further investigate the real space structure of the superconducting gap. Additional kagome systems have recently been discov- ered to host and might also be for the pseudospin mechanism, including bi- layer kagome systems [ 78], ferromagnetic kagome metals [79], and lanthanum-based materials [ 80]. has also been seen in Dirac surface states of insulators [ 82{86], which might also be ex- plained by this theory. In materials such as transition metal dichalcogenides like MoS 2and MoTe 2, and few- layer stanene [ 87], the presence of spin orbit the Dirac points and/or introduces a but the Dirac physics persists nonetheless. In a forthcoming paper, we have examined the eects of spin- orbit coupling on the possible superconducting states, and nd that the pseudospin mechanism results in in these systems as well. As a general guiding principle, promising candidate materials for this mechanism are those with strongly localized orbitals { which lead to larger values of the couplings giand lower values of the critical doping { and larger Dirac-like Fermi surfaces { which enhance the screening eects. VI. authors thank P. Brouwer, M. Scheurer and P. Kotetes for helpful discussions. MG acknowledges support by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program under grant agreement No. 856526, and from the Deutsche (DFG) Project Grant 277101999 within the CRC network TR 183 Entangled States of Matter (subproject A03 and C01), and from the Danish National the Danish Council for Independent Research | Natural Sciences. H. D. Scammell acknowledges funding from ARC Centre of Excellence FLEET. TL from the DFG within the CRC network TR 183.10 [1]M. Sigrist, Introduction to Unconventional , AIP Conference Proceedings 789, 165 (2005). [2]S. Maiti and A. V. Chubukov, from repulsive interaction, AIP Conference Proceedings 1550 , 3 (2013). [3]A. J. Leggett, theoretical description of the new phases of liquid3He, Rev. Mod. Phys. 47, 331 (1975). [4]R. B. Laughlin, Magnetic induction of in high- , Phys. Rev. Lett. 80, 5188{5191 (1998). [5]T. Senthil, J. B. Marston and M. P. A. Fisher, Spin quantum Hall eect in unconventional Rev. B 60, 4245{4254 (1999). [6]B. Horovitz and A. Golub, Superconductors with symmetry: Spontaneous magnetization and quantum Hall eects, Phys. Rev. B 68, 214503 (2003). [7]A. P. Mackenzie and Y. Maeno, The of Sr 2RuO 4and the physics of spin-triplet pairing, Rev. Mod. Phys. 75, 657{712 (2003). [8]W. Kohn and J. M. Luttinger, ew mechanism for Phys. Rev. Lett. 15, 524{526 (1965). [9]G. E. Volovik, edge states in superconductors with time inversion symmetry breaking, J. Exp. Theor. Phys. Lett. 66, 522{527 (1997). [10]N. Read and D. Green, Paired states of fermions in two dimensions with breaking of parity and and the fractional quantum Hall eect, Phys. Rev. B 61, 10267 (2000). [11]R. Nandkishore, L. S. Levitov and A. V. Chubukov, Chi- ral from repulsive interactions in doped graphene, Nat. Phys. 8, 158-163 (2012). [12]Y. Peng, Y. Bao, and F. von Oppen, Boundary Green functions of topological insulators and Rev. B 95, 235143 (2017). [13]J. Langbehn, Y. Peng, L. Trifunovic, F. von Oppen and P. W. Brouwer, Re ection symmetric second-order topo- logical insulators and Phys. Rev. Lett. 119, 246401 (2017). [14]M. Geier, L. Trifunovic, M. Hoskam and P. W. topological insulators and an order-two crystalline symmetry, Phys. Rev. B 97, 205135 (2018). [15]M. Geier, P. W. Brouwer and L. Trifunovic, Symmetry- based indicators for topological Bogoliubov-de Phys. Rev. B 101, 245128 (2020). [16]L. Trifunovic and P. W. Brouwer, Higher-order bulk- boundary correspondence for topological Phys. Rev. X 9, 011012 (2019). [17]L. Trifunovic and P. W. Brouwer, Higher-order topologi- cal band structures, physica status solidi (b), 2020. [18]K. Shiozaki, Variants of the symmetry-based Ono, H.-C. Po and H. Watanabe, Rened for topological superconductors in all space groups, Sci. Adv. 6, eaaz8367 (2020). [20]E. Khalaf, Higher-order topological insulators and protected by inversion symmetry, Phys. Rev. B 97, 205136 Zhang, W. B. Rui, A. Calzona, S.-J. Choi, A. P. Schnyder, B. Trauzettel, Topological and computation based on second-order topolog-ical Phys. Rev. Research 2, J. Gray et al., Evidence for Helical Hinge Zero Modes in an Fe-Based Superconductor, Nano Lett. 19, Choi et al., Evidence of higher-order topology in multilayer WTe 2from Josephson coupling hinge states, Nat. Mater. 19, 974{979 (2020). [24]J. Labbe and J. Bok, in La 2CuO 4: A Theoretical Model Eu- rophys. Lett. 3, 1225 (1987). [25]I. E. Dzyaloshinskii, Superconducting transitions due to Van Hove singularities in the electron spectrum, Zh. Eksp. Teor. Fiz. 93, 1487 (1987) [Sov. Phys.-JETP 66, 848 (1987)]. [26]J. Friedel, The high- : a J. Phys. Condens. Matt. 1, 7757 (1989). [27]R.S. Markiewicz, Survey of the Van Hove Scenario for With Special Emphasis on Pseudogaps and Striped Phases, J. Phys. Chem. Solids 58, 1179 (1997). [28]S. Maiti and A. V. Chubukov, Renormalization group ow, competing phases and the structure of gap in multiband models of iron-based Phys. Rev. B 82, 214515 (2010). [29]J. L. McChesney et al., Extended van Hove Singularity and Superconducting Instability in Doped Graphene, Phys. Rev. Lett. 104, 136803 (2010). [30]M. Kiesel et al., Competing many{body unconventional in graphene, Phys. Rev. B 86, 020507(R) Zhang et al.,Higher Order Topology and in Fe(Se,Te) Phys. Rev. Lett. 123, 167001 (2019). [32]X. Zhu, Second-order Topological Superconductors with Mixed Pairing, Phys. Rev. Lett. 122, 236401 (2019). [33]S. Franca, D. V. Efremov and I. C. Fulga, Phase topological superconductor, Phys. Rev. B 100, 075415 (2019). [34]Z. Wu, Z. Yan and W. Huang, Higher-order : possible realization in Fermi gases and Sr2RuO 4, Phys. Rev. B 99, 020508 (2019). [35]B. Roy, Higher-order topological superconductors in P-, T-odd quadrupolar Dirac materials, Phys. Rev. B 101, 220506 (2020). [36]J. Ahn and B.-J. Yang, Higher-Order Topological of Spin-Polarized Fermions Phys. Rev. Research 2, 012060 Zhang and S. Das Sarma, Intrinsic Topological in Thin Films of Iron-based Phys. Rev. Lett. 126, 137001 (2021). [38]A. Chew, Higher-Order Topological in Twisted Bilayer Graphene, arXiv:2108.05373 Y.-T. Hsu et al., higher order topo- logical in monolayer WTe 2, Phys. Rev. Lett. 125, 097001 (2020). [40]Y. Wang, M. Lin and T. L. Hughes, Weak-pairing higher order topological Phys. Rev. B 98, 165144 (2018).11 [41]T. Li, J. Ingham and H. D. Scammell rticial in a honeycomb super- lattice, Phys. Rev. Research 2, 043155 (2020). [42]B. R. Ortiz et al., CsV 3Sb5: AZ2Topological Kagome Metal with a Superconducting Ground State, Phys. Rev. Lett. 125, 247002 (2020). [43]C. C. Zhu et al., Double-dome under pressure in the V-based Kagome metals AV 3Sb5(A = Rb and K), Y. Chen et al., Double Superconducting Dome and Triple Enhancement of T cin the Kagome 3Sb5under High Pressure, Phys. Rev. Lett. 126, 247001 (2021). [45]B. R. Ortiz et al., in the Z2kagome metal KV 3Sb5, Phys. Rev. Materials 5, 034801 (2021). [46]S. Ni et al., nisotropic superconducting properties of Kagome metal CsV 3Sb5, Chin. Phys. Lett. Chen et al., Roton pair density wave and unconven- tional strong-coupling in a metal, Liang et al., charge density wave and robust zero-bias conductance peak inside the vortex core of a kagome 3Sb5Phys. Rev. X 11, 031026 (2021). [49]A. Altland and M. R. Zirnbauer, ovel Symmetry Classes in Mesoscopic Hybrid Struc- tures, Phys. Rev. B 55, 1142 (1997). [50]C-K Chiu et al., Classication of topological quantum matter with symmetries, Rev. Mod. Phys. 88, Kitaev, Periodic table for topological insulators and AIP Conference Proceedings 1134 , P. Schnyder, S. Ryu, A. Furusaki, and A. W. W. Ludwig, Classication of topological insulators and in three spatial dimensions, Phys. Rev. B, 78, 195125 (2008). [53]J. Gonz alez, F. Guinea, and M. A. H. behavior from interaction, Phys. Rev. B 59, R2474 (1999). [54]D. T. Son, Quantum critical point in graphene ap- proached in the limit of innitely strong Coulomb in- teraction, Phys. Rev. B 75, 235423 (2007). [55]E. H. Hwang and S. Das Sarma, Dielectric and plasmons in two-dimensional graphene, Phys. Rev. B 75, 205418 (2007). [56]V. N. Kotov, B. Uchoa and A. H. Castro Neto,
=Nex- pansion in correlated graphene, Phys. Rev. B 80, Fulde, R. A. Ferrell, in a eld, Phys. Rev. 135, A550{A563 (1964). [58]A. I. Larkin, Y. N. Ovchinnikov, onuniform state of Sov. Phys. JETP 20, 762 (1965). [59]B. Roy and I. F. Herbut, nconventional on honeycomb lattice: the theory of Kekul e Physical Review B, 82, 035429 (2010). [60]S. Tsuchiya, J. Goryo, E. Arahata and M. condensation and intra-valley pairing states in honeycomb Dirac systems, Phys. Rev. B 94 , Honerkamp, Density Waves and Cooper Pairing on the Honeycomb Lattice, Phys. Rev. Lett. 100, Ma, Z. Huang, F. Hu and H.-Q. Lin, Pairing in graphene: A quantum Monte Carlo study, Phys. Rev. B 84, 121410(R) (2011). [63]W. Wu, M. M. Scherer, C. Honerkamp and K. Le Hur, Correlated Dirac particles and on the honeycomb lattice, Phys. Rev. B 87, 094521 (2013). [64]A. A. Vladimirov, D. Ihle and N. M. Plakida, of strongly correlated electrons on the honey- comb lattice, Eur. Phys. J. B 92, 135 (2019). [65]J. Zhao, Electronic correlations in the normal state of the kagome superconductor KV 3Sb5, Phys. Rev. B 103, L241117 (2021). [66]M. Kang et al., Twofold van Hove singularity and origin of charge order in topological kagome 3Sb5, arXiv:2105.01689 Jiang et al., nconventional chiral charge order in kagome superconductor KV 3Sb5, Nat. Mater. (2021). [68]H. Li et al., Rotation symmetry breaking in the normal state of a kagome superconductor KV R. Ortiz et al., ew kagome prototype of KV 3Sb5, RbV 3Sb5, and CsV 3Sb5 Phys. Rev. Mat. 3, 094407 (2019). [70]H. Zhao et al., Cascade of correlated electron states in a kagome superconductor CsV 3Sb5 X. Li et al., Observation of Density Wave without Acoustic Phonon Anomaly in Kagome Superconductors AV (2021). [72]T. Qian et al., Revealing the competition between charge- density wave and in CsV strain arXiv:2107.04545 (2021) [73]M. H. Christensen et al., Theory of the inAV3Sb5kagome metals arXiv:2107.04546 Tan et al., Charge density waves and electronic prop- erties of superconducting kagome metals, Phys. Rev. Lett. 127, 046401 (2021). from Julian Ingham to everyone: 3:02 PM [75]T. Park, M. Ye and L. Balents Electronic instabilities of kagome metals: saddle points and Landau theory, Phys. Rev. B 104, 035142 (2021). [76]X. Wu et al., ature of unconventional pairing in the kagome superconductors AV3Sb5, Lin and R. M. Nandkishore, Kagome from Pomeranchuk uctuations in charge density wave metals, arXiv:2107.09050 Baidya et al., Interplay of Magnetism and Topological in Bilayer Kagome Metals, Phys. Rev. Lett. 125, 026401 (2020). [79]L. Ye et al., Massive Dirac fermions in a metal, Nature 555, 638{642 (2018). [80]C. Mielke et al., odeless kagome in LaRu 3Si2, Phys. Rev. Materials 5, 034803 (2021). [81] H. D. Scammell, J. Ingham and T. Li, to appear. [82]G. Bian et al., Topological nodal-line fermions in spin- orbit metal PbTaSe 2, Nat. Comm. 7, 10556 (2016). [83]M. Neupane et al., Observation of the state in a Nat. Comm. 7, 13315 L., Xu, SY., Xia, Y. et al. Observation of in a superconducting doped topological Phys 6, 855{859 (2010). [85]Z. Wang et al., in Tl a Topological Insulator, Chem. Mater. 28, Q. Han et al., Electronic structure of a topological insulator Sr-doped Bi 2Se3, Appl. Phys. Lett. 107, 171602 (2015).[87]M. Liao et al., in few-layer stanene Nat. Phys. 14, 344{348 (2018). [88]We note that the pairing term only possesses U(1) spin symmetry, however the normal state dispersion retains its original SU(2) spin symmetry. It was noted in [ 52] that such a state generically falls in class SYMMETRY RELATIONS FOR THE EFFECTIVE FIELD THEORY We may explicitly derive the interaction parameters Vabin Eq. 3 of the main text by projecting the extended Hubbard interactions onto the valley and pseudospin eigenstates u(r). We (A1) The Hubbard interactions may be expressed in terms of the Wannier orbitals r(x) (withxbeing a 3D is the Coulomb interaction, and we have chosen the Wannier orbitals to be purely real. The satisfy U(r1;r2;r3;r4) may use the representation of time reversal symmetry (Eq. (2) in the main text) T= Kwith =xxto derive a (A3) Applying this to the factors u11(r1) andu33(r3) in (A1) we (A4) We may express this in the (A5) Since the adjoint basis elements Jaare Hermitian and transform with denite sign under time reversal, we nd thatVab= 0 for any operator Jawhich is odd under time reversal. We may derive further symmetry constaints using the (A6) where  is a symmetry operation acting on coordinates and Uis its unitary representation. Applying a ri!riin (A1) we (A7) Since there always exists a symmetry operation under which JaandJbtransform with opposite signs unless a=b, we nd thatVabis a diagonal matrix.2 A2. LANDAU-GINZBURG ANALYSIS OF THE SUPERCONDUCTING GAP SYMMETRY The generic mean eld Hamiltonian to account for all pairing possibilities, in the upper band, (A8) The gap takes any of the forms presented in Table II, which contains the valley, spatial and spin structure. The spin structure employs standard notation of ); d0=1 2(j#ij#i ): (A9) The analysis presented in the main text states that the gap functions can be chosen as simply one of the sets; for the gap structure discussed in Section IV it is important that we can choose the d-vector along one axis
/dysy, and that a non-unitary choice of
kis energetically penalized. We explicitly show that this is the case below. We compute the free energy to quartic order from (A8), and are the Cooper channel Born amplitudes, explicitly given in (5), and the trace includes summation over momentum and frequency. Each gap structure entering Table II has a critical temperature set by the coupling , where= `  for while = intervalley pairing. We restrict our attention to work within each degenerate set of gaps separately. In most cases, the valley structure is trivially traced out { the free energy reduces to separate cases based on the overall spin degeneracy d
sand degenerate chiral phase factors eik. Using this logic, there are only ve distinct cases to (A11) Case (iii) enumerates the two physically distinct valley structures that share the same spin structure and chirality, but themselves are not degenerate, i.e. have dierent Tc. The diering valley structures of these gaps does not aect the contribution toFthat is quartic in
; the valley structure enters through and aects the value of Tcbut not the stability analysis. The brackets in ( v) accounts for the four physically distinct gap structures which are degenerate, i.e. have the same value of and hence have same critical temperature, c.f. Table II. We combine these three gap functions into a vector ( v0;v2;u1), which will use below in the free energy the free energy to quartic order in the gaps, F1[d0s0] +b jd0+j4+ =a3(TTc3)(d
d) +bn (d +
d+)2+ (d
d)2+ coecients ai>0;b> 0;i= 1;2;3. We summarize the main (i) is as-wave singlet paired state. The free energy establishes stability of this phase. *Case (ii) corresponds to an intervalley p+ipstate, in which electrons from opposite valleys pair with a we nd that coexistence of opposite chiralities is energetically penalized so that the system spontaneously chooses one chirality, e.g. d0 +s0e+ik, therefore breaking time reversal. This state exhibit rst order topology with gapless edge states. *Case (iii) concerns two triplet paired states for which the free energy has the same form; from F3[d
s] we nd that non-unitary paring is penalised, i.e. dis purely real (or purely imaginary); the dvector a direction, breaking spin SU(2) symmetry, but the condensate has vanishing magnetization. The states included in this case are: intravalley p+ip, in which electrons undergo p+ippairing in one valley and pippairing in the other, and intervalley s-wave, in which electrons from opposite valleys undergo triplet s-wave pairing. Both states respect time-reversal symmetry. *Case (iv) examines spin-triplet p-wave intervalley pairing. From F4[d
seik], similar to case ( ii), the resulting state spontaneously breaks time reversal symmetry, resulting in a rst order topological p+ipstate e.g.d+e+ik. As in case ( iii), we again nd the d-vector is purely real (or imaginary), and hence this state breaks time but does not support a (v) looks at a degenerate manifold of s-wave,d+id, and a mixed state with s-wave in one valley and a d+idpairing in the opposite valley, the latter of which breaks time reversal. The intravalley d-wave state is { a time reversal invariant combination of d+idpairing in one valley, and didpairing in the other { and exhibits second order topology with gapless corner states.4 A3. SCREENED INTERACTIONS In this appendix we discuss the screened interactions. We begin by presenting general expressions for the polarization operators, and making some general comments about regulating UV divergences in eective quantum eld theories. We then explicitly calculate the intra and inter valley current-current We present the formulae for the screened couplings, and plot the frequency dependence of the resulting scattering amplitudes for Cooper pairs. We conclude by discussing alternative regularizations for computing the A. RPA equations for the screened interactions are eV(!;q) =V(q) (A12) where  is the polarization the vertices J=fJ1;:::;J as per Section ??of the main text. We choose units in which v= 1, and upon performing the frequency integral by presented in [A1] result in (!;q) = (A14) The rst term { the interband polarization operator, denoted  +{ contributes when = 0, while the second term { the intraband polarization operator, denoted  { only contributes when 6= 0. Explicitly, we write  +(!;q) = polarization operator 11is the standard charge polarization operator for graphene [ A2{A4], which describes the screening of the density-density interaction 00 00. The function 22is the pseudospin polarization operator which describes the screening behaviour of the interaction zz zzwhich was rst calculated in [A1] (denoted there as zz;00due to dierent notation and a dierence choice of basis for the pseudospin states). The remaining functions describe the screening of the intra and intervalley chiral currents. The current-current graphene has been discussed in [ A5{A7], though to the best of our knowledge the intervalley has not been previously investigated. Below, it will be convenient for us to derive expressions for both functions, and present them in a form somewhat more general than those already in the literature. All polarization operators may all be written in terms of four basic functions, for which we will introduce simple notation in what follows. Use the spacetime indices J= (00;0i) wherei=x;y, and denote 22as zz. Considering ie the chiral current screening 33;34;44, we then write this function as ij and decompose into longitudinal and transverse parts: ij(!;q) = + chiral current polarization operator obeys an important identity due to electromagnetic gauge invariance. In the Dirac theory, the electromagnetic current operator J=so that the vector potential of electromagnetism A5 appears in the Hamiltonian as the perturbation eJ
A. Gauge invariance, or equivalently the conservation of charge, can be shown to imply the Ward = 0 (A18) From this we = = 0 Combining =qiqjij(!;q) =q4?(!;q) (A19) which gives us an expression for the longitudinal part of the current-current susceptibility in terms of the =!2 q400(!;q) (A20) Comparing our below results with the expression for 00cite [ A1] shows our expressions satisfy this relation. We shall also derive an expression for the transverse part of the susceptibility in the dimensional regularization + zz(!;q) (A21) which therefore allows the current-current screening operator to be written entirely in terms of the density and pseudospin responses. As is common quantum eld theories, the polarization operators are formally divergent quantities and However, the physical origins of the divergences and the eects of regularization dier between the polarization operators. The dimensional regularization scheme has the eect of simply setting all UV contributions to zero, leaving just the eects of those degrees of freedom in the Dirac eective theory near the K-points. Since actual materials are UV-completed by a lattice, placing the theory on a lattice is a more physical regularization scheme, and gives the physical relation between the numerical values of the couplings computed by band structure methods and the couplings in the eective theory. Firstly, the charge susceptibility 00(!;q) obeys the exact compressibility sum rule, 00(!= 0;q!0) =0where 0is the density of states at the Fermi level, which means that cuto-dependent quantities never appear in any scheme. For small frequencies and momenta, the regularized polarization operator must be the same as that calculated in dimensional the current susceptibilities ij(!;q) obey the Ward identity discussed above, as a result of the fact that is the current operator in the Dirac theory, which also excludes cuto-dependent contributions. However, away from the K-points, the current operator is no longer given by , which will result in UV-dependent contributions to ij(!;q). In the lattice regularization scheme, this means that one can make changes to the lattice { for instance by modifying the dispersion near the point { that will result in a dierent value for ij(!;q), but these UV-contributions originate from physics away from the Dirac point where the Ward identity applies. Similar to 00(!;q), for small frequencies and momenta, the regularized ij(!;q) must be the same in lattice and dimensional pseudospin susceptibility zz(!;q), by contrast, obeys no such constraint and receives a constant and negative cuto dependent contribution in a lattice regularization. We may distinguish between two types of cuto dependence { contributions which originate from physics near the Dirac point, and contributions which originate from physics away from the Dirac point. While only the latter aects ij(!;q), as we explained above, both types of in zz(!;q); one way to see this is to use a hard cuto  in the Dirac theory, where (restoring the Dirac velocityv) one nds 00(!= 0;q!0) ==(2v) +0. Thus by solely modifying physics near the Dirac point { for instance by changing the velocity { these UV contributions which originate from physics near the Dirac point are changed. As a result, these contributions which appear in a lattice regularization are important even at and momenta, and are described in the main text. When working with an eective eld theory, one always performs calculations in terms of UV-independent if one wishes to perform a microscopic calculation of the couplings giin(??){ by computing the wavefunctions of a material for e.g. through band structure diagonalization and then taking matrix elements of the Coulomb interaction { the couplings which appear in the eective quantum eld theory will be related to those values through in which ij(!;q) and zz(!;q) receive contributions from physics away from the Dirac point, and zz(!;q) receives an additional constant contribution from physics near the Dirac point.6 B. Intravalley current-current polarization operator We now turn to a calculation of ij(!;q). We begin by calculating the interband part through using the rst on the denominator in Eq. (A16). Using the (1x)B)2(A24) we we now use relativistic notation l= (E;k),p= (!;q), in mind that this will aect the numerator as well), and Wick rotating E!iE, the expression corresponding numerator (before Wick rotation) [xq+k]
) (A27) Now we substitute perform the pseudospin trace using the 2ACij+
= x(x1)p2. Using rotational symmetry to replace ( li)2!1 3l2, we arrive at ij =
)2(A29) where we account for the valley and spin trace through a factor N= 4. Integrating using Eq. (A23) , we nd the imaginary part Im ij = ==8, we arrive at Im +qiqj (!q) (A31) Similarly using Eq. (A22), the real part is calculated from Eq. (A29) to be Re +qiqj (q!) (A32) The functions ij ?( ;q) and ij k( ;q) are respectively the transverse and longitudinal components of the From above, we see the their interband parts the density response J=J=0through the formalism above, one may see that the above expression for ij ksatises the Ward identity (A20). Let us now perform the intraband calculation. The numerator in Eq. (A15) is evaluated using Eq. (A28), so that ij +(!;q) = (A35) The denominator equals e!2+ 2se!kq2+ 2sk
q. Shifting the angle of integration the angle of the vector q, and denoting the angle of integration hereafter as simply , we havek
q!kqcos. Using the evenness of the denominator in , the numerator may be simplied to arrive at ij +(!;q) = 2k2cosij+ 2k2cos 2se!kq2+ 2skqcos(k) (A36) where ^qi=qi=q. To proceed, we make use of the =Z2 01 a+ib0 + (A37) and the cosd 2= 1aI(a) (A38) Z2 0cos 2 a+ cosd 2=2a+ (2a21)I(a) (A39) Evaluating (A36), we dene denote ~I() = 2kqI(). Then ij +(!;q) = (1 +2k q+f2k2+ (A40) Making the substitution 2 p= 2k+se!, we get ij +(!;q) (A41) We arrive at ij +(!;q) =  +  + (A44)8 As stated earlier, the latter function is related to the density response through the Ward identity. Direct comparison with the appendix of [ A1] shows that the Ward identity is satised by the above expression. One also sees that the inter and intraband polarization operators separately satisfy Eq. (A21). In summary, one nds ij(!;q) = + + zz(!;q) (A45) k(!;q) =!2 q400(!;q) (A46) as claimed in the previous section. C. Intervalley polarization and performing the trace changes the sign of the terms containing factors of z. If we denote this intervalley polarization operator ij IIthen one nds the simple result ij II;?= ij ?(A48) In other words, the transverse and longitudinal response are reversed between intra and intervalley. The same trace relations show that 22= 55= 66= zz, ie the screening of the operators is the same as that of zz. D. Solution to the RPA equations in the onsite limit We have the +g5(x x+y y) (+ + +) (A49) We consider the simple case of only onsite interactions, so only UAAAA6= 0; in this limit one nds g3=g4= 0 for honeycomb and g2=g5= 0 for kagome lattices. In each case the RPA equations decouple, so that each giis The result is the screened (A50) for honeycomb (A51) for kagome.9 E. Cooper channel scattering amplitudes 1. On-site limit for honeycomb systems For honeycomb systems, the `-wave scattering amplitude ) + On-site limit for kagome systems For kagome we ) (kp)2(1 + + (A53) F. Eliashberg self-energy corrections, the linearized frequency dependent gap equation is given
`(i!n;k) (A54) with the Eliashberg ) 1)Tare the fermionic Matsubara frequencies [ A8{A10]. The frequency dependent coupling (!;q) is dened analogously to the static coupling in the main text: (!;q) = (!;q) for intravalley pairing or (!;q) = for intervalley pairing. The functions
, Kand are tensors in spin and valley space { spin and valley indices are left implicit. The critical temperature can found by nding the value of Tfor which the above linear equation for
has an eigenvalue of 1=T. The zero temperature gap can also be found from solving
`(!0) is the`-wave coupling evaluated onshell, k=!;p=!0. These equations can be derived through general assumptions about the analytic behavior of the scattering amplitude (see e.g. [ A8]). The requirements of10 FIG. A3: The p-wave intravalley coupling function ands-wave intervalley coupling function `=0 inter(k;p) for honeycomb (left column) and kagome (right column) systems in the onsite limit. We use ne structure constant 2 0e2=r= 0:3, bare couplings g2= 0:6;g3= 1:4;g4= 0:8;g5= 1:4, and for the intrarvalley coupling functions use g1= 0:4 and for the intervalley coupling functions g1= 0:2. The scattering amplitudes have been regularized and smoothed to avoid are inessential to the solution of the gap equation, as discussed in text. unitarity and causality imply that as a function of frequency the amplitude must be analytic in the upper half plane, and that the amplitude has branch cuts corresponding to the threshold of particle-hole production. While as a function of !0, these occur either for !0vFjkpj, which provides a negligible contribution, or close to the edge of the particle hole continuum !0vFjkpj, which is signicant for a range of scattering angles kp0 which only becomes signicant when pkF. As is shown by examination of the formulae in the previous subsection, the scattering amplitudes have a behavior. Below, we plot the p-wave inter and intra valley on-shell couplings { `=1 intra = well as`=1 inter = { for kagome and honeycomb to illustrate this phenomenon. Treating the scattering amplitude as a step function in frequency space, the frequency dependent gap equation can therefore be solved as in the Anderson-Morel treatment of the electron-phonon problem [A11], the result being a renormalized coupling appearing in the exponential form of Tcand an eective frequency cuto, c.f. the treatment in S3 of [A1].11 A4. REAL SPACE FORM OF THE p+ipGAP FUNCTION In this appendix we derive the real space form of the p+ipgap function used for exact diagonalization in Section IV. Starting with the momentum space mean eld Hamiltonian, we h.c. an electron in the upper band,
kis an overall factor depending only on the magnitude k,is the phase of the pair density wave, and ( dx;dy;dz) is a constant 3D vector with unit length. The normal dispersioon kis spin independent, which allows us to perform a spin rotation so that ( dx;dy;dz) =i(0;1;0), and the into two spin-diagonal h.c. (A58) Since the Hamiltonian is a sum of independent spin blocks, we discard the spin index henceforth. We now convert this expression to real space, in which the Hamiltonian h.c. the full electron creation operator (as compared to fy kwhich creates an electron with momentum kin the upper band). We assume that the only hopping term present connects nearest neighbors, i.e. that H(r;r0) =tifr andr0are nearest neighbors and zero else. It remains then to calculate the function
( r;r0). In momentum space, the pairing term for each spin block is given relate the creation operator for the upper band to the full electron creation operator is the wavefunction for the upper Dirac band. Substituting this relation into the pairing (A62) from which it follows immediately that
(r;r0) (A63) The upper band eigenstates of the Dirac Hamiltonian H=vk
are given in the coordinate representation in terms of the wavefunctions eiK1
r'(r) by uk;+(r) =1p 2 '++(r) =u k;+(r) are periodic functions under lattice translations. Thus
(r;r0) the summation over kand introducing the   1(r) (A66) we + + maximum for r= 0 and goes to zero over length scales small for smallrand increases to a maximum at r=kF. To simplify
( r;r0) for the purposes of exact diagonalization, we will neglect the pairing correlations at large separations, discarding the term containing  1and set  0!
0with
0 being some average value. With these simplications we nally arrive at the real space h.c.;
(r;r0) = (A68) Note that
( r;r0) =
(r0;r) =
(r;r0), and with treal, the Hamiltonian is invariant under complex with our claim in the main text that this superconducting phase lies in class BDI. Below we present
( r;r0) explicitly for the honeycomb and Kagome lattices. The procedure is simply to calculate the eigenfunctions '(r) of the normal state Hamiltonians at the Dirac points, and insert into Eq. (A68). A. Honeycomb lattice For the honeycomb lattice, we have '++(r) therefore get
(r;r0) = sin(K
(r+r0) +) (A70) and
0h sin(K
(r+r0) +)cy rcy r0+ H.c.i (A71) as per Eq. (19). B. Kagome lattice We have '(r)

2108.10897

Tommy Li

Tommy Li, Max Geier, Julian Ingham, Harley D. Scammell

Higher-order topological from repulsive interactions in kagome and honeycomb systems

2D Mater. 9 015031 (2022)

cond-mat.supr-con cond-mat.mes-hall

http://creativecommons.org/licenses/by/4.0/

We discuss a pairing mechanism in interacting two-dimensional that intrinsically leads to a second order topological with a spatially modulated gap. When the chemical potential is close to Dirac points, oppositely moving electrons on the Fermi surface undergo an interference phenomenon in which the Berry phase converts a interaction into an effective attraction. The topology of phase manifests as gapped edge modes in the and Majorana Kramers pairs at the corners. We present which constrain the possible form of the in these systems and classify the possible superconducting phases which result. Exact diagonalization of the Bogoliubov-de Gennes the existence of gapped edge states and Majorana corner states, which strongly depend on the spatial structure of the gap. Possible applications superconducting kagome metals AV$_3$Sb$_3$ (A=K,Rb,Cs) are discussed.

2 Li, Fu & Dorfman matic or broadband waves (Malara & Velli 1996; Shoda et al. 2018), coherent or incoherent waves (Umeki & Terasawa 1992), with either parallel or oblique wave propagation (Verscharen et al. 2012; Gao et al. 2013a), single or multiple ion species (Gao et al. 2013b), and in quiescent or turbulent background (Shi et al. 2017; Fu et al. 2018). In the context of the Earth's foreshock and the solar wind, studies also show observational ev- idence of PDI (e.g., Spangler et al. 1997; Narita et al. 2007; Dorfman et al. 2017; Bowen et al. 2018), which has been simulated to illustrate, for example, the gen- eration of density uctuations (Tanaka et al. 2007), the origin of low-frequency Alfv enic spectrum (R eville et al. 2018), the eect of wind (Shoda et al. 2018), and the evolution of switchbacks (Tenerani et al. 2020). Despite these extensive studies, most of the and simulations have essentially considered a periodic innite system. This is achieved using the pe- riodic boundary condition where the waves with a localized plasma (e.g., Del Zanna et al. 2001; Nariyuki et al. 2008; Gao et al. 2013a; Shi et al. 2017; Tenerani et al. 2020; Gonz alez et al. 2020). While this setup is useful for addressing the temporal dynamics of the instability, e.g., conrming the temporal growth rates (Shi et al. 2017), actual AWs observed in space and laboratory are usually nite in time and space, and the associated interactions are nonperiodic (Gigliotti et al. 2009). These discrepancies may complicate the interpre- tation of laboratory or spacecraft observation of AW dy- namics; for example, a instability was recently observed in a laboratory plasma, while the PDI was mysteriously missing despite its signicant growth rates in the low-beta environments (Dorfman & Carter 2016). Pruneti & Velli (1997) rst lifted the constraint by considering wave injection into an open system using 1D/2D MHD simulations. Del Zanna et al. (2001) further extended it to 3D. However, these simulations involved only a few Alfv en the wave direction, thus were limited in dynamics pertinent to the nonperiodic inter- actions. Notice that while the open boundary condition has been used in some solar-wind simulations to account for the global wind structure, the MHD framework is adopted to handle the vast space-time scales at the ex- pense of ignoring kinetic eects (e.g., Tanaka et al. 2007; Shoda et al. 2018; R eville et al. 2018). The objective of this paper is to address how a nite- size AW packet decays in a large-scale nonperiodic sys- tem of uniform plasmas. This basic setup has funda- mental interest and can also be realized in with well-controlled plasma and wave con-ditions (e.g., Gigliotti et al. 2009). We consider a wave packet of tens of Alfv en wavelengths in size and examine its nonperiodic interaction with a low-beta plasma over a large time scale. This study is carried out using with an absorption boundary condition for the wave elds. The large simulation domain considered here removes the constraints on the integer wave num- bers as imposed under the periodic system. The also allows us to keep the ion kinetic eects (e.g., wave damping, ion heating) that are absent from MHD simulations. The paper is organized as follows. In Section 2, we describe the simulation model, in particular the imple- mentation of eld masking for the absorption Novel PDI dynamics from the are then detailed in Section 3, as well as its in uences on the wave propagation. In Section 4, we dis- cuss the results in relation to spacecraft and and brie y mention future MODEL The simulations are performed with H3D, a hybrid code that ion particles while treating electrons as a massless neutralizing uid (Karimabadi et al. 2006). Pe- riodic boundary conditions were assumed by the code. In order to realize the absorption boundary condition, we apply a masking technique (Umeda et al. 2001) to the eld (2) whereEis the electric eld, Bmagnetic eld, pressure, and ui;ni;qiare the ion uid velocity, density, and charge, respectively. The the time step and
tis the step size. The mask consists of two layers on both sides of the z boundaries ( zis along the mean magnetic eld), and its proleFmis dened zm 1;  the domain size along z,mthe length of either mask layer, and ra ratio factor controlling the slope of the mask. In H3D time subcycling is used forParametric decay of Alfv enic wave packets 3 advancing the Beld with Equation (2), thus the mask- ing is applied only at the nal subcycle. It is important that the background magnetic eld Bzis not masked. Equation (1) is also left unmasked as the Eeld will be automatically damped with B. It is found that the masking works signicantly better with smaller r(i.e., smaller gradients of Fm) but requires thicker mask lay- ers to fully damp the waves, as was also observed for full Maxwell eld solvers (Umeda et al. 2001). Using r= 0:1;m= 5diwe achieved 99% absorption of the eld energy upon 3dipropagation into the mask lay- ers, where di=c=!piis the ion inertial length and !pithe ion plasma frequency. Notice that particles are still periodic, making the simulation less reliable at high wave amplitudes when strong plasma heating may in- duce fast particle circulation in the simulation domain. We carry out essentially 1D simulations by using one cell for the xdirection, four cells for y, and cells along the zdirection. The results have on yand are therefore averaged over this coordinate. We take zmax= 4400di1100for simu- lations with the eld masking, where 0= 2vA=!0= 40diis the Alfv en wavelength. In all simulations, the cell size is
x=
y=
z= 1diand time step
t= 0:01 1 ci= 4!1 piwhere ciis the ion satisfying !pi= ci= 400. These parameters imply weakly dispersive AWs with !0= ci= 0:157. A forward propagating (along + z) AW packet of eld ~B=B 0= (Bx=B0)(z) u=vA~B=B 0;(3) is initialized inside the domain next to the left mask layer (i.e.,zm), 1. The waveEeld is determined from the eld equations. The wave envelope (z) consists of ramps on sides and a plateau in andlr;lpare the length of the ramp and plateau segment, respectively. The full packet size is Sz=lp+ 2lr. Notice that, despite the left-hand polarization, ion cyclotron resonance is not a concern here since !0= ci1. We have also run the simulation with right-hand polarization which gives verysimilar PDI dynamics. The plasma ions (protons) are sampled by 1000 macroparticles per cell. The initial density noise, = 00:5%, is suitable for triggering the PDI over a reasonable time scale. We have conrmed very similar results using 64 particles per cell where the PDI develops slightly earlier due to larger initial have the same initial temperature as ions and follow the adiabatic equation of state Te=n e1 e = const where e= 5=3. Therefore, the plasma beta is =i+ e= the ion thermal speed. The comparable may imply strong ion Landau damping. As we shall see, this kinetic eect mostly aects the nonlinear stage of PDI (i.e., when the sound wave grows to a high amplitude), leading to substantial ion heating. We have checked that the linear stage of PDI growth is well approximated by the uid picture; the single- uid MHD (Derby Jr 1978; Goldstein 1978) and two- uid Hall MHD (Hollweg 1994) growth rates are also comparable due to the small 1 rst presents an overview of two contrast sim- ulations, highlighting the dierence between the more realistic absorption boundary condition (upper panels a-c) and the usual periodic system (lower panels d-f). In the former case, a square wave packet of full size Sz= 400and small ramps lr= 10is launched at 50z450. The latter periodic case has a domain, zmax= 50, which is fully occupied by a wave of constant amplitude initially. Both cases have the same wave amplitude B=B 0= 0:1 and plasma beta= 0:02. Signicant dierences are found in the space-time evolution of density uctuation (a, d) and wave magnetic eld (b, e). Under the absorption bound- ary condition the wave packet undergoes a strong decay at the early stage ( with the packet head being little distorted. Correspondingly, the density uctuation is largely localized and In contrast, the periodic case shows mainly temporal dynamics where the density uctuation grows or decays almost uniformly across the domain. To measure the strength of PDI, we separate the waves into forward- and backward-going components using the normalized Elsasser variables dene the associated wave energies and cross helic- ity where the summa- tion integrates over space and, in the absorption bound- ary case, includes also the absorbed waves in order to ensure energy conservation. By comparing between Fig- ure 1(c) and (f), it is seen that about 60% of the pump energy (E+) is decayed during the early stage in the ab-4 Li, Fu & Dorfman Figure 1. Space-time evolution of (a, d) density uctuation = 0and (b, e) transverse x-component of the wave (c, f) Time evolution of forward wave energy E+(black solid), backward wave energy E(red dashed), and dotted) where the wave energies are normalized to the initial value of E+. The upper panels (a-c) are obtained for the absorption boundary condition and the bottom panels (d-f) for a periodic system. Detailed parameters are provided in the main text and notice the dierent axis ranges of the space-time plots. sorption boundary case, while virtually the entire E+is depleted for the periodic system. This dramatic dif- ference is also revealed by the evolution of the cross helicity which ips exactly in the latter case, indicat- ing the dominance of backward waves as also seen in Figure 1(e) during 15 T0< t < 30T0. Despite this dif- ference, about 80% of the pump energy lost is trans- ferred to the backward wave in both cases. To inter- pret that, we cast the frequency matching condition for the PDI in the quantum form ~!0=~!+~!, the energy apportioning between the sound wave (!) and backward lower-sideband AW ( !). For = 0:02 in the weakly nonlinear regime, the mode of the maximum growth rate corresponds to !=! 0= 25% and !=!0= 75% (Derby Jr 1978). The consistency in en- ergy apportioning thus suggests that both cases contain the same PDI nature despite of very dierent dynam- ics. The two cases also share similar growth rates of the density wave (not shown), but their maximum den- sity uctuation amplitudes dier as the two start with dierent pump energies. The absorption boundary case quickly evolves towards a steady state ( = 0:1 in this case) after the initialstrong decay. However, the periodic system could un- dergo several ips of until the wave energies in both directions exhaust. This latter phenomenon is appar- ently due to the repeated wave interaction with the lo- cal plasma forced by the periodic boundary condition. Thus, investigations using a periodic system will signi- cantly overestimate the energy conversion from pump to daughter waves in the PDI, and are therefore of limited use in unraveling the PDI dynamics of nite-size AWs in more realistic nonperiodic systems. Let us now focus on the absorption boundary case and identify how the instability emerges, cascades, and saturates. Figure 2(a) shows progressively the wave sep- aration and associated density uctuation. Initially, the wave is dominated by the square pump sitting close to the left boundary. The maximum growth rate for the present case is theoretically predicted to be max 0:1!0. As a result, the PDI grows exponentially fol- lowing exp( maxt) in the linear stage. At t= 20T0, it is seen that the instability has already strongly devel- oped: the back portion of the forward wave is much de- pleted, a signicant backward wave appears, and a large- amplitude density uctuation excited. Despite that, theParametric decay of Alfv enic wave packets 5 Figure 2. Snapshots of (a) the forward wave envelope jz+j(black), backward wave envelope jzj(red), the initial wave to the current time step at the Alfv en speed, dashed line), density uctuation = 0(blue), (b) parallel electric eld Ez(black) and parallel temperature Tk(orange), and (c) longitudinal ion velocity distribution (red) at t= to bottom) for the absorption boundary case presented in Figure 1. The magenta curves in (a) are smoothed indicating any large-scale modulations. The grey curves in (c) represent the initial velocity distribution at t=0. pump front remains largely intact, but is gradually tai- lored to have a low-amplitude precursor and a smooth down-ramp prole towards the back. We term this en- tire front wave portion as the residual packet, because it stays rather stable against further propagation as shown att= 40;60T0. The formation of this steady state may be qualitatively understood as follows. In the frame comoving with the AW packet, its front portion is al- ways fed with unperturbed plasmas. During the transit of the plasma through the wave, the density perturba- tion grows in a convective manner, becoming stronger at later times or towards the back of the wave. The develops to sucient levels once it reaches the position corresponding to the back end of the resid- ual packet. This leads to the decay of the back portion of the pump wave, leaving only the residual packet intact. Notice that the residual packet propagates at a speed smaller than the nominal Alfv en speed vA=B0=p00 as seen by the delay relative to the dashed envelopes. The delay is found to be partly attributed to the nite- frequency eect, i.e., we obtain the group velocity vg= d!=dkk= (!=kk)=(1 the dis- persion relation parallel wave vector along the background magnetic eld. The wave bandwidth of the nite packet is among the other eects that would cause the additional delay.Another remarkable feature is the emergence of a wave spike, trailing behind the pump (i.e., outside the as- sumed unperturbed envelope outlined by the dashed line). It has a large amplitude, twice that of the pump wave. This is the secondary decay, or inverse cascade as termed in some studies with the periodic system (e.g., Del Zanna et al. 2001). It was generated by a sec- ondary PDI process where the backward AW serves as the pump. This secondary decay mostly occurs when the initial decay was strongly developed around t= 20T0. While secondary decays are also found in the periodic system, they never form such an isolated shape, involv- ing both a large amplitude and a narrow it may imply new potentials for secondary decays in real nonperiodic environments. The detailed mech- anism of their formation is worthy of a separate the nal steady state of the pump is reached, the PDI does not have enough time to develop as evidenced by the negligible density uctuations ( = 0<5%) ex- cited within the modulated wave envelope. On the other hand, the large-amplitude density waves produced dur- ing the initial decay remain localized because of the small wave speed ( 0:13vA) under the low-beta condi- tion; see also the slope of the wave trace in Figure 1(a). However, at later times, the density uctuations grad-6 Li, Fu & Dorfman ually evolve into large-scale cavitation as indicated by the magenta curves in Figure 2(a). This is caused by localized ion heating. To see that, Figure 2(b) presents how the parallel electric eld (black curves) develops with the compressible sound wave. Eventually the par- allel wave eld vanishes due to Landau damping of the sound wave. As a consequence, the ions are heated up to a parallel temperature Tk(orange) several times of the initial value T0 k, whereas the perpendicular low throughout. The locally heated ions spread out longitudinally due to thermal expansion, causing the density cavitation. This long-term dynamics is also a re ection of the pressure balance, P/nT, where a region tends to have lower densi- ties. The above heating eect is also evident from the velocity distribution as shown in Fig. 2(c). While the distribution approximately maintains the Maxwellian- type prole, attening of the distribution occurs be- tweenvz= [1;2:2]vi. This is consistent with the sound wave speed 5=3. The spatio-temporal (instead of just temporal) dy- namics, along with the reduced energy transfer and lo- calized density and ion our central new insights into the PDI process under the more realistic nonperiodic conditions. The spatial separation of Alfv enic uctuations from and our new estimates of the eciency of energy conversion and new insights on ion heating may have important implications for laboratory or space- craft observations. For these applications, it may be imperative to understand better the Alfv enic propaga- tion, especially how the nal steady state depends on the plasma and wave conditions. As a rst step, we investigate the eects of the wave envelope prole and packet size on the Alfv enic propaga- tion. Figure 3 displays the nal steady states of several cases with dierent wave proles or packet sizes (e.g., il- lustrated by the dashed lines). All of them are obtained under= 0:02 andB=B 0= 0:1. In the case of sine- square waves (i.e., no wave plateau region or lp= 0) the wave amplitude refers to the peak value. Case (a) is the baseline that we have discussed above, and it has an eective packet size (measured at full width at half max- imum, FWHM) of Se z= 390initially. Compared with that, case (b) is only diered by a smaller packet size (Se z= 190). We see virtually the same residual packet size for the two cases, Se r'120. This illustrates the irrelevance of initial packet size on the steady state so along as the initial packet is larger than its nal resid- ual packet, because the latter corresponds to only the front portion of the wave. However, larger initial packet Figure 3. Snapshots of jz+j(black), jzj(red),= 0 (blue), and the initial wave envelope (translated to the cur- rent time step at the Alfv en speed, dashed line) for dierent cases (from top to bottom) at t= 60T0, whereSzis the full packet size and Se zthe eective packet size measured at FWHM. sizes means more energies to be converted due to the PDI. This is evidenced by the stronger density uctua- tions and more intense cascade spike found in case (a) as compared with (b). For an initial packet smaller than the nal state size, no PDI should develop and the wave is stable throughout because of the short transit time. This is conrmed by case (c) where Se z= 7:30<120. To clarify that the result is not due to the change of the initial wave envelope (i.e., from square to sine-square), we double the packet size [case(d)] and nd that the sine-square wave becomes unstable again because its ef- fective sizeSe z= 14:60>120. These additional com- parisons show the importance of packet size (or transit time), rather than the envelope, on the PDI excitation. The envelope may still play a role through the pondero- motive force, but its eect is generally small here due to the large packet size (or small envelope gradients) used. Finally, Figure 3(e) presents a periodic case having a wave size comparable to the simulation domain of the absorption boundary cases. Again, the instability de- velops uniformly across the domain, and no nal steady state or enhanced wave cascades can be decay of Alfv enic wave packets 7 Figure 4. (a-c) Snapshots of jz+j,jzj, and h= 0i(smoothed over 1 0alongzaxis) for cases of increasing B=B 0(from top to bottom) and of increasing (from left to right), shown at t= 60T0. The cases in the second row correspond to the same maximum growth rate (marked as dashed lines in d-f) despite their dierent B=B 0and. (d) Maximum growth rates versus B=B 0, solved from single- uid MHD theory under dierent plasma beta . (e) The wave number kand (f) frequency !of the compressible ion acoustic wave corresponding to the maximum growth rates versus B=B 0. The shaded areas indicate the range within which the growth rates are the above insights, we next focus on the simple square wave prole and extend the investigation to a broad range of wave amplitude and plasma beta. The simulation results are summarized in Figure 4(a-c) in terms of the nal states and their corresponding . Furthermore, following the single- uid the- ory (Derby Jr 1978; Goldstein 1978) which is justied here due to the small pump dispersion, the parameters B=B 0andadequately dene the dispersion relation of the PDI, !=!(k). The latter can be solved to nd the growth rate, central frequency, and unstable band- width of the instability measured by the sound wave. These calculations, central to understanding the PDI and the resulting Alfv enic propagation, are presented in Figure 4(d-f) in terms of the maximum growth rate max and the bandwidth corresponding to positive . Of the most importance is the growth rate which denes the timescale relevant for the PDI development. In general, more signicant decays happen with increasing growth rates, corresponding to larger B=B 0and Figure 4(d). This is veried by the overall pattern of the simulation results orderly arranged in Figure 4(a-c). Closer to the top-right having smaller B=B 0and higher , the wave is less prone to PDI. The trend reverses to- wards to bottom-left (having larger B=B 0and smaller) where strong decays are found even at the early stage of the its importance, the growth rate does not exclusively dene the PDI-driven Alfv enic evolution. As seen in Figure 4(e-f), the central !andkof the sound wave (solid curves) corresponding to the maxi- mum growth rates also show strong dependence with . For weakly nonlinear regimes (i.e., not too large B=B 0), one has !=! 0<1 andk=k02=(1 +p), where!=! 0increases and k=k0decreases with , respec- tively. Therefore, following the energy apportioning more energy goes into the backward AWs at smaller for a given pump depletion (i.e., xedNunits of energy quanta). In addition, one cascading into smaller numbers upon each secondary decay. The more limited k-space at higher(which corresponds to smaller initial jk=k0j) may make secondary decays more dicult. This phe- nomenon was identied in periodic systems (Del Zanna et al. 2001) where the minimum wavenumber is limited by the box size. It may also hold for the present nonpe- riodic condition when the packet size is large enough to allow for multiple decays. From the above anal- ysis, both the frequency and wavenumber matching8 Li, Fu & Dorfman Figure 5. (a) The eective residual packet size, Se r, versus the maximum growth rate of the PDI under dierent . All simulation results (point markers) are obtained with the square wave of an initial size Sz= 600. The wave amplitudes B=B 0 are chosen to have the maximum growth rate fall between 0 :06!0and 0:16!0. The solid curves represent tting by Equation (4), where the coecients C() are presented in (b) as a function of (vertical lines indicate errorbars). The solid curve in (b) is a tting ofC() using the 0:36dependence. (c) All simulation data presented in (a) are rearranged here using the combined axis =(B=B 0)2. The solid line represents Equation (5). The shaded areas in (b, c) come from the uncertainties in tting C(). conditions suggest that the secondary wave decays fa- vor smaller where more energy goes into the back- ward wave having larger k=k0. This is supported by the simulation results where the secondary decays are stronger from column (c) to column (a) with decreasing ; for example, notice the prominent spike observed for = 0:02;B=B 0= 0:1. We have checked that the wave power in the low-frequency range (corresponding to the secondary decays) indeed increases with decreasing even in cases without a clearly visible cascade spike. A nal contributing factor to the Alfv enic dynamics is the unstable mode bandwidth, i.e., the shaded areas in Figure 4(e-f). The theory predicts rapidly increas- ing bandwidths with B=B 0, becoming even larger than the central as B=B 0!1. The presence of multiple modes may cause strong compe- tition among each other and make the wave envelopes more spiky. These features can be seen by comparing top to bottom panels in Figure 4(a-c). The above analyses may point out the criteria for strong secondary decays with good wave coherency (e.g., narrow bandwidths and smooth envelopes). These phe- nomena would favor low-beta plasmas where the is small and the wavenumber is large, such that more energy goes into the backward waves with suciently large wavenumbers. Besides, the wave am- plitude should not be too large, involving much mode competition and making the interactions less coherent. It should neither be too small requiring a large timescale for the PDI to we have assumed in the above analyses a cer- tain pump depletion, which corresponds to the total en-ergy lost to both the backward wave and sound wave. It can be dened as D= The quan- tity the portion of energy that can be stably transported over large distances and should be useful for estimating the energy transport e- ciency and the limit of PDI-related plasma heating. As identied above, several factors may contribute to the Alfv enic dynamics, making analytical treatment of Dor Se ra challenge. Here, we resort to simulations for its dependencies on the wave amplitude and plasma beta. We scan over a range of parameters where the maxi- mum growth rate falls within 0 :06< growth rates typically involve large B=B 0and broad mode competition, making the nal state less Even lower growth rates may require a long time for PDI to develop suciently, a challenge to the simulations. Our results are summarized as point markers in Figure 5(a). Despite the complexity of a simple inverse-square dependence of the residual packet size versus the maximum growth rate is found (4) where the coecient C() has a power-law dependence on, as displayed in Figure 5(b). To more vividly illustrate the -dependence, the second row of Figure 4(a-c) shows three cases of dierent B=B 0 anddesigned to have the same growth rate [marked as the dashed lines in Figure 4(d-f)]. Despite the same growth rate, the nal evolved states of the three cases are quite dierent, with signicantly more pump depletion at decay of Alfv enic wave packets 9 In the low-amplitude low-beta limit, one may approx- imate the growth rate as improved tting from that suggested by Sagdeev & Galeev (1969). By substituting it into Equation (4), we arrive at Se r 0= 11:5C0:64 (B=B 0)2'5:8 (B=B 0)2; (5) which explicitly associates the residual packet size with the plasma beta and wave amplitude. Notice a dier- ence ofp 2 in wave amplitude when applying Sagdeev & Galeev's scaling who studied linearly polarized AWs. We have veried Equation (5) by arranging all simula- tion data in the combined axis =(B=B 0)2as displayed in Figure 5(c). It is seen that they follow the linear scal- ing reasonably well, especially when the uncertainty in the coecient of Equation (5) is considered (the shaded area). The >0:1 cases (i.e., the open points) are o, conrming the applicability of Equation (5) to only the low-beta limit. In addition, high-amplitude cases typ- ically involve broad mode bandwidths and spiky mod- ulations of the wave envelope, making identication of the residual state more this study, we have relaxed the usual condition and investigated the parametric de- cay dynamics of nite-size AWs in a large-scale low- beta plasma under the more realistic nonperiodic bound- ary condition. By simulating the dynamics over large space and time scales, our 1D hybrid simulations unravel novel dynamics that were not appreciated in the periodic system before, including strongly space-time reduced pump energy transfer, and and ion heating. These results may have important implications for spacecraft and laboratory robustness and importance of the PDI have stim- ulated the search of their existence in space over the past decades, for example, in the wave eld upstream of the Earth's bow shock (Spangler et al. 1988; Narita et al. 2007). However, observational signa- ture in the magnetic spectrum has been limited due to the dominance of the pump wave and the close separa- tion of the pump and sidebands in frequency space (Tsu- rutani et al. 1987; Agim et al. 1995). The density spec- tral feature, on the other hand, may provide signature of PDI because it is clearly sep- arated from the magnetic signature in both wavenum- ber and frequency in low-beta plasmas (Spangler et al. 1997). As our simulations have shown, however, simul- taneous observation of both magnetic and density sig- natures may not always be possible because the can be separated in space and time from the Alfv enic uctuation. This may explain the missing de- cay line signature reported in Spangler et al. (1997) and help improve the search by establishing proper Furthermore, the conjunction observation from multiple spacecraft in the heliosphere (Velli et al. 2020) may identify missing signatures for PDI by and density signatures at dierent locations along the same Parker spiral. In addition, our simulations have shown that the front portion of the wave packet may evolve into a steady state under suitable wave and plasma parameters. This is be- cause the instability has not enough time to grow in the wave packet. It also indicates ecient energy trans- port of a short wave train over long distances even in an environment where the PDI has a considerable growth rate. By combining simplied theories and numerical scans, we have identied well-behaved scaling dependen- cies of the residual packet size with both the growth rate, wave amplitude, and plasma beta. These scaling laws are valid for relatively low-beta low-amplitude regimes and should be useful for further analytical treatment on the energy transport. The AW propagation under the absorption boundary condition in higher-beta regimes, which may apply to certain regions of the heliosphere (Gary 2001), is beyond the scope of the present work. Finally, we have elucidated the major factors and cor- responding parameter space that would in uence the strength of decays and cascades. These insights should stimulate new interest in the basic PDI dynamics, es- pecially in their highly nonlinear regime. For exam- ple, the PDI may modify the spectral evolution of in- verse cascade (Chandran 2018). However, it should be pointed out that while our study has adopted a basic setup of fundamental importance, it is limited by the 1D geometry. High-dimensional eects such as & Carter 2016), oblique wave propa- gation (Vi~ nas & Goldstein 1991; Verscharen et al. magnetosonic decays, and cascades (Del Zanna et al. 2001; Gonz alez et al. 2020) are among immediate next steps to check on the features identied in this work. We may also consider more realistic wave injection from the bound- ary for a nite period of time (Del Zanna et al. 2001). The latter would have direct correspondence with labo- ratory experiments (Gigliotti et al. 2009) and may help resolve some puzzles related to the experiments (e.g., Dorfman & Carter 2016). This work is supported by NSF/DOE Partnership in Basic Plasma Science and Engineering under the10 Li, Fu & Dorfman grant DE-SC0021237. XF is also supported by the LANL/LDRD program and DOE/OFES. The the Texas Advanced Computing Center (TACC) at The University of Texas at Austin and theNational Energy Research Scientic Computing Cen- ter (NERSC) for providing HPC and visualization re- sources. NERSC is supported by the Oce of Science of the U.S. Department of Energy under Contract Y., Vinas, A., & Goldstein, M. 1995, Journal of Geophysical Research: Space Physics, 100, 17081 Araneda, J. A., Marsch, E., & Vinas, A. F. 2007, Journal of Geophysical Research: Space Physics, 112 Bowen, T. A., Badman, S., Hellinger, P., & Bale, S. D. 2018, The Astrophysical Journal Letters, 854, L33 Chandran, B. D. 2018, Journal of plasma physics, 84 Chen, L., & Zonca, F. 2016, Reviews of Modern Physics, 88, 015008 De Pontieu, B., McIntosh, S., Carlsson, M., et al. 2007, Science, 318, 1574 Del Zanna, L., Velli, M., & Londrillo, P. 2001, Astronomy & Astrophysics, 367, 705 Derby Jr, N. 1978, The Astrophysical Journal, 224, 1013 Dorfman, S., & Carter, T. 2016, Physical review letters, 116, 195002 Dorfman, S., Hietala, H., Astfalk, P., & Angelopoulos, V. 2017, Geophysical Research Letters, 44, 2120 Fu, X., Guo, F., Li, H., & Li, X. 2020, The 890, 161 Fu, X., Li, H., Guo, F., Li, X., & Roytershteyn, V. 2018, The Astrophysical Journal, 855, 139 Gao, X., Lu, Q., Li, X., Shan, L., & Wang, S. 2013a, Physics of Plasmas, 20, 072902 Gao, X., Lu, Q., Tao, X., Hao, Y., & Wang, S. 2013b, Physics of Plasmas, 20, 092106 Gary, G. A. 2001, Solar Physics, 203, 71 Ghosh, S., Vi~ nas, A., & Goldstein, M. 1993, Journal of Geophysical Research: Space Physics, 98, 15561 Gigliotti, A., Gekelman, W., Pribyl, P., et al. 2009, Physics of Plasmas, 16, P., & Sridhar, S. 1995, The 438, 763 Goldstein, M. L. 1978, The Astrophysical Journal, 219, 700 Gonz alez, C., Tenerani, A., Velli, M., & Hellinger, P. 2020, The Astrophysical Journal, 904, 81 Heyvaerts, J., & Priest, E. R. 1983, Astronomy 117, 220 Hollweg, J. V. 1994, Journal of Geophysical Research: Space Physics, 99, V. 1995, Physica Scripta, 1995, 113Karimabadi, H., Vu, H., Krauss-Varban, D., & Omelchenko, Y. 2006, in Numerical Modeling of Space Plasma Flows, Vol. 359, 257 Longtin, M., & Sonnerup, B. O. 1986, Journal of Geophysical Research: Space Physics, 91, 6816 Malara, F., & Velli, M. 1996, Physics of Plasmas, 3, 4427 Matteini, L., Landi, S., Del Zanna, L., Velli, M., & Hellinger, P. 2010, Geophysical research letters, 37 Narita, Y., Glassmeier, K. H., Fr anz, M., Nariyuki, Y., & Hada, T. 2007, Nonlinear Processes in Geophysics, 14, 361, doi: Y., Matsukiyo, S., & Hada, T. 2008, New Journal of Physics, 10, 083004 Pruneti, F., & Velli, M. 1997, in Fifth SOHO Workshop: The Corona and Solar Wind Near Minimum Activity, Vol. 404, 623 R eville, V., Tenerani, A., & Velli, M. 2018, Journal, 866, 38 Sagdeev, R. Z., & Galeev, A. A. 1969, Nonlinear J.-I., & Sonnerup, B. O. 1983, Journal of Geophysical Research: Space Physics, 88, 9069 Shi, M., Li, H., Xiao, C., & Wang, X. 2017, Journal, 842, 63 Shoda, M., Yokoyama, T., & Suzuki, T. K. 2018, Journal, 860, 17 Spangler, S., Fuselier, S., Fey, A., & Anderson, G. 1988, Journal of Geophysical Research: Space Physics, 93, 845 Spangler, S. R., Leckband, J. A., & Cairns, I. H. 1997, Physics of Plasmas, 4, 846 Suzuki, T. K., & Inutsuka, S.-i. 2005, The Letters, 632, L49 Tanaka, S., Ogino, T., & Umeda, T. 2007, Journal of Geophysical Research: Space Physics, 112 Tenerani, A., Velli, M., Matteini, L., et al. 2020, Journal Supplement Series, 246, 32 Terasawa, T., Hoshino, M., Sakai, J.-I., & Hada, T. 1986, Journal of Geophysical Research: Space Physics, 91, 4171 Tsurutani, B. T., Thorne, R. M., Smith, E. J., Gosling, J., & Matsumoto, H. 1987, Journal of Geophysical Research: Space Physics, 92, 11074

2108.10913

Feiyu Li

Feiyu Li, Xiangrong Fu, and Seth Dorfman

Parametric decay of Alfv'enic wave packets in nonperiodic low-beta plasmas

11 pages, 5 figures

physics.plasm-ph astro-ph.SR physics.space-ph

http://creativecommons.org/licenses/by/4.0/

The parametric decay of finite-size Alfv'en waves in nonperiodic low-beta plasmas is investigated using one-dimensional hybrid simulations. Compared with the usual small periodic system, a wave packet in a large system under the absorption boundary condition shows different decay dynamics, including reduced energy transfer and localized density cavitation and ion heating. The wave dynamics are influenced by several factors of the the growth rate, central wave frequency, and unstable bandwidth. A final steady state of the wave packet may be achieved when the instability does not have enough time to develop within the residual packet, and the packet size shows well-defined scaling dependencies on the growth rate, wave amplitude, and plasma beta. Under the proper conditions enhanced secondary decay can also be excited in the form of a narrow, amplified wavepacket. These results may help interpret laboratory and spacecraft observations of Alfv'en waves, and refine our understanding of associated energy transport and ion heating.

2 Jiaxuan Chen, Shenglin Gui 1 Introduction The information in people's daily lives has shown an explosive growth nowadays due to the rapid development of the Internet. For these large-scale and disorderly text datasets, how to dig out the most valuable information has always been a key research topic in the eld of natural language processing [1]. Clustering are sort of unsupervised learning method, which can divide a large amount of unknown text information into a set of intelligible clusters. By nding such from corpus, the search range can be reduced on a large scale and target information can be easily found though ltering. The most widely used text model in the early days is the Vector Space Model (VSM) [4], where each document is represented as a list of weighted terms, based on the document [2]. However, VSM given TF-IDF only can exploit the advantages of text information from the word level, it ignores the potential semantics in the text. Thus result in reducing the value of the information. To address these problems, topic modeling methods are M. Aiello et al.[3] investigates several popular topic modeling Latent Dirichlet Allocation (LDA), mallet LDA, Hierarchical (HDP), etc. LDA model considers the potential topic of the document to cluster the text, however, this technology is still based on the discrete representa- tion of the text and does not consider the context. Thus, it ignores the of the word. Since word location is a kind of key information in the text and the semantic will be very dierent if the location of the word is et al. proposed Word2vec[5] and Doc2vec[15], which are two types of dis- tributed word embeddings (DWE). DWE takes the context of words into account and learns the semantic and grammatical information. However, since word and vector are one-to-one relationship in DWE model, it is dicult to optimize the algorithm itself in the face of some text with unsatisfactory processing eect. In recent years, more optimization text feature selection algorithms has been proposed to improve clustering performance. Hujino et al. [6] utilized Jensen- Shannon divergence to measure documents similarity. Laith et al. [2] text clustering algorithm (MKM) to improve VSM perfor- mance. However, these two methods respectively use just one type of text repre- sents algorithm to select features from documents corpus, both of which de facto loss some critical and useful text features. Hu et al.[1] proposed a hybrid method which integrate VSM and LDA model, this model combine the advantage of these two models by hybrid the weighted similarity matrices between documents cor- pus. However, bunches of experiments need to be conducted to obtain appropri- ate weight parameters, which consumes large amounts of time. Hassani et al. [8] proposed a text feature extract technique based on Nonnegative Matrix Factor- ization and Latent Semantic Analysis (NMF-FR). NMF-FR employ NMF method to reduce dimensionality of term-document matrix which is extracted by Latent Semantic Analysis (LSA) model, then agglomerate term vectors to create a new feature space. This method take the dimensionality of feature vectors into account and improve the performance of LSA text represent algorithm. However, NMF- FR ignore the potential semantics in the text, thus result in text information loss and aect agglomeration performance. Hybrid harmony search algorithm (FSH- STC)[7] and hybrid particle swarm optimization algorithm (H-FSPSOTC) [9], etc. also widely used in text clustering area. FSHSTC based on harmony search algo-Hybrid Multisource Feature Fusion for the Text Clustering 3 rithm, which displace uninformative options with a replacement set of ne options to improve feature selection performance. H-FSPSOTC apply genetic operators to improve particle swarm optimization (PSO) algorithm performance. Neverthe- less, the performance of these algorithms uctuate wildly in dierent The solution this paper proposed, in contrast, keep leading performance in each experiment. In this paper, a Hybrid Multisource Feature Fusion (HMFF) framework is proposed by fusing two kinds of feature extraction algorithms based on and distributed representation. HMFF analyzes document features from multiple perspectives to ensure the stability and accuracy of the algorithm when processing dierent kinds of documents and eectively avoid the failure of the algorithm. Below we summarize the contributions of this paper: 1) We proposed a new text feature fusion framework, HMFF, which integrate with three dierent text representation source to extract more from input samples. The benchmark experiments results show HMFF performs overall better than other competitors. 2) The previous text clustering algorithms usually compared their performance on dataset with known cluster count, however, most of them did not show how to be applied on dataset in the wild with unknown cluster count. It is a huge challenge for text representation algorithms to obtain good clustering eect on a stochastic wild dataset with unknown cluster count. In this paper, we show how to apply HMFF framework on COVID-19 public opinion dataset from Twitter apart from standard datasets. Our HMFF shows a better performance than other algorithms under various cluster number settings. 3) HMFF framework has good scalability, hence it could easily integrate with additional feature sources into HMFF feature matrix from other text represen- tation algorithms to enlarge feature coverage of corpus without overall time. 2 Preliminaries In this section, we brie y show the basic information which will be used in this paper, including text preprocessing technique, vector space model, topic model, word embedding and K-means clustering. 2.1 Text preprocessing is a key work in text clustering, document retrieval or related to natural language processing. The purpose of text prepro- cessing is to transform the original text into a format that the program can under- stand and handle easily. The preprocessing of English documents usually consists of POS tagging, stemming and stop words removal. 1) i) Tagging words with POS tags according to their meanings and context content. Eective POS tagging can help to improve the eciency of tok- enization.4 Jiaxuan Chen, Shenglin Gui ii) Splitting a text document into tokens and deleting empty statements and documents, in which each word or symbol is called a token. 2) Stemming is the process of reducing word by removing the ax to get the most common way to write words, namely, the roots. The Porter stemmer is the common stemming method embraced in content mining [11]. 3) Removing high frequency words, such as he, she, I, etc. Because these words appear in almost every text. They don't bring symbolic features to the docu- ment. In addition, this kind of words will also aect the feature extraction of the document, resulting in the accuracy of the subsequent document is reduced. The list of forbidden words can be found in this website1. 2.2 Vector Space Model In the late 1960s, Salton and others rst proposed the vector space model (VSM). VSM is a common model that easy to implement which describe documents as a series of vectors composed of keywords. In this process, each text is abstracted into a vector. The length of the vector equals to the number of document feature words, and the value is the weight of the feature value. VSM represents documents as follows: V(d) (1) wherewi(d) is the TF-IDF weight of dictionary token iin document d. TF-IDF[12] is a common weighting technology in information retrieving and data mining. The method is based on statistics and is used to calculate the im- portance of words in corpus. Compared with one-hot method, TF-IDF can words from common but unimportant words. Let ni(d) be the count of occurrences of each term in each text, the following equation is used to expressed Term and Inverse Document =TFIDFi(d) the number of all texts in whole dataset, t(d) is the number of all tokens in textdanddfimeans how many text contains the token i. 2.3 Topic model Topic model is a kind of statical model for using the abstract opics to repre- sent documents. Latent Dirichlet Allocation is a classical topic model that was proposed by BLEI in 2002[13]. It is a generative model based on probability. In the topic model, the text is composed of selected words from certain topics, and the specic words in the text can re ect the specic topics. Therefore, in Latent Dirichlet Allocation model, each topic is considered as the probability distribution of keywords, and each document is represented as the probability distribution of several topics in the shows the Graphical model representation of LDA. The gure utilizes boxes to represent replicates, i.e. repeated entities. The outer plate Multisource Feature Fusion for the Text Clustering 5 documents, while the inner plate represents the repeated word positions in a given document, and each position is associated with a choice of topic and word[14]. In the gure, only wis greyed, because it is the only observable variable, and the others are all latent 1 Graphical model representation of LDA Suppose there are Ddocuments, and each of them has Niwords, where i2 f1;:::;Mg. When generating the whole corpus, LDA assumes the following steps: 1) Choose 'kDirichlet (), which is the word distribution of the topic k2 f1;:::;Kg. 2) For each document (), which is the topic distribution of the document. ii) For each word (i), which is the topic of word wij. b) Choose wijMultinomial ('zij) There are two major approaches to learn an LDA model, which are and Gibbs sampling. The learning process of LDA ends up with a set of statistical parameters, including ;; and', that maximize the log likelihood of the corpus. In other words, the learning algorithm maximizes the ;;) Distributed word Word Embedding[15] is an approach of representing words in a corpus by mapping them into an n-dimensional vector space, in which words with similar meaning are expected to lie close to each other. Word2vec and Doc2Vec are two typical distributed word embedding algorithms, which are suitable for data in dierent formats. 2.4.1 Word2vec is a word vectorization method based on two-layer neural network. Given enough training data, word2vec can t all the words into an n-dimensional space6 Jiaxuan Chen, Shenglin these leather jacketsfine fine fine fine I am selling these fine leather jackets (a) Word2vec ( Skip-gram ) (b) Doc2vec ( PV-DBOW ) Fig. 2 two typical word embedding There are two major strategies in doing so, continuous bag of words (CBOW) and Skip-gram. Skip-gram strategy predicts a word's meaning using a random close-by word, while CBOW takes into account of the whole context of the word dened by a window size[16]. In practice, the Skip-gram model has proved to have better overall Doc2vec paragraph vector method[17] is an unsupervised algorithm, which can learn xed length feature representation from variable length text (such as sen- tences, paragraphs and documents). The algorithm trains the predicted words in the document to use a single dense vector to represent each document. It is in- spired by the word vector learning, paragraph vector can predict the next word according to the given context samples from the paragraph. The original paper mentioned two methods: distributed memory model of paragraph vectors, PV-DM and distributed bag of words version of paragraph vectors, compared with bag of words model, Doc2vec has some unique advan- tages. Firstly, the feature vectors can inherit the semantic information of words. Secondly, word order is considered like the N-gram model on the basis of small context, which retains more information about paragraphs, including word order. For the data set in word bag format, due to the context information of words can- not be obtained, Word2vec is used as a part of feature extraction. Thus, for the data sets that can express word order, Doc2vec technology incline to have K-means text clustering algorithm[19] is a basic partition method in clustering analysis. Firstly, the samples are roughly clustered, and then the clustering results are according to some correction principle until the clustering results areHybrid Multisource Feature Fusion for the Text Clustering 7 reasonable. Each data points inside the dataset is a vector of document algorithm mainly consists of the following four steps: 1. be clustered, ktexts are randomly selected as the initial clustering center For each sample xiin the data set, calculate its distance to the each cluster center and assign it to the cluster corresponding to the cluster center with the shortest distance. 3. For each category aj, recalculate its cluster centers Repeat step2 and step3 until the specied number of iterations Stepmax is reached or the termination condition jEn+1Enjis satised,where the standard measure function xiis the text in the center ofCi. 3 Proposed method In this section, we present the details about proposed text feature fusion frame- work, Hybrid Multisource Feature Fusion (HMFF) framework, which consists of three main parts: feature representation of multi-model, mutual similarity ma- trices construction, dimension reduction and features fusion. Notice that in this section, the whole document dataset is already assumed to be preprocessed and be tokenized into token-document dataset Dtoken . As HMFF framework shown in Fig.3, we construct mutual similarity matrices for each feature source and fuse discriminative features from mutual by reducing dimensionality to generate HMFF features DHMFF (nh), , wherenis the number of documents and his the dimension of hybrid features. In order to detect the performance of HMFF framework, k-means clustering al- gorithm is congured to partition DHMFF into groups. Accuracy, F-measure and Silhouette Coecient are used as clustering judgements to measure the with other similar 3 owchart of applying the proposed HMFF framework. The left dotted rectangle shows main text preprocessing steps that convert raw datasets to Dtoken . The main parts of HMFF framework are presented at middle rectanglen, which generate HMFF feature-document matrix DHMFF from token-document dataset Dtoken . And comparative evaluation indexes are shown in the right dotted rectangle to measure HMFF framework performance.8 Jiaxuan Chen, Shenglin Gui 3.1 Features representation of multi-model In this selection we will introduce the mathematical representation of under VSM mvsm, topic model mtopic and distributed word embedding memb. By using text representation method mi, we convert the dataset from a token-document format Dtoken to a feature-document as = (dn1;mi)T;Vec (dn;mi)T] (4) wheremi2 number of document in the the vector that represent document djunder model mi. 3.1.1 Features from Vector Space Model Vector Space Model (VSM) use TF-IDF weight as features to represent docu- ments, and all documents using one dictionary which include every tokens from the dataset. In order to ensure the consistency of the document feature vector length, we assign the weight of the tokens in the dictionary that does not appear in document dito zero. Hence, dican be represented in an Nvsm-dimensional vec- tor space, where Nvsmis total number of unique tokens from the whole dataset. Letnbe the number of documents in a collection, document diunder VSM mvsm is represented as = (tNvsm;di)) represents the weight of token tj in document di, it can be measured under TF-IDF encoding or one hot encoding. 3.1.2 Features from topic model Documents can be represented as probability distribution of each topic under topic model, the number of topic determines the length of the feature vector. In this paper we use topic coherence measurement to choose optimal topic score[34] was proposed by Michael in 2015, this evaluation model is more reasonable than the traditional log perplexity measurement of topic model. We useCUMass [34] as coherence score, which based on document to evaluate whether one topic model is good or not. CUMass two dierent words in the dataset, P(wi;wj) is the proportion of documents that contain both words wiandwjto all documents in the dataset, p(wj) is the proportion of documents that contain words wjto all documents in the dataset, Nis the top words number which are chosen from a topic. In this formula, P(wi;wj) andP(wj) are determined by topic number and other parameters are all xed when a dataset and topic model are given. A higherHybrid Multisource Feature Fusion for the Text Clustering 9 score means that the topic has better interpretability in semantics, and the seman- tics within the topic are more coherent. Then we can use this formula to calculate an optimal topic number Ntopic. Ntopic =argmaxfCUMassg (7) Letnbe the number of documents in a collection, document diunder Topic model mtopic can be represented in the andw(tj;di) represents the of topic tjof document di,nis the number of documents, ttopic is topic num which equals to optimal topic num Ntopic. 3.1.3 Features from distributed word and Doc2vec are two types of distributed word embeddings. The former is suitable for corpus in format of tokens, and if we have word order information of each document, we can use Doc2vec to extract paragraph information directly. i) In Word2vec model, each unique token kin corpus is mapped to an numerical vector as = (tNw2v;k)) is the value of feature tjof the vector of token k, 1jNw2v. Let suppose N(;D) is total number of unique tokens from the whole dataset D, 1kN(;D). Dimension of word vector Nw2vis determined by the number of hidden layer nodes in neural network that trained model mw2v. Then we use one `focus word' to represent document diby averaging the word vectors of all tokens in di. The representation of diunder Word2vec model mw2vis as represents the jth feature in average word vector of is the number of tokens in di. ii) In Doc2vec model, text feature representation is considered from the perspec- tive of paragraphs. Compared with Word2vec, Doc2vec can learn representations from variable-length texts, text representation format under this model is shown as: Vec(di;md2v) = (tNd2v;di)) andw(tj;di) represents the value of featurestjin vector of document di. Finally, we give document under distributed word embedding representation (in- clude Word2vec model and Doc2vec model) a unied format as = Jiaxuan Chen, Shenglin Gui 3.2 Mutual similarity matrices construction In HMFF framework, feature fusion from dierent algorithms is a key step, there- fore we need design a unied document feature representation format for each feature source. We construct mutual similarity matrix which shows documents to represent document for the following reasons: 1) Dimension of similarity matrix is only determined by document number nin the dataset D. Compared with feature-document matrix Matfeat(mi), in which dimension is de- termined by text representation algorithm's type, it makes preparation for dimen- sion reduction and features fusion. 2) We can choose optimal similarity measures to match various text representation model, in this way more features are kept in the matrix. By calculating the mutual similarity between each two documents under spe- cic distance formula in this section, we can construct one mutual similarity matrix for each type of document representation algorithm in the


1i;jn,nis document number in a collection, and Sim(di;dj) is similarity score between dianddj. Note that the value range of Sim(di;dj) is between 0 and 1. 3.2.1 Cosine-Euclidean formula The Cosine similarity measure and Euclidean distance measure are commonly used in text clustering domain to calculate mutual similarity between each feature vector of Let dianddjbe the two documents in the dataset, the similarity is calculated under these two formulas by the following. Cos (di;dj) = is the weight of feature tkin document di,w(tk;dj) is the weight of featuretkin document the summation of all features weight square of 1 toNmodel technique use a kind of multi-objectives method to improve the per- formance of similarity matrix, which combines Cosine Similarity Euclidean Distance measure[eq15], namely, Cosine-Euclidean function. the range of Cos(di;dj) is between 0 and 1. small value means high mutual similarity between twoHybrid Multisource Feature Fusion for the Text Clustering 11 documents. This paper use Min-max feature scaling to normalize Euclidean dis- tance and convert the function to measure mutual similarity between documents but not formula get advantages of both measurement together hence are able to to make mutual similarity algorithm become more precision and robust. The equation is portrayed as =CE(di;dj ) =Cos(di;dj) Jensen-Shannon divergence For Topic model, the documents is represented by topic probability vector of the Dirichlet distribution. Since cosine similarity and Euclidean distance are not suitable to measure similarity between two probability we divergence instead of these two measurements to calculate the document similarity, so as not to lose the advantage of topic model. To mea- sure similarity between document dianddjin topic model, Jenssen-Shannon di- vergence is dened as andw(tk;dj) represents the probability distribution of topic tk of document JS(di;dj) is in range [0 ;1] in the expression. 0 indicates the two distributions are the same, and 1 shows that they are nowhere similar. In consideration of uniform format, we convert it to mutual as = 1JS(di;dj) (19) 3.3 Dimension reduction and features fusion The feature selection issue in HMFF framework can be dened as integrating fea- tures which are extracted from mutual similarity matrices Matsim(mi), In this process, discriminative feature directions re- quire to be conrmed from each mutual similarity matrix. Since mutual is a non-negative real symmetric matrix, all its eigenvectors of are orthogonal to each other. The truncated of documents similarity Matsim(mi) is dened as thenkmatrix whose jth columns is the orthogonal the diagonal matrix whose diagonal elements are the k largest corresponding eigenvalues, j,j= 1;:::;k andkn(documents number in the dataset). This can be much quicker and more economical than calculating12 Jiaxuan Chen, Shenglin Gui all the eigenvalues and eigenvectors in the Matsim(mi) and are able to retain most of information of the normalized keigenvectors vjin theQmi (nk), eigenvalues jbe the weight of j. We extract truncated features mi jmi jfrom each text represen- tation method miand stitch them together, as a complete fused feature matrix to represent the whole dataset. This is called HMFF feature matrix in this paper. And we describe the overall procedure of HMFF framework in Algorithm 1. DHMFF = 1 HMFF framework Input: all tokens of dataset Dtoken Output: HMFF features of dataset DHMFF 1: determine text representation models M;mi2Mbased on format of document Dtoken . 2:formi2Mdo 3: fordi2Dtoken do 4:Vec(di;mi) representdiinto a numerical representation under model miaccord- ing to [Eq5-Eq12] 5: end represent whole dataset according to build similarity matrix according to [Eq13-Eq19] 8: return Sim(di;dj) =CE(di;dj) 9: else return Sim(di;dj) = 1JS(di;dj) wheredi;djare two dierent documents in the dataset. 10: end calculate truncated with klargest eigen- values for 13:DHMFF integrate all truncated features mimifrom dierent model miaccording to [Eq21] 4 Experiment details We have programmed the HMFF technique using Jupyter Notebook (Python 3.8.5 64-bit) to perform the document clustering mechanism. In this section we illus- trate the evaluation measures, then we explain experiment details on At last we utilize HMFF technique to conduct extended experiment about COVID-19 twitter dataset. 4.1 Evaluation measures Since benchmark datasets has been labeled with categories, the comparing assess- ments were conducted utilizing one internal evaluation measure, silhouette coef- cient, and two external evaluation measures, accuracy(Ac) and measures are the common evaluation criteria used in the domain of the text clustering to evaluate the clusters Multisource Feature Fusion for the Text Clustering 13 4.1.1 internal evaluation Coecient is a typical internal evaluation method of clustering eect. It combines the two factors of cohesion and separation, which is used to evaluate the impact of dierent algorithms or dierent operating modes of algorithms on the clustering results on the basis of the same original data[24]. For text record point i2Ci, the denition of Cohesion is: a(i) (22) whered(i;j) is the distance between text points iandjin the cluster Ci. And the Separation of text data point i2Ciis: b(i) = (23) Considering the coecient and separation, a silhouette value of one text data pointiis dened as below: s(i) 1(24) The value of the silhouette coecient of the clustering result is between [-1,1]. The larger the value, the closer the similar samples are. The farther the dierent samples are, the better the clustering eect[25]. 4.1.2 external evaluation is widely used in the elds of statistical classication which and precision(P) in text clustering[26]. The F-measure controls for the clusterrand category sis decided by the =nr;s nr;R(r;s) one of clustering category and one of real class respectively, nr;s is the number of samples both in the real class sand in the cluster r.nris the number of samples of the cluster r, andnsis the number of samples in class s. Then the F value between cluster rand category sis dened as follows. F(r;s) +P(r;s)(26) and letnis the number of all texts, nrmeans the number of smaples in cluster r. F-measure for all clusters is calculated by the (27)14 Jiaxuan Chen, Shenglin Gui The Accuracy (Ac) measurement[26] is one of the common external measure- ments used to compute the percentage of correct assigned documents to each cluster according to the following (28) whereP(r;s) represent the precision value of real category sin clusterr. K is the number of all clusters. 4.2 Benchmark evaluation We use eleven real-word benchmark datasets, and their statistics are summarized in Table 1. The rst six datasets [27] are in bag-of-words format after the The rst dataset(DS1), called the Computer Science Technical contain 299 documents which are composed by abstracts and tech- nique reports. CSTR belong to 4 areas: Articial Intelligence, Robotics, Theory and Systems. SyskillWebert is the second which contain 333 doc- uments and composed by web pages about Goats, Sheep, Biomedical, and Bands. Oh15 is the third dataset (DS3)[29], is part of the OHSUMED collection[30] and belong to ten topics. The forth, fth, sixth datasets (DS4-DS6) are Tr11, Tr12, Tr41[29] respectively, they are all from TREC and the topics correspond to the documents that were judged relevant to particular queries[27]. And the last ve datasets are subsets of which was orig- inally collected by Ken Lang. It contains over 18,000 documents in 20 groups. The groups can be divided into 5 categories, where each category has In order to separate results and reduce computation, we divided the dataset and consider each separately. In the Table, The seventh dataset (DS7), called 20 Newsgroups Computer (20COMP),which contains the group of pc-hardware, mac.hardware, windows.x that are related to com- puter technique. The eighth dataset (DS8), called 20 Newsgroups Politics the group of misc, guns, mideast that are related to politic problems. The ninth dataset (DS9), called 20 Newsgroups Miscellaneous (20MISC), which con- tains the group of autos, motorcycles, baseball, hockey that are related to sports and vehicle. The tenth dataset (DS10), called 20 Newsgroups Religion (20REL), which contains relision.misc, atheism, christian that are related to religion sub- ject. The eleventh dataset(DS11), called 20 Newsgroups Science (20SCI), which contains the groups of crypt, electronics, med and space that are related to sci- ence technique. And these ve datasets are in the format of raw documents, so we utilize standard text preprocessing to clean the datasets and extract tokens as introduction in 2.1. 4.2.1 Parameters setting for the proposed algorithm When we use HMFF technique, parameters need to be set for each document rep- resentation measures in Feature Representation Step for dierent datasets. The evaluation experiments are conducted in order to compare with the baseline and ve similar advanced feature selection algorithms. Since these algorithms are usedHybrid Multisource Feature Fusion for the Text Clustering 15 Table 1 An overview of the datasets used in our Abbreviation Documents Terms Tokens Classes S-Index[27] reference DS1 CSTR 299 1725 24858 4 0.752 [27] DS2 SW 333 4339 58085 4 0.964 [28] DS3 Oh15 913 3100 95521 10 0.857 [29] DS4 Tr11 414 6429 437143 9 0.937 [29] DS5 Tr12 313 5804 311111 8 0.955 [29] DS6 Tr41 878 7454 357606 10 0.933 [29] DS7 20COMP 4582 64139 420789 5 0.732 [8] DS8 20POL 2287 41254 284113 3 0.825 [8] DS9 20MISC 3648 39836 204655 4 0.757 [8] DS10 20REL 2195 40149 246276 3 0.889 [8] DS11 20SCI 3617 54373 316480 4 0.831 [8] Documents only statistics non-empty text number in the dataset. Terms means the number of unique tokens in the dataset. Tokens means the number of all tokens in the dataset. S-index means silhouette coecient[32] of each classes in each parameters which perhaps result in error in the comparative experi- ments. We strive to ensure the parameter selection under technique is unied to ensure the nal results is comparable. In this experiment, All the parameters set- ting for HMFF and comparing algorithms refers to experts papers as and repeated experiments conducted by ourselves. Table 2 Parameter Symbol Value Random Seed  1 Topic num for topic model Ntopic 5185 Iterations for topic model topic 50 Window size for Word2vec model w2v 1 Vector dimension for Word2vec model tw2v 200 Iterations for Word2vec model w2v 100 Window size for Doc2ec model d2v 5 Vector dimension for Doc2Vec model td2v 200 Iterations for Doc2Vec model d2v 100 truncated eigenvectors number for HMFF framework k 3 Max iterations of k-means algorithm kmeans 300 Since neural network are utilized in distributed word embeddings and both topic model and K-means have random initiate step, we x Random Seed = 1 to ensure the repeatability of experimental results. And we set truncated eigenvectors number k= 3 in the HMFF frame- work. Although set dierent numbers may have a better performance in specic dataset, we still x this parameter to maintain model 2 shows all the parameter we set in the experiment. For the topic model, as the optimal number of topics is related to coherence score in the dierent data set, we change topic number from 2 to 500 for topic model in each dataset, and calculate each model's coherence score in Fig. 4, and choose the model which has higher score as the optimal model in Table 3.16 Jiaxuan Chen, Shenglin Gui 0 25 50 75 100 125 150 175 200 topic score (a) DS1LDA model 0 25 50 75 100 125 150 175 200 topic score (b) DS2LDA model 0 25 50 75 100 125 150 175 200 topic score (c) DS3LDA model 0 25 50 75 100 125 150 175 200 topic score (d) DS4LDA model 0 25 50 75 100 125 150 175 200 topic score (e) DS5LDA model 0 25 50 75 100 125 150 175 200 topic score (f) DS6LDA model 0 25 50 75 100 125 150 175 200 topic score (g) DS7LDA model 0 25 50 75 100 125 150 175 200 topic score (h) DS8LDA model 0 25 50 75 100 125 150 175 200 topic score (i) DS9LDA model 0 25 50 75 100 125 150 175 200 topic score (j) DS10LDA model 0 25 50 75 100 125 150 175 200 topic score (k) DS11LDA model Fig. 4 coherence score with topic number for dierent datasets Table 3 Topic number for dierent datasets Dataset DS1 DS2 DS3 DS4 DS5 DS6 DS7 DS8 DS9 DS10 DS11 Ntopic 185 5 25 15 5 10 5 95 5 65 25 4.2.2 Experiment results and discussion Since HMFF framework integrate text representation algorithm from VSM, topic model and distributed Word embeddings (include Word2vec model and Doc2vec model), the rst experiment is to examine optimal measurement of each model before conduct integration to let HMFF become eective. Figure 5 demonstrates the performance of each models with dierent measure- ments on these benchmark datasets. To be specic, we evaluated Vector Space models, Topic models and Word2vec models on the datasets (DS1-DS6), and eval- uate Doc2Vec models on the datasets (DS7-DS11) as these datasets retain word order. The result shows that applying Cosine-Euclidean measure on VSM and two distributed word embeddings will let models become more eective. And Mallet LDA with Jensen-Shannon measurement performance better among topic models. We conduct comparative evaluations on two dierent formats of And the experiments are carried out to compare HMFF with other algorithms in each formats. For accurate results and statistical com- parisons, we carried out the programs more than 20 times. This number as selected based on the literature which it can sucient to validate the proposed are bag-of-words format dataset. We choose Word2vec as word em- bedding model in HMFF word embedding module, and compared this proposed method to K-mean algorithm (MKM), Mallet LDA hybrid PSO algorithm (H-FSPSOTC) and harmony search algo- rithm (FSHSTC). In the second format, all the datasets are retain word order. We choose Doc2vec as word embedding model in HMFF framework. are MKM, MalletLDA, FSHSTC and Nonnegative Matrix text clustering (NMF-FR). This section shows the performance of each algorithm based on clusters quality. We can see the results of proposed Hybrid Multisource Feature Fusion (HMFF)Hybrid Multisource Feature Fusion for the Text Clustering VSM models measure CE topic models performanceLDA (JS measure) Mallet LDA (JS measure) HDP (JS Word2vec models measure CE Doc2vec models measure CE measure Fig. 5 Performance of each models with dierent measurements. Note that the experiment use Distance formula, Similarity formula and Cosine-Euclidean formula (CE measure) to measure VSM [Fig.(a)], Word2vec model [Fig.(c)], Doc2vec model [Fig.(d)] respectively. And compared three topic models of LDA, Mallet LDA[35] and HDP[36]. All topic models used Jensen- Shannon divergence to measure. This experiment use Accuracy (Ac) and F-measure (F) outperformed all types of basic algorithms shows in Fig.5 based on ex- ternal measurement. As expected, the HMFF framework improved the text clus- tering performance when compared with utilizing text represents methods alone. In Table 4, the result shows HMFF have the best performance in DS1, DS2, DS4 and DS5. FSHSTC and MalletLDA win in DS3 and DS5 respectively. And result in Table 5 shows that our proposed method outperforms the other algorithms in DS8, DS9 and DS10, and in the second position in the rest datasets. algorithms may have better performance in specic dataset but have low rank in another dataset, which means these algorithms become less eective or even malfunctioned in these situations. For example, FSHSTC ranks 1st on DS3 and DS10, but falls to forth place on DS1 and DS7. In comparison, HMFF wins in 7 of 11 benchmark datasets and still keep a high ranking on the rest 4 datasets as well, which performs overall better than other competitors. 4.3 Wild dataset clustering Apart from benchmark tests, we also create a wild real environment as an to test HMFF framework performance in practice. In the , the HMFF framework was utilized on a twitter dataset about COVID- 19 discussion, the dataset consists of 32132 documents and 353668 terms after text preprocessing. We set Doc2vec as word embedding model to build HMFF and set topic number Ntopic = 137 for malletLDA, since when Ntopic = 137 topic model has the highest coherence score. Then we compared dierent type of text selection algorithms including HMFF, FSHSTC, NMF-FR, MKM, H-FSPSOTC and MalletLDA. Cluster count change from 4 to 10, which is an explainable range.18 Jiaxuan Chen, Shenglin Gui Table 4 Algorithms performance on bag-of-words format Dataset Measure MKM MalletLDA H-FSPSOTC FSHSTC HMFF [2] [35] [9] [7] DS1 Accuracy 0.6128 0.5483 0.6792 0.6060 0.7640 F-measure 0.6041 0.5265 0.6010 0.5808 0.1368 0.1261 0.0289 || 0.2968 Rank 2 5 3 4 1 DS2 Accuracy 0.7501 0.6684 0.5538 0.7846 0.8560 F-measure 0.7203 0.6848 0.5944 0.7537 0.2501 0.5160 0.0008 || 0.2918 Rank 3 4 5 2 1 DS3 Accuracy 0.5383 0.4774 0.3966 0.5915 0.5746 F-measure 0.4906 0.4564 0.3429 0.5472 0.2199 0.0344 0.0246 || 0.1294 Rank 3 4 5 1 2 DS4 Accuracy 0.5635 0.5274 0.6594 0.6148 0.7721 F-measure 0.4659 0.4294 0.5239 0.5311 0.2813 0.0476 0.0076 || 0.3372 Rank 4 5 2 2 1 DS5 Accuracy 0.5567 0.5873 0.5632 0.6542 0.7153 F-measure 0.5348 0.5152 0.5358 0.6475 0.1956 0.0384 0.1119 || 0.2794 Rank 4 5 3 2 1 DS6 Accuracy 0.5554 0.6947 0.4110 0.5815 0.6437 F-measure 0.4720 0.5594 0.2985 0.4536 0.3333 0.2009 0.0065 || 0.2365 Rank 3 1 5 3 2 Mean rank 3.17 4.00 3.83 2.33 1.33 Final rank 3 5 4 2 1 For each dataset, we use Accuracy, F-measure and Silhouette to evaluate cluster's quality, and shows algorithms' comprehensive performance rank based on above three judgement. As H-FSPSOTC and FSHSTC only use Accuracy and F-measure as judgements, so silhou- ette scores of these two methods in the table are omitted. At last we statistic algorithms' rank in dierent dataset and give Final rank. Although twitter public opinion data is completely stochastic and has no de- nite categories, which brings huge challenge to the text clustering method, Figure 6 shows HMFF features still could obtain almost the highest silhouette score on each cluster count settings. Furthermore, all these six algorithms unanimously show the most appropriate cluster number of this dataset is 7. 5 Conclusions and future work This paper presents a new feature fusion framework which hybrid three dier- ent source of text representation algorithm to extract comprehensive information from documents, namely, Hybrid Multisource Feature Fusion (HMFF) framework. And these HMFF features are used by K-means clustering algorithm to clusters. In fact, our HMFF framework could run with other cluster- ing algorithm other than K-means. Six datasets with bag-of-words format and ve datasets retaining word order are used for performance and comparing evaluations. All these data sets are public benchmarking text datasets, which are used most widely in comparing algorithms. Five competitors come from recently published papers and compared with HMFF. The experimental results show that the per-Hybrid Multisource Feature Fusion for the Text Clustering 19 Table 5 Algorithms performance on token order retained corpus format Dataset Measure MKM MalletLDA NMF-FR FSHSTC HMFF [2] [35] [8] [7] DS7 Accuracy 0.4131 0.3950 0.6370 0.4015 0.5476 F-measure 0.3760 0.3586 0.6114 0.3741 0.0995 0.1170 || || 0.2154 Rank 3 5 1 4 2 DS8 Accuracy 0.7158 0.4987 0.6897 0.6945 0.7638 F-measure 0.5783 0.4190 0.5808 0.5921 0.2084 0.0421 || || 0.3027 Rank 2 5 4 2 1 DS9 Accuracy 0.5960 0.5408 0.8121 0.6314 0.7830 F-measure 0.5518 0.4955 0.7194 0.5819 0.1492 0.1628 || || 0.3038 Rank 4 5 2 3 1 DS10 Accuracy 0.6287 0.5206 0.4860 0.6700 0.6581 F-measure 0.5416 0.4421 0.4045 0.6213 0.2145 0.0285 || || 0.0655 Rank 3 4 5 1 2 DS11 Accuracy 0.5142 0.4759 0.8035 0.5210 0.8293 F-measure 0.4351 0.3845 0.7137 0.4513 0.3098 0.0523 || || 0.3200 Rank 4 5 2 3 1 Mean rank 3.20 4.80 2.80 2.60 1.40 Final rank 4 5 3 2 1 All the judgements are same as which in Table 4. As NMF-FR and FSHSTC only use Accuracy and F-measure as judgements, so silhouette scores of these two methods in the table are omitted. 4 5 6 7 8 9 10 cluster scoreSilhouette score with different cluster 6 algorithms performance on dierent cluster number formance of HMFF framework is generally better than other algorithms in each format, which means multisource features of HMFF enhance the accuracy and steadiness of clustering eect. At last, we give a case of applying HMFF frame- work in the wild corpus, and high adaptability of HMFF framework is shown in the experiment result.20 Jiaxuan Chen, Shenglin Gui We should note that HMFF framework could accommodate other feature sources from more feature representation algorithms but rather these three sources in section 3.1 only. HMFF framework has high scalable architecture so that fea- ture vectors from additional sources could be easily combined into the HMFF fea- ture matrix without signicantly increasing overall time. The 3 sources of text representation algorithms in section 3.1 have de facto made su- perior performance to other competitors. We believe that integrating with more eective feature sources helps to further enhance clustering performance. In our future work, we will keep seeking more eective feature sources to be integrated into HMFF Liu X , Xiong H , N Shen. A hybrid model of VSM and LDA for text 2nd IEEE International Conference on Computational Intelligence and IEEE, 2017. 2. Abualigah L M , Khader A T , Al-Betar M A . text using K-mean algorithm[C]// International Conference on Computer Science & Information Technology. IEEE, 2016. 3. Aiello L M , Petkos G , Martin C , et al. Sensing Trending Topics in Twitter[J]. on Multimedia, 2013, Kalogeratos A , Zagorisios P , Likas A . Improving Text Stream Clustering using Term Burstiness and ACM, 2016:1-9. 5. Mikolov T , Chen K , Corrado G , et al. Ecient Estimation of Word Representations in Vector Space[J]. Computer Science, 2013. 6. Fujino I , Hoshino Y . Making Topic Words Distribution More Accurate and Ranking Topic Signicance According to the Jensen-Shannon Divergence from Background International Publishing, 2015. 7. Abualigah L M , M Al-Diabat, Shinwan M A , et al. Hybrid Harmony Search Algorithm to Solve the Feature Selection for Data Mining Applications[M]. John Wiley & Sons, Ltd, 2020. 8. Hassani A , Iranmanesh A , Mansouri N . Text mining using nonnegative matrix factoriza- tion and latent semantic analysis[J]. Neural Computing and Applications, 2021:1-22. 9. Abualigah L M , Khader A T . Unsupervised text feature selection technique based on hy- brid particle swarm optimization algorithm with genetic operators for the text Journal of Supercomputing, 2017, 73(11):1-23. 10. De S , Dey S , Bhattacharyya S . Recent Advances in Hybrid Metaheuristics for 2020. 11. Karaa W , N Grib^ aa. Information Retrieval with Porter Stemmer: A New Version for English[M]. 2013. 12. Jones K S . Jones, K.S.: A Statistical Interpretation of Term Specicity and its Application in Retrieval. Journal of Documentation 28(1), 11-21[J]. Journal of Documentation, Blei D M , Ng A Y , Jordan M I . Latent Dirichlet Allocation[C]// Advances in Processing Systems 14 [Neural Information Processing Systems: Natural and Synthetic, NIPS 2001, December 3-8, 2001, Vancouver, British Columbia, Canada]. 2001. 14. Hao P Y , Kung C F , Chang C Y , et al. Predicting stock price trends based on nan- cial news articles and using a novel twin support vector machine with fuzzy Soft Computing, 2020:106806. 15. Mikolov T , Chen K , Corrado G , et al. Ecient Estimation of Word Representations in Vector Space[J]. Computer Science, 2013. 16. Goldberg Y , Levy O . word2vec Explained: deriving Mikolov et al.'s method[J]. arXiv, 2014. 17. QUOC LE, TOMAS MIKOLOV. Distributed Representations of Sentences and Doc- uments[C]. //International Conference on Machine Learning.:Curran Associates,

2108.10926

Jiaxuan Chen

Jiaxuan Chen and Shenglin Gui

Hybrid Multisource Feature Fusion for the Text Clustering

21 pages

cs.CL

http://creativecommons.org/licenses/by/4.0/

The text clustering technique is an unsupervised text mining method which are used to partition a huge amount of text documents into groups. It has been reported that text clustering algorithms are hard to achieve better performance than supervised methods and their clustering performance is highly dependent on the picked text features. Currently, there are many different types of text feature generation algorithms, each of which extracts text features from some specific aspects, such as VSM and distributed word embedding, thus seeking a new way of obtaining features as complete as possible from the corpus is the key to enhance the clustering effects. In this paper, we present a feature fusion (HMFF) framework comprising three representation of multimodel, mutual similarity matrices and feature fusion, in which we construct mutual similarity matrices for each feature source and fuse discriminative features from mutual similarity matrices by reducing dimensionality to generate HMFF features, then k-means could be configured to partition input samples into groups. The experimental tests show our HMFF framework outperforms other recently on 7 of 11 public benchmark datasets and has the leading performance on the rest 4 benchmark datasets as well. At last, we compare HMFF framework with those competitors on a COVID-19 dataset from the wild with the unknown cluster count, which shows the clusters generated by HMFF framework partition those similar samples much closer.

and messy. People speak with accents that make speech recognition difficult, not to mention com- mon spoken dialogue artifacts such as false prosody, filled time-buying and other extra-linguistic further illustrate how challenging spoken dia- logue can be, Schlangen (2020) gives an example of a fairly mundane dialogue between two people walking together in a park: a. Ann: Look at the dog, over at the Bert : I know, isn't it cute! c. (later, they meet a mutual friend) d. Bert : We just saw man carrying a dog! e. Ann: The cutest poodle ever! f. Bert: Actually, I don't think that was a poodle. It was too tall. I think it was one of the participants in this dialogue, say Bert, were to be replaced by an automated system, then Bert would need to be able to handle deictic refer- ences to external objects (a), identify what a dog is (a), know that poodles and labradoodles are types of dogs to the degree that Bert knows the two similar types and can express confi- dence in that knowledge (f; requiring as noted by Schlangen), be able to talk about events that took place in the past about an entity that is no longer perceptually available (d-e), and use anaphoric references such as the word it(b). While it is true that developing working systems re- quires ample data and engineering, handling these types of phenomena together in a single system fruit for research. It is our obser- vation that these kinds of dialogue artifacts are not well-known in the broader research community, to say nothing of higher-level pragmatic issues such as to endowing robots with spoken lan- guage abilities Due to the above-mentioned com- plexities of robots and spoken language, it is not enough to plug existing SDS systems into exist- ing robot platforms and expect any degree of nat- ural, robust, or even useful communication. Fo- cusing just on automatic speech recognition ( ASR) as a part of SDS,ASR in a setting with a robot is wrought with its own challenges: the robot actua- tors could make noise and the changing acoustics of the room due to the robot's location could af- fect speech recognition, to say nothing of the full-duplex nature of co-located spoken interaction; i.e., the robot needs to always be listening even if it's talking or carrying out an action. Beyond speech recognition, natural language un- derstanding ( NLU) is challenging even if is perfect because people don't usually speak in complete, grammatical sentences and the fact that words and phrases refer to real, physical things that could be co-present with the human and the robot or abstractly refer to a known entity. The NLU plays an important role of mapping from speech or transcription to a computable one that is often tied to the task. What should the robot say, and when? Should the robot stick to the task or offer freer chat caba- bilities? Natural language generation ( NLG) is very challenging because what the system says is how the system informs the user of its state or move the dialogue forward towards a shared task goal. It's not enough to say something grammatical, it also needs to be coherent and contribut to the for speech and human-robot inter- action On the same argument that one cannot simply attach the best-performing ASR to a robot and expect natural communication, one cannot sim- ply put humans and robots in the same room to work together on some task and expect even on narrowly-defined tasks. Part of the problem is that the abilities of robots is limited, but the other and perhaps more vexing problem is on the side of the humans: when humans are confronted with robots and told to work with them, even in lab settings, humans immediately have ex- pectations of what the robots should do, and hu- mans also go so far as to anthropomorphize robots for gender, age, social status, among other things (in some cases, only based on how the robots ap- pear, not how they behave), and these expectations and perceptions have implications for the kinds of things robots can do with humans. This intricate area of how humans perceive robots and the nu- ances of how robots and humans interact with each other (using speech or any other means) is the focus of the field of human-robot interaction ( HRI). Both SDSand robots take actions, and in order to take actions they must make decisions. For a SDS, that is often handled in a dialogue manager or dialogue state tracker. Robots likewise have mod- ules that make high-level decisions on goals and low-level decisions on how to reach those goals,like move an arm or move a camera. Should deci- sions about what the system should say and what it should do be made jointly, or separately? It is our goal that the lessons learned from roboti- cists, SDSresearchers (including speech, but also textual modalities) and HRIresearchers can join forces to work towards humans and robots collab- orating on terms that are (at least to some degree) safe and natural for the humans involved. The purpose of this paper is to synthesize what has happened recently in this cross-cutting research area to identify and overcome the and open questions to language com- munication with robots. In the following section, we outline some recent work including workshops and publications that contribute to the discussion, then propose next steps for the growing SLIV AR community. 2 Recent Efforts Two U.S. National Science Foundation workshops were convened in October, 2019: Future Direc- tions Workshop, Toward User-Oriented Agents: Re- search directions and Challenges (Eskenazi and Zhao, 2020) which focused on the role of and how to make them more participants of the workshop identified broad areas and themes for future directions (taking note of common pitfalls (Balentine, 2007)) infrastructure, dynamic views of user- agent interaction, and made several in building low-cost dialogue systems, multi- modal, grounded, and situated interaction, robust and flexible dialogue management, and as good actors. The second workshop Spo- ken Language Interaction with Robots (Marge et al., 2020, 2022) focused on speech and the complexi- ties thereof when robots are involved. The partici- pants identified recommendations relating to eight different themes: 1.First, meeting human needs requires work on new challenges in speech technology and user experience this requires better models of the so- cial and interactive aspects of language use. 3.Third, for robustness, robots need communication with users and bet- ter handling of uncertainty, including simul- taneous consideration of multiple hypotheses and goals.4.Fourth, more powerful adaptation methods are needed, to enable robots to communicate in new environments, for new tasks, and with diverse user populations, without or the collection of massive training data. 5.Fifth, since robots are embodied, speech should function together with other commu- nications modalities, such as gaze, gesture, posture and motion. 6.Sixth, since robots operate in complex envi- ronments, speech components need access to rich yet efficient representations of what the robot knows about objects, locations, user, and other since robots operate in real time, their speech and language processing com- ponents must also. 8.Eighth, in addition to more research, we need more work on infrastructure and resources, in- cluding shareable software modules and inter- nal interfaces, inexpensive hardware, and diverse corpora. A Dagstuhl (Germany) Seminar convened in Jan- uary 2020 on the topic of SLIV AR.4This resulted in organization of other events such as a special ses- sion on robots and dialogue (RoboDial 2.0) at the SIGDIAL 2020 conference,5a workshop on natural language generation at the HRI 2020 conference,6 and a workshop ROBOTDIAL at the IJCAI primary goal of the Dagstuhl Seminar was to provide discussion and establish a community. Discussions revolved around scenarios for human-agent / evaluation, architectures, and situated lan- guage understanding. Other recent, related events include a AAAI Symposium on Natural Communi- cation for Human-Robot Collaboration (2017),8 and RoboNLP workshops including spatial lan- guage ID 20021 and links to prior workshops many have contributed in different ways to the convergence between robotics, SDS, and HRI research, two recent articles highlight recent and opportunities at these Tellex et al. (2020) surveys the use of natural language from a robotics point of view including natural language issues that are common on robots and current state of the art with focus on tasks such as robot navigation. Kragic et al. (2018) makes the case that though there has been progress, things like full autonomy, learning in people, and safe and flexible performance in varied environments remain elusive. 3 Basic et al. (2020) identified 5 basic require- ments for robot-ready SDS. These requirements are largely at the level of framework and processing; not all tasks and applications will need all of nor is the list comprehensive Section below), but a goal that ro- bust and natural dialogue will need and as a starting point for the opportunities that follow in the : robot components are modular and individual modules must be able to integrate with : robots are situated having many sensors along with the SDSspeech input; those sensors must be inte- grated : robot and SDSmodules are expensive; modules should be able to easily communicate with each other in a distributed : modules must be able to pro- cess input quickly and : sensors must be temporally aligned; i.e., synchronized in time 4 Tools and Platforms: A common research platform should be developed or ex- tended for intelligent agent The collection of situated and mul- timodal corpora is an important future requirements were derived from discussions in Tasks: With a shared corpora, shared tasks will promote effec- tive and comparable Metrics need to be designed that account for the nuances of dialogue with in- telligence agents like robots. *Science and Engineering: Work to elucidate the fundamental questions in real-time both scientific and engineering. 4.2 Multimodal, grounded and range of modalities used in today's dialog systems should be broadened, for example by including prosody, haptics, or motion capture as input dialog should play a central role in future dialog systems research. *The advances of continuous (i.e., processing should be generalized to incremental Audio and Speech exploit context and expectations in speech creating a speech recognition system focused on the issues encountered in robotics (e.g., splitting speech signals and speech in noise where the noise may come from the robot and/or the surrounding exploit prosody, emotion, and mental state from the speech signal. *Use audio scene and sound event analysis to better understand the environment. 4.4 Mutual understanding between robot and human *Develop language understanding models for robots that resolve referential ambiguities, par- ticularly to objects and lo- cations in situated increase the fidelity of how robots represent the state of the human and understanding should be based on the user's goals and the broader space of clarification and recovery strategies in spoken language inter- action with co-present agents, including when and how. *Research should be highly linguistics, pragmatics and on language not only as a way to achieve human-like behaviors, but also as a way to support limited but highly usable com- munications abilities. *Work to better characterize the list of commu- nicative competencies most needed for robots in various scenarios 4.5 Robust and should be inspired by human dia- log techniques in order to efficiently address others goals. Discourse context (as well as physical context) should be preferred to mak- ing explicit representations of context should be designed in a way that enables them to be shared with other domains, tasks and should support different user style of interaction preferences * Include functionality. * Make components robust to not only on improving better core com- ponents, but also on cross-cutting issues and those that have fallen between the cracks *Make systems and components adaptable to users. 4.6 are often narrow, but should be as generalizable as are dependant on the target user pool / should be sensitive to aspects of the individual user (e.g. their cognitive or behavioral. change, their speaking proficiency) 4.7 Dynamic Views of User-Agent research from generic, static views of the user, the agent and their relationship to personalized and dynamic models *Move from human-computer to the agent learns about different emphasize scenarios that involve longer conversations (beyond 1-2 turns) and/or more frequent interactions. from the same user or groups of engineer user perceptions Building Dialogue Systems in Low-Resource Conditions *A systematic way to introduce prior knowl- edge as an input to data-driven effort to make pretraining and transferable models well-documented, thor- oughly tested and widely robust and reusable tools for solved tasks, e.g. wizard-of-oz interface, data collection tools, human evaluation platform). *A principled approach for system design, data collection, resources management, and evalu- ation for new AI projects with limited/no data *New machine learning methods that can bet- ter utilize knowledge from other domains or unannotated data. *Shared tasks that standardize the evaluation of dialog agents under low resources setting, and regular benchmark comparison across Intelligent Agents as Good Actors *Research plans within scientific proposals should raise and address at least one of students and newcomers to the field on ethical issues. *Develop outreach efforts for increasing public awareness in collaboration with target every NSF/government proposal or grant that includes the release of data, the data should have a data sheet. 4.10 Policy *Fund spoken language interaction with robots as its own area of research. * Prefer evaluation-based on use cases. *Support many kinds of research and develop- ment to overcome the barriers to data novel public-private partnerships for open source software 5 Discussion Taken together, we highlight some of the opportuni- ties and complementary goals from within the and robotics in Figure 1. The figure shows a human interacting with a robot in a co- located environment with two objects; the dashed lines show perceptions, solid lines and arrows show

2108.10931

Casey Kennington

Casey Kennington

The State of SLIVAR: What's next for robots, human-robot interaction, and (spoken) dialogue systems?

6 pages

cs.CL

http://creativecommons.org/licenses/by-sa/4.0/

We synthesize the reported results and recommendations of recent workshops and seminars that convened to discuss open questions within the of robotics, human-robot interaction, and spoken dialogue The goal of this growing area of research interest is to enable people to more effectively and naturally communicate with robots. To carry forward opportunities networking and discussion towards concrete, projects, we encourage interested parties to consider participating in future virtual and in-person discussions and workshops.

2 ZHENKUN LI, YI XIE, AND BOYU ZHANG To see that (1.1) follows from (1.2), take M=YnB3and let be a simple closed curve on @B3. Another interesting special case of (1.2) is when M is a link complement. Let Lbe a link in S3, letN(L) be an open of L, letM=S3nN(L), and let @Mbe the union of a pair of oppositely oriented meridians for every component of L. In this case, by [Juh06, Proposition 9.2], inequality (1.2) is equivalent (1.3) whereKis an arbitrary knot in S3nL. It is natural to ask about the prop- erties of the pair ( L;K) when the equality of (1.3) holds. This question has been studied by Ni [Ni14] and Kim [Kim20]. Ni [Ni14] proved that if Lis the unlink and the equality of (1.3) holds, then L[Kis the unlink. (This property was then used in [Ni14, Proposition 1.4] to prove that the rank of [HFK detects the unlink.) Kim [Kim20, Proposition 4] proved that if the equality of (1.3) holds and L[Kis non-split, then Lis non-split and the linking number of Kwith every component of Lis zero. It is conjectured by Kronheimer and Mrowka [KM10b, Conjecture 7.25] that sutured Heegaard Floer homology is isomorphic to sutured instanton Floer homology. This paper studies an analogue of (1.2) for instanton Floer homology. From now on, all instanton Floer homology groups are An inequality on instanton Floer homology analogous to (1.2) was proved by Li-Ye [LY20, Proposition 3.14]: let ( M; ),K, and Kbe as above, we (1.4) Remark 1.2.To see that (1.4) follows from [LY20, Proposition 3.14], remove a small 3{ball B3fromMcentered on Ksuch thatB3Kis an arc. Let be a simple closed curve on @B3separating the two points in @B3K. Consider the balanced sutured manifold ( MnB3; [), and letT=KnB3 be a vertical tangle in MnB3. Then (MnN(K); K) can be obtained from (MnB3; [) by removing a neighborhood of Tand adding a meridian of Tto the suture. By [LY20, Proposition 3.14] and [Li18, Proposition = this paper, we will study the properties of Kwhen the equality of (1.4) is achieved. To simplify notation, we introduce the following 1.3. Suppose (M; ) is a sutured manifold. Suppose Kis a knot inMsuch that [ K] = 02H1(M;@M ;Q). LetN(K)Mbe an open tubular neighborhood of K. Let Kbe the union of and a pair of oppositely oriented meridians on @N(K). We say that Kisinstanton Floer minimal , = FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 3 Suppose (M; ) is a connected balanced sutured manifold. Let f ig1is be the components of , then the homotopy class of every idenes a con- jugacy class [ i] in1(M). Dene 0 1(M) the normal subgroup of 1(M) generated by the conjugacy classes [ 1],


, [ s]. The rst main result of this paper is the following theorem. Theorem 1.4. Let(M; )be a connected balanced sutured manifold with SHI(M; )6= 0: SupposeMis decomposed (1.6) where each Miis an irreducible sutured manifold (in particular, @Miis non- empty), and Yis a closed manifold. Here, Yis allowed to be S3or a knot in the interior of Mthat represents the zero element in 0 1(M). IfKis instanton Floer minimal, then Kis contained in 1.5.If the knot Krepresents the zero element in 0 1(M), then [K] = 02H1(M;@M ;Q), and hence Denition 1.3 is applicable to K. Moreover, when Y=S3in (1.6), we have the following result. Theorem 1.6. Suppose (M; )is a balanced taut sutured manifold or a connected sum of balanced taut sutured manifolds. Let Kbe a knot in the interior ofMthat represents the zero element in 0 1(M). ThenKis instan- ton Floer minimal if and only if Kis the a link in S3andKis a knot in S3nL. By (1.7) The following result characterizes all the pairs ( L;K) such that the equality of (1.7) is 1.7. LetLS3be a link and KS3nLbe a knot. = 2
dimCKHI(L) (1.8) if and only if Kis the unknot in S3nL. Theorem 1.7 has the following immediate 1.8. SupposeLis ann{component link in S3. ThenLis an unlink if and only if dimCKHI(L) = 2n1. Proof. SupposeKis a knot in S3, then by [KM10b, Proposition 7.16], (1.9)4 ZHENKUN LI, YI XIE, AND BOYU the unknot if and only if dim CKHI(K) = 1. IfLis the unlink with ncomponents, then dim CKHI(L) = 2n1by Theorem 1.7 and induction on n. On the other hand, suppose Lis a link with ncomponents such that dim CKHI(L) = 2n1. LetK1;


;Knbe the components of L. Then by (1.7) and (1.9), we = everym<n . Therefore by Theorem 1.7, Km+1is an unknot everym<n , soLis the unlink.  Theorems 1.4, 1.6 and 1.7 will be proved in Section 3.3. 2.Sutured manifolds and sutured manifolds and tangles. This subsection reviews some terminologies on balanced sutured manifolds and 2.1 ([Gab83], [Juh06]) .LetMbe a compact oriented 3-manifold with boundary. Suppose @Mis an embedded closed oriented that (1) every component of @Mcontains at least one component of , (2) [ ] = be an open tubular neighborhood of , and letR( ) = @MnA( ). LetR+( ) be the subset of R( ) consisting of components whose boundaries have the same orientation as , and letR( ) =R( )nR+( ). We say that ( M; ) is a balanced sutured manifold if it satises the The 3{manifold Mhas no closed 2.2. Suppose (M; ) is a balanced sutured manifold. (1) A tangle in (M; ) is a properly embedded 1-manifold A component T0ofTis called vertical A tangle Tis called vertical if every component of Tis vertical. (4) A tangle Tis called balanced 2.3.In the following, we will use the notation T(M; ) to indicate that the tangle Tis in (M; ). Remark 2.4.Note that a tangle T(M; ) is, a priori, un-oriented. When the tangle is vertical, we may orient each component of Tso that it goes fromR+( ) toR( ). Remark 2.5.The tangle Tin Denition 2.2 is allowed to be empty.ON FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 5 Denition 2.6 ([Sch89]) .Suppose (M; ) is a balanced sutured is a (possibly empty) tangle, and SMis an oriented surface so that If Sis connected, dene theT{Thurston norm ofSto be xT(S) = the number of intersection points between TandS. and S1, ...,Snare the connected components of S, then norm ofSto be xT(S) =xT(S1) a subset. For a homology class 2H2(M;N ;Z), norm ofto be xT() = minfxT(S)jSis properly embedded in M; @SN;[S] = []2H2(M;N ;Z)g: A surfaceSproperly embedded in Mis said to be if xT(S) =xT([S]): Here [S] is the homology class in H2(M;@S ;Z) represented by S. WhenT=;, we simply write xTasx. Denition 2.7 ([Sch89]) .Suppose (M; ) is a balanced sutured is a (possibly empty) tangle. We say that ( M; ) isT-taut, if (M; ;T ) satises the following andR( ) are both incompressible in MnT. (3)R+( ) andR( ) are both. (4) Every component of Tis either vertical or closed. WhenT=;, we simply say that ( M; ) manifold hierarchy. Suppose (M; ) is a balanced sutured manifold and T(M; ) is a tangle. Suppose SMis an embedded surface. Under certain mild conditions, generalized the concepts from [Gab83] and introduced a process to decompose the triple ( M; ;T ) into a new triple ( M0; 0;T0), and this process is called a sutured manifold decomposition . Extending the work of Gabai in [Gab83], Scharlemann [Sch89] also proved the existence of sutured for aT-taut balanced sutured manifold ( M; ). In this paper, we will use a modied version of the sutured manifold hierarchy, which is presented in the current 2.8 (c.f. [GL19]) .Suppose (M; ) is a balanced sutured man- ifold andT(M; ) is a tangle. An oriented properly embedded surface SMis called admissible in (M; ;T ) if (1) Every component of and transversely.6 ZHENKUN LI, YI XIE, AND BOYU ZHANG (2) For every component T0ofT, choose an arbitrary orientation for T0, then all intersections between T0andSare of the same sign. (3)Shas no closed components. (4) The is even. The following theorem is essentially a combination of [Sch89, Theorem 4.19] and [Juh08, Theorem 8.2]. Theorem 2.9. Suppose (M; )is a balanced sutured manifold. Suppose K is a knot in the interior of Mand viewKas a tangle in (M; ). then there is a nite sequence of sutured manifold )S1 (M1; 1;K) ::: (Mn1; n1;K)Sn (Mn; n;T) so that the following is true. (1)Fori= 1;:::;n , the surface Siis an admissible surface in 1;:::;n1, we a vertical tangle in The proof of the theorem is essentially a combination of the proofs of [Sch89, Theorem 4.19] and [Juh08, Theorem 8.2]. We will only sketch the proof here and point out the adaptions to our current setup. First, suppose there is a homology class 2H2(M;@M ;Z) so that @()6= 02H1(@M;Z). The argument in [Sch89, Section 3] then produces a surfaceSproperly embedded in Mso that (1) [S] = []2H2(M;@M ;Z). (2) The sutured manifold decomposition of ( M; ;K ) alongSyields a taut sutured manifold. Note thatSmust have a non-empty boundary. We further perform the following modications on S: a) Disregard any closed components of S. This makes the decomposi- tion alongSresulting in a balanced sutured manifold. b) Apply the argument in [Gab83, Section 5] to make Ssatisfying the following extra property: for every component VofR( ), the intersection of @SwithVis a collection of parallel oriented non- separating simple closed curves or arcs. Note the argument in [Gab83, Section 5] happens in a collar of @MinsideM, and the tangle Kis in the interior of M. So the argument applies to our case. c) After step b), we can make Sadmissible by further isotope @Son @Mvia positive stabilizations on S, in the sense of [Li19, Denition 3.1]. Note that by denition, a positive stabilization creates a pair of intersection points between @Sand , and thus it is possible to mod- ifySso that every component of FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 7 is even. By [Li19, Lemma 3.2], we know that after positive the sutured manifold decomposition along Sstill yields a K{taut (orT{taut, ifSK6=;) balanced sutured the resulting surface after the above modications. If S1has a non-trivial intersection with Kthen we are done. If S1K=;, we then know thatS1is admissible and the sutured manifold )S1 (M1; 1;K) yields aK-taut balanced sutured manifold ( M1; 1). We can run the above argument repeatedly. By [Sch89, Section 4], the triple ( M1; 1;K) has a smaller complexity than ( M; ;K ), so this decomposition procedure must end after nitely many steps. There are two possibilities when the 1 . We have a surface Snthat intersects Knontrivially, and S1, ..., Sn1are all disjoint from K. After the decomposition along Sn, the knot Kbecomes a collection of arcs Tinside (Mn; n), and we have a sutured manifold (Mn; n;T) so that (Mn; n) isT-taut. Note that Thas no closed components and by Condition (4) in Denition 2.7, Tis a vertical tangle. Hence the of the theorem are veried. Case 2 . All decomposing surfaces Snare disjoint from K, and we end up with a balanced sutured manifold ( Mn; n) so the zero map. In this case, the injective. It then follows from a straightforward diagram chasing (see, for example, [Hat07, Lemma 3.5]) that @Mnmust be a union of 2-spheres. Let ^Mnbe the component of Mnthat contains K, and let ^Sbe a sphere in^Mnthat is parallel to one of the boundary components. Then ^Sis a reducing sphere for ^MnnK, therefore ( Mn; n) cannot be K{taut, which yields a contradiction.  2.3.Instanton Floer homology. Suppose (M; ) is a balanced Kronheimer and Mrowka constructed the sutured instanton Floer homology for the couple ( M; ) in [KM10b]. Given a balanced tangle T (M; ) (see Denition 2.2), an instanton Floer theory for the triple ( M; ;T ) was constructed in [XZ19b]. In this subsection, we review the constructions of these instanton Floer (M; ) is a balanced sutured manifold and T(M; ) is a bal- anced tangle. We construct a tuple ( Y;R;L;u ) as follows: Pick a that @F= :8 ZHENKUN LI, YI XIE, AND BOYU ZHANG Pickn2 pointsp1, ...,pnonF. Letube an embedded simple arc connecting p1andp2. Glue given by a dieomorphism from A( ) to [1;1] . The of two 1gF are closed surfaces. pick an thatf(eTR+) =f1gu. a closed oriented 3-manifold Y. LetRYbe the image of R. LetLYbe the image of T, thenLis a link inY. We will also abuse notation and let udenote the image of 2.10. The tuple ( Y;R;L;u ) is called a closure of (M; ;T ). By the construction of Kronheimer and Mrowka in [KM11], there is an instanton homology group I( Y;L;u ), which is a C{vector space, associated to the triple ( Y;L;u ). The surface Rinduces a complex linear map orb(R) : I(Y;L;u )!I(Y;L;u ): Given a point p2YnL, there is a complex linear map (p) : I(Y;L;u )!I(Y;L;u ): orb(R) and(p) are commutative to each other. From now on, if is a linear map on I( Y;L;u ), we will use Eig( ;) to denote the of with eigenvalue in I(Y;L;u ), Denition 2.11 ([XZ19b, Denition 7.6]) .Suppose (M; ) is a balanced sutured manifold, and T(M; ) is a balanced tangle. Let ( Y;R;L;u ) be an arbitrary closure of ( M; ;T ) such thatjLRjis odd and at least 3 (this can always be achieved by increasing the value of nin the construction of the closure). Let p2YnLbe a point. Then the sutured instanton Floer homology of the triple ( M; ;T ), which is denoted by SHI( M; ;T ), is ) = ) The next statement guarantees that the isomorphism class of SHI( M; ;T ) is well-dened:ON FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 9 Theorem 2.12 ([XZ19b, Proposition 7.7]) .For a triple (M; ;T ), the iso- morphism class of the sutured instanton Floer homology SHI(M; ;T )as a complex vector space is independent of the choice of the closure of (M; ;T ) and the point p. We also have the following non-vanishing theorem: Theorem 2.13 ([XZ19b, Theorem 7.12]) .Suppose (M; )is a balanced sutured manifold and T(M; )is a balanced tangle so that (M; )isT{ taut. Then we have SHI(M; ;T )6= 0: Now we establish a theorem relating sutured instanton Floer homology to sutured manifold decompositions, which is analogous to [KM10b, Propo- sition 6.9]: Theorem 2.14. Suppose (M; )is a balanced sutured manifold and T (M; )is a balanced tangle. Suppose further that S(M; )is an admis- sible surface in the sense of Denition 2.8 and we have a sutured )S (M0; 0;T0): Then there is a closure (Y;R;L;u )of(M; ;T )and a closure (Y;R0;L;u )of (M0; 0;T0)so that the two closures share the same 3-manifoldY, the same linkLand the same arc u. Furthermore, there is a closed oriented thatR0is obtained from RandSby a double curve surgery in the sense of [Sch89, Denition 1.1] . Finally, let p2YnL be a point, then we have SHI(M; ;T ) = = (2.2) Proof. The proof of the theorem is almost the same as that of 6.9]: the constructions of the closure ( Y;R;L;u ) and the ex- tension Sfollows verbatim from the constructions in the proof of 6.9]. Equation (2.1) is precisely the denition of SHI( M; ;T ) in Denition 2.11. The verication of Equation (2.2) is also similar to the proof of [KM10b, Proposition 6.9], except that the surface in the current setup might intersect Ltransversely at nitely many points, so we need to replace [KM10b, Corollary 7.2] in the proof of [KM10b, Proposition 6.9] by [XZ19b, Proposition 6.1] for our current setup.  Remark 2.15.The original proof of [KM10b, Proposition 6.9] only works with a surface is even (or else it is impossible to extendSto a closed surface Sin the closures of ( M; ).) This is the10 ZHENKUN LI, YI XIE, AND BOYU ZHANG reason why we needed to introduce the denition of admissible surfaces in Denition 2.8. This subtlety was not explicitly mentioned in [KM10b]. 2.4.A module structure on SHI.We introduce an additional module structure on SHI( M; ;T ) whenT6=;. As before, suppose ( M; ) is a balanced sutured manifold and T(M; ) is a balanced tangle. Pick an arbitrary closure ( Y;R;L;u ) of (M; ) such thatjLRjis odd and at least 3 (so that the condition of Denition 2.11 is satised). Pick a point p2YnLand a point q2T. By [XZ19b, (2), Proposition 2.1], the point qinduces another complex linear map (q) : I(Y;L;u )!I(Y;L;u ) and we have (q)2= 0. Moreover, (q) commutes with (R) denes a complex linear map on SHI(M; ;T ) = squares to zero. Therefore, we can view SHI( M; ;T ) as where the action of Xis dened by (q). The proof of [XZ19b, Theorem 7.12] applies verbatim to verify the following theorem: Theorem 2.16. For a triple (M; ;T ), the isomorphism class of the Floer homology SHI(M; ;T )as is indepen- dent of the choice of the closures of (M; ;T ). Moreover, if q;q02Tare in the same component of T, then the module structure induced by qis isomor- phic to the module structure induced by q0.  Forq2T, we will call the C[X]=(X2){module structure on SHI( M; ;T ) dened as above the module structure induced by q. Remark 2.17.IfThas only one component, then the isomorphism class of the module structure is independent of q. In this case, it makes sense to refer to the isomorphism class of the module structure on SHI( M; ;T ) the point q. Corollary 2.18. Suppose (M; )is a balanced sutured manifold and T (M; )is a balanced tangle. Suppose further that S(M; )is an admis- sible surface in the sense of Denition 2.8 and we have a sutured )S and letq02T0be the image of qinM0. If the module structure on SHI(M; ;T )induced by qis a free then so is the module structure on by q0. Proof. This follows directly from Equation (2.2) and the fact that the with the action of (S). ON FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 11 2.5. SHI with local coecients. We also introduce a version of Floer homology dened with local coecients. Let ( M; ) be a balanced sutured manifold and let T(M; ) be a tangle. Suppose LT is a link in the interior of M. Let (Y;R;L0;u) be a closure of ( M; ;T ) odd and at least 3. Let Lbe the local system for the triple ( Y;L0;u) associated with LL0 as dened in [XZ19, Section 2]. By denition, Lis a local system over Forh2C, dene L(h) = L R=(th), then L(h) is a local system over the ring R=(th)=C, and hence I( Y;L0;u; L(h)) is aC{linear space. Let p2YnL0be a point. Denition 2.19. Dene SHI( M; ;T ; L(h)) to be the simultaneous gener- alized eigenspace of orb(R) and(p) in I(Y;L0;u; L(h)) with and 2 the proof of [XZ19b, Theorem 7.12] applies verbatim to verify that: Theorem 2.20. ForM; ;T;L as above and h2C, the isomorphism class of the sutured instanton Floer homology SHI(M; ;T ; L(h))as aC{linear space is well-dened. In other words, it is independent of the choice of the closures of (M; ;T )and the point p.  Lemma 2.21. lim ; ): Proof. This is an immediate consequence of [XZ19, Lemma 4.1 (1)] and the observation that L(1) is the trivial local system over C, thus SHI(M; ;T ; L(1)) = SHI( M; ;T ):  Lemma 2.22. SupposeT0is a tangle in (M; ), andL;L0Mare two links in the interior of Mthat are homotopic to each other in Mand are both disjoint from T0. and suppose h =2f ; L(h)) = dim C(M; ;T0; L0(h)) Proof. This is an immediate consequence of [XZ19, Proposition 3.2].  Lemma 2.23. SupposeTis a tangle in (M; )andUis an unknot in MnT. Then for every h2C, we = 2
dimCSHI(M; ;T ): Proof. LetBbe a 3{ball in Mn(T[U), letbe a simple closed curve on @B. LetU0S3be an unknot in S3, letB0be a 3{ball in S3nU0, let0be a simple closed curve on @B0. By the connected sum formula for SHI (c.f. [Li18, Proposition 4.15]), we U(h))12 ZHENKUN LI, YI XIE, AND BOYU ZHANG = U(h)) = 2
dimCSHI(M; ;T dim CSHI(M; ;T U0(h)): Since (S3nB0;0) is a product sutured manifold, we = 1: Let c(h) = dim is a function depending on h, and we ): for allM; ;T . Recall that Tis allowed to be empty. Now we invoke the computations of annular instanton Floer homology AHI with local coe- cients from [XZ19, Example 3.4]. Since AHI can be regarded as a special case of sutured instanton Floer homology, by [XZ19, Example 3.4], we must havec(h) = 2 for all h.  Remark structure introduced in Section 2.4 cannot be extended to SHI with local coecients. The reason is that [XZ19, (2)] may no longer hold for instanton Floer homology with local coecients, so we may not have (q)2= 0 on SHI( M; ;T ; of Floer minimal and module structure. This subsection proves the following theorem. The result is inspired by the work of Wang [Wan21]. Theorem 3.1. Suppose (M; )is a balanced sutured manifold with SHI(M; )6= 0 (3.1) andKis a knot in the interior of M. ViewKas a balanced tangle in (M; ). IfMnKis irreducible, then SHI(M; ;K )is not free as a SinceMnKis irreducible, by (3.1) and the adjunction inequality for instanton Floer homology [KM10b, Theorem 7.21], we conclude that ( )S1 (M1; 1;K) ::: (Mn1; n1;K)Sn (Mn; n;T)ON FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 13 be the sutured hierarchy given by Theorem 2.9. In particular, Tis a vertical tangle and ( Mn; n) isT-taut. Assume SHI( M; ;K ) is a free Applying Corollary 2.18 repeatedly and applying Theorem 2.13, we conclude that for every point q2T, the module structure on SHI( Mn; n;T) induced by qis a any closure ( Y;R;L;u ) of (Mn; n;T) such thatjRLjis odd and at least 3. Let N(R)=[1;1]Rbe a closed tubular neighborhood of R inY. SinceRT6=;, we can choose qso that it is contained in N(R). Isotope the arc uso at one point p0. To simplify notation, we will identify N(R) with [1;1]R. After shrinking N(R) and taking isotopies for Landu, we may assume that (1)L(R) is given by is given by we can cut Yopen along the two surfaces re-glue via the dieomorphism from by the identity map on R. This yields a disjoint union Y1tY2, where (1)Y1is given by ( Yn[1;1]R)=, whereis the identication given by ([1;1]R)=, whereis the same identication as above. ThenY1is dieomorphic to Y, andY2is dieomorphic to S1R. LetR1be the image off1gRinY1, letL1,u1be the images of Clearly ( Y1;R1;L1;u1) is dieomorphic to ( Y;R;L;u ). InY2, letR2=f0gR, and in [XZ19b, Theorem 6.4], there is a cobordism an isomorphism (as C{linear spaces) e: the notation I( Yi;Li;uijRi) is dened in [XZ19b, Denition 6.3]. Re- call that we assumed the base point qis in the neighborhood [ 1;1]R. Let q22Y2be the image of q. Then(q2) acts By Section 2.4, we have (q2)2= 0. On the other hand, by [XZ19b, Proposition implies that (q2) = 0 on I( Y2;L2;u2jR2), so(q2) denes the zero action on the tensor the cobordism map eintertwines the (q2){action ZHENKUN LI, YI XIE, AND BOYU ZHANG with the(q){action on SHI( Mn; n;T). Thus we conclude that the action ofX(which equals (q)) on SHI(Mn; n;T) must also be zero. As a conse- quence, the module structure on SHI( Mn; n;T) induced by qis not a free which yields a contradiction.  3.2.Floer minimality and module structure. Recall that for a sutured manifold (M; ), the group 0 1(M) is dened by Equation (1.5) as a quotient group of1(M). The main result of this subsection is the following theorem. Theorem 3.2. Let(M; )be a balanced sutured manifold. Suppose Kis a knot in the interior of Mthat represents the zero element in 0 1(M). IfK is instanton Floer minimal (see Denition 1.3), then SHI(M; ;K )is a free module over C[X]=(X2)(see Section 2.4). We rst establish the following rank inequality for instanton homology. Lemma 3.3. Suppose (M; )is a balanced sutured manifold, and Kis a knot in the interior of Msuch thatKrepresents the zero element in 0 1(M). Then we (3.2) Proof. For each i, let 0 i; 00 i be a pair of simple closed curves on @Mparallel to i. Moreover, let 00 ihave the same orientation as , let 0have the as , and let 0be placed between and 00. LetUbe an unknot in M. The proof of [KM10, Theorem 3.1] implies = 2m
dimCSHI(M; ) = 2m
dimCSHI(M; ;K ) = 2m
dimCSHI(M; ;U ): (3.5) Decompose A( 0) asA( ) For each i, letDibe a disk, let'ibe an dieomorphism from [ and ^ be the image of[i( i[ 00 i) on@^M. Then every component of @^M contains at least one component of ^ . Moreover, it is straightforward to verify that ( ^M;^ ) is a balanced sutured manifold. Let Ti= a tangle, we can view T=T1[:::[Tmas a vertical tangle inside ( ^M;^ ). Let ( ^MT;^ T) be the balanced sutured manifold obtained from (^M;^ ) by removing a neighborhood of Tand adding a meridian of every component of Tto the suture. It is straightforward to check that ( ^MT;^ T) is simply dieomorphic to ( M; 0). Hence by [XZ19b, Lemma 7.10], we ^M;^ ;T):ON FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 15 In fact, the proof of [XZ19b, Lemma 7.10] also applies verbatim to ^M;^ ;T[U): By Lemma 2.23, we have dimCSHI( ^M;^ ;T[U; U(h)) = 2
dimCSHI( ^M;^ ;T) for everyh2C. SinceKrepresents the zero element in 0 1(M), it is homo- topic toUin^M. By Lemma 2.22, when h =2f 1g, we have dimCSHI( ^M;^ ;K[T; K(h)) = dim CSHI( ^M;^ ;K[U; ^M;^ ;K[T; K(h)) = 2
dimCSHI( ^M;^ ;T) for allh =2f 1g. Therefore, by Lemma = dim CSHI( ^M;^ ;T) = by (3.3) and (3.4), we the lemma is proved.  Now we can prove Theorem 3.2. Proof of Theorem 3.2. We dene a new instanton Floer homology as follows. Regard Kas a tangle inside ( M; ). Pick a clo- sure (Y;R;L;u ) of (M; ;K ). Recall that from the construction of closures in Section 2.3, Lhas the the image of the arcs [ the closure. Now dene L0=K[[L1to be a link in Y, whereis a meridian ofK, and dene u0=u[v, wherevis an arc connecting Kwith. := 3.4.If we replace ( M; ) by a closed 3-manifold Yand disregard R,L1, andu, then the above construction is the same as the denition of I(Y;K) introduced by Kronheimer and Mrowka in [KM11].16 ZHENKUN LI, YI XIE, AND BOYU ZHANG K = 0 [ v] = 1. The knotK,K, and the unoriented skein triangle. Now apply the un-oriented skein triangle from [KM11, Section 6 and 7] to a crossing between Kand, as shown in Figure 1. The result is the following exact )f//SHI(M; ;K mapfhas been computed explicitly in [Xie, Proposition 3.4], and we have f=(q) : SHI(M; ;K )!SHI(M; ;K ) for a base point q2K. Since(q)2= 0, we (3.6) and the equality holds if and only if rank((q)) ): (3.7) Let Kbe the union of with a pair of oppositely oriented meridians of Kon@N(K). The proof of [KM11, Proposition 1.4] (also [XZ19b, Lemma 7.10]) applies verbatim to give the following (3.8) Recall that Kis assumed to be instanton Floer minimal, = 2
dimCSHI(M; ): (3.9) By Lemma 3.3, we (3.10) By (3.6), (3.8), (3.9), (3.10), we conclude ) = dim CSHI(M; ;K);ON FLOER MINIMAL KNOTS IN SUTURED MANIFOLDS 17 therefore the equality holds for (3.6), and hence (3.7) holds. Recall that the action ofXequals(q) in the C[X]=(X2){module structure of SHI( M; ;K ). Hence by (3.7), SHI( M; ;K ) is a free  3.3.Proofs of the main results. This subsection proves Theorems 1.4, 1.6, and 1.7. Proof of Theorem 1.4. By Theorem 3.2, SHI( M; ;K ) is a non-trivial free module over C[X]=(X2). Hence by Theorem 3.1, MnKis as M=M1#M2, whereKM2after isotopy. IfM2is a closed manifold, then the desired result holds. If M2is a manifold with boundary, since SHI( M; )6= 0, by [KM10b, Theorem 7.21], R+( ) and R( ) are both norm-minimizing, therefore M2is also a balanced By the connected sum formula for SHI, we ) = the desired result follows from induction on the number of factors in the prime decomposition of M.  Proof of Theorem 1.6. By the connected sum formula for SHI and the non- vanishing theorem for taut sutured manifolds [KM10b, Theorem 7.12], we have SHI(M; )6= 0. By Theorem 1.4, there exists a 3{ball B3Msuch thatKB3. Hence we can view Kas a knot in the northern hemisphere of S3. Dene S3(K) =S3nN(K), and let @S3(K) be a pair of oppositely of K. Let K@(MnN(K)) be the union of and a pair of oppositely oriented meridians on @N(K). The connected sum formula for SHI = instanton Floer minimal if and only the unknot detection result for I(S3;K) [KM10b, Proposition 7.16], the above equation holds if and only if Kis the unknot.  Proof of Theorem 1.7. LetM=S3nN(L), then the prime not contain any non-trivial closed component. Let @M be the suture consisting of a pair of oppositely oriented meridians for each component of L, then0 1(M) = 0. By [KM10b, Proposition 7.16], we have SHI(M; )6= 0, therefore Msatises the conditions of Theorem 1.6, and hence (1.8) holds if and only if Kis the unknot in M. David Gabai, Foliations and the topology of 3-manifolds , J. Dierential Geom. 18(1983), no. 3, 445{503. MR723813 Sudipta Ghosh and Zhenkun Li, Decomposing sutured monopole and instanton oer homologies , v2 (2019).18 ZHENKUN LI, YI XIE, AND BOYU ZHANG [GL21] Joshua Evan Greene and John Luecke, Fibered simple knots , Allen Hatcher, Notes on basic 3-manifold topology , 2007. [Hed11] Matthew Hedden, On oer homology and the berge conjecture on knots admitting lens space surgeries , Transactions of the American Mathematical Society 363 (2011), no. 2, 949{968. [Juh06] Andr as Juh asz, Holomorphic discs and sutured manifolds , Algebr. Geom. Topol. 6(2006), 1429{1457. Andr as Juh asz, Floer homology and surface decompositions , Geom. Topol. 12 (2008), no. 1, 299{350. Juhyun Kim, Links of second smallest knot Floer homology , P. B. Kronheimer and T. S. Mrowka, Instanton Floer homology and the Alexan- der polynomial , Algebr. Geom. Topol. 10(2010), no. 3, 1715{1738. ,Knots, sutures, and excision , Journal of Dierential Geometry 84 (2010), no. 2, 301{364. [KM11] ,Khovanov homology is an unknot-detector , Publ. Math. Inst. Hautes Etudes Sci. 113(2011), 97{208. MR2805599 [Li18] Zhenkun Li, Contact structures, excisions, and sutured monopole Floer homol- ogy, arXiv:1811.11634 (2018). [Li19] ,Knot homologies in monopole and instanton theoreis via sutures v6 (2019). [LY20] Zhenkun Li and Fan Ye, Instanton Floer homology, sutures, and Heegaard dia- grams , arXiv: 2010.07836 (2020). [Ni14] Yi Ni, Homological actions on sutured Floer homology , Math. Res. Lett. 21 (2014), no. 5, 1177{1197. MR3294567 [NW14] Yi Ni and Zhongtao Wu, Heegaard Floer correction terms and rational genus bounds , Adv. Math. 267(2014), 360{380. MR3269182 [OS04] Peter Ozsv ath and Zolt an Szab o, Holomorphic disks and knot invariants , Adv. Math. 186(2004), no. 1, 58{116. Jacob Rasmussen, Lens space surgeries and l-space homology spheres (2007). [RR17] Jacob Rasmussen and Sarah Dean Rasmussen, Floer simple manifolds and L- space intervals , Advances in Mathematics 322(2017), 738{805. [Sch89] Martin Scharlemann, Sutured manifolds and generalized Thurston norms , J. Dif- ferential Geom. 29(1989), no. 3, 557{614. MR992331 [Wan21] Joshua Wang, Link oer homology also detects split links , Bulletin of the Society (2021). [Xie] Yi Xie, Earrings, sutures and pointed links , to appear in International Mathe- matics Research Notices. [XZ19a] Yi Xie and Boyu Zhang, Classication of links with Khovanov homology of min- imal rank , arXiv:1909.10032 (2019). [XZ19b] ,Instanton Floer homology for sutured manifolds with tangles (2019).

2108.10933

Zhenkun Li

Zhenkun Li, Yi Xie, and Boyu Zhang

On Floer minimal knots in sutured manifolds

18 pages, 1 figures, comments are welcome!

math.GT

http://creativecommons.org/licenses/by/4.0/

Suppose $(M, gamma)$ is a balanced sutured manifold and $K$ is a knot in $M$. It is known that the rank of the sutured Floer homology of $Mackslash N(K)$ is at least twice the rank of the sutured Floer homology of $M$. This paper studies the properties of $K$ when the equality is achieved for instanton homology. As an application, we show that if $Lsubset S^3$ is a fixed link and $K$ is a knot in the complement of $L$, then the instanton link Floer homology of $Lcup K$ achieves the minimum rank if and only if $K$ is the unknot in $S^3ackslash L$.

2 0 2 4 6 8 10-4-20246 FIG. 1. (Color online) Memory kernels for the two-level sys- tem using the two projection schemes resulting in master equations Eqs. (4)-(5). The initial condition is chosen as ^r(0) =j0ih0j, and the parameters are D=12=13 and e=5=13. Both master the same dynamics for the population s0(t)(inset). For brevity, we use a role like an external force in a Langevin this force arises due to a need to satisfy a +s1(t) =1, which used to be enforced by properties of the memory kernel, K0;n(t) +K1;n(t) =0. While our preceding discussion has been fully general al- beit abstract, let us be more concrete by working with the simple example of a TLS with states j0iandj1inot cou- pled to any bath. The system is described by the Hamilto- nian ^H=e^tz+D^tx, where ^ta(a=x;z) are Pauli matrices. For the first case as outlined above, we use forn=0;1 and the initial condition ^r(0) =j0ih0jto get explicit memory kernels (see Ref. =K1;1(t;P0+P1) the second case, we obtain =2D2cos(2wt); w=q e2+ memory kernels are quite distinct, as shown in Fig. 1, but the population dynamics they generate are identical. It is straightforward to compute ^r(t) the above model Hamiltonian and verify that the following func- tion solves Eqs. (4) and (5) when s0(0) =1: s0(t) (8) Equivalence of the dynamics between Eqs. (4) and (5) should imply relations between the memory kernels in the two equations. It is convenient to work with Fm;m(t;P)such that its time-derivative generates the memory m=0;1. The condition that the functions F(t)with respect to different projectors must generate the same to the =Fm;m(t;P0+P1) identity is easily demonstrated by working with eF, the Laplace transform of F, and applying Dyson's identity, Using Eq. (10) in Laplace space, it is straightforward to show that the s0(t)solving Eq. (4), es0(z) identical to the one solving Eq. (5), see Ref. [4]. We can turn back to the TLS and note that the =F1;1(t;P0+P1) =F1;1(t;P0) satisfy Eq. (10). The structural difference between the master from projections and conservation laws-is not limited to a system of two states. While our point is most striking for a TLS, differences can also appear for general D-level systems whenever the conservation of total population allows for the reduction of one additional dof. It is important to remember that the memory kernels associated with are not independent from each other. and complicated, relations like Eq. (10) show that the memory kernels contain information in common about the underlying unitary dynamics. This from conservation laws, the presence of which is explicitly ignored by the projection operations. Thus, our intuition is that different evolution equations can arise depending whether one imposes constraints or projects out dofs (see Ref. [4]). Thus, there is no guarantee of a quantum master equation to describe the reduced dynamics of an observable. We note that a related observation have been made about Markovian quantum master Material : See supplementary material for 1) a quick derivation of the Nakajima-Zwanzig equation with generic projection superoperators; 2) a general demonstration of how inhomogeneous terms in the appear if conservation laws are explicitly imposed; 3) a generalization of the time-nonlocal relation Eq. (10) forD-level systems; and 4) details on the two-level : This work was supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Materials Sciences and under Contract No. within the Physical Chemistry of Inorganic Nanostructures The codes that support this study are available from upon reasonable request. Author Declarations : Conflicts of Interest: The authors have no conflicts to dis- close. 1S. Nakajima, On Quantum Theory of Transport Phenomena: Steady Diffu- sion, Progress of Theoretical Physics 20, 948-959 (1958). 2R. Zwanzig, Ensemble method in the theory of The Journal of Chemical Physics 33, 1338-1341 (1960). 3We put bath in quotations since we are using it as a shorthand for degrees of freedom we do not care about, rather than it being an infinitely large reservoir that is unchanged by coupling to a small system. 4Details can be found in the Supplementary Material at URL. 5T. Levante and R. Ernst, Homogeneous versus inhomogeneous master equations, Chemical Physics Letters 241, 73-78 Material: Nonuniqueness of generalized master equations for a single observable This supplement is divided into four parts: SM I: We review a general derivation of Nakajima-Zwanzig equations for general choices of projection superoperators using only the assumption of an autonomous, first order linear homogeneous dynamical system. SM II: We show generically how structurally distinct Nakajima-Zwanzig master equations can arise from imposing in the form of conserved charges. SM III: We show how to prove equivalence of dynamics between the different master equations presented in the main text, and generalize the time-nonlocal relation between memory kernels (Eq. (10) of the main text) to D-state systems. SM IV: We give details on the two-level system example stated in the main text, deriving master equations and their kernels from four different approaches to obtaining reduced dynamics. These four approaches differ in how one imposes dynamical constraints and uses the Nakajima-Zwanzig equation for reduced dynamics. I. DERIVATION OF THE NAKAJIMA-ZWANZIG EQUATION In a closed system, time evolution in the Schrodinger picture is governed by the Liouville-von Neumann (1) Laplace transforming both sides (2) where in the last line we used the fact that 1=P+Q. Working in Laplace space is convenient, partly because it circumvents the need to deal with manipulating integrals. For instance, we can directly interpret the product of two as a convolution. We proceed via Dyson's identity in algebraic (3) where Xand(XY)are assumed to be invertible. Applying this twice on Eq. (2) gives ^er(z) in the last line we have applied Dyson's identity to the first term on the RHS of the first line. To arrive back at a we multiply both sides by zand subtract this equation can be partially closed by applying Pon both back into the time domain, and making judicious use of the identities Qf(LQ)Q=Qf(QLQ function fdefined in terms of a power series, we arrive at d dtP^r(t) =i(PLP (5) This is the Nakajima-Zwanzig equation governing the dynamics of a subset P^rof all the degrees of freedom. Note that because this derivation works in Laplace space, the same steps will hold for a general dynamics described by z^er(z)^r(0) +^eq(z) eLeff(z) extends the Nakajima-Zwanzig approach to open quantum systems as well. Interestingly, one is therefore able to recur- sively iterate Nakajima-Zwanzig reduced dynamics, in a similar vein to how one might iterate decimation procedures in group. II. STRUCTURALLY DISTINCT GENERALIZED QUANTUM MASTER EQUATIONS DUE TO a finite dimensional Hilbert space of Dnumber of states, which includes bath degrees of freedom. Suppose there are Nexplicitly identifiable conserved quantities in the dynamics, i.e. there are N1 traceless ^C(k)such that for some initial condition ^r(0), the following holds at all fork=2;:::; N: (7) In addition, (8) Without loss of generality, let us assume that these operators form an orthonormal set with respect to the Hilbert-Schmidt for all k;k0=1;:::; N: (9) This translates into Nlinear constraints for the matrix elements of (10) For simplicity, let us again consider a single distinguished which we will attempt to derive generalized master equations. We will assume that the operator j0ih0j ^IBis from the set of Nconserved operators ^Ck. One step of the Gram-Schmidt procedure allows us to orthogonalize operator j0ih0j ^IBwith respect tof^C(k)gto (12) whereNis a normalization factor. The remaining of the full density matrix ^rcan be the same way, which we will call r1;:::; rD2N1. If we vectorize the full density matrix and turn superoperators into matrices, ^r(t) =) !r(t) (13)3 then the orthonormal set of operators constructed above defines a time-independent rotation matrix RRRsuch that RRR !r(t) (14) Because the basis-change is the Liouville equation (in matrix form) is transformed as d dt !r(t) =iLLL !r(t) =)d dtRRR !r(t) !r(t) (15) iLLL0RRR !r(t): At this point we can project on to the distinguished variable using the (16) which results in the Nakajima-Zwanzig !r(t) !r(0) !r(tt): This turns into a scalar equation after one projects on to the distinguished component on both sides. Because the value of observable is real, the first term on the right hand side must be zero and one is generally left with d dts0(t) (18) where q(t) =1 !r(0) (19) K(t) =1 (20) Observe that the inhomogeneity q(t)in general cannot be set to zero even if we assume that the full density matrix is ^r(0) =^rS ^rB. This is because QQQ0will pick out the conserved quantities qkin Eq. (14), which are generally this with the Nakajima-Zwanzig equation projected on to the j0ih0jcomponent of the system's reduced density matrix, where the inhomogeneity can be set to zero under factorized initial conditions. Thus the presence of inhomogeneous terms in some forms of the generalized master equation follows from a requirement to explicitly satisfy dynamical constraints. This formalizes the intuition we sketched out in the main text.4 III. TIME-NONLOCAL RELATIONS BETWEEN MEMORY KERNELS OF DIFFERENT PROJECTORS AND EQUIVALENCE OF GENERATED DYNAMICS 1. Warm up: Two-level system We will follow the convention of the main text, defining for n=0;1 the (21) and memory (22) Taking an initial state ^s(0) =j0ih0j, we have the following master equations for the two projection memory kernels can be written as a derivative of functions F(t), where Fm;n(t;P) Km;n(t;P) =d dtFm;n(t;P): (24) Note that Fm;n(0;P0) =Fm;n(0;P0+P1) =0. We now work in Laplace space for Fusing the larger projector P0+P1, which has a complementary projector =eFn;n(z;P0) (27) where we have used Dyson's identity in obtaining the second equality. By using the identity F0;n+F1;n=0, we therefore =eFn;n(z;P0) (28) We arrive at Eq. (10) in the main text by transforming back to the time domain. We can cast the right hand side of Eq. (28) solely in terms of eF's for the projector P=P0+P1. Taking n=1, we find eF1;1(z;P0) (29) Inserting this into the above identity for n=0 (30) This relation allows us to directly prove equivalence of the dynamics from both forms of master equations Eq. (23). Note that, in terms of the Ffunctions, the solutions to these equations for the initial condition s0(0) =1 is given in the Laplace domain as es0(z;P0) (32)5 Inserting Eq. (30) into Eq. (31) immediately yields Eq. (32), which implies that the dynamics generated by the two approaches are identical. Note that this follows solely from the assumption that the projectors for the memory kernels are P0andP0+P1. We observe that one can derive a direct relation between the memory kernels of the two projection schemes by using the derivative relation K(t) =PtF(t) =)zeF(z) =eK(z). Doing so in Eq. (30) (33) which in the time domain can be written with the help of an auxiliary function G(t)such that K0;0(t;P0) =K0;0(t;P0+P1) G(0) =1:(34) 2. General case The case where the system dof has Dnumber of states and is coupled to a bath is more involved, but the ideas are the same as those for the two-level system. Therefore, we will only define the necessary notation here, and then give the final result analogous to Eq. (30). We will name the distinguished component of the system's reduced density matrix to be the population of thej0istate. We will define the projectors on to the jmihnjcomponent of the system's reduced density matrix (35) We will take a general projector on to ND2number of matrix elements of the system's reduced density (38) Thus we will project on to some subset of matrix elements of the system's reduced density matrix, taking care to include we shall rewrite the Nakajima-Zwanzig equation (Eq. (5)) using the generalized force =PLeitQLP =)K(t;P) =d dtF(t;P) (39) and integration by parts, assuming the initial condition is factorized such that Q^r(0) =0: d dtP^r(t) =i(PLP (40) In Laplace space, we can formally write the (42)6 We can cast Eq. (42) in a more familiar form by vectorizing operators and turning superoperators into matrices. For {z !s: (46) Following the same steps as in the previous section, the generalization of Eq. (30) is eFFF(z;P00) (47) with the matrix version of the projector (48) Proving the equivalence of the generated dynamics for the j0ih0jcomponent of the system's reduced density matrix is easiest with the following form of the above =h 1eFFF(z;P00) (49) IV. DETAILS ON TWO-LEVEL SYSTEM EXAMPLE In the main text, we showed two methods of deriving master equstions and memory kernels. Here we shall fill in the details of their derivation, as well as to present another approach to derive two-term master equations with entirely different memory kernels. At least for the almost pathologically simple example of the two-level system, this highlights the possibility of nonuniqueness in the memory kernels, even if the structures of the master equations are the same. We start with the (50) If we represent the density operator of the qubit as a (51) the equation of motion for the density matrix (53) Note that we will be working in more generality here than in the main text, which focused only on the diagonal components of the density matrix No projection; reduction via two the unitarity constraint, s00(t) +s11(t) =1, there are three independent dynamical quantities whose equations of motion =2ie(s10s01) =2iD(s10s01) (56) where one immediately finds an additional conserved quantity corresponding to the energy +D(s10(t) +s01(t)): (57) From Eq. (57) we must therefore have s10(t) +s01(t) +s01(0) : (59) At this point there are two ways to reduce the dynamics further, each resulting in a different master equation. 1. Method 1 We can remove the degree of freedom s10(t) +s01(t)exactly in Eq. (54) using the dynamical constraint Eq. (58) to = +E D (61) This can be formally integrated as s10(t)s01(t) = (s10(0)s01(0)) (62) This can be substituted into Eq. (56) to we specialize to the case where s01(0) =s10(0) =0 and use the unitarity condition 1 =s00(t)+s11(t), this can be =tZ 0 4(D2+e2) +4Ee s00(t) + 4(D2+e2) +4Ee s11(t)dt =tZ 0 2(D2+e2) +2Ee s00(t) + 2(D2+e2) +2Ee s11(t)dt: (65)8 Thus using only dynamical constraints, we obtain memory kernels K0;0(t) = 2(D2+e2) +2Ee D=) K1;0(t) =2(D2+e2) +2Ee D=) K1;1(t) = remark that even though these kernels generate the correct dynamics, they are somewhat pathological in that some terms in the master equation grow without bound over time. For the initial condition ^s(0) =j0ih0j=)E=e, we find that these memory kernels indeed satisfy Eq. (34), i.e., d dtG(t) =1 (67) =)G(t) (1=2)D2 (68) K0;0(t;P0) (69) 2. Method 2 Using the same dynamical equations, we can write the energy constraint as s10(t) +E D: (70) We shall isolate only the equations of motion for Eq. (53). Using the forms of the two constraints as written above, we have d dts00(t) (72) We can formally solve for s01(t)in terms of s00(t), which gives s01(t) (73) Thus the equation for s00(t)can be closed to give d dts00(t) =i(eE) (74) Reality of the population means that we can focus only on the real parts, which give the master (75) which implies, k0;0(t) = 2D2cos(2et) =) k1;0(t) =2D2cos(2et) =) k1;0(t) = again, these memory kernels satisfy Eq. (34) with G(t) (77)9 B. Nakajima-Zwanzig master equation 1. Projection on to j0ipopulation We define projection superoperators ( n=0;1) to have (78) such that we isolate only the dynamics of snn(t). This can be used in the Nakajima-Zwanzig (5) We see that the first term of Eq. (5) is (81) where the last line follows from the vanishing of the expectation value of the commutator. We choose the second term of Eq. (5) to be zero by setting Q0^s(0) =0=)^s(0) =j0ih0j. Thus with this projector, the Nakajima-Zwanzig equation Eq. (5) ^r(tt)0E : (82) We consider only the j0ih0jcomponent in this equation (cf. Eq. (5) in the main (84) In order to explicitly compute this memory kernel, we first note that it can be written as the derivative of a simpler =d dtF(t;P0): (86) We proceed by vectorizing the density (87) In this basis, the projection operator Q0and the Liouvillian are respectively represented andiL=0 B@0 iDiD 0 iD2ie 01 CA: (88)

2108.10937

Nathan Ng

Nathan Ng, David T. Limmer, Eran Rabani

Note: Nonuniqueness of generalized quantum master equations for a single observable

J. Chem. Phys. 155, 156101 (2021)

10.1063/5.0068331

quant-ph physics.atom-ph physics.chem-ph

http://creativecommons.org/licenses/by/4.0/

When deriving exact generalized master equations for the evolution of a reduced set of degrees of freedom, one is free to choose what quantities are relevant by specifying projection operators. However, obtaining a does not always need to be achieved through projections--one can also use conservation laws for this purpose. Such an operation should be considered as distinct from any kind of projection; that is, projection onto a single observable yields a different form of master equation compared to that resulting from a projection followed by the application of a constraint. We give a simple example to show this point and give relationships that the different memory kernels must satisfy to yield the same dynamics.

2 Charangan Vasantharajan, Uthayasanker Thayasivam task to purify social media platforms but also conduc- tive to promote positive development in the society[5]. It is impossible to identify such content manually from the large amounts of data being generated every mo- ment on social media, so there is a need to moderation of the content. This has en- couraged a significant number of researchers who are in the domain of Natural Language develop robust computational systems to limit the spread of offensive content using the goal of our current study is to classify offen- sive language from a T amil code-mixed dataset of com- ments and posts collected from Y ouT ube[7]. The com- ments/posts contain multiple types of offensive content in the dataset. F or this purpose, a multi-class clas- sification method is to be adapted to our task. Re- cent studies in multilingual text classification are led by networks like BER T [8] and XLM- RoBER T a [9]. An additional advantage of these archi- tectures is that it yields good results even with the low- resourced languages. This aspect is beneficial to T amil due to not having properly established datasets. T o ac- complish our task, we examined a few transfer- learning models like BER T, DistilBER T, and XLM-RoBER T a as well as a deep learning model CNN-BiLSTM. How- ever, there is a caveat directly using the models on the romanized or code-mixed text: the transformer models are trained on languages in their native script, not in the romanized script in which users prefer to write on- line. W e solve this issue by using a novel way to convert such sentences into their native language while preserv- ing their semantic meaning - selective translation Our important contributions are as fol- lows. - Proposed selective translation and text conversion in romanized and which can be extended to other romanized and code-mixed contexts in any language. - Experimented and analyzed the effectiveness of fine- tuning and the ensemble of mod- els for offensive language identification in code-mixed and romanized scripts. The rest of the article is organized as follows, Sec- tion 2 presents an overview of the articles proposed on Offensive Language Detection. Section 3 entails a de- tailed description of the task and dataset. This is fol- lowed by the explanation of proposed models in Section 4. The experiments and results are explained in Section 5 and Section 6. Section 7 analyzes our results by using a confusion matrix. The paper concludes by comparing and highlighting the main findings with a benchmark.2 Related W ork Offensive content in social media might trigger objec- tionable consequences to its user like mental health problems, and suicide attempts [1]. T o keep the social media ecosystem, coherent researchers and try to develop computational models to identify and classify offensive content and restrain it within a short period. So, there is a lot of research has already been done. In the early stages, computational models were cre- ated by using support vector machine(SVM), Naive Ba- yes, and other traditional machine learning approaches [10]. [11] investigated which can be effect more on pretrained models BER T and Machine Learning Ap- proaches for Offensive Language Identification in En- glish, Danish and Greek languages. They trained the models using Logistic Regression, Multinomial Naive Bayes, Random F orest and Linear Support V ector Ma- chine. BER T model and Logistic Regression performed well in their task. [12] addressed all code-mixed and imbalance distribution related issues and came up with a solution to analyze sentiments for a class data using sampling technique combined with levenshtein distance metrics. F urthermore, this paper compared the performances of different machine learning algorithms namely , Random F orest Regression, XGBoost classifier, Support V ector Machine and Naive Bayes Classifier using F1- Score. [13] explored in Kannada-English by the approaches such as Logistic V ector Machine(SVM), Multinomial Naive Ba- yes, K-Nearest Neighbors(KNN), Decision T rees(DT), and Random F orest (RF). The performance of these models has not reached a state of the art as they could not detect the semantic and contextual relationship in texts. This problem was mitigated with the arrival of word embeddings and Recurrent Neural On the other hand, to encourage sentiment analy- sis tasks and overcome the non-availability of the code- mixed dataset in T amil-English, [2] created an anno- tated corpus containing 15744 comments and posts from Y ouT ube. In addition to this, they examined a bench- mark system with various machine learning algorithms such as Logistic Regression(LR), Support V ector Ma- chine(SVM), K-Nearest Neighbour(K-NN), Decision T r- ee(DT), Random F orest(RF), Multinominal Naive Bay- es, 1DConv-LSTM, and BER T Multilingual. In this task, the Logistic regression, random forest classifiers, and decision trees were provided better results com- pared to other algorithms. Moreover, [15] introduced a methodology to detect Sinhala and English words in code-mixed data by presenting a language detectionTowards Offensive Language Identification for Tamil Code-Mixed YouTube Comments and Posts 3 model with XGB classifier. [16] created a newly code- mixed dataset with sentence and word-level annotation for Sinhala-English language pairs collected from F ace- book comments, chat history , and public posts. According to the social media explosion, [17] per- formed an analysis to separate the general profanity from the hate speech in social media. They presented a supervised classification system that used word n-grams, and word skip grams. Their model was able to achieve 80% accuracy on a huge dataset which contained English tweets annotated with three labels (hate speech (HA TE), offensive language but no hate speech (OFFENSIVE), and no (OK)). [18] addressed the problem of identification by creating an annotated corpus using comments retrieved from T witter Posts for the Greek language. They experimented with seven differ- ent classification models: GRU with Capsule [19], and BER T [8]. In this task, LSTM and GRU with well compared to all other models in terms of macro-f1, because bidirectional LSTMs and GRUs can hold contextual information from both past and future, creating more robust classification systems[20]. How- ever, the fine-tuning approach with the BER Cased model did not provide good results. In recent years, model such as BER T[8], DistilBER T[21], XLM-RoBER T a[9] gained m- ore attention to identify and classify offensive texts through contextual and semantic learning. These models can classify code-mixed texts of dif- ferent languages with astonishing accuracy[2]. [22] ex- amined transformer networks to predict named entities in a corpus written both in native as well as in Ro- man script for English-Hindi language pairs. Based on the above analysis, most of the approaches except the works done by [2] and [18]. They used BER T [8] as their pre-trained model but it did not give state-of-the-art results on code-mixed text. After a deep dive into previous studies and obser- vations about recent developments, we chose models and a deep learning-based model for language experiment as an extension work of [23]. 3 Datasets The T amil-English dataset used in this work was taken from a competition - Offensive Language Dravidian Languages [7] (henceforth ODL). The ob- jective of the ODL task was a multi-class classifica- tion problem with multiple offensive categories (Not- Offensive, Offensive-T Offensive-T Offe- nsive-T Not-T amil). The dataset provided is scrapped entirely from the Y ouT ube posts/comments of a multilingual community where code-mixing is a prevalent phenomenon. This de- rived dataset comprised of all six types of No code-mixing, Inter-sentential Code-Mixi- ng, Only T amil (written in Latin script), morphological level (written in both Latin and T amil script), Intra-sentential mix of English and T amil (writ- ten in Latin script only) and Inter-sentential and mix. T o understand the statistics of mis- spelled and code-mixed words, we compared with an existing pure language vocabulary available in the Dak- shina dataset [24]. W e find the proportion of the (OOV) words (including code-mixed, En- glish, and misspelled words) in the dataset as 85.55%. The dataset consists of comments in six Not-Offensive : Comments which are not rude, or profane. e.g.: I am waiting vera level trailer i ever seen 2. Offensive-T : targeting an Sarkar T easer likes ah beat pana payale Iththu ponavane 3. Offensive-T : offensive com- ments targeting a group. e.g.: Wig mandaiyan fans,vaaya mudikittu irukkanum 4. Offensive-T : offensive com- ments targeting an issue, an organization, or an event other than the previous two Inimae ajith fans evanadhu.... Copy adiku- radha pathi pesuvenga... 5. : offensive comments tar- geting no one. e.g.: Kola gaandula irruke mavane kollama vida maatan 6. Not-T amil : comments not in T amil. e.g.: Hindi me kab release hoga plz comment W e present the dataset statistics in table 1 and vi- sualize in Figure 1. 4 System Description 4.1 dataset described in the section 3, needs some pre- processing to be able to adapt to machine learning al- gorithms. The pre-processing used in this paper are as follows:4 Charangan Vasantharajan, Uthayasanker Thayasivam Label T rain Dev T est 25425 31933190 2343 307315 2557 295288 454 6571 2906 356368Not-Offensive Offensive-T Offensive-T Offensive-T Not-Tamil 1454 172160 Total 35139 43884392 T able 1: Number of instances for each class in train, validation and test sets. The imbalance of the dataset depicts a realistic picture observed on social media 1: Class distribution on T raining set. Dataset is highly imbalanced where a number of instances in Not- Offensive class is much higher compare to other classes. - Firstly , the words in the script are employed Stemming, removing snip words to preserve context to the users' intent, and Lower cas- ing the Romanized words. This step is not per- formed for words written in T amil script as there is no such casing used in T amil script. - Secondly , we remove emojis and emoticons from text, as well as remove all special characters, and user mentions as they generally do not carry any semantic meaning. F urther, since we use transformer models (contex- tual models), it is observed that stop words receive a similar amount of attention as non-stop words, so we did not remove stop words. W e provide some examples for the above techniques below. 1. Inda music ha engaiyoo keta mariyae Inda music ha engaiyoo keta att aaireevkkaattttu pyeel :-D att aaireevkkaattttu pyeel 3. #6Million Views @5Days..!! #200K Came In Quick Baby ..!! #V arlaam #V aralaaam V aa Views Came In Quick Baby V aa4. ThalaivanSTR <3 #Vjs <3 #A V <3 #AS ThalaivanSTR 4.2 Selective translation and is a novel idea we have used to get a proper rep- resentation of text in the native script for the final neural architecture training. The pseudo-code for the proposed algorithm is given in Algorithm 1. The pri- mary need for this step is as follows: Recent advance- ments in state-of-the-art multilingual NLP tasks are led by T ransformer architectures like mBER T and Distilm- BER T which are trained on multiple languages in the native script but not in the romanized script. Hence to reap better results by fine-tuning these text is to be in a native script (for example, T amil text in T amil script). T o convert text into the native script, we cannot rely on neural translation systems, particularly in tweets where users tend to write informally using multiple lan- guages. Also, translating romanized non-English lan- guage words into that language does not make any sense in our context. F or example, translating the wordTowards Offensive Language Identification for Tamil Code-Mixed YouTube Comments and Posts 5 V althukkal(which means congrats in English) directly into T amil would seriously affect the entire in which the word is present because V althu- kkal will be treated as an English word. In many cases, valid translations from English to a non-English lan- guage would not be available for words. So, as a solution to this problem, we propose selec- tive transliteration and translation of the text. In this conversion pipeline, we checked the word in three steps as follows and applied rel event If the word is in English dictionary , then we trans- lated the word into native T amil script 2. W e ignored the word if the detected language is na- tive T amil 3. If the word not processed by the above both cases, then we transliterated the romanized word into na- tive T amil. The segregation of English words and native lan- guage words in each sentence is done using a big cor- pus of English words from NL TK-corpus1. The idea of this selective conversion is based on the observation that in romanized native language comments(like T an- glish), users tend to use English words only when they can convey the meaning better with the English word or when the corresponding native language word is not much used in regular conversations. F or example, the word movie is preferred by T amil users over its corre- sponding T amil word. The translation of words is done using Google T rans- late API2, and transliteration is done with the help of AI4Bharat T ransliteration3. The detection of language script is carried out with a help of NL TK words. In some tweets, the direct translation of intermedi- ate English words present in them can affect the seman- tics negatively , and it may be better to keep them as they are. So, we have experimented with this varia- tion of the STT algorithm - only transliterate the non- English words into their corresponding language and keep the English words as they are. W e observed that the cross-lingual nature of the mod- els can efficiently handle these types of mixed Older me --- engineer aganum miss Me know --- Y edhuku engineering Output: paillly ennnainnn epaarriyaallr aknnnum ec- lvi ennnainnn etriyum etuku epaarriyiyl apadiye kaathukku panchachayum poora raththam STT Output: appttieey kaatukku pnyceecyum ekaattung- kttaa kaatu puuraa rttm 4.3 Models Recent studies show that pre-trained word embeddings and fine-tuning of state-of-the-art T ransformer archi- tectures show better performance in text to classical machine learning approaches like N-gram features with bag of words models [25]. So, we directed our entire focus on using the word em- beddings of transformer architectures both pre-trained and finetuned for text classification along with CNN- BiLSTM. The text obtained using selective translation and transliteration is used in further steps. 4.3.1 CNN-BiLSTM This is a hybrid of bidirectional LSTM and CNN archi- tectures that extracts a new feature vector from the per- character feature vectors for each word by employing a convolution and a max-pooling layer. F or each word, these vectors are concatenated and fed to the BiLSTM network and then to the output layers. with Doc2V ec embedding achieved very high re- sults for sequence classification tasks [26], thus we use GLoV e4. 1 Algorithm for Selective Translation and Transliteration of code mixed and romanized languages Input : romanized or Code -mixed text T and desired native language for the final script L Output Text in native script 1: procedure STT(T,L) This is the STT Initialization: EngWords = Set of all English words words = LOOP Process for i - 0 to len(words ) do word = words[i] if detect (word) = L then continue else if wordinEngWords then words[i] = translate(word, L) else 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: words[i] = L) return joinWords (words ) 6 Charangan Vasantharajan, Uthayasanker Thayasivam 4.3.2 mBER models of BER T (mBER T) are largely ba- sed on the BER T architecture [8] pre-trained on the Wikipedia dump of top 100 languages using the same training strategy that was employed to BER T, i.e., Ma- sked Language Modeling (MLM) and Next (NSP). T o account for the data imbalance due to the size of Wikipedia for a given language, ex- ponentially smoothed weighting of data was performed during pre-training data creation and word piece vocab- ulary creation. This results in high resource languages being under-sampled, while low resourced languages be- ing DistilBER T DistilBER T [21] follows the same architecture of that of BER T [8], while reducing the number of layers by a factor of 2. DistilBER T follows a triple loss lan- guage modeling, which combines cosine distance loss with knowledge distillation for it (student) to learn from the larger pre-trained natural language model (teacher) during pretraining. It has 40% fewer parameters than BER T, runs 60% faster while preserving over 95% of BER T's performances as measured on the GLUE lan- guage understanding XLM-RoBER T a XLM-RoBER T a [9] is a large multi-lingual language model, trained on 2.5TB of cleaned Common Crawled data in 100 languages. It can be recognized as a union of XLM [27] and RoBER T a [28]. The training process involves sampling streams of text from different lan- guages and masking some tokens, such that the model predicts the missing tokens. Using SentencePiece [29] with a unigram language model [30] subword tokeniza- tion is directly applied on raw text data. Since there are no language embeddings used, this allows the model to better deal with code-switching. XLM-RoBER T a man- ifested remarkable performance in various multilingual NLP tasks. 4.3.5 ULMFiT ULMFiT stands for Universal Language Model Fine- tuning for T ext Classification and is a transfer learn- ing technique that involves creating a Language Model that can predict the next word in a sentence, based on unsupervised learning of the WikiT ext 103 ULMFiT model uses multiple LSTM layers, with dropout applied to every layer and developed as the A WD-LSTM architecture [31]. W e prepared our dataset in the format that F astAI6requires it to be in. F as- tAI provides simple functions to create Language Model and Classification data bunch . Creating a data results in pre-processing of text, includ- ing vocabulary formation and tokenization. W e created a language model with the A WD-LSTM model. Then we use transfer learning to update the model to train on a specific dataset that may be quite small. Final stage is the classifier fine-tuning which preserves the and updates the high-level weights using gradual unfreezing of the layers. 5 Experiment Setup All our models are culled from HuggingF library and the models' parameters are as stated in T able 2. The architecture of deep learning and models are depicted in Figure 2 and Figure 3 respectively . Hyperparameter V Function Max Len Batch Rate Loss Function Epochs 10 T able 2: Common parameters for the models that we used during our experiments. 5.1 CNN-BiLSTM In our CNN model(Figure 2), the first layer is the GloV e8 Embedding layer that uses 100-dimension vectors to represent each word. The next layer is the bidirec- tional LSTM layer followed by the 1D Then the output of the Con1D layer is fed into a max-pooling layer, and dropout for regularization, re- spectively . Finally , because this is a classification prob- lem, we use a Dense output layer with six neurons and a SoftMax activation function to make predictions for the six classes in the Offensive Language Identification for Tamil Code-Mixed YouTube Comments and Posts 7 Figure 2: : Proposed hybrid CNN network involves using Convolutional Neural Network (CNN) layers for feature extraction on input data combined with bidirec- tional LSTMs to support sequence prediction. 5.2 mBER is a pre-trained BER T multilingual model having approximately 110M param- eters with 12-layers, 768 hidden-states, and 12-heads9 was used. During the fine-tuning step, layers were integrated into the model and then fed embeddings as the input to it which leads to the increase in the information being fed which results in the improvement of the context and precision. 5.3 DistilBER T DistilBER T is a small, fast, cheap, and light T rans- former model trained by distilling BER T base on of Wikipedia in 104 different languages. W e fine-tuned distilbert base multilingual cased m- odel having 6 layers, 768 dimension, 12 heads, and 134M parameters (compared to 177M parameters for mBER T-base). 5.4 XLM-RoBER T a W e used XLM-RoBER T a-base, a pre-trained multi- lingual model that has been trained on more than 100 languages for our sequence classification. This model has 12 Layers, 768 Hidden, 12 attention heads, and 270M parameters and is trained on 12.2 GB of monolin- gual T amil in addition to 300.8 GB of English corpus[9]. This allows the model for effective cross-lingual trans- fer learning. As we were primarily dealing with native T amil script data, it was effective. 5.5 ULMFiT W e did more epochs after unfreezing all the layers of the pre-trained model. This process will train the whole neural network rather than just the last few layers. T o train the classifier, we used a technique called W e started by training the last few layers, then went backward and unfreeze and trained layers before. This method involves fine-tuning a model (LM) A WD-LSTM to a new dataset in such a manner that it does not forget what it pre- viously learned. A WD-LSTM language model has an embeddings size of 400 and 3 layers which consists of 1150 hidden activations per layer. It also has a batch Figure 3: T ransformer based Models' Architecture. T o recreate this image, we used a source image from [8] .8 Charangan Vasantharajan, Uthayasanker Thayasivam size of 64 and is fine-tuned by adhering to the learning rates by freezing a few of the layers and dropouts with a multiplier of 0.5 were applied. 6 Results and Analysis This section presents the experimental results of all three models explained in Section 5. The results with weighted average are presented in terms of and F1-score of class not-offensive, Offensive- T , Offensive-T Offensive-T Not-T amil in T able 3. A particular model is identified as best if it has reported the highest weighted average of precision, recall, and F1-score. The results on the test set are tabulated in T able 3. The evaluation metric of this task is weighted av- erage F1-score due to account for the class imbalance in the dataset. Among the mentioned models, ULM- FiT and mBER T-BiLSTM produced good weighted avg F1-scores of 0.7346 and 0.7341 on the test dataset, re- spectively . CNN-BiLSTM and XLM-RoBER T a models gave almost similar F1-scores of 0.6112 and 0.6111 re- spectively on the same dataset. W e submitted Distil- BER T model as it has obtained an F1-score of 0.6032, and it is slightly less than the F1-score of CNN-BiLSTM and XLM-RoBER T a. According to T able 3, the ULM- FiT model has marginally outclassed mBER T-BiLSTM and other models. Models like CNN-BiLSTM, DistilBER T, and XLM- R gave similar and poor results on the test dataset. One of the reasons for the poor performance of these models is the imbalance in the distribution of the classes. In the dataset, most of the comments/posts belong to the class of not-offensive while the other classes like not-T have a small classifica- tion of texts. These models performed better on the majority class and poorly on the minority classes. XLM- RoBER T a gave better results on the validation but due to the class imbalances, it could have underperformed on the test set. It is observed that the CNN-BiLSTM model also performed poorly . Models Precision Recall F1-Score 0.7287 0.7682 0.7401 0.7182 0.7579 0.7341 0.5741 0.6754 0.6112 0.5275 0.7263 base DistilmBERT cased 0.5146 0.7169 0.6032 T able 3: W eighted F1-scores according to the models on the test dataset.In the CNN-BiLSTM model, the convolution layer was not capturing the correlations and patterns within the input. Moreover, the BiLSTM layer did not apprehend the dependencies within the attributes extracted by the CNN layer, which has led to the poor performance of the model. F or the word embeddings, we used GloV e embedding which did not perform well on the CNN. Fine-tuning the transformer model Multilingual BER T- BiLSTM has resulted in good performance than Distil- BER T. ULMFiT models attained a better performance in predicting the minority classes as well. The major reasons for the better performance of ULMFiT over other models are due to its superior fine-tuning meth- ods and learning rate scheduler. 7 Error Analysis T able 3 shows that ULMFiT and mBER T-BiLSTM are the best-performing languages models for T amil. T o get more insights, we present a detailed error analysis of these models carried out by using the confusion matrix (Figure 4). In the view of Figure 4, only NF and NT classes have a high true-positive rate (TPR) while in other classes, texts have mostly been misclassified as not- offensive. However, among all six classes, none of the classes got 0% TPR on the other hand, our model iden- tified only 1 OTIO class among 71 texts more than 60% wrongly classified as NF. The high-class imbalance sit- uation might be the reason for this vulnerable outcome. This resulted in the rate of of OTIO class that is more than 98%. Figure 5 reveals that the model is biased towards NF class. The model failed to identify none of the 71 OTIO texts. The insufficient number of instances have been confirmed with ULMFiT and mBER T. Moreover, among 288 OTIG texts, only 34 have been Meanwhile, we observe that only NF, OU, and NT classes have a higher TPR wherein most of the cases, the model misclassified as NF class (Figure 5) as same as in the Figure 4. In contrast, OTIO and OTII classes have shown rates of 100% and more than 99%. In comparison to the models, the deep learning model CNN-BiLSTM shows a slightly dif- ferent confusion matrix (Figure 6). The model gave high TPR for only NF classes while in other classes, texts have mostly been misclassified as same as in the 5, OTIO got 0% TPR, and pre- dicted only 2 of 160 NT classes with more than 98% rate. Overall, a lower number of test datasets for OTIO class mostly led to 100% and we observed a class imbalance problem inTowards Offensive Language Identification for Tamil Code-Mixed YouTube Comments and Posts 9 Figure 4: Confusion matrix of the best model - ULMFiT. 0-Not Offensive(NO), 1-Offensive 2-Offensive T argeted Insult T argeted Insult 4-Not T amil(NT), 5-Offensive T argeted Insult Other(OTIO) the dataset which has a consequential impact on in some instances, the STT (1) may give bad translations. F or instance, when the sentence is en- tirely in English, STT algorithm translates every word, affecting the meaning of the sentence because, in gen- eral, the word-to-word translation of a sentence will not give the correct meaning in the target language. This is a drawback of our STT algorithm. But if we Figure 5: Confusion matrix of the best model - mBER T- BiLSTM. 0-Not Offensive(NO), 1-Offensive 2-Offensive T argeted Insult T argeted Insult 4-Not T amil(NT), 5-Offensive T argeted Insult 6: Confusion matrix of the best model - CNN- BiLSTM. 0-Not Offensive(NO), 1-Offensive 2-Offensive T argeted Insult T argeted Insult 4-Not T amil(NT), 5-Offensive T argeted Insult the real-world problem, most of the corpus text is in code-mixed and romanized format reduces the issues from this drawback. Notably , the ULM- FiT and mBER T-BiLSTM models perform even better than other models (DistilBER T and XLM-RoBER T a) and deep learning-based approach Benchmark Systems W e participated in a shared task on offensive in Dravidian languages which was organized as a Codalab competition10to encouraging research in under resourced Dravidian languages (T amil, Malay- alam, and Kannada). Many participants connecting to all three languages had submitted their solutions. T a- ble 4 shows the overall results and teams which were placed in the top positions in the benchmark for the T amil language. The team hate-alert [32] got first position followed by the team indicnlp@kgp [33] in the second place. As per the F1 scores, out of 31 teams, our ULMFiT and mBER T-BiLSTM models (indicated as Current T ask in the T able 4) stand between OFFLangOne and es teams, lagging the top model only by 0.04 points. This is proven that mBER T along with the BiLSTM reached the state-of-the-art results compared to the Charangan Vasantharajan, Uthayasanker Thayasivam T eam Name Precision Recall F1-Score 0.78 0.78 0.78 0.75 0.79 0.77 0.75 0.77 0.76 0.75 0.79 0.76 0.75 0.78 0.76 0.74 0.77 0.75 0.73 0.78 0.75 0.74 0.77 0.75 0.74 0.77 0.75 0.74 0.77 0.75 0.74 0.75 0.75 0.72 0.76 0.74 0.74 0.75 0.74 0.71 0.76 0.73 0.74 0.73 0.73 0.71 0.74 0.72 0.70 0.73 0.71 0.72 0.77 0.71 0.67 0.71 0.69 0.64 0.62 0.62hate-alert [33] ZYJ123 [34] SJ-AJ [35] NLP@CUET [37] hub [38] MUCS [39] bitions [40] IIITK [41] OFFLangOne [42] Current T ask cs [43] hypers [23] SSNCSE-NLP [44] JUNLP [45] IIITT [47] CUSATNLP [50] 0.54 0.73 0.61 T able 4: Comparison with the benchmark leader board based on F1-score along with other evaluation and Recall). Our model is indicated as Current T ask . systems and can be also observed that ULMFiT has effectively fine-tuned the language model for our task. 9 Conclusion In this work, we have described and analyzed and neural network models to detect the of- fensive language in code-mixed romanized social media text in T amil. A novel technique of selective transla- tion and transliteration is proposed to deal with code- mixed and romanized offensive language T amil. This technique is flexible and can be extended to other languages as well. F or classification, with clas- sical classifiers on top of pre-trained embeddings, we applied and fine-tuned models such as mBER T, DistilBER T, and XLM-RoBER T a, which provides an astonishing rise in accuracy than DL-based methods like CNN-BiLSTM W eighted f1 score increased from 0.6112 to 0.7341. W e experimented with learning techniques to leverage the offensive speech dataset from resource-rich languages. Our work also points to the usefulness of T ransformer architec- tures, particularly mBER T, for low resource languages like T amil. Our proposed models show an average per- formance of 74% F1-weighted score across the dataset. 10 F uture W ork In the above work, we implement STT algorithm to translate and transliterate other languages into T amilnative script, even we could reach a bounded results. It may be due to the lower amount of dataset that is used to fine-tune the model and difficulty in find- ing a T amil word embedding like GloV English has the best word embeddings and pre- trained models, we have planned to modify our STT algorithm to translate and transliterate our dataset to English as one of the future works. Since T amil is a low- resourced language, there are a small number of word embeddings available for it. So, we have decided to cre- ate a word embedding for T amil with consideration of semantic, syntax, and morphology it is decided to deal with imbalanced data in the training and development sets by applying over- sampling to increase the number of samples for It is assumed that the above approaches will increase the model's performance in the e would like to express our thanks to Mr.T emcious F ernando12for their help- ful suggestions to improve and clarify this Offensive Language Identification for Tamil Code-Mixed YouTube Comments and Posts 11 Code A vailability All software used in the above article are available from the public data repository at GitHub of Interest On behalf of all authors, the corresponding author states that there is no conflict of Vigna, F.D., Cimino, A., Dell'Orletta, F., Petrocchi, M., Tesconi, M.: Hate me, hate me not: Hate speech detection on facebook. In: ITASEC (2017) 2. Chakravarthi, B.R., Muralidaran, V., Priyadharshini, R., McCrae, J.P.: Corpus creation for sentiment analysis in code-mixed Tamil-English text. In: Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Computing for Under-Resourced Languages (CCURL), pp. 202-210. European Language Resources France (2020). URL Suryawanshi, S., Chakravarthi, B.R., Arcan, M., Buitelaar, P.: Multimodal meme dataset (MultiOFF) for content in image and text. In: Proceedings of the Second Workshop on Trolling, Aggression and Cyber- bullying, pp. 32-41. European Language Resources Asso- ciation (ELRA), Marseille, France (2020). URL Jose, N., Chakravarthi, B.R., Suryawanshi, S., Sherly, E., McCrae, J.P.: A survey of current datasets for code- switching research. 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS) pp. 136-141 (2020) 5. Chakravarthi, B.R.: HopeEDI: A multilingual hope speech detection dataset for equality, diversity, and inclusion. In: Proceedings of the Third Workshop on of People's Opinions, Personality, and Emotion's in Social Media, pp. 41-53. Association for Barcelona, Spain (Online) (2020). URL Kumar, R., Ojha, A.K., Lahiri, B., Zampieri, M., Malmasi, S., Murdock, V., Kadar, D. (eds.): Proceedings of the Sec- ond Workshop on Trolling, Aggression and Language Resources Association (ELRA), Mar- seille, France (2020). URL Chakravarthi, B.R., Priyadharshini, R., Jose, N., Ku- mar M, A., Mandl, T., Kumaresan, P.K., Ponnusamy, R., R L, H., McCrae, J.P., Sherly, E.: Findings of the shared task on offensive language identification in Tamil, Malay- alam, andKannada. In: Speech and Language Technologies for Dravidian Lan- guages, pp. 133-145. Association for Computational Lin- guistics, Kyiv (2021). URL Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: In: Proceedings of the 2019 Conference ofthe North American Chapter of the Association for Compu- tational Linguistics: Human Language Technologies, Vol- ume 1 (Long and Short Papers), pp. 4171-4186. Asso- ciation for Computational Linguistics, Minneapolis, Min- nesota (2019). DOI URL Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzman, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representa- tion learning at scale. arXiv preprint M., Trieschnigg, D., Ordelman, R., de Jong, F.: Improving cyberbullying detection with user context. pp. pp 693-696 (2013). DOI A., D., T.: SSN_NLP_MLRG at 12: Offensive language identification in English, Dan- ish, Greek using BERT and machine learning approach. In: Proceedings of the Fourteenth Workshop on Semantic Eval- uation, pp. 2161-2170. International Committee for Com- putational Linguistics, Barcelona (online) (2020). Srinivasan, R., C.N, S.: Sentimental analysis from imbal- anced code-mixed data using machine learning approaches. In: Distrib Parallel Databases (2021) (2021). DOI URL Hande, A., Priyadharshini, R., Chakravarthi, and offensive language detection. In: Proceedings of the Third Workshop on Computational Modeling of People's Opinions, Personality, and Emotion's in Social Media, pp. 54-63. Association for Computational Linguis- tics, Barcelona, Spain (Online) (2020). URL Aroyehun, S.T., Gelbukh, A.: Aggression detection in so- cial media: Using deep neural networks, data augmen- tation, and pseudo labeling. In: Proceedings of the First Workshop on Trolling, Aggression and pp. 90-97. Association for Santa Fe, New Mexico, USA (2018). Smith, I., Thayasivam, U.: Language detection in In: Asian Language Processing (IALP), pp. 228-233 (2019). DOI Thayasivam, U., Smith, I.: Sinhala-English code-mixed data analysis: A review on data collection process. In: 2019 19th International Conference on Advances in ICT for Emerging Regions (ICTer), vol. 250, pp. 1-6 (2019). DOI Malmasi, S., Zampieri, M.: Detecting hate speech in social media. In: Proceedings of the International Conference Re- cent Advances in Natural Language Processing, RANLP 2017, pp. 467-472. INCOMA Ltd., Varna, Bulgaria (2017). DOI URL Pitenis, Z., Zampieri, M., Ranasinghe, T.: Offensive lan- guage identification in Greek. In: Proceedings of the 12th Language Resources and Evaluation Conference, pp. 5113- 5119. European Language Resources Association, Mar- seille, France (2020). URL Hettiarachchi, H., Ranasinghe, T.: Emoji powered In: Proceedings of the International Conference Natural Language Processing (RANLP 2019), pp. 474-480. INCOMA Ltd., Varna, Bulgaria (2019).12 Charangan Vasantharajan, Uthayasanker Thayasivam DOI URL Chakravarthi, B.R., Priyadharshini, R., Muralidaran, V., Suryawanshi, S., Jose, N., Sherly, E., McCrae, J.P.: Overview of the track on sentiment analysis for Dravid- ian languages in code-mixed text. In: Forum for Infor- mation Retrieval Evaluation, FIRE 2020, p. 21-24. Asso- ciation for Computing Machinery, New York, NY, USA (2020). DOI URL Sanh, V., Debut, L., Chaumond, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv e-prints arXiv:1910.01108 (2019) 22. Priyadharshini, R., Chakravarthi, B.R., Vegupatti, M., Mc- Crae, J.P.: Named Entity Recognition for code-mixed In- dian corpus using meta embedding. In: 2020 6th Inter- national Conference on Advanced Computing and Com- munication Systems (ICACCS), pp. 68-72 (2020). C., Thayasivam, Offensive lan- guage identification in Dravidian code-mixed YouTube comments and posts. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, pp. 195-202. Association Linguistics, Kyiv (2021). URL Roark, B., Wolf-Sonkin, L., Kirov, C., Mielke, S.J., Johny, C., Demirsahin, I., Hall, K.B.: Processing south asian lan- guages written in the latin script: the dakshina dataset. In: LREC (2020) 25. Zhang, Y., Jin, R., Zhou, Z.H.: Understanding bag-of- words model: A statistical framework. International Jour- nal of Machine Learning and Cybernetics 1, 43-52 (2010). DOI Rhanoui, M., Mikram, M., Yousfi, S., Barzali, S.: A cnn- bilstm model for document-level sentiment analysis. Ma- chine Learning and Knowledge Extraction 1(3), 832-847 (2019). DOI URL Lample, G., Conneau, A.: Cross-lingual language In: NeurIPS (2019) 28. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: RoBERTa: A Robustly Optimized BERT Pretraining Ap- proach. arXiv e-prints arXiv:1907.11692 (2019) 29. Kudo, T., Richardson, J.: SentencePiece: independent subword tokenizer and detokenizer for neural text processing. In: Proceedings of the 2018 Confer- ence on Empirical Methods in Natural Language Process- ing: System Demonstrations, pp. 66-71. Association Linguistics, Brussels, Belgium(2018). URL Kudo, T.: Subword regularization: Improving neural net- work translation models with multiple subword candidates. In: Proceedings of the 56th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pp. 66-75. Association for Computational Lin- guistics, Melbourne, Australia (2018). DOI URL Howard, J., Ruder, S.: Universal Language for Text Classification. arXiv (2018) 32. Saha, D., Paharia, N., Chakraborty, D., Saha, P., Mukher- jee, A.: En- sembling strategies for offensive lan-guage detection. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Lan- guages, pp. 270-276. Association for Computational Lin- guistics, Kyiv (2021). URL Kedia, K., Nandy, A.: Indicnlp@kgp Offensive in Dravidian languages. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, pp. 330-335. Association Linguistics, Kyiv (2021). URL Zhao, Y., Tao, X.: Offensive language identification based on XLM-RoBERTa with DPCNN. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, pp. 216-221. Association Linguistics, Kyiv (2021). URL Jayanthi, S.M., Gupta, A.: Task-adaptive pre-training of multilin- gual BERT models for offensive language In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Lan- guages, pp. 307-312. Association for Computational Lin- guistics, Kyiv (2021). URL NLP- Sharif, O., Hossain, E., Hoque, detection from multilingual code-mixed text using transformers. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, pp. 255-261. Association Linguistics, Kyiv (2021). URL Li, Z.: multimodal transformers for meme Tamil language. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Lan- guages, pp. 352-356. Association for Computational Lin- guistics, Kyiv (2021). URL Huang, B., Bai, Y.: and classify offensive text in multilingual code mix- ing in social media. In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Lan- guages, pp. 203-209. Association for Computational Lin- guistics, Kyiv (2021). URL Balouchzahi, F., B K, A., Shashirekha, offensive language identification. In: Proceedings of the First Workshop on Speech and Language Technolo- gies for Dravidian Languages, pp. 323-329. Association for Computational Linguistics, Kyiv (2021). URL Tula, D., Potluri, P., Ms, S., Doddapaneni, S., Sahu, P., Sukumaran, R., Patwa, P.: Ensemble of multilingual language models with pseudo labeling for offence detection in In: Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Lan- guages, pp. 291-299. Association for Computational Lin- guistics, Kyiv (2021). URL Ghanghor, N., Krishnamurthy, P., Thavareesan, S., Priyadharshini, R., Chakravarthi, B.R.:

2108.10939

Charangan Vasantharajan

Charangan Vasantharajan and Uthayasanker Thayasivam

Towards Offensive Language Identification for Tamil Code-Mixed YouTube Comments and Posts

13 pages

10.1007/s42979-021-00977-y

cs.CL

http://creativecommons.org/licenses/by/4.0/

Offensive Language detection in social media platforms has been an active field of research over the past years. In non-native English spoken media users mostly use a code-mixed form of text in This poses several challenges in the offensive tasks, and considering the low resources available for Tamil, the task becomes much harder. The current study presents extensive multiple deep learning, and transfer learning models to detect on YouTube. We propose a novel and flexible approach of and transliteration techniques to reap better results from fine-tuning and ensembling multilingual transformer networks like BERT, Distil- BERT, and XLM-RoBERTa. The experimental results showed that ULMFiT is the best model for this task. The best performing models were ULMFiT and mBERTBiLSTM for this Tamil code-mix dataset instead of more popular transfer learning models such as Distil- BERT and XLM-RoBERTa and hybrid deep learning models. The proposed model ULMFiT and mBERTBiLSTM yielded good results and are promising for effective offensive speech identification in low-resourced languages.

I. migration of particles in microchannels has caught extensive attention in the last two decades due to its numerous applications in cell sorting, fractionation, filtration, the chief reason for observing the underlying physical phenomena in microfluidic systems [5-9]. The importance of these forces has motivated many researchers to analyze or measure them within the microchannel. Di Carlo et al.have derived the inertial lift on particles and studied the effects of channel Reynolds number and particle size on it; they have shown that by increasing Reynolds, the magnitude of lift coefficient decreases near the wall and increases near the channel center. Also, the particle equilibrium positions shift toward the center as its size increases and its rotational motion is not a key component of the inertial lift [10]. Using lift force profiles, Prohm and Stark have investigated and categorized the particle focusing points and demonstrated that the stable fix points lie on either the diagonal or main axes of the channel cross-section [11]. A fast numerical algorithm combined with machine learning techniques has been proposed to predict the inertial lift distribution acting on solid particles over a wide range of operating parameters in straight microchannels with three types of geometries by specifying the cross-sectional shape, Reynolds number, and particle size there have been attempts to derive analytical relationships for the A simple formula using data fitting and least square was obtained to investi- gate the relationship between the lift and particle size and Reynolds number; according to the proposed criterion, particle focusing does not occur for too small particles or too low Reynolds numbers [13]. Asmolov et al.illustrated that the velocity of finite-size particles near the channel wall is different from that in the undisturbed flow and then reported a generalized expression for the lift force at Re 20 [14]. Another study has proposed a gen- eralized formula for the inertial lift acting on a sphere that consists of 4 terms: lift, slip-shear lift, and correction of the shear-gradient lift; the authors have further confirmed that wall and shear-gradient are the main features of the lift [15]. Moreover, there are examples of works concentrating on the effect of particle shape. For instance, Zastawny et al.presented the great influence of shape both by changing the experienced values of forces and torques and modifying the Reynolds at which the transition 2to unsteady flow happens [16]. Further extension on previous theories and analytical works resulted in an analytical expression capturing the weak, inertial lift on an moving near a wall [17]. Most of the studies on lift forces in the microchannels have focused on solid particles size, etc. Therefore, there are very few examples presenting the whole lift force profiles acting on deformable particles such as droplets and bubbles and studying the effect of their corresponding parameters like Capillary number on the force values. For example, Chen et al.have extensively studied the inertial migration of a deformable droplet in a rectangular microchannel, but their presented lift force profile only considers one value for particle Weber number (a measure for particle deformability) [18]. Rivero-Rodriguez have plotted the inertial lift on a rigid bubble at different Reynolds numbers, and in the pure Capillary regime where the inertial effects are absent, a lift profile is presented for different Capillary numbers [19]. However, their work lacks a similar profile visualizing the total lift force in the most general nonlinear regime. The obtained lift force profiles are mainly the result of either some experimental measure- ments [10, 20, 21] or applying a feedback control in the numerical code to fix the position of particle [22], capsule [23], or drop [18]. Nevertheless, in this paper, we present a method for lift force calculation at different Capillary numbers that solely depends on the trajectory of the drop. In addition, the importance of exploiting an oscillatory flow in the microchannel for working with sub-micron particles [24], having direct control over their focal points and tuning them depending on the flow oscillation frequency [25], and more effective separation and sorting strategies [26] has already been presented in the literature. Thus, we will expand our lift force analysis to include both steady and oscillatory regimes at various where the latter is completely missing in the literature. We will then try to fit analytical expressions to the obtained lift profiles for different cases and present a scheme to predict this expression over a continuous range of input parameters. 3II. METHODOLOGY A single droplet with density and viscosity ratios of one is placed in a laminar flow of 1. simulated using Front-tracking method [27] as elaborated in detail in our previous work [28]. The pressure gradient in the xdirection has a constant magnitude of P0for the steady flow and a varying strength of P0cos(!t)for the oscillatory flow. The periodic is applied in the xdirection, and the no-slip condition is applied on the walls The axes of symmetry have been avoided. Parameters WandUc (maximum velocity of the steady case) are used as the characteristic length and In other words, the period), and !=2 T. Three dimensionless parameters describe the dynamics of the drop: (i) Reynolds number, Re=Uc2W , whereandare the density and viscosity, Capillary number, Ca=Uc , in which is the surface tension, and (iii) the frequency ( !). The drop has a constant size ofa W= 0:3with a spherical initial shape, andRe= in the x,y, andzdirections, respectively, and with 29578 triangular elements for of the drop. FIG. 1: Schematic of the problem setup 4The active dominant forces on the migrating drop in the wall-normal direction are the inertial and lift and lateral drag forces. The direction of the inertial lift at drop locations far from the wall is towards it, and the deformation lift pushes the drop towards the channel center [29]. The direction of the lateral drag is the negative of its migration velocity sign, assuming that the carrier fluid is stationary in the [29]. Therefore, if we assume the positive direction to be the one from the center to the wall, the force balance on the drop according to Newton's second law is the (1) The drag force is computed based on its definition + 0:1Re0:75 rel); (4) WhereCDis the drag coefficient, vris the relative velocity between the drop and the fluid (which is essentially its migration velocity), and Adis the frontal projected area of the drop. Equation 3 is consistent with the findings of [30-34] and those of [35] at a viscosity ratio of one. Although eq. 3 is derived for steady flows, researchers have shown that the drag coefficient in unsteady flows depends heavily on an unsteady parameter that includes the density ratio [36, 37]. Since the density ratio in the present study is one, the parameter becomes zero, and hence, CDfor unsteady flows (including can be approximated as the one for steady flows using this equation. The parameter Adis calculated based on the projected area of the drop on a plane having a normal vector parallel to its migration velocity. Thus, the value of this parameter varies at This procedure leads to a more precise computation of the drag a viscosity ratio of one, the lift for a drop that has a distance higher than its diameter from the closest wall leads to the following compact form [38, = the drop capillary number, Vavgis the average velocity of the carrier fluid across the channel, and dis the distance of the drop from the channel center. The linear dependency of this force with respect to the distance din the specified region is also confirmed in [19]. The lift force analysis in this work is solely based on the drop trajectory. Therefore, to get a lift profile that spans a wide range of d, the drop is released from two different 0:46andz= 1(the upper release) *y= 0:98andz= 1(the lower release) The upper release is chosen such that the whole range of studied dfalls within the validity domain of the deformation force equation (eq. 5). This enables us to plug eq. 5 into the force balance equation (eq. 1) to get the inertial force once the total force is calculated as elaborated below. We will compare the inertial force at different Cavalues for the steady flows in the results section. The lower release initial location is slightly off from the channel center since it is also an equilibrium point, and if a drop is placed there, it does not move at all [18]. The initial zcomponent for both releases is on the main axis for faster These different initial locations do not alter the drop equilibrium position [40-43]. The results of each parameter computation for both releases will be combined to reflect its overall behavior within the channel migration velocity and acceleration of the drop is calculated by taking the first and second temporal derivatives from its trajectory numerically. Since time-step varies through- out the simulations to keep the number at 0.9, the following equa- tions are used to obtain the corresponding derivatives +hi hi1); (7) In anddiandtidenote the distance from center and time at the current step. Both non-uniform finite difference schemes have a the first derivative from the steady flow trajectory at the lowest Ca(Ca= 0:09) results in a very noisy curve that is impossible to interpret. Therefore, we use an regression by minimizing the sum of squared errors to fit the trajectories with analytical expressions, from which we can take first and second derivatives analytically. The trajectory from the upper release is very similar to an exponential decay. Therefore, we fit a curve with the following form to it: d(t)cebt+k; (8) Wherec,b, andkare all constants that should be determined after completing the opti- mization. The constant kis essentially the drop equilibrium distance from the center. The trajectory from the lower release looks like the sigmoid logistic function. Consequently, we use the following equation as its analytical general form: d(t)c 1 (9) Where again, c,b,k, andoffsetare the regressor constants. The regression fits to from the upper and lower release have very high R2scores of 0.99 as plotted in fig. 2. This figure further confirms that the drop focuses at the same dregardless of its initial location. Once the migration acceleration is derived following the aforementioned steps, it will be multiplied by4 3a3, which is the total constant mass of the drop with the initial spherical shape, to get the total force. By subtracting the calculated drag force from the total force, the total lift force can be obtained. III. RESULTS AND DISCUSSION In this section, we report the results of a single deformable droplet simulations in the previously introduced microchannel that contains both steady and oscillatory carrier fluid. As we are interested in studying the effects of oscillation frequency and Capillary number on the lift force, we fix the Reat a value of 10.Caranges between 0:09and1:67, and for 7FIG. 2: Regression fits to both steady trajectories at Ca= 0:09 oscillatory cases, !values are chosen such that for a channel with a cross-section of 100m and water as the working fluid at room temperature, the frequency ranges between 2Hz and 200Hz, which is mostly referred to in the literature [45]. The validation of our as well as grid and domain independence studies, are discussed in detail in our previous work [28]. Figure 3 illustrates the dimensionless mass flow rate over dimensionless time. It can be seen that while the steady regime has the largest constant flow rate in a single direction, the average of oscillatory flow rates in each half of a periodic cycle decreases by increasing the frequency [25, 28]. Although the average of a sinusoidal function in half of a period is constant regardless of its oscillation frequency the lower maximum absolute value of the flow rate at higher frequencies is the chief reason for the 4 visualizes the dimensionless time-dependent frontal projected area of the Adin equation 2) as it migrates toward its lateral equilibrium position traveling both upper and lower-release trajectories. The first thing we note is that in the transient stage before focusing, the drop has a higher average projected area while traveling the upper trajectory (fig. 4b) compared to the one in the lower trajectory (fig. 4a) in each of the flow regimes correspondingly. This is because the drop experiences more shear and deforms easier 8FIG. 3: Flow rate versus time at Ca= 1:67andRe= 10 (a) (b) FIG. 4: Transient cross-stream frontal projected area of the drop when it is released from a (a) lower initial location, and an (d) upper initial location when traveling the upper trajectory. Moreover, in each subfigure, the average of Ais lower at a higher frequency since the average deformation parameter decreases by increasing the frequency [28]. It is important to note that the minimum projected area of the drop in the steady flow is its initial value when the drop is still undeformed and has a spherical shape; in the oscillatory cases, this minimum value occurs when the direction of the flow changes in each periodic cycle. Also, as expected, the drop at higher Cadeforms more and has a higher projected area. This is why Ca= 1:67is used for visualization here among all the other cases in the present study. 9Figure 5 demonstrates the dimensionless total lift coefficient as a function of the dimen- sionless distance of the drop from the channel center in the steady flows and at different Ca. Similar to [18], all of the lift coefficients in this work are obtained by dividing the derived lift force, according to the introduced methodology in the previous section, by a factor in whichVavg;sis the average of flow velocity across the channel cross-section in the steady flow. As expected, we observe that each lift curve has a stable at the corresponding drop focal point. Furthermore, at each Ca, the maximum pos- itive total lift occurs when the drop migration velocity is also maximum. This maximum value is the highest at the lowest Ca. In addition, the maximum negative total lift is at the initial location of the upper trajectory, and its absolute value is the highest for higher Caexcept for Ca= 1:67. This is because as we go further up from the channel center and the drop focal point, the deformation lift becomes the dominant force. According to equation 5, this force is larger at higher Ca. Also, the negative lift sign in this region is due to the direction of the force, which is toward the center. The drop at Ca= 1:67is released from an initial location closer to the center compared to other cases because it has the highest deformability among all. When it was released from the same location as that of the others, it experienced an extremely large deformation that led to its break up. Therefore, the selected initial point for Ca= 1:67is the furthest possible one from the center that results in the largest possible deformation of the drop throughout its upper trajectory without its break up. Consequently, since the drop in this case starts to travel from a closer distance from the center, it has a lower maximum negative total lift compared to Ca= 1andCa= 0:5(please refer to eq. 5 that shows the dependence of the deformation force on the drop distance from the center). Since the principal hypothesis underlying equation 5 is that the wall effect is negligible due to the large distance of the drop from it [29, 46, 47], we can assume that the shear- gradient force is the dominant component of the inertial lift in our study. After subtracting the calculated deformation lift based on this equation from the obtained total lift (fig. 5), we can derive the inertial lift, as shown in fig. 6. We see that the inertial lift as we increase the Canumber. This could make sense as the more deformed shape of the drop can help further increase the difference between the relative velocities of the fluid with respect to the drop on the channel wall and center sides, which is the chief reason for the shear-gradient force existence [29]. According to eq. 5, the deformation force is a linear 10FIG. 5: Total lift coefficient for the steady flows at different Ca function of the drop distance from the center and is larger for higher values of Ca. Because of this trivial conclusion, a plot of this force is not depicted here. FIG. 6: Inertial lift coefficient for the steady flows at different Ca Total lift curves acting on the drop in steady and different oscillatory flows at a few 7. Ineachsubfigure, the larger are both the amplitude of oscillations and the distance between two (e.g. maximum or minimum in the oscillatory cycle) on two consecutive periodic 11(a) (b) (c) (d) (e) FIG. 7: Total lift coefficients for steady and oscillatory flows with different frequencies at (a)Ca= 1:67, (b)Ca= 1, (c)Ca= 0:5, (d)Ca= 0:33, and (e)Ca= 0:25 cycles. Hence, similar to steady regimes, the maximum absolute values of oscillatory occur when the drop migration velocities are maximum as well. Similarly, the lift oscillations around the drop focal point and near the lower initial point are lower because the drop migration velocities are minimum at those locations. 12The moving averages of the total lift coefficients in fig. 7 are plotted in fig. 8. computing the average of lift in each corresponding oscillatory cycle is chosen to be the middle value of din that period. Therefore, since oscillatory cycles with lower frequencies have longer periods, the averaged lift curves at lower frequencies cover a shorter length of d avg(a later beginning and a sooner ending). We first note that the obtained averaged lift curves for the oscillatory flows are not necessarily as smooth as that of the steady lift at the corresponding Ca. This observation becomes more pronounced as we increase !or decrease Ca. Nevertheless, these fluctuations on the curves are to those of the original oscillatory lifts (fig. 7). Additionally, although the average of inertial and lift forces decrease separately by increasing ![28], the difference between them (fig. 8) does not follow the same pattern. This confirms the existence of a drop focal point with an extremum distance from the channel center at an intermediate frequency, as elaborated in our previous work [28]. Despite this, the average of lift is the largest in the steady flow and the smallest in the oscillatory flow with the at each Cain our study. By taking another close look at fig. 7, we realize that each of the oscillatory total can be fitted using an expression that comprises of a base curve, which can be best fitted by a 4th order polynomial, combined with the absolute value of some Both the amplitude of oscillations and oscillatory periods can be controlled by the drop migration velocity and have direct relationship with it. In other words, the proposed expression can have the following +b gd4(t) +hd3(t) +kd2(t) +ld(t) the time-dependent drop distance from the channel center and its migration velocity, respectively, and andqare the constants to be determined while performing the optimization. The constant mis just placed for achieving higher accuracy for the fits, and it ends up to be almost zero compared to other constants in the expression. The resultant curves are plotted in fig. 9 against their corresponding data (fig. 7a) for Ca= 1:67and!= 0:01,!= 0:1,!= 0:5, and!= 1withR2scores of 0.99, 0.99, 0.99, and 0.97, respectively. Similar curves having the same proposed of 0.97 or higher and capable of capturing all the infinitesimal details are 13(a) (b) (c) (d) (e) FIG. 8: Averaged total lift coefficients for steady and oscillatory flows with at (a) Ca= 1:67, (b)Ca= 1, (c)Ca= 0:5, (d)Ca= 0:33, and (e)Ca= 0:25 obtained for other Canumbers as well. To further extend the total lift prediction to more general cases within a continuous range ofCaand!, we consider the steady and averaged oscillatory lifts (fig. 8) for regression with a 4th order polynomial here. In other words, the expression in the first bracket of eq. 14(a) (b) (c) (d) FIG. 9: The fitted total lift coefficients to the ground-truth data using the regression at Ca= 1:67and (a)!= 0:01, (b)!= 0:1, (c)!= 0:5, and (d) != 1 10 is replaced with a value of 1. Using the analytical set of equations 11, we can derive the unknown coefficients a,b,c,g, andhof the polynomial analytically. In these equations, the subscripts for dandfdenote their locations. For instance, the subscript max represents where the magnitude of total (or averaged) lift is maximum and its first derivative is zero. The obtained polynomials with this approach have R2scores of around 0.9 or higher for the 15cases presented in fig. 3bd2 max+ 2cdmax+g= andflastare already available for the cases in fig. 8 to solve the system of equations 11 for them. However, we use the processes (MFGP) method to predict these unknown parameters for any given double inputs of 0:25Ca1:67and 0!1. MFGP is a Bayesian that does a casual inference on a set of high and low-fidelity datasets, and it is extremely effective if there are strong correlations between them [48]. This method is described in detail in our previous work, and it is carried out to predict the distance of the drop equilibrium position from the channel center (deq)withR2of 0.99 and root mean squared error (RMSE) of 0.01 in that work [28]. Here, we refer to the data for the cases in fig. 8 except for Ca= 1as our high-fidelity data. We generate similar data for all the cases in that figure, but with a grid of 1287676in thex,y, andzdirections, having 13038 triangular elements for the discretization of the drop. We consider this data as our low-fidelity dataset. Therefore, we have a total of 25 low and 20 points, which satisfies the required nested structure to apply MFGP on the data [48]. We randomly allocate 5 data points of the entire high-fidelity dataset as our test set since the high-fidelity response is our main target. We train the algorithm on the remaining 40 training data points and evaluate its performance on the test set. We repeat this procedure 30 times and compute the average of evaluation metrics so that the selection of test sets does not significantly affect the overall algorithm I presents the average of R2and RMSE on our 6 remaining unknown parameters after completing the aforementioned steps. We can see that the trained algorithm is capable of predicting the intended parameters with very high accuracies. Especially, the of fmaxandflastis useful for determining the maximum and minimum values of 16averaged total lift for any given input in the range, respectively. The slightly less for dmax(i.e. where the maximum averaged lift occurs) is because of the in its values among different cases. This is unlike the consistent pattern that exists for other parameters for a combination of !values across different Canumbers. The similar lower accurate prediction for dlastis due to the lack of data points between Ca= 1 andCa= 1:67since all the cases with Ca1have the same upper release initial the R2score of around 0.8 for this prediction is still high, and it can help us determine the furthest starting point from the center and the widest traveling region of a droplet with 1Ca1:67in the microchannel so that it undergoes the largest without breaking up. Parameter R2RMSE dmax 0.79 0.0140 dlast 0.78 0.0213 dfirst 0.97 0.0007 flast 0.92 0.0177 fmax 0.99 0.0015 ffirst 0.99 0.0003 TABLE I: MFGP averaged performance metrics on 30 randomly chosen test sets for the parameters required for the determination of the analytical averaged total lift CONCLUSIONS The dynamics of particles and biological cells in microchannels has caught many re- searchers' attention because of several biomicrofluidic applications it has. The owes its behavior mainly to the presence of different lift forces in such channels. Hence, many scientists have dedicated their time to calculate or measure these forces. How- ever, most of these works have focused on analyzing the lift forces acting on solid and non- deformable particles and studied the effects of parameters such as particles' size, Reynolds number, etc on them. Consequently, such analysis on deformable droplets or bubbles and studying the effects of varying parameters like Capillary number is almost missing in the literature. In this work, we have extended such analysis to the case of a single in the channel. We have calculated the main components of the lift force based on a unique methodology that merely depends on the drop trajectory. To do so, first, the 17drop migration velocity and its frontal projected area as it travels its lateral trajectory have been computed to calculate the drag force in the wall-normal direction accurately. After applying Newton's second law on the drop, the total lift profile is obtained over a region where the drop has a distance higher than its diameter from the wall. It has been observed that the total lift has a higher maximum at a lower Capillary, and its minimum decreases as we increase the Ca. The inertial and lift forces both increase by increasing the Canumber. Moreover, since the oscillatory flows within the previously shown to enable working with sub-micron biological particles as well as in- troducing new focal points for them, we have also included these flow regimes in our analysis and investigated the effects of oscillation frequency on the lift in addition to the We have seen that for all cases, the total lift and for oscillatory ones, the ampli- tude of oscillations are both higher when the drop migration velocity is higher. At each Ca, the steady lift and moving averages of oscillatory ones at different !have also been compared. It has been shown that the steady lift has the largest magnitude, and the av- erage of oscillatory one with the highest frequency in this study has the smallest there is not a constant decreasing pattern in the average of lift by increasing the frequency, which is why the drop focuses furthest from the channel center at an intermediate !. Additionally, an accurate mathematical expression has been proposed that captures the detailed total oscillatory lift curves at various !withR2scores of 0.97 or higher. Finally, the multi-fidelity Gaussian processes has been used to accurately predict the 7 required to define a simple 4th order polynomial to fit the steady and lifts with R2scores of about 0.9 or higher for any given Caand!within the ranges of D. R. Gossett, W. M. Weaver, A. J. Mach, S. C. Hur, H. T. K. Tse, W. Lee, H. Amini, and D. Di Carlo, Analytical and bioanalytical chemistry 397, 3249 (2010). [2] M. Toner and D. Irimia, Annu. Rev. Biomed. Eng. 7, 77 (2005). [3] D. R. Gossett, T. Henry, S. A. Lee, Y. Ying, A. G. Lindgren, O. O. Yang, J. Rao, A. T. Clark, and D. Di Carlo, Proceedings of the National Academy of Sciences 109, 7630 (2012). [4] A. Karimi, S. Yazdi, and A. Ardekani, Biomicrofluidics 7, 021501 (2013). 18[5] D. Di Carlo, Lab on a Chip 9, 3038 (2009). [6] J. M. Martel and M. Toner, Annual review of biomedical engineering 16, 371 (2014). [7] S. R. Bazaz, A. Mashhadian, A. Ehsani, S. C. Saha, T. Kruger, and M. E. Warkiani, Lab on a Chip 20, 1023 (2020). [8] S. Connolly, D. Newport, and K. McGourty, Biomicrofluidics 14, 031501 (2020). [9] D. Stoecklein and D. Di Carlo, Analytical chemistry 91, 296 (2018). [10] D. Di Carlo, J. F. Edd, K. J. Humphry, H. A. Stone, and M. Toner, Physical review letters 102, 094503 (2009). [11] C. Prohm and H. Stark, Lab on a Chip 14, 2115 (2014). [12] J. Su, X. Chen, Y. Zhu, and G. Hu, Lab on a Chip (2021). [13] Q. Wang, D. Yuan, and W. Li, Micromachines 8, 197 (2017). [14] E. S. Asmolov, A. L. Dubov, T. V. Nizkaya, J. Harting, and O. I. Vinogradova, Journal of Fluid Mechanics 840, 613 (2018). [15] C. Liu, C. Xue, J. Sun, and G. Hu, Lab on a Chip 16, 884 (2016). [16] M. Zastawny, G. Mallouppas, F. Zhao, and B. Van Wachem, International Journal of Multi- phase Flow 39, 227 (2012). [17] F. O. Mannan and K. Leiderman, Physical Review Fluids 5, 044102 (2020). [18] X. Chen, C. Xue, L. Zhang, G. Hu, X. Jiang, and J. Sun, Physics of Fluids 26, 112003 (2014). [19] J. Rivero-Rodriguez and B. Scheid, Journal of Fluid Mechanics 855, 1242 (2018). [20] C. A. Stan, L. Guglielmini, A. K. Ellerbee, D. Caviezel, H. A. Stone, and G. M. Review E 84, 036302 (2011). [21] J. Zhou and I. Papautsky, Lab on a Chip 13, 1121 (2013). [22] A. H. Raffiee, A. M. Ardekani, and S. Dabiri, Journal of Non-Newtonian Fluid Mechanics 272, 104166 (2019). [23] C. Schaaf and H. Stark, Soft Matter 13, 3544 (2017). [24] B. R. Mutlu, J. F. Edd, and M. Toner, Proceedings of the National Academy of Sciences 115, 7682 (2018). [25] A. Lafzi, A. H. Raffiee, and S. Dabiri, Physical Review E 102, 063110 (2020). [26] P. P. Schonewill, Oscillatory Flow: Effect on Transverse Diffusivity and Inertial Migration of Particles and Bubbles , Ph.D. thesis, University Of Notre Dame (2008). [27] S. O. Unverdi and G. Tryggvason, Journal of computational physics 100, 25 (1992). 19

2108.10947

Ali Lafzi

Ali Lafzi and Sadegh Dabiri

A numerical lift force analysis on the inertial migration of a deformable droplet in steady and oscillatory microchannel flows at different Capillary numbers and oscillation frequencies

physics.flu-dyn

http://creativecommons.org/licenses/by/4.0/

Inertial migration of deformable particles has become appealing in recent years due to its numerous applications in microfluidics and biomedicine. The physics underlying the motion of such particles is contingent upon the presence of lift forces in microchannels. This importance initiated several works to analyze and quantify such forces acting on particles. However, since most of such attempts have focused on solid and non-deformable particles, we extend the lift force analysis for the case of deformable droplets and study the effects of Capillary numbers on their dynamics in this paper. Furthermore, flows as an alternative for steady currents within has been proved to be beneficial by introducing new for the particles. Therefore, the present analysis includes the oscillatory regimes and identifies the effects of oscillation frequencies on lift forces as well. We then propose an expression that mimics the lift force behavior in oscillatory flows accurately. Finally, we introduce a procedure to derive and predict a simple expression for the steady and averaged oscillatory lift for any given combination of Capillary number and oscillation frequency within a continuous range.

complexes ([ 35],[23]), complexes of graph homomorphisms between cycles [ 24], connections to Stiefel manifolds ([ 22], [65]), other notions of box complexes ([ 19], [21], [56], [72]), applications to hypergraphs ([ 1], [43], [51], [57], [71]), as well as connections to statistical physics ([ 13], [31]) and random graphs [ 44]. The 1-skeleton of Hom(T,G)has also been studied in the context of of graphs ([ 27], [58]) and mixing of graph colorings (for the case G=Kn) and problems ([15], [16], [70]). As Kozlov [ 48] has pointed out, there is a notion of a Hom complex in any category where the set of morphisms between two objects can be described as a subset of some collection of 'vertex set' mappings. In this case there is notion of a between two objects, and the set of all s naturally forms a partially ordered set (poset), which in turn gives rise to a topological space. In [ 57] Matsushita considered Hom complexes of ' r-sets', which and other generalizations of graphs. He showed how such complexes can be modeled by simplicial sets and how many results from Hom complexes of graphs, including the of [27], can be extended to this setting. In this work we apply these ideas to the class of directed graphs (or digraphs for short). For two digraphs GandH, we consider the complex- - --Hom(G,H)that parametrizes all directed f:G-H. We will see that many of the categorical properties of such complexes that were satisfied in the undirected setting carry over to this context. Some of our results follow directly from the general theory of homomorphism complexes alluded to above, but in many cases we require new tools and constructions that seem specific to the directed graph setting. At the same time we will see that the- - --Hom complexes of digraphs exhibit behavior that is not seen in the undirected setting, providing insight into the nuances of the category of directed graphs. We consider homomorphism complexes of digraphs to be a natural area of study in its own right, but we see how these concepts also connect to several other areas of existing research. For instance in [11] Braun and Hough study a morphism complex associated to maximal chains in a graded poset. These complexes can be recovered as- - of digraphs by considering the Hasse diagram G(P)of the underlying poset P, and choosing Tto be a directed path. Hence our complex- - --Hom(G,H)for arbitrary digraphs GandHgeneralize this complexes of digraphs also make an appearance in commutative algebra. In [30] the first author and Engstr om showed that minimal resolutions of a class of monomial ideals they call cointerval are supported on complexes that can be described as directed This extends work of Nagel and Reiner [ 62] where a complex of boxes was shown to support resolutions of (squarefree) strongly stable ideals. The idea of using to describe resolutions of monomial ideals was further investigated by Braun, Browder, and Klee in [ 10], where they considered ideals defined by nondegenerate morphisms complexes. It is our hope that a thorough understanding of homomorphism complexes of digraphs may lead to further applications of this kind. In addition, the 1-skeleton of- - a natural model to study regarding homomorphisms of digraphs. In many areas of graph theory, one is interested in the relationships among solutions to a given problem for a specific graph. A defines how one can move from one solution to another, leading to a reconfiguration graph of solutions. This has practical applications when exact counting of solutions is not possible in reasonable time, in which case Markov chain simulation can be used. For these questions diameter, realizability, and algorithmic properties of the configuration graph are typically studied (see [ 61] for a survey). In our context a pair of homomorphisms of be considered adjacent if fandgagree on all but one vertex, defining graph that corresponds to the 1-skeleton of- - --Hom(G,H). For undirected graphs, the connectivity and diameter of this graph are well studied with many results and open questions, but it seems that the analogous questions for digraphs have not been the connectivity and higher topology of- - --Hom(G,H)is also a natural place to consider a notion of homotopy and other categorical properties for directed graphs. For any two digraphs G andH, the 0-cells of- - --Hom(G,H)are given by the homomorphisms G-H, and hence paths in the 1-skeleton of- - a natural notion of homotopy between them. This perspective was investigated for undirected graphs by the first author in [ 27] and with Schultz in [ 32], where the resulting notion was called x-homotopy. In both the undirected and directed setting, homotopy can be described by certain paths in the exponential graph HG, and can also be recovered via a certain 'cylinder' object. For digraphs, there is a subtlety regarding which notion of path in HGone considers, which leads to a hierarchy of homotopies in this setting. These constructions relate to and extend existing theories from the literature (see for instance [3], [41]). 1.1. Our results. We next give a brief overview of our contributions. Our first collection of results involve structural properties of the- - --Hom complexes that parallel those of Hom complexes in the undirected setting, and which will be used throughout the paper. Many of these involve graph operations that induce homotopy equivalences on the relevant complexes. Although we usually think of the- - --Hom complexes as topological spaces (in fact polyhedral complexes), we will typically work with them as posets. In this setting we will often describe our homotopy equivalences in terms of poset maps and a notion of homotopy equivalence of posets, which in turn induce strong homotopy equivalences on the underlying topological spaces. We refer to Section 2.2 for more details. In Proposition 3.2 we prove that the strong homotopy type of- - --Hom complexes are preserved under products in the second coordinate. In Propositions 3.3 and 3.7 we apply results of Matsushita from [ 57] to show that- - --Hom complexes behave well with respect to internal hom adjunction , and also how a notion of directed folding preserves strong homotopy type. Our main results in the setting can be summarized as follows, where we refer to Sections 2 and 3 for any undefined terms. Theorem (Propositions 3.2, 3.3, 3.7) .For digraphs A,B, andC, we have strong homotopy --Hom(A,BxC)- --Hom(A,B)x- --Hom(A,C); (2)- --Hom(AxB,C)- a directed folding, then for any graph Hwe have strong homotopy --Hom(H,G)- --Hom(G{v},H)- --Hom(G,H). As an important consequence of (2) we can recover- - --Hom(G,H)(up to strong homotopy type) as the clique complex of a certain graph G(HG)associated to GandH. See Remark 3.4 for details. Our next results involve notions of out- and in-neighborhood associated to a digraph Gthat mimic the Lov 'asz construction in the undirected setting (see Definition 4.1). In Theorem 4.3 we prove that both complexes are homotopy equivalent to complex- - --Hom(K2,G)(and hence to each other). In Proposition 4.7 we prove that any simplicial complex can be recovered up to isomorphism as--N(G)for some digraph G. This stands in contrast to the undirected setting, where the neighborhood complex must be to a space with a free Z2-action. In the undirected graph setting, the Mycielskian (G)of a graphGhas important applications to the study of homomorphism complexes. As Csorba [ 18] has shown, the neighborhood complex of (G)is homotopy equivalent to S(N(G)), the suspension ofN(G). Motivated by this, we introduce a number of directed versions of the Mycielskian (see Definition 4.8), and in Proposition 4.9 we determine their effect on the homotopy type of the directed neighborhood complex. In particular, we prove that for any digraph Gthere is a homotopy a certain directed Mycielskian of GandS(-)denotes suspension. We next address the effect that directed bipartite Definition 4.10) have on the topology of directed neighborhood complexes. In Proposition 4.11 we prove that if Gis a digraph notcontaining a copy of--Km,n(for anym+n=d) then the a retract onto a complex of dimension at most d-3. Again this mimics an for neighborhood complexes of undirected graphs, first established by Kahle in [44]. Our main result in this section is Theorem 4.12. Here we prove that if Gis any simple digraph on at then ~Hi(--N(G))=0for alli>=n, and that this property holds for any induced subcomplex (such a complex is said to be n-Leray ). We also show that this result is tight in the sense that there exists a digraph with--N(Tm)Sn. Our results for neighborhood complexes of digraphs can be summarized as follows. Theorem (Theorems 4.3, 4.12, Propositions 4.7, 4.9, 4.11) .For any digraph G, we have the neighborhood satisfies the following Any simplicial complex Xcan be realized as--N(G)for some choice of For any digraph G, we have a homotopy subgraphs) If a digraph Gdoes not contain a copy of--Km,n(for anym+n=d), then the a strong deformation retract onto a complex of dimension at most d-3. (4)(Leray property) If Gis a simple digraph on at most 2n+2vertices, next turn our attention to homomorphism complexes of the form- - --Hom(G,Tn), whereTnis a tourmanent (directed complete graph) on nvertices. Homomorphisms G-Tnare used to define a notion of oriented chromatic number o(G)and is a natural place to look for results that compare to Hom complexes of undirected graphs. In particular, we are interested in reconfiguration questions of homomorphisms into tournaments as an analogue of the well-studied question of mixings of (undirected) graph colorings. In this context one is interested in the connectivity and diameter of the complex- - --Hom(G,Tn). 4We mostly study the case that Tn=--Knis an acyclic (ortransitive ) tournament. Our main result in this section is Theorem 5.2, where we prove that for any digraph G, the complex- - empty or collapsible. This again stands in stark contrast to the undirected setting, where even the connectivity of Hom(G,K3)is a subtle question (see Section 5.1 for more discussion). connectivity of- - --Hom(G,--Kn), it is then natural to consider the diameter of its 1-skeleton. In Theorem 5.4 we prove that (- - either empty or has diameter at most |V(G)|. In the special case that G=--Kmis itself a transitive tournament we can say more about the topology and polyhedral structure of- - In Proposition 5.6 we show that such complexes can be recovered as certain mixed subdivisions of a dilated simplex, and are hence homeomorphic ball for any m<=n. Our results for complexes of homomorphisms into transitive tournaments can be summarized as follows. Theorem (Theorems 5.2, 5.4, Proposition 5.6) .Let--Kndenote the transitive tournament on nvertices. Then we have (1)For any digraph G, the complex- --Hom(G,--Kn)is empty or --Hom(G,--Kn)is nonempty, then the diameter of its 1-skeleton then- homeomorphic to a mixed subdivision of the dilated simplex mDn-m. It is an open question to determine the possible homotopy types of- - --Hom(G,Tn)for other choices of tournaments Tn. In Proposition 5.1 we show that any sphere Sncan be recovered up to homotopy type as- - --Hom(K2,Tn)for some choice of tournament Tn. Our last collection of results involves applications of- - --Hom complexes to notions of homotopy for directed graphs. For any two digraphs GandH, one can see that a path in- - to a certain 'bi-directed' path in the directed graph HG, whereHGis the internal hom object associated to the categorical product. This defines a notion of bihomotopy of di, a special case of the strong homotopy ofr-sets as developed by Matsushita [ 57]. Using this we show that foldings of digraphs are intimately related to bihomotopy and show that bihomotopy equivalence of digraphs is characterized by topological properties of the- - --Hom complexes. In Theorem 6.6 we see that a digraph Gis dismantlable if and only if- - for all digraphs T, providing a directed graph analogue of a result of Brightwell and Winkler from [12]. Other notions of paths in HGlead to increasingly weaker notions of homotopy for di. The existence of a directed path inHGdefines a notion of dihomotopy f-g, whereas a path in the underlying undirected graph of HGdefines a line homotopy fg. Dihomotopy is an example of a 'directed homotopy theory', whereas line homotopy is related to constructions of Grigor'yan, Lin, Muranov, and Yau from [41]. Our results can be summarized as follows. Theorem (Propositions 6.3, 6.9, Theorem 6.6) .For any two digraphs GandH, paths in the rise to notions of bidirected homotopy-, directed homotopy-, and line any pair of digraph homomorphism f,g:G-H, we have a strict to a notion of bihomotopy equivalence which can be characterized in terms of the topology of- --Hom bihomotopy type, and each bihomotopy equivalence class contains a unique stiff and only if- --Hom(T,G)is connected for any digraph T. In the undirected setting, an important application of neighborhood and more general homomor- phism complexes was the notion of 'topological lower bounds' on chromatic number. These results came from a general obstruction theory based on the Z2-equivariant topology of the In the directed graph setting we will see how a similar obstruction theory can be achieved, where the most natural example is given by taking the oriented cycle as C3. We refer to Section 7 for more discussion, where we also consider some open questions. 1.2. Paper outline. The rest of the paper is organized as follows. In Section 2 we discuss basic properties of the category of directed graphs and review some relevant tools from poset topology. Here we also provide the precise definition of the complex- - --Hom(G,H)and discuss some examples. In Section 3 we establish structural results regarding the- - --Hom complexes, including product and adjunction formulas. In Section 4 we define the out- and in- neighborhood a digraph G, and establish the homotopy equivalence with- - --Hom(K2,G). Here we establish the other topological properties above, including universality results, the directed analogue of the Mycielski construction, the dependence on bipartite subgraphs, and vanishing of higher homology. In Section 5 we study homomorphism complexes of the form- - --Hom(G,Tn), whereTnis a tournament. Here we prove that- - --Hom(G,--Kn)is empty or contractible for any digraph G, and establish the bound on its diameter. We also discuss the connection to mixed subdivisions for the case of G=--Km. In Section 6 we define the various notions of homotopy of directed graphs and discuss how they relate to topological properties of the- - --Hom and other complexes. Finally in Section 7 we consider open questions and directions for future We thank Nikola Milic 'evi'c, Raphael Steiner, and Hiro Tanaka for We are also grateful to the anonymous referees for their very helpful comments and corrections which greatly improved the paper. The first author is partially supported by Grant # 964659 . The second author is partially supported by the Start-up Research Grant SRG/2022/000314 from SERB, DST, India and by the Research Initiation Grant (RIG), IIT Bhilai, India. 2. D EFINITIONS AND OBJECTS OF STUDY 2.1. The category of directed graphs. We briefly review some basics of directed graphs. It seems these ideas are mostly well-known and scattered throughout the literature, but we collect the main ideas here. For us a directed graph (ordigraph of a vertex set V(G)and an Hence our digraphs have at most one (directed edge) from any vertex to another, and may have loops of the form(v,v). Note that we do allow both Unless otherwise specified the vertex set V(G)will be finite. If Gis a digraph, we let Godenote the subgraph of Ginduced on the set of looped vertices. If (v,w)is an edge in G, we 6will often write v~wand say that ' vis adjacent to w' (note that the order of vandwmatters here). IfGis a digraph and vV(G), we define the out-neighborhood and in-neighborhood any two digraphs GandH, a(directed) graph homomorphism is a vertex set mapping f: V(G)-V(H)that preserves adjacency, so that if (x,y)E(G)we have(f(x), f(y))E(H). We will sometimes refer to a digraph homomorphism as a digraph map. We let- - the set of all digraph homomorphisms from GtoH. IfGandHare digraphs, the productGxHis the directed graph with vertex set with adjacency given by ((g,h),(g',h')) E(GxH)if(g,g') This is indeed the categorical product in this category of digraphs, in the sense that given homomorphisms exists a unique map the natural is the case with usual (undirected) graphs, the category of directed graphs enjoys an internal hom which is adjoint to the product. Given digraphs graph HGis the directed graph with vertex set given by all vertex set mappings f:V(G)-V(H)with adjacency given by(f,g)is a directed edge if whenever (v,v')E(G)we With this definition one can check that for any digraphs G,H, andK, we have a natural bijection of sets :- - - a homomorphism f:GxH-Kis sent to(f):G-KH, Collapses, posets and strong homotopy. We assume the reader is familiar with basic notions of simplicial complexes and homotopy theory, but we review some concepts that we will need. For a simplicial complex Xwe let ~Hi(X)denote itsith integral reduced homology group. For a faceX, the linklkX()is the subcomplex defined will often identify a simplicial complex Xwith its geometric realization and for instance will writeXYto denote that XandYare homotopy equivalent. For any integer nwe simplex. For a finite set S, we letDSdenote simplex whose vertices are the elements of S. SupposeXis a simplicial complex and ,Xare faces such that andis the only maximal face (facet) in Xthat contains . A simplicial collapse ofXis the simplicial complex Y obtained from Xby removing all those simplices ofXsuch that. Hereis called a free face ofand(,)is called a collapsible pair . We denote this collapse by XuY. Observe that ifXuYthen the geometric realizations of XandYare homotopy equivalent. In this article, to mean that Xcollapses onto a subcomplex generated by the we are mostly dealing with (realizations of) partially ordered sets (posets) and poset maps, we recall some relevant constructions specific to this context. First note that if Xis a simplicial complex 7itsface posetP(X)is a partially ordered set given by inclusion. On the other hand, if P=(P,<=) is a poset, we use O(P)to denote the order complex ofP, by definition the simplicial complex on vertex simplices given by all chains pi0< pi1<***< pid. Slightly abusing notation we use|P|to denote the geometric realization of the simplicial complex O(P). Furthermore we will often speak about topological properties of Pitself, by which we mean those |P|. Finally note that if Xis a simplicial complex (or more generally a regular CW complex) then the barycentric subdivision of the complex X. IfPandQare posets, a poset map map )f:P-Qis a function of the underlying sets such that for all x,yP, A poset map f:P-Q induces a simplicial map O(f):O(P)-O(Q)of simplicial complexes and hence a the underlying geometric realizations. For posets PandQ, we define the productPxQto be the poset with elements {(p,q):pP,qQwith posets, we define Poset(P,Q)to be the poset with elements given by all poset maps f:P-Q, with relation every element xP. We say that two poset homotopic if they are in the same connected component of strongly homotopic, then the simplicial complexes O(P)andO(Q)are in the sense of [7], and in particular |P|and|Q|are simple homotopy equivalent. In particular, if allpPthen we say that cis a closure map . In this case c:P-c(P)induces a strong deformation retract of the associated spaces, see [ 8] for more discussion. We will also need the following poset fiber theorem, first established by Babson and Kozlov in [6]. Lemma 2.1. [6, Proposition 3.2] Let:P-Qbe a map of finite posets. If contractible for all qQ, and *for a maximal element, thenis a homotopy Discrete Morse theory. We will need some basics from discrete Morse theory, as developed by Forman in [ 37]. The theory is typically described in terms of operations on a simplicial complex but can equally well be phrased in terms of underlying posets, which is the perspective we use here. For a poset Pwith covering relation , apartial matching onPis by definition a subset SPand an injective function :S-PSsuch that(x)xfor allxS. The elements of P(S(S))that are not matched are called critical . A matching on SPisacyclic if there does not exists a the property that 2.2. [37, Proposition 5.4] LetXbe a polyhedral complex (or any regular CW-complex) and supposeYXis a subcomplex. Then the following are exists a sequence of collapses XuY. *There exists a partial acyclic matching on P(X)with critical cells given by P(Y). 2.4. Homomorphism complexes. We next turn to the main definition of the paper. For this we follow closely the construction of the Hom complex of undirected graphs as studied in [ 6]. Here if GandHare directed graphs we define a to be a map that if(v,w)E(G)we have(v)x(w)E(H). We letDV(H)denote the simplex whose vertex set is V(H), and useC(G,H)to denote the polyhedral complex given by the direct The cells of C(G,H)are given by direct products of 2.3. SupposeGandHare digraphs. Then- --Hom(G,H)is the polyhedral subcomplex of C(G,H) with cells given by all s :V(G)-2V(H){}. An in- --Hom(G,H)if and only if for all (x,y)E(G)we have(u,v)E(H)for anyuxandvy. In particular, the vertex set of- --Hom(G,H)is given by the set- --Hom0(G,H)of all digraph that for digraphs GandHthe set of all s naturally forms a poset P(G,H), allvV(G). The poset P(G,H)can be seen to be the face poset of the regular CW-complex- - --Hom(G,H). As above we let |P(G,H)|denote the of this poset. We then have that |P(G,H)|is the barycentric subdivision of- - --Hom(G,H), so that in particular |P(G,H)|and- - --Hom(G,H)are homeomorphic. In many of our proofs we will think of- - --Hom(G,H)as poset, by which we mean the poset P(G,H)described above. As we mentioned above, Kozlov has pointed out that there is a notion of a for any category of objects where the underlying morphisms are set maps satisfying some condition. Cells are defined as set valued functions where each restriction is a morphism in the category, see [ 48] for more details. A notion of homomorphism complex for ' r-sets' has been studied by Matsushita in [57]. We further discuss this notion in Section 3. IfGandHare directed graphs, we let GandHdenote the underlying undirected graphs. We note that- - since an element - - --Hom(G,H)is in particular a of the underlying undirected graphs. We can also recover any complex graphs via our construction in the 'usual' way of embedding graphs into the category of directed graphs. Namely, given undirected graphs GandHwe construct digraphs ^Gand^Hwhere for each edge in the underlying graph we introduce a directed edge in both directions. One can then see that- - Hence the construction of- - --Hom complexes for digraphs is in particular a generalization of the theory for graphs. We also point that if PandQare graded posets then a strictly order preserving poset map P-Q can be thought of as a homomorphism of digraphs G(P)-G(Q), whereG(P)is the Hasse of as a digraph. Hence our construction of- - --Hom complexes for digraphs generalizes the work of Braun and Hough in [ 11], where the topology of complexes of maximal chains in a graded poset is studied. 2.4.1. Examples. In this section we provide some examples of- - --Hom complexes, some of which will be referred to in later sections. Here we the directed path graph 1-2-***-n, directed cycle graph 1-2-***-n-1, and the transitive n-tournament (see Section 5 for a formal definition), respectively. We have the following easy observations. *- - a disjoint union of s-r+1points ifs>=r, and is empty otherwise. 9*- - a disjoint union of spoints ifsdividesr, and is empty otherwise. *- - a path on nvertices. For the next examples we define graphs C1 3andO6as the digraph on vertex set the digraph given by the 1-skeleton of the octahedron with orientation as in Figure 5. 1,3 1,4 2,4 2,31,2 3,4 - - - 1. Examples of- - --Hom complexes between acyclic tournaments. T5- - --Hom(--C3,T5)- - 2. The tournament T5and two of its- - --Hom complexes. T6- - 3.The tournament T6and its complex of morphisms from the acyclic 7,4,1 7,4,1 1,5,27,5,2 7,4,2 7,5,27,4,2 - - strip FIGURE 4. The tournament T7and its complex of morphisms from the 3-cycle--C3. 2 14 3 - 5.The octahedral graph O6and its complex of morphisms from the 3-cycle--C3. 3. S TRUCTURAL PROPERTIES OF- - --HOM COMPLEXES In this section we collect some structural results regarding directed homomorphism complexes, many of which extend analogous properties for Hom complexes of undirected graphs. We first point out some immediate properties of the- - --Hom complexes, first established in the context of undirected graphs by Babson and Kozlov in [6]. *For any two digraphs GandH, we have that- - --Hom(G,H)is a polyhedral complex (and hence a regular CW complex); *The cells of- - --Hom(G,H)are products of simplices; *The map- - --Hom(G,-)is a covariant, while- - --Hom(-,H)is a contravariant, functor from the category of directed graphs to the category of topological spaces. If f- - --Hom0(A,B)is a graph homomorphism, we let f*:- - --Hom(G,A)-- - --Hom(G,B), f*:- - --Hom(B,H)-- - the induced cellular (poset) maps; *In particular, if a group Gacts on a digraph Gthen for any digraph H,Gacts on the complex- - map induced by composition - - --Hom(G,H)x- - --Hom(H,K)-- - --Hom(G,K) is a topological (poset) map. Recall that for digraphs the digraph whose vertex set and edge set are given by disjoint unions. The homomorphism complex interacts with this operation in a predictable way. Proposition 3.1. For any digraphs A,B,Cwe have an isomorphism of posets - --Hom(AB,C)- --Hom(A,C)x- connected and has at least one edge, we have an isomorphism of posets - --Hom(A,BC)- --Hom(A,B)- The first formula follows from the fact that any by its restriction to AandB. For the second formula, suppose that Then if(x)V(B)then we must have (y)V(B). it follows that xV(A)(x)V(B). # As is the case with Hom complexes of undirected graphs the- - --Hom construction also interacts well with products in the second 3.2. SupposeT,G, andHare digraphs. Then we have a natural strong homotopy equivalence of posets - --Hom(T,GxH)-- --Hom(T,G)x- The proof is more or less identical to the proof of Proposition 3.8 from [ 27], we sketch the argument here. Let P=- - - --Hom(T,G)x- - --Hom(T,H)denote the First define a poset map for any(,)Qand any vertexvV(T). It is clear that iis injective. Next define a poset map c:P-Qas follows. If PandvV(T), we have minimal Definec()=x, these subsets. One can check that these are well-defined poset maps and also that which the result follows. # We will see that- - --Hom complexes also behave well with respect to the internal hom in Section 2, as well as a notion of 'folding' for digraphs. Again this mimics the situation for undirected graphs and, as pointed out by Matsushita in [ 57], applies in more general contexts as well. We will import some results from [ 57] that can applied to digraphs. For this we first review some relevant concepts. For a fixed positive integer r, anr-setis a of a set V(X)and ther-times direct product of V(X). Note that a directed graph is simply an r-set withr=2, whereas an r-uniform hypergraph is an r-set whose relation set R(X)is closed under theSr-action onV(X)r. 12In [57] Matsushita develops a theory of homomorphism complexes of r-sets that specializes to the- - --Hom complexes of digraphs discussed here. He establishes several properties of such complexes, in many cases extending the proofs given for homomoporhism complexes of undirected graphs given by the first author in [27]. In particular, we have the 3.3. [57, Lemma 4.4] SupposeA,B,Care digraphs. Then we have a natural strong of posets - --Hom(AxB,C)-- the graph with a single vertex {v}and a single edge (loop) (v,v). Note that for any digraphGwe haveG=Gx1G1xG, so that from Proposition 3.3 we get a strong - - --Hom(G,H)- - --Hom(1xG,H)- - 3.4. The complex- --Hom(1,HG)can alternatively be described as follows. Form an undirected graph G(HG)on the looped vertices of HGwith adjacency f~gif both directed edges (f,g)and(g, f)exist in the digraphHG. Then- with X(G(HG)), the clique complex of the graph G(HG). Hence we have- 3.5. One can use Proposition 3.3 and Remark 3.4 to prove that the functor- limits, in the sense that if Dis any finite diagram of digraphs with limit limDwe have a - follows from the fact that taking a clique complex of a graph preserves limits, and that one can view G(-)as a functor to the category of reflexive undirected graphs that is right adjoint to inclusion. Details can be found in Proposition 3.7 of [27]. In the case of undirected graphs, the notion of a folding proved useful in analyzing the homotopy type of Hom complexes, for instance in showing that any at least one edge [ 6]. In [ 47] Kozlov provided a simple proof that foldings type in both entries of the Hom complex of undirected graphs. As was established by Matsushita a similar result holds for r-sets and hence can be applied to- - --Hom complexes of digraphs. Let us first set some 3.6. SupposeGis a digraph with vertices vandwsuch Then the vertex vis said to be dismantlable, and the mapping v--wdefines a graph a directed folding. A digraph Gis dismantlable if there exists a sequence of that results in the single looped vertex 1. Results from [57] can then be applied to obtain the 3.7. [57, Theorem 5.6] LetGbe a digraph and suppose f:G-G{v}is a directed folding. Then for any digraph Hwe have natural strong homotopy equivalences of posets f*:- --Hom(H,G)-- --Hom(G{v},H)-- --Hom(G,H). 13We note that Matsushita's proof of this result for r-sets relies on an auxiliary result regarding strong homotopy ofr-sets. We further discuss this notion in Section 6. Example 3.8. As an example of the folding result we refer to Figure 2. Here we see that the vertex 4of Hence the folding 4--3induces a homotopy equivalence of the relevant- --Hom 3.9. In the context of digraphs, foldings exhibit some perhaps unexpected behavior. For a path of arbitrary length (in fact any finite tree) can be folded down to a single edge, so that all paths are in some sense equivalent as far as the Hom complexes are concerned. In the directed case this is no longer the case, as a directed path 1-2-***-nhas no dismantlable vertex. On the other hand we note that of alternating sinks and sources can be folded down to a single directed edge. We discuss applications of this in Section 6. 4. N EIGHBORHOOD COMPLEXES AND OTHER CONSTRUCTIONS In his original study of topological bounds on chromatic number, Lov 'asz [ 52] introduced the neighborhood complex N(G)of an undirected graph G, by definition the simplicial complex on the vertices ofGwhere a face is given by all subsets that have a common neighbor. It was later shown by Babson and Kozlov [ 6] that the neighborhood complex could be recovered as a in the sense that there exists a homotopy equivalence N(G)Hom(K2,G). We will see that a similar result holds in the directed setting. We start with a 4.1. SupposeGis a digraph. The out-neighborhood complex--N(G)is the simplicial complex on vertex with facets given by the out 6. A graph G, along with its out- and in-neighborhood complexes. Note that vertices vV(G)with indeg(v)=0are not vertices of the simplicial are still used to define the 4.2. The class of directed neighborhood complex includes neighborhood complexes of as a special case, in the following sense. Given an undirected graph G, define a digraph--Gwith with a directed edge from and only if vwE(G). Then one can check refer to Figure 6 for an example. Next let P(--N(G))denote the face poset of the complex. We then have the following 4.3. For any digraph G, we have a strong homotopy equivalence - particular,- homotopy the single edge of K2. Define a poset map :- - eachv(1), so that is well-defined. We first show that for each BP(--N(G)), the set-1(B)has a maximal element. For this let BP(--N(G)). we have the maximal element of -1(B). Here, we use the we prove that for each the poset -1(B)- - a maximal element. Let - --Hom(K2,G)such - the maximal element of -1(B)- - The result then follows from Lemma 2.1. # For any digraph G, we can equally well define the in-neighborhood complex--N(G)on vertex with facets given by the in- allvV(G). See Figure 6 for an example. As one might expect from Theorem 4.3 this does not give anything new. Proposition 4.4. For any digraph G, we have a strong homotopy Using similar arguments as in Theorem 4.3, it is easy to see that the map :- - as()=(1)is a strong homotopy equivalence. The result then follows from the composition *, whereis the map from the proof of Theorem 4.3. We also include a proof of homotopy equivalence that uses ideas that go back to Dowker [33]. For the given digraph G, define a bipartite graph BGon vertex set V0V1, whereV0= {(v,0) 0}andV1={(w,1) 0}, and only if vwE(G). Using the notation of [ 8, Theorem 10.9], we have and from this it follows # Remark 4.5. There is yet another way to prove Theorem 4.3 and Proposition 4.4 that we wish to sketch. For this letGandHbe digraphs and suppose IGis an independent set (a collection of vertices with no edges between them). As in Csorba [20]we can define a complex- of all extend to G. Following the proof of [20, Theorem 2.36] one can show that- --HomI(G,H)and- --Hom(G,H)are homotopy equivalent. Now suppose G=K2is a directed edge 1-2. On the one hand if we let I={1}we see that- --HomI(K2,H)is isomorphic as a simplicial complex if we take I={2}we get- the case of undirected graphs, for any tree Tthe neighborhood complex N(T)is to 0-dimensional sphere (two isolated points) since Tfolds down to an edge. In the case of directed trees, we have something 4.6. For any directed tree T, the complex--N(T)is homotopy equivalent to a wedge of 0- dimensional spheres. Proof. We prove this by induction on the number of vertices of T. If|V(T)|=2, then the statement is clear. Suppose Tbe a directed tree on more than 2vertices and let vbe a leaf vertex in If E(T), then a free pair. In the former and in the latter case the onto--N(T{v}). In either case the result follows from induction. again the result follows from induction. # A natural question to ask is what other kinds of topology the directed neighborhood complexes can attain. In the undirected setting Csorba has shown [ 19, Theorem 7] that the homotopy types of neighborhood complexes N(G)of simple graphs (without loops) are precisely those of finite Hence this puts restrictions on the possible homotopy types of N(G)(for have odd Euler characteristic). On the other hand the first author has shown in [ 29] that if one allows loops on the vertices, then for any finite connected graph T(with at least one edge), any homotopy type can be achieved as Hom (T,G)for some choice of G. For the case of digraphs, there is a similar universality for the directed neighborhood complexes, even when one restricts to simple directed 4.7. SupposeXis any finite simplicial complex. Then there exists a digraph G(without loops) such that--N(G)X, an isomorphism of simplicial SupposeXis a finite simplicial complex on vertex set V. For our digraph G, we take its vertex setV(G)to consist of the set Vas well as an additional vertex vFfor each facet FX. For the edges of G, we construct a directed edge from the vertex vFto each vertex vVwherevF. We see that--N(G)is a simplicial complex on vertex set V, with facets given by the facets of X.# What if we restrict to neighborhood complexes of tournaments (oriented complete graphs)? Is it still true that any homotopy type can be achieved as--N(Tn)for some choice of tournament Tn? We address this question in Section 5. 4.1. Mycielskian and other constructions. In the context of undirected graphs, the notion of the Mycielskian of a graph plays an important role in the study of homomorphism complexes. To recall the definitions, let Jndenote the undirected graph on vertex set {0,1,...,n}with well as the loop 00. For a graph Gand integer n>=2, the Mycielskian graph Mn(G)is defined as the quotient allg,g'G. The notation (G)is also sometimes used to indicate Mycielskian construction preserves the property of being triangle-free but increases the chromatic number. In particular, the Gr otzsch graph =M2(C5)is a triangle-free graph with chromatic number 4. By applying the Mycielskian construction repeatedly to a triangle-free starting showed in [ 60] that there exist triangle-free graphs with arbitrarily large [18, Theorem 3.1] Csorba proved that for every (undirected) graph Gand everyr>=1, the homomorphism complex equivalent to S(Hom(K2,G)), the suspension of Hom(K2,G). We next discuss a version of the Mycielskian construction for digraphs, where now several choices can be made regarding 4.8 (Directed Mycielskian) .Define the directed graphs I1,I2, andI3as in Figure 7. For a digraphG, we define 1 2 (A)I10 1 2 (B)I20 1 2 (C)I3 FIGURE 7. The graphs I1,I2, 4.9. SupposeGis a digraph . Then we have a homotopy pt}, and an isomorphism of simplicial a simplex of dimension m. In particular, we In addition we have a homotopy the suspension of the topological space X. Proof. It is easy to observe LetP andQdenote the face posets respectively. Define a poset a maximal element and hence is contractible for any qQ. Further, if a maximal element of the poset f-1(q)P<=p. Therefore, from Lemma 2.1 we get note a simplex on nvertices, we The fact by the following Bipartite graphs and vanishing homology. In the case of undirected graphs, results of Lov 'asz [52] and Kahle [ 44] imply that if Gis a graph not containing a complete bipartite subgraph Ka,b for somea+b=m, then the neighborhood complex N(G)admits a strong deformation retract to a complex of dimension at most m-3(and hence cannot have homology in degree larger than m-3). In particular, if the max degree of Gisd(so thatGis missing a subgraph of the form have homology in degree larger than d-1. We prove here that a similar result holds in the case of out-neighborhobd complex of any digraph. We start with a 4.10. For integers m,n define--Km,nto be the complete bipartite digraph, the directed graph on vertex edge 4.11. SupposeGis a digraph not containing a copy of--Km,nfor somemandn. Then the a strong deformation retract onto a complex of dimension at most m+n-3. Proof. Our argument follows the strategy of Kahle's proof for the undirected case [ 44, Lemma 2.3]. LetPbe the face poset of--N(G)and Clearly,is an order preserving poset map. Also, it is easy to see that for any XPwe Thus, from [ 8, Corollary 10.12] and the discussion below that, we get that induces a strong homotopy equivalence between letr=m+n-3and suppose Xr+2Xr+1***X1is a face ofO((P))of dimension r+1. Fori[r+2], letYi=--NG(Xi). By definition we have Xi=(Zi)for someZiP. that Yr+2Yr+1***Y1. SinceY1, we have that Ymcontains at least mvertices at least If m<=n, thenYnandXnspan a subgraph of G which contains a copy of--Km,n. Otherwise YmandXmspan a subgraph of Gwhich contains a copy of--Km,n, a contradiction to our assumption on G. We conclude that the dimension of O((P))is at mostr. # In the case of undirected graphs, the fact that bipartite subgraphs of a graph Ginfluences the topology of the neighborhood complex N(G)allowed Kahle [ 44] to study the behavior of complexes. Our Proposition 4.11 suggests that one can study--N(G)of random digraphs in a similar way. We note that Proposition 4.11 also of course holds for the in-neighborhood complex--N(G). As an application of these results we obtain a vanishing theorem for homology of In what follows we say that a digraph Gissimple if there is at most one directed edge between any pair of vertices (with loops 4.12. SupposeGis a simple digraph on at most m=2n+2vertices. Then ~Hi(--N(G))=0for all i>=n. Proof. First note that a vertex of outdegree less than n+1. Without loss of generality, let m we add the appropriate number of isolated vertices) and the vertices of outdegree less than n+1. This implies that outdegG(j)>n, so that indegG(j)<=nfor the digraph obtained from Gby removing all out-edges from vertices 1,2...,t . thatDSdenotes a simplex on vertex set S. Proposition 4.4, we know Therefore, (2) letG2be the digraph obtained from G1by removing all in-edges from vertices t+1,t+ 2,...,m . It is again clear indegG(j)<=n, we have indegG1(j)<=nfor hence (3) Equations (2)and (3), to prove our result, it is enough to show that all i>=n.In fact we will show that the a strong deformation retract onto a complex of dimension at most n-1. Observe that, in G2, all the edges are from the set {t+1,...,m}to the set{1,...,t}. Therefore, if t<=nthen the complex--N(G2)is of dimension at most n-1. Ift >n+1, thenG2does not contain a copy of--Kn+1,1and therefore the result follows from Proposition 4.11. Now let t=n+1. not contain a copy a simplex on vertex set [t]. In the first case the result again follows from Proposition 4.11 and the second case is trivial. The result follows. # Recall that a simplicial complex Xisn-Leray if~Hi(Y)=0for all induced subcomplexes YXand i>=n. Equivalently Xisn-Leray if ~Hi(lkX())=0for alli>=nand for any face X. As a corollary to Theorem 4.12 we get the 4.13. SupposeGis a simple digraph on at most m=2n+2vertices. Then the Without loss of generality assume V(G)=[m](once again if|V(G)|<m we add isolated vertices). Let be a face the set of vertices of alli[t]. Define a directed graph only It is easy to observe that Using Theorem 4.12, we get that alli>=nand hence the result follows. # 5. T OURNAMENTS AND letKndenote the complete (undirected) graph with vertex set [n]={1,2,...,n}. Recall that atournament is some choice of orientations on the edges of Kn. In this context we will use Tnto denote an arbitrary tournament. We let--Kndenote the , the directed graph with vertex set[n]and with edges(i,j)for alli-jifi < j . Up to isomorphism this is the unique acyclic tournament on nvertices, and is also sometimes called the transitive study of tournaments has a long history in combinatorics. Many of the important properties of tournaments were first established by Landau [ 49] in his study of dominance relations in flocks of chickens. More recently tournaments have found applications in voting theory [ 50] and social choice theory [ 9]. A primary statistic of a tournament is its outdegree sequence (c1,c2,...,cn), sequence of outdegrees of the vertices of Tn. A well-known theorem of which sequences arise as outdegree sequences of tournaments. In this section we study the connectivity and more general topology of the complexes- - --Hom(G,Tn) for various tournaments Tnand digraphs G. We first consider the case G=K2, where- - --Hom(G,Tn) is recovered up to homotopy type The goal here is to understand of the neighborhood complexes in terms of combinatorial properties of the Tn. First note that, if Tn=--Knis transitive, that neighborhood a simplex, and hence contractible. If Tnis not transitive, it is not so clear what (if any) restriction there might be on the topology of--N(Tn). In Table 1 we have listed all isomorphism types of tournaments on 5 vertices, along with their outdegree sequence and the ranks of the homology of their neighborhood complex. Note that two tournaments can have the same outdegree sequence but have neighborhood complexes with different homotopy types. Also note that from Theorem 4.12 we cannot have nontrivial homology in rank 2 or above. Adjacency (out neighbors) in T5 Outdegree sequence ~Hi(--N(T5)) 1: [], 2: [1], 3: [1, 2], 4: [1, 2, 3], 5: [1, 2, 3, 4] (4, 3, 2, 1, 0) 0: 0, 1: 0, 2: 0, 3: 0 1: [3], 2: [1], 3: [2], 4: [1, 2, 3], 5: [1, 2, 3, 4] (4, 3, 1, 1, 1) 0: 0, 1: 0, 2: 0, 3: 0 1: [], 2: [1, 4], 3: [1, 2], 4: [1, 3], 5: [1, 2, 3, 4] (4, 2, 2, 2, 0) 0: 0, 1: 0, 2: 0, 3: 0 1: [4], 2: [1], 3: [1, 2], 4: [2, 3], 5: [1, 2, 3, 4] (4, 2, 2, 1, 1) 0: 0, 1: 0, 2: 0, 3: 0 2: [], 2: [1], 3: [1, 2, 5], 4: [1, 2, 3], 5: [1, 2, 4] (3, 3, 3, 1, 0) 0: 0, 1: 0, 2: 0 1: [], 2: [1, 4, 5], 3: [1, 2], 4: [1, 3], 5: [1, 3, 4] (3, 3, 2, 2, 0) 0: 0, 1: 0, 2: 0 1: [4], 2: [1], 3: [1, 2], 4: [2, 3, 5], 5: [1, 2, 3] (3, 3, 2, 1, 1) 0:Z, 1: 0, 2: 0 1: [4], 2: [1], 3: [1, 2, 5], 4: [2, 3], 5: [1, 2, 4] (3, 3, 2, 1, 1) 0: 0, 1: 0, 2: 0 1: [4, 5], 2: [1], 3: [1, 2], 4: [2, 3], 5: [2, 3, 4] (3, 2, 2, 2, 1) 0: 0, 1: 0, 2: 0 1: [4], 2: [1, 5], 3: [1, 2], 4: [2, 3], 5: [1, 3, 4] (3, 2, 2, 2, 1) 0: 0, 1: Z, 2: 0 1: [5], 2: [1, 4], 3: [1, 2], 4: [1, 3], 5: [2, 3, 4] (3, 2, 2, 2, 1) 0:Z, 1:ZxZ, 2: 0 1: [4, 5], 2: [1, 5], 3: [1, 2], 4: [2, 3], 5: [3, 4] (2, 2, 2, 2, 2) 0: 0, 1: Z TABLE 1.A list of all tournaments T5on 5 vertices, with their outdegree sequence and homology groups of--N(T5). These computations were done using Sage [64]. 20One can check that the third to last entry T5in Table 1 in fact Our next result says that any n-sphere, for n>=1, can be realized (up to homotopy type) as--N(T2n+3)for some choice of tournament 5.1. Supposen>=1is an integer. Then there exists a tournament onto the boundary of a simplex of dimension n+1. In We will describe the graphs T2n+3explicitly. For this let Forn=1, define the edges of T5according to as Observe onto the boundary of the can be done by collapsing the free we define the edges of T2n+3as ifi=1. Recall that we use <1,...,t> to denote the simplicial complex with facets 1,...,t. Observe that in the tournament T2n+3defined above, we eachi[n]. This implies . Our strategy now will be to collapse each--NT2n+3(i), fori>=n, down to a distinct with vertices among the set [n+2]. a free pair we get . Next we see that a free pair in X1, which implies that . Forn>2 , we now collapse the order of increasing i, using the following sequence of free after we get <[n+1]> . <[n+1]> . Finally (or if n=2), we the following sequence of free we 5.1 demonstrates that there exist tournaments on 2n+3vertices whose out neigh- borhood complex has nonzero homology in dimension n. Note that by Theorem 4.12 this is the minimum number of vertices of a digraph that could achieve this, if m<=2n+2. 5.1. Reconfiguration for tournaments. We next consider the notion of reconfiguration in the context of homomorphisms of digraphs. As discussed in Section 1 the concept of reconfigurations of colorings of undirected graphs has been extensively studied. Here for an n-colorable graph G, then-color graph of G, by definition the graph with node set given by the of G, with two adjacent whenever the corresponding colorings differ on precisely one vertex ofG. One can see that Cn(G)is equivalent to the one-skeleton of Hom(G,Kn). The question of connectivity of Cn(G)(and hence of Hom(G,Kn)) is of interest, as it describes how one can move from one coloring of Gto another by local moves involving changing the color at one vertex. As discussed in Section 1 it also relevant in the context of sampling from colorings of a graph. In [15] Cereceda, van den Heuvel, and Johnson show that if an undirected graph Ghas thenCn(G)is always disconnected. Other the hand, if n>3 there chromatic number nfor whichCn(G)is connected, and also n-chromatic graphs Hfor whichCn(H)is disconnected. In [ 16] the same group of authors characterize the which Hom(G,K3)is connected. They further show that the problem of deciding the connectedness of Hom(G,K3)forGa bipartite graph is coNP-complete, but that restricted to planar bipartite graphs, the question is answerable in polynomial time. We now apply these ideas to digraphs, using the 1-skeleton of the- - --Hom complex to define our notion of reconfiguration. Although the concept makes sense for homomorphisms G-Hof will restrict to the case that H=Tnis a tournament. This is motivated by the notion ofn-colorings of graphs and digraphs. For instance, if Gis a digraph one defines the number o(G)as the minimum nfor which there exists a homomorphism f:G-Tn for some tournament on nvertices. We refer to [ 45] for more details regarding oriented Here we are interested in connectivity properties of- - --Hom(G,Tn)for various choices of tournaments Tn. Our first result says the situation is very well understood for the case that Tn=--Kn is 5.2. SupposeGis a digraph, and let--Kndenote the transitive tournament on nvertices. Then- --Hom(G,--Kn)is either empty or We prove the statement by induction on n. Note that by Proposition 3.1 we can assume that G is connected. The statement is clearly true for n=1and similarly we see that- - --Hom(G,--K2)is either empty, a vertex (in the case that Ghas at least one edge and admits a homomorphism G---K2), or an edge (in the case that Gis an isolated vertex). Not let n>=3and suppose that- - implies that the graph Gis acyclic. Let the set of all vertices of Gwith outdegree 0. SinceGis acyclic,A. Now letSbe the set of all s f- - then it is easy to see that every f- - IfS, we will collapse- - the complex {f- - define a 22matching:S-- - that (f)- - --Hom(G,--Kn)for eachfSand(f)f. We now show that S(with the map defined above) is an acyclic partial matching. For a contradiction, suppose Sis not an acyclic matching and let a minimal cycle. Since we conclude that ft=f1, which is a contradiction to the minimality of the cycle. Hence, from Proposition 2.2 we have - - --Hom(G,--Kn)u- - - for any i[r-1], using similar arguments as above we can show that {f- - eachj[i]}u{f- - - - eachj[r]}- - The result then follows from induction. # Remark 5.3. One may wonder if there exists other acyclic graphs Gwith the property that- or collapsible for every digraph T. We observe that not every acyclic graph Ghas this property, since in the proof of Proposition 4.7 we see that any homotopy type can be realized as- --Hom(K2,G)for some acyclic (and in fact transitive) graph G. Another natural question to ask is whether there exists a graph Gwith the property that- and collapsible for any T. First note that such a graph must contain a looped vertex since in particular it must admit a graph homomorphism from any graph T. We address this question in Section 6. Theorem 5.2 in particular implies that the one-skeleton (- - connected for any G, assuming some homomorphism G---Knexists. Note that a digraph Gis acyclic if and only if G---Kmfor somem. In this case any homomorphism G---Kmcan be 'reconfigured' into any other. In this context one is interested in the diameter of the configuration space, and here we can give a tight bound. Theorem 5.4. SupposeGis a digraph, and let--Kndenote the transitive tournament on nvertices. Then either empty or has diameter at most |V(G)|. Proof. Without loss of generality assume- - --Hom(G,--Kn)is nonempty. From Theorem 5.2 we know that- - --Hom(G,--Kn)is connected, and here we will explicitly construct a path between any two elements of length bounded by |V(G)|. For this let f,g- - thatfg. Define a 5.5. For any digraph Gsuch that- --Hom(G,--Kn)is nonempty, we have the following. (1)h- --Hom0(G,--Kn), and (2)fandgare connected to hin- establish Claim 5.5(1), suppose his not a homomorphism, so that there exists an edge (u,v) E(G)such thath(v) <=h(u). Sincefandgare homomorphisms, we have f(u)< Without loss of generality suppose f(u)<=g(u), so thath(u)=f(u)(by definition of h). f(v), in particular we have that h(v)f(v)and implies that which contradicts the fact that f(u)<=g(u). A similar argument works when h(u)=g(u). Therefore h:G---Knis a homomorphism. We now show that fis connected to hin- - --Hom(G,--Kn). Clearly, g(v)< f(v)for allvA. g(x) otherwise. We claim that the iare edges in the complex- - give a path from vertex fto vertexh. The fact that they are 1-dimensional and give the desired path is clear, and we are left to verify that iis a for each i[t]. For this, we have to show that for We split the proof in four In this case from the fact that f- - somej1,j2[i]. f(ai)and g- - --Hom(G,--Kn), we get that j1< Sincei(v)=f(v), the condition g(u)< f(u), along with the fact thatf,g- - --Hom(G,--Kn), implies that we get that f(u)<=g(u). Thus the result follows from the facts that g- - f(v). Observe that the path from ftohobtained above is of length A similar path of length be constructed from gtoh. Hence we get a path fromftogof This establishes the second part of Theorem 5.4. # We note that Theorem 5.4 is tight since for instance the one-skeleton of- - <n . See Section 5.2 for more details regarding these complexes. Now suppose Gis a connected digraph with oriented chromatic number o(G)=3, so thatG admits a homomorphism to--K3or to the3-cycle--C3. Note that for any digraph G(regardless of chromatic number),- - --Hom(G,--C3)is either empty or has at least three connected components, each of which is a point. To see this note that--C3has no vertex vsatisfying addition, if f1- - --Hom(G,--C3)we obtain two other elements f2andf3by rotating the images of the vertices of G. Hence with Theorem 5.4 this gives a complete picture of mixing of digraphs with directed chromatic number 3. Similarly, if Gis connected and satisfies o(G)=2, then we have a complete understanding of- - --Hom(G,T3)for any choice of T3. Namely, we have that- - and connected (in fact contractible) whereas- - --Hom(G,--C3)is nonempty and has at least 24three components, each of which is a point. With these observations we obtain digraph analogues of the results in [15] and [16]. In general, if Ghas oriented chromatic nforn>=4the connectivity of- - --Hom(G,Tn)will depend on the tournament Tn. Note that for a fixed tournament Tn, the complex- - --Hom(Tn,Tn)is nonempty and has no edges. Hence- - --Hom(Tn,Tn)will be connected if and only if Aut(Tn), the ofTn, is trivial. 5.2. Transitive tournaments and mixed subdivisions. In this section we further study the spe- cial case of homomorphism complexes between transitive tournaments. From Theorem 5.2 we know that- - contractible whenever m<=n, but here we employ some notions from geometric combinatorics to precisely describe its structure as a polyhedral complex. We first review some notions from the theory of mixed subdivisions. Let polytopes which we assume to be The Minokowski sum is defined as cell of the Minkowski sumIk i=1Piis any full-dimensional polytope some (not necessarily full dimensional) polytope whose vertices are a subset of the vertices of Pi. Amixed subdivision ofP=Ik i=1Piis a collectionSof Minkowski cells whose set- theoretical union is P, and whose intersections are given as Minkowski sums, so that if Minkowski cells then CiC' iis a face ofCiand ofC' i. Now letDbdenote the b-dimensional simplex given by where eifor i=1,...,b are the standard basis vectors in Rband e0=0is the origin. If each Pi=Di, then the Minkowski sumIa i=1Piis simply the dilation aDbof a simplex, and any Minkowski cell can be thought of as an ordered set (S1,...,Sa), The study of mixed correspond to triangulations of the product of simplices Da-1xDbunder the Cayley trick) is an active area of research, with connections to flag arrangements [ 2], the geometry of products of minors of a matrix [4], tropical geometry [26], and determinantal ideals [69]. There is a certain mixed suddivision of aDbthat has a particularly nice combinatorial description. For this let Ka,b+1denote the (undirected) complete bipartite graph on vertex sets [a][b+1]. A spanning tree of Ka,b+1then defines a Minkowski cell where each Siis given by the neighbors of the vertex i=1,...,a . Embed the vertices of Ka,b+1in the plane using A spanning tree of if its straight line drawing in this embedding does not have crossings. The collection of all Minkowski cells that correspond to non-crossing spanning trees of Ka,b+1defines a mixed subdivision Ta bof aDbthat we call the staircase subdivision . We refer to [25] for more 5.6. For any integers 2<=m<=n, the complex- homeomorphic to Tm n-m. In particular,|- homeomorphic to a ball of dimension n-m. Proof. The proof of this statement is implicit in [ 30] but we spell out the details here. Recall that the maximal cells in the subdivision Tm n-mcorrespond to noncrossing spanning trees of the where as above Km,n-m+1has vertex set[m][n-m+1]. Each spanning tree T corresponds to a mixed cell given for all i=1,...m the set of vertices adjacent to i, and denote the simplex determined by vertices in S. SinceTis noncrossing, we have that if xSiandySjwithi < j , thenx<=y. Hence, if . . . get an subsets of[n]with the property that if xSiandySjwith i < j thenx < y . These are precisely the maximal cells (corresponding to s) in- - The result follows. # Example 5.7. For examples of the complexes Tm n-mwe refer to Figure 1, where the cases of depicted. Remark 5.8. We note that there exist contractible- --Hom(Tm,Tn)that do not arise from transitive tournaments. For instance, the tournament T4in Figure 8 has the cycle 132 but we see that- --Hom(T2,T4)is - 8. Non-transitive 4-tournament T4such that- - --Hom(T2,T4)is contractible 6. H OMOTOPY OF DIRECTED GRAPHS In this section we use the tools developed above to construct various notions of homotopy for graph homomorphisms of digraphs. Again this runs parallel to the theory for undirected graphs as worked out in [ 27], where a notion of x-homotopy was introduced. The basic idea is to use the 1-skeleton of the relevant homomorphism complex to define a notion of homotopy between graph homomorphisms. In the undirected case we have that the complex Hom(G,H)is to the clique complex on the looped vertices of the graph HG, and hence one can equally well work with(HG)owhen considering homotopy. This idea was further explored by the first author and Schultz in [32], as well as by Chih and Scull in [17]. As we will see, a fundamental difference in the directed setting is that there are a number of natural notions of paths in the exponential graph HG, only one of which corresponds to paths in the corresponding homomorphism complex. In particular, if one considers the digraph structure ofHG, a notion of direction can be incorporated into the theory. In recent years the concept of a directed homotopy theory for discrete structures has found applications in topological data science and the study of concurrency, see for instance [ 36]. The category of digraphs provides a context where a notion of directed homotopy naturally arises, and seems to be a useful testing ground for such ideas. To motivate our definitions we recall some analogous constructions for undirected graphs discussed in [ 27]. Here two graph homomorphisms if there exists a path fromftogin the (undirected) homomorphism complex Hom(G,H), or equivalently a path 26along looped vertices in the graph HG. In [ 27] this was seen to be equivalent to existence of a graph homomorphism (or homotopy )F:Gx- -In-Hwith the property that Here- -Inis a looped path of length n(which we will also think of as a digraph, see below). Thexnotation is used to emphasize that the underlying product involved in the internal homHGis the categorical product. This distinguishes the construction from other approaches (for instance [3]) where the cartesian product was used. We next explore these constructions for digraphs. In this context there are a number of ways one may wish to define the notion of homotopy, based on paths in various parametrization spaces and different notions of the interval object Infor digraphs. For our first notion of homotopy we consider (usual, topological) paths in the homomorphism 6.1. homomorphisms of digraphs. We will say that denoted f-g, if there exists a path from ftogin the complex- --Hom(G,H). In Remark 3.4 we discussed the homotopy equivalence- - Recall that the latter complex is the clique complex of the (undirected) graph G(HG)), whose vertices are the looped vertices of the digraph HG, and where f~gif both edges(f,g)and(g, f)exist. For any non-negative integer n, we define the graph- -Into be the digraph with vertex set that- -Inis a bidirected path with loops on every vertex, see Figure 9). We then have the following 6.2. homomorphisms of digraphs. Then fandgare bihomotopic if and only if there exists a non-negative integer nand a graph homomorphism F:Gx- -In-Hsuch notion of bihomotopy leads to an equivalence relation on the set of digraphs in the usual way. Namely, a homomorphism of digraphs f:G-His abihomotopy equivalence if there exists a homomorphism g:H-Gsuch In [ 57] Matushita considers a notion of strong homotopy ofr-sets, extending the notion of x-homotopy of undirected graphs developed by the first author in [ 27]. Recall that a 2-set recovers the notion of a directed graph, and one can see that strong homotopy equivalence for 2-sets is equivalent to our notion of bi. From [57] we then get the following characterization of bihomotopy 6.3. [57, Corollary 5.2] Supposef:G-His a homomorphism of digraphs. Then the following are homomorphism fis a bihomotopy any digraph T, the poset map f*:- --Hom(T,G)-- --Hom(T,H)is a strong homotopy any digraph S, the poset map f*:- --Hom(H,S)-- --Hom(G,S)is a strong homotopy equivalence. For the case ofx-homotopy of undirected graphs, foldings were essential in defining elements of a x-homotopy class. We will see that something similar holds for bihomotopy of digraphs. From Proposition 3.7 we have the following. Lemma 6.4. SupposeGis a digraph and vGis dismantlable. Then the inclusion G{v}-Gis a bihomotopy equivalence. 27A digraph is said to be stiffif does not contain any dismantlable vertex. As a corollary to the above results we get the following, see also Corollary 5.10 from [ 55] for the analogous 6.5. SupposeGandHare both stiff digraphs. Then a homomorphism f:G-His a bi if and only if it is an isomorphism. As a consequence of Corollary 6.5 each bihomotopy class [G]contains a unique (up to iso- morphism) stiff representative. This representative is a single looped vertex 1exactly when and in this case we have that- - --Hom(T,G)- - --Hom(T,1)is collapsible for any graph T. In [ 12] Brightwell and Winkler prove that an undirected graph Gisdismantlable if and only if Hom(T,G)is nonempty and connected for any graph T. We have something similar in the 6.6. SupposeGis a digraph. Then the following are dismantlable (in the directed sense of Definition 3.6); (2)The stiff representative of [G]is a single looped vertex 1; (3)The complex- --Hom(T,G)is nonempty and contractible for any graph T; (4)The complex- --Hom(T,G)is nonempty and connected for any graph T. Proof. The from Lemma 6.4, (2)=(3)follows from Proposition 3.7, and(3)=(4)is immediate. We prove (4)=(1)by induction on n=|V(G)|. Note that Gmust have a looped vertex since in particular- - --Hom(1,G)is nonempty. If G=1we are done, and otherwise we have that the constant homomorphism c:G-1is distinct from the identity i:G-G. By assumption we have that- - --Hom(G,G)is connected and hence there exists a path fromctoiin the complex- - --Hom(G,G). But note that if j:G-Gis any homomorphism that is adjacent to iin- - --Hom(G,G)we have some vertex vGwherej(v)vand where j(w)=wfor all wv. HenceGadmits a folding G-G{v}. By our assumption and Proposition 3.7 we have that- - --Hom(T,G{v})is connected for any T, and by induction we conclude that G{v}is result follows. # 6.1. Other notions of homotopy. We next consider other constructions of homotopy of digraphs, based on different notions of paths in the exponential digraph HG. For the first we take advantage of the fact that HGis itself a directed graph. Recall that for a digraph Twe useToto denote the subgraph of Tinduced on the looped vertices. We note that the vertices of (HG)ocorrespond to the graph homomorphisms G-H. Definition 6.7. homomorphisms of digraphs. We say that denoted f-g, if there exists a directed path from ftogin the graph(HG)o. We note that-is reflexive and transitive, but does not define an equivalence relation since it is not symmetric. Also note that if we let--Indenote the directed path on n+1looped vertices, then f-gif and only if there exists an integer n>=1and a graph homomorphism has some perhaps unexpected properties. For example, if fi, fj:---Inare the inclusions of of a looped vertex 1given by 1--i,1--j, thenfi-fjif and only if i<=j. Similarly, it is not entirely clear how one should define 'dihomotopy equivalence' between two digraphs. For our third (and final) notion of homotopy, we allow anypath in the underlying 6.8. homomorphisms of digraphs. Then fandgare if there exists a path from ftogin the underlying undirected graph of (HG)o. The existence of a homotopy f-gis again equivalent to a digraph map F:GxIn-Hwith the desired projections. In this case Inis a path on n+1looped each edge appears in one of either direction (so that for all iwe have exactly one of an edge ). This approach to intervals is also taken in [ 41], where the resulting notion of homotopy is based on the Cartesian product of digraphs. In [ 41] the graph Inis called a line digraph and the set of all line digraphs on n+1vertices is denoted by In. We refer to figure 9 for an illustration of the various notions of interval graphs discussed so far. The homotopy theory of digraphs developed in [ 41] was inspired by a homology theory of by the authors in previous papers. In [ 41] it was shown that these theories interact in the expected ways, including invariance of homology under homotopy equivalence. In addition, the theory extends the A-homotopy of (undirected) graphs studied in [ 3] to the setting of digraphs. In a similar way the line-homotopy theory of digraphs described here can be seen as a thex-homotopy of [ 27] to the setting of digraphs. We do not know if there is a theory in our context, see Section 6.2 for more discussion. 0 1 2 3 (A)- -I30 1 2 3 (B)--I30 1 2 3 (C) An element ofI3 FIGURE 9. The various interval objects involved in homotopy of di 6.9. homomorphisms of digraphs. Then we these implications are strict. Proof. We have seen that f-gif and only if the there exists a path along bidirected edges in the graph(HG)o. The definitions of-and-similarly involve increasingly larger classes of paths in (HG)o, and hence the implications hold. To see that the implications are in fact strict, we refer to in Figure 10. In this case the complex- - of 6 isolated points and hence each graph homomorphism G-Hsits in its own bihomotopy equivalence class. However the directed edges in the graph HGleads to other notions of homotopies. For not dihomotopic to bihomotopic to g. Here we let G=(a,b)and usef=(x,y)to denote the images of (a,b)in a homomorphism f:G-H. # 29a b (A) The graph G20 1 3 4 5 (B) The graph H(0,1) (2,3) (1,0)(3,2) (4,5) (5,4) (C)- - (1,0)(3,2) (4,5) (5,4) (D) The 10. An illustration of the implications of our various homotopies. 6.2. Higher homotopies. We have seen that bihomotopies of digraph homomorphisms correspond to paths in the one skeleton of the complex- - --Hom(G,H). Hence we can take advantage of the topology of- - --Hom(G,H)to define higher categorical structures. In particular, our results from Section 3 implies that the category of directed graphs is enriched over the category of posets (and hence topological spaces). For the case of undirected graphs, this is further studied in [ 28] and [ 32]. It would be interesting to construct a topological space that extends paths for our other notions of of homotopy. In the case of line homotopy, a natural candidate is provided by the directed clique complex of a directed graph. This notion has recently found applications in network theory, see for instance [ 39] and [ 54]. To recall the definition suppose Gis a digraph. Then the directed the simplicial complex on vertex set given by all1<=i < j<=n}. Note that a directed n-clique is simply a subgraph of Gthat is isomorphic to--Kn. Note that the edges in--X((HG)o)are precisely the (directed) edges in (HG)oso a path to a (not necessarily) directed path in (HG)o. The notion of line-homotopy equivalence between homomorphisms of digraphs naturally leads to a definition of line-homotopy equivalence of digraphs G-H, and it is an open question how this relates to the topology of the directed clique complex. For instance, if G-Hare line digraphs, is it true homotopy equivalent for any digraphT? More generally, do the results from Proposition 6.3 hold in this setting? If so this would lead to a homology theory for digraphs that is invariant under line homotopy 6.10. Thex-homotopy theory for undirected graphs developed in [28]has close connections to the notion of strong homotopy type as introduced by Barmak and Minian in [7]. Here one can see that the strong collapse of a simplicial complex from [7], when restricted to a clique complex X(G)of a to foldings ofG. In[28]it is shown that foldings of a graph Gin generate thex-homotopy type of a graph and in particular preserve the homotopy type of Hom(-,G). It would be interesting to see how the notion of directed homotopy and directed foldings can be generalized to a setting of directed . 7. F URTHER THOUGHTS In this section we discuss some areas of future research and collect some open Topology of complexes into tournaments. In Section 5 we saw that if--Knis the on nvertices then for any digraph Gthe complex- - --Hom(G,--Kn)is empty or contractible. On the other hand we saw in Lemma 5.1 that for any nthere exists a tournament such that--N(T2n+3)is homotopy equivalent to an n-sphere Sn. What are the other homotopy types that can be achieved? For instance, we do not know of any tournament Tnfor which the neighborhood torsion in its homology. In a related question, does the something about the combinatorics of the tournament? In the context of undirected graphs, the maximum degree of a graph Ginfluenced the In particular, VCuk 'ic and Kozlov [ 23] proved that if Ghas max degree at A stronger result in terms of degeneracy by Malen in [ 53]. We note that something similar cannot hold for digraphs, at least for arbitrary tournaments. In particular, it is not hard to see that for any nthere exists a tournament Tn for which--N(Tn)- - --Hom(K2,Tn)is Obstructions to graph homomorphisms. In the context of undirected graphs, the topology of the neighborhood (and more general homomorphism) complexes of a graph Gleads to lower bounds on the chromatic number (G). In particular, by results of Lov 'asz [ 52] and we have for any graph Gthe following an odd cyle on 2r+1vertices, and for a topological space X,conn(X)is the connectivity ofX. These results follow from an obstruction theory for graph homomorphisms that utilizes the Z2-equivariant topology of the relevant Hom complexes. In this context the free group action on the topological spaces is induced by a Z2-action on the graphs K2andC2r+1. By takingG to beK2(resp.C2r+1) one can see that the bounds in Equation 4 are tight. In the language of Hom complexes we say that graphs (see [32] and [46]). It is a natural question to ask whether there is a similar obstruction theory coming from the homomorphism complexes of digraphs. In this context there is no nontrivial action on the new tools seem to be required. However, if we view the edge K2as a directed 2- cycleC2, the next natural nontrivial group action arises if we consider the oriented cycle--C3as a potential test graph. We then have a free Z3group action on--C3which induces a free Z3-action on the complex- - --Hom(--C3,G)for the case of loop-free G. Since- - --Hom(--C3,-)is functorial, we can then employ Dold's Theorem (see for instance [55]). Theorem 7.1 (Dold) .LetGbe a non-trivial finite group. Suppose where the action ofGonYis free and dimY<=n. Then ifXisn-connected there cannot exist a continuous G-equivariant of the dimension of- - --Hom(--C3,H)for any digraph Hthen leads to an for homomorphisms G-H. For example, in Figure 2 we have a tournament T5with the property that- - --Hom(--C3,T5)is 1-dimensional. Hence if Gis any digraph with - there cannot exist a homomorphism G-T5. Similarly for T7in Figure 4 we have that if- - there cannot exist a homomorphism G-T7. 317.3. Other directions. There are a number of other properties and applications of- - --Hom complexes that one could explore. Some of these have been mentioned in previous sections but we collect them here. *Other graph classes - In addition to tournaments it would be interesting to investigate the neighborhood complexes of other classes of digraphs. Do the combinatorial properties of these digraphs play a role in the topology (and diameter) of the resulting candidates include Eulerian digraphs, digraphs of bounded width, etc. *Directed cops and robbers - In the undirected setting one defines a pursuit-evasion game on the vertices of a graph G, and it is known that a finite Giscop-win if and only if Gis dismantlable [ 63]. We do not know if there is a similar characterization (or even a similar game) for digraphs. *Gibbs measures and long range action - In [ 12] Brightwell and Winkler study a notion of long range action as well as Gibbs measures on the set of graph homomorphisms graphs. They again relate these properties to dismantlable graphs and We do not know if similar constructions apply in the directed theories - As we have discussed, the homotopy theory of digraphs introduced in [41] was developed in tandem with an underlying homology theory. These be seen as digraphs analogues of A-theory for undirected graphs studied in [ 3]. In all of these cases, the underlying product is the cartesian product of (di)graphs. Our three notions of homotopy described in Section 6 are based on the categorical product and lead to new constructions. For instance, for any two digraphs GandHwe can define the G-bihomology groups of Hto be ~Hi(- - --Hom(G,H)). Our results from Section 6 then is invariant under bihomotopy equivalence of digraphs. Similarly, we can define the G-line homology groups of Hto be conjecturally (see Section 6.2) is invariant under line homotopy equivalence of digraphs. Similar recently considered by Bubenik and Mili 'cevi'c in [ 14] although we are not aware of the precise categories - In any category with a notion of equivalence, a natural question to ask if one can impose a model category structure. In the context of undirected graphs, Matsushita [58] described a model category structure using Z2-homotopy equivalence of as weak equivalences. In [ 34] Droz constructed model categories with other notions of weak equivalences. On the other hand, in [ 40] Goyal and Santhanam showed no model category structure exists if one takes x-homotopy as weak equivalence and inclusions as cofibrations. It is an open question whether one can use any of the three notions of homotopy equivalence for digraphs discussed here in the context of a model category structure. Model categories have been studied in the context of directed homotopy for instance in work of Gaucher rings - For any simplicial complex Xon vertex set[n], one defines the Stanley- Reisner the monomial ideal in Rgenerated by nonfaces of X. In Corollary 4.13 we saw that if Gis a simple digraph on at most m=2n+2vertices then the neighborhood This has consequences for the ideal I(--N(G)), and in particular the regularity of the factor ring R/I(--N(G)). It would be interesting to study 32other algebraic invariants of these rings, and to relate them to combinatorial properties of the underlying digraph Alon, P . Frankl, L. Lov 'asz, The chromatic number of Kneser hypergraphs , Trans. Amer. Math. Soc. 298(1986), no. 1, pp. 359-370. 2 [2]F. Ardila, S. Billey, Flag arrangements and triangulations of products of simplices , Adv. Math. 214, (2007), no. 2, pp. 495-524. 25 [3]E. Babson, H. Barcelo, M. de Longueville, and R. Laubenbacher, Homotopy theory of graphs , J. Algebraic Comb. 24 (2006), pp. 31-44. 3, 27, 29, 32 [4] E. Babson, L. Billera, The geometry of products of minors , Discrete Comput. Geom. 20(1998), pp. 231-249. 25 [5] E. Babson, D. N. Kozlov, Proof of the Lov' asz Conjecture , Ann. of Math. 165(2007), no. 3, pp. 965-1007. 1, 31 [6]E. Babson, D. N. Kozlov, Complexes of graph homomorphisms , Israel J. Math. 152(2006), pp. 285-312. 1, 8, 11, 13, 14 [7]J. A. Barmak, E. G. Minian, Strong homotopy types, nerves and collapses , Discrete Comput. Geom. 47(2012), pp. 301-328. 8, 30 [8] A. Bj orner, Topological methods , Handbook of combinatorics 2(1995), pp. 1819-1872. 8, 15, 18 [9]F. Brandt, M. Brill, and P . Harrenstein, Tournament Solutions . Chapter 3 in: F. Brandt, V . Conitzer, U. Endriss, J. Lang, A. Procaccia, Handbook of Computational Social Choice, Cambridge University Press, 2016. 20 [10] B. Braun, J. Browder, S. Klee, Cellular resolutions of ideals defined by nondegenerate simplicial homomorphisms , Israel J. Math. 196(2013), pp. 321 - 344. 2 [11] B. Braun, W. K. Hough, Homomorphism complexes and maximal chains in graded posets , European J. Combin. 81 (2019), pp. 178-194. 2, 9 [12] G. R. Brightwell, P . Winkler, Gibbs Measures and Dismantlable Graphs , J. Comb. Theory Ser. B 78(2000), no. 1, pp. 141-166. 5, 28, 32 [13] G. R. Brightwell, P . Winkler, Graph homomorphisms and long range action , Graphs, morphisms and statistical physics, DIMACS Ser. Discrete Math. Theoret. Comput. Sci. 63, Amer. Math. Soc., Providence, RI, 2004, pp. 29-47. 2 [14] P . Bubenik, N. Mili homology and persistent homology using VCech's closure spaces , preprint . 32 [15] L. Cereceda, J. van den Heuvel, M. Johnson, Connectedness of the graph of , Discrete Math. 308 (2008), pp. 913-919. 2, 22, 25 [16] L. Cereceda, J. van den Heuvel, M. Johnson, Mixing 3-colourings in bipartite graphs , European J. Combin. 30(2009), no. 7, pp. 1593-1606. 2, 22, 25 [17] T. Chih, L. Scull, A homotopy category for graphs , J. Algebraic Comb. 53(2021), pp. 1231-1251. 26 [18] P . Csorba, Fold and Mycielskian on homomorphism complexes , Contrib. Discrete Math. 3(2008), pp. 1-8. 4, 17 [19] P . Csorba, Homotopy types of box complexes , Combinatorica 27(2007), no. 6, pp. 669-682. 2, 16 [20] P . Csorba, Non-Tidy Spaces and Graph Colorings . PhD thesis, ETH Z urich, 2005. 15 [21] P . Csorba, C. Lange, I. Schurr, A. Wassmer, Box complexes, neighborhood complexes, and the chromatic number , J. Combin. Theory Ser. A 108(2004) pp. 159-168. 2 [22] P . Csorba, F. Lutz, Graph coloring manifolds , in: Algebraic and Geometric Combinatorics, Contemporary Mathe- matics 423, American Mathematical Society, Providence, RI, 2007. 2 [23] S.VCuk 'ic, D. N. Kozlov, Higher connectivity of graph coloring complexes , Int. Math. Res. Not. 2005 , no. 25, pp. 1543-1562. 2, 31 [24] S. Lj. VCuk 'ic, D. N. Kozlov, The homotopy type of the complexes of graph homomorphisms between cycles , Discrete Comput. Geom. 36(2006), no. 2, pp. 313-329. 2 [25] J. De Loera, J. Rambau, F. Santos, Triangulations, Structures for Algorithms and Applications , Algorithms and Computation in Mathematics, vol. 25, Springer-Verlag Berlin Heidelberg, 2010. 25 [26] M. Develin, B. Sturmfels, Tropical convexity , Documenta Math. 9(2004), pp. 1-27. 25 [27] A. Dochtermann, Hom complexes and homotopy in the category of graphs , European J. Combin. 30Issue 2 (2009), pp. 490-509. 2, 3, 12, 13, 26, 27, 29 [28] A. Dochtermann, Homotopy groups of Hom complexes of graphs , J. Combin. Theory Ser. A 116, (2009), pp. 180-194. 30 [29] A. Dochtermann, The universality of Hom complexes of graphs , Combinatorica 29(2009), pp. 433-448. 16 33[30] A. Dochtermann, A. Engstr om, Cellular resolutions of cointerval ideals , Mathematische Zeitschrift 270(2012), pp. 145-163. 2, 25 [31] A. Dochtermann, R. Freij-Hollanti, Warmth and edge spaces of graphs , Adv. Appl. Math. 96(2018), pp. 176-194. 2 [32] A. Dochtermann, C. Schultz, Topology of Hom complexes and test graphs for bounding chromatic number , Israel J. Math. 187(2012), pp. 371-417. 1, 3, 26, 30, 31 [33] C. H. Dowker, Homology Groups of Relations , Ann. of Math. (2) 56(1952), pp. 84-95. 15 [34] J. M. Droz, Quillen model structures on the category of graphs , Homology Homotopy Appl. 14(2012), no. 2, pp. 265-284. 32 [35] A. Engstr om, A short proof of a conjecture on the higher connectivity of graph coloring complexes , Proc. Amer. Math. Soc. 134(2006), no. 12, pp. 3703-3705. 2 [36] L. Fajstrup, E. Goubault, M. Raussen, Algebraic Topology and Concurrency , Theoretical Computer Science 357, (2006), pp. 241-278. 26 [37] R. Forman, Morse theory for cell complexes , Adv. Math. 134(1998), no. 1, pp. 90-145. 8 [38] P . Gaucher, Six model categories for directed homotopy , Categories Gen. Algebraic Struct. 15(2021), no. 1, pp. 145-181. 32 [39] C. Giusti, E. Pastalkova, C. Curto, V . Itskovm, Clique topology reveals intrinsic geometric structure in neural correlations , PNAS 112(2015), no. 44, pp. 13455-13460. 30 [40] S. Goyal, R. Santhanam, (Lack of) model structures on the category of graphs , Appl. Categ. Structures 29(2021), no. 4, pp. 671-683. 32 [41] A. Grigor'yan, Y. Lin, Y. V . Muranov, S. T. Yau, Homotopy theory of digraphs , Pure and App. Math. Quat. 10(2014), pp. 619-674. 3, 5, 29, 32 [42] P . Hell, J. Ne VsetVril, Graphs and Homomorphisms, Oxford Lecture Series in Mathematics and Its Applications 28, Oxford University Press, 2004. 1 [43] K. Iriye, D. Kishimoto, Hom complexes and hypergraph colorings , Topology and Appl. 160(2013), pp. 1333-1344. 2 [44] M. Kahle, The neighborhood complex of a random graph , J. Combin. Theory Ser. A 114(2007), pp. 380-387. 2, 4, 18 [45] A. V . Kostochka, E. Sopena, X. Zhu, Acyclic and oriented chromatic numbers of graphs , J. Graph Theory 24(1997), no. 4, pp. 331-340. 22 [46] D. N. Kozlov, Chromatic numbers, morphism complexes, and Stiefel-Whitney characteristic classes , in: Geometric Com- binatorics (eds. E. Miller, V . Reiner, B. Sturmfels), IAS/Park City Mathematics Series 14, American Providence, RI; Institute for Advanced Study (IAS), Princeton, NJ (2007). 31 [47] D. N. Kozlov, A simple proof for folds on both sides in complexes of graph homomorphisms , Proc. Amer. Math. Soc. 134 (2006), no. 5, pp. 1265-1270. 13 [48] D. N. Kozlov, Combinatorial Algebraic Topology, Algorithms and Computation in Mathematics 21, Springer, Berlin, 2008. 2, 9 [49] H. G. Landau, On dominance relations and the structure of animal societies. III. The condition for a score structure , Bulletin of Mathematical Biophysics 15(1953), no. 2, pp. 143-148. 20 [50] J-F. Laslier, Tournament solutions and majority voting , with a preface by Herv 'e Moulin. Studies in Economic Theory 7, Springer-Verlag, Berlin, 1997. 20 [51] C. Lange, G. Ziegler, On generalized Kneser hypergraph colorings , J. Combin. Theory Ser. A 114(2007), no. 1, pp. 159-166. 2 [52] L. Lov 'asz, Kneser's conjecture, chromatic number, and homotopy , J. Combin. Theory Ser. A 25(1978), no. 3, pp. 319-324. 1, 14, 18, 31 [53] G. Malen, Homomorphism complexes and k-cores , Discrete Math. 341(2018), no. 9, pp. 2567-2574. 31 [54] P . Masulli, A. E. P . Villam, The topology of the directed clique complex as a network invariant , SpringerPlus, 5(2016), no. 1. 30 [55] J. Matou Vsek, Using the Borsuk-Ulam theorem. Lectures on topological methods in combinatorics and geometry , with A. Bjorner and G.M. Ziegler, Universitext, Springer-Verlag, Berlin, 2003. 28, 31 [56] J. Matou Vsek, G. M. Ziegler, Topological lower bounds for the chromatic number: A hierarchy , Jahresber. Dtsch. pp. 71-90. 2 [57] T. Matsushita, Morphism complexes of sets with relations , Osaka J. Math. 53(2016), pp. 267-283. 2, 3, 5, 9, 12, 13, 27 [58] T. Matsushita, Box complexes and homotopy theory of graphs , Homology Homotopy Appl. 19(2017), no. 2, pp. 175-197. 2, 32 34

2108.10948

Anton Dochtermann

Anton Dochtermann and Anurag Singh

Homomorphism complexes, reconfiguration, and homotopy for directed graphs

34 pages, 10 figures; V2: some changes in notation, clarified statements and proofs, other corrections and minor revisions incorporating comments from referees

European J. Combin. 110 (2023)

math.CO math.AT

http://creativecommons.org/licenses/by/4.0/

The neighborhood complex of a graph was introduced by Lov'asz to lower bounds on chromatic number. More general of graphs were further studied by Babson and Kozlov. Such `Hom complexes' are also related to mixings of graph colorings and problems, as well as a notion of discrete homotopy for graphs. Here we initiate the detailed study of Hom complexes for directed For any pair of digraphs graphs $G$ and $H$, we consider the polyhedral complex $ ext{Hom}(G,H)$ that parametrizes the directed $f: G ightarrow H$. Hom complexes of digraphs have applications in the study of chains in graded posets and cellular resolutions of monomial ideals. We study examples of directed Hom complexes and relate properties to certain graph operations including and foldings. We introduce a notion of a neighborhood complex for a digraph and prove that its homotopy type is recovered as the Hom complex of homomorphisms from a directed edge. We establish a number of results regarding the topology of directed neighborhood complexes, including the dependence on directed bipartite subgraphs, a digraph version of the Mycielski construction, as well as vanishing theorems for higher homology. The Hom complexes of digraphs provide a natural framework for reconfiguration of homomorphisms of digraphs. Inspired by notions of directed graph colorings we study the connectivity of for $T_n$ a tournament. Finally, we use paths in the internal hom objects of digraphs to define various notions of homotopy, and discuss connections to the topology of Hom complexes.

2 Morey et al. it requires at least 0.8 Gyr to grow a 109MSMBH from a 100 Mseed with a ducial radiative eciency of 10% (Shakura & Sunyaev 1973) and accretion at the Eddington limit, i.e. Lbol=Ledd. It remains an open question as to the validity of this simple model and the possibility of additional physics leading to more complicated light curves (e.g. Di Matteo et al. 2005; Springel et al. 2005; Hopkins et al. 2005; Novak et al. 2011; Davies et al. 2019a). At high redshifts, the age of the universe is compara- ble to the timescale required for SMBHs to form (Volon- teri 2010). However, measurements of the timescales of quasar activity have proven to be dicult. At lower red- shifts ofz24 previous studies have estimated the duty cycle of quasars, which denotes the total fraction of time compared to the age of the universe that galax- ies spent as luminous quasars and therefore represents an upper limit on the quasars' lifetime tQ, by compar- ing the number density of quasars to the abundance of dark matter halos via clustering studies (Efstathiou & Rees 1988; Haiman & Hui 2001; Martini & Weinberg 2001; Martini 2004). These studies rely on the measure- ment of the abundance of quasar host dark matter halos in comparison to the abundance of luminous quasars to estimate the quasars' duty cycle (White et al. 2008). However, these methods have yielded only weakly con- strained estimates of 106109years, since cluster- ing estimates are susceptible to high uncertainty due to variance in parameters governing how quasars populate dark matter halos (Shen et al. 2009; White et al. 2012; Conroy & White 2013; Cen & Safarzadeh 2015). Other estimates of quasar activity timescales have been made using an extension of the Soltan argument (Soltan 1982), where the quasar luminosity function can be com- pared to quiescent SMBHs observed in galaxies locally, which have led to an estimate of the quasars' duty cycle to between 107108years (Yu & Tremaine previous studies have used the quasars' light echo transverse to the line{of{sight to measure the time delay between the onset of the quasar emission and changes in opacity of the intergalactic medium (Adel- berger 2004; Hennawi & Prochaska 2007; Schmidt et al. 2017, 2018; Bosman et al. 2020) to constrain the lifetime of quasars to be studies have shown how the lifetime of quasars tQcan be inferred by means of the proximity zone re- gion observed in rest-frame UV and optical spectra of high redshift quasars (Khrykin et al. 2016; Eilers et al. 2017, 2018; Khrykin et al. 2019; Davies et al. 2019b; Eilers et al. 2020). The proximity zone is a region of en- hanced transmitted ux in the vicinity of quasars that has been ionized by the quasar's own radiation (Ba-jtlik et al. 1988; Haiman & Cen 2001; Wyithe et al. 2005; Bolton & Haehnelt 2007; Lidz et al. 2007; Bolton et al. 2011; Keating et al. 2015). Estimating quasar life- times from the HI proximity zone is a technique that has primarily been used to estimate tQof high red- shift quasar populations, but recent studies have used the Helium II line at z'34 to measure individ- ual quasar lifetimes and the quasar lifetime et al. 2021; Worseck et al. 2021; Khrykin et al. 2019). The intergalactic medium has a nite response time to the quasar's radiation and therefore the level of enhanced ux due to ionized gas around the quasar is highly sensitive to the lifetime of quasars. Eilers et al. (2017, 2018) used proximity zones in the HILyforest to estimate the lifetime for z'6 quasars. The same method can be applied to proximity zones in the He IILyforest to estimate the lifetimes of quasars at low redshift ( z'34), as was done in Khrykin et al. (2019). Eilers et al. (2017) found a population of quasarsz'6 that show very small proximity zones and therefore are likely to be very young objects ( a simple light-bulb light curve for the quasars, i.e. the quasar turns on abruptly and emits at a constant luminosity during its entire lifetime tQ, the probability of detecting a quasar age of that these young quasars corresponded to '10% of the population, this suggested a ducial lifetime of tQ= 106yr for the quasar population (Eilers et al. 2017, 2020). Davies et al. (2019a) expanded upon this work by developing a semi-analytic model of quasar proximity zones, including non-equilibrium interactions between the IGM and the quasar, using hydrodynamical radia- tive transfer simulations to simulate the proximity zone behavior on dierent timescales. The observed distribu- tion of proximity zones by Eilers et al. (2017) was sim- ilar to the modeled distribution for tQ'106using the lightbulb model. As part of this analysis, Davies et al. (2019a) considered more general light curves, not just the lightbulb model, and found consistent results for the lifetime, again conrming evidence for tQ'106yr. Davies et al. (2019b) measured the lifetime of individ- ual high redshift quasars during the epoch of reioniza- tion (EoR), and measured the age of these quasars to By using the neutral IGM as a counter of the ionizing photons emitted by the quasar, they were able to constrain the lifetimes of two z&7 quasars. By measuring the number of emitted ionizing photons, they were able to directly measure the radiative much smaller than the canonical 10% value given by Eddington limited accretion. This study notes, however, that if this behavior is instead caused by UVThe effective lifetime of the quasar population at z6 3 obscured quasars, these radiative eciencies are consis- tent with a fraction of UV obscured quasars greater than 82% at redshift z'7. However, further work is needed to characterize whether this mechanism is indeed radia- tive ineciency or UV lifetime estimates are more than an order of magnitude shorter than expected and cause tension with the current model for the growth of SMBHs. A poten- tial solution could be the presence of very massive initial black hole seeds (i.e. Mseed>1000M), vastly exceed- ing the mass of Population III stellar remnants (Lodato & Natarajan 2006; Visbal et al. 2014; Habouzit et al. 2016; Schauer et al. 2017). As mentioned above, very low radiative eciencies could also explain such short lifetimes, with radiative eciency rate suggested by other studies (Volonteri et al. 2015; Trakhtenbrot et al. 2017; Davies et al. 2019b), or highly UV obscured black hole growth phases (Eilers et al. 2018; Davies et al. 2019b). Additional estimates of tQ, both for individual quasars and for the entire pop- ulation, could provide important constraints on i.e. provide constraints on the mass of the initial seeds or the specic radiative eciency rate . In this study, we will estimate the eective lifetime of the high-redshift quasar population by means of the composite transmitted ux within the proximity zone region. To this end, we assembled a data set of 15 quasar spectra atz6 with a similar absolute magnitude, i.e. a similar ionizing radiation output. By comparing this stacked proximity zone prole to outputs from simulations of quasars at dierent lifetimes, we constrain the eective lifetime of the quasar population. This paper is organized as follows: in SS2 we discuss our quasar sample and the selection criteria. In SS3 we describe our quasar continuum estimation procedure and the radiative transfer models. In SS4, we discuss our stacking procedure, as well as the forward model- ing of uncertainties, and present our nal result. We summarize our results in SS5. Throughout this paper, we assume a at CDM cos- mology ofh= 0:685, m= 0:3, and = 0:7, which is consistent within the 1 errorbars from Planck Collab- oration (2020). 2.DATA SAMPLE For our analysis we select quasars at 5 :8< z < 6:6, and an absolute magnitude at 1450 A in the a narrow range of We require that all objects in our sample have spectroscopy at optical and possibly also near-infrared (NIR) wavelengths. Additionally, we select quasars for which precise redshift estimates (
v.100 km s1) based on sub-mm observations of the [C II] emission line at 158 m (rest= 1900:548 GHz) or the CO(6{5) emission line at 3 mm ( rest= 691:473 GHz) from the Atacama Large Millimetre Array (ALMA) or the NOrthern Extended Millimeter Array (NOEMA) at the Institute de Radioastronomie Millim etrique (IRAM) are available, since uncertainties on the quasars' sys- temic redshift contribute the largest source of uncer- tainty to the proximity zone measurements (Eilers et al. 2017, 2020). Furthermore, we remove all quasars from the sample that show broad absorption line features (BALs) or proximate damped Ly systems (pDLAs), which would cause additional absorption in the proxim- ity zone region and would bias our results. These selec- tion criteria result in 15 objects with a mean redshift of hzi= 6:23, and mean magnitude of overview of our data sample is shown in Table 1. Of the quasars in our sample, eleven were observed with the X-Shooter instrument at the Very Large Tele- scope (VLT) (Vernet et al. 2011), two where observed using the DEep Imaging Multi-Object instrument at the Keck observatory Faber et al. (2003), one was observed using the Echellette Spec- trograph and Imager (ESI) instrument also at the Keck observatory (Sheinis et al. 2002), and one was observed using the Folded-port InfraRed Echellette (FIRE) spec- trometer instrument at the Magellan telescope (Simcoe et al. 2013). The individual detector resolutions R=
 areR8800 for the X-Shooter instrument Vernet et for the FIRE et al. 2013),R5000 for the DEIMOS instrument (Faber et al. 2003), and R5400 for the ESI instrument (Shei- nis et al. 2002). 2.1. Data Reduction All optical and NIR spectroscopic data observed and Keck/DEIMOS were reduced ap- plying the open source python spectroscopic data re- duction package PypeIt (Prochaska et al. 2020). For the reduction we rst subtract the sky emission, which was performed on the 2D images by including both im- age dierencing between dithered exposures whenever these were available, and a B-spline tting procedure (e.g. Bochanski et al. 2009). We apply an optimal spec- trum extraction technique (Horne 1986) to extract the 1D spectra. All individual 1D spectra are ux calibrated using the standard stars LTT 3218 (for spectra observed with VLT/X-Shooter) and G191B2B or Feige 110 (for spectra taken with Keck/DEIMOS). These uxed 1D spectra are stacked and a telluric model and the quasar PCA model from Davies et al. (2018a) are jointly tted to the stacked spectra using telluric model grids from4 Morey et al.Table 1. Overview of our data sample. Name RA [hms] DEC [dms] Redshift M 1450 SNR PID PI PSOJ011+09 00:45:33.566 +09:01:56.93 6.4693 -26.85 VLT/X-Shooter 7 0101.B-0272(A) 02:26:01.873 +03:02:59.24 6.5412 -27.28 VLT/X-Shooter 11 0102.A-0154(A)/ 0100.A-0625(A) 04:21:38.050 -26:57:15.72 6.1877 -27.25 VLT/X-Shooter 19 098.B-0537(A) 08:42:29.438 +12:18:50.47 6.0763 -26.91 VLT/X-Shooter 26 097.B-1070(A) 10:34:46.509 -14:25:15.86 6.0681 -27.41 VLT/X-Shooter 15 096.A-0418(B) 10:36:54.190 -02:32:37.940 6.3809 -26.59 VLT/X-Shooter 9 098.B-0537(A) 11:43:38.347 +38:08:28.823 5.8367 -26.69 Keck/DEIMOS 15 2017A U078 11:48:16.64 +52:51:50.2 6.4189 -26.81 Keck/ESI 37 U43E, U11E, U2E Becker C70E H45aE Cowie C97E 13:06:08.259 +03:56:26.19 6.0337 -26.81 VLT/X-Shooter 38 084.A-0390(A) 13:19:11.291 +09:50:51.49 6.1330 -26.88 VLT/X-Shooter 52 084.A-0390(A) 13:35:50.81 +35:33:15.82 5.9012 -26.67 Keck/DEIMOS 22 2017A U078 14:28:21.371 -16:02:43.73 6.1498 -26.93 Magellan/FIRE 9 15:09:41.781 -17:49:26.68 6.1225 -27.14 VLT/X-Shooter 37 085.A-0299(A) 21:32:33.178 +12:17:55.07 6.5872 -27.06 VLT/X-Shooter 14 098.B-0537(A) 23:56:32.452 -06:22:59.31 6.1719 -26.79 VLT/X-Shooter 19 098.B-0537(A) Farina Note |The sub-mm redshift estimate for PSOJ036+03 was obtained by Venemans et al. (2015), the redshift estimate for J1319+0950 was obtained by Wang et al. (2013), the redshift estimate for SDSSJ1335+3533 was obtained by Wang et al. (2013), the redshift estimates for SDSSJ1143+3808, PSOJ359-06, PSOJ158-14, and PSOJ011+09 were obtained by Eilers et al. (2020), and all others were published in Decarli et al. (2018). The SNR values are estimated using a pixel scale of 0.05 A/pixel in the rest frame of the quasars.The effective lifetime of the quasar population at z6 5 the Line-By-Line Radiative Transfer Model (LBLRTM1; Clough et al. 2005; Gullikson et al. 2014). We then match the visible (VIS) and NIR regions for the VLT/X-Shooter spectra. To this end, we compute the mean value of all pixels between 10070 and 10270 A in the observed frame, and then scale the NIR to a constant factor such that the two means are equal. In this step, we do not normalize the spectra to unity, but merely scale the NIR spectra to be matched with the VIS spectra. For the FIRE spectra, which do not require matching between NIR and visible parts, we use the FIREHOSE2, developed by Simcoe et al. (2010). The procedures used for data reduction include cosmic- ray rejection, at-elding, wavelength calibration using OH sky lines (calibrated to vacuum wavelength), nonlin- earity correction, optimal extraction of Horne (1986), combining individual multiple echelle orders, and heliocentric cor- rections. This pipeline performs sky subtraction via B-spline model of the sky, based on the technique developed in Bochanski et al. (2009). We reduce the ESI spectra using the ESIRedux as part of the XIDL4suite of in the Interactive Data Language (IDL). This pipeline is very similar to FIREHOSE . More details to this pipeline for high-redshift quasars can be found in Eilers et al. (2017). In the end we mask all spectral regions that are af- fected by telluric absorption, namely between 13000 and 15200 A, 17500 and 20000 A, and beyond 22500 A in the observed frame. Figures 1 and 2 show the nal spectra of all 15 quasars in our sample. We estimate the SNR of the spectra by taking the median pixel-based ratio of the signal to noise between 1265 and 1295 A in the rest frame of the quasar, due to the absence of any emission lines in this region. We scale the pixel-based SNR for each of the instruments assuming Gaussian white noise to the X-shooter resolution to compare the SNR mea- surements with a common resolution in Table 1. 3.METHODS We aim to estimate the lifetime of the quasar popula- tion as a whole by means of their composite proximity zone region. Thus, in order to obtain the stacked trans- mission prole within the quasars' proximity zones, we rst have to normalize the spectra by their Once the quasars are continuum normalized, we stack them between 1190 and 1218 A in the rest{ frame, in order to obtain an estimate for the composite ux transmission prole in the proximity zone region. After stacking the quasars, we compare this stacked ux prole to 1-d radiative transfer models. 3.1. Quasar Continuum Normalization To estimate the continuum of the quasar within the proximity zone region, we t for the quasar of Ly , as this side is unaected by IGM ab- sorption. We then use this red-side estimate to predict the blue-side continuum, which is aected by et al. 2005; Paris et al. 2011; Davies et al. 2018a; Bosman et al. 2021). As a rst step to estimate the quasar continua we normalize the matched spectra to unity at 1290 2:5 A, for ease of tting. Then, we sigma clip outlying val- ues at the 3 level from the smoothed spectra (10 pixel lter), shown in Figures 1 and 2, using the noise vec- tors from each spectra as an estimate of the . To esti- mate the quasar continua, we make use of the analysis (PCA) method developed by Davies et al. (2018a), where we use 10 red-side basis between 1200 and 2850 A for most quasars in the sample) and 6 blue-side basis spectra 1177 and 1280 A). For SDSSJ1143, SDSSJ1148 and SDSSJ1335 we use 10 truncated red-side basis spec- tra, instead spanning between 1200 and 1450 A to ac- commodate these quasars where we do not have principal components were determined on a sam- ple of 12;764 low-redshift quasars at z2, for which the blue-side continuum is easier to estimate due to the lower level of absorption by the intergalactic medium. Davies et al. (2018a) noticed a high degree of correla- tion or anti-correlation between the red-side coecients and blue-side coecients, allowing the blue side coe- cients to be estimated from the red side. This can be done using the relationship bi=P10 j=1rjXji, where the biandrjare blue side and red side PCA and Xjiis a projection matrix determined by the correlations or between the blue and red side coecients from the training sample. Using the MCMC ane-invariant ensemble sampler emcee (Foreman-Mackey et al. 2013), we t each spec- trum using 10 basis spectra redwards of the Ly emis- sion line spanning between 1220 A and 2800 A in the rest frame, as well as a redshift oset zbetween the of the quasar and the best redshift estimate for the PCA. We rst take the median of the posterior dis- tribution to estimate the median red side coecients,6 Morey et 5.8367 6000 8000 10000 12000 14000 16000 18000 20000 6.4189 0.0 0.2 0.4 0.6 0.8 1.0 Observed Wavelength [ 1. Quasar spectra from our sample. We mask the regions of telluric absorption in grey. All uxes (black) and noise vectors (grey) have been inverse-variance smoothed with a 10 pixel lter. We also show the quasar continuum estimates tted on the red-side (red) and the blue-side prediction (blue) from our PCA analysis described in section 3.1.The effective lifetime of the quasar population at z6 6.5872 6000 8000 10000 12000 14000 16000 18000 20000 6.1719 0.0 0.2 0.4 0.6 0.8 1.0 Observed Wavelength [ 2. Same as Fig. 1 for the remaining objects in our sample. we then use the projection matrix Xjito estimate the coecients for the blue-side continuum. However, un- like Davies et al. (2018a) who developed this machin- ery for studying IGM damping wings in z > 7 quasar spectra (Davies et al. 2018b), we match the blue and red side basis spectra at 1220 A instead of at 1280 A(Eilers et al. 2020). This is due to the fact that all quasars in our sample are at z.6:6, and thus we do not expect signicant absorption redwards of the Ly  emission line due to IGM damping wings. We set a at prior on the PCA components to be the 3 bound- aries on each of the components determined by Davies8 Morey et 5.8367 1180 1190 1200 1210 6.1719 1180 1190 1200 1210 6.4189 0.0 0.2 0.4 0.6 0.8 1.0 Rest-Frame Wavelength [ Normalized Flux Figure 3. Continuum normalized spectra (black) along with continuum normalized noise vectors (grey) within our sample. The continuum normalized spectra are shown between 1177 and 1222 A, except for PSOJ217-16, where we only show between 1190 and 1222 A, due to detector noise. All spectra have been inverse variance smoothed with a 3 pixel lter. The grey dashed line marks the Ly line at 1215.67 A.The effective lifetime of the quasar population at z6 9 et al. (2018a), and allow a prior z2[0:01;0:03] for all quasars except for PSO J323+12, where we extend the prior in order to allow for a better continuum t. The quasar spectra and their are shown in Fig. 1 and 2. The continuum nor- malized spectra that are obtained when dividing by the estimated continuum emission are shown for each quasar in Fig. 3. The quasar continuum normalization procedure intro- duces additional uncertainty into our analysis procedure due to the fact that the continuum estimation can either be an overestimate or underestimate of the true contin- uum. It can be seen that the PCA based method un- derestimates the continuum for SDSSJ1143+3803 since the continuum normalized ux is clearly above a ux of unity close to the Ly line. However, the estimated bias in the PCA is very small, i.e. <1% (Davies et al. 2018a; Bosman et al. 2021) and thus the this is balanced by the roughly symmetric over- estimation of other quasar continua, which can be less obviously identied as it decreases the continuum nor- malized ux near Ly . We address this uncertainty in the quasar continuum estimates in detail in section 4.1. 3.2. Radiative Transfer Simulations For each quasar in our sample we use a 1-d radia- tive transfer (RT) code (Davies et al. 2016) to simulate the eect of ionizing radiation emitted by the quasar along the line-of-sight (Bolton & Haehnelt 2007). We take 1000 skewers originating from the most massive dark matter halos with halo masses of 4 the cosmological hydrody- namical simulation Nyx with a box of 100 Mpc h1on a side (Almgren et al. 2013; Luki c et al. 2014). The RT code computes the time{dependent ionization and re- combination rates of six particle species ( e, HI, HII, HeI, He II, He III), as well as the heating and cooling of the IGM due to the expansion of the universe and inverse Compton cooling o the cosmic microwave background. The Nyx simulation has outputs at z= 5:5, 6:0 and 6:5. We choose the output closest to the quasar's emis- sion redshift zemand re-scale the physical densities by (1 +zem)3. We used the Lusso et al. (2015) compos- ite composite to relate M1450to the specic ux at the Lyman limit, and then compute the ionizing ux by as- suming a spectral slope of =1:70. Previous stud- ies of quasar lifetimes based on proximity zones have only considered the extent of the proximity zone (with the notable exception of Davies et al. (2018b, 2019a)), which is dened as the location at which the ux smoothed to a resolution of 20 A in theobserved frame drops below the 10% ux (Fan et al. 2006). Therefore, the precise ioniza- tion rate of the ultraviolet background (UVB) did not play a crucial role, since the total ionization rate was always dominated by the quasar at the 10% ux trans- mission level. In this work, however, we consider the complete ux prole within the vicinity of the quasar, which extends below the 10% ux transmission level. At these lower ux transmission levels the ionization rate of the UVB is comparable to the ionization rate of the quasar and therefore needs to be considered, although the UVB should not matter too much given that we are assuming uniformity. Nonetheless, we re-scale the ionization rate of the UVB radiation UVB to match the expected strength at the quasar's redshift based on Haardt & Madau (2012). We then simulate the composite quasar spectrum as a function of quasar lifetime, assuming a light-bulb light curve, ranging from log10(tQ=yr) = 18:9 in steps of log10(tQ=yr) = 0:1. Each simulated spectrum is tuned to that quasar's particular redshift and luminosity, and the UVB was calibrated at the particular redshift to match Haardt & Madau (2012). For each step in the simulation, we compute 1000 dierent spectra at dier- ent sight lines, i.e. we have a model grid of 15 spectra, since there are 15 dierent quasars, 1000 dierent sightlines, and 80 dierent values of the lifetime, from 1.0 to 8.9 in log tQspace. At each time step we stack the simulated spectra by averaging the ux pixels across all 15 quasars and across all 1000 sight lines, since each simulation is computed on the same velocity grid. The simulated stacked spectra for a few selected lifetimes are shown in the left panel of Fig. Stack of the Quasars' Proximity Zones We now stack the uxes within the proximity zone region in order to create a compos- ite ux transmission prole. To this end, we dene a regularly spaced grid between 1190 A and 1218 A with a grid spacing of 0 :5A. We then bin the wavelengths and the ux pixels within this new wavelength grid, i.e. we take an average of all the wavelength pixels and ux pixels within a particular wavelength bin to obtain the stacked ux prole. The full composite spectrum within the proximity zone region from the 15 quasars in our data sample is shown in the right panel of Fig. 4. 4.2. Forward-Modeling of are two possibilities to estimate uncertainties on the ux pixels within the stacked proximity zone, i.e. either via bootstrapping of the Morey et al. 1190 1195 1200 1205 1210 1215 Rest Frame Wavelength [ Normalized FluxtQ= 102years tQ= 104years tQ= 106years tQ= 108years 1190 1195 1200 1205 1210 1215 1220 Rest Frame Wavelength [ Normalized FluxtQ= 4. Left panel: Proximity zone prole as predicted by radiative transfer simulations for log10(tQ=yr) =2, 4, 6, and 8. The shaded regions around each of the mean models show the 1 uncertainty range as described in section 4.2. The grey dashed line marks the Ly line at 1215.67 A.Right panel: Plot of the quasar stack, with a bin size of 0.5 A, with the best t model withtQ= 105:7yr shown in purple. spectra, or via forward modeling the various sources of uncertainty onto the simulated quasar spectra. Forward modeling the sources of uncertainty does not give us an estimate of the covariance directly from the data, rather we use covariance estimated from the forward models in performing inference of the quasar lifetime tQ. This gives a better estimate of the covariance than estimating it directly from the data, with the only drawback that the uncertainty on the stack is now to the limited size of our data sample, the covari- ance matrix obtained via bootstrapping is very noisy, and thus we forward{model the uncertainties onto the simulated quasar spectra from the RT simulations by means of forward modeling. We then estimate the co- variance between the dierent ux bins on the sample of 10;000 mock stacks with uncertainties. For each stack, we incorporated four types of uncertainty: cosmic vari- ance, uncertainty from the PCA continuum in the systemic redshift, and spectral noise. We assume that the full uncertainties on the stack can be approximated as a multivariate Gaussian distribu- tion about a mean model. We will now describe in de- tail how these uncertainties are being modeled in the quasar stack. We repeat this uncertainty analysis for each lifetime, so we obtain 80 dierent samples of 10 ;000 mock spectra. We then assume that these mock stacks are from a Gaussian distribution and then estimate the covariance of this distribution directly using the mock stacks. The rst source of uncertainty that we incorporated is uncertainty from cosmic variance. Thus, for each quasar in the stack we randomly chose one of the 1000 the cosmological simulation box, ensuring thatevery simulated stack consists of 15 dierent sightlines. In order to mimic the spectral resolution of the data, we convolve each spectrum with a Gaussian kernel with a standard deviation in velocity space given by vel= c 2:355R, wherecis the speed of light, Ris the resolution of the spectrograph, and the factor of 2 :355 arises from the full width at half maximum (FWHM) to conversion. Following this, we interpolate the convolved spectrum onto the wavelength grid of the data for the specic quasar in question. That is, we aim to generate simulated data for each quasar that has the same wavelength grid as the actual data, that way we can use the noise from the data to simulate Davies et al. (2018a) procedure produces a co- variance matrix for the continuum prediction, which we can directly sample from to create mock realizations of the uncertainty in the continuum estimate. We then scale the simulated spectra for each quasar by a draw from this distribution. In or- der to incorporate uncertainties in the systemic of the quasars, we assume an uncertainty on the systemic redshift of
v= 100 km s1, and randomly draw from a normal distribution with a mean of = 0 and the corresponding standard deviation (Eilers et al. 2020). The simulated quasar spectra are then shifted to the rest-frame wavelength grid based on this new red- shift estimate. Finally, in order to incorporate spec- tral noise, we directly sample the noise vectors from the data, as we have already interpolated the simulated con- volved spectra onto the wavelength grid. To be clear, we are generating sets of 15 quasars each on the wave- length grid of the data, allowing us to directly sampleThe effective lifetime of the quasar population at z6 11 1190 1195 1200 1205 1210 1215 Wavelength [ [ 5. Correlation matrix for log10(tQ=yr) = 5:7 com- puted from Monte-Carlo sampling the distribution of quasar stacks directly from the simulated quasar spectra from ra- diative transfer models. the noise without re-scaling, after which we can stack the 15 quasars within a set using our binning the radiative transfer models are dependent on the quasar lifetime, as an input parameter, we per- formed this Monte-Carlo sampling of mock stacks for each value of the lifetime on a grid between 1.0 and 8.9 in log10years, with a grid spacing of
log10tQ= 0:1. This means that we computed 80 dierent for each lifetime dependent model, which we used when we constructed our likelihood function over the log lifetime grid. For each lifetime dependent model we use the same set of random draws for the Monte Carlo sampling to ensure that the likelihood function varies smoothly with the model parameter, rather than being in uenced by the stochasticity of the draws. We show an example correlation matrix for the model cor- responding to tQ= 105:7yr in Fig. 5. 4.3. Quasar Lifetime Estimate We assume that the data is drawn from a by one of the MC estimated covariance ma- trices, centered at the mean model for that Additionally, we also assume that each quasar in the stack has the same lifetime, i.e. our models are made up of stacks of quasars of the same lifetime (see SS5 for a discussion of this our assumptions, we construct a lifetime depen- dent likelihood (tQ)) ; (3) 5.5 6.0 6.5 7.0 log10(tQ) DensityFigure 6. Posterior PDF of the average quasar lifetime tQ. Vertical dashed lines indicate the median (orange), 68% credible interval (red), and 95% credible interval is the likelihood of observing the data given a particular model tQ,kis the number of wave- length bins in the stacked spectrum, and ( tQ) is the model dependent covariance matrix. However, we only knowr(tQ) and (tQ) at discrete points in since we have outputs of the RT simulation only at discrete time steps. We calculated the likelihood func- tion at the discrete points within the log10(tQ) space (i.e. where our covariance matrices are dened) and interpo- lated the likelihood function to estimate the the grid-points. To construct our likelihood we consider wavelengths from 1190 to 1217 A in the rest-frame. This cuto at 1190 A avoids potential biases in the very low ux regime, where the ux transmission is completely dom- inated by the UVB. To obtain the posterior probability density function (PDF) for the lifetime, we use a at prior for tQ2 [10;108:9] yr, meaning that the posterior is merely the normalized version of the likelihood. We normalize the likelihood function using trapezoidal rule numerical inte- gration to obtain the result shown in Fig. 6. We obtain a measurement of the eective lifetime of the quasar pop- ulation of log10(tQ=yr) = the median and 68th (95th) percentile of the posterior This best t model is shown in the right panel of Fig. 4. 5.SUMMARY & DISCUSSION In this paper we estimate the eective lifetime of the quasar population at z6 by means of the average ux transmission prole in the proximity zone region around the quasars. To this end, we stack the continuum-12 Morey et al. 1180 1190 1200 1210 Wavelength [ Normalized Fluxs= 0.0 s= 1.0 s= 2.0 Figure 7. Modeled stacks of the 15 quasars in the stack for a mean lifetime of log10(tQ=yr) = 5:7, but for quasar widths = 0;1;2, along with 1 the spectra of 15 quasars at 5 :8<z < 6:6 with similar absolute luminosities, i.e. 27:4< M 1450< 26:6. For all objects in this data set precise estimates from sub-mm emission lines are avail- able, since uncertainties in the systemic redshifts con- tribute the largest source of uncertainty to proximity zone estimates (Eilers et al. 2017). This average com- posite transmission prole is then compared to simu- lated transmission proles from 1-d RT simulations at dierent eective quasar lifetimes. Our best estimate for the eective lifetime of the quasar population as a whole based on the zone prole is log10(tQ=yr) = from the median of the posterior as well as the 68th (95th) percentile. In order to estimate the uncertainties on the stacked spec- trum we forward-model the uncertainties arising due to cosmic variance, the PCA the estimate of the quasars' systemic redshifts, as well as spectral noise onto the simulated quasar sightlines. Our measurement of the eective lifetime for the quasar population is consistent with previous work, which had estimated an average quasar lifetime of tQ 106yr (Khrykin et al. 2019; Davies et al. 2019a). Our results are also consistent with the observed fraction of young quasars within the early universe with life- times oftQ.104105yr, which was determined to be 5% .fyoung.10% (Eilers et al. 2020; Eilers et al. 2021). This is because we expect to see of quasars to shine for 10% of the lifetime, if the quasars' ages are uniformly sampled from [0 ;tQ]. In reality it is likely that the lifetime of the quasar pop- ulation can not be described by a single value, but ratherby a distribution of lifetimes with a mean of tQ=tQ and standard deviation tQ. Because we do not mea- sure, or even consider the full distribution of we note that our measurement of the eec- tive lifetime will be comparable to the average lifetime only in the regime where the variance of the quasar life- time distribution is small. Khrykin et al. (2016) showed that a large variance of the quasar lifetime an underestimate of the lifetime, because quasars with lifetime less than tQimpact the proximity zone more. Additionally, if we have a very young quasar in our stack, this could cause our estimate to be an un- derestimate of the lifetime. This is due to the fact that the intergalactic gas reaches ionization equilibrium with the quasar's radiation for lifetimes longer than the equi- libration timescale, i.e. tQ> teq3104yr. This means that our measurement of the eective lifetime of the quasar population could potentially mean quasar lifetime, if the width of the distribution is very wide. In Figure 7 we show modeled stacks of the same mean lifetime (log10(tQ=yr) = 5:7) for the distribution of quasar lifetimes, but with dif- ferent standard deviations. As we can see our method is not really sensitive to the width of the quasar , unless the width of the lifetime distribu- tion is large, i.e. log10tQ&2, in which case our method would underestimate the average quasar lifetime. How- ever, recent results suggest only a very modest stan- dard deviation of the distribution of quasar et al. 2021), suggest- ing that our measurement of the eective quasar re ects the average lifetime of the quasar pop- ulation as a whole. Another potential source of uncertainty in our mea- surement is the systematic bias in the PCA Davies et al. (2018a) estimate that the continuum overestimates the true quasar continuum on average by 1%. However, this is a rel- ative error on the ux, i.e. at 10% ux transmission the estimated error is 0 :1% and hence this bias has only negligible eects on our nal results. Our lifetime measurement is more than an order of magnitude shorter than expected from models for SMBHs (Volonteri 2010, 2012). A potential explanation for such short average quasar life- times could be ickering quasar light curves instead of the simple light-bulb light curves, implying that there could be multiple epochs of quasar activity and con- current black hole growth. Since the transmitted ux within the proximity zone around quasars at z6 is only sensitive to the last on-time our composite prox-The effective lifetime of the quasar population at z6 13 imity zone spectrum from this analysis is not sensitive to potential previous phases of quasar several analysis have shown previously that even ickering quasar light curves cannot explain the discrepancy between the short average quasar lifetimes and the time required to explain the SMBH growth with the current exponential growth model. For instance, Davies et al. (2019b) showed that the damping wing feature observed in quasar spectra at z >7, where the surrounding IGM still has a very high neutral gas frac- tion, provides a constraint on the total number of ion- izing UV photons emitted into the IGM integrated over cosmic time, irrespective of the light-curve shape. Their analysis shows that the integrated time these quasars have been emitting UV radiation is still consistent with a light bulb light-curve with however, the IGM is highly ionized and the quasar spectra do not exhibit a damping wing. Hence the proximity zones are sensitive only to the on-time of the last accretion episode, which is only consistent with the total quasar lifetime for light-bulb light curves. (Davies et al. 2019a) has analyzed the distribution of z6 proximity zones in the context of various quasar light curves. While they nd a good agreement be- tween the observed distribution of proximity zones and model predictions when assuming a simple light-bulb light curve, ickering quasar light curves with multiple black hole growth phases cannot be excluded for z6 quasars. Our presented results at z6, the distribution of proximity zones at z6 (Davies et al. 2019a), the frac-tion of very young quasars within the quasar et al. 2020; Eilers et al. 2021), as well as the damping wing analysis at z >7 Davies et al. (2019b), are all in good agreement and point towards quasar life- times of log10(tQ)'106. These measurements suggests that either a signicant fraction of the black hole growth occurs in UV obscured, dust-enshrouded growth phases in the early universe, or it could indicate mass accretion rates, such that more of the accreted mass contributes to the growth of the black holes rather than the quasars' UV emission (Eilers et al. 2018; Davies et al. 2019b). In future work, we aim to explore possible mechanisms by which quasars can ac- crete sucient mass while only emitting UV on average. Additionally, we will expand this work to measure the properties of the full quasar lifetime distribution from the individual measurements, as suggested in (Khrykin et al. 2021). These studies will provide further insights for our understanding of SMBH formation in the early universe and whether or how the accretion mechanisms change with cosmic authors would like to thank the UROP oce at MIT for generous funding of this work. ACE acknowledges support by NASA through the NASA Hubble Fellowship grant #HF2-51434 awarded by the Space Telescope Science Institute, which is op- erated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract K. L. 2004, The Astrophysical Journal, 612, 706, doi: A. S., Bell, J. B., Lijewski, M. J., Luki c, Z., & Van Andel, E. 2013, The Astrophysical Journal, 765, 39, doi: S., Duncan, R. C., & Ostriker, J. P. Journal, 327, 570 Ba~ nados, E., Venemans, B. P., Mazzucchelli, C., et al. 2018, Nature, 553, 473, doi: J. J., Hennawi, J. F., Simcoe, R. A., et al. of the Astronomical Society of the Pacic, 121, 1409, doi: J. S., & Haehnelt, M. G. 2007, Monthly Notices of the Royal Astronomical Society: Letters, 381, doi: J. S., Haehnelt, M. G., Warren, S. J., et al. 2011, Monthly Notices of the Royal Astronomical 416, 70, doi: S. E. I., Kakiichi, K., Meyer, R. A., et al. 2020, The Astrophysical Journal, 896, 49, doi: S. E. I., Durov c kov a, D., Davies, F. B., & Eilers, A.-C. 2021, MNRAS, 503, 2077, doi: R., & Safarzadeh, M. 2015, Astrophysical Journal Letters, 798, 2, doi: S. A., Shephard, M. W., Mlawer, E. J., et al. 2005, Journal of Quantitative Spectroscopy and 91, 233, doi: C., & White, M. 2013, Astrophysical Journal, 762, doi: Morey et al. Davies, F. B., Furlanetto, S. R., & McQuinn, M. 2016, Monthly Notices of the Royal Astronomical Society, 457, 3006, doi: F. B., Hennawi, J. F., & Eilers, A.-C. 2019a, Monthly Notices of the Royal Astronomical Society, 14, 1, doi: 2019b, The Astrophysical Journal, 884, L19, doi: F. B., Hennawi, J. F., Ba~ nados, E., et al. 2018a, The Astrophysical Journal, 864, 143, doi: 2018b, The Astrophysical Journal, 864, 142, doi: R., Walter, F., Venemans, B. P., et al. 2018, Journal, 854, 97, doi: Matteo, T., Springel, V., & Ilernquist, L. 2005, Nature, 433, 604, doi: G., & Rees, M. J. 1988, Monthly Notices of the Royal Astronomical Society, 230, 5P, doi: A.-C., Davies, F. B., Hennawi, J. F., et al. 2017, Journ, 840, 24, doi: A.-C., Hennawi, J. F., & Davies, F. B. 2018, Journal, 867, 30, doi: A.-C., Hennawi, J. F., Davies, F. B., & Simcoe, R. A. 2021, arXiv e-prints, A.-C., Hennawi, J. F., Decarli, R., et al. S. M., Phillips, A. C., Kibrick, R. I., et al. 2003, in Instrument Design and Performance for Telescopes, ed. M. Iye & A. F. M. Moorwood, Vol. 4841, International Society for Optics and Photonics (SPIE), 1657{1669, doi: X., Strauss, M. A., Becker, R. H., et al. 2006, The Astronomical Journal, 132, 117, doi: D., Hogg, D. W., Lang, D., & Goodman, J. 2013, Publications of the Astronomical Society of the Pacic, 125, 306, doi: K., Dodson-Robinson, S., & Kraus, A. 2014, The Astronomical Journal, 148, 53, doi: F., & Madau, P. 2012, The Astrophysical Journal, 746, 125, doi: M., Volonteri, M., Latif, M., Dubois, Y., & Peirani, S. 2016, Monthly Notices of the Society, 463, 529, doi: Z., & Cen, R. 2001, in Astronomical Society of the Pacic Conference Series, ed. M. Umemura & H. Z., & Hui, L. 2001, The Astrophysical Journal, 547, 27, doi: J. F., & Prochaska, J. X. 2007, The 655, 735, doi: P. F., Hernquist, L., Cox, T. J., et al. 2005, Journal, 630, 716, doi: K. 1986, Publications of the Astronomical Society of the Pacic, 98, 609, doi: K., Visbal, E., & Haiman, Z. 2020, Annual Review of Astronomy and Astrophysics, 58, 27, doi: L. C., Haehnelt, M. G., Cantalupo, S., & Puchwein, E. 2015, Monthly Notices of the Society, 454, 681, doi: I. S., Hennawi, J. F., McQuinn, M., & Worseck, G. 2016, The Astrophysical Journal, 824, 133, doi: I. S., Hennawi, J. F., & Worseck, G. 2019, Monthly Notices of the Royal Astronomical Society, 484, 3897, doi: I. S., Hennawi, J. F., Worseck, G., & Davies, F. B. 2021, Monthly Notices of the Royal 505, 649 Lidz, A., McQuinn, M., Zaldarriaga, M., Hernquist, L., & Dutta, S. 2007, The Astrophysical Journal, 670, 39, doi: G., & Natarajan, P. 2006, Monthly Notices of the Royal Astronomical Society, 371, 1813, doi: c, Z., Stark, C. W., Nugent, P., et al. 2014, Monthly Notices of the Royal Astronomical Society, 446, 3697, doi: E., Worseck, G., Hennawi, J. F., et al. 2015, Monthly Notices of the Royal Astronomical Society, 449, 4204, doi: P. 2004, Coevolution of Black Holes and University Press), 169 Martini, P., & Weinberg, D. H. 2001, The 547, 12, doi: C., Ba~ nados, E., Venemans, B. P., et al. 2017, The Astrophysical Journal, 849, 91, doi:

2108.10907

Karna Morey

Karna A. Morey, Anna-Christina Eilers, Frederick B. Davies, Joseph F. Hennawi, and Robert A. Simcoe

Estimating the effective lifetime of the $zsim6$ quasar population from the composite proximity zone profile

15 pages, 7 figures, accepted by ApJ

10.3847/1538-4357/ac1c70

astro-ph.GA astro-ph.CO

http://creativecommons.org/licenses/by/4.0/

The lifetime of quasars can be estimated by means of their proximity zone sizes, which are regions of enhanced flux bluewards of the line observed in the rest-frame UV spectra of high-redshift quasars, because the intergalactic gas has a finite response time to the We estimate the effective lifetime of the high-redshift from the composite transmitted flux profile within the proximity zone region of a sample of $15$ quasars at $5.8leq zleq 6.6$ with precise systemic redshifts, and similar luminosities, i.e. and thus a similar instantaneous ionizing power. We develop a Bayesian method to infer the effective lifetime from the composite robust estimates of various sources of uncertainty on the spectrum. We estimate an effective lifetime of the quasar population as a whole = 5.7^{+0.5 (+0.8)}_{-0.3 (-0.5)}$ given by the median and $68$th ($95$th) percentile of the posterior probability distribution. While our result is consistent with previous quasar lifetime studies, it challenges on the current model for the growth of holes (SMBHs) located in the center of the quasars' host galaxies, which requires that quasar lifetimes are more than an order of magnitude longer.

2 tions established within the clock model. Finally, in Sec. V we summarize the main findings reported in the article. II. CHIRAL CLOCK MODEL: QUANTUM PHASE TRANSITION IN THE ISOLATED SYSTEM Let us start by considering the ZNsCCM for Mquantum rotors [30, 47-50], i.e. Mquantum systems with levels. Each individual rotor can be seen as a spin- (Ns1)=2 or as particles occupying the Nsvertexes of a regular polygon. Letfjkiigdenote the orthogonal basis of the Hilbert space of the ithrotor, with k=0;:::Ns1 corresponding to the Nsdirections along which the angular momentum can point (or vertices of the polygon), with jk+Nsii=jkii. The CCM is then described by the fis the control parameter that accounts for the rela- tive weight between the free and interaction terms, and 'jthe so-called chiral phases. We assume periodic boundary condi- tions, so that M+1=1. Here the operators and, in the vertexes are defined 0 0


0 0!0


0 0 0!2


0 0 0 0::: 0 0 0 0 0  1 0 0


0 0 0 1 0


0 0 0 0 1


0 0 0 0 0:::0 0 0 0 0


1 1 0 0 The first term of the Hamiltonian encodes the dynamics of the individual rotors and gives rise to tunnel- ing currents between their internal levels (cf. Fig. 1). For a particle at the vertices of a regular polygon, the tunneling cur- rents can be visualised as describing the hopping of the single system between such vertices induced by the rotor The second term in the Hamiltonian encodes the interaction between nearest neighbors. The model possesses a global ZNssymmetry and, classi- cally, presents two phase transitions in 2D [51]. The inter- action potential breaks a specific rotational symmetry as discussed in Refs. [30, 34]. This is a necessary condition for the emergence of the currents, as we will also see in the following (cf. Sec. III A). In this context, the order parameter of the model is the total magnetization From now on, we will focus on the minimal configuration allowing for namely the case of Ns=3. It should be noted that, our model is similar to the one investigated in Ref. [50] where the role ofandwas interchanged. In Refs. [49, 50], the structure of the phase diagram of the CCM with Ns=3 and homogeneous chiral phase 'j='was investigated in detail, showing that for small values of 'there is a direct transition from the ordered ( f1=2) to a disordered phase ( f1=2). For large chirality ( '>= 6), the two phases are separated T 1FIG. 1. Schematic representation of the CCM with periodic bound- ary conditions, interacting with thermal reservoirs with T1andT2, and chiral hopping interactions governed by the phase 'j. an incommensurate phase. In Appendix A we provide a brief summary of the QPT taking place in the ground state of Hccm, while we refer to Refs. [49, 50] for a thorough inspection of the model's critical features. III. OPEN SYSTEM DYNAMICS: CORRELATIONS VS CURRENTS We are interested in exploring the physics of the CCM when interacting with thermal baths. In particular, we consider the case in which each rotor is in contact with an reservoir and partition our system in two of even (e) and odd (o) rotors, respectively. The in- verse temperature of the two sub-lattices is set to be and we will assume e,o, in general, thus realizing a configuration (cf. Fig. 1 for a schematic illustration). As it will be shown later on in this Section, the temperature di erence gives rise to ther- mally driven mechanical currents in the system that are sus- tained asymptotically in time. The system thus evolves to- wards a non-equilibrium steady-state (NESS), whose proper- ties we now aim at describe the open system dynamics via the local (GKLS) master (3) with local in terms of the jump operators jMiwith jk=0;


;Ns1. The transition rates the local detailed (5)3 where Ej=hjjHccmjji. For a generic bosonic bath, we with 0;(6) and where gis a microscopic rate. Before proceeding fur- ther, a note is in order. As it is well known, local master equations can be problematic from a thermodynamic point of view [53, 54]. However, it should be noted that this conclusion has been recently challenged by a stream of works [34, 55-57] pointing towards a reconciliation of local master equation In particular, it has been shown that the lo- cal master equation is not, in general, at odds with the second law of thermodynamics as far as the proper expression for the heat currents is considered. In the specific case under study, we can split the Hamiltonian in its diagonal and in thefjjigbasis as Hccm=HD+HND, which allows us to introduce the individual energy the dual ofDm. It is useful to remark that the standard definition of heat flux when dealing with a local mas- ter equation would read using .Qmleads in general to violations of the second law of thermodynamics (cf. Ref. [53] for an exam- ple). However, it is the weighted sum of .QD;m's that enters the second law of thermodynamics and gives a positive rate . = the second law [34], and one should really focus on the individual currents. On the other hand, the .QND;mcan be as- sociated to a work rate within a microscopic collisional model framework [57]. For further details on this contruction we re- fer the interested reader to [34]. A. NESS of the GKLS Master Equation and Particle Currents From the numerical diagonalization of the Lioville super- operator on the right-hand-side of the GKLS master equation, we obtain the unique steady-state SSof the CCM interact- ing with independent thermal baths. Such state is in general a NESS, however its nature is determined by the choice of pa- rameters of the model. Note that, although the ground a QPT, such abrupt transition is blurred in this open quantum system setting. In order to quantify nature of the steady-state we turn to look at quantum particle currents in the system. The definition of quantum particle currents in general is a non trivial task. A formal characterisation has been carried out in [30] where the authors also investigate a CCM with M=2 rotors. For a classical particle hopping on a graph, one can readily define the probability current between any two vertices on the graph which j0pj0; (8) -5 0 5 10 15 0 0.2 0.4 0.6 0.8 1(a) (b) (c) m f1 2 3 4 -5 0 5 10 15 0 0.2 0.4 0.6 0.8 1<Jth> 0 2 4 6 0 0.2 0.4 0.6 0.8 1<Jtun> m ph/p1 2 3 4 -6-4-2 0 2 4 6 0 0.2 0.4 0.6 0.8 1<Jth> m ph/p1 2 3 4-1 0 1 0 0.05 0.1-1 0 1 0 0.05 0.1FIG. 2. Illustration of the non-trivial behavior of the NESS tunneling and thermal currents each rotor in a chain with M=4 and Ns=3. Panels (a) and (b) [(c) and (d)] illustrate such currents as functions of with g=0:2. Panels (a) and (b) have been obtained taking fixed staggered chiral phases 'j=(1)j=2, while the insets show the behavior of currents in the region close to f=0. Note whilehJth 0!1i,0 for f=0. Panels (c) and (d) have been obtained taking f=1=2, and'j=(1)j', here the currents display a 2=Nsperiodicity in '. See main text for further details and Fig. 3 for the total currents. where pjis the instantaneous probability of finding the par- ticle at vertex j, and Wj j0is the transition rate from j0toj. In Ref. [30], the quantum analogous of this classical current was defined as the sum of the tunnelling and thermal (10) where xj=jjihjjis the projector onto a generic state jji, and the sum runs over all the possible transitions between pair of states of the system. Note that the thermal current reduces to the classical probability current (8) in the classical limit. Fur- thermore, in order to simplify the analysis of the dynamics, in the following we will only allow jumps between states where only one spin is rotated, that is proceeding further, we shall discuss the of these NESS currents. In order to simplify the notation, in the following we will omit the subscript j!j0in the steady state currents hJtun;th j!j0i, as in the NESS the of the mthrotor are independent of the specific initial and final position considered. For '=k=Ns(k2Z) the cur- rents for each individual rotor m=1;:::; M,8fand regardless of the temperature di erence among sub-lattices 0 0.1 0.2 0.3 0.4 0 0.2 0.4 0.6 0.8 1(a) (b) (c) T fM=4 M=6 0 0.5 1 1.5 2 0 0.2 0.4 0.6 0.8 0 0.1 0.2 0 0.2 0.4 0.6 0.8 1f=1/2 f=1/2<Jtun> 0 0.5 1 0 0.2 0.4 0.6 0.8 1<Jth> 3. Total NESS currents and M=6 rotors ande=1,o=1:1 with g=0:2. Panels (a) and (b) are for 'j=(1)j=2, while panels (c) and (d) are for f=1=2. In if',k=Nsand f,0;1, when
T=0. In a similar fash- ion, the thermal current fulfills for ',k=Nsand f,1, whilehJthiT=0 for
T=0. Fig. 2 show the individual steady-state currents as a function of for each of the individual rotors in a CCM with and g=0:2, which already reveal a non-trivial behavior. A similar behav- ior is found for di erent parameter combinations. Inspection of Fig. 2 suggests that, as a function of f, the maximum of jhJtunimjis reached at f'0:45, which is very close to the value f'0:46 at which a QPT occurs in the ground state of the CCM at thermodynamic limit [cf. Appendix A]. How- ever, the thermal current hJthimis maximized for a slightly smaller value of f. Also, the insets in Fig. 2 (a) and (b) show that, in the classical limit f=0, the tunnelling current is van- ishing in both sub-lattices, while the rotors exhibit the same non-zero thermal rotational frequency. Although not explic- itly shown, the mean square value of the thermal current has a maximum at f1=2 while the analogous quantity for the thermal current gets the value of 2f2=3 independently of '. This value suggests that all clock states are equally popu- lated at the NESS. However, as shown in Sec. IV, such state is not a maximally mixed one as it brings about coherence and non-trivial correlations among the individual rotors. Fig. 3 also illustrates the total currents hJtun;thiTfor the same parameters as Fig. 2 for M=4 and 6 rotors and
T,0. Note that for fixed Mthe total thermal current is larger than the tunnelling one. Furthermore, for the two sizes here almost constant for increasing number of ro- tors, its value suggesting that for large Mthe total tunneling current will be negligible with respect to the thermal one. Hence, in the thermodynamic limit, one should 0 1 2 0 0.2 0.4 0.6 0.8 1oddeven(a) (b)Q. D fbe=1 be=0.2 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1Q. 4. Diagonal (a) and non-diagonal (b) heat currents, as defined by Eqs. (7), respectively, as a function offfor the NESS of the CCM with M=4 rotors and staggered chiral phases 'j=(1)j=2 and temperatures. We have taken o= 1:1,g=0:2, ande=0:2 and 1, as specified in the legend of the figure. In panel (a) we show the heat currents for each even and odd rotors. For e=1,.QD<0 for both sub- lattices, which implies a large non-diagonal heat current. Note now turn our attention to the steady state heat currents, as given by Eqs. (7). Here the heat currents are positive when flowing from the bath(s) to the system. The results, for two dierent sets of system parameters, are shown in Fig. 4. As previously done, we have chosen the even sub-lattice to be in contact with the hot bath. We observe that, for a gradient, the diagonal heat currents are both neg- ative. This can be understood as follows. First, the first law - written in the - is valid. Sec- ond, we recall that the non-diagonal heat current within the framework of the collisional model, to the work done or produced when switching on and o the inter- action of the system with the colliding particles making up the environment [34, 57]. Thus, the situation in Fig. 4 where .QD;m<0 for all rotors is compensated by a large and corresponding to a net amount of work done on the system that is then dissipated in both the cold and hot baths. One can understand this result also noticing that, when f>0 the Hamiltonian in Eq. (1) is not diagonal in the basis jji. Thus, Eq. (3) will introduce coherence in the steady state, resulting in a non-zero non-diagonal heat current, as given by the second line of Eq. (7). For a larger temperature gradient, and one of the two tem- peratures relatively high, the heath currents exhibits a more classical behaviour with a net diagonal current from the hot to the cold baths and a reduced non-diagonal heat current. In the ground state of the system (1) the thermally for any f: At T=0 there is no heat current to sustain the rotational motion. However, the tun- nelling in principle be non-vanishing in the ground state: Eq. (9) is indeed the discrete counterpart of the Schr odinger probability current, as discussed in [30]. We find nevertheless that also hJtunimvanishes in the ground state of (1) for any f. In Appendix B we consider the rotated model of (1) with !and!, and interestingly find that the hJtunimshown a critical-like behaviour in the ground state, be- ing non-zero for f.fc.5 IV . CONNECTING CURRENTS TO THEORETIC QUANTITIES The connection between the location, in parameter space, of the quantum critical point of the CCM and that of the op- timal particle currents is suggestive of a potential role of col- lective quantum phenomena in the establishment of the non- equilibrium features of the system. In this Section we explore such suggestion further by making use of a toolbox of infor- mation theoretic figures of merit that have been used, in the past, to explore the interplay between quantum critical phe- nomena and non-classicality [46, 58, 59]. In doing so, we un- veil the intrinsically collective nature of the features that have been highlighted in our analysis so we will consider the von Neumann entropy of a subsystem Aof a compound A[B, which is defined denotes the partial trace over B. Another relevant measure is the negativity NA[60], which is able to quantify entanglement and is given the spectral decomposition of the partially transposed state with respect to subsystem A. The total amount of correlations (classical and quantum) shared between the bipartitions AandBcan be quantified using on the mutual (13) where SA[Bis the von Neumann entropy of the state of the whole compound. In addition, we shall compute the coher- ence of the system state using the L1norm the density matrix entries in the clock-state ba- sis. Finally, we will use the quantifier of multipartite quan- tum correlations provided by the so-called global quantum discord [46, the reduced state of the ithrotor, ()=P jjjis a projector operator acting on the global state, andi(i) the corresponding projector acting on the single- rotor states. Following Ref. [58], we choose a collection of single-particle rotation op- erators, while the operator acting on the ithrotor (16) where i;na) is a vector of naangles, is a vector of generators of rotations 0.8 1.2 1.6 2 2.4 0 0.2 0.4 0.6 0.8 1(a) (b) (c) (d)SA fph=0, be=1 ph=p/2, b e=0.2 ph=p/2, b e=1 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1I(A:B) f 0 1 2 3 4 0 0.2 0.4 0.6 0.8 1C f 0 1 2 3 4 5 0 0.2 0.4 0.6 0.8 1G (x103) fFIG. 5. Information measures of the NESS for M=4 rotors as a function of the control parameter f. Panels (a), (b), (c) and (d) show the von Neumann entropy SA, mutual information Cand global discord, respectively, with Adenoting the first half of the chain, namely, rotors 1 and 2. Di erent points (and colors) correspond to distinct values of the staggered chiral phase, such that'j=(1)j', and local temperature e, while g=0:2 and o=1:1. See main text for further details. for the single rotor. We have considered the na=8 Gell- Mann 33 matrices as generators of rotations. The mini- mum in Eq. (15) is obtained by varying the set of angles fig, i=1;:::M, through an annealing algorithm. These instruments are all very informative of the quantum critical features of the ground-state QPT [44, 45] in the CCM (cf. Appendix A). Here however we are mainly interested in the NESS properties: In such an open quantum system, critical features become blurred or disappear altogether. This might lead one to naively think that no connection could be established. Yet, the interplay between temperature gradient between sub-lattices, currents, and correlations reveal a Fig. 5 shows the behavior of these quanti- ties for di erent parameters. Contrary to the CCM ground- state, these quantities show a smooth dependence on f, which suggests that it is not in quantities that a behavior reminiscent of a critical one should be sought. How- ever, it is interesting to observe that both SAandI(A:B) have an inflexion point in the region where we expect the crit- ical value of fto occur, which indicates a qualitative change in trend taking place around f'0:46. On the other hand, the global quantum discord shows the quantumness acquired by the NESS away from f=0;1, which correlates with the amount of coherence C. However, while the coherence for the chiral model becomes maximum at f'0:46, the global discord peaks at a slightly smaller value of f. This should be compared with figures 2 (a) and (b), and 3-(a) and (b), showing a similar behaviour for the tunnelling and the ther- mal current. This suggests that the amount of 0 0.1 0 0.2 0.4 0.6 0.8 1(a) be=1 phj=(-1)jp/2, T/DT 0 0.05 0.1 0 0.2 0.4 0.6 0.8 T/DT x10 FIG. 6. Mutual information and current a function of the control parameter for M=4 rotors and'j=(1)j=2. Panel (a) and (b) corresponds to e=1 and e=0:5, respectively, such that o=1=(Te+
T) with The current susceptibility is multiplied by a factor 10 for a better representation. Both quantities feature a qualitatively measured by the coherence and the global discord, is an essential ingredient for the CCM to work as a thermal machine, thus converting thermal currents into mechanical currents. Note however, that although SScontains coherence for f,0;1, its maximum value is significantly smaller than in the ground-state where Fig. 8(d)]. Similar behavior is observed for other choices of ', also for '=k=Nswith k2Z. In addition, all these quantities inherit the periodicity 2 =Nsin the phase '. Finally, we stress that NA=08f;', in contrast to the ground-state negativity (cf. App. A). The observed behavior of the mutual information correlates with that of the total current Fig. 3(b) and Fig. 5(b)). Building on this observation, we investigate the thermal susceptibility of the total current of the mutual information,
I(A:B)=
Twhen the in- crement in the total current between 0 <j
Tj1 and
T=0 (equal temperatures for both sub-lattices), and equivalently for the mutual information. Note that since In order to il- lustrate this susceptibility, we fix e=1=Teand change o=1=(Te+
T) forj
Tj1. In Fig. 6 we show two exam- ples of the mutual information and current susceptibility for
T=0:001 and di erent Te, for a fixed 'j=(1)j=2. feature a qualitative similar behavior, as well as for other choices of the parameters. We stress however that for di erent choices of '(oro) one may revert their relative sign. In addition, one should note that the total current van- ishes for'=k=Nswith k2Z, while the mutual information does not. V . CONCLUSIONS We have addressed the link between the emergence of NESS currents in a chiral few-body model and critical features of the corresponding model at the ther-modynamic limit: The response of the system, in terms of currents, is maximum at the working point where a QPT is predicted to occur. This is also well captured by the behav- ior of genuinely multipartite information theoretic quantities, such as global quantum discord, and provides strong numeri- cal evidences of the possible role that collective quantum phe- nomena play in the non-equilibrium response of this inter- esting interacting model. Such link will be explored further in future works through the investigation of possible e ects in work-extraction games aimed at achieving ergotropic per- formance from the that we have addressed the investigation of the dynamical properties of the ground state of the model (1) and its variations is an interesting open question. In particular we find that can arise in a rotated version of (1) with a finite num- ber of rotors. Whether such currents persist in the thermo- dynamic limit and exhibit a critical behaviour are questions worthy of future gratefully acknowledges the financial support of The Faculty of Science and Technology at Aarhus a Sabbatical scholarship and the hospitality of the Quantum Technology group, the Centre for Molecular and Optical Physics and the School of Mathematics and Physics, during his stay at Queen's Univer- sity Belfast. AB acknowledges the hospitality of the Institute for Theoretical Physics and the Nonequilibrium quantum dy- namics group at Universit at Stuttgart, where part of this work was carried out. RP and MP acknowledge the support by the SFI-DfE Investigator Programme (grant 15 /IA/2864), the Ero- pean Union's Horizon 2020 FET-Open project and TEQ (766900). MP acknowledges support by the Leverhulme Trust Research Project Grant UltraQuTe (grant RGP-2018-266) and the Royal Society Wolfson Fel- lowship (RSWF /R3/183013). AB also acknowledges support from H2020 through the MSCA IF pERFEcTO (Grant Agree- ment nr. 795782) and from the (DFG, German Research Founda-tion) project number BR 5221 /4-1. Appendix A: Critical ground state features in the chiral clock model As already noted, the CCM exhibits a ZNssymmetry. In or- der to exploit this symmetry, it is handy to remap the Hamil- tonian as!and!(as in Ref. [50]), so that (A1) allows us to split the Hilbert space in Nssub-spaces. The ground state is contained in the subspace with eigenvalue 0 0.2 0.4 0.6 0.8 1(a) 0.44 0.46 0.48B f FIG. 7. (a) Binder cumulant Bfor the ground state of the CCM with Ns=3 for di erent system sizes (from M=4 to 10 rotors) as func- tion of fand with'j==8 (solid lines) and staggered chiral (dashed lines). The vertical dotted line indicates the critical value fcwhere the QPT takes place, reported in Ref. [50]. Panel (b) shows a zoom close to the region where Binder (close to fc). For homogeneous chiral phase, the the cross- ing approaches the reported value fc. The intersections in Bfor stag- gered chiral phases suggests that location of fcis shifted to a slightly larger value. In the case of Ns=3, the operator reads as U=  ndenotes the projector on the cor- responding subspace. One can use this symmetry to reduce the dimension of the Hilbert space. In  0Hccmy 0: (A2) contains the ground state 8fwith a well defined symme- try. For f<fc, the ground state becomes Ns-fold degen- erate. In [50] the ground-state energy critical exponents of theHccmwere investigated. For completeness, here we just provide a brief summary of the critical features of such a model. In particular, note that the ground-state order within the 0subspace is the ground state of H0 ccm. As custom- ary in phase transitions, one needs to re- sort to m2andm4, which clearly reveal the for f<fc(and thus the QPT). Moreover, the location of the QPT can be witnessed by looking at the energy gap
or Binder cumulant B[62]. The energy gap between the ground and first excited state closes at fcfollowing the universal scal- ing law [43]
jffcjz: (A4) where zare critical exponents of the QPT. The Binder cumu- lant is defined as [62] B=1 2 evaluated over the ground state, This quantity has been proven very useful to locate the critical point fc(see for exam- ple Refs. [63, 64]). Applying finite-size scaling arguments, B is expected to become size independent at fc. Hence, the QPT takes place at the value of fat which the Binder cumulant Bfor di erent system sizes Mintersect, although still yield small deviations to the In Fig. 7 we show the resulting Binder cumulant Bfor the ground state of the CCM for Ns=3 for the case of a staggered and homogeneous chiral phase 'j=(1)j=8 and 'j==8, respectively. The location of the QPT, i.e. fc, for the homogeneous chiral phase is consistent with the reported value in Ref. [50], fc=0:46267 which is indicated by a dot- ted vertical line, while fcappears to be shifted to a slightly larger value for a staggered chiral phase. The signatures of the QPT are already evident even for the considered system sizes M.10. In addition, in Fig. 8 we show the quantum on the CCM ground-state as a function of the con- trol parameter f, namely, von Neumann entropy SA, negativ- ityNA, coherence C, mutual information I(A:B) and global quantum discord G. The system is split in half, so that A refers to the first two rotors for M=4. All the a QPT taking place at f0:46 [50]. Compare results with those discussed in the main text for the NESS. It is worth noting that the non-chiral model 'j=0 has a critical field fc=1=2 [50], thus the chirality lowers the value of the field required to achieve the disordered phase, as one would expect. Appendix B: Tunneling current in a rotated CCM model As commented in the main text, while the of the CCM model remain unaltered upon the the tunneling current di erent. Note that the definition of the tunnel- ing current Jtun(j!j0) given in Eq. (9) is independent of the specific choice of the Hamiltonian. In particular, staggered chiral phases 'j=(1)j', the tunneling cur- rent in its ground state is non zero. Moreover, the behavior of Jtunresembles that of a critical quantity across a phase transi- tion. This is plotted in Fig. 9 for Mfrom 4 to 12 rotors, which indicate a sharp transition around the QPT.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1(a) (b) (c) (d) (e)SA fM=4 M=6 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1NA fM=4 M=6 0 0.5 1 1.5 2 2.5 3 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1C/(N sM) fM=4 M=6 0 0.5 1 1.5 2 2.5 0 0.2 0.4 0.6 0.8 1G fM=4 M=6 FIG. 8. Quantum information measures of the ground state of the CCM with periodic boundary conditions and staggered chiral phase 'j= (1)j=2 for M=4 and 6 rotors, which unveil the QPT taking place in the system. From left to right, von Neumann entropy SA, negativity NA, mutual information I(A:B), coherence C(rescaled over the total Hilbert space dimension NM s), and global discord G, respectively. The system is split in half, so the partition Aincludes the first M=2 rotors, namely, rotors 1 and 2 for M=4 and 1, 2 3 for M=6. 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1(a) (b)<Jtun> 0 0.1 0.2 0.3 0.4 0.4 0.5 0.6<Jtun> FIG. 9. (a) Tunneling current for the odd rotors hJtuniin the ground state of the CCM ~Hccm, given in Eq. (B1), with 'j=(1)j=2 and as a function of f. The tunneling current for even rotors is reversed in sign, i.e.,hJtuni. Panel (b) shows a zoom closer to the transition point to signal the sharper behavior of T. B. Batalh ~ao, A. M. Souza, L. Mazzola, R. Auccaise, R. S. Sarthour, I. S. Oliveira, J. Goold, G. De Chiara, M. Paternostro, and R. M. Serra, Phys. Rev. Lett. 113, 140601 (2014). [2] F. Brand ~ao, M. Horodecki, N. Ng, J. Oppenheim, and S. Wehner, PNAS 112, 3275 (2015). [3] A. M. Alhambra, L. Masanes, J. Oppenheim, and C. Perry, Phys. Rev. X 6, 041017 (2016). [4] M. Brunelli, L. Fusco, R. Landig, W. Wieczorek, J. G. Landi, F. L. Semi ~ao, A. Ferraro, N. Kiesel, T. Donner, G. De Chiara, and M. Paternostro, Phys. Rev. Lett. 121, 160604 (2018). [5] G. T. Landi and M. Paternostro, (to appear in Rev. Mod. (2020). [6] R. Koslo , J. Chem. Phys. 80, 1625 (1984). [7] N. Linden, S. Popescu, and P. Skrzypczyk, Phys. Rev. Lett. 105, 130401 (2010). [8] O. Abah, J. Rossnagel, G. Jacob, S. De ner, F. Singer, and E. Lutz, Phys. Rev. Lett. 109, 203006 (2012). [9] J. Rossnagel, O. Abah, F. Schmidt-Kaler, K. Singer, and E. Lutz, Phys. Rev. Lett. 112, 030602 (2014). [10] J. Rossnagel, S. T. Dawkins, K. N. Tolazzi, O. Abah, E. Lutz, F. Schmidt-Kaler, and K. Singer, Science 352, 325 (2016). [11] J. P. S. Peterson, T. B. Batalh ~ao, M. Herrera, A. M. Souza, R. S. Sarthour, I. S. Oliveira, and R. M. Serra, Phys. Rev. Lett. 123, 240601 (2019). [12] D. von Lindenfels, O. Gr ab, C. T. Schmiegelow, V . Kaushal, J. Schulz, M. T. Mitchison, J. Goold, F. Schmidt-Kaler, andU. G. Poschinger, Phys. Rev. Lett. 123, 080602 (2019). [13] R. Dorner, J. Goold, C. Cormick, M. Paternostro, and V . Ve- dral, Phys. Rev. Lett. 109, 160601 (2012). [14] G. Watanabe, B. P. Venkatesh, P. Talkner, M.-J. Hwang, and A. del Campo, Phys. Rev. Lett. 124, 210603 (2020). [15] A. Hartmann, V . Mukherjee, W. Niedenzu, and W. Lechner, Phys. Rev. Research 2, 023145 (2020). [16] Z. Fei, N. Freitas, V . Cavina, H. T. Quan, and M. Esposito, Phys. Rev. Lett. 124, 170603 (2020). [17] T. Fogarty and T. Busch, Quantum Sci. Technol. 6, R. B. S, V . Mukherjee, U. Divakaran, and A. del Campo, Phys. Rev. Research 2, 043247 (2020). [19] R. Filliger and P. Reimann, Phys. Rev. Lett. 99, 230602 (2007). [20] V . Dotsenko, A. Maciolek, O. Vasilyev, and G. Oshanin, Phys. Rev. E 87, 062130 (2013). [21] A. Argun, J. Soni, L. Dabelow, S. Bo, G. Pesce, R. Eichhorn, and G. V olpe, Phys. Rev. E 96, 052106 (2017). [22] V . Mancois, B. Marcos, P. Viot, and D. Wilkowski, Phys. Rev. E97, 052121 (2018). [23] H. C. Fogedby and A. Imparato, EPL (Europhysics Letters) 119, 50007 (2017). [24] A. Imparato, J. Stat. Mech. Theory Exp. 2021 , 013214 (2021). [25] A. Mari, A. Farace, and V . Giovannetti, J. Phys. B At. Mol. Opt. Phys. 48, 175501 (2015). [26] A. Roulet, S. Nimmrichter, J. M. Arrazola, S. Seah, and V . Scarani, Phys. Rev. E 95, 062131 (2017).

2108.10955

Ricardo Puebla

Ricardo Puebla, Alberto Imparato, Alessio Belenchia, Mauro Paternostro

Open Quantum Rotors: Connecting Correlations and Physical Currents

9 pages, 9 figures; comments welcome!

Phys. Rev. Research 4, 043066 (2022)

10.1103/PhysRevResearch.4.043066

quant-ph cond-mat.stat-mech

http://creativecommons.org/licenses/by/4.0/

We consider a finite one-dimensional chain of quantum rotors interacting with a set of thermal baths at different temperatures. When the interaction between the rotors is made chiral, such a system behaves as an autonomous thermal motor, converting heat currents into non-vanishing rotational ones. Such a dynamical response is strongly pronounced in the range of the for which the ground state of the system in the thermodynamic limit exhibits a quantum phase transition. Such working points are associated with large quantum coherence and multipartite quantum correlations within the state of the system. This suggests that the optimal operating regime of such motor is one of maximal quantumness.

2 6. The Blackett Laboratory, Department of Physics, Imperial College London, London SW72AZ, United Kingdom 7. Departamento de Fisica, Facultad de Ciencias Exactas y N aturales, Universidad de Buenos Aires, Int. Guiraldes 2620, CABA, Argentina -muenchen.de ABSTRACT Several fields of applications require a reliable characterization of the photothermal response and heat dissipation of nanoscopic systems , which remains a challenging task both for modeling and experimental measurements . Here, we present a new implementation of anti-Stokes thermometry that enables the in situ photo thermal characterization of individual nanoparticles (NPs) from a single hyperspectral confocal image . The method is label -free, applicable to any NP with detectable anti -Stokes emission , and does not require any prior information about the NP itself or the surrounding media . With it, w e first studied the photothermal response of spherical gold NPs of different sizes on glass substrates, immersed in water , and found that heat dissipation is mainly dominated by the water for NPs larger than 50 nm . Then, the role of the substrate was studied by compar ing the photothermal response of 80 nm gold NPs on glass with sapphire and graphene , two materials with high thermal conductivity. For a given irradiance level, the NPs reach temperatures 18% lower on sapphire and 24% higher on graphene than on bare glass. The fact that the presence of a highly conductive material such as graphene leads to a poorer thermal dissipation demonstrates that interfacial thermal resistances play a very significant 3 role in nanoscopic systems , and emphasiz e the need for in situ experimental thermometry techniques. The developed method will allow addressing several open questions about the role of temperature in plasmon -assisted applications, especially ones where NPs of arbitrary shapes are present in complex matrixes and environments. KEYWORDS Anti-Stokes , nanothermometry, optical printing, metal , graphene, plasmonics , metallic nanoparticles . TEXT Plasmonic nanoparticles (NPs) are widely used in several fields of research due to their outstanding optoelectronic properties.1,2 Among these properties, they stand out for being remarkably efficient converters of light into heat,3,4 with absorption cross -sections up to several times their geometrical size. In some cases, plasmon -assisted heating is the primary reason for using plasmonic NPs ,5 such as for photothermal therapy,6,7 drug delivery and release,8 solar steam generation,9 or photothermal microscopy.10 Alternatively, heating can be a n undesired side effect in many other applications like ultra-sensitive (bio) sensing ,11,12 non-linear optics13 or integr ated optoelectronics .14 In some other areas , the role of th ermal effects remains unclear. For example, many efforts have recently been put in the field of plasmon assisted chemistry to disentangle thermal effects from other plasmon derived phenomena such as hot -carriers generation or electromagnetic field enhancem ent.15-19 A similar situation occurs in plasmon ic optical tweezers, where optical forces are non -trivially intertwined with thermal transport caused by thermophoresis, convection, Brownian motion , or All these application s of plasmonic NPs 4 have in common the necessity of accurate characterization of the NP photothermal and heat dissipation response. However, modeling or measuring heat transport in the nanoscale is not straightforward. The challenge in modeling resides in having an accurate description of the irregular inte rfaces, surface facets or molecular rendering most thermal simulations only approximate. Measuring the temperature of a nanometric object is also a challenging task.3 Conventional methods such as infrared lack the necessary spatial resolution. Scanning Thermal Microscopy reaches a 100 nm resolution using a miniature thermocouple at the tip of an atomic force microscope .28 However, it is technica lly complex, limited to surfaces, and invasive. Optical methods based on temperature -sensitive luminescen t probes are an appealing alternative. Organic dyes, rare -earth doped (nano)crystals, or quantum dots29 may be used to report temperature through changes in the ir spectrum,30 intensity,31 lifetime or anisotropy32. Yet another optical method relies on determining the temperature from variations in the refractive index .33 However, retrieving accurate temperature measurements of plasmonic NPs using these methods is not always possible . On the one hand, if the probes are not attached to the NP s, the limited spatial resolution leads to temperature values that are a weighted average of the NP and surrounding temperature .34 On the other hand, if the probes are attached to the NP surface, they may modify both the thermal dissipation and the a bsorption cross -section of the plasmonic NPs . Recently, a new optical thermometry method has emerged with the potential to overcome these issues , exploiting the (PL) of plasmonic NPs35. In particular, the NP temperature is determined by analyzing the temperature -dependent anti-Stokes (AS) emission in various ways22,36 -39. The photo thermal response of an array of gold nanodisks and inverted pyramids was measured by Xie et al.36 and Hugall et al.37, respectively . Later, Carattino et al.38 5 and Cai et al.40 studied individual supported gold nanorods obtained by colloidal synthesis. Jones et al.22 characterized single bowtie gold NPs fabricated on fused silica by electron beam lithography . Finally, Hogan et. al.41 studied Au nanocylinders on a Au film. These implementations of anti -Stokes thermometry differ not only on the target structures but also in the experimental methods and assumptions required for data modeling. Here, a new implementation of anti-Stokes thermometry that is universally applicable to any NP with detectable PL, and requires no assumption or previous characterization of the NP is presented . The method retrieves the photothermal coefficient of individual NPs from a single PL hyperspectral image . With it, a systematic study of the photothermal response of supported single gold NPs of different sizes in a water environment is performed . The role of the thermal conductivity of the substrate was investigated by performing measurements on NPs supporte d on glass, sapphire and graphene -coated glass substrates . RESULTS AND DISCUSSION Hyperspectral anti -Stokes nanothermometry Figure 1a summarizes the experimental set -up. Heating and PL excitation were performed simultaneously with a continuos wave 532 nm laser in a custom -made sample -scanning confocal microscope. The laser was focused near to its diffraction limit (beam waist 0=(342 +-5) nm). Scanning was achieved with a closed -loop piezoelectric stage. By a cquiring a single confocal hyperspectral PL image it is possible to probe the AS emission over a range of temperatures. During scanning , the NP is exposed to a range of irradiance levels (Figure 1b) . The integration time at each position was set to 2 s, well beyond the transient time required by the NP to reach 6 thermal steady state, which is in the order of ns to us. Thus, the measurement at each pixel corresponds to a constant steady -state temperature , with its tempera ture-dependent PL spectral signature, as depicted in Figures 1 c. Figure 1 : Method to obtain spectra vs. irradiance from a single hyperspectral image. (a) Scheme of the experimental set -up. BS: beam -splitter. PD: photodiode. NF: notch filter. FM: flippe r mirror. (b) Confocal image of an 80 nm Au NP color coded in terms of irradiance on the NP. (c) Schematic illustration of the heating of a NP during scanning, along with example PL spectra from pixels 1 to 3 marked in (b) . All measurements were performed far from saturation and a linear power dependence of Stokes PL emission was observed, typical of a one-photon process39 (Figure S1 ). Explo iting this fact, the irradiance of each pixel was determined as follows : the maximum irradiance of the central pixel was calculated taking into account the incident laser power and the beam geometry. The irradiance of the other pixels was assigned proportionally to their integrated Stokes signal. Then, pixels were grouped according to their irradiance in N equally distributed bins . Thi s improves the signal quality, especially for low irradiance (peripheral) pixels; they have a lower signal but are present 7 in larger numbers. Figure 2a shows a NP PL image color -coded in N=10 irradiance intervals, along with the corresponding PL spectra , background subtracted . The anti-Stokes (AS) and Stokes emission of each group (i) can be modeled as: (1a) (1b) where IiAS(,exc,T) and IiS(,exc) are the AS and Stokes PL emission at a wavelength , under excitation with a laser of wavelength exc and irradiance Iiexc. fPL(,exc) is the intrinsic PL emission spectrum, n(,exc,T) is the temperature -dependent distribution of states responsible for the anti -Stokes emission . Extracting the NP temperature , included in n, from equation (1a) requires knowledge of fPL, which is hard to calculate from first principles because that requires detailed knowledge about the mechani sm. Also , several experimental factors have been identified to influence fPL such as excitation wavelength,42 the electronic band structure of the material ,38 and the photonic density of states (PDOS)42. For th ese reason s, several experimental approaches ha ve been used to determine fPL in the context of AS thermometry . Carattino et al. proposed that in the case of Au nanorods fPL follows the surface plasmon resonance with a Lorentzian shape, and used the PL emission measured at a different excitation w avelength to find the parameters of the resonance .38 Recently, Cai et al. approximated fPL with the measured scattering cross section of Au nanorods .40 However, these approaches require additional for each NP and make approximations that are not generally valid for every plasmonic NP. For example , these approximations are not valid for nanospheres, where the plasmon resonances are typically broader and closer to interband transitions.43 As a result, the AS emission spectrum is highly dependent on the excitation wavelength44 and may not match the scattering spectrum.40 In order to derive the temperature from IiAS without needing an explicit 8 expression for fPL, a clever strategy was implemented by Xie et al.36 and followed by Jones et al.22 and Pensa et al.45 By taking the ratio Qi,jAS= IiAS IjAS between two spectra obtained at different temperatures, the factor cancels out. This holds if fPL does not depend on temperature, an assumption that is validated by the rather temperature -independent spectra of the Stokes emission (see Figure S2). The function n(T) expresses the energy distribution of the the rmally available states that provide the extra energy for AS emission. I t has been already established in the context of AS emission of plasmonic NPs under continuous wave excitation that n(T) is well described by a Bose -Einstein (BE) (2) Where kB is the Boltzmann constant and T is the temperature (differences between lattice and electronic temperature are negligible under the experimental conditions of this work , see Section 1.4 of the ESI). Furthermore, Hogan et al.41 used a n occupation function including a Bose -Einstein plus a Fermi -Dirac (FD) distribution to estimate the contribution from hot electrons . They f ound that the maximum contribution of hot electrons is smaller than 3% and that the FD term is only necessary to describe AS emission at energy shifts larger than 0.25 eV ( 2000 cm-1), which is usually beyond the range of detectable anti -Stokes emission of single NPs. Therefore, under the experimental conditions of this work , a BE distrib ution is a correct representation of n(T). Further discussion on alternative dis tribution functions can be found in Section 1.5 of the ESI. 9 Figure 2: Data analysis to obtain the photothermal coefficient from a single hyperspectral PL image. (a) PL spectra of an 80 nm Au NPs at ten different irradiance levels obtained from an hyperspectral confocal image (inset). (b) AS emission of each group normalized by their excitation irradiance. (c) Color solid lines: Ratios Qi,1AS of every AS spectra divided by the spectr um of group 1 (0.2 mW/um2). Dashed black lines: F its using Eq. 3. (d) Calculated temperature as a function of the excitation irradiance . Solid red line: mean photothermal coefficient <> =63 Kmm2mW/ with a standard deviation of 3 Kmm2mW/ . Grey lines: Values obtained for ten repetitions of the measurement . 10 Figure 2b shows the AS spectra IiAS, normalized by the incident irradiance of Iiexc. They do not overlap because they correspond to different temperatures of the Au NP. Then, ratio s Qi,jAS= IiAS IjAS between every pair of spectra are computed , which including equation (2) take the form : Qi,jAS= IiAS ) kBTj -1 eE()-E(exc ) kBTi -1 (3) Conveniently , this expression does not depend on the intrinsic PL emission spectrum fPL. If the range of explored temperatures is not so large, it can be assumed that the temperature of the NP increases linearly with excitation intensity . Ti=T0+ Iiexc (4) where T0 is the temperature of the particle in the absence of light and is the photothermal coefficient. Combining Eq. 3 and 4 gives Iiexc]-1 (5) which is an expression with two free parameters, T0 and . If one is known, the other can be calculated from a fit to the data . As an example, Figure 2c shows all ratios Qi,1AS using irradiance j=1 (the lowest) as a reference, together with fits to Eq. 5 using T0=295 K, the room temperature during measurements . For each pair of i rradiances {i,j} a value of i,j is extracted from a fit to Qi,jAS(). Then, the average value can be determined, which enable s the determination of the NP temperature as a function of the irradiance through Eq. 4. The reproducibility of the method was tested by determining from ten consecutive hyperspectral images of the same NP . Figure 2d shows the temperature vs. irradiance curves corresponding to each of the individual determinations 11 of , along with its mean value of <> =63 Kmm2mW/ . The standard deviation of the ten measurements was of 3 Kmm2mW/ (4%). In summary, the photothermal coefficient of single NP s can be determined from one hyperspectral image . In turn , delivers the NP temperature for any given irradiance. It is important to remark that no characterization of fPL was required, eliminating possible sources of error. For example, i f fPL is approximated by the scattering cross -section of the NP, very high and unrealistic temperatures are obtained (in the range of thous ands of degrees , see Section 1.3 on the ESI) . Using NPs as nanothermometers Hyperspectral images of 80 nm Au NPs were acquired at six different temperatures of the sample, Tset, ranging from 22 to 70 degC (details on Materials and Methods) . The obtained data was analyzed in the two possible ways. First, as before, was determined at the different temperatures by fitting Qi,jAS() to Eq. 5, while keeping T0=Tset as a fixed parameter. The resulting curves of temperature vs. irradiance corresponding to the obtained <> (10 images) are shown in Figure 3a. The slightly different slopes of the curves are in agreement with the variability observed at room temperature (Figure 2d). This is better seen in Figure 3b , where <> is plotted as a function of Tset. The variability of <> is contained within +- 1 standard deviation around the mean value obtained at room temperature , supporting the hypothesis of being independent of temperature (Eq. 3). The second way to analyze the data is to take as a known parameter and extract the surrounding temperature T0. In this way, if the photothermal coefficient of a NP is known, the NP can be used as a nano -thermometer. Figure 3c shows the determined T0 vs. Tset using <> obtained at room temperature f or this NP (62.5 Kmm2mW/ ). The determined temperature follows faithfully the externally set temperature, though it becomes less accurate at higher temperatures . This is 12 attribute d to larger measurement errors associated with heating the sample, such as fast er sample drift. Figure 3: Au NPs as (a) Temperature vs. irradiance for an 80 nm Au NP on a glass substrate, immersed in water at different temperature s Tset. (b) Ext racted photothermal coefficient vs. Tset. The red dashed line indicates <> at room temperature ( 62.5 Kmm2mW/ ) and the grey band corresponds to +-1 standard deviation s (4.6 Kmm2mW/ ). (c) Determined T0 vs. Tset using a fixed photothermal coefficient =62.5 Kmm2mW/ . Error bars in (b) and (c) represent the standard error of the mean. Size-dependent photothermal response of spherical Au NPs Next , the method was applied to characterize the photothermal response of spherical gold NPs of different sizes deposited onto a glass substrate and surround ed by water. NPs with four different nominal diameters of 48, 64, 80 and 103 nm were used . Arrays of individual Au NPs were fabricated through optical printing21,48 -50 (further information can be found in our previous works and in Materials and Methods ). The NPs of 48, 80 and 103 nm were whereas t he 64 nm ones were co nventional, commercially available NPs . An illustrative dark -field image of the optically printed Au NPs is show n in Figure 4a. Representative scattering spectra of individual NPs of each size are shown in Figure 4b (solid 13 lines ), along with the corresponding calculated spectra using Mie theory , considering the NPs immersed in water (no substrate). Figure 4: Photothermal response of spherical Au NPs as a function of size. (a) Dark -field image of spherical gold NPs optically printed on glass , immersed in water. Scale bar = 1 um. (b) Scattering spectra of single NPs of each size . Solid lines: Experimental data. Dashed lines: Mie theory considering the NPs in water (no substrate) . (c) Histograms of photothermal coefficients. The m ean (red dashed lines) and an interval corresponding to +-1 standard deviations (grey band) are shown . (d) Experimental and calculated photothermal coefficient vs. diameter . Error bars in diameter represent the standard deviation of NPs size as measured by TEM and the standard deviation of . 14 The photothermal coefficient was determined at least 40 times for each NP size, with a maximum of 90 times. Figure 4c shows the resulting histograms of . The width of the distributions is significantly larger than the variability in the determination of for a single NP, which reflects the size distribution of the NPs. The narrowest distributions of are obtained for 48 nm and 80 nm ultra-smooth NPs, consistent with their narrower size distribution, shown in Figure S 5. Figure 4d shows the mean <> vs. the diameter of the NPs. Further insight into the heat dissipation is obtained if is decomposed into its optical and thermal contribution s: ()=abs() (6) where abs is the absorption cross -section and the effective heat dissipation factor , which depends on the geometry and thermal properties of the system. The simplest way to model is to consider the NPs as a sphere with radius a, immersed in a homogeneous env ironment with thermal conductivity :3 =4a (7) The same holds for the optical properties of the NPs . The simplest way of estimating abs is through Mie theory , considering the NPs as gold spheres immersed in a uniform medium (no substrate). This approximate calculation reproduces well the scattering cross -section ( Figure 4b ). Dashed and dotted lines in Figure 4 d show the predicted values of for the spherical NPs in homogenous environment s, water or glass, using Mie theory for abs and Eq. 7 for . Remarkably , this simple calculation for water describes the experimental trend , especially for the large r NPs. This appears reasonable since for large NPs most of the surface is in contact with the water . A more accurate calcula tion of the absorption cross -section , considering the presence of the substrate, was performed using a finite element method solver (FEM, see Section 2.2 of the ESI 15 for further details). Figure 4d shows the predicted values of using FEM absorption cross -sections and as Eq. 7, in water (orange squares) and glass (green triangles). Again, a reasonable agreement between calculations and experimental results is obtained for in water. However, the predicted values are smaller than the experimental ones for large Au NPs, indicating that the simple model of Eq. 7 overestimates . This result is rather surprising , as it is the opposite of expected . This model ignores the glass subs trate which is a better thermal conductor than water (glass = 1 WK m/ , water =0.6 WK m/ ). Thus, any correction made to include the substrate would lead to a larger , and as a consequence to a smaller , enlarging , even more, the differences with the expe rimental data. The only way to obtain a lower value for would be to consider the thermal resistances of the Au-water, Au -substrate and water -substrate interfaces (Kapizka resistance s RK). However, considering these three parameters require precise knowledge of the interfaces and the geometric boundaries, and thus increase the complexity of the model. To summarize, t he results presented so far emphasize the importance of interfaces in heat transport at the nanoscale , highlighting the necessity of in situ experimental methods. Photothermal response of 80 nm Au NPs on different substrates To gain further insight into the role of substrates in heat dissipation, the photothermal response of 80 nm Au NPs was characterized on two highly conductive substrates : sapphire and graphene - coated glass. Sapphire has a thermal conductivity an order of magnitude higher than glass (sapphire 20 WK m/ ). Graphene is known for its outstanding thermal conductivity . When a graphene monolayer is deposited onto SiO 2, its in -plane thermal conductivity can be up to two orders of magnitude higher than glass. For example, Seol et al.54 reported graphene@ SiO 2= 600 WK m/ and Li et al.55 reported 840 WK m/ . 16 Arrays of 80 nm ultra-smooth Au NPs were optically printed onto these two substrates (see Materials and Methods for details). Sapphire substrates were treated with polyelectrolytes following the same procedure used with the glass substrates . A monolayer of graphene was deposited onto glass substrates using a wet transfer method (further details in Materials and Methods ). Polyelectrolyte was not necessary in the graphene substrates because nonspecific deposition of Au NPs was not observed during the optical printing process , probably due to hydrophobic repulsion . In the graphene substrates, prior to the optical printing of NPs, Raman spectra were acquired to confirm the presence of a monolayer of graphene in the printing area (Figure S 7a). The characteristic Raman peaks of graphene are still visible in the PL measurements used for thermometry ( Figure S7b), indicating that graphene is not damaged by the optical printing process. Figure 5a shows example dark-field images of the arrays of NPs on the different substrates . Figure 5b shows the mean temperature increase vs. irradiance whereas Figure 5c shows the histograms of photothermal coefficients for the NPs on each substrate . The mean photothermal coefficient for the NPs supported on sapphire was (51 +- 1 ) Kmm2mW/ . This value is lower than on glass , which is reasonable given its higher thermal conductivity. Interesting ly, a substrate with thermal conductivity 20 times higher leads to a reduction of <> of only 18%. Sapphire has a significantly higher refractive index than glass , and therefore the absorption spectr um of the Au NPs is red-shifted with respect to glass. FEM simulations were perfo rmed to take this effect into account . abs(at =532 nm) was estimated to be 16.9 103 nm2 on glass and 18.9 103 nm2 on sapphire. Considering this and using Eq. 6, the calculated heat dissipation factor for a NP on sapphire is 36% higher than for glass . These results are consistent with the ones obtained for 80 nm NPs on glass, where it was observed that the 17 thermal transport is dominated by the surrounding water . It is important to note that the glass and sapphire surfaces were treated in the same manner with polyelectrolytes . Thus, it is reasonable to expect a similar surface chemistry and Kapitza resistance RK56,57 between the substrate and the water . On the contrary, RK between the NP and the substrate might be different. The obtained photothermal coefficient for the NPs supported on graphene -coated glass is <> = (77+-1) Kmm2mW/ . Surprisingly, for a given irradiance level, the NPs on graphene reach temperatures 24% higher than on bare glass. It was observed that the scattering spectra of NPs on the graphene substrate red -shifts with respect to glass, on average by 12 nm ( Figure S 8). Using Mie theory with an effective refractive index for the medium, it was estimated that abs(= 532 nm) is reduced by 7% with respect to glass (Figure S8b). Taking this into account , it was estimate d that the heat dissipation factor for a NP on graphene is 75% of the value for bare glass . Instead of improving heat dissipation, the g raphene monolayer acts as an insulator, in spite of its great heat conductivity. This is ascribe d to the RK at the graphene interface . Contrary to the hydrophilic surface of polyelectrolyte -coated glass, graphene -coated glass is highly hydrophobic.58 RK is inversely proportional to the work of adhesion56,59, and therefore a higher RK resistance is expected between the water and the hydrophobic graphene -coated glass. Moreover, very high values of RK between graphene and SiO 2 interfaces have been measured,60 including a dramatically five -order of magnitude increase in RK when the graphene is not in full contact with glass but presents some corrugation of nm -sized height. 18 Figure 5: The role of the substrate in the photothermal coefficient. (a) Example dark -field images of the optically printed arrays of NPs on the different substrates. (b) Temperature vs. irradiance for 80 nm Au NP on glass, sapphire and graphene substrates, immersed in water. (c) Histograms of photothermal coefficients. The m ean photothermal coefficients (red line) and corresponding standard error of the mean are <> glass =(62+-1), <> sapphire =(51+-1) and <> graphene = (77+-1) Kmm2mW/ . The grey band corresponds t o a +-1 standard deviations interval. CONCLUSIONS A new implementation of anti -Stokes nanothermometry was presented that allows the measurement of the photothermal coefficient of single NPs from an individual hyper spectral image . Unlik e previous implementations, this thermometry method can be used to probe in situ the photothermal response of any type of NP, provided it delivers detectable 19 anti-Stokes emission. It uses excitation at a single wavelength and fixed power, and does not require any extra characterization or prior knowledge about the NP scattering or PL spectra. The method retrieves with a precision of 4 %. Alternatively , if is known, or previously determined, the same method serves to use N Ps as nano -thermometers to sense the temperature of the surroundings . Using this method, the photothermal response of spherical Au NPs on glass substrates, immersed in water was characterized . It was found that for NPs with diameters larger than 50 nm , can be modeled by a homogeneous environment of just water , indicating only a minor role of the substrate in heat dissipation. This finding was consistent with measurements on 80 nm Au NPs deposited on substrates with nominally higher thermal conductance. On sapphire , was found to be 18% smaller than on glass. Taking into account the change in absorption induced by the sapphire substrate, it was estimate d that the heat dissipation for the NPs on sapph ire was only 36% higher than for NPs on glass , despite sapphire has a thermal conductivity 20 times higher. NPs on graphene -coated glass presented a 24% larger than on bare glass. Considering the change in absorption induced by the graphene substrate, it was determine d that NPs on graphene -coated glass dissipate heat 25% less efficiently than on bare glass. This is a rather surprising result, at least at first sight: including a material with extremely high heat conductivity leads to a NP more thermally isolated. However, this can be explained by the presence of high thermal resistances at the hydrophobic graphene interface . Further studies to disentangle these parameters would unlock the potential of graphene materials for thermal management in the nanos cale. Overall, the results of this work reveal the complexity of heat dissipation around NPs and highlights the value of in situ experimental methods able to quantitatively assess the photothermal response of nanosystems . At the same time, they provide a guideline and a method for future 20 research. In this context, we believe that the in situ thermometry method presented here will be useful to address open questions about the role of temperature in plasmon -assisted applica tions , such as photocatalysis or optical manipulation of NPs. MATERIALS AND METHODS Ultra -pure water was employed in all cases (18 MO*cm, Milli -Q(r), Millipore). Gold nanoparticles: Different sized AuNPs capped with CTAB (positively charged) were prepared and purified accordingly to Ref 52. The diameters were (48 +- 2) nm, (80 +- 5) nm and (103 +- 3) nm). Au NPs of (64 +- 5) nm in diameter stabilized with NPC privative capping agent (negatively charged) were purchased from NanoPar tz. All NPs sizes were characterized by transmission and/or scanning electron microscopy (CMA, FCEyN, University of Buenos Aires, Argentina). In order to reach the adequate concentration to optically print the NPs21,48,49, the positively charged NPs were diluted using a 2 mM CTAB solution, and a 1.5 mM NaCl solution was employed for the 64 nm NP s. Substrate : chloride (PDDA, MW ~ 400.000 - 500.000) and sodium polystyrene sulfonate (PSS, MW = 70.000) were purchased from Sigma - Aldrich and used without further purification. Soda -lime glass (Paul Marienfeld GmbH & Co, DE) and sapphire (PI -KEM, UK, Aluminum Oxide <0001>) substrates were cleaned using Hellmanex III(r) 0.2% for 10 min in an ultrasound bath at room -temperature. After rinsing with deionized water and acetone, they were drie d in an oven at 85 degC for 2 h s. Their surface was plasma -activated using a plasma cleaner (Diener, Zepto) at 75 W and 0.5 mbar for 3 min. Finally, positive -charged substrates were produced by immersion of the glass/ sapphire in a PDDA solution (1 mg/ml in 0.5 M NaCl) for 15 min and afterwards they were rinsed with ultrapure water several times. Substrates were stored in water for no longer than one week. Positively charged substrates were used to print 21 the CTAB capped Au NPs. For the negatively charged Au NPs (64 nm), the previously functionalized PDDA substrates were immersed in a PSS solution (1mg/ml in 0.5 M NaCl) for 15 minutes and rinsed again with ultrapure water. Graphene transfer on glass substrates: Monolayer graphen e sheets were deposited on soda -lime glasses by a wet transfer method.61 First, glass substrates were cleaned as previously desc ribed . Graphene -on-copper foils by chemical vapor deposition (CVD) were obtained from Graphenea Inc62. Graphene -copper -foils were coated with polystyrene (PS, MW = 290 .000 g/m ol) by drop casting from a solution of PS in toluene and dried at 75 degC. Then, the underlying copper support was etched using a solution of hydrochloric acid (1.4 M) and hydrogen peroxide (0.5 M). PS - graphene sheets were transferred onto the glass substra te and dried at 75 degC. Finally, PS was removed using toluene and the graphene -coated substrates were annealed at 200 degC overnight. Experimental methods : Au NPs were optically printed onto each substrate according to the process described in our previous wo rks. 21,48,49 Briefly, the Au NP suspension is placed on top of the functionalized substrat e having the same charge as the NPs in an open chamber and printed using a 532 nm laser (Ventus, Laser Quantum) using a 60x water -immersion (Olympus) objective with a NA 1 (beam waist 0=(342 +-5) nm). Printing power was adjusted to each system in order to avoid morphological changes during the optical printing process. Typically, 0.9 mW was used to print onto glass, 0.78 mW for sapphire and 1 mW for graphene. For nanothermometry , laser power was kept rather low to elude long -term irradiation damage. Hyperspectral confocal images of 48 nm, 64 nm, 80 nm and 103 nm AuNPs were acquired using 0.80 mW, 0. 36 mW, 0. 31 mW and 0.2 0 mW, respectively. The heating stage was built using two he aters (MP800, Caddock Electronics) in contact with a stainless -steel sample holder. A temperature sensor (LM35, Texas Instruments) was used to monitor the stage temperature. A PID controller was implemented using 22 an Arduino Uno R3 board for data acquisitio n and control via a Python -based software. A settling time of 30 min was left after the temperature set -point was reached and before the acquisition of hyperspectral images at different temperatures. The entire process of optical printing, hyperspectral im aging and analysis was fully automated by Python routines. ASSOCIATED CONTENT Supporting Information The Electronic Supporting Information (ESI) contains d etails about linearity of Stokes emission, comparison of different temperature -dependent distributi on functions and their ratios, e stimation of the difference between lattice and electronic temperatures , histograms of maximum scattering wavelength of different sized Au NPs , information about calculations of optical properties and Raman spectra of graphe ne. The following files are available free of charge (pdf). AUTHOR INFORMATION Corresponding Author -muenchen.de Present Addresses + Instituto de Nanosistemas, UNSAM -CONICET, Av. 25 de Mayo 1021, San Martin 1650, Argentina. Author Contributions 23 The manuscript was written through contributions of all authors. All authors have given approval to the fi nal version of the manuscript. ACKNOWLEDGMENT M.B., I.L.V., L.P.M and F.D.S a re thankful for the financial support from ANPCyT (PICT 2013 - 0792 , PICT 2014 -3729 and PICT 2017 -0870). M.B. acknowledges CONICET for a postdoctoral scholarship and L.P.M. for a doctoral scholarship. J.G. acknowledges the European Commission for the Marie -Sklodowska -Curie action (797044) and the PRIME programme of the German Academic Exchange Service (DAAD) with funds from the German Federal Ministry of Education and Research (BMBF ). D.P. and V.G. thank Prof. Dr. Kannan Balasubramanian for his help in the graphene transfer method . S.S. acknowledges funding from the German Cancer Foundation (DKH, Deutsche Krebshilfe) within the Priority Program 'Visionary Novel Concepts in Cancer Research' . S.A.M and E.C acknowledge funding and support from the Deutsche (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2089/1 - 390776260, the Bavarian program Solar Energies Go Hybrid (SolTech ) and the C enter for NanoScie nce (CeNS). E. C. acknowledge s support from the European Commission through the ERC Starting Grant CATALIGHT (802989). REFERENCES (1) Maier, S. A. Plasmonics : Fundamentals and Applications ; 2004; Vol. 677. (2) Giannini, V.; Fernandez -Dominguez, A. I.; Heck, S. C.; Maier, S. A. Plasmonic Nanoantennas: Fundamentals and Their Use in Controlling the Radiative Properties of 24 Nanoemitters. Chem. Rev. 2011 , 111 (6), 3888 -3912. (3) Baffou, G. Heating Metal Nanoparticles Using Light ; Cambridge University Press, 2017. (4) Kuppe, C.; Rusimova, K. R.; Ohnoutek, L.; Slavov, D.; Valev, V. K. Hot in Plasmonics: Temperature -Related Concepts and Applications of Metal Nanostructures. Adv. Opt. Mater. 2020 , 8 (1). (5) Jauffred, L.; Samadi, A.; Klingberg, H.; Martin Bendix, P.; B. Oddershede, L. Plasmoni c Heating of Nanostructures. Chem. Rev. 2019 , 119 (13), 8087 -8130. (6) Huang, X.; Jain, P. K.; El -Sayed, I. H.; El -Sayed, M. A. Plasmonic Photothermal Therapy (PPTT) Using Gold Nanoparticles. Lasers Med. Sci. 2008 , 23 (3), 217 -228. -007-0470 -x. (7) R. K. Ali, M.; Wu, Y.; A. El -Sayed, M. Gold -Nanoparticle -Assisted Plasmonic Photothermal Therapy Advances Toward Clinical Application. J. Phys. Chem. C 2019 , 123 (25), 15375 -15393. (8) Croissant, J.; I. Zink, J. Nanovalve -Controlled Cargo Release Activated by Plasmonic Heating. J. Am. Chem. Soc. 2012 , 134 (18), 7628 -7631. (9) Neumann, O.; Neumann, A. D.; Tian, S.; Thibodeaux, C.; Shubhankar, S.; Muller, J.; Silva, E.; Alabastri, A.; Bishnoi, S. W.; Nordlander, P.; et al. Comb ining Solar Steam Processing and Solar Distillation for Fully Off -Grid Production of Cellulosic Bioethanol. ACS Energy Lett. 2017 , 2 (1), 8 -13. (10) Selmke, M.; Braun, M.; Cichos, F. Photothermal Single -Parti cle Microscopy: Detection of a Nanolens. ACS Nano 2012 , 6 (3), 2741 -2749. 25 (11) Langer, J.; Jimenez de Aberasturi, D.; Aizpurua, J.; A. Alvarez -Puebla, R.; Auguie, B.; J. Baumberg, J.; C. Bazan, G.; E. J. Bell, S.; Boisen, A.; G. Brolo, A.; et al. Present and Future of Surface -Enhanced Raman Scattering. ACS Nano 2019 , 14 (1), 28 -117. (12) Taylor, A. B.; Zijlstra, P. Single -Molecule Plasmon Sensing: Current Status and Future Prospects. ACS Sensors 2017 , 2 (8), 1103 -1122. (13) Butet, J.; Brevet, P. -F.; J. F. Martin, O. Optical Second Harmonic Generation in Plasmonic Nanostructures: From Fundamental Principles to Advanced Applications. ACS Nano 2015 , 9 (11), 10545 -10562. (14) Spinelli, P.; Ferry, V. E.; van de Groep, J.; van Lare, M.; Verschuuren, M. A.; Schropp, R. E. I.; Atwater, H. A.; Polman, A. Plasmonic Light Trapping in Thin -Film Si Solar Cells. J. Opt. 2012 , 14 (2), 24002. (15) Zhou, L.; Swearer, D. F.; Zhang, C.; Robatjazi, H.; Zhao, H.; Henderson, L.; Dong, L.; Christopher, P.; Carter, E. A.; Nordlander, P.; et al. Quantifying Hot Carrier and Thermal Contributions in Plasmonic Photocatalysis. Science (80 -. ). 2018 , 362 (6410), 69 -72. (16) Sivan, Y.; Baraban, J.; Un, I. W.; Dubi, Y. Comment on Quantifying Hot Carrier and Thermal Contributions in Plasmonic Photocatalysis. Scien ce (80 -. ). 2019 , 364 (6439), eaaw9367. (17) Zhan, C.; Liu, B. W.; Huang, Y. F.; Hu, S.; Ren, B.; Moskovits, M.; Tian, Z. Q. Disentangling Charge Carrier from Photothermal Effects in Plasmonic Metal Nanostructures. Nat. Commun. 2019 , 10 (1), 1 -8. -019-26 10771 -3. (18) Baffou, G.; Bordacchini, I.; Baldi, A.; Quidant, R. Simple Experimental Procedures to Distinguish Photothermal from Hot -Carrier Processes in Plasmonics. Light Sci. Appl. 2020 , 9 (1), 108. -020-00345 -0. (19) Gargiulo, J.; Bert e, R.; Li, Y.; Maier, S. A.; Cortes, E. From Optical to Chemical Hot Spots in Plasmonics. Acc. Chem. Res. 2019 . (20) Cuche, A.; Canaguier -Durand, A.; Devaux, E.; A. Hutchison, J.; Genet, C.; W. Ebbesen, T. Sorting Nanoparticles with Intertwined Plasmonic and Thermo -Hydrodynamical Forces. Nano Lett. 2013 , 13 (9), 4230 -4235. (21) Gargiulo, J.; Brick, T.; Violi, I. L.; Herrera, F. C.; Shibanuma, T.; Albella, P.; Requejo, F. G.; Cortes, E.; Maier, S. A.; Stefani, F. D. Understanding and Reducing Photothermal Forces for the Fabrication of Au Nanoparticle Dimers by Optical Printing. Nano Lett. 2017 , 17 (9), 5747 -5755. (22) Jones, S.; An dren, D.; Karpinski, P.; Kall, M. Photothermal Heating of Plasmonic Nanoantennas: Influence on Trapped Particle Dynamics and Colloid Distribution. ACS Photonics 2018 , 5 (7), 2878 -2887. (23) Zemanek, P.; Volpe, G.; Jonas, A.; Brzobohaty, O. Perspective on Light -Induced Transport of Particles: From Optical Forces to Phoretic Motion. Adv. Opt. Photonics 2019 , 11 (3), 577. (24) Setoura, K.; Okada, Y.; Werner, D.; Hashimoto, S . Observation of Nanoscale Cooling Effects by Substrates and the Surrounding Media for Single Gold Nanoparticles under CW - Laser Illumination. ACS Nano 2013 , 7 (9), 7874 -7885. (25) Rodriguez -Sevilla, P.; Prorok, K.; Bedna rkiewicz, A.; Marques, M. I.; Garcia -Martin, A.; 27 Garcia Sole, J.; Haro -Gonzalez, P.; Jaque, D. Optical Forces at the Nanoscale: Size and Electrostatic Effects. Nano Lett. 2018 , 18 (1), 602 -609. (26) Schmidt, A . J.; Alper, J. D.; Chiesa, M.; Chen, G.; Das, S. K.; Hamad -Schifferli, K. Probing the Gold Nanorod -Ligand -Solvent Interface by Plasmonic Absorption and Thermal Decay. J. Phys. Chem. C 2008 , 112 (35), 13320 -13323. (27) Pollack, G. L. Kapitza Resistance. Rev. Mod. Phys. 1969 , 41 (1), 48 -81. (28) Majumdar, A.; Carrejo, J. P.; Lai, J. Thermal Imaging Using the Atomic Force Microscope. Appl. Phys. Lett. 1993 , 62 (20), 2501 -2503. (29) Jaque, D.; Vetrone, F. Luminescence Nanothermometry. Nanoscale 2012 , 4 (15), 4301 - 4326. (30) Rycenga, M.; Wang, Z.; Gordon, E.; Cobley, C. M.; Schwartz, A. G.; Lo, C. S.; Xia, Y. Probing the Photothermal Effect of Gold -Based Nanocages with Surface -Enhanced Raman Scattering (SERS). Angew. Chemie - Int. Ed. 2009 , 48 (52), 9924 -9927. (31) Chiu, M. J.; Chu, L. K. Quantifying the Photothermal Efficiency of Gold Nanoparticles Using Tryptophan as an in Situ Fluorescent Thermometer. Phys. Chem. Chem. Phys. 2015 , 17 (26), 17090 -17100. (32) Baffou, G.; Kreuzer, M. P.; Kulzer, F.; Quidant, R. Temperature Mapping near Plasmonic Nanostructures Using Fluorescence Polarization Anisotropy. Opt. Express 2009 , 17 (5), 3291. (33) Baffou, G.; Bon, P.; Savatier, J.; Polleux, J.; Zhu, M.; Merlin, M.; Rigneault, H.; Monneret, 28 S. Thermal Imaging of Nanostructures by Quantitative Optical Phase Analysis. ACS Nano 2012 , 6 (3), 2452 -2458. 021/nn2047586. (34) Carlson, M. T.; Khan, A.; Richardson, H. H. Local Temperature Determination of Optically Excited Nanoparticles and Nanodots. Nano Lett. 2011 , 11 (3), 1061 -1069. 021/nl103938u. (35) Dulkeith, E.; Niedereichholz, T.; Klar, T. A.; Feldmann, J.; Von Plessen, G.; Gittins, D. I.; Mayya, K. S.; Caruso, F. Plasmon Emission in Photoexcited Gold Nanoparticles. Phys. Rev. B - Condens. Matter Mater. Phys. 2004 , 70 (20), 1 -4. (36) Xie, X.; Cahill, D. G. Thermometry of Plasmonic Nanostructures by Anti -Stokes Electronic Raman Scattering. Appl. Phys. Lett. 2016 , 109 (18). 063/1.4966289. (37) Hugall, J. T.; Baumberg, J. J. Demonstra ting from Au Is Electronic Inelastic Light Scattering of a Plasmonic Metal: The Origin of SERS Backgrounds. Nano Lett. 2015 , 15 (4), 2600 -2604. 021/acs.nanolett.5b00146. (38) Carattino, A.; Caldarola, M.; Orrit, M. Gold Nanoparticles as Absolute Nano Lett. 2018 , 18 (2), 874 -880. 021/acs.nanolett.7b04145. (39) He, Y.; Xia, K.; Lu, G.; Shen, H.; Cheng, Y.; Liu, Y.; Shi, K.; Xiao, Y. -F.; Gong , Q. Surface Enhanced Anti -Stokes One -Photon Luminescence from Single Gold Nanorods. Nanoscale 2015 , 7 (2), 577 -582. 039/C4NR04879B. (40) Cai, Y. Y.; Sung, E.; Zhang, R.; Tauzin, L. J.; Liu, J. G.; Ostovar, B.; Zhang, Y.; Chang, W. S.; Nordlander, P.; Link, S. Anti -Stokes Emission from Hot Carriers in Gold Nanorods. Nano Lett. 2019 , 19 (2), 1067 -1073. 021/acs.nanolett.8b04359. (41) Wu, S.; Hogan, N.; Sheldon, M. Hot Electron Emission in Plasmonic Thermionic 29 Converte rs. ACS Energy Lett. 2019 , 4 (10), 2508 -2513. 021/acsenergylett.9b01857. (42) Cai, Y. Y.; Liu, J. G.; Tauzin, L. J.; Huang, D.; Sung, E.; Zhang, H.; Joplin, A.; Chang, W. S.; Nordlander, P.; Link, S. of Gold Nanorods: Purcell Effect Enhanced Emission from Hot Carriers. ACS Nano 2018 , 12 (2), 976 -985. 021/acsnano.7b07402. (43) Yorulmaz, M.; Khatua, S.; Zijlstra, P.; Gaiduk, A.; Orrit, M. Luminescence Quantum Yield of Single Gold Nanorods. Nano Lett. 2012 , 12 (8), 4385 -4391. 021/nl302196a. (44) Jollans, T.; Caldarola, M.; Sivan, Y.; Orrit, M. Effective Electron Temperature Measurement Using Time -Resolved Anti -Stokes . 2020 . (45) Pensa, E.; Gargiulo, J.; Lauri, A.; Schlucker, S.; Cortes, E.; Maier, S. Spectral Screening of the Energy of Hot Holes over a Particle Plasmon Resonance. Nano Lett. 19 (3), 1867 -1874. 021/acs.nanolett.8b04950. (46) Bayle, M.; Combe, N.; Sangeetha, N. M.; Viau, G.; Carles, R. Vibrational and Electronic Excitations in Gold Nanocrystals. Nanoscale 2014 , 6 (15), 9157 -9165. 039/C4NR02185A. (47) Carles, R.; Bayle, M.; Benzo, P.; Benassayag, G.; Bonafos, C.; Cacciato, G.; Privitera, V. Plasmon -Resonant Raman Spectroscopy in Metallic Nanopar ticles: Surface -Enhanced Scattering by Electronic Excitations. Phys. Rev. B 2015 , 92 (17), 174302. 103/PhysRevB.92.174302. (48) Gargiulo, J.; Cerrota, S.; Cortes, E.; Violi, I. L.; Stefani, F. D. Connecting Metallic 30 Nanoparticles by Op tical Printing. Nano Lett. 2016 , 16 (2). 021/acs.nanolett.5b04542. (49) Gargiulo, J.; Violi, I. L.; Cerrota, S.; Chvatal, L.; Cortes, E.; Perassi, E. M.; Diaz, F.; Zemanek, P.; Stefani, F. D. Accuracy and Mechanistic Details of Optical Printing of Single Au and Ag Nanoparticles. ACS Nano 2017 , 11 (10), 9678 -9688. 021/acsnano.7b04136. (50) Urban, A. S.; Lutich, A. a.; Stefani, F. D.; Feldmann, J. Laser Printing Single Gold Nanoparticles. Nano Lett. 2010 , 10 (12), 479 4-4798. 021/nl1030425. (51) Ruan, Q.; Shao, L.; Shu, Y.; Wang, J.; Wu, H. Growth of Monodisperse Gold Nanospheres with Diameters from 20 Nm to 220 Nm and Their Core/Satellite Nanostructures. Adv. Opt. Mater. 2014 , 2 (1), 65 -73. https:/ (52) Yoon, J. H.; Selbach, F.; Langolf, L.; Schlucker, S. Ideal Dimers of Gold Nanospheres for Precision Plasmonics: Synthesis and Characterization at the Single -Particle Level for Identification of Higher Order Modes. Small 2018 , 14 (4), 1 -5. 002/smll.201702754. (53) Lee, Y. J.; Schade, N. B.; Sun, L.; Fan, J. A.; Bae, D. R.; Mariscal, M. M.; Lee, G.; Capasso, F.; Sacanna, S.; Manoharan, V. N.; et al. Ultrasmooth, Highly Spherical Monocrystalline Gold Particles for Precision Plasmonics. ACS Nano 2013 , 7 (12), 11064 -11070. 021/nn404765w. (54) Seol, J. H.; Jo, I.; Moore, A. L.; Lindsay, L.; Aitken, Z. H.; Pettes, M. T.; Li, X.; Yao, Z.; Huang, R.; Broido, D.; et al. Two -Dimensional Phonon Transport in Supported Graphene. Science (80 -. ). 2010 , 328 (5975), 213 -216. 126/science .1184014. (55) Li, Q. Y.; Xia, K.; Zhang, J.; Zhang, Y.; Li, Q.; Takahashi, K.; Zhang, X. Measurement of 31 Specific Heat and Thermal Conductivity of Supported and Suspended Graphene by a Comprehensive Raman Optothermal Method. Nanoscale 2017 , 9 (30), 10784 -10793. 039/c7nr01695f. (56) Shenogina, N.; Godawat, R.; Keblinski, P.; Garde, S. How Wetting and Adhesion Affect Thermal Conductance of a Range of Hydrophobic to Hydrophilic Aqueous Interfaces. Phys. Rev. Lett. 2009 , 102 (15), 156101. 103/PhysRevLett.102.156101. (57) Ge, Z.; Cahill, D. G.; Braun, P. V. Thermal Conductance of Hydrophilic and Hydrophobic Interfaces. Phys. Rev. Lett. 2006 , 96 (18), 1 -4. 103/PhysRevLett.96.186101. (58) Rafiee, J.; M i, X.; Gullapalli, H.; Thomas, A. V.; Yavari, F.; Shi, Y.; Ajayan, P. M.; Koratkar, N. A. Wetting Transparency of Graphene. Nat. Mater. 2012 , 11 (3), 217 -222. 038/nmat3228. (59) Gonzalez -Valle, C. U.; Paniagua -Guerra, L. E.; Ramos -Alva rado, B. Implications of the Interface Modeling Approach on the Heat Transfer across Graphite -Water Interfaces. J. Phys. Chem. C 2019 , 123 (36), 22311 -22323. 021/acs.jpcc.9b05680. (60) Zhao, W.; Chen, W.; Yue, Y.; Wu, S. In -Situ Two -Step Raman Thermometry for Thermal of Monolayer Graphene Interface Material. Appl. Therm. Eng. 2017 , 113, 481-489. 016/j.applthermaleng.2016.11.063. (61) Iost, R. M.; Crespilho, F. N.; Zuccaro, L.; Yu, H. K.; Wodtke, A. M.; Kern, K.; K. Enhancing the Electrochemical and Electronic Performance of CVD - Grown Graphene by Minimizing Trace Metal Impurities. ChemElectroChem 2014 , 1 (12), 2070 -2074. 002/celc.201402325. (62) Munoz, R.; Gomez -Aleixandre, C. Review of CVD Synthesis of Graphene. Chemical Vapor 32 Deposition . John Wiley & Sons, Ltd December 1, 2013, pp 297 -322. 002/cvde.201300051.

2108.10954

Julian Gargiulo Dr.

Mariano Barella, Ianina L. Violi, Julian Gargiulo, Luciana P. Martinez, Florian Goschin, Victoria Guglielmotti, Diego Pallarola, Sebastian Schl\"ucker, Mauricio Pilo-Pais, Guillermo P. Acuna, Stefan A. Maier, Emiliano Cortes, Fernando D. Stefani

In Situ Photothermal Response of Single Gold Nanoparticles Through Hyperspectral Imaging AntiStokes Thermometry

physics.optics physics.app-ph

http://creativecommons.org/licenses/by/4.0/

Several fields of applications require a reliable characterization of the photothermal response and heat dissipation of nanoscopic systems, which remains a challenging task both for modeling and experimental measurements. Here, we present a new implementation of anti-Stokes thermometry that enables the in situ photothermal characterization of individual nanoparticles (NPs) from a single hyperspectral confocal image. The method is label-free, applicable to any NP with detectable anti-Stokes emission, and does not require any prior information about the NP itself or the surrounding media. With it, we first studied the photothermal response of spherical gold NPs of different sizes on glass substrates, immersed in water, and found that heat dissipation is mainly dominated by the water for NPs larger than 50 nm. Then, the role of the substrate was studied by comparing the photothermal response of 80 nm gold NPs on glass with sapphire and graphene, two materials with high thermal conductivity. For a given irradiance level, the NPs reach temperatures 18% lower on sapphire and 24% higher on graphene than on bare glass. The fact that the presence of a highly conductive material such as graphene leads to a poorer thermal dissipation demonstrates that interfacial thermal resistances play a very significant role in nanoscopic systems, and emphasize the need for in situ experimental thermometry techniques. The developed method will allow addressing several open questions about the role of temperature applications, especially ones where NPs of arbitrary shapes are present in complex matrixes and environments.

I. decay is an inherently quantum mechanical process. The likelihood for the existence of a decaying (unstable) state at a given point of time is known as the and must be calculated within the framework of quantum mechanics. The most intriguing fact that followed from the theoretical studies is that the survival at all times follow the exponential decay law [1] mostly observed in the laboratory. The decay law is expected to be quadratic at small times [2{5] and a power law at large times [6{9]. The theoretical claims led experimental nuclear and particle physicists to (see [10] and references therein) with nuclei such as222Rn,60Co and56Mn with half-lives ranging from hours to days. In spite of performing observations for several half- lives only an exponential decay law was measured at all times. The unique experiment where the non-exponential behaviour at large times was conrmed involved the measurement of the luminescence decays of many dissolved organic materials after pulsed laser excitation [11]. Experimental evidence for short time non-exponential decay was found in a experiment [12] where ultra-cold sodium atoms were trapped in an optical potential created by a standing wave of light. On the theoretical side, the decay law has been investigated using various dierent formalisms in literature (see [13] for a comparison of approaches). Of great interest is the calculation of the critical time for the transition from the quadratic to the exponential at small times and the exponential to the power law at large times. Predictions of the critical times are useful in deciding the feasibility of experimentally observing a non-exponential decay [13]. One of the most commonly used formalism for the calculation of survival probabilities of unstable states is the method introduced by Fock and Krylov (FK) [14]. In this method, the survival amplitude (modulus-squared of which gives the survival probability) is evaluated as a Fourier transform of the density of states (DOS) in the resonance. Thus, the DOS is indeed the crucial quantity required in the calculation and must satisfy certain conditions [6] for the correct physical behaviour of the survival probability. The DOS can in principle be constructed using the poles and residues of the resonances, in a model independent way as was shown in [15]. An essential feature of the DOS is the existence of a threshold factor which ensures the correct power law at large times. In [13], the present authors obtained the expressions for the DOS using formalisms other than the FK. In the present work, we revisit 2the calculation of the survival probability with the FK framework to discover some subtle features of the survival probability and constraints on the density of states. One of the is the fact that the survival probability can only be an even function of time. The result has consequences for the standard determination of the critical transition time at small times by expanding an exponential in all powers of t[2, 3]. In fact, the result we obtain is a particular case of the Wiener Khinchin theorem which tells us that the and the autocorrelation function (constructed from the density of states) are cosine Fourier transforms of each other. Following this result, we investigate the behaviour of the survival probability at small times with realistic examples of resonances from nuclear physics. Constructing a functional form of the DOS with the desired physical features we notice that the absence of an energy dependent form factor (which is usually included in a model dependent way) in the DOS can lead to unphysical results. Finally, the obtained are demonstrated in a more visual way by applying the expressions obtained to realistic resonances. We shall begin by very brie y introducing the Fock-Krylov (FK) method which can in principle be used to describe the decay of any resonant state, whose density of states as a function of energy is known. The formalism has been a common tool for investigating the behaviour of survival probabilities in literature. This includes the 1958 paper of Khaln [1], [2] on the Zeno eect, [3] on the short time behaviour of P(t), a review [6] and more recently Ref. [16] where the authors discuss the rue face of quantum decay. There exist approaches in literature which do not use the FK method but rely on a potential based formalism which is useful in studying tunneling decays (see [17] and references therein) and some interesting aspects of the latter [18, 19] (we refer the reader to [20] for a the next section, using the FK formalism, we show that it is not possible to nd a density of states (DOS) corresponding to a purely exponential decay. Section III derives the relation between the survival probability and the autocorrelation function of the DOS which leads to a particular case of the Wiener-Khinchin theorem. Section IV derives the expression for the survival probability of a system at small times. Here we emphasize that a consequence of the particular form of the Wiener-Khinchin theorem derived earlier is the evenness of the survival probability. A functional form of the density of states is considered in Section V and the dependence of the results on the choice of the form factor appearing in the density 3of states is discussed in Section VI. Expressing the survival probability P(t) as a function of the number of oscillations performed n, we discover interesting features of P(n) with one of them being that the transition time from the small time non-exponential to the exponential decay law happens at a time when the survival probability has completed one oscillation. These results are compared with those obtained from other approaches in literature. In section VII, we present results for the critical times and behaviour of the transition regions from the non-exponential to the exponential at small times and the exponential to the power law behaviour at large times. II. DENSITY OF STATES ASSOCIATED WITH THE EXPONENTIAL COMPO- NENT OF THE SURVIVAL PROBABILITY The Fock-Krylov (FK) method has been widely used in literature for the analysis of unstable states [16, 21{25]. We refer the reader to [13] for details of the derivation and begin here with the survival amplitude given by, A(t) (1) where(E) is the density of states (DOS) in the resonance and is a real positive function. In the FK method, one constructs (E) by rewriting the initial state in terms of the of the decay products (see Eqs (4) - (10) in [13]). The commonly used Breit- Wigner distribution is an example of such a DOS. Though the FK method with an energy dependent DOS has been extensively used in literature ([1, 2] to quote a few), there exist other approaches for the evaluation of survival probabilities. We refer the reader to [13] for a comparsion between dierent approaches such as the Green's function method, Jost function based formalism and the Fock-Krylov formalism. In [13], the survival amplitude within the Green's function approach was rewritten in a form similar to that of the Fock-Krylov method to obtain, A(t) the sum over nis over all poles. Given the form of the above amplitude, one may identify the quantity in the square brackets as a density of states as in Eq. (1). The DOS standardly used in the FK method is dierent from the quantity in the square brackets. The authors in [13] showed that restricting the sum to only fourth quadrant poles and one isolated resonance, the above density (referred to as GCthere) is a sum of two terms (see Eqs (69) to (72) in [13]). The rst term is an energy derivative of the phase shift and the second term is negligible for narrow resonances. Thus, the Fock-Krylov and Green's function method [26] agree for isolated narrow resonances. The survival amplitude allows us to compute the so-called the survival probability P(t), which measures the probability that the state of a system is in its initial state at a time, t>0: P(t) =A(t)2: (2) As a natural manifestation of the normalization condition, we have, A(0) = 1 and P(0) = 1. It is known that the survival probability is split in three well-dened regions: the small time region where P(t) follows a quadratic law, the intermediate time region where P(t) is dominantly exponential and the large time region where P(t) is dominated by a power law [1, 6, 27, 28]. Since some systems have long intermediate exponential regions (for decays), it would be suitable to describe those decays such that the density of states gives an exponential survival probability only. However, in the following, we shall show that such a density of states does not exist. We start from the complex (3) whereCis the contour of integration shown in Fig. 1: Contour of integration for the survival amplitude The exponential function (or sum of exponential ones) is obtained when the density of states has simple poles in the fourth quadrant. Since (z) must be a positive real function 5for positive real z, its poles have to be complex conjugate pairs on the right semiplane, i.e., (z) has simple poles (s)=2; and!s>0 for alls. Fors >0 the poles are in the fourth quadrant, and for s<0 they are in the rst quadrant. Let also R(zs) be the residue of (z) at the pole z=zs. As a result of these conventions, we get the following (5) R(zs<0) =R(zs>0): (6) In the appendix A, we show from the integral (3) that A(t) (A5) The integral (A5) lets us easily separate the exponential and the non-exponential component of the survival amplitude. Calling these respective components, Ae(t) andAne(t), from the integral (A5) we get: Ae(t) (7) Ane(t) (8) If the survival amplitude were purely exponential, Eq. (8) would be zero, and therefore the density of states would vanish along the negative imaginary axis. Since an analytic function in a region is zero when the function vanishes along a continuous curve in this region [29], we deduce that (z) = 0 for all points of a domain on the zplane which contains some segment of the negative imaginary axis. Since this domain is arbitrary, the density of states would be zero over the complex zplane and then, the survival amplitude would be zero for all t. However, this is imposible because the density of states has to be dierent from zero over the real positive axis and has to have simple poles for Re z >0. In addition to that, the could not be zero for all tbecause the normalization condition would no longer be satised (apart from such an amplitude losing the meaning of a survival amplitude). Hence, there does not exist a density of states such that the survival amplitude (taken as a of this density of states) and the corresponding survival probability are few words to put this result in context with the well-known seminal result of Khaln [1] are due. Using the Paley-Wiener theorem, Khaln showed that a survival amplitude given as the Fourier transform of a semi-nite density is not attainable for arbitrary functions M(t), where the survival amplitude is given by A(t) =M(t) exp(it). The result obtained in this section is in some sense a reverse statement of what Khaln states. We show that assuming a purely exponential survival probability, jA(t)j2, it is not possible to nd a density of states, the Fourier transform of which gives the survival amplitude as in our Eq. (1). The result is relevant since most observed decays are exponential and it would be desirable to look for such a density of states. Apart from this, it has consequences for the calculation of the autocorrelation function which we shall see later in Section III. It is worth noting that the above result is based on a simple argument of analytic continuation and does not rely on the Paley-Wiener theorem. Eq. (A5) lets us descompose the survival probability, P(t), in a suitable way. Therefore, from the denition of P(t) and Eq. (A5), we can write P(t) 2 (9) If we do not focus on the non-exponential behaviour at very small times, we can see that the survival probability is split into three terms: the rst one comes from the of A(t), the second one comes from the non-exponential component of A(t) (at large times) and the last one is nothing but the interference between the components of A(t). These terms are denoted respectively as Pe(t),Pne(t) andPi(t) and are called (in the same order) the exponential, non-exponential and interference survival probability. From the Eqs (7) and (8), those terms are given (11) Pi(t)2 (12) This particular descomposition of the survival probability is useful for analyzing the tran- sition of the decay law from the exponential to the power law at large times. The analysis of the latter transition using this splitting can be seen in [13]. In section VII we shall study 7the small time transition region. Though the above division of regions is in general valid, there exist exceptions where the decay law is non-exponential at all times. This happens in the case of broad resonances, of which the meson is a good example [8]. Other examples can be found in [30, 31]. III. SURVIVAL PROBABILITY IN TERMS OF THE AUTOCORRELATION OF THE DENSITY OF STATES Although it is easier to compute the survival probability simply as the modulus squared of the survival amplitude, it would be desirable to get an expression involving some function of the density of states. We shall show that this function is the autocorrelation function of the density of states. Furthermore, the expression will be a particular case of the Wiener- Khinchin theorem1. Using (2), P(t) =A(t)A(t) (13) Making the change of variables E=xandE0E=y, the new region of integration Sis given Integrating rst over xand then over y, the integral (13) is split in two integrals: P(t) in the second double integral another change of variables, i.e., region of integration is transformed to one given by x0Eminandy00. Thus, P(t) the integrals can be added. Therefore, P(t) =Z1 0dy2 (16) The survival probability is then the cosine Fourier transform of the function R(y) = (17) 1For details about this theorem, see [32] and [33]. 8R(y) is the autocorrelation function of the density of states. In order to obtain the function as an inverse cosine Fourier transform of P(t), we multiply Eq. (16) by cosy0tand then we integrate with respect to tbetweent= 0 andt=1so that Z1 0P(t) = 2R(y): (18) To summarize, the survival probability and the autocorrelation function of the density of states are a pair of cosine Fourier transforms given by P(t) =Z1 0R(y) cosytdy; (19) R(y) =2 Z1 0P(t) cosytdt: (20) A word of caution about the lower limit of integration of the autocorrelation function of the density of states is in order here. Without loss of generality, we take its value as Emin= 0. This is equivalent to shifting the origin of the energy scale at the threshold of the system and is translated in a change of variable in the integral (17) such that R(y) = (21) Eqs. (19) and (20) still remain valid. Although the aim of this work is to study the characteristic features of the survival prob- ability of an unstable quantum system, some applications of the autocorrelation function are worth mentioning. In the case of a system with a discrete spectrum, it is simple to compute the autocorrelation function by taking the Fourier transform of the survival probability. Let n be a discrete eigenstate of a system with Hamiltonian Hand initial statei2, and let En be their corresponding eigenenergies: Hn =Enn , wheren= 0;1;2;::: and the discrete energies are labeled such that E0< E 1< E 2


. The survival amplitude and probability are then given as A(t) (22) 2In order to avoid any confusion related to the notation for the initial state used at the beginning of this article, this state is denoted asi only in this section. 9P(t) (EnEm)t: (23) The corresponding autocorrelation function can be computed by inspection: Eq. (23) is obtained from Eq. (19) by multiplying the rst sum on the right hand side of (23) by (y) and substituting cos ( Therefore, R(y) (24) Since the autocorrelation function is a convolution-like function, the authors in [34{36] use Eq. (24) together with random matrix theory for nding signatures of classical chaos in quantum systems. They do so by comparing numerical results with experiment. We pointed out in Section II about the impossibility of deriving a density of to a purely exponential survival probability. This implies that we would not be able to use the autocorrelation function of the density of states given by Eq. (21) if we wanted to identify the exponential contribution of the survival probability. Even if we showed that it is impossible to nd a density of states corresponding to a survival probability, one can always nd a density of states which leads to a survival probability with an exponential component in addition to the non-exponential ones as given in Eqs (7) and (10). In spite of having found a functional form for the density of states, the general case of the exponential component of a sum consisting of several poles is dicult to work with. Let us therefore consider the simpler case of a narrow with a long intermediate region of exponential decay. In this case, P(t) can be approximated roughly by its exponential component. Using zsas given by (4), Ae(t) and Pe(t) from (7) and (10) can be written as Ae(t) (25) Pe(t) = (26) and Eq. (20) gives us the approximate autocorrelation function: R(y) =2 Z1 0P(t) (27) 10If we wish to take other resonances into account, we can improve the approximation of function. From Eq. (10): P(t)Pe(t) = (28) Using Eq. (20), the autocorrelation function is approximately equal to: R(y) SURVIVAL PROBABILITY OF A SYSTEM FOR SMALL TIMES A consequence of Eq. (19) is the evenness of P(t). This property can be checked directly from the denition of the survival =A(t)A(t) (30) Expanding the temporal evolution operators in a Taylor series: P(t) (31) wherepn(H) is given by: pn(H) (32) The sum given by (32) has the withAmgiven (33) Since the above sum is zero for nodd,pn(H) is zero for nodd. Thus, P(t) = 1 the above, we can say that not only is the survival probability an even function, but also its Taylor expansion around t= 0 contains even powers of tonly. As a = 0 and P00(0) =p2(H) =2 (
H)2 0<0; Furthermore, Eq. (34) implies that the survival proba- bility must follow a quadratic law for small times, meaning, P(t) = (36) Another implication of the above result is that it is not possible to have a purely probability because the Taylor expansion of the exponential survival 0 has both even and odd powers of t. We can see this fact if we expand the Eq. (7) in Taylor series: Ae(t) Therefore, the corresponding Taylor expansion for Pe(t) will be: Pe(t) (38) lets us interpret the role of the non-exponential and interference terms of the survival probability. Adding the Taylor expansion of both functions to the expansion of Pe(t), the normalization condition and the evenness of the survival probability should be ensured. We shall see that this actually happens. Since the Taylor expansion of the interference terms of the survival probability depend on the Taylor expansion of survival amplitude, we can expand A(t) use Eq. (37). Thus, we get the required expansion, i.e., Ane(t) =A(t)Ae(t) (39) Therefore, the Taylor expansion of the non-exponential survival probability is Pne(t) addition, the Taylor expansion of the interference part of the survival probability is Pi(t) = is gratifying to nd that adding the Eqs. (38), (40) and (41), we obtain after some algebra the Taylor expansion (34). Note that the expansions (40) and (41) are given in terms of the poles and residues of the DOS and the expectation values of the integer powers of the Hamiltonian at the initial state. Finally, owing to the condition that the coecients pn H
have to be nite, some restric- tions must be taken into account when the Fock-Krylov formalism is used. The nite if the expectation value of Hnat the initial state exists for all n= <1; n = 0;1;::: (42) The conditions (42) claim that the density of states is such that these integrals must converge for each value of n, and hence we infer that the convergence of the integrals will be possible if there exists a real and positive function g(E) such that it is part of the density of states. This function is indeed the form factor introduced often in literature. Since there are no analytic methods for obtaining the form factor, it is common to see phenomenological procedures in the literature (see for instance [6{8, 14, 15, 37]). We note that the existence of the form factor is a consequence of the evenness of the survival probability. A word of caution regarding the moments of His in order before ending this a general short time dependence of the form, A1 +btc, wherebandcare nite constants and requiring the niteness of the moments of H, it was noted in [38] that out the possibility of the short time t1=2behaviour (sometimes found in literature such as [2] and [39]) of A(t) since it implies an innite time derivative of Aatt= 0. For the derivatives to be nite, c1. We also note that studies of the short time behaviour within a quantum eld theoretic (QFT) approach as in [40, 41] nd that the energy depends on the rst two moments of His innite. In a more recent QFT however, the author introduces a cut-o parameter  such that there is no divergence for times t1=. This time scale determines the renormalization of the bare state and formation of the quasiparticle state. In connection with the anti-Zeno eect, the author also determines the energy uncertainty at a given time. A discussion of the niteness of the moments and the short time behaviour within solvable models can be found in [42], where the authors found that the expansion of the survival probability in terms of resonant states predicts the possibility of a t3=2short time behavior, which follows from the fact that in general the energy moments of the Hamiltonian may diverge. The authors considered the expression, P(t) = 1(t=), with parameters, andto adjust the short time behaviour of calculations using experiment. Two sets, = 2,= 12.55sand= 3/2 and = 23.15swere found to agree with the data at small times [12]. V. DENSITY OF STATES FOR CONTINUUM SPECTRA All the results obtained so far in this work indicate that knowing the poles and residues of the density of states and its form factor as well is necessary if we wish to construct (E) through the Mittag-Leer theorem. The latter arms that a meromorphic function can be constructed by knowing its poles and residues. Since the calculation of the survival a Fourier transform which must be performed by going over to the complex energy plane, we must consider the behaviour of the density of states in the complex energy evidence suggests that the large time behavior of the survival probability follows a power law that comes from a branch point of (E) atE= 0 such thatjArgEj<. This can be established by supposing that the density of states has an asymptotic 0 in the E!0+: Eq. (43) in Eq. (A5) and applying the Watson's lemma [43], we nd that the survival amplitude and probability have, for larges times, the asymptotic 1
t+n+1; (44) P(t) =j0j2 + (45) It is known from scattering theory that at large times P(t)/t(2l+3)[6] for a resonance in thelthpartial wave. This would imply =l+ 1=2 which as we will see later is consistent with conditions imposed on . In Section III we noted that the evenness of P(t) follows from the Wiener-Khinchin theorem and indeed the small time behaviour was consistent with However, the large time be an even function. Such a strange behavior is however common in many asymptotic expansions of even functions. As a rst example, consider the function f(z) = (z6+z2+ 1)1=2, which, for large values of jzj, behaves as f(z)z3. A second example is provided by the Bessel function of the rst kind of order 2 n,J2n(x) whose asymptotic expression for large xisq 2 xcos (x 4n) [44]. Finally, the integral g(u) the asymptotic form g(u)ieiu2=ufor largeu[45]. It is like these functions forget how they were raised in t= 0 and change completely as t!1 . We must note that in spite of the above, there is in general no contradiction with the result following from the Wiener- Khinchin theorem since it is the total survival probability (and not just a which must be an even function. Eq. (45) shows that the nature of the branch point is determined by the exponent of the power law and this feature must be included if we want to obtain an expression for the density of states. In summary, the density of states can be built if we know (i) its poles residues, (ii) its form factor and (iii) the threshold factor which also denes the exponent 2 + 2 of the survival probability for large times. Since the density of states can be decomposed into the product of E, with > 0, an analytic form factor g(E) withg(0)6= 0, and a meromorphic function whose poles are the same as those of the density of states, the deduction of a generic expression for the density 15of states starts from the function F(z) (46) which is the meromorphic component of the density of states, is analytic at the origin and has simple poles at z=zsandz=E. LetCNbe the circlejzj=RNsuch thatjzsj< RN fors= 1;2;:::N and contains the pole z=E. From the residue and >0, the integral vanishes when Ntends to innity. Hence, (E) (48) Since the density of states has to be a real function, the form factor must have the in the complex z-plane:g(zs<0) =g(zs>0). Using this property with Eqs (5) and (6), we can write an alternative form for (E): (E) (49) where (zs) is dened as (zs) =2R(zs) g(zs); (50) and due to Eq. (5) and the property of the form factor, it satises (zs<0) = (zs>0). Since the density of states satises the normalization condition, we have Z1 0(E)dE= 1: (51) The density of states is such that the integrals (42) will be nite, <1; n = 1;2;


: (52) Since the non-exponential survival probability given by Eq. (8) is proportional to the of the density of states on the negative imaginary axis, by the existence theorem of the Laplace transform, if there exist real constants Mandsuch that(iy) has the integral converges for t> [46]. From Eq. (53) 16Sincet0, the integral (8) must be convergent for t0, and this means that = 0. Hence, the above inequality (54) There, the factor1 2was absorbed into the constant M. A curious consequence of the con- dition (54) is that the form factor cannot be a Gaussian function, because if g(E) the left hand of Eq. (54) would be not bounded over the negative imaginary axis. Finally, notice that the form factor could be obtained by solving the integral equation (51) subject to the conditions (52) and (54). In many of the physical examples of unstable states such as radioactive nuclei, the poles are such that their imaginary parts are much less that their real parts, i.e., Im zsRezs. These poles are referred to as narrow resonances . In such cases, the coecients (zs) do not depend on the residues of the density of states and is a constant equal to i[13, 15]. Thus, for narrow (56) Using Eq. (55), Eq. (49) transforms to (E) =Eg(E) (57) VI. SPECIFIC FORM FACTORS We shall now discuss the advantages or disadvantages of choosing some particular form of the form factor in the density of states to obtain the survival amplitude at small times. Our two choices are the exponential form factor (commonly used for particle and and a constant which is used in a potential description of unstable states. 17A. Exponential form factor This is one popular choice in the literature because it allows closed expressions to be obtained easily. In this case the form factor g(E) takes the form: g(E) =ebE; b> 0: (58) Even if there is no physical reason for choosing this particular form, it satises the require- ments imposed on the form factor, i.e., it ensures the convergence of the integrals (42) and the condition (54) will be satised if 0 < 1. Substituting Eq. (58) in Eq. (49), the density of states is given by (E) =(E) (59) We shall rst calculate the expectation values of the powers of the Hamiltonian at the initial state by substituting Eq. (49) in Eqs (60) From the identity + >1;jArgj<; (61) where (;z) is the incomplete gamma function3, we (1)n+1 1 1, the residues of the density of states satisfy 1 1: (63) 3Actually, there are two dierents incomplete gamma functions: the lower and upper ones [44]. Here, we are using the latter one, which is dened as (;z) =Z1 zt1etdt; Re>0: 18Eq. (62) must incorporate Eq. (63) in order to ensure the normalization condition. Thus, Hn 0= (1)n 1 ++n
1 (64) In the case of narrow resonances, Eq. (55) gives us an expression for the residues of the density of states. Therefore, Eq. (64) takes the following form: Hn 0= (1)n 1 ++n
1 the other hand, we can calculate the survival amplitude using the identity (61): A(t) Even though the Eqs (64) and (65) allow us to compute the coecients of the Taylor expansion, we have to choose a suitable value for the parameter b. It must be chosen such that the inequality (35) is satised, be positive. the range of the values of bfor which (
H)2 0>0 analytically is it is better to achieve this by employing numerical and graphical methods. In order to ilustrate this procedure, the following example is worked out for one isolated, Introducing the parameters xs, the normalized pole s, and the for time, , as xsImzs Rezs; (67) szs Rezs= 1ixssgn (s); (68) 2tImzs; (69) the termsbzsandpzsin Eqs (64), (65) and (66) have to be replaced (70) pzs= The parameter xs1 indicates that a pole represents a narrow res- onance, and it is related to the oscillation frequency fsof the survival probability as by [13, 48]: fs=1 4xs. This frequency allows us to measure the time of the decay in terms of the number of oscillations nthat the survival probability has (72) 19which is suitable for the description of the survival probability at small times. The choice of a suitable value of b(orbs) depends on the sign of (
H)2 0. Forxs= 0:1 and= 1=2, (
H)2 0is negative if 3 :51104< bs<0:55. If we choose a value of bssuch we would expect the survival probability to take values greater than one. We illustrate this feature in Fig. 2 by computing P(t) withbs= 0:1. 0 5 10 15 20 25 [P(n)] 0.00 0.25 0.50 0.75 [P(t)]0 5 10 15 20 25 30 35t FIG. 2: Survival probability for a system described by an isolated resonance such that xs= 0:1, = 1=2, andbs= 0:1 as a function of the number of oscillations n, and the dimensionless variable for timegiven by the Eq. (69). The dashed line represents the exponential component of the survival probability. In this case, we can see that the survival probability is greater than one approximately at the rst one-fourth of the oscillation. Since the survival probability is convex at t= 0, the form factor with the value of bschosen forces P(t) to increase until a maximum. This behavior makes no physical sense. This situation enables us to pick bssuch that (
H)2 0>0 for two intervals. Moreover, the larger the value of bs, the shorter the critical time that characterizes the transition from the exponential to the power law behavior. Such a sensitivity of the critical time to the value ofbsis however seen only for large values of bs. Ifbs1 we do not face the problem of a variable critical time since the eect of the form factor over P(t) is weak. However, the convergence of the survival probability depends on the condition Re p >0, and then, a singularity is located at p= 0. Hence, we should expect that the survival strong variations when both tandbsare close to zero. In Fig. 3 we plot the survival probability for xs= 0:1 and a relatively large value of bs, i.e.,bs= 1. Note that this value of bsensures a survival probability less than one for n0. 0 5 10 15 20 25 [P(n)]0 5 10 15 20 25 30 [P(t)] 0.00 0.25 0.50 0.75 [P(n)] FIG. 3: Survival probability for a system described by an isolated resonance (with xs= 0:1, = 1=2, andbs= 1) as a function of the number of oscillations n, and the dimensionless variable for timegiven by the Eq. (69). The dashed line represents the exponential component of the survival 4 shows the survival probability for xs= 0:1 andbs= 104. As expected, the survival probability is less than one for n0, its critical time (for the transition from the exponential to the power law) has changed with respect to the previous example, i.e., it happens around 15 to 20 oscillations and near t= 0, the survival probability decreases from 1 to approximately e2in less than two thousandth of a period of oscillation of P(t). Knowing how to choose a suitable value of the parameter b( orbs), we shall now show what would happen with the survival probability for small times for a realistic case, namely, the decay of8Be(0+) into two alpha particles for S-waves and assuming an exponential form factor. In this case, the real part of the pole is, Re zs= 92 keV and the imaginary part is 5:6=2 = 2:8 eV. Thus, xs= 3105, and= 1=2. The survival amplitude oscillates with a period of 4 :6104mean lifetimes. Here, the sign of (
H)2 0is negative if 4:81011< bs<0:5. Since a tiny value of bsdoes not describe the survival 0 because of the singularity at p=b+it= 0, the appropiate value of bsshould 210 5 10 15 20 25 [P(n)] 0.00 0.25 0.50 0.75 [P(n)]0 5 10 15 20 25 30 [P(t)] 0.0 0.2 0.4 0.6 0.8 1.0 [P(t)] 0.000 0.001 [P(n)]FIG. 4: Survival probability for a system described by an isolated resonance ( xs= 0:1,= 1=2, andbs= 104) as a function of the number of oscillations n, and the dimensionless variable for time given by the Eq. (69). The upper plot shows P(n) for 30 oscillations, and the lower one shows P(n) for one oscillation. The dashed line represents the exponential component of the . satisfybs>0:5. Hence, we choose bs= 1, orb=bs=Rezs= 10:83(MeV)1. For this value ofbs, the survival probability can be seen in Fig. 5. Let us look at the prominent features of this decay for the parameters chosen above. The survival probability P(t) is less than one for the value of bsthat we chose, P(t) reaches the 220 30000 60000 90000 120000 150000 [P(t)]0 10 20 30 40 50 60 70 80t 0.0 0.2 0.4 0.6 0.8 [P(n)]FIG. 5: Survival probability for the decay of8Be(0+) into two alpha particles with bs= 1 as a function of the number of oscillations n, and the dimensionless variable for time given by the Eq. (69). The dashed line represents the exponential component of the survival time region after 150000 oscillations approximately (around 70 mean lifetimes) and the intermediate region starts after the survival amplitude has completed one oscillation (in 4:6104mean lifetimes). Both examples point to the fact that the exponential behavior at small times is after the survival amplitude has completed one oscillation. Although this could seem to be a coincidence, we shall show in section VII that the intermediate region indeed starts approximately after the rst oscillation is completed by the survival probability for narrow isolated resonances. B. Constant form factor Even though a constant factor is the easiest and simplest choice, we shall show that this option implies some contradictions. The form factor g(E) takes the form: g(E) = 1; (73) 23and the density of states is given by (E) (74) From the normalization condition, the density of states must 1 (75) From a mathematical point of view, the integral in Eq. (75) converges if 1< < 0, and we have >0 because the density of states must be analytic at E= 0. Assuming that  takes values in the range 0 < < 1, we can write Eq. (75) as Z1 0(E)dE= 1 (76) and we can see that the integral is not nite because of the integral in the rst term of the right side of Eq. (76) 0;0< < 1: (77) From the Eqs (76) and (77), we have Z1 0(E)dE= 1 (s): (78) In conclusion, if 0 <  < 1 and because of the normalization condition, the residues and poles of the density of states have to satisfy the (s)=sin ; 0: (80) Although the conditions (79) and (80) seem arbitrary, they appear naturally in systems under the in uence of a central potential of nite range whose decay is calculated by either resonant states or Jost functions [13]. Moreover, substituting (E) given by Eq. (75) into Eq. (42), we get the expectation values of the integer powers of the Z1 0Em+p1dE! (81) will be nite 0; p = 0;1;:::;m: (82) Since this condition should be valid for all positive m, it can be written 0; p = 0;1;:::;m;::: (83) The expectation value of Hnat the initial state is therefore given (s); m = 0;1;2;::: (84) It is worth to point out that it is possible to ensure the convergence of the integrals (42) by imposing conditions over the poles and residues of the density of states given by Eqs (82) and (84). For p= 0, we get the condition (80) which was obtained by the As a result, the survival amplitude can be obtained from Eq. (A5) and the identity (61). After a short algebra, we get, A(t) < 1:(85) There are some advantages to studying the decay of a system using a density of states with a constant form factor. Firstly, it lets us separate the exponential behavior from both the survival amplitude and probability. Secondly, it lets us deal with isolated resonances because the conditions (79) and (80) give us an expression for the residue in terms of its associated pole only (see Eq. (88) below). Lastly, the descriptions of the survival probability in the intermediate and large time regions are simpler and allow us to study in detail the transition from the former to the latter region in a suitable way. However, using this formalism for the description of P(t) for small times for one isolated resonance gives us a result that disagrees 25with Eq. (34). Let us assume that the isolated resonance has the pole, with it. Thus, its residue is obtained from the conditions (63) and (64) to 0; (87) Solving Eqs (86) and (87), the residue is given by, R(zr) (88) SinceR(zr) has to satisfy the conditions (66), we = 0;for allp= equation can be written as, sin pArgzr = 0;for allp= 0;1;2;:::. The only choice that satises this equation for all values of pand for the range of values that Re zrand Imzr can take is: Arg zr= 0, a result that implies Im zr= 0 and hence a resonance without a width. If one would still insist to continue with the description of such a resonance and examine the residue, one would =1 2i; (90) and the density of states would be given by (E) = (91) From the Plemelj-Dirac formula4 (E) =(ERezr); (92) 4The Plemelj-Dirac formula the principal value [49]. 26the survival amplitude reduces to A(t) (93) and the decay probability will be equal to one, implying no decay. This means that it is not possible to obtain the survival probability for small times starting with the energy density of a system which is an isolated resonance. VII. TRANSITION REGIONS AND CRITICAL TIMES We mentioned before that the survival probability has three well-dened regions: the small times region where P(t) is dominantly quadratic, the intermediate region where P(t) is approximately exponential and the large time region where P(t) displays a power law be- haviour. However, there is not really a sharp separation between the three regions but rather an oscillatory transition region from the quadratic to the exponential and the exponential to the power law. Associated with these regions are the critical times , which indicate when the transition starts and ends or when the transition is happening. In this section, we shall explore these aspects. A. Critical Time and the Transition Region from Small to Intermediate Times In order to nd the critical time of transition, tcs, from the quadratic to the we begin by approximating P(t) given by (36), at small times, (94) Focussing on the intermediate region, we note that from Eqs (7) and (28), we know the exponential component of both the survival amplitude and probability. In order to simplify the notation in what follows, we recall that R(zs) =2iR(zs); and rewriting Eqs (7) and (28) and expressing the poles in terms of their real and imaginary parts, we get, Ae(t) (96) In order to study the transition, the simplest choice is to approximate Pe(t) by taking the slowest decreasing term which is associated with the fourth-quadrant pole of the density of states having the smallest absolute value of the imaginary part. We call this the decay's dominant pole ,zd=di!d=2, so that!dmust satisfy the following: !d=2 (97) The intermediate time survival probability can now be written approximately (98) The transition time t=tcscould be dened as the time for which the probabilities given by Eqs (94) and (98) are equal. Thus, tcssatises the (99) It is convenient to dene the dimensionless quantity, cs=!dtcs, which, after some allows us to write Eq. (99) as, f(cs) (100) where we have introduced the parameter dened (101) The problem is thus reduced to nding the values of csfor whichf(cs) is equal tojR(zd)j2. The next step is to study under what conditions will Eq. (100) have a solution. Therefore, we need to study some properties of this function in the interval 0 cswhere the survival probability for small times is positive. One can easily check these properties to be: (i) Forcs= 0,f(0) = 1 for all values of andcs=,f() = 0. (ii) Sincef0(0) = 1, the straight line which is tangent to f(cs) at0has a slope of 1 and is independent of the value taken by . (iii)f(cs) has a maximum in the interval (0 ;) given by+=p 1 +21, and the value of the function at that point is f(+) =f+= 2+e+=2. it is concave. But, for p 2, it 28tcsf(tcs) a f+<11FIG. 6: Sketch of the function f(cs). has an in ection point at csi=p 2+ 22 and it is convex for cs<iand concave for cs>i. The above features of f(cs) in the interval 0 csare represented graphically in the gure 6. The maximum of f(cs) allows us to deduce a criterion for knowing how many solutions Eq. (100) can have: (i) IfR(zd)2>f +, there is no solution. (ii) If f+=R(zd)2, there is one solution. (iii) If 1 <R(zd)2< f +, there are two solutions. (iv) IfR(zd)2< f +<1, there is one solution. These cases are represented graphically in Fig. 6. The about this denition of the transition time is that there exist dierent solutions that depend on the modulus squared of the residue of the density of states evaluated at the dominant pole. Unlike the denitions adopted for the transition time from the exponential to the power law behavior, there is no physical criterion for choosing one particular solution in this case. In other words, this denition of the critical time for the transition from the small to the intermediate time region does not provide a unique solution and it is not possible to choose one of them from physical arguments. At the end of section VI A, we conjectured based on the examples developed there that the intermediate time region had been reached once the survival probability completes its This would mean that the modulus of the non-exponential survival amplitude at small times would be smaller than the modulus of the exponential survival amplitude once 29the survival probability completes its rst oscillation, and this fact may give us a pointer to dene the critical time. Furthermore, this pointer can be easily studied owing to the fact that the frequency of oscillations of both the survival probability and the survival amplitude are the same (which shall be shown in the section VII C), which permits us to introduce the number of oscillations of the survival probability in order to analyze the temporal behavior of the exponential and non-exponential survival amplitudes. Let us then write Ane(t) given by Eq. (8) and use the expression (49) for the density of states for one isolated (102) Using the notation introduced by Eqs (67), (68) and (69) in section VI A for writing the time in terms of the number of oscillations and making the change of variable y!yRezs, we get, Ane(n) the case of large nand writing the asymptotic expansion of Ane(n) using Wat- son's lemma, we 1
2n
+1; n!1 (104) Note that for n= 1, 2nwhich is around 6.28 can be considered large and thus we could say that the above equation is valid for n1. Although obtaining an analytic estimation of the error in using the above expression for all n1 is dicult, numerical tests validate this approximation. From Eqs. (7), (67), (68) and (69), the exponential survival amplitude can be written as, Ae(n) =2iR(zd) exp 2idn
: n1; is dened as: C(xd;) This observation gets better for where xs1. It allows us to establish approximately when the small time Therefore, we dene the critical time for the transition from the quadratic small time behaviour of the decay law to the intermediate exponential one as the time for which the survival amplitude has completed its rst oscillation . From Eq. (72), the critical time in the dimensionless units dened before is given by, cs= 4xd: (108) On the other hand, we shall now provide expressions for describing not only the small time survival probability, but also the transition from this region to the exponential regime. We start by computing the ratio Ane(n)=Ae(n) once again from the Eqs. (103) and (105), we are dealing with narrow resonances ( xd1) and since n < 1, (zd) anddeixd. Hence, Eq. (109) (110) Without the knowledge of the specic form of the form factor, we cannot make more However, as ngoes from 0 to 1, we could expect that the integral decreses because of the exponential factor e2ny. Calling this integral N(n), i.e., N(n) (111) we xs1;0<n< 1: (112) Both the survival probability and the transition from small to intermediate times can be described through Eq. (112). Furthermore, the survival probability at small times will be given by P(t) =Ae(n) +Ane(n)2= +e2inN(n)2; (113) expression that shows the oscillatory nature of the transition. An example of how the transition from small to intermediate time happens can be seen in the gure 7 for xd= 0:01 31and an exponential form factor with bd= 1. We can see that the survival probability for small times given by the Eq. (111) (dashed line) agrees with the survival from the eqs. (103) and (105) (solid line). We also see that the transition ends nearn= 1 by comparing with the exponential survival probability (dotted line). 0.0 0.5 1.0 1.5 [P(t)]0.0 0.05 0.10 0.15 0.20 0.25t FIG. 7: Survival probability for small times for a system with xd= 0:01, and an exponential form factor with bs= 1 as a function of the number of oscillations n, and the dimensionless variable for timegiven by the Eq. (69). The solid line represents the complete survival probability, the dashed line represents the survival probability given by the Eq. (113), and the dotted line represent the exponential component of the survival probability. In the section VII C, we shall use the asymptotic expansion of Ane(n) once again and we shall deduce the conditions on the number of oscillations, n, for the intermediate and its transition to the power law at large times. Apart from the above observations, we can also see that since the survival probability is convex downward at n= 0 and it becomes convex upward near n= 1=2, the quadratic law will be valid approximately in the rst half-cycle of the survival probability, and the transition from the small to the intermediate time region occurs in the second half-cycle. 32B. Comparison with other of the small time behaviour of the survival probability have a long history with the interest being particularly enhanced by the possibility of the so-called quantum Zeno eect [2]. Here we shall compare the results obtained in the present work with some of those in literature for the behaviour of the survival probability at short times and its transition to the exponential behaviour at intermediate times. Before performing such a comparison, we must emphasize that most of the comparisons in literature do not take into account the fact that the survival probability P(t) must be an even function of t. For example, in [2], on the basis of models, the authors propose that as 6= 1: (114) The critical time is typically determined by comparing the expansion of the small times, namely, exp( t) = the quadratic behaviour, P(t) = 1(
H)2t2. Performing such a comparison, Ghirardi et al., obtained an expression [4] for the critical transition time from the quadratic to the exponential decay law
Eis the uncertainty of the Hamiltonian evaluated at the initial state and is the width of the resonance. The authors in [4] initially obtained a result based on an by Fleming [50] and showed that the above result coincided with the rst one in the case of
E . We shall brie y describe the derivation of this result and rewrite it in order to compare it with the result of the present work. Fleming derived an inequality which provided a lower bound for the survival probability such that, jA(t)jcos (t
E);0t 2
E: (116) Assuming that the density of states is described as a narrow resonance and satises the Breit- Wigner form near the resonance, the critical time G cswas dened in [4] as the intersection of the functions cos2Dande. Hence, (117) whereD=
E= , the approximate solution to Eq. (117) given in the paper is: G cs=1 D2: Eq. (118) to our notation: =!d, and (119) The right side of the above equation is nothing but the square of the parameter (120) In the Fock-Krylov language, the uncertainty of the energy depends on the density of states and hence depends on the form factor too. Therefore, the critical time derived in [4] is in principle form factor dependent. Similarly, csdened by (100) is also dependent on the form factor used. Our denition (108) of the critical time, i.e., cs= 4xd, however, depends only on the poles of the resonance and is hence independent of the form factor in the density of states. If we rewrite this critical time in terms of xd=!d=2d, we 0 because of the energy-time uncertainty. We can deduce that cs!0 whenxd!0. Finally, the critical time measured in terms of the number of oscillations (122) Critical times of Ref. [4] and our Eq. (100) for dierent values of xd=!d=(2d), using an exponential form factor with bs= 2 in the density of states for an isolated narrow resonance are compared with our form factor independent denition (108) in Table I. From Table I, we can infer that the critical times of Ref. [4] are one order of magnitude less than the critical times given by (108) whereas several orders of magnitude smaller than those of (100). The number of oscillations seem to be constant as long as xdis small but start increasing beyond xd= 0.1. Since the result of Ref. [4] is expected to be valid for , a comparison with the above results beyond xd= 0.1 may not be only the cases below xd=0.1, one can draw a general conclusion that the survival probability completes about half an oscillation (i.e., nG cs0:4) before going over to the dominant exponential decay law in contrast to ncs= 1 of (108). However, the value of nG csis form factor dependent and increases with decreasing value of bs. For the lowest times Number of 0:1257 0:396 1:639 1010:639 0:765 1:257 0:509 0:609 TABLE I: Comparison of the critical transition times, cs, from the non-exponential to the expo- nential decay law at small times within the approach of Ref. [4] and that of the present work (Eqs (100) and (108)). nG cs(122) are the number of oscillations performed by the survival reaching the exponential region. The last column gives the number ncscorresponding to the solution of (100). The number corresponding to cs(108) is always ncs= 1. value ofbs= 0:65 for example, nG cs0:9. The results given in the third and sixth column, arising from Eq. (100) are grossly dierent from those of (121) and (108). One should be cautious while drawing conclusions from the results of (100) since (i) values of Eq. (100) are strongly form factor dependent (for example, there is no solution for an exponential form factor with bs= 1 even if it falls in the allowed range of values of bs and gives reasonable results for G cs) (ii) the solution of (100) is not always unique and there are cases as mentioned above when there exists no solution at all. As a consequence, (108) which depends solely on the pole value of the resonance, should be considered as the of csof the present work. A brief comment about the increase in the number of oscillations as well as the existence of a bigger exponential region in P(t) with decreasing xdis in order here. These on the coupling constant in the decay (see [51] for a demonstration with a nice model). Explicit examples of strong and weak decays conrm the same [52]. Taking the ratioxdto the extremes, a large value of xdimplies no exponential decay at all for the resonance (see for example [8] for the case of the broad resonance) and a tiny xdimplies an exponential decay for a large region of t. Though one cannot use the idea of a for tunneling decays in nuclear physics such as that of8Be to two4He nuclei, one can see in Fig. 5 that with xd= 3105, the survival probability displays an extremely large exponential region and a huge number of oscillations as expected. To complete the comparison, we nally note that Fleming [50] dened the lifetime of the resonance as is the value of the autocorrelation function R(y) of the present work (see Eq. (21)) at y= 0. C. Critical time and the transition region between intermediate and large times The intermediate time region is characterized by its dominant exponential nature and in the same way, the large time region is ruled by a strong power law. One must however take into account the oscillatory nature of the survival probability too. The oscillatory nature has given rise to a lot of debate in literature [13, 48, 53, 54]. Our aim in this section is to obtain expressions that incorporate the oscillatory feature into the description of these regions for systems described by one isolated resonance. In order to do this, we will use the ideas and formalism described in [13] for studying the transition from the intermediate to the large time region. In ref. [13], the survival amplitude given by Eq. (9) is written as the product of the sum of the exponential and non- probabilities, and a modulating function I(t): P(t) = Pe(t) +Pne(t) I(t); (123) where I(t) = 1 +2 1 +2 (124) Dening the function I(t) seems to be the appropriate path for incorporating the in the intermediate and large time regions. To begin with, notice that of P(t) comes from the from the Eqs. (7) and (8), that the oscillation of the last terms comes from the expo- nential component of the survival amplitude, mainly, from the term the survival amplitude oscillates with a frequency equal a constant, I(t) oscillates at the same frequency as well, and so does the . On the other hand, substituting Eq. (106) in Eq. (124), we get, I(n) = 1 +2 n> 1;(125) In Fig. 8 we show a plot of I(n) as a function of nwithxd= 0:1,= 0:5, and using an exponential form factor with bd= 1. 5 10 15 20 0.1, bd= 1 FIG. 8: Modulating function given by Eq. (125) as a function of the number of oscillations n. Here,xd= 0:1,= 0:5 and an exponential form factor with bd= 1 is used. 37For the intermediate time region, the term be negleg- ible with respect to the term and, On the other hand, the large time region satises the opposite conditions5. From Eqs (123) and (125), the survival probability for the former region is given by P(n) = the survival probability for the latter region is given by: P(n) + Eqs (126) and (127) incorporate the oscillatory nature of the survival probability, we do not have characteristic times that allow us to say when the intermediate or large time region starts or ends, or when the transition is happening. The rst step for times is to dene the critical time (in terms of the number of oscillations) nclof the transition from the intermediate to large times. This is given by the intersection of the exponential and non-exponential survival probability [13, 39, (128) This equation has two solutions and the critical time is dened as the largest one. Another property of this time is related with the nature of the resonance, i.e., the narrower the resonance is, the larger is the critical time6. The next step is to analyze the critical points of the function m(n) (129) It is simple to verify that, for n>1,m(n) has a maximum at n=ncl, and has a or in dimensionless units, = 2(+ 1). These two times let us dene where the transition and the large time regions are. The rst interval 1 <n<+1 2xd 5A demostration of this property can be seen in the appendix B. 6Our aim is not to discuss the details or demostrate the properties of the critical time for the transition from the intermediate to the large time region. In [13], the authors present a complete analysis of this for the case= 1=2, and these properties are similar for the range of values that can take here, so that is to the intermediate time region. The second n < to the transition region: it is dened such that the critical time is the middle point of the interval. Finally, the interval n > to the large time region. Summarizing: i) The intermediate time region is dened in the interval 1 < n <+1 2xd, and the is given by Eq. (126). ii) The transition region from the intermediate to the large time region is dened in the n< 2ncl+1 2xd, and the survival probability is given by P(n) (130) whereI(n) must be given by Eq. (125). iii) The large time region is dened in the interval n > 2ncl+1 2xd, and the is given by Eq. (127). In Fig. 9, we show a plot of P(n) as a function of nwithxd= 0:1,= 0:5 and using an exponential form factor with bd= 1. The diferent regions have been identied by SUMMARY AND CONCLUSIONS The time evolution of an unstable state is investigated by studying the properties and behaviour of the survival probability, P(t), of a decaying state at all times. It is well known that the quantum mechanical description of P(t) leads to ve distinct regions, namely, the quadratic form at small times, the dominant exponential decay at intermediate times, the power law at large times and the two transition regions between the exponential and non- exponential decay laws at small and large times. Even though the topic as such has been studied in quite detail over decades, digging deeper into the behaviour of P(t) in the ve regions allowed us to nd some subtle features as well as restrictions in the form of P(t) and the input density of states (DOS) used to calculate P(t). Working within the commonly used Fock-Krylov formalism and taking into account the conditions derived, each of the above mentioned ve regions of the survival probability are studied mathematically as well 390 5 10 15 20 [P(t)]0 10 20 30t 0 1 2 3 4 [P(n)]FIG. 9: Survival probability with parameters xd= 0:1,= 0:5, and an exponential form factor withbs= 1 as a function of the number of oscillations n, and the dimensionless variable for time given by the Eq. (69). The dotted line is the small time region, the dash-dotted line is the intermediate time region, the solid line is the transition from the intermediate to large time region, and the dashed line is large time region. as numerically with physical examples. Some of the main observations and conclusions are listed below: i) In the FK formalism, the survival amplitude, A(t) is written as a Fourier transform of the DOS. Since the non-exponential decay regions are hard to observe and the expo- nential decay law dominates the observations in the real world, it would be useful to construct a DOS leading to a purely exponential decay. We show that the latter is The survival probability and the autocorrelation function of the density of states are shown to be a pair of cosine Fourier transforms. This is a particular case of the Wiener- Khinchin theorem. iii) A consequence of the previous result is the evenness of P(t). 40iv) A functional form of the DOS which depends on the pole values, a threshold factor and an energy dependent form factor is provided. Consideration of particular form factors with physical examples of resonances leads to the observations that (i) a constant form factor leads to unphysical values of the energy uncertainty, (ii) the (54) derived in this work does not allow the usage of a Gaussian form factor and (iii) the commonly used exponential form factor can be used with restrictions on the parameter in the exponential. v) Transition regions from the small time quadratic law to the exponential decay law and from the exponential to the large time power law are studied using the constructed DOS and the mathematical conditions derived in this work. Expressions for the critical times of transition are provided and compared with existing literature. vi) An interesting novel feature introduced in this work is the description of the and hence survival probability in terms of the number of oscillations per- formed. The small time quadratic behaviour is found to go over to the when the survival probability has completed one oscillation. Analytical expres- sions for the number of oscillations performed in each of the regions of P(t) are given in terms of the pole values and the threshold factor in the DOS. An insight into the mathematical construction of the survival probability alongwith the dierent constraints and implications of the conditions derived for the behaviour of P(t) over the entire region of its evolution from small to large times is thus provided in the present work within the Fock-Krylov framework. Since the time evolution of an unstable state is often studied in literature by focussing on a particular region of P(t) or even a of its behaviour, such a complete mathematical evaluation should prove useful for the focussed studies in future. Appendix A: Descomposition of the survival amplitude as a sum of and an the complex (A1) 41whereCis the contour of integration shown in Fig. 1. Since (z) satises the Jordan's lemma, using the residue theorem we get, Z AO+Z OB= (A2) However, for AO,z=x. (A3) and (A4) The survival amplitude can be written as: A(t) (A5) Appendix B: Ratio of the exponential and non-exponential survival amplitude Our aim is to show that the exponential region and large times. Let (n) be the function (n) the derivative with respect to nand calculating the possible critical points: 0(n) =(n)+ 1 n2xd = 0;)n=+ 1 2xd: It is easy to show that this critical point is a maximum. Since (0) = 0 and (1) = 0, this function starts to rise up until a maximum, and then starts falling down. These properties allow us to study the function 1 =(n), In this case, when nincreases from zero, 1 =(n) falls down until n= (+ 1)=(2xd), and then it starts to rise up indenitely. There should be some values of nsuch that(n) =1 (n). The values of n where this condition is valid are nothing but the critical points for the transition from the exponential to the power law. Since we are interested in narrow resonances, we know that the second solution of Eq. (128) goes to zero when xdgoes to zero too. This implies that even forn= 1(n)1 (n). In conclusion, (n)>1 (n)for 1<n<n cl, wherenclis the critical point for large time transition. 421 5 10 15 20 10: Ratio of the exponential and non-exponential survival amplitudes, and its inverse as a function of the number of oscillations nof the survival probability. In Fig. 10, we plot (solid line) and 1 =(dashed line) for xd= 0:1 andbs= 1. Finally, the gure 11 plots 1 + m(n), wherem(n) is given by Eq. (129) and written in terms of(n) asm(n) = 2=((n)+(n)1) wherem(n) (black line) has been computed using Eq. (129) and using the approximations deduced for . The dashed line is when m(n)2= in the exponential region and the dotted line is when m(n)2in the power law region. Appendix C: Eect of other poles on Pe(t) An interesting and typical case is a system whose survival probability can be as a sum of the exponential terms only. After certain time, this sum should reduce to one term associated to the dominant isolated resonance. In order to see how this reduction law happens, we write Eq. (96) such that the dominant pole is explicit, i.e., Pe(t) + + 5 10 15 20 +m(n) xd= 0.1, bd= 1FIG. 11: Modulating function as a function of the number of oscillations nof the survival we can approximate the exponential survival amplitude by taking all the terms associ- ated to the dominant is dened as the function M(t) = 1 + (C3) The modulating function M(t) (which should not be confused with the modulating function dened by Eq. (124)) contains information about how the reduction occurs. As is expected, this process does not depend on the dominant pole only but it also requires the other poles of the density of states for its description. If we use a mean eld approach in the sense of approximating the sum in Eq. (C3) by taking the nearest pole to the dominant one, the modulating function takes the form: M(t)1 + (C4) 44or expressing the time in dimensionless (lifetime) units of the dominant resonance, i.e., =!dt, we have,Pe() where, M() = 1 + (C5) From Eq. (C5), we infer that the survival probability experiment an oscillation with a frequency!t= (d+1d)=!d, and the survival amplitude goes to the exponential term associated to the dominant pole at a rate of 2( !d+1=!d1)1per lifetime unit (see Fig. 12: Sketch of the transition of Pe() to the exponential law of the dominant pole. [1] L. A. Khaln, JETP 6, 1053 (1958). [2] C. B. Chiu, E. C. G. Sudarshan and B. Misra, Phys. Rev. D 16, 520 (1977). [3] J. Levitan, Phys. Lett. A 129, 267 (1988). [4] G. C. Ghirardi et al., Il Nuevo Cimento 52A, 421 (1979). [5] D. Boyanovsky, Ann. of Phys. 405, 176 (2019). [6] L. Fonda, G. C. Ghirardi and A. Rimini, Rep. Prog. Phys. 41, 587 (1978). [7] N. G. Kelkar, M. Nowakowski and K. P. Khemchandani, Phys. Rev. C 70, 024601 (2004). [8] N. G. Kelkar and M. Nowakowski, J. Phys. A 43, 385308 (2010). 45[9] E. Torrontegui, J.G. Muga, J. Martorell and D.W.L. Sprung, Adv. in Quantum Chemistry 60, 485 (2010). [10] E. B. Norman, S. B. Gazes, S. G. Crane and D. A. Bennett, Phys. Rev. Lett. 60, 2246 (1988). [11] C. Rothe, S. I. Hintschich and A. P. Monkman, Phys. Rev. Lett. 96163601 (2006). [12] S. R. Wilkinson et al., Nature 387, 575 (1997). [13] D. F. Ram rez Jim enez and N. G. Kelkar, J. Phys. A: Math. Theor. 52, 055201 (2019). [14] V. Fock and N. Krylov, JETP 17, 93 (1947). [15] D. F. Ram rez Jim enez and N. G. Kelkar, Ann. Phys. 396, 18 (2018). [16] K. Urbanowski, Acta Phys. Pol. B 48, 1847 (2017); K. Urbanowski, Eur. Phys. J. D 71, 118 (2017); K. Raczynska and K. Urbanowski, Acta Phys. Pol. B 49, 1683 (2018). [17] G. Garc a-Calder on and R. Romo, Annals of Phys. 424, 168348 (2021). [18] F. Giacosa, P. Ko scik and T. S owinski, Phys. Rev. A102, 022204 (2020). [19] T. Koide and F. M. Toyama, Phys. Rev. A66, 064102 (2002). [20] C. Anastopoulos, Int. J. Theor. Phys. 58890 (2019). [21] M. Nowakowski, Int. J. Mod. Phys. A 14, 589 (1999). [22] F. Giraldi, J. Phys. A 52, 415301 (2019) [23] Giraldi, F. Eur. Phys. J. D 73, 239 (2019). [24] W. van Dijk and F. M. Toyama, Phys. Rev. A 100, 052101 (2019). [25] D. Boyanovsky, Annals Phys. 405 (2019) 176{201. [26] G. Garc a-Calder on, A. M attar and J. Villavicencio, Phys. Scripta T151 , 014076 (2012). [27] H. Nakazato, S. Pascazio. Modern Physics Letters A, Vol 10, No. 40 (1995), 3103{3111. [28] H. Nakazato, M. Namiki, S. Pascazio. International Journal of Modern Physics B, Vol 10, No. 3 (1996), 247{295. [29] P. M. Morse and H. Freshbach. Methods of Theoretical Physics, Part I , McGraw-Hill Book Company, 1953. [30] G. Garc a-Calder on and J. Villavicencio, Phys. Rev. A 73, 062115 (2006). [31] G. Garc a-Calder on and R. Romo, Phys. Rev. A 93, 022118 (2016). [32] R. L. Allen and D. W. Mills. Signal Analysis , IEEE press, 2004. [33] D. Brandwood. Fourier Transform in Radar and Signal Processing , Artech House, 2003. [34] Y. Alhassid and R. D. Levine. Phys. Rev. A 46, 4650 (1992). [35] Y. Alhassid and N. Whelan. Phys. Rev. Letters. 70, 572 (1993). 46

2108.10957

Diego Fernando Ram\'irez-Jim\'enez

D. F. Ram\'irez Jim\'enez and N. G. Kelkar

Formal Aspects of Quantum Decay

47 pages, 12 figures

Phys. Rev. A 104, 022214 (2021)

10.1103/PhysRevA.104.022214

quant-ph hep-ph nucl-th

http://creativecommons.org/licenses/by/4.0/

The Fock-Krylov formalism for the calculation of survival probabilities of unstable states is revisited paying particular attention to the on the density of states, the Fourier transform of which gives the survival amplitude. We show that it is not possible to construct a density of states corresponding to a purely exponential survival amplitude. he $P(t)$ and the autocorrelation function of the density of states are shown to form a pair of cosine Fourier transforms. This result is a particular case of the Wiener Khinchin theorem and forces $P(t)$ to be an even function of time which in turn forces the density of states to contain a form factor which vanishes at large energies. Subtle features of the from the non-exponential to the exponential at small times and the exponential to the power law decay at large times are discussed by as a function of the number of oscillations, $n$, performed by it. The transition at short times is shown to occur when the survival probability has completed one oscillation. The number of oscillations depend on the properties of the resonant state and a complete description of the evolution of the unstable state is provided by determining the limits on the number of oscillations in each region.

1. drive is a mechanism based on General Relativity which in theory allows for massive particles to be propelled throughout the spacetime with glob al superluminal speeds [1, 2]. The theory describes this possibility by means of a localized spac etime bubble , that would contain a local lightcone where the particle would follow special relativity, that is, move locally with speeds smaller than light. H owever, the metric is such that the warp bubble moves along a geodesic that creates an expansion of spacetime behind it and a contraction in front of it in such a way that an observe r outside the warp bubble sees it moving with superluminal speed. In the original paper by M. Alcubierre the warp drive metric was esta blished without solving the Einstein equations [1]. The Einstein tensor components we re calculated and it was noticed that negative energy density would be required to crea te the warp bubble, thus violating the weak and dominant energy conditions. Ford and Roman [3] calculated via quantum inequalities the amount of n egative energy required for the warp drive to be possible, concluding that it would be a massive to achieve. Pfenning and Ford [4] also concluded that it wou ld be necessary an enormous amount of energy for the warp drive to be possible. They obtained a quantity ten orders of magnitude greater than the mass-energy of the entire visible Universe, also with negative [5] discussed the possibility of a massive particle moving in sp ace faster than a photon, arguing that this is not possible due to limitations on globally hy perbolic with feasible physical assumptions. He is the creator of a specific with devices that would allow massive particles to travel bet ween two points in space with superluminal velocities without the need for tachyons. E veret and Roman [6] coined the name for this spacetime topology as the Krasnikov tube . They generalized the metric designed by Krasnikov by proposing a tube in the direction of t he particle's path, connecting the start and end point. Inside this tube, the spacetim e is flat and the lightcones are opened to allow the one direction superluminal travel. The Krasn ikov tube also requires huge amounts of negative energy density. Since the tube is designe d to not possess closed timelike curves, it would be theoretically possible to construct a two w ay that could work as a time machine. The energy-momentum te nsor (EMT) for the Krasnikov metric is positive in some regions. Both the metric and the o btained EMT were thoroughly analyzed in Refs. [7, 8]. Van de Broeck [9] made a relevant contribution to warp drive theor y by a small modification of the original Alcubierre geometry would re duce, to a few solar masses, the total negative energy necessary for the creation o f the warp bubble distortion of spacetime. This result have led van de Broeck to suppose that othe r geometrical modifica- tions of the Alcubierre's geometry for the warp drive could also redu ce the amount of energy necessary to create a warp drive bubble in the same way. Natario [10] stated that the spacetime contraction and expansion of the warp bubble is a peculiar consequence of the warp drive metric. Hence, he designe d a spacetime where occurs forthewarp drivebubble. Loboa ndVisser [11, the center of the warp bubble proposed by Alcubierre needs t o be massless [see also 2Refs. 13, 14]. They proposed a linearized model for both Alcubierre and Natario proposals and demonstrated that for small speeds, the energy stored in th e warp fields must be a significant inside thewarpbub ble. geodesics moving according to far away obs ervers when inside a gravitational wave field. Lee and Cleaver [16, 17] analyzed how external radiation might affect the Alcubierre warp bubble to turn it unsustainable. The y also claimed that a warp field interferometer could not detect spacetime distortions . Mattingly et al. [18, [20] discussed further curvature invariants for warped and Martire [21] claimed that any warp drive spacetime consis ts of a shell of regular or exotic material moving inertially with a certain speed, also r eaching at a class of subluminal spherically symmetric warp drives. Lentz [22] and Fell and Heisenberg [23] advanced superluminal capable soliton solutions with positive energy warp drives. Santiago et al. [24, 25] argued that only comoving timelike Eulerian observers s atisfy the weak whereas this not the case for all timelike observers. Fur thermore, they claimed that all physically reasonable warp drives will violate the null and weak energy set up positive energy warp drives that satisfy the weak energy by the fact that neither the original paper by Alcubierre , nor the subsequent ones cited above, did actually solve the Einstein equations using the w arp drive metric, we proceed to investigate possible solutions for a dust particle energy momentum tensor [26]. Our results showed that solutions of the Einstein equations for the Alcubierre warp drive metric having dust as source connect them in a particular case the w arp drive geometry to the well-known Burgers equation, which describes the dynamics o f the waves moving through an inviscid fluid. Hence, shock waves appear to be vacuum s olutions of the endowing the warp drive metric [26]. In our second paper [27] we investigated solutions for the warp driv e metric having the perfect fluid and a special case of anisotropic fluid with heat flux, bu t both with zero cos- mological constant in the Einstein equations. The resulting solutions indicate that positive matter densities are possibly capable of generating superluminal sp eeds. In our third paper [28] a charged dust was used as source EMT for the Alcubierre metr ic and the which included the cosmological constant. We obtained so lutions connecting the electric energy density with the cosmological constant and, again, some solutions were found having positive matter density and satisfying the energy conditions . Motivated by the results we obtained in Ref. [28] we have pondered t hat even though the Alcubierre warp drive metric is a vacuum geometry, the warp bubble w ould be created by geometry alone, or if a vacuum energy would make it possible through other material sources of energy and momentum. Hence, in this paper we went back to the p erfect fluid source but included the cosmological constant in the Einstein equations as an ad ditional flexibility and geometrical properties for the solutions. We calculated the Einstein equations and analyzed the null divergenc e of the energy mo- mentum tensor together with the validity requirements for the ene rgy conditions inequalities to be satisfied. We found that the perfect fluid with the cosmologica l constant as source for the Alcubierre warp drive results in four sets of differential equatio ns, two of them are very 3similar and raise the possibility for the shift vector to be a complex fun ction in one case, depending on the ( t,y) coordinates, and in another case depending on the ( t,z) coordinates. The other two sets of solutions are identical to each other and simila r to the solution we found in Ref. [26], except that now there is a cosmological constant coupled with the Burg- ers equation and, again, the warp drive is connected to shock wave s solutions. Considering that the zero pressure reduces the perfect fluid to the dust EMT , the solution for this case is identical to the one we found in Ref. [26], namely, the vacuum solutio n of the connecting the warp drive to shock waves. The plan of the paper is as follows. Section 2 presents a brief review o f the basic equations and concepts of the warp drive theory, and in Section 3 we solve the Einstein equations and calculated the covariant divergence for the EMT. Section 4 discuss es the energy and their validity for the warp drive with L /negationslash= 0 and the perfect fluid as a source. In section 5 we analyze incoherent matter as a source assuming tha t this is as a special case considering the perfect fluid with null pressure. In section 6 we dep ict our conclusions and final remarks. 2. WARP DRIVE BASIC CONCEPTS 2.1. Warp drive metric The warp drive metric is a generic metric in a foliated spacetime given by the (2.1) wheredtis the proper time lapse, ais the lapse function that controls the amount of the spatial metric for the hypersurfaces. The lapse function aand the shift vector biare functions of the spacetime coordinates to be determined, gijis a on each one of the spacelike hypersurfaces and these feat ures make this hyperbolic. Throughout this paper Greek indices will range fr om 0 to 3, whereas the Latin ones indicate the spacelike hypersurfaces and will range f rom 1 to 3. We have the following choices for Eq. (2.1) [1], a= 1, (2.3) b2=b3= 0, (2.4) gij=dij. (2.5) Hence, the warp drive metric is given is the velocity of the center of the bubble moving along the curve xs(t), given by, vs(t) =dxs(t) dt. (2.7) 4The function f(rs) is the warp drive regulating form function . It describes the shape of the warp bubble, which is given by the constants to be determined. The function rs(t) defines the distance from the center of the bubble [ xs(t),0,0] to a generic point ( x,y,z) on the surface of the bubble, given by the following equation, rs(t) (2.9) From Eq. (2.9) one can see that the warp bubble is perturbed in a one -dimensional manner because of the term x-xs(t). 2.2. Einstein tensor components The components of the Einstein tensor with a cosmological constan t for the warp drive metric in Eq. (2.1) are given by the expressions below, G00= (2.13) G11= as in Eq.(2.3). Also noticed that we incorporated the cosmo- logical constant into the Einstein (2.20) 2.3. Energy conditions revisited The components for the Eulerian (normal) observers' 4-velocities are given by, ua= (1,-b,0,0), ua= (-1,0,0,0). (2.21) Using these results into the Einstein (2.22) results in an expression concerning the energy conditions. From Eq s.(2.21) and considering that the only non-zero terms of Eq. (2.22) are G00,G01andG11, we obtain the Eqs.(2.10), (2.11) and (2.14) into Eq. (2.23) the resu lt may be written as, Tabuaub= (2.24) The bubble radius is given by using Eq. (2.9). So, Eq. (2.24) is given by, Tabuaub= (2.25) This result is similar to the one found by Alcubierre [1], with the differen ce that Ref. [1] did not consider the cosmological constant. Considering the result s in Ref. [1] we realized that both the weak and dominant energy conditions would be violated [26] if the bubble was formed. However, these same energy conditions would be satisfied in the case of a vacuum solution, which discloses the new result that the warp drive metric is a vacuum solution for the Einstein equations. Besides, the Burgers equation is connecte d to this geometry where shock waves are partial solutions. Here, with the inclusion of the co smological constant it may be possible that the weak and dominant energy conditions could b e satisfied if L is positive and large enough in Eq. (2.25). 63. MATTER CONTENT ENERGY-MOMENTUM TENSORS 3.1. Perfect fluid energy momentum tensor For Eulerian observers 4-velocity ua= (1,-b,0,0) andua= (-1,0,0,0) the perfect fluid EMT for those observers is given by the expression below, Tab= (u+p)uaub+pgun, (3.1) whereuis a scalar function that represents the matter density, pis the fluid pressure, and gunis the metric tensor. It must be noted that the dust EMT is a particu lar case for the perfect fluid with null pressure. From the Einstein tensor components, Eqs. (2.10) to (2.19), and t he perfect fluid EMT it is possible to write all the components of Einstein equations. After s ome algebraic work we found the following set of equations, 4 3L = , 16p(T01+bT11) = 0, = 8p(T33+T22) = 16pp, = 0, = 0, (3.8) b yb z= 16pT23= 0, (3.9) 3.2. Solving the Einstein equations with Lfor the perfect fluid From Eq. (3.9) it follows that either b/z= 0, orb/y= 0, or both vanish. From Eqs. (3.7) and (3.8) it is easy to see that b/xcan also be zero. Those cases reveals which we will discuss in detail as follows. 7Case 0 andb x= this case Eqs. (3.2) to (3.9) simplify to, L = 4(L-8pu), (3.12) The set ofthese last equations implies that theshift vector bisnot uniquely defined. It is a complex valued function that depends only on ( t,y) spacetime coordinates. In the case of the dust EMT as p= 0, the warp drive metric is no longer a vacuum solution as it was found in [26], because the existence of the cosmological con stant as originated a solution that does not consider shock wave s via the 0 andb y= this case one has to solve the following equations, L = 8pu= 8pp= 0, 0. (3.14) Eq. (3.14) is the Burgers equation that connects the warp drive to shock waves, as discussed in Ref. [26]. The cosmological constant, fluid pressure an d matter density are all equal to zero and the warp drive metric (2.6) is a vacuum solut ion for the Einstein equations. Case 0 andb x= this configuration, the set of Eqs. (3.2) to (3.9) simplify to L = 4(L-8pu), (3.17) The above set of equations are very similar to Case 1a, where the sh ift vector is a complex valued function and it is not uniquely defined, but in this case bdepends on the (t,z) 0 andb y= this case, one has to solve the following equations L = 8pu= 8pp= 0, 0, (3.19) which is the same case as Case 1b. 3.3. Divergence for the perfect fluid EMT Calculating the divergence for the perfect fluid EMT, one arrives at the following (3.23) Besides, imposing the null divergence condition, Eqs. (3.20) to (3.23 ) implies that the pressure pdoes not depend on the spatial coordinates. Considering cases 1a and 2a there is another partial differential equation to solve, bu x+u t= 0, (3.24) and for Cases 1b and 2b Eq.(3.20) is trivially satisfied since u=p= 0. 94. ENERGY CONDITIONS 4.1. Weak Energy Conditions For this case the EMT at each point of the spacetime must obey the in (4.1) for any timelike vector u(uaua<0) and any null zero vector k(kaka= 0). For an observer with unit tangent vector vat a certain point of the spacetime, the local energy density measured by any observer is non-negative [29]. For the perfect flu id EMT the , (4.2) and the weak energy condition from Eq.(4.1) is satisfied if the matte r density uis positive. This is also the case for the dust EMT. 4.2. Dominant Energy Conditions For every timelike vector ua, the following inequality must be a non-spacelike vector, and the following condition must also |Tab|,for eacha,b . (4.4) Evaluating the first condition for the perfect fluid EMT we have that , Tabuaub=u. (4.5) The other condition FaFais given by the (4.6) Hence, the dominant energy condition is satisfied for u >0, as can be seen in Eq. (4.5). Besides, Eq.(4.6) is always satisfied no matter the sign of the matte r density. This condition also holds true if one considers the dust EMT as a particular case for the perfect fluid with null pressure. 4.3. Strong Energy Conditions For the strong energy condition the (4.7) 10is true for any timelike vector u. Computing the strong energy condition in Eq.(4.7) (4.8) and the strong energy condition stated in Eq.(4.7) is satisfied if 3 p+u>=0. The same is valid for the dust EMT, considering p= 0 for the perfect fluid, if u>=0. 4.4. Null Energy Conditions The null energy conditions are satisfied in the limit of null observers. For the null vector kthe following conditions must be any null vector ka. (4.9) Assuming that the following null vector kais given by, ka= (a,b,0,0), (4.10) we have that the relation between the components aandbare obtained by solving kaka= 0. The two solutions given (4.11) Then, the null energy condition (4.12) and the null energy condition may be satisfied if the following condition s are written as u+p>=0. (4.13) Eq.(4.13) is also true for the dust EMT if one considers it as a particu lar case for the perfect fluid with zero pressure, then the null energy condition is s atisfied for the dust if the matter density is positive. 5. DUST AS A PARTICULAR CASE FROM THE PERFECT constant that are widely known [29]. Considering the d ust EMT as a particular case for the perfect fluid by imposing the pressure pto be zero, the energy conditions for the warp drive metric and the dust EMT would be trivially satisfied, sinc e for this case, the solution of the Einstein equations is a vacuum solution [26]. 11TABLE I: Summary results for the perfect fluid energy conditi ons Energy condition Results Weak u>=0 Strong u>=0 Dominant u+3p>=0 Null u+p>=0 Table II summarizes the solutions of the Einstein equations for the p erfect fluid EMT with the cosmological constant and the warp drive metric. As can be seen there are two types of solutions and each is divided in two sub cases. Solutions 1b an d 2b are identical and require that L = p=u= 0, where the two solutions are the ones already found in Ref. [26] for the dust of non interacting particles EMT. This led to a v acuum solution of the Einstein equations and the connection between shock waves an d the warp drive via the Burgers 1a and 2a in table II have structures very similar to the one s of the same type of equations, but for the solution 1a the shift vector is a function o f both the time and the y-spatial coordinates, i.e., b=b(y,t). For solution 2a it is a function of both the time and thez-spatial coordinate, i.e., b=b(z,t). If we consider the dust solution as a particular case of perfect fluid with the imposition that the pressure is zero, we have that the four sets of partial d ifferential equations in Table II become a solution for the warp drive metric and the dust EMT with t he In the case of dust EMT there is no longer a set of equat ions 1a and 2a to be solved, only 1b and 2b, that are identical to the ones appearing in Re f. [26]. Even with a cosmological constant the dust EMT seems to be not a stable sourc e of matter, energy and momentum for the warp drive. 6. CONCLUSIONS AND FINAL REMARKS In this work we investigated how the presence of a cosmological con stant would affect the solutions of the Einstein equations endowed with the Alcubierre warp drive metric and the perfect fluid EMT as the source. Firstly, we solved the Einstein equa tions and obtained two solutions, Cases 1b and 2b, that are similar to the solutions we found for the dust particle without the cosmological constant [26], and two other solutions, Ca se 1a with b=b(t,y) andCase 2a with b=b(t,z), having the following equation of state relating L, the matter density uand the fluid pressure p: L = 2p(3u-p). The presence of the cosmological constant allows the shift vector to be a real valued function as can be seen from Eqs. (3.11) and (3.12) for Case 1a in Ta ble II, and Eqs. (3.16) and (3.17) for Case 2a, namely, L -8pu>=0 and L-8pp>=0. If we do not consider the cosmological constant, then the shift ve ctor would become a complex valued function for Cases 1a and 2a in Table II. The energy c onditions are all satisfied for the perfect fluid if the conditions in Table I are satisfied . Solutions 1b and 2b shown in Table II connect the Burgers equation to both the warp dr ive and the perfect fluid 12Case Condition Results 1)b z= 01a)b x= 0L = 0 (null 0L = 8pu= 8pp= =h(t) Null divergence is trivially satisfied This is the solution found in Ref. [26] 2)b y= 02a)b x= 0L = 0 (null 0L = 8pu= 8pp= =h(t) Null divergence is trivially satisfied This is the solution found in Ref. [26] TABLE II: Summary of all solutions of the Einstein equation w ith the cosmological constant and the Alcubierre warp drive metric having the perfect fluid EMT as mass-energy source. This table is also valid for the dust particle if considered as a particu lar case for the perfect fluid with In Ref. [26] we found this intrinsic relationship between the warp drive and shock waves by solving Einstein equations for the warp drive metric and the dust particle EMT, but we concluded that there is an impossibility of coupling the dust as a source in this case. So, the presence of shock waves would imply that the Alcubierre met ric shown in Eq. (2.6) is a vacuum solution for the warp drive. In Ref. [1] the Einstein equa tions were not solved, 13since the metric was merely guessed with a form function (see Eq. 2.8 ) that rules the warp bubble shape. The results found here led us to a kind of prescription where the war p drive requires more complex forms of matter than dust in order to obtain stable so lutions. In onesofthisseries ofpapers clear and that negative matter density may not be a strict requ irement to obtain warp speeds. The shift vector in the direction of the warp bubble moveme nt creates a coupling in the Einstein equations that requires off-diagonal terms in the EMT s ource. In the light perluminal interstellar travel could be understood as a complex distributio n of energy, matter and momentum sources that could stabilize the warp drive geometry, allo wing then thanks CNPq (Conselho Nacional de Desenvolvimento Cient' ifi co e Tecnol' ogico), Brazil's federal scientific supporting agency, for partial financial support: grant M. Alcubierre, The warp drive: hyper-fast travel within general relativity , Class. Quant. Grav. 11 (1994) L73, M. Alcubierre and F.S.N. Lobo, Warp drive basics . In: Wormholes, Warp Drives and F. Lobo (ed). Fundamental Theories of Physics, vol 189a (2017) 257-279. Springer, Cham., L.H. Ford and T.A. Roman, Quantum field theory constrains traversable wormhole geome tries, Phys. Rev. D 53 (1996) 5496, M.J. Pfenning and L.H. Ford The unphysical nature of warp drive , Class. Quant. Grav. 14 (1997) 1743, S.V. Krasnikov Hyperfast interstellar travel in General Relativity , Phys. Rev. D 57 (1998) 4760, A.E. Everett and T.A. Roman, A superluminal subway: The Krasnikov tube , Phys. Rev. D 56 (1997) 2100, F.S.N. Lobo and P. Crawford, Weak energy condition violation and superluminal travel , in Fern' andez-Jambrina L., Gonz' alez-Romero L.M. (eds) Cu rrent Trends in Lecture Notes in Physics, vol 617, (2003). Sp ringer, Berlin, F.S.N. Lobo and P. Crawford, Weak energy condition violation and superluminal travel , Lect. Notes Phys. 617 (2003) 277, C. Van Den Broeck, A warp drive with more reasonable total energy , Class. Quant. Grav. 16 (1999) 3973,

2108.10960

Marcelo Byrro Ribeiro

Osvaldo L. Santos-Pereira, Everton M. C. Abreu and Marcelo B. Ribeiro

Perfect fluid warp drive solutions with the cosmological constant

15 pages, LaTeX. Accepted for publication in The European Physical Journal Plus. arXiv admin note: text overlap with arXiv:2101.11467

Eur. Phys. J. Plus 136 (2021) 902

10.1140/epjp/s13360-021-01899-7

gr-qc

http://creativecommons.org/licenses/by/4.0/

The Alcubierre metric describes a spacetime geometry that allows a massive particle inside a spacetime distortion, called warp bubble, to travel global velocities. In this work we advance solutions of the Einstein equations with the cosmological constant for the Alcubierre warp drive metric having the perfect fluid as source. We also consider the particular dust case with the cosmological constant, which generalizes our previous dust solution and led to vacuum solutions connecting the warp drive with shock waves via the Burgers equation, as well as our perfect fluid solution without the cosmological constant All are also analyzed. The results show that the shift vector in the direction of the warp bubble motion creates a coupling in the that requires off-diagonal terms in the energy-momentum it seems that to achieve superluminal speeds by means of the Alcubierre warp drive spacetime geometry one may require a and distribution of energy, matter and momentum as source in order to produce a warp drive bubble. In addition, warp speeds seem to require more complex forms of matter than dust for stable solutions and that negative matter may not be a strict requirement to achieve global superluminal speeds.

Contents 1 Introduction 1 2 Performance Dependencies 3 2.1 Energy Resolution Dependencies 5 3 Discussion 8 1 some methods have been developed to calibrate such devices. However, another trend in calorimeter design pushes the number of readout channels to amounts which warrant We have started experimenting with applications of machine learning models - particularly networks or CNNs - to reconstruct energy from the many signal channels coming from such devices. The increasing channel count reduces occupancy which reduces the sample size available in any one channel for calibration; the increasing channel count also stresses the available for calibration activities. For example, a Cu/Si test stand 113 m3with scheme contains 187500cells. If this scheme is extended to a barrel plus endcap configuration at LHC detector size scales, the combined system contains computation, data usage, and speed of data collection. This development raises several concerns in relation to calibrating and extracting physical In this paper, we confine ourselves to matters related specifically to calibration. Primarily, we focus on studying a CNN's sensitivity to variations of the input data in order to determine the feasibility of any sort of calibration. In general, we have two options to correct the CNN's reconstructed energy. We can adjust the calorimeter readout channel values, the inputs to the model, which we shall call pre-model or we can adjust the predicted energy from the CNN reconstruction, the CNN output, which we shall call post-model. Historically, experiments have employed linear reconstruction methods, simple sums of cell responses, resulting in simple multiplicative factors to calibrate. A full-size, combined electromagnetic and hadronic high granularity calorimeter of LHC exper- imental proportions would have upward of 10s of millions of cells. Attempting to calibrate millions of channels via traditional techniques seems like an unreasonably large task. The degrees of 1 for ourselves since we now have to deal with a non-linear energy reconstruction as well. We studied the effect of miscalibration and noise on a CNN trained on a Cu/Si module having a224 cm3segmentation. In a previous paper, we described the reconstruction performance of this design [2]. The CNN we trained operates as a correction to the sum of energy of the the convolutional layers of the CNN consume the raw calorimeter cell energies and we inject the sum of cell energies at a later layer of the network. We found this architecture gives us on hadron initiated showers. The data consist of single charged pions ranging up to Ci, it adds a global bias b, and a stochastic noise specific to the cell and event. The bias band noise carry units of MeV to add to the base GEANT4 energy (1.1) We also applied a zero suppression of 06 MeVto the adjusted energy values which is it to the miscalibrated energy results in a fluctuation around the nominal value. We draw and noise values from the beta distribution with both parameters set to two, B22, and then we transform the variables to achieve the desired mean and scale of variation /one.sup. We adopted this approach because it sufficiently perturbs the CNN's inputs while mitigating the effects of events arising from long tails. We have no expectation to derive quantitative results that transfer to other detectors or models, and this approach allows us to deduce transferable, factors, (1.2) =s12B221o >>1sc1,sc 1/4and>>s s 1/4 for miscalibration and noise, correlation of noise and calibration amongst cells results from physical aspects of a detector design. Lab bench calibration of sub-modules, common electronics and read out boards for groups of channels, upstream material, etc. lead to groups of read out channels showing similar noise and response characteristics. In order to draw random noise values and miscalibration factors from with symmetric parameters greater than 1 has both the median and mode equal to1 2[4]. - 2 -Table 1. The calibration, noise, and bias parameters scanned in a Scale Noise Scale [MeV] Bias [MeV] 0. 0. 0. 0.15 0.1 -0.15 0.2 0.1 0.15 0.25 0.3 0.3 distribution that captures the variable dependencies, we turn to vine copulae [5, 6]. Modeling a joint distribution with a copula allows us to consider the univariate, marginal distributions separately from the variable ( Vine copulas further help us to structure the variable by only considering pairs of partial correlations. We divided the 505075test stand module into smaller sub-module sections; we chose a vine structure such in a way that mimics common detector readout electronics 1 lists the parameter values we scanned in a grid to study the CNN sensitivity. We applied the parameters to modify the GEANT4 simulated cell energies prior to scoring the CNN. the respective elements. 2 Performance 1 shows the mean response as a function of the initiating pion energy for the indicated in the legend. As was discussed in [2], the CNN acts like a correction to Since the zero suppression prevents negative contributions and the miscalibration factor is both multiplicative and positive, the energy sum is always positive. We also notice in Figure 1 the significant over-estimation of energy by the CNN in some scenarios. Typically, the sum of visible energy is a small fraction of the energy of the initiating particle. Consequently, when the can simply re-calibrate the response. - 3 -Figure 1. The mean CNN response against initiating pion energy for all of the scanned calibration parameters. This collection of curves gives some visual justification for our assumption of linearity. We fit linear models, ^EaEbeam,b, to the response data. This parameterization gives us Oneexpects a1andb0. Performing this fit for each point in the parameter scan reveals the reconstruction's sensitivity to variations of the input data with respect to the training data. Figure 2 summarizes the results. Figure 2. Summary of response fits: the subplots show the distributions of slope and intercept of the CNN response against the calibration, noise, and bias scales. The error bars represent the sample standard deviation. - 4 -In Figure 2, we can see that noise and bias have a stronger influence than the calibration scale. We can also see the impact of the zero suppression, 06 MeV, in the relationship that noise scale has on the fit parameters. We then fitted independent linear models to the response slope, a, and to the intercept, b. We investigated fitting aas dependent upon b, but found it has a statistically insignificant we rejected the dependence. Equations 2.1 and 2.2 enumerate the final estimated (2.2) We find a relatively weak relationship between the bias and the slope compared to that of noise scale. Interestingly, we observe an interaction between the noise scale and bias in both fits. This implies that controlling noise should take the greater share of concern in controlling accuracy. These results are promising in that we see not only linear responses throughout the scan but we also see linear relationships between the scanned parameters and response parameters. This result justifies a multiplicative calibration factor for post-model calibration; it also implies that we can as noise scale. 2.1 Energy Resolution we also studied the impact of miscalibration, noise, and bias on the reconstructed of the CNN. This resolution study is based on a reduced beam energy scale, >>0125 1/4GeV, to reduce the influence of leakage in higher energy events. Figure 3 shows the results of this scan in the traditional plot of 5 -Figure 3. CNN reconstructed energy resolution vs.the inverse square-root of the beam energy for noise, and bias parameters. We fit these resolution curves with a three term parameters corresponding to constant, stochastic, and noise terms respectively. No implements the least-squares methodology. Figure 4 summarizes the estimated parameters. - 6 -Figure 4. The estimated resolution fit parameters are summarized (mean and standard deviation) along , noise, and bias dimensions explored. In order to understand the CNN's energy resolution, we fitmodels to estimate the effects the CNN 2.4, 2.5, and 2.6 report the final estimated effects. We follow a step-wise process to remove insignificant variables from the (2.6) The of the resolution parameters ensures non-negative estimated values. We note that in the high noise and bias region ( suppressed in favor of the N parameter. energy resolution is no longer characterized by energy dependent Poisson statistics; it is rather - 7 miscalibration scale. 3 Discussion We find the consistent linearity of the CNN response under various input conditions encouraging for future use of these types of reconstruction models. It is particularly noteworthy that the observed up to 500GeV extends well beyond the upper bound of the model build data that had an upper limit at 125GeV; that is to say the CNN response linearity extrapolates beyond the builddata. and additive bias offsets. In fact, inverting the response fits discussed above achieves such a calibration. With such an approach, one can calibrate a single test beam stand or improve may be a boon to future experiments. Nevertheless, we should remain somewhat cautious because we have no guarantee that all CNNs will have the same linear response. We also note that comprehending the impact of miscalibration, noise, and bias on the is more complicated. Nevertheless, we can conclude from the results the to our initial expectations. When noise and bias are under control, the resolution follows bias are not under control, the resolution is characterized almost purely by the noise term. The transition between low and high noise-bias regions may be better understood if we had modeled tails of the miscalibration and noise scales. However, since the general trend contains no surprises, we do not think expending additional resources to study this would result in any interesting finding. This point should be particularly salient when one considers that the particular numerical results do not translate form detector configuration to results should be read with the caveat that miscalibration, noise, and bias effects do not depend on depth. Therefore, this study does not model well the effects of noise induced by for future work to study how a CNN performs given real test beam Texas Tech University, Office of the Vice President for Research and O. Ganel, R. Wigmans, On the calibration of longitudinally segmented calorimeter systems, and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 409 (1) (1998) 621-628. - 8 -

2108.10963

Christopher Cowden

N. Akchurin, C. Cowden, J. Damgov, A. Hussain, and S. Kunori

Perspectives on the Calibration of CNN Energy Reconstruction in Highly Granular Calorimeters

physics.ins-det hep-ex

http://creativecommons.org/licenses/by/4.0/

We present a study which shows encouraging stability of the for a simulated high granularity calorimeter module reconstructed by a CNN model to miscalibration, bias, and noise effects. Our results also show an intuitive, quantifiable relationship between these factors and the calibration parameters. We trained a CNN model to reconstruct energy in the calorimeter module using simulated single-pion events; we then observed the response of the model under various miscalibration, bias, and noise conditions that affected the model input. From these data, we estimated linear response models to calibrate the CNN. We also quantified the relationship between these factors and the calibration parameters by regression analysis.

(VFA) approach, e.g. Deoni et al. [2], Helms et al. [3]. A common assumption when computing is that certain multiplicative factors, such as the signal intensity modulation imposed by the receiver coil's net sensitivity prole, are constant across the weighted volumes. However, this is invalid if motion occurs between the volume acquisitions. In the case of neuroimaging, rigid body co-registration can be used to realign the brain but will not correct for the dierential coil sensitivity modulation, which in R1maps computed with the VFA approach can lead to mean absolute error approaching 20% [4]. A correction scheme has previously been proposed by Papp et al. [4] and validated for R1mapping at 3T. The position-specic net receive sensitivity is estimated from two rapid low-resolution magnitude images, received on the body and array coils respectively prior to each VFA acquisition. The more of the body coil is used as a reference to compute the net receiver sensitivity, which is then removed from the VFA acquisitions. This approach eectively assumes that the body coil's modulation is volumes instead of that of the array coil. This in itself is a potential limitation, as is the at body coils at higher eld strengths. Here we propose an alternative whereby we estimate the relative sensitivity between volumes. This approach does not fully remove the receiver's sensitivity modulation but does remove the bias that introduces in quantitative metrics. Only the calibration images obtained with the array coil are required, i.e. less data than the originally proposed method [4]. To validate the approach, we focus on R1 maps computed with the multi-parameter mapping (MPM) protocol [5]. We rst compare performance with the established method of Papp et al. at 3T [4] and then demonstrate a reduction of inter-scan motion artefacts at 7T under a range of dierent motion conditions. We further demonstrate that, unlike at 3T, the transmit eld B+ 1also exhibits substantial position-specic variability at 7T. As a result, the most were obtained by accounting for both position-specic transmit and receive sensitivity eects. While we validate this approach in the context of R1mapping, it has much more general potential and can be applied to other mapping methods that combine data from multiple volumes. 2 Methods 2.1 Theory R1mapping can be achieved by acquiring spoiled gradient echo volumes with at least two dierent ip angles [2, 3]. At a given spatial location, the image intensity, I, for a given nominal ip angle exp(TRkR1); (1) wheresis the receive sensitivity, is the proton density, fTis the transmit eld, R1is the rate, TR is the repetition time, and kindexes the VFA acquisition. Co-registration allows for inter- scan motion by realigning anatomical structure across acquisitions. Under the small ip angle with two nominal ip angles ( k=f1,2g),R1can be computed as (2) Typically, it is assumed that s1=s2and the sensitivities simplify out. However, this assumption is invalid if inter-scan motion has occurred leading to substantial bias in R1estimates [4]. This can be avoided by accounting for the relative sensitivity across positions:
1;2=s1=s2. Substitution for s1in equation (3) 2The method of Papp et al. [4] did not include the relative sensitivity but referenced to an additional acquired on the body coil, assuming that the body coil modulation was is commonly assumed that the transmit eld is suciently smooth as to be considered i.e.fT1=fT2, such (4) However, in this work we show that this assumption does not hold at 7T, and that incorporating transmit eld estimates maximises the precision of R1. Ratio approach The calibration data used to correct inter-scan motion artefacts comprised rapid low resolution, coil-combined magnitude images acquired immediately prior to each high resolution VFA acqui- sition. These images, fxkgK k=1, assumed to have been rigidly co-registered to the same space, can be written as the product of a common image rand a net sensitivity eld fskgK k=1. The relative sensitivity, k;refcan be computed with respect to one of the calibration acquisitions, used as a
k;ref (5) Dividing each VFA acquisition by its relative sensitivity
k;refresults in a common modulation, sref, which, although less homogeneous than the body coil used by Papp et al. , more faithfully restores the validity of assuming common modulation when computing R1. The assumption that ris common, such that k;ref=
k;refholds only if there are no eld eects. Simulations were used to explore the validity of this approach This ratio approach risks noise amplication, particularly in regions of ratio (SNR). This is combatted by isotropically smoothing xkandxrefbefore taking their ratio. A potentially more robust alternative is to cast the computation of the relative coil sensitivities, and a common image modulated by them, as an inference problem in a probabilistic generative model of xk that incorporates noise and can also embed knowledge about the spatial smoothness of the generative modelling approach allows coils with arbitrary sensitivity to be incorporated, e.g. coils with more (array) or less (body) spatial variation, or both concurrently (array + body) if available. A priori knowledge about the expected smoothness of the sensitivity can be incorporated at the level of coil type (body versus array) via appropriate tailoring of a regularisation parameter, . Images acquired with the body coil will have a atter sensitivity eld modulation, which can be incorporated by setting details are given in Appendix A. 2.2 One participant (F, 31 years) was scanned at 3T (MAGNETOM Prisma, Siemens, using a body coil for transmission and either the body coil or a 32-channel head array coil for reception. Three additional participants (2F, 1M; 32 - 41 years) were scanned at 7T (MAGNETOM Terra, Siemens, Erlangen, Germany) using an 8-channel transmit, 32-channel receive head array coil (Nova MA, USA) in a quadrature-like (TrueForm) mode. All data were acquired with approval from the UCL research ethics committee. MPM Datasets MPM data were acquired using a multi-echo spoiled gradient echo sequence with ip angles of 6(PD-weighted, PDw) and 26(T1-weighted, T1w), a TR of 19.5 ms and an RF of 117with a total dephasing gradient moment per TR of 6 . Eight echoes were acquired with TE ranging from 2.56 ms to 15.02 ms in steps of 1.78 ms using a bandwidth of 651 Hz/pixel. Data were acquired with a nominal 1 mm isotropic resolution over a eld of view of 160 mm right-left and 192 mm in the and directions. Elliptical sampling and partial Fourier, with factor 6/8 in each phase-encoded direction, were used to accelerate the acquisition, leading to a scan time of 5 minutes per volume. A B+ 1map was estimated by acquiring a series of spin and stimulated echoes using 3First et 1: 3T Example. The acquired calibration images, xkhave dierent orientation due to between acquisitions (1st row). Co-registration can align the images spatially, but does not correct for their dierential sensitivity eld modulation (2nd row), visible via their ratio, x1=x2=
1;2. The method from Papp et al. [4] estimates and corrects this modulation using an additional body-coil image (3rd row). When one is not available, relative signal dierences can be corrected for using the relative modulation
1;2 (4th row). Alternatively, the joint log-likelihood of a generative forward model that embeds the from a mean image, r, to native space can be maximised to determine the mean image and modulating sensitivities, sk;r, that best explain the acquired images xk(5th row). The generative produces a similar relative modulation ( s1;r=s2;r=
1;2) but allows for the corrected images to have the minimal modulation of the mean described 3T and 7T protocols [6, 7]. These data were acquired with 4 mm isotropic in a total acquisition time of 3 minutes 48 seconds, and a further 1 minute for B0mapping. For inter-scan motion correction, additional single echo acquisitions were acquired prior to each VFA acquisition with a ip angle of 6, TE = 2.4 ms, TR = 6.5 ms, a bandwidth of 488 Hz/pixel and no acceleration schemes. At 3T, these data were acquired, receiving sequentially on the array and body coils, with 8 mm isotropic resolution leading to a scan time of 6 s per volume. To capture the greater in the net sensitivity eld at 7T, the resolution was increased to 4 mm isotropic leading to a scan times of 18 s per volume, but acquired only on the array coil due to the absence of a body coil. Motion Conditions Two MPM datasets were acquired to dene baseline reproducibility. then instructed to move to a new, arbitrary position within the connes of the coil. A localiser was acquired and the eld of view repositioned as necessary to ensure appropriate brain coverage in the new position. A third MPM data set was then contrast For eachR1map computed from data across two positions, i.e.with inter-scan motion, two dierent corrections for transmit eld inhomogeneity were performed. The rst assumed the transmit eld was identical across head positions (Eq. (4)) and in the same position as the PD-weighted volume. The second used position-specic B+ 1maps (Eq. were computed using the hMRI toolbox [8], which uses the small ip [3] and corrects for imperfect spoiling [9], which always used the B+ 1map acquired in the space of the PDw acquisition. Maps were computed with and without inter-scan motion using all possible PDw and T1w combinations. To ease comparisons, all maps were constructed in the space of the rst PDw volume. Rigid transformations between all volumes and the rst PDw volume were estimated using SPM12 (Wellcome Centre for Human Neuroimaging) having rst corrected for intensity non-uniformity and skull- stripped the images. R1maps were computed with and without the proposed inter-scan motion At 3T, R1maps were also computed with the method of Papp et al. [4]. An isotropic kernel of 12mm { the default in the hMRI toolbox { was used to smooth the s prior to computing the relative (proposed) and absolute (Papp et al. ) sensitivities. The approaches used array = 107,body= 109, and 15 metric Three (two from position one and the third from position two) R1maps, with no motion corrections applied, were averaged to produce a `ground-truth' map, ^R1. All of the available R1maps for each participant and condition (motion/no motion) were assessed against this reference to quantify the error, and its variability. The set of R1maps used to compute the reference were also segmented to create a mask selecting those voxels with a mean probability of being in WM, GM or CSF greater than 50%. For participant 3, the cerebellum was excluded, using the SUIT toolbox [10] in SPM, as a failure caused by excessively large o-resonance. For the Nvoxels within the mask, the mean absolute error, MAE, for each R1map was computed with respect to the `ground-truth' map, ^R1as: MAE (6) These errors are reported as percentages. 2.3 Simulation study Validity of assumptions Eq (5) assumes that the calibration data are insensitive to changes in the transmit eld across positions such that k;ref=s1=s2. Here we test the validity of this assumption via the small ip angle approximation and allowing for position-specic transmit and receive elds, the calibration images can be written (7) 5Figure 2: Misestimation of the true relative sensitivity (
1;2) by the ratio of calibration images ( 1;2), as a function of transmit elds. Colours encode@ @
1;2, which is 1 when either fT1=fT2(black doted line) dotted the repetition of the acquisition, i.e.the calibration data for each high resolution VFA acqui- sition, and cdenotes the sequence settings. Considering the ratio method for then be written more fully as: 1;2=

the ratio of the calibration images, 1;2, to equal the relative sensitivity,
1;2, we 1 (9) We note that the Ernst angle is E= 2TRcR1, and rearrange to =fT1fT2(fT2fT1) (10) This condition is met when is no change in transmit eld, or when conditions are highlighted in Fig. 2 which shows@ @
1;2as a function of fTjforR1= 0:84s1. 1 is within 3% for a broad range of values centred on both acquisitions being at the error Numerically, errors in the R1estimates were computed for the empirically observed range of relative transmit and receive elds. The median proportion of error arising from transmit or receive eld changes was computed over this 4D parameter space. 2.4 Code code used to t the generative model is available at . A modied version of the hMRI toolbox that integrates this approach and enables B+ 1correction on a per-contrast basis is available at . The ratio approach can be performed natively with the hMRI toolbox: . The source code to reproduce the simulation gures is available at: . 6No arrayPapp et al. Ground-truthR1 Map R1 histogram Error 0.41.4 1 1.5 0.5 1 1.5 0.5 1 1.5 0.5 1 1.5 0.5 1 1.5 0.5 1 1.5Figure 3: Results at 3T. The rst row shows example R1maps constructed with each method. The second row shows (normalised) error maps with respect to the ground-truth map ^R1. The third row shows area) of R1within the GM (green) and WM (purple), and their log-Normal t (solid line); these histograms display probability distributions and therefore integrate to 1. Generative / only array used only the array coil images in the generative modelling framework, whereas Generative / both the array and body coil images using coil-specic regularisation for the smoothness of the sensitivity modulation. 3 Results Exemplar images, relative sensitivities, and results from the generative modelling are shown in gure 1. The R1and error maps obtained at 3T and 7T are shown in gures 3 and 4 respectively. The means and of the MAE are reported in Table 1. The dierential impact of correcting for transmit and receive eld eects is illustrated in gure 5. This shows R1and error maps without motion, and with motion (i) no correction, (ii) correction only for receive eld eects, (iii) only for transmit eld eects, or (iv) for both eects in combination. 3.1 3T Validation The net motion is summarised as the across the three orthogonal axes, of the translations or rotations independently. The net translational and rotational motion in the no motion condition was 0.8 mm and 0.3 degrees. These were increased to 1.2 mm and 18.1 degrees in the inter-scan motion case. When the T1w and PDw volumes were acquired in the same position, the MAE captured the which was approximately 3% at 3T and 4-5% at 7T. In the absence of overt motion, correcting for the dierential sensitivity modulation did not substantially change the MAE. In the presence of overt motion, the MAE rose to 10%. It was reduced to 4.7% by the method of Papp et al. and to less than 7Table 1: MAE (mean s.d. across repeats, in %) with respect to the average reference ^R1. Dataset Motion No correction Ratio + body)Papp et al. 3T #1 No 3.00.1 3.0 0.1 3.0 0.1 3.1 0.2 3.7 0.2 Yes 10.10.8 4.4 0.4 4.3 0.3 4.3 0.1 4.7 0.3 Dataset Motion No correction Ratio Generative B+ 1per contrastRatio & B+ 1per & B+ 1per contrast #1 No 4.00.3 4.1 0.4 4.1 0.3 4.1 0.2 4.1 0.4 4.1 0.3 Yes 5.80.7 4.8 0.3 4.9 0.4 4.9 0.3 3.8 0.3 3.9 0.3 7T #2 No 4.80.2 4.9 0.3 4.9 0.3 4.8 0.2 5.0 0.2 4.9 0.2 Yes 8.40.4 6.1 0.3 6.2 0.3 8.0 0.4 5.3 0.1 5.5 0.1 #3 No 5.10.2 5.0 0.2 5.0 0.2 5.0 0.2 5.1 0.3 5.1 0.2 Yes 13.41.9 7.9 1.1 8.3 1.2 11.8 0.9 6.3 0.3 6.8 0.3 4.4% by the proposed correction schemes, with or without incorporating the body coil in the approach. The histograms in gure 3 conrm that the method did not introduce any bias to Extension to 7T At 7T, the range of motion varied across participants. The net translational and rotational motion in the no motion conditions did not exceed 1.6 mm and 1.0 degree respectively. In the inter-scan motion cases, the net translation ranged from 1.6 to 7.9 mm, while the net rotation ranged from 2.7 to 11.2 motion led to more apparent artefacts. The overall amplitude of motion dictated the increase in MAE, which reached a maximum of 13.4% under the tested conditions (c.f. motion summaries in gure 4 and MAE in table 1). The proposed correction scheme reduced the MAE (5-8%), though not to the level of no overt motion. The variability of the transmit eld, B+ 1, across head positions was found to be much higher than at 3T. Incorporating position-specic B+ 1maps reduced the MAE (5-12%) even without correcting for the dierential receive sensitivity modulation. The greatest reductions in MAE were obtained by correcting for position-specic transmit reaching 4-7%, converging on the level obtained in the absence of overt motion (i.e. 4-5%). 3.3 Comparison of Methods Overall, the ratio and generative modelling approaches to correcting the eects of dierential relative sen- sitivities in R1maps performed similarly. The MAE was marginally lower for the generative at 3T (0.1%) and for the ratio approach at 7T (0.5%). However, these dierences were small relative to the variability across cases (Table 1). 3.4 Numerical R1Error At 3T the relative transmit eciency ranged from 0.97 to 1.04, whereas the relative receive eld ranged from 0.84 to 1.18. At 7T the relative transmit eciency ranged from 0.85 to 1.18 under comparable motion conditions. These ranges were used in the simulations which revealed that inter-scan motion caused error as high as 130%. Over the 4D parameter space investigated, a median of 29% of the error was caused by transmit eld eects and 71% by receive eld eects. Fig. 6 shows a plane of this error as the relative transmit and receive elds change. Position-specic fToers only (Fig. 6 B). Larger error reduction arises from receive eld correction (Fig. 6 C). Combining both (Fig. 6 D) shows receive eld eects are removed but transmit sensitivity remains (when @=@
1;26= 1). 80.2 1.4 s-1Correction:R1 Map NoError with inter-scan +50% Generative Ratio Generative Ratio No Figure 4: Results at 7T. The rst column shows the ground-truth map ^R1for the three participants sorted based on the magnitude of inter-scan motion. Uncorrected and corrected R1maps are shown in the middle with the corresponding (normalised) error maps with respect to the ground-truth on the right. Correction is only applied for net receive sensitivity modulation and not for transmit eld eects. Rows 1 to 3 of the gure correspond to datasets 1, 2 and 3 as reported in Table 1. The net motion is summarised as the across the three orthogonal axes, of the translations or rotations independently. In the absence of overt motion, the average displacements between the VFA scans, across the group, were 1 mm and 0.6 degrees for translations and rotations respectively, with a maximum translation of 1.6 mm and a maximum rotation of 1.0 degree. 9B1+ per Contrast R1 map Error 0.21.4 Sensitivity Correction With Inter-scan Motion Figure 5: Combining receive sensitivity and B+ 1correction at 7T, for participant 3 (last row in gure 3). The rst row shows an example R1map without and with inter-scan motion, before and after net correction, and employing a separate B+ 1map for each contrast in the case of inter-scan motion. The second row shows (normalised) error maps with respect to the ground-truth map ^R1. In this motion correction was performed with the generative modelling approach. Figure 6:R1error (in percentage of the true ^R1= 0:84s1) as a function of the relative transmit eld fT1=fT2 and relative receive sensitivity
1;2between two head positions. The four panels show this error with dierent degrees of correction: (A) none, (B) correction for position-specic transmit eld, (C) correction for position specic receive sensitivity, (D) both corrections. Note that minimal error remains even with both corrections, as position-depend transmit eects lead to inaccuracies in the estimation of
1;2by1;2. 104 Discussion We have introduced methods for correcting inter-scan motion artefacts in quantitative MRI that do not rely on the availability or spatial homogeneity of a body coil. The approaches are based on estimating the modulation across positions, and successfully reduced error in R1maps at both 3T and 7T. At 3T, the proposed approaches outperformed a previously established correction method [4]. This can be attributed to the fact that the method of Papp et al. assumes that the reference modulation of the body coil is independent of position, whereas the proposed methods do not. Instead they specically account for therelative sensitivity across positions thereby restoring consistent modulations. In the motion conditions tested here, both proposed approaches (ratio with Gaussian smoothing, or generative modelling) produced comparable improvements in in the presence of Equally importantly, when there was no overt motion neither method decreased was at a level in keeping with previous reports for similar resolution MPM data [11, 5]. The ratio method benets from its simplicity, but may be vulnerable to low SNR given that it denes one calibration image as the reference (denominator in equation (5)). The alternative generative has the benet of inherently adapts to variable SNR by estimating the position-specic net sensi- tivity modulation relative to a common image, which is their barycentre mean. This common image dictates the nal modulation of allthe corrected volumes. The generative model can also easily incorporate any additional data, e.g. body coil images as done at 3T, which attens the nal modulation. registration could be interleaved with model tting [12] to reach a better global optimum. Finally, the generative model could naturally be integrated with any tting approach that denes a joint probability over all acquired data, such as Balbastre et al. [13] in the context of MPM. The impact of movement on the eective transmit eld has previously been investigated in the context of specic absorption rate management [14, 15, 16, 17]. An important additional nding of the present work is the impact this can have on R1estimates at 7T, which was negligible at 3T as demonstrated methods are specically designed for the correction of inter-scan motion and therefore cannot motion, which may be more likely to occur coincidentally with inter-scan motion, e.g. with uncompliant participants. Although the dominant source of error in R1was related to receive eld eects, the MAE was further reduced by additionally accounting for the of the transmit eld. However, acquiring a B+ 1map at multiple positions comes at a cost of increased scan time and inevitably leads to a greater temporal separation between the calibration data and those volumes it is used to correct. Issues such as this, coupled with other uncorrected eects, e.g. B0eects (no reshimming was performed during the experiments), may underlie the fact that the corrections implemented do not reduce the MAE quite to the level of no motion. This nding recapitulates those of Papp et al., though the discrepancy is lower in this work, which is likely because the assumption of a at body coil receive sensitivity is no longer made. However, even with the combined receive and transmit eld corrections, the MAE is never reduced to the level of no motion. This is in line with the simulations, which show that transmit eld eects remain in the calibration data and propagate into the R1estimates. An additional limitation of the generative model is its reliance on a Gaussian noise assumption, which is violated in the background (but not in the tissue, given the high SNR of the calibration scans). Although we did not nd this violation to hamper sensitivity estimation in the present study, the model could nonetheless be modied to incorporate a Rice or noncentral Chi likelihood [18]. 5 motion causes serially acquired weighted volumes to be dierentially modulated by sensitivities leading to substantial errors when they are combined to compute quantitative metrics. We 11have demonstrated the ecacy of two methods at reducing these artefacts in the context of R1mapping. The proposed methods do not require a body coil making them ideally suited for use at 7T, and can be extended to the computation of other quantitative metrics, such as magnetisation transfer saturation [19], that similarly assume constant modulation across multiple weighted acquisitions. Given that the application of this correction does not degrade reproducibility in the no inter-scan motion condition we would recommend it always be Nikolaus Weiskopf, Luke J Edwards, Gunther Helms, Siawoosh Mohammadi, and Evgeniya magnetic resonance imaging of brain anatomy and in vivo histology. Nature Reviews Physics , pages 1{19, 2021. [2] Sean CL Deoni, Brian K Rutt, and Terry M Peters. Rapid combined T1 and T2 mapping using acquisition in the steady state. Magnetic Resonance in Medicine: An Ocial Journal of Society for Magnetic Resonance in Medicine , 49(3):515{526, 2003. [3] Gunther Helms, Henning Dathe, and Peter Dechent. Quantitative FLASH MRI at 3T using a of the Ernst equation. Magnetic Resonance in Medicine: An Ocial Journal of Society for Magnetic Resonance in Medicine , 59(3):667{672, 2008. [4] Daniel Papp, Martina F Callaghan, Heiko Meyer, Craig Buckley, and Nikolaus Weiskopf. Correction of inter-scan motion artifacts in quantitative r1 mapping by accounting for receive coil sensitivity eects. Magnetic resonance in medicine , 76(5):1478{1485, 2016. [5] Nikolaus Weiskopf, John Suckling, Guy Williams, Marta Morgado Correia, Becky Inkster, Roger Tait, Cinly Ooi, Edward T Bullmore, and Antoine Lutti. Quantitative multi-parameter mapping of R1, PD*, MT, and R2* at 3T: a multi-center validation. Frontiers in neuroscience , 7:95, 2013. [6] Antoine Lutti, Chloe Hutton, J urgen Finsterbusch, Gunther Helms, and Nikolaus Weiskopf. Opti- mization and validation of methods for mapping of the radiofrequency transmit eld at 3T. in Medicine , 64(1):229{238, 2010. [7] Antoine Lutti, Joerg Stadler, Oliver Josephs, Christian Windischberger, Oliver Speck, Chloe Hutton, and Nikolaus Weiskopf. Robust and fast whole brain mapping of the RF transmit eld B1 at 7T. PLoS One , 7(3):e32379, 2012. [8] Karsten Tabelow, Evelyne Balteau, John Ashburner, Martina F Callaghan, Bogdan Draganski, Gunther Helms, Ferath Kherif, Tobias Leutritz, Antoine Lutti, Christophe Phillips, et al. hMRI{A toolbox for quantitative MRI in neuroscience and clinical research. Neuroimage , 194:191{210, 2019. [9] C Preibisch and R Deichmann. In uence of RF spoiling on the stability and accuracy of T1 mapping based on spoiled FLASH with varying ip angles. Magnetic Resonance in Medicine: An Ocial Journal of the International Society for Magnetic Resonance in Medicine , 61(1):125{135, 2009. [10] J orn Diedrichsen. A spatially unbiased atlas template of the human cerebellum. NeuroImage , 33(1): 127{138, 2006. ISSN 1053-8119. [11] Tobias Leutritz, Maryam Seif, Gunther Helms, Rebecca S Samson, Armin Curt, Patrick Freund, and Nikolaus Weiskopf. Multiparameter mapping of relaxation (R1, R2*), proton density and saturation at 3 T: A multicenter dual-vendor reproducibility and repeatability study. Human brain mapping , 2020. [12] John Ashburner and Gerard R Ridgway. Symmetric dieomorphic modeling of longitudinal structural MRI. Frontiers in neuroscience , 6:197, 2013. 12[13] Yael Balbastre, Mikael Brudfors, Michela Azzarito, Christian Lambert, Martina F Callaghan, and John Ashburner. Model-based multi-parameter mapping. Medical Image Analysis , 73:102149, 2021. [14] Emre Kopanoglu, Cem M Deniz, M Arcan Erturk, and Richard G Wise. Specic absorption rate impli- cations of within-scan patient head motion for ultra-high eld MRI. Magnetic resonance in medicine 2020. [15] Morgane Le Garrec, Vincent Gras, Marie-France Hang, Guillaume Ferrand, Michel Luong, and Nicolas Boulant. Probabilistic analysis of the specic absorption rate intersubject variability safety factor in parallel transmission MRI. Magnetic resonance in medicine , 78(3):1217{1223, 2017. [16] S Wolf, D Diehl, M Gebhardt, J Mallow, and O Speck. SAR simulations for high-eld MRI: how much detail, eort, and accuracy is needed? Magnetic resonance in medicine , 69(4):1157{1168, 2013. [17] Gunamony Shajan, Mikhail Kozlov, Jens Homann, Robert Turner, Klaus Scheer, and Rolf Pohmann. A 16-channel dual-row transmit array in combination with a 31-element receive array for human brain imaging at 9.4 T. Magnetic resonance in medicine , 71(2):870{879, 2014. [18] Divya Varadarajan and Justin P Haldar. A framework for Rician and non-central chi MR images. IEEE transactions on medical imaging , 2015. [19] Gunther Helms, Henning Dathe, Kai Kallenberg, and Peter Dechent. High-resolution maps of magne- tization transfer with inherent correction for RF inhomogeneity and T1 relaxation obtained from 3D FLASH MRI. Magnetic Resonance in Medicine: An Ocial Journal of the International Society for Magnetic Resonance in Medicine , 60(6):1396{1407, 2008. [20] John Ashburner. A fast dieomorphic image registration algorithm. Neuroimage , 38(1):95{113, 2007. A Generative Model of Images The proposed generative model is similar to that of Ashburner and Ridgway [12] (section 2.3), deformations. The magnitude images of the calibration can be written as the voxel-wise product of a mean image rand the net sensitivity additive noise, approximated as Gaussian with variance 2 k, which is assumed to be uncorrelated across xk. The net sensitivity elds can be written as diagonal matrices Sk= diag ( sk), such that the corresponding is: p(xkjsk;r) (11) If the images have been co-registered but not resliced, a mean space can be dened by computing the barycentre of all aligned orientation matrices [12]. The linear operation of resampling an image from mean to acquired space can be encoded by the matrix Ak2RNNsuch that the conditional probability (12) While this is the approach taken in practice, the following derivation is restricted to the resliced case for clarity. The magnitude of each sensitivity eld is unknown since the intensity scaling depends on many param- eters. However, the sensitivities are known to vary smoothly in space, which is captured by a that penalises the eld's bending energy [20], which integrates the squared curvature of the sensitivity eld, making it invariant to intensity scaling. The net sensitivities are encoded by their logs, ( i.e., sk= exp zk), such that invariance under shifts in log-space implies invariance under scales in which is also incorporated into the prior. In a discrete setting, computing the bending energy re- duces to the quadratic term zTLz; The prior distribution over sensitivities is therefore dened as a over their logs: p(zk) (13) wherekis an image-specic regularisation factor. The joint model likelihood, or its negative log, is obtained by combining the likelihood in (11) and the prior in This is minimised with respect to fzkgK k=1andr. Neglecting terms that do not depend on these variables, yields the objective +k 2zT kLzk + const: with respect to the mean image, r, while keeping the sensitivities xed gives a voxel-wise ( n) closed-form update: rn (15) The have no closed-form solution necessitating an iterative method. The objective function is not everywhere convex, but the likelihood term resembles that of a nonlinear least-squares problem, which can be solved using Gauss-Newton optimisation. Gauss-Newton is a modication of Newton-Raphson that uses Fisher's method of scoring, which amounts to replacing the Hessian at any point with its value at the optimum. With R= diag ( r) and Xk= diag ( xk), the gradient gkand Hessian Hkused in the Newton- Raphson iteration +kLzk; scoring)0+kL: (17) This ensures that the Hessian used in the Newton-Raphson iteration is positive denite, but does not ensure that the iteration monotonically improves the objective function. We therefore replace the with the more robust (18) zk zkP1 kgk; (19) which has been shown to yield monotonic convergence [13]. The inversion in equation (19) is performed with a full multi-grid solver that leverages the sparsity and structure of the preconditioner [20]. Finally, a global scaling eld s= exp z, applied to both the sensitivities ( skn skn=sn) and mean image (rn snrn), ensures that the product sknrnis unchanged. Keeping only terms of the objective function that depend on this scaling eld (20) By dierentiating, the optimal scaling eld (21) 14

2108.10943

Ya\"el Balbastre

Ya\"el Balbastre, Ali Aghaeifar, Nad\`ege Corbin, Mikael Brudfors, John Ashburner, Martina F. Callaghan

Correcting inter-scan motion artefacts in quantitative R1 mapping at 7T

eess.IV cs.CV

http://creativecommons.org/licenses/by/4.0/

Purpose: Inter-scan motion is a substantial source of error in and can be expected to increase at 7T where $B_1$ fields are The established correction scheme does not translate to 7T since it requires a body coil reference. Here we introduce two alternatives that outperform the established method. Since they compute relative do not require body coil images. Theory: The proposed methods use coil-combined magnitude images to obtain the relative coil sensitivities. The first method efficiently computes the via a simple ratio; the second by fitting a more model. Methods: $R_1$ maps were computed using the variable flip angle (VFA) approach. Multiple datasets were acquired at 3T and 7T, with and without motion between the acquisition of the VFA volumes. $R_1$ maps were constructed with the proposed corrections, and (at 3T) with the correction scheme. Results: At 3T, the proposed methods outperform the baseline motion artefacts were also reduced at 7T. However, converged on that of the no motion condition if transmit field effects were also incorporated. Conclusion: The proposed methods simplify inter-scan motion correction of $R_1$ maps and are applicable at both 3T and 7T, where a body coil is typically not available. The open-source code for all methods is made publicly available.

andm(x,t) gives the density of players whose remaining stock is xat timet. The parameter e measures the degree to which firms compete. If e= 0, every firm is a monopolist, while if eis very large, all firms are nearly interchangeable in the eyes of con and uniqueness of solutions to System (1.1) has be en established under the assumptions we make in this article, thanks to the results found in [16, 19 , 17], cf. [18, 15]. In [8], Chan and Sircar proposed a numerical method to solving (1.1) that con sists of assuming the following be derived by differentiating System (1.1) ktimes with respect to e, lettinge= 0, and then solving. Notice that setting e= 0 decouples the system. Thus u(k)andm(k) are computed by solving two equations separately, avoiding the computational difficulties arising from the coupling. To establish rigorously the accuracy of this numerical meth od, one must answer the question: is the solution to System (1.1) differentiable with respect to e? In the present article, we provide an affirmative answer to this question under generic assumpti ons on the data. In particular, we establish that there exists e0>0 small enough such that for every kN, the solution to System (1.1) isktimes differentiable with respect to efor 0<=e<=e0. See Theorem 1.1 below. More precisely, we establish that the difference between the solut ion and its kth order Taylor expansion is bounded by Ckek+1for some constant Ck. On the other hand, the constants Ckthat appear in our analysis blow up very fast as k- , and for this reason we conjecture that the infinite series expansion does not in fact converge (see Remark 4.1). While the numerical method of Chan and Sircar gives a practic al application, our work is also motivated by the theory of forward-backward linear systems of equations, which turn out to be of fundamental importance in mean field game theory. In their seminal work [3], Cardaliaguet et al. show that one can established well-posedness of the master equation by the following steps: (1) solve the mean field game system, (2) linearize the system by f ormally differentiating with respect to the measure variable m, and, most crucially, (3) proving estimates on solutions to which have as a corollary that the solution from step 1 is diffe rentiable with respect to the initial measure. We do not study the master equation in th is article. However, our follows exactly the same steps, where instead of li nearizing with respect to an measure variable, we differentiate with respect toe. At an abstract level, the system of linear equations we study has the = 0, w(x,T) = 0, 0<=x< (iv)w(0,t) =u(0,t) = 0, Ph are given functions, and (1.5) known function such that b(e)-0 ase-0. Ourmain mathematical contribution in this article is a set of a priori estimates on solutions to System ( 1.4). (See Section 2.) The estimates share three types in common with [3, Section 3.3]: (1) energ y estimates, derived by using the duality between equations (i) and (ii); (2) sta ndard Schauder estimates on parabolic equations; and (3) estimates on uin the space dual to a certain H older space. Since the coupling depends on integral terms involving the unknowns, the energy estimates are technical than for standard mean field games. In particu lar, they require a fourth type of estimate, namely (4) estimates see Section 2.1. Additionally, because the boundary conditions in our model are of Dirichlet type, t he estimates on uin the dual of a H older space require additional care compared to similar e stimates on a torus; see Section 2.3. All of our results hold in one space dimension, on which the probl em was originally posed by Chan and Sircar (cf. [22]). In higher space dimensions, the Dirichle t boundary conditions pose an for the purposes of regularity, since there would b e a corner in the domain, although the interior estimates are expected to hold as in the one-dimens ional case. This is a technical aspect we do not address in the present work. Mean field games were introduced circa 2006 by Lasry and Lions [27] and by Caines, Huang, and Malham' e [23]. Since then, both theory and applications hav e been well-studied. For a we refer the reader to the texts [2, 5, 6]. Econom ic models are a common application of mean field game theory. See e.g. the overviews in [1, 28, 13] , and for the particular example of exhaustible resource production see [8, 9, 22]. When the e quilibrium strategy depends on the distribution of controls, as is often the case for economics applications, we give the name mean field game of the controls [4] (orextended mean field games [12]) to the resulting mathematical model. The theory of partial differential equations for mean field gam es of controls has been developed in [4, 25, 24, 14, 12, 20]. The present work is, to the best of ou r knowledge, the first study of parameter sensitivity for a mean field game of controls. Some of the estimates presented here will be useful in a forthcoming study of the master equation for a m ean field game of controls with absorbing boundary conditions [21]. The remainder of this article is organized as follows. In the rest of this introduction, we define some notation and present the main result. Section 2 is the co re of the paper and provides a priori estimates on systems of equations with the abstract form (1. 4). In Section 3 we prove existence and uniqueness of solutions to (1.4). Finally, in Section 4, we prove the main Throughout this manuscript, Lp(D) (1<=p<= ) will denote the usual Lebesgue space on a domain Dwith standard norm, denoted either by simply We will often consider the space-time domain D= (0,)x[0,T], on which we consider the space such that the map inLp(0,T). The norm is given by (1.6) the space the space of H older continuous functions usuch that the following norm is finite: (1.7) an integer, the space of ktimes differentiable functions Ca. It is a Banach space with the norm (1.8) by standard interpolation results is equivalent to th e the space the space of all H older that the following norm is finite: (1.9) t,s[0,T], integers, we can also define In particular, the space the standard H older space for parabolic equations (cf. [26]); its norm can be written (1.10) any function f(x,t;e), we denote the k-th partial derivative of f(x,t;e) with respect to eby (1.11) the convention that f(0)=f. In studying System (1.4), we will frequently suppress notat ion and write G(ux,m;e),G(e), or even Gto provided that no ambiguity arises. Analogous statements hold for that depend on multiple of the main result. Throughout this paper, we assume that conditions hold on the data so that System (1.1) has a unique solution ( u,m) satisfying u,m Sufficient conditions are provided in [17], cf. [16]. We wi ll denote this solution by (u(0),m(0)). In addition to smoothness, we will also need to assume the i nitial density m0has finite first if we differentiate (1.1) ktimes with respect to e, we obtain the = 0, u(k)(x,T) = 0, =m(k)(0,t) = 0, clarity, we note that the full expression for F(k)is given that for a(k)(e) reader may wonder why, in writing F(k),Jk, andKk, we have suppressed the arguments u(j) andm(j)forj < k. This is because we will be arguing inductively as follows: t o prove that we may assume that u(j)andm(j)are known functions for j <k. We now state our main result. Theorem 1.1. There exists e0>0small enough (see Equation (2.5)) such that for each System (1.12)has a unique classical solution (u(k),m(k)), which satisfies the identity (1.18) ( u(k),m(k)) the solution to (1.1). That is, the formal differentiation carried out on uandmto obtain System (1.12)is justified. Moreover, there exists a constant Ck>0such that for sufficiently small. 2.A priori estimates In this section we present our main mathematical contributi on by proving new a priori estimates for a coupled forward-backward system of linear partial differen tial equations (1.4). It turns out that a useful first step for such systems is to prove energy estima tes, which are derived by using the duality between equations (i) and (ii). Howeve r, because of the in (1.4), the energy estimates are not useful with out certain a priori estimates on the whereuis the solution to the Fokker-Planck type We derive these estimates first in Section 2.1, wh ich may have independent interest to the reader interested in Fokker-Planck equations with a source . Section 2.2 then provides the desired energy estimates, which we apply throughout the rest of the s ection. In section 2.3 we arrive at 5further estimates for the Fokker-Planck equation in the dua l to a H older space. Finally, in Section 2.4 we give a priori estimates that establish full parabolic priori estimates on Fokker-Planck equations with a source. In this subsection, we collect some estimates on the Fokker-Planck equation (1.4) (ii), in particular with respect to the L1 norm and firstmoment in the space variable. These results hav e very little to do with the of the coupled system (1.4) and can be stated abstr actly for a Fokker-Planck equation with a source. The proofs of these results are given in Append ix A. Proposition 2.1. =u0(x), = 0, known assume the constant C(0,)depends only on andT; (b) the constant C(0,)depends only on andT. Proof.See Appendix A.1. /square The following proposition could be stated for an abstract Fo kker-Planck equation like (2.1); never- theless, it is given in the form below to make it more obvious h ow it may be applied later on. Proposition 2.2. Define the constant (2.2) C0:= as in(A.5), and suppose usatisfies (ii)of System the constants C1,C2(0,)depend only particular, if T <, then there is no restriction on l, and Appendix A.1. estimates. Byenergy estimates we mean specifically an estimate on the Before stating our result, we first define an upper bound on th e pa- rameterethat we will need in the proof. With bas in (1.16) and C0as in (2.2), let e0>0 be such that (2.5) becauselim e-0+b(e) = 0. a consequence of the method of proof for the energy estimat e. Specifically, the upper bounds are motivated by (2.16), (2.17), and the value of C2, the constant appearing in Proposition 2.2--see the proof of Proposition 2.2 for an explicit 2.3. Suppose (w,u)satisfies Then for e[0,e0]the following energy estimate is a constant that depends only on r; andC2(0,)is a constant that depends only andT. As an immediate begin by first equality follows via differentiation under the integ ral sign, the second equality follows from substituting in the equations for wt-rwandutgiven by System (1.4), and the last via integration by parts. Next, unpacking G, (2.8) then have the following three estimates. the first inequality follows by two applications of You ng's inequality, and the last inequality via the Cauchy-Schwarz inequality. Second, let d>0 to be chosen later. last inequality above follows from a generalized Young' s inequality. Third and first inequality follows via the Cauchy-Schwarz inequal ity and Young's inequality. The last inequality follows via another application of the Cauchy-S chwarz inequality. Together with (2.9), 8the above three estimates make future calculations less cluttered define (2.14) f(t):= because w(x,T) =u(x,0) = 0, integrating the left-hand side of (2.13) in time and re arranging the inequality, we thatb(ex), and subsequently f(t), converges to 0 uniformly in taseanddgo to 0. Let d1>0 to be determined shortly. We can take eanddsmall enough such that (2.16) enough. Consequently, by Proposition some constants C1,C2(0,) which depend only andT. We now If necessary, we may take esmaller so that b(ex)<1/10 as well. This, together with (2.16), demonstrates that by choosing e0as in Equation (2.5), we find for all e[0,e0] that (2.18) 1 clean things up, the restriction on d1impliesd and we can restrict dfrom below so that As such, we find This yields the desired result. /square 9As a consequence of the above energy estimate, we find that the maximum of wis by the square root of the maximum of wx. Corollary 2.4. Suppose (w,u)satisfies System Then for e[0,e0], (2.21) |w(x,t)| a constant depending only andT;C2(0,)is a only on r, andT; andk[0,)is a constant depending solely such a way that k= 0. Proof.Begin by defining (2.22) (1.4) hence (2.23) implies (2.25) the standard maximum principle on v, we find that (2.26) similar argument, where we define (2.27) shows that (2.28) e-rtw(x,t)>= all several careful applications of the Cauchy-Schwarz i nequality we that the fourth inequality follows from Proposition 2. 2, so thatC1,C2(0,) are constants that depend only andT. Also, the constant Ccan be given by (2.31) combining this estimate with (2.29) we obtain (2.32) |w(x,t)| all (x,t)[0,)x[0,T]. Now, using the energy estimate, established in Propositi on 2.3, there exist constants C3,C4(0,), depending only andTsuch that all (x,t)[0,)x[0,T], which is the desired result. in the dual of C1+a.In this subsection we want to provide a priori estimates on u, the solution to (ii) of System (1.4), in the space (2.34) Ca/4([0,T]; (C1+a)*) for a given a(0,1). Such estimates are required to deduce the time-regulari ty of integral We first use duality methods to obtain estimates on uin the space (2.35) Ca/4([0,T]; the space of all ph C1+asuch thatph(0) = 0. These estimates rely on for the primal problem, which are found in Appendi x A.2. Proposition 2.5. Then for e[0,e0]there exists a constant C(0,), depending exclusively l,a, andT, such the constant kdepends solely such a way that k= 0. Proof.Fix somet1[0,T]. Consider the dual PDE to 0, =ph(x) C1+a ([0,)), x [0,) ps(0,t) = 0, t [0,t1], for somel>0. By Lemma A.2, we have (2.38) some constant C0depending only on l,a, so that the above equation implies (2.39) 0. We observe first equality follows via differentiation under the integ ral sign, the second equality follows from substituting in the equations for and the last equality follows via integration by parts. Consequently, with t2[0,t1] arbitrary, we we have used (2.38), and where (2.42) assumption and the fact that e-lt Ca/2([0,T]), we the constant C1depends on landa. Turning our attention now to e2, we first use the Cauchy-Schwarz inequality to be estimated in the following first inequality follows from a combination of using the t riangle inequality, unpacking the definition of G, and recalling The second inequality uses the fact The third inequality is a consequence of the Cauchy-Schwar z this initial estimate on e2, we find the constants C2,C3(0,) depend only l, andT, and we define (2.46) second inequality in (2.45) is a consequence of Proposit ion 2.2, followed by an application of Proposition 2.3. Returning to (2.44), it follows in (2.43). Hence (2.41) and the two estimates in (2.43) and (2.47) show call (2.50) sett2= 0, so that u(x,t2) = 0. Hence (2.48) particular, since t1andphwere arbitrary, we lett1,t2[0,T] be arbitrary and without loss of generality take t1>=t2. the last integral, we the last inequality follows from (2.52) and (2.38). Co mbining this with what we found in (2.48), the equation in (2.53) ast1,t2, andphwere is the desired result. /square Notice that any ph C1+acan be written as the sum of a constant and an element of C1+a , since ph-ph(0) C1+a . Therefore the following proposition will complete our est imates ofuin the 2.6. Define (2.57) e(t):=/integraldisplay 0u(x,t)dx. Then for there exists a constant k[0,), and a constantC(0,), depending solely l,aandTsuch that for all e[0,e0] (2.58) the constant kdepends exclusively such a way that k= 0. Proof.We begin by applying (A.30) with the identifications u:=u,b:=F(0), and n:= get (2.59) u(x,t) first step will be to prove a bound on the (A.35) and H older's inequality, we apply (2.62) and (A.4) to (2.59) and get (2.63) as in (A.5). Multiply both sides of (2.63) by e-ltfor somel >0 to be chosen. Define fl(t) =e-ltf(t) andgl(t) =e-ltg(t). We see (2.65) wish to the other hand, using Proposition 2.3, H older's inequal ity, and recalling constants depending only on s,l, andT, andk'is as in (2.46). We combine these estimates to the supremum, we get (2.70) setting l= (2C1p1/2)2, we yields (2.72) some constants A,B(0,), depending only l, andT, and (2.73) now turn to H older estimates. By integrating (2.59) with respect tox, we discover that (2.74) e(t) want H older estimates on each term. Let 0 <=t0<=t1<=T. Fore1, = for allx, it follows 2(2s2p)-1/2. We (2.72) and computing the remaining integrals, we obta some constant C(0,). We proceed similarly for e2. = (A.35) and (2.72), we again constants depending only l, andT. Using the same reasoning as above, we H older's inequality and Proposition 2.3, we a similar computation, we (2.80) and (2.86), we can find some constant /tildewideC(0,), depending only andT, such yields (2.58), as desired. /square With Propositions 2.5 and 2.6 established, we now have the co mplete duality estimate on u. Corollary 2.7. Then for e[0,e0], one some constant C(0,), depending exclusively l,aandT; solely such a way that k= 0. 2.4.FullC2+a,1+a/2regularity. We are now in a position to establish full parabolic regulari ty in classical spaces for solutions to System (1.4). Theorem 2.8. Leta(0,1/6]be such that (u(0),m(0)) C2+2a,1+a. Ca,a/2andPh C1+a,a/2. Then for e[0,e0], one has (2.90) w there exist constants C(0,)andk[0,), depending solely on the data such that (2.91) exclusively on such a way that k= 0whenever Ph = Ps = 0 . In addition, we have u an estimate (2.92) by defining (2.93) that after expanding terms (i) of System (1.4) reads (2.94) = 0. Note thatb(ex) Ca/2([0,T]), and by assumption, Ps As such, using classical estimates [26, Theorem IV.5.1], we the constant C0(0,) depends on the data. Focus one2first. We by Corollary C1(0,), dependingsolely and k'[0,), such away that forces k'= 0. On the other hand, since m(0)satisfies a (homogeneous) Fokker-Planck equation, by classical arguments (cf. [21, Lemma 3.1]) combined with the estimates of Proposition 2.6 (cf. [21, Lemma 3.7]), we have an estimate on the Now a similar calculation as in (2.96) the constant C2(0,) depends solely Therefore, returning to (2.95), we By interpolation of H older spaces, there exists a constant C'>0 such as in the last line of (2.99). As such, (2.100) along with (2 .99) show second inequality follows from Corollary 2.4, so that th e constants C3,C4(0,) depend only on the data, and k''[0,) depends solely such a way that k''= 0. The constants C''(0,) andk2[0,) can be given by (2.102) C'':= by Young's (C'')2+ Hencewhas the desired regularity with in terms of the data as stated in (2.91). Turning our attention to u, we begin by observing that (ii) of System (1.4) implies (2.105) already have estimates on the coefficients in Ca,a/2. As for the right-hand side, we someC1(0,), depending solely on the data, and k1[0,) depending on such a way that k1= 0 whenever Ph = Ps = 0. Finally, since Ph C1+a,a/2, we obtain via classical estimates [26, Theorem some constant C0(0,) depending only on the data, some constant C2(0,) depending only on the data andk2[0,) depending on a way that k2= 0 whenever Ph = Ps = 0. This demonstrates the desired the required bound given in (2.91). Finally, we appeal to Proposition 2.1 and the estimates just obtained to find that, for all K Ph+(G-wx)m(0), and where Cis a constant on the data. take the supremum over get an estimate on the H older Combine this with the estimates above to obtain (2.92). and uniqueness for the abstract system In this section, we use a fixed point argument to demonstrate e xistence of solutions to (1.4). we provide a statement of the Leray-Schauder fixed point theorem, taken from [11, Theorem 11.6, p. 286]. Theorem 3.1 (The Leray Schauder Fixed Point Theorem) .LetXbe a Banach space, and let T:X x[0,1]- Xbe a compact mapping satisfying T(x,0) = 0for allx X. Suppose there exists a that (3.1) all(x,l) X x[0,1]satisfying T(x,l) =x. Then the mapping T1:X - X, given by T1(x) =T(x,1), has a fixed point. 21We now aim to construct a suitable Banach space along with a co mpact operator to satisfy the hypotheses of Theorem 3.1. Fixa(0,1/6] such that u(0),m(0) C2+2a,1+a. Define the Banach space norm (3.2) Ps Ca,a/2and Ph be given functions with xPhL t(L1 x) and Ps(0,T) = 0. For anym Xandl[0,1], letu=T(m,l) be determined by solving the = 0, w(x,T) = 0, 0<=x< (iv)w(0,t) =u(0,t) = 0, 0<=t<=T, where we recall that Gis defined in (1.5). The subsequent lemma establishes properties of Tthat are necessary for applying Theorem 3.1. Lemma 3.2. The operator T, defined in (3.3), is well-defined, continuous, and break the proof into three parts. Throughout the proof, we will refer to the results found in [26], especially Theorems IV.5.1-3, as classical resul To show that the operator Tis well-defined, suppose m Xandl[0,1]. By Lemma B.1 there exists a unique solution w C2+a,1+a/2to Equation (3.3)(i) (with given in (iii)-(iv)). In particular, G(wx,m;e) (which does not depend on x) is esti- mated in Ca,a/2, and hence via classical results, Equation (3.3)(ii) has a u nique solution u, which is estimated in C2+a,1+a/2. In addition, as a consequence of Proposition 2.1 (a), with (3.4) find that uL t(L1 x). Using Equation (2.109), we can deduce that u well. This ultimately shows that u X, and hence Tis To demonstrate continuity of the operator T, start by letting {(mj,lj)} j=1 X x[0,1] be a sequence such that . Let (wj,uj) be the solution to system (3.3) with mreplaced by mjandlreplaced by lj, so classical estimates (see also Lemma B.1), the sequences in C2+a,1+a/2, and therefore both wj,ujand their derivatives are uniformly bounded By the Arzel` a-Ascoli Theorem and diagona lization, there exists some ( a subsequence that (3.5) (wjk,ujk)-(w,u) ,R>0. In particular, wjk,ujkand their derivatives converge pointwise and are uniformly bounded. Also, because we also have mj-minL t(L1 x), we see that (3.6) letk- in the system solved by ( wjk,ujk) to see that ( w,u) solves System (3.3), i.e. u= T(m,l). Finally, by the argument that appears in the next step, we d educe that T(m,l) inX. It follows that Tis Finally, to demonstrate compactness of the operator T, bounded sequence in X x[0,1]. If again we let ( wj,uj) be the solution to system (3.3) with m replaced by mjandlreplaced by lj, then as in the previous step we see that the sequences uniformly bounded in C2+a,1+a/2, and there exists a subsequence that (3.5) holds. It remains to show that will appeal to Proposition 2.1. First, note that ujk-u(k= 1,2,...) satisfies (2.1) of Proposi- tion 2.1 with (3.7) bk:= have that bkis uniformly bounded in L; we need to show that nkandxnkare in L t(L1 x). By assumption, Ph L t(L1 x) andxPhL t(L1 x). Next, by (3.5), there exists a constantC1(0,) such allkN. Also, by assumption, {mjk} k=1is uniformly bounded in L t(L1 x), so there exists some constant C2(0,) such allkN. It follows allkN. Note that to obtain the inequality we also used We now find that (3.12) 1 for some constant M1(0,), independent of k. Also, Proposition 2.1 (b)ensures that xm(0) L t(L1 x), and thus (3.13) 2 for some constant M2(0,), independent of k. Therefore, the PDE that (2.1), also satisfies the hypotheses of Proposition 2 .1 uniformly for all kN. 23We then (3.14) = (3.5), for each fixed K >0, (3.15) ask- . By Proposition 2.1 asK- for allkN, and we see that Similarly, let - ujk=ujk-u. For have, cf. this we deduce that hence also in by [7, Lemma 2.2], we have (3.19) only on Combining this estimate with (3.5), we deduce Since {ujk}is bounded in C2+a,1+a/2, we deduce that we have that (3.20) ujk-uinXask- , as desired. We have thus shown that the operator Tis continuous and compact. This completes the proof. /square With Proposition 3.2 established, we may now appeal to the Le ray-Schauder fixed point theorem to obtain an existence result for System (1.4). Theorem 3.3. There exists a unique classical solution (w,u)to System endeavor to use the Leray-Schauder fixed point theorem, i. e., Theorem 3.1. For that purpose, a quick inspection reveals that when l= 0, we have T(u,0) = 0 for all m X. Further- more, Lemma 3.2 shows that Tsatisfies the compactness requirement of the Leray-Schaude r fixed point theorem. Now, suppose that there exists uXsuch thatu=T(u,l), so that (3.3) has a 24solution (w,u) withu=m. Then via the a priori estimates established in Section 2, sp ecifically the regularity results of Theorem 2.8 and Proposition 2.1 (a), we find that indeed, (3.21) some constant C(0,). We conclude that there exists a fixed point of the operator T, corresponding to l= 1. We deduce that there exists a solution to (1.4), as wanted . To show uniqueness, let ( w1,u1) and(w2,u2) besolutions to (1.4). Consider thedifference( Then because System (1.4) is linear, we find that ( w,u) satisfies System (1.4) with Ps and Ph identically 0. As such, Theorem 2.8 implies (3.22) subsequently w=u= 0. Therefore, w1=w2andu1=u2, hence the solution to System (1.4) is unique. This completes the proof. /square 4.Proof of Theorem 1.1 We now present the proof of Theorem 1.1. Proof of Theorem 1.1. Weinducton kandbreak theproofinto . Step 1.LetnNand suppose the following induction hypothesis. For allk there exists a solution ( u(k),m(k)) to (1.12), as well as constants Ck,C* k(0,), depending solely on the data and on ( u(j),m(j)) forj <k, such that a consequence of the inductive hypothesis, for all kNwith 0<=k<n, we can we define the constant (4.3) order to apply our abstract results from Section 3, we intr oduce the following notation. We will separate out the n-th order terms appearing in JnandKnby as (4.4) the exact formulas of be easily gleaned from (1.13), (1.14), and (1.15), it is more important to know that entirely of lower order terms, and hence in light of the induction hypothesis (see (4.1) i n Section 4). 25Using (4.2) and recalling the formulas for JnandKnin (1.13) and (1.14), we may using the product rule, we find thebase case, k= 0, is established in [17] (cf. [16]). (Note be established, even on an unbounded domain, by the same argume nt as in the proof of Theorem 2.8, cf. Equation (2.109).) Assume then that the induction hypot hesis holds for k=n-1. Observe that System (1.12) can be written as System (1.4) with w:=u(n),u:=m(n), and (4.6), (4.7), and (4.9) show that the induction hy pothesis guarantees Ps Ca,a/2and Ph . Therefore, by Theorem 3.3, there exists a unique solution to System (1.12), which by Theorem 2.8 must satisfy (4.1) for k=n. Remark 4.1. The constants Cnderived from this inductive argument are likely to blow up ve ry quickly. Assuming they are greater than 1, they will be an inc reasing sequence, and an inspection of the argument above reveals that (4.10) in turn implies that the ratio Cn/Cn-1converges to +, which not converge on any radius. Step 2.Next, letnNand suppose the following induction the induction hypothesis of Step 1holds. Aditionally, for all k {0,1,...,n-1}, suppose there exist constants solely on lower order terms and the data such the casek=n, it can beshown that System (1.4) with Ps := Ph:= quick inspection reveals that the induction hypothesis im such, the estimates in Theorem 2.8 show that as well. Step 3.LetnNand suppose the following induction the induction hypothesis of Step 2holds. Additionally, suppose there exist constants depending solely on lower order terms and the data such 0,1,...,ndefine (4.18)-hu(k+1)(e) well-defined objects via Step 1. As it happens, when k=nthe pair (w(n),u(n)) satisfies System (1.4) with Ps := Ph:= such, we find that the induction hypothesis together with Step 2imply that Using the estimates in Theorem 2.8, we ultimately see that This yields the desired result concerning differentiabili ty with respect to e. /square Appendix A.Some technical proofs of results from Section 2 A.1.Proofs for Section 2.1. Proof of Proposition 2.1. LetS(x,t) denote the heat kernel, given by (A.1) S(x,t):= Duhamel's principle and integration by parts, we have u=u1+u2, where u1(x,t) u2(x,t) c':= letM >1 to be determined later. the first inequality follows directly from (A.2); the s econd inequality follows from (A.4) and noting that for ally(0,). Define (A.7) B:= (A.6) second inequality follows from the substitutions and es timate ifM >1, then 2e-Mt t<=1 for allt>0. Thus, (A.8) implies (A.10) the other hand, since S(*,t) is a probably density, we (A.11) to (A.10), take a supremum on both sides over all t[0,T], and we find that (A.12) >1 large enough so that M Consequently, we find that (A.13) given the definition of Bin (A.7), establishes part (a). For part (b)we argue similarly. First, note that (A.14) 1. 30Knowing this and using (A.4), we can as in (A.5). Note that the second equality above follows fr om (A.14), and the last inequality from (A.4). We now use Fubini's Theorem to estima second inequality follows directly from (A.15). Let M >1 to be determined later and define (A.17) B':= then estimate in a similar manner as in the proof of part similar fashion, we we have also used that S(*,t) is an even function. We calculate (by a change of Add together (A.18) and (A.19) to a constant that depends only on its existence is guaranteed by part (a)of this proposition. Now, taking a supremum on both sides of ( A.20) over all t[0,T], then choosingM >1 large enough so that M we derive (A.21) the definition of B'in (A.17), the above estimate implies part (b)of the proposition. /square Before proving Proposition 2.2, we present an abstract lemm a. Lemma A.1. LetA>=1,B,d>0be given constants. Suppose f,g:[0,)-[0,)are functions that satisfy (A.22) for any l>1 dln(A), we =f(t)+g(t), so that (A.22) reads simply (A.24) all arbitrary t>0 Use (A.24) n+1 times to get (A.25) max{s,0}. Note that t-(j+1)d we define ph(s) Then (A.25) and Multiply (A.26) by e-lt, then integrate from 0 to T to now observe (A.28) to (A.27), we implies (A.23). /square We now apply Proposition A.1 to the Fokker-Planck equation. Proof of Proposition 2.2. First we will treat uas a solution of the abstract Fokker-Planck equation (2.1) by identifying b:=F(0)andn:= We start with the following formula: for the same calculations as in (A.6), we defined in (A.5). we can write (A.32) H older's inequality, we we square both sides and use the inequality ( a+b)2<=2(a2+b2) to get (A.34) now first inequality follows from the fact The second from the Cauchy-Schwarz inequality and the fact tha Now, define (A.37) see that (A.36) and (A.34) yield (A.38) factoring + and + From (A.38), we see that Proposition A.1 applies with A:= 2, andB:= choose d>0 small enough such that (A.39) 2 d1/2B choose l>1 dln(2) large enough such that (A.40) 1 a by Proposition A.1 we second inequality is obtained by the simple (2.3) holds. Now, in the case that Tis finite, we can estimate for all l(0,C0], (A.44) using (2.3) with l= 2C0, we the last inequality follows, because e-2C0t<=e-ltfor allt[0,T]. This completes the proof. /square A.2.An abstract H older estimate. In this subsection we will give an abstract result on H older regularity for a parabolic equation with Dirichlet boundar y conditions and bounded argument is in the spirit of [3, Lemma 3.2.2.], but we cove r the case of an unbounded domain. The result stated below is also meant to allow for a possibly i nfinite time horizon, though in the present work we will not exploit this. Lemma A.2. Fix0<a<1. Letube a solution u(0,t) = 0, u(x,0) any positive constant, VandFare a bounded continuous functions, and u0 C1+a (D) (i.e.u0 C1+a(D)andu0(0) = 0). Then (A.47) independent of T. As a corollary, we have (A.48) will start by assuming u0= 0. By a standard application of the maximum principle we have that Letph(x) be a smooth function, and observe g(x,t) a time t >0, and setv(x,t) Then (A.49) ~g, where (A.52) ~ g(x,t) anya>0 and letph(x) Note that phsatisfies the following It follows that (A.54) remains bounded as p- large. By the potential estimates in [26, Secti on IV.3] we have an estimate of the the constants do not depend on t. On the other hand we have (A.56) interpolation (see e.g. [26, Lemma II.3.3]) we some constant C, whered >0 is sufficiently small. Choosing d>0 small enough, we by a Sobolev type embedding theorem [26, Lemma II.3.3] w e have (A.59) We can rewrite this locally Lipschitz with constant depending on l, we can write this the fact thatph' phis bounded by 1 and is globally Lipschitz with 1, we also since1 phis globally Lipschitz with Lip(1 /ph) = 1, and bounded on [ a-1,a+1] with an upper bound of 2, we deduce estimate is independent of tanda. Lettingtandavary through all the positive integers, and sincepcan be determined through a, it follows that (A.65) now remove the assumption that u0= 0. Letwbe the solution 0, w(x,0) =u0(x). Asl >0, by the maximum principle, we have Then [26, Theorem 10.1]; this estimate does not depend on time because of the global in time Lbound. 37This establishes We then take the derivative in xof Equation (A.66) and apply the same argument as above tow xto let ^ube the solution zero initial conditions. Then by (A.65) we ^u+wis the solution to (A.46), the claim is proved. Finally, to prove (A.48), note that (A.47) immediately impl ies (A.69) we also (A.69) and (A.70) to get (A.48). /square Appendix B.Nonlocal Existence Lemma For this lemma, recall the definition norm (B.1) B.1. Leta(0,1/2)be such that u(0),m(0) C2+2a,1+a. Letm X,uT given. Consider the backward parabolic =uT(x), = 0, defined as in (1.5). We assume the usual compatibility conditions of first order : (B.3) uT(0) = = 0. Then there exists a unique classical solution u the boundary value problem (B.2). Moreover, the following estimate holds: (B.4) only on the be given. Note on the C2+2a,1+anorms ofu(0)andm(0). By [26, Theorem IV.5.2], there exists a unique solution the boundary value =uT(x), = 0. 0<=t<=T, and it satisfies the estimate (B.7) will denote w=F(u). Our goal it to show that Fis a contraction on a suitably defined metric space, and use this to prove (B.2) has a then define Note = 0, = 0. the definition of Gin (1.5), we see that there is a constant C1, depending only on the data, such this with classical estimates from [26, Theorem I V.5.2], we see that there is some constantC2, depending only on the data, such that (B.10) interpolation on H older spaces, we deduce that there exists a constant C3, depending only onC2, such that (B.11) we define ~ w(x,t) which satisfies (B.13) the maximum principle to ~ wto deduce now apply all of these estimates only on the time interval [ T-t,T] for somet >0. We now set t=1 8C3erTC1. Define Ytto be the space endowed with the norm (B.16) /bardblw/bardblYt:=/bardblw/bardbl+2C3/bardblw/bardblL([0,)x[T-t,T]). Observethat Ytis aBanach space. Moreover, by theabove estimates, F:Yt- Ytisa HenceFhas a unique fixed point u, which is a solution to (B.2) and satisfies the estimate (B.4) on the time interval [T-t,T]. However, Tis arbitrary. We can now partition the interval [0 ,T] into subintervals that are each at most tin length, i.e. 0 = t0< t1<***< Apply the same argument on each replacing the final condition uT(x) withw(x,tj), for each jstarting with Nand going down to 1. (Cf. the proof of [10, Proposition 3.11].) In this w ay we obtain a solution uto Equation (B.2), which indeed satisfies (B.4). Uniqueness of this solu tion follows from uniqueness on Achdou, F. J. Buera, J.-M. Lasry, P.-L. Lions, and B. Moll ,Partial differential equation models in macroeconomics , Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 372 (2014). [2]A. Bensoussan, J. Frehse, and P. Yam ,Mean field games and mean field type control theory , Springer, 2013. [3]P. Cardaliaguet, F. Delarue, J. Lasry, and P. Lions ,The Master Equation and the Convergence Problem in Mean Field Games: (AMS-201) , Annals of Mathematics Studies, Princeton University Pres s, 2019. [4]P. Cardaliaguet and C.-A. Lehalle ,Mean field game of controls and an application to trade crowdi ng, Mathematics and Financial Economics, 12 (2018), pp. 335-36 3. [5]R. Carmona and F. Delarue ,Probabilistic Theory of Mean Field Games: vol. I, Mean Field FBSDEs, Control, and Games , Stochastic Analysis and Applications, Springer Verlag, ( 2017). [6] ,Probabilistic Theory of Mean Field Games: vol. II, Mean Fiel d Games with Common Noise and Master Equations , Stochastic Analysis and Applications. Springer Verlag, ( 2017). [7]A. Cesaroni and M. Cirant ,Concentration of ground states in stationary mean-field gam es systems , Analysis & PDE, 12 (2018), pp. 737-787. [8]P. Chan and R. Sircar ,Bertrand and Cournot mean field games , Applied Mathematics & Optimization, 71 (2015), pp. 533-569. [9] ,Fracking, renewables, and mean field games , SIAM Review, 59 (2017), pp. 588-615. [10]M. Cirant and A. Goffi ,On the existence and uniqueness of solutions to time-depend ent fractional MFG , SIAM Journal on Mathematical Analysis, 51 (2019), pp. 913-9 54. [11]D. Gilbarg and N. S. Trudinger ,Elliptic partial differential equations of second order , springer, 2015. [12]D. A. Gomes, S. Patrizi, and V. Voskanyan ,On the existence of classical solutions for stationary exte nded mean field games , Nonlinear Analysis: Theory, Methods & Applications, 99 (2 014), pp. 49-79. [13]D. A. Gomes, E. Pimentel, and L. Nurbekyan ,Economic models and mean-field games theory , 30CBM, IMPA, 2015. [14]D. A. Gomes and V. K. Voskanyan ,Extended deterministic mean-field games , SIAM Journal on Control 54 (2016), pp. J. Graber ,Linear quadratic mean field type control and mean field games w ith common noise, with application to production of an exhaustible resource , Applied Mathematics & Optimization, 74 (2016), pp. 459-48 6. 40

2108.10962

Philip Jameson Graber

P. Jameson Graber and Marcus Laurel

Parameter sensitivity analysis for mean field games of production

math.AP math.OC

http://creativecommons.org/licenses/by/4.0/

We study a mean field game system introduced by Chan and Sircar (AMO, 2015) to model production of an exhaustible resource. In particular, we study the sensitivity of the solution with respect to a parameter $arepsilon$, which measures the degree to which producers are interchangeable. We prove that on some interval where $arepsilon_0 > 0$, the solution is infinitely differentiable with respect to $arepsilon$. The result is based on a set of new a priori estimates for forward-backward systems of linear equations.

(a)(b)Systematic Bias(c)Map to Quantum AlgorithmMap Program to Physical Qubits Compile to Quantum Gate Circuit)Execute Quantum TrialsCast to HamiltonianEmbed to Hardware TopolgyGenerate Quantum Machine Instruction (QMI)Execute Single TrialsFigure 1: Steps involved in solving a problem using (a) gate-model quantum computers and (b) quantum annealers. (c) Energy histogram of a 2000-qubit optimization benchmark executed on D-Wave QA (in logscale). The QA can quickly identify the region of the ground state energy (overlapping region) but the solution is far from the global optima due to systematic bias. range of real-world but not limited to, planning [80], scheduling [90,91], constraint satisfaction prob- lems [16], Boolean satisfiability (SAT) [13, 84], matrix factor- ization [68], cryptography [39, 74], fault detection and system diagnosis [75], compressive sensing [12, 14], control of auto- mated vehicles [43], finance [27], material design [49], and protein folding [59]. Although promising, QA hardware suf- fers from various drawbacks such as noise, device errors, lim- ited programmability and low annealing time, which degrade their reliability [4,9,56]. Addressing these limitations enhancements that may span generations of QAs. Therefore, software techniques to improve the reliability of QAs is an important area of research recent hardware and software enhancements, ex- isting QAs may fail to find the global minima for certain problems [9]. For example, Figure 1(c) shows the energy his- togram of a 2000-qubit optimization problem on a D-Wave QA. We can think of QA as a machine that samples from a Boltzmann distribution such that samples with lower energy values are exponentially more likely to be observed [9, 93]. In theory, QA can find the optimal solution with a very high probability [65]. However, in this example, we observe that although the QA can quickly identify the region of the global optima, the best solution from the QA is far from the global optima. As users run only a single QMI, the program is sub- jected to a similar noise profile for all trials, resulting in a systematic bias . Our experiments show that running more trials or relying on existing error-mitigation schemes cannot overcome this bias. Unfortunately, systematic bias solutions far from the global optima and limits the reliability of QAs for practical applications. In this paper, we propose Ensemble Quantum , an effective scheme for mitigating systematic bias and improving the reliability of QAs by running an ensemble of QMIs with controlled perturbations. EQUAL is based on the insight that running the same QMI for all trials projects QAs to a very similar noise profile and bias. Instead, EQUAL uses an ensemble of QMIs that subjects the system to different noise profiles and biases. Generating effective ensembles ofQMIs is non-trivial and our design focuses on addressing it.1 To generate ensembles, EQUAL creates new Perturbation Hamiltonians , and adds them to the orig- inal problem Hamiltonian. Every perturbation Hamiltonian adds noise to the original Hamiltonian and the QMI obtained from this process is a perturbed variation of the original QMI. The challenge in this step is that adding extremely small per- turbations will have no impact on the systematic bias, whereas adding large perturbations can significantly change the land- scape of the original Hamiltonian. In the worst case, the final perturbed Hamiltonian may correspond to a problem com- pletely different from the one at hand. Thus, there exists a trade-off between the ability to eliminate systematic bias and the correctness of a Hamiltonian. Ideally, we want a per- turbed Hamiltonian that can eliminate systematic bias without altering the characteristics of the problem Hamiltonian signif- icantly. To address this challenge, EQUAL exploits the fact that QAs only allow a limited precision of coefficients for a Hamiltonian due to hardware limitations. For every ensemble, EQUAL draws the coefficients of the corresponding perturba- tion Hamiltonian randomly at a range just below the so that adding the Perturbation Hamiltonian may only shift the coefficients of the QMI (post truncation) to one of the neighboring quantization levels and not impose signifi- cant changes to the landscape of the original Hamiltonian. We also analyze existing error-mitigation approaches for QAs. Our characterization experiments on D-Wave shows that the SQC [9] post-processing technique is highly effective for D-Wave. We compare EQUAL with SQC and show that the two schemes can be combined for even greater benefit. The resulting design, EQUAL+ , provides significantly better fidelity than EQUAL and SQC standalone. As the SQC post- processing relies only on classical computations, EQUAL+ does not incur any additional trials compared to EQUAL. Our evaluations on D-Wave's 2041-qubit QA show that EQUAL bridges the difference between the baseline and the 1The problem of systematic bias is similar to correlated errors on gate- model quantum computers that can be addressed by mapping programs to different sets of physical qubits on the same machine [86], inserting different SWAP routes [86], or using different machines [71]. The equivalent step for QAs would be to use multiple embeddings. However, this is not viable for QA and we discuss the details in Section 6. 2ideal by an average of 14% (and up to 26%). EQUAL+ further bridges the difference between the baseline and the ideal by an average of 55% (and up to 68%). Overall, this paper makes the following show that there is a systematic bias associated with each QMI, running on QA, that deviates the annealing pro- cess from achieving the ground state of the and produces sub-optimal solutions. 2.We propose EQUAL (Ensemble Quantum Annealing) to mitigate the bias by forming multiple perturbed copies of a given QMI and running each for a subset of trials. 3.We propose an effective method to generate the perturbed copies while retaining the structure of the problem by leverag- ing the hardware imperfections from limited precision. 4.We propose EQUAL+ that combines EQUAL with existing SQC error-mitigation to further improve the reliability. 2. Background and Motivation 2.1. Quantum computing is a computational paradigm that stores and processes information using quantum bits or qubits . The state of a qubitjyican be represented as a superposition of its two basis a aandbare complex probability amplitudes associated with the basis states. Similarly, an N-qubit system exists in a superposition of 2Nbasis states. This exponential scaling in state space with a linear increase in qubits enables Currently, two types of quantum platforms are available to users through cloud services [6, 23, 57]--gate- based quantum computers and quantum Quantum Computers: A gate-based quantum computer executes a predefined sequence of quantum gate op- erations to transform the initial state of the qubits to the desired state by changing the superposition. The sequence of quantum gates is known as a quantum circuit. Quantum computers from IBM, Google, and others use gate-based model. Quantum Annealers: Quantum annealing is a solving combinatorial optimization problems that runs on classical computers [5, 24, 30, 44, 65, 66]. Quantum Annealers are single instruction machines for solving combinatorial op- timization problems. Unlike gate model quantum computers, where we directly change the state of qubits via quantum gates, QAs control the environment, and qubits evolve to remain in the ground state (i.e., a configuration with the lowest en- ergy value) of a Hamiltonian (or energy/cost function) [4, 24]. Quantum annealers, such as the ones from D-Wave, are analog systems that can only minimize the following energy jzizj; 2: (a) Evolution of Quantum Annealers (QAs) over time (b) A cropped view of the D-Wave 2041-qubit QA with its con- nectivity graph where the nodes denote qubits and edges rep- resent couplers (or connectivity between two qubits). where Nis the number of qubits, hi2Rspecifies the of qubit the coupler weight between qubits iandj;andziis the variable that can take its value fromf1;+1g[9, 13, 56]. Ever since the introduction in 2011, QAs have rapidly scaled in size up to few thousand of qubits, as shown in Figure 2(a) and promise advantage for a wide range of Operation Model of QA To execute a program on a QA, the problem is cast to a Hamil- tonian such that its global minimum represents the optimal solution of the problem at hand. This step computes the co- efficients handJ, denoted in Equation (1), corresponding to the quantum machine instruction (QMI) to be executed on the QA. Executing the QMI on a QA returns a zN1gas a potential minimum of the corre- sponding energy function. Unfortunately, executing a QMI only once may not result in the ground state of the Hamiltonian due to noise in the system [9]. Thus, in practice, the process of executing the single QMI is repeated for thousands of trials. The sample with the lowest energy is reported as the solution. 2.3. The Opportunity: Solving Large Problems with QA Google Sycamore is a state-of-the-art 54-qubit computer that can outperform even the most power- ful supercomputer for some tasks [7]. We compare the perfor- mance of the D-Wave 2041-qubit QA and Google Sycamore for 18 different Max-Cut problems. The Max-cut problems used in this evaluation correspond to the (SK) Model [81] and uses up to 17 qubits [37, 79, 89]. These are some of the hardest benchmarks on Sycamore as fully connected graphs require many SWAP operations to overcome the limited connectivity. Running the same benchmarks on the D-Wave QA requires only 102 qubits (less than 1% of the qubits). Figure 3 shows the value of the solution obtained from both machines. We use the same weighted graphs from prior work [37] that result in negative cut values. The performance of the D-Wave QA is comparable to Google Sycamore and both of them are successful in finding the optimal cut at small problem sizes. However, the performance of Google Sycamore degrades with 32 4 6 8 10 12 14 16 18 Max-Cut Cut Value Google Sycamore DWave QAFigure 3: Comparison of 54-qubit Google Sycamore [37] and 2041-qubit D-Wave Quantum problem size [37] due to an increase in SWAPs and circuit depth. Furthermore, due to the limited capacity of 54-qubits, the problem size for Sycamore is limited to no more than 54 nodes. However, as QAs are much qubits), we can use them to solve larger problems more relevant to real-world applications and exceed the size of near-term gate-based quantum computers. For example, Figure 4 shows the performance of the D-Wave QA for Max- Cut problems corresponding to the SK Model using up to 60 qubits. For each problem, the D-Wave QA can find the optimal cut value. To determine the optimal cut value, we evaluate all possible combinations for problems using up to 25 qubits and use the best estimate for the larger problems. 18 19 20 21 22 23 24 25 30 40 50 60 Max-Cut Problem Cut Value Optimal Solution DWave QA Figure 4: Performance of D-Wave QA for larger Max-Cut prob- lems corresponding to the SK Model 2.4. The Challenge: Hardware and Software QAs look promising for various applications, limit us from solving real-world problems on them. 2.4.1. Hardware-Level time: The probability of find- ing the ground state of a Hamiltonian using a QMI with increasing annealing time [4] and theoreti- cally, many hard problems may require large annealing the annealing time on current QAs is in the order of microseconds [9, 56] as qubits can retain their state only for a short span of time. Increasing the annealing time on QA hardware causes qubits to decohere and lose their state. Noise and limited connectivity: Thermal noise and opera-tional errors add unwanted perturbations during annealing and prevent QAs from reaching the ground state of a Hamilto- nian [4]. QAs also suffer from sparse connectivity between qubits, as shown in Figure 2. To address the same drawback on gate-model quantum computers, compilers insert that interchange the state of physically adjacent qubits [60, 88, 95]. However, QAs cannot use a similar ap- proach as they use only one QMI. Instead, we embed the problem graph to match the target device topology where mul- tiple physical qubits represent a program qubit with This can reduce the effective capacity of QAs. Limited precision and range of coefficients: Casting a prob- lem to a Hamiltonian and generating the corresponding QMI coefficients can require a double-precision large precision impacts the performance of converters (DACs) used on the real QAs which slows the annealing process. Therefore, existing QAs trade-off precision to achieve lower annealing times and trun- cate the QMI coefficients post casting to match the in hardware. This subjects the QMI to and the reduced precision QMI actually executed on a QA can be slightly different from the QMI that we origi- nally intended to run, leading to a ground state that may not represent the solution of the problem at hand [9, 26, 78]. 2.4.2. Software-level programmability: QA can only minimize a function and any input problem must be cast to a Hamiltonian. Unfortunately, casting is non-trivial due to lack of standardized algorithms and often comes with [13,56]. Additionally, QAs can only execute a single QMI that performs the annealing step and therefore, fine- grained optimizations at the are of Embedding: To overcome limited connec- tivity, a problem graph is embedded in the QA to match the device topology. Finding the best embedding is NP- hard [18, 21, 25, 34] and existing algorithms can take several hours despite approximations. Also, our studies show they often fail for large programs. 2.5. Impact of Trials on Energy Residual In theory, QAs should find the global optima of a problem with high probability [4, 65]. However, in reality, QAs often fail to find the global optima for large problems due to noise and imperfect control. Moreover, the limited QAs forces users to run a single QMI for thousands of in a bias. As a user runs a single QMI for all trials, the noise profile is similar throughout execution, resulting in similar quality outcomes due to the inherent bias in the noise profile. We refer to this bias as Systematic Bias . Figure 5 shows the Energy Residual (ER) [46, 47] for an optimization problem on D-Wave QA. ER compares the gap between the energy of the solution from a noisy QA and the global minima. The energy of the best solution from a noisy 4QA remains far from the global optima even after running 1 million trials. This non-zero ER occurs due to systematic bias, and is particularly severe for large 5: Energy Residual of an optimization problem on D- Wave QA with increasing number of trials. 2.6. Goal of this Paper Hardware and software limitations of QAs cause programs to encounter a systematic bias during execution which cannot be bypassed by executing more trials. Ideally, we want QAs to be free from systematic bias. In this paper, we propose Ensemble Quantum Annealing (EQUAL) that uses an ensemble of QMIs (with different biases) to mitigate systematic bias. We discuss the evaluation methodology before discussing our solution. 3. Evaluation Methodology We discuss the evaluation infrastructure used in this paper. 3.1. Quantum Platform and Baseline For our evaluations, we use the 2041-qubit quantum annealer from D-Wave Systems via Amazon BraKet cloud service [6]. We use the default annealing time (i.e., 20 mseconds) and schedule recommended for this system. For the baseline, we use 100,000 trials for each benchmark. For EQUAL, trials are equally split between QMIs. Thus, EQUAL requires the same number of trials as the baseline. 3.2. Benchmarks We use random weighted Max-Cut problems, similar to Quan- tum Approximate Optimization Algorithms [28] used on gate- based quantum computers. For the benchmarks, we draw the Hamiltonian coefficients of the QMIs from the standard normal distribution (a mean of 0 and standard deviation of 1). This approach is a common practice used in prior works related to benchmarking QAs [9, 11, 19, 24, 78]. To avoid the impact of embedding on our evaluations, we directly use the connectivity graph of the D-Wave QA. Thus, the number of program qubits in benchmarks is equal to the number of phys- ical qubits on the QA. As the size of benchmarks the size of existing gate-model quantum computers, we cannot compare our results with them.3.3. evaluate the reliability of QA using Energy Residual (ER) . The best solution from a QA is the outcome with the minimum energy. ER computes the energy gap between the minimum energy (Emin)obtained on a QA with respect to the global energy minimum (Eglobal)of the application as follows: Energy Residual (ER) (2) Ideally, when the best solution obtained on a QA corresponds to the ground state of the problem Hamiltonian, ER is zero. Thus, a lower value (closer to zero) for ER is desirable. The challenge in computing the ER for random large bench- marks spanning 2000+ qubits is that finding the ground state of the Hamiltonian is non-trivial. To overcome this challenge and still enable a fair comparison, we perform intensive using state-of-the-art tools [10] and approximate the global optimum of our benchmark problems. Recent stud- ies have shown that this algorithm can estimate the ground state of Chimera based Hamiltonians [21,42] (such as the ones considered in our paper) with a very high probability.2 4. EQUAL: Ensemble Quantum Annealing The vulnerability of a program to systematic bias results from limited programmability and the current execution model of QAs where the same QMI is executed for thousands of trials. This subjects each trial to a similar noise profile on the QA and the entire execution suffers from the same inherent bias. Our proposed solution EQUAL --Ensemble Quantum a different approach. Instead of a single QMI, EQUAL generates an ensemble of QMIs that subjects the to different noise profiles and therefore, biases. When results are aggregated, ensembles enable us to improve the quality of solutions. 4.1. Challenges in Generating Ensembles in EQUAL There is potential to generate ensembles during any one of the three phases that a problem goes through before execution on a physical QA hardware: (1) casting, (2) embedding, and (3) QMI generation. Generating ensembles during casting was previously studied in the context of Boolean [13] and binary compressive sensing [12] problems on QAs. Unfortunately, these methods exploit the features of the casting algorithms. Therefore, this approach has limited applicability and is hard to generalize for QAs. The other alternative approach is to use an ensemble of embeddings for a given problem. However, this approach too has its limitations. Firstly, finding the best embedding is an NP-hard problem in itself [18, 21, 25, 34]. Secondly, schemes for QAs use several approximations and 2The techniques used to derive the best estimate of the ground state energy of a Hamiltonian requires intensive classical computing resources and could take up-to days for problem sizes with a few thousand qubits. We discuss more on this in Section 6 5Cast to Hamiltonian, Embed to Hardware, Generate QMIQMI = HExecute on QA Controlled on QA HardwarePost Generate Ensemble of QMIsQMI1 = H + /u1D6FF1 = H1QMI2 = H + /u1D6FF2 = H2QMIm = H + /u1D6FFN = 6: Overview of EQUAL. EQUAL creates an ensemble of QMIs by adding controlled perturbations to the original QMI. It executes the original QMI as well as the ensemble of QMIs separately on the QA hardware and returns the outcome with the lowest energy value. EQUAL can also optionally leverage existing post-processing error mitigation schemes (EQUAL+). may or may not be able to determine an ensemble of embed- dings of similar quality [18, 21, 25, 34]. Our studies show that existing embedding algorithms often fail to find an adequate number of embeddings, particularly for problems at scale that require 2000+ qubits. Thirdly, even if it is possible to find multiple embeddings, they are often of inferior quality and require larger chains of physical qubits to represent a program qubit with higher connectivity. This makes the more vulnerable to noise compared to the best em- bedding. Thus, generating ensembles at the embedding step is non-trivial. Instead, EQUAL focuses on generating ensembles at the and produces multiple QMIs. 4.2. Overview of Design Figure 6 shows an overview of EQUAL. It relies on perturbations to the original QMI. For each ensem- ble, EQUAL generates a Perturbation Hamiltonian , denoted byd:Each of these Perturbation Hamiltonians creates a new QMI when added to the original Hamiltonian. For example, if EQUAL generates mensembles of QMIs, it generates mper- turbation Hamiltonians, namely d1;d2;:::;dm:The 1, QMI 2to QMI m--are obtained by adding the original Hamiltonian (say H)and the respective In other words, the ensemble of QMIs now cor- responds to the perturbed versions of the original Generating Ensembles via Controlled an effective perturbation Hamiltonian is non-trivial. If the perturbations add too little noise, the resulting Hamil- tonian will be too close to the problem Hamiltonian and en- counter similar bias. Alternately, too large perturbations re- sult in a Hamiltonian significantly different from the problem of interest and can produce infeasible results. For example, Figure 7(a) shows the landscape of an example Figure 7(b) shows that injecting an extremely Hamiltonian significantly changes the landscape of the original problem. Thus, there is a trade-off between the effectiveness of a perturbation Hamiltonian to reduce bias and its ability to alter the problem Hamiltonian. To challenge and generate an effective ensemble of QMIs, EQUAL exploits the device-level characteristics of 7: (a) Landscape of an example optimization problem. (b) The resultant landscape differs significantly from (a) when an extremely noisy perturbation is imposed. (This figure is for illustrative purposes only. Hamiltonians and QAs can only deal with discrete optimization Exploiting Hardware Characteristics of QAs Recollect that casting a problem to a Hamiltonian can re- quire double-precision representation of the Hamiltonian co- efficients. Unfortunately, real QAs can only support a small range and precision of coefficients due to the limitations im- posed by the digital to analog converters (DACs) used on QAs. If the precision of the coefficients are too large, the DACs are too slow which eventually slows the controlling modules of QAs and is not desirable. To bridge this gap, post the casting step, the coefficients of the QMI are truncated to match the precision supported by the hardware. While this is a limitation on QAs, EQUAL leverages it to its advantage and draws the coefficients of the perturbation Hamiltonian randomly at a range that is below the supported precision so that adding the perturbation Hamiltonian only shifts the coefficients of the QMI (post truncation) to one of the neighboring and thus, does not significantly alter the problem land- scape. More specifically, let bbe the number of bits used for representing coefficients of a physical QA. For every ensem- ble, EQUAL draws a uniform random number set all coefficients of the Perturbation Hamiltonian to be r. 64.3.2. Profiling QAs to estimate Hardware the precision of the coefficients supported on real devices is unavailable to programmers. Determining this precision is vital for the performance of EQUAL. Drawing the perturbation Hamiltonian coefficients far below the introduces large noise and may alter the Hamilto- nian landscape significantly. Alternately, drawing them far above the supported range may not have any effect post trunca- tion. To tackle this challenge, EQUAL profiles the QA using random benchmarks to estimate the precision supported by QAs. In this experiment, we truncate all coefficients of the benchmark for 2;3;:::; 16bits precision and execute the corre- sponding QMIs. Figure 9 shows the relative Energy Residual of the truncated QMIs with respect to the original problem (without truncation). Our profiling experiments with multi- ple benchmarks show that the hardware is likely limited by 7-8 bits of precision. Thus, EQUAL generates coefficients of ensembles 4 6 8 10 12 14 16 Precision of 9: Relative Energy Residual of QMIs with truncated co- efficients with respect to the original problem QMI for bits val- ues of precision. (Lower is better) 4.4. Execution on QA Hardware EQUAL splits the trials between the ensemble of QMIs equally, including the original QMI (without executes them separately on the QA hardware. Our de- fault design uses 10 ensembles of QMIs, and allocates 10,000 trials for every ensemble. We do a more rigorous for the number of trials and ensembles in Section 5.8. By default, the outcome with lowest energy is deemed as the solution for problems executed on QAs. In the baseline, this corresponds to the outcome with the lowest energy obtained by executing the original QMI. As EQUAL executes multiple QMIs, the outcome with the lowest energy among all the QMIs is returned as the solution. Also, as EQUAL runs the ensemble of perturbed QMIs in addition to the original program QMI, the final solution is guaranteed not to perform worse than the baseline, assuming there are no sampling errors. Note that the solution with the minimum energy corresponds to an outcome that may come from a single QMI. For the baseline this corresponds to the original QMI, whereas for EQUAL it comes from one or more of the QMIs in the which QMI corresponds to the best solution is not known a-priori and EQUAL must execute the entire ensemble. 4.5. Results for Energy Residual Energy Residual (ER) computes the gap between the energy obtained from the best outcome on a QA with the global op- tima. Figure 8 compares the ER of the individual benchmarks for baseline and EQUAL. We observe that the ER in the baseline for all benchmarks, whereas improves for EQUAL as more QMIs are executed. As the QMIs are gen- erated using random controlled perturbations, some of them may result in higher ER compared to the baseline due to a different noise profile at run time. However, the ensemble overall enables EQUAL to reach a better solution. In the worst case, EQUAL performs similar to baseline as the original pro- gram QMI is executed too. We observe that the fidelity of the baseline saturates with more trials, whereas the diversity of EQUAL helps it keep on improving with additional trials. Figure 8: Trends in Energy Residual for the baseline and EQUAL for all the 10: Histogram of energy values from the outcomes on the QA for benchmark B1 using the Baseline and EQUAL. The solution from EQUAL is closer to the ideal solution compared to the baseline solution ( d2<d1). The histograms for the baseline and EQUAL largely overlaps which indicates that EQUAL does not significantly alter the problem 11 shows the ER of EQUAL for our on D-Wave 2041-qubit QA relative to the baseline. We observe that EQUAL bridges the difference between the baseline and the ideal by an average of 14% (and up to 26%). QAs deal with industry-scale optimization problems where even a minuscule improvement has a tremendous impact in terms of practical advantage such as saving millions of dol- lars [1, 22, 27] in the context of scheduling and planning ap- plications or finding better candidates for drug discovery [59] and material science [49]. Thus, the quality of solutions is of utmost importance. B1 B2 B3 B4 B5 B6 11: Energy Residual of six random benchmarks on D- Wave QA hardware using EQUAL relative to the baseline. 4.6. Case-Study: How EQUAL Reduces Systematic Bias Figure 10 shows the histograms of the energy values obtained by running benchmark B1 for the baseline and EQUAL. The goal of QAs is to obtain the outcome corresponding to the ground state energy. We observe that the optimal solution is at a distance d1from the ground state and EQUAL produces a solution at distance d2that is closer to the ground state energy (d2<d1) by minimizing the impact of bias. We also observe that distributions for both the baseline and EQUAL overlap largely, indicating that the ensemble of QMIs do not largely alter the original Hamiltonian corresponding to our problem. We make similar observations for other benchmarks. An ideal QA is a machine that samples from the whereas the distribution obtained from a real QA hardware is different due to noise. The best solution obtainedby a QA depends on the overlapping region between the ideal and the noisy distributions. From Figure 10, we observe two potential approaches to get closer to the global optima. the energy histogram of the Hamiltonian such that it covers a broader search space. Second, by shifting the energy histogram towards the ideal solution. Note that both of these techniques must ensure that the properties of the original program Hamiltonian remain unaltered. EQUAL uses the first approach. The performance of EQUAL can be improved further if we could shift the histogram closer to the ideal solution. We explore combining EQUAL with existing error mitigation schemes to obtain the advantage of both flattening the histogram and shifting the histogram towards the ground state. 5. Combining EQUAL with are generated by only adding controlled perturba- tions to the problem Hamiltonian. Therefore, they have to shift the noisy distribution from a QA towards the ideal distribution even if a large number of ensembles are used. Alternately, large perturbations may significantly change the landscape of the problem. Instead, we take an and explore existing error-mitigation schemes that introduce a shift in the energy histogram. 5.1. Primer on Error-Mitigation Schemes for schemes for QAs can be classified into (1) software and (2) hardware schemes. Software schemes refer to optimizations performed during the casting and embedding steps (pre-processing techniques) or modifications on the out- comes obtained from QAs (post-processing techniques). On the other hand, hardware-based schemes control the device- level parameters on QAs to reduce the impact of errors. We characterize the impact of these error-mitigation tech- niques individually and combined with each other to under- stand their effectiveness in (1) eliminating systematic bias on their own and (2) shifting the noisy distribution of the QA towards the ideal distribution. For our analysis, we choose (a) spin reversal transform, (b) longer inter-sample delay, and (c) single-qubit correction. Spin-reversal transform is a candidate for a software pre-processing technique. On the other hand, inter-sample delay is a device-level to programmers to reduce the correlation trials on a QA. Lastly, Single-Qubit Correction (SQC) [9] is a post-processing technique which leverages the insight that for a given QMI, QAs can quickly recognize the neighborhood of the ground state even if they fail to get to the ground state [9]. We perform characterization studies for these three error mitigation schemes (see Appendix A) and found that SQC is the most effective scheme, and therefore we use SQC as the error mitigation scheme for our study. 0 1 1 Cost Function for neighborsKeep best neighbor, if it is betterQMI outcome1 1 00 0 00 1 1 Figure 12: Single Qubit Correction Post-Processing [9] 5.2. Overview of SQC is analogous to the gradient descent scheme but only applicable to discrete optimization problems. Instead of com- puting the gradient for determining the direction of the move in every iteration, SQC uses a greedy approach and moves to a neighbor (i.e., an outcome that is one Hamming distance away from the current solution) with the lowest energy value. Figure 12 illustrates the overview of an iteration of SQC. For each candidate outcome generated by executing the QMI, SQC finds the one Hamming distance away neighbors and computes their energy values. If any neighbors can obtain a lower energy value than the candidate itself, the neighbor is retained and the current solution is discarded. When multiple neighbors obtain lower energy values, the best neighbor is retained. The process is repeatedly executed until we cannot find any new neighbor that has better quality. 5.3. EQUAL+: Combining EQUAL and SQC Figure 13 shows an overview of EQUAL+. EQUAL+ applies SQC on the outcomes of each QMI and obtains the best out- come for each QMI. The process is performed for each QMI in parallel. Once applying SQC on each QMI converges, the final output of EQUAL+ is picked as the candidate with the lowest energy among all the individual best candidates from the QMIs. The time to converge depends on several factors such as the size of the problem, number of outcomes, quality of the outcomes. However, our evaluations show that EQUAL+ converges within a few seconds even for large benchmarks such as the ones used in our this greedy approach helps locate neighbors from current outcomes that were not produced by the QA originally. With each neighbor located, EQUAL+ shifts the outcome dis- tribution towards the ideal solution (global optima). Note that QMISQCBest Best Outcome from all 13: Overview of EQUAL+ design. It applies the algorithm on the outcomes from each QMI in parallel. Finally, it selects the best outcome from all the QMIs as the output SQC is effective on its own, the diversity of EQUAL+ is essential to improve its search space. The capability of SQC alone to introduce new outcomes is limited by the quality of outcomes from the QMI. In EQUAL+, the ensembles enable us to explore a much larger neighborhood compared to apply- ing SQC alone. In the end, EQUAL+ may discover a solution from one of the weakest outcomes corresponding to one of the weakest QMIs (sub-optimal outcome that did not correspond to the best solution in any of the QMIs). Note that EQUAL+ is versatile, and any other post- processing candidate that introduces the desired shifting prop- erty in the energy distribution may be used. We use SQC for its performance and low time complexity. 5.4. Analysis of Overheads We discuss the overheads for both EQUAL and EQUAL+. EQUAL generates the ensemble of QMIs prior to execution on the QA. As the perturbed Hamiltonian only adjusts the co- efficients of the original problem QMI, the ensemble does not need to re-perform the casting or embedding step. can take up to several hours and may fail for this overhead and limitation is entirely avoided by EQUAL. EQUAL also requires the programmer to esti- mate the precision of the hardware using a set of However, profiling need not be done for As the precision supported is only once for each QA hardware is enough and the same information can be re-used for multiple applications. For exe- cution on the QA, EQUAL requires the same number of trials as the baseline and therefore, does not incur any overhead of additional trials. EQUAL+ incurs some additional overheads for the post- processing step as it applies the SQC heuristic algorithm on all the outcomes obtained from all the QMIs. The space complex- ity of the post-processing phase in EQUAL+ is linear with the number of qubits [9]. As SQC is iteratively applied on every outcome of a QMI, the time complexity depends on the num- ber of outcomes which is equal to the number of trials in the worst-case (assuming each trial generates a unique outcome). The post-processing for each QMI is done in parallel. Our studies show that EQUAL+ converges within a few iterations and the post-processing step for EQUAL+ only takes a few seconds. Therefore, the overheads are 14: Histogram of energy values from the outcomes on the QA for benchmark B1 using EQUAL and EQUAL+. 5.5. Results for Energy Residual Figure 15 shows the Energy Residual of EQUAL+ relative to the baseline. We also compare against EQUAL and SQC standalone. We observe that EQUAL+ improves the ER by 0.45 compared to the baseline on average and by up to 0.32. In other words, EQUAL+ improves the quality of solutions by 55% on average and up to 68%. B1 B2 B3 B4 B5 B6 SQC EQUAL+ Figure 15: Energy Residual using EQUAL+ relative to the base- line. We also compare with EQUAL and SQC standalone. 5.6. Results for Validation of Precision Selection We draw the perturbation Hamiltonian coefficients in the on profiling across a wide range of values. We confirm that this approach is robust by conducting at application level. Figure 16 shows the ER of some benchmarks relative to baseline when the precision range is varied. We confirm that 8 to 9 bits of precision is more robust compared to others and on average outperforms the others. 5.7. Case-Study: How EQUAL+ reduces Systematic Bias Figure 14 shows the histograms of energy values of benchmark B1 for EQUAL and EQUAL+. We observe that the optimal solution is at a distance d1from the ground state energy in EQUAL. EQUAL+ exploits the shifting property of SQC to obtain a solution at distance d2and is closer to the ground state energy ( d2<d1). Note that EQUAL+ shifts the towards the ideal solution and achieves the intended goal. As EQUAL+ applies post-processing on the outcomes from the QMIs, the introduced shift in the histogram does not alter the original problem Hamiltonian. B1 B2 B3 B4 B5 bits 7-8 bits8-9 bits 9-10 bitsFigure 16: Energy Residual of random benchmarks on D-Wave QA hardware using EQUAL relative to the baseline for different values of bit precision used for generating QMIs. 5.8. Impact of Number of Ensembles We study the impact of the number of ensembles on the effec- tiveness of EQUAL using a single benchmark problem. For a given trial budget of 100K trials, we choose two modes for EQUAL. In the first instance, we use 10 QMIs and run each of them for 10K trials each. In the second instance, we use 100 QMIs and run each of them for 1K trials each. Figure 17 shows the ER for the baseline and these two instances of EQUAL. Note that we access the QA device through cloud services and a more rigorous sensitivity analysis in terms of QMIs and trials is challenging. We observe that executing more QMIs introduces more randomness and makes them vulnerable to sampling errors. EQUAL with 10 QMIs achieve a sweet spot between the baseline and EQUAL with a large number of en- sembles such that we have both diversity as well as for each QMI to reduce sampling errors. 0K 20K 40K 60K 80K 100K Number of QMIs100 QMIs Figure 17: Energy Residual for benchmark (B1). The a single QMI for all 100K trials. EQUAL has 10 QMIs for 10K trials each or 100 QMIs executed for 1K trials each. 106. Related Work Both gate-model quantum computers and QAs promise sig- nificant advantages for a wide range of applications [7, 12 - 14, 16, 17, 27, 28, 39, 43, 45, 49, 59, 68, 74 -76, 80, 84, 90, 91]. Thus, developing error-mitigation policies is an active area of research for both QAs and gate-model quantum computers. We discuss prior works and compare against schemes that Priors works using Ensembles The potential of ensembles has been explored for both gate- model quantum computers and QAs. Ensemble policies for Gate-model quantum computers : Systematic bias in QAs is similar to correlated errors on gate- based quantum computers. To tackle these errors on gate- model quantum computers, recent studies propose the use of ensemble of mappings that maps a program to different sets of physical qubits and SWAP routes on the same [86] or [71]. This process produces functionally identical copies of the same program but are only executed a similar approach for EQUAL is non-trivial due to the complexities involved in the embedding process, partic- ularly for problems at scale. Obtaining alternate embeddings is non-trivial, may fail or result in inferior quality. Instead, EQUAL uses an ensemble of QMIs by introducing while minimizing the alterations in the function- ality of the original problem policies for QAs : Using ensembles in QAs have been investigated at the casting level for two different appli- cations. However, as each application uses its own this approach cannot be generalized. EQUAL on the other hand, avoids such assumptions and is applicable irrespective of the problem at hand. In an- other study, Mohseni et al. proposed a multi-level embedding scheme that uses a diverse encoding of qubits to generate en- sembles [58]. However, this approach reduces the capacity of the QA significantly. It is also not scalable as it to the embedding step, which already takes several hours for current systems. Moreover, our studies show that finding alternate embeddings frequently fail or result in embed- dings of inferior quality for large problems. EQUAL avoids these overheads by introducing diversity post embedding. 6.2. Software error mitigation policies These techniques are either applied prior to the execution of the QMI (pre-processing) or after the QMI is executed Pre-processing schemes transform the problem QMI at the casting or embedding level such that it is less vulnerable to errors during execution time [9, 12, 13, 73]. Pre- processing schemes are analogous to compiler-level optimiza- tions on gate-model quantum computers [2,3,15,31,32,50 -52, 60, 61, 61 -63, 69 -71, 82, 85 -88, 95]. Unlike QAs, computers have higher programmability and allowprogrammers to leverage fine-grained optimizations at the in- struction scheduling level. Post-processing schemes for QAs exploit the fact that even if a QA cannot generate the solution with the lowest energy, it quickly locates the the optimal solution might reside. By modifying the outcome obtained from the QA using classical heuristic algo- rithms, post-processing schemes can significantly improve the quality of solutions [9, 19, 33]. One of the most schemes is Multi-Qubit Correction (MQC) [9]. We use this scheme to obtain the best known estimate of the ground state energy in our evaluations. However, this algo- rithm requires significant classical computational overheads and may take up to days to obtain a better quality the performance of MQC depends on the quality of the outcomes obtained from the QA and both EQUAL and EQUAL+ can benefit from it. Post-processing algorithms improve the application fidelity even for gate- based quantum computers [20, 40, 71, 72]. 6.3. Hardware-level error studies have proposed improvements to the itself [4, 65] but implementing them requires fabri- cating newer QAs. Few other schemes reduce the impact of noise on the annealing process. Examples include controlling the delay between two consecutive trials and qubit prepara- tion times [41]. Exposing the hardware characteristics and enabling device-level controls to the programmer has been shown to be effective even for gate-model systems [8]. 7. Annealers (QAs) have thousands of qubits and are promising for a wide range of applications. Unfortunately, QA suffers from both hardware and software limitations. There- fore, there is increasing interest in developing software policies to tackle these limitations. However, our studies show that executing a program on a QA makes it vulnerable to a system- atic bias that cannot be overcome by increasing the number of trials or relying on existing error-mitigation schemes. This paper proposes EQUAL--Ensemble software framework that creates multiple per- turbed copies of an input problem by injecting to the original problem Hamiltonian. By exe- cuting an ensemble of quantum machine instructions (QMIs), EQUAL projects the program to different noise profiles and therefore, different biases. Our evaluations using the 2041- qubit D-Wave QA show that EQUAL bridges the the baseline and the ideal by an average of 14% (and up to 26%), without requiring any additional trials. We also propose EQUAL+ that exploits the properties of existing er- ror mitigation schemes for enhanced performance. EQUAL+ bridges the difference between the baseline and the ideal by an average of 55% (and up to 68%). 11A. Appendix: Characterizing Error Mitigation While designing EQUAL+, we have various options regarding how to combine existing error-mitigation schemes. To under- stand suitable candidates, we perform several whose results are described next. A.1. Spin-Reversal vs. Single Qubit Transform (SRT) is a pre-processing which flips randomly selected qubits to mitigate analog errors of QAs [73]. Figure 19 shows the Energy Residual of a benchmark executed on D-Wave QA when spin reversal transform is applied stan- dalone and in the presence of SQC. We observe that both spin reversal and SQC reduces the ER, and the performance of SQC is comparable to spin-reversal transform. 1K 2K 3K 4K 5K 6K 7K 8K 9K 10K Number of Residual QA Hardware SQCSpin Reversal Spin Reversal + SQC Figure 18: Energy Residual of a random benchmark with Spin Reversal Transform and Single Qubit Correction. A.2. Inter-Sample Delay vs. Single Qubit the (ISD) which increases the delay between successive QA reads to reduce the Figure 19 shows the ER of a benchmark exe- cuted on D-Wave QA under default ISD and when the ISD is increased. We also study the performance in the presence of SQC. We observe that similar to spin-reversal transform, while longer ISD reduces the ER, SQC outperforms and can be combined for further benefits. 1K 2K 3K 4K 5K 6K 7K 8K 9K 10K Number of Residual QA ISD Longer ISD + SQC Figure 19: Energy Residual of a random benchmark with longer ISD and Single Qubit Correction. We leverage the insights from this study to investigate if EQUAL can benefit from post-processing too. The of relying on post-processing schemes is that it does not add any overheads in the casting or embedding Ayanzadeh was supported by the NSF Computing In- novation Fellows (CI-Fellows) program. Poulami Das was supported by the Microsoft Research PhD fellowship. This research was partially supported by the Office of the Vice Chancellor for Research and Graduate Education at the Univer- sity of with funding from the Wisconsin Alumni Research K Ahuja, Claudio B Cunha, and Guvenc , Sahin. Network models in railroad planning and scheduling. In Emerging theory, meth- ods, and applications , pages 54-101. INFORMS, Alam, Abdullah Ash-Saki, and Swaroop Ghosh. methodologies for quantum approximate In 2020 53rd Annual IEEE/ACM International Symposium on (MICRO) , pages 215-228. IEEE, Alam, Abdullah Ash-Saki, and Swaroop Ghosh. An effi- cient circuit compilation flow for quantum approximate In 2020 57th ACM/IEEE Design Automation Conference (DAC) , pages 1-6. IEEE, 2020. [4]Tameem Albash and Daniel A Lidar. Adiabatic quantum of Modern Physics , 90(1):015002, Amara, D Hsu, and John E Straub. Global energy mini- mum searches using an approximate solution of the imaginary time schrodinger equation. The Journal of Physical Chemistry , 1993. [6]Amazon. Amazon Braket - Explore and experiment with https://aws :amazon :com/braket/ , 2022. Arute, Kunal Arya, Ryan Babbush, Dave Bacon, Joseph C Bardin, Rami Barends, Rupak Biswas, Sergio Boixo, Fernando GSL Brandao, David A Buell, et al. Quantum supremacy using a pro- grammable superconducting processor. Nature , Asfaw, Thomas Alexander, Paul Nation, and Jay Gambetta. Get to the heart of real quantum hardware. https://www , 2019. [Online; Ayanzadeh, John Dorband, Milton Halem, and Tim Finin. Multi- qubit correction for quantum annealers. Scientific Reports , 11, 2021. [10] Ramin Ayanzadeh, John Dorband, Milton Halem, and Tim Finin. Multi qubit correction (mqc) for quantum annealers, 2021. Python imple- mentation of MQC. [11] Ramin Ayanzadeh, Milton Halem, John Dorband, and Tim greedy algorithms. arXiv preprint arXiv:1912.02362 , 2019. [12] Ramin Ayanzadeh, Milton Halem, and Tim Finin. An ensemble ap- proach for compressive sensing with quantum annealers. In IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Sym- posium , pages 3517-3520. IEEE, 2020. [13] Ramin Ayanzadeh, Milton Halem, and Tim Finin. annealing: A hybrid quantum learning automata. , 10(1):1-11, 2020. [14] Ramin Ayanzadeh, Seyedahmad Mousavi, Milton Halem, and Tim Finin. Quantum annealing based binary compressive sensing with matrix uncertainty. arXiv preprint arXiv:1901.00088 , 2019. [15] George S Barron and Christopher J Wood. Measurement error mitigation for variational quantum algorithms. arXiv , 2020. [16] Zhengbing Bian, Fabian Chudak, Robert Brian Israel, Brad Lackey, William G Macready, and Aidan Roy. Mapping constrained optimiza- tion problems to quantum annealing with application to fault in ICT , 3:14, 2016. [17] Rupak Biswas, Zhang Jiang, Kostya Kechezhi, Sergey Knysh, Sal- vatore Mandra, Bryan O'Gorman, Alejandro Perdomo-Ortiz, Andre Petukhov, John Realpe-Gomez, Eleanor Rieffel, et al. A nasa perspec- tive on quantum computing: Opportunities and challenges. , 64:81-98, 2017. [18] Tomas Boothby, Andrew D King, and Aidan Roy. Fast clique minor generation in chimera qubit connectivity graphs. Quantum , 15(1):495-508, 2016. 12[19] Ajinkya Borle and Josh McCarter. On post-processing the results of quantum optimizers. In International Conference on Theory and Practice of Natural Computing , pages 222-233. Springer, 2019. [20] Sergey Bravyi, Sarah Sheldon, Abhinav Kandala, David C Mckay, and Jay M Gambetta. Mitigating measurement errors in arXiv preprint arXiv:2006.14044 , 2020. [21] Jun Cai, William G Macready, and Aidan Roy. A practical heuristic for finding graph minors. arXiv preprint arXiv:1406.2741 , 2014. [22] Brian Carlson, Yonghong Chen, Mingguo Hong, Roy Jones, Kevin Lar- son, Xingwang Ma, Peter Nieuwesteeg, Haili Song, Kimberly Sperry, Matthew Tackett, et al. Miso unlocks billions in savings through the application of operations research for energy and ancillary Interfaces , 42(1):58-73, 2012. [23] International Business Machines Corporation. Universal Quantum Computer Development at IBM:. http://research , 2021. [Online; accessed Arnab Das and Bikas K Chakrabarti. Colloquium: Quantum anneal- ing and analog quantum computation. Reviews of Modern Physics , 80(3):1061, 2008. [25] Prasanna Date, Robert Patton, Catherine Schuman, and Thomas embedding qubo problems on adiabatic quantum Information Processing , 18(4):1-31, 2019. [26] John E Dorband. Extending the d-wave with support for higher preci- sion coefficients. arXiv preprint arXiv:1807.05244 , 2018. [27] Nada Elsokkary, Faisal Shah Khan, Davide La Torre, Travis S Hum- ble, and Joel Gottlieb. Financial portfolio management using d-wave quantum optimizer: The case of abu dhabi securities exchange. Techni- cal report, Oak Ridge National Lab.(ORNL), Oak Ridge, TN (United States), 2017. [28] Edward Farhi, Jeffrey Goldstone, and Sam Gutmann. A optimization algorithm. arXiv preprint arXiv:1411.4028 , 2014. [29] RP Feynman. Simulating physics with computers. of Theoretical Physics , 21(6), 1982. [30] AB Finnila, MA Gomez, C Sebenik, C Stenson, and JD Doll. a new method for minimizing multidimensional physics letters , 1994. [31] Pranav Gokhale, Yongshan Ding, Thomas Propson, Christopher Win- kler, Nelson Leung, Yunong Shi, David I Schuster, Henry Hoffmann, and Frederic T Chong. Partial Compilation of Variational Algorithms for Noisy Quantum Machines. In Proceedings of the 52nd Annual IEEE/ACM International Symposium on , pages 266-278. ACM, 2019. [32] Pranav Gokhale, Ali Javadi-Abhari, Nathan Earnest, Yunong Shi, and Frederic T Chong. Optimized Quantum Compilation for with OpenPulse. arXiv preprint arXiv:2004.11205 , 2020. [33] John K Golden and Daniel O'Malley. Pre-and post-processing hydrologic inverse analysis. Quantum Infor- mation Processing , 20(5):1-18, 2021. [34] Timothy D Goodrich, Blair D Sullivan, and Travis S Humble. Optimiz- ing adiabatic quantum program compilation using a Quantum Information Processing , 17(5):1-26, 2018. [35] Google. Google Quantum AI. :google/ , 2022. [Online; accessed Lov K Grover. A fast quantum mechanical algorithm for database search. In Proceedings of the twenty-eighth annual ACM symposium on Theory of computing , pages 212-219, 1996. [37] Matthew P Harrigan, Kevin J Sung, Matthew Neeley, Kevin J Satzinger, Frank Arute, Kunal Arya, Juan Atalaya, Joseph C Bardin, Rami Barends, Sergio Boixo, et al. Quantum approximate optimization of non-planar graph problems on a planar superconducting Physics , 17(3):332-336, 2021. [38] Honeywell. Honeywell Quantum Solutions. , 2021. [Online; ac- cessed Feng Hu, Lucas Lamata, Mikel Sanz, Xi Chen, Xingyuan Chen, Chao Wang, and Enrique Solano. Quantum computing cryptography: Find- ing cryptographic boolean functions with quantum annealing by a 2000 qubit d-wave quantum computer. Physics Letters A , IBM. Measurement Error Mitigation. https://qiskit :html , 2010. [Online; accessed D-Wave Systems Inc. D-wave ocean software documentation. https: //docs :ocean :dwavesys :com/en/stable/ , 2022. [Online; ac- cessed D-Wave Systems Inc. The first and only quantum computer built for business. https://www :dwavesys :com/ , 2022. [Online; Daisuke Inoue, Akihisa Okada, Tadayoshi Matsumori, Kazuyuki Ai- hara, and Hiroaki Yoshida. Traffic signal optimization on a square lattice with quantum annealing. Scientific reports , 11(1):1-12, 2021. [44] Tadashi Kadowaki and Hidetoshi Nishimori. Quantum annealing in the transverse ising model. Physical Review E , 58(5):5355, 1998. [45] Abhinav Kandala, Antonio Mezzacapo, Kristan Temme, Maika Takita, Markus Brink, Jerry M Chow, and Jay M Gambetta. quantum eigensolver for small molecules and quantum magnets. Nature , 2017. [46] Hamed Karimi and Gili Rosenberg. Boosting quantum annealer per- formance via sample persistence. Quantum Information Processing , 16(7):166, 2017. [47] Hamed Karimi, Gili Rosenberg, and Helmut G Katzgraber. Effec- tive optimization using sample persistence: A case study on and various monte carlo optimization methods. Physical Review E , 96(4):043312, 2017. [48] Andrew D King, Jack Raymond, Trevor Lanting, Sergei V Isakov, Masoud Mohseni, Gabriel Poulin-Lamarre, Sara Ejtemaee, Isil Ozfidan, Anatoly Yu Smirnov, et al. Scaling advantage over path-integral monte carlo in quantum simulation of magnets. Nature communications , 12(1):1-6, 2021. [49] Koki Kitai, Jiang Guo, Shenghong Ju, Shu Tanaka, Koji Tsuda, Ju- nichiro Shiomi, and Ryo Tamura. Designing metamaterials with quan- tum annealing and factorization machines. Physical Review Research , 2(1):013319, 2020. [50] Hyeokjea Kwon and Joonwoo Bae. A hybrid to mitigating measurement errors. arXiv , 2020. [51] Gushu Li, Yufei Ding, and Yuan Xie. Tackling the Qubit Mapping Prob- lem for NISQ-Era Quantum Devices. arXiv preprint arXiv:1809.02573 , 2018. [52] Junde Li, Mahabubul Alam, and Swaroop Ghosh. Large-scale quan- tum approximate optimization via arXiv , 2021. [53] Seth Lloyd. Universal quantum simulators. Science , pages Andrew Lucas. Ising formulations of many np problems. Frontiers in physics , 2:5, 2014. [55] C McGeoch and P Farre. The d-wave advantage system: An report, Tech. Rep. (D-Wave Systems Inc, Burnaby, BC, Canada, 2020). [56] Catherine C McGeoch. Theory versus practice in computing. Theoretical Computer Science , 2020. [57] Microsoft. Azure Quantum - Quantum Service | Microsoft :microsoft , 2022. [Online; accessed Naeimeh Mohseni, Marek Narozniak, Alexey N Pyrkov, Valentin Ivan- nikov, Jonathan P Dowling, and Tim Byrnes. Error suppression in adiabatic quantum computing with qubit ensembles. npj , 7(1):1-10, 2021. [59] Vikram Khipple Mulligan, Hans Melo, Haley Irene Merritt, Stewart Slocum, Brian D Weitzner, Andrew M Watkins, P Douglas Renfrew, Craig Pelissier, Paramjit S Arora, and Richard Bonneau. on a quantum computer. bioRxiv , page 752485, 2020. [60] Prakash Murali, Jonathan M Baker, Ali Javadi-Abhari, Frederic T Chong, and Margaret Martonosi. Noise-adaptive compiler mappings for noisy quantum computers. In Proceedings of International Conference on Architectural Support for Programming Languages and Operating Systems , pages Prakash Murali, Norbert M Linke, Margaret Martonosi, Ali Javadi Abhari, Nhung Hong Nguyen, and Cinthia Huerta Alderete. Architect- ing noisy quantum computers: a real-system study. IEEE Micro , 40(3):73-80, 2020. [62] Prakash Murali, Norbert Matthias Linke, Margaret Martonosi, Ali Javadi Abhari, Nhung Hong Nguyen, and Cinthia Huerta real-system quantum computer studies: Architectural com- parisons and design insights. In 2019 ACM/IEEE 46th Annual Interna- tional Symposium on Computer Architecture (ISCA) , pages 527-540. IEEE, 2019. [63] Prakash Murali, David C McKay, Margaret Martonosi, and Ali Javadi- Abhari. Software Mitigation of Crosstalk on Noisy Quantum Computers. arXiv preprint arXiv:2001.02826 , 2020. 13

2108.10964

Ramin Ayanzadeh

Ramin Ayanzadeh, Poulami Das, Swamit S. Tannu and Moinuddin Qureshi

EQUAL: Improving the Fidelity of Quantum Annealers by Injecting Controlled Perturbations

quant-ph cs.AR cs.ET

http://creativecommons.org/licenses/by/4.0/

Quantum computing is an information processing paradigm that properties to speedup computationally hard promising, existing gate-based quantum computers consist of only a few dozen qubits and are not large enough for most applications. On the other hand, existing QAs with few thousand of qubits have the potential to solve optimization problems. QAs are single instruction machines and to execute a program, the problem is cast to a Hamiltonian, embedded on the hardware, and a single quantum machine instruction (QMI) is run. and imperfections in hardware result in sub-optimal solutions on QAs even if the QMI is run for thousands of trials. The limited programmability of QAs mean that the user executes the same QMI for all trials. This subjects all trials to a similar noise profile throughout the execution, resulting in a systematic bias. We observe that systematic bias leads to sub-optimal solutions and cannot be alleviated by executing more trials or using existing error-mitigation schemes. To address this challenge, we propose EQUAL (Ensemble Quantum Annealing). EQUAL generates an ensemble of QMIs by adding controlled perturbations to the program QMI. When executed on the QA, the ensemble of QMIs steers the program away from encountering the same bias during all trials and thus, improves the quality of solutions. Our evaluations using the 2041-qubit D-Wave QA show that EQUAL bridges the difference between the baseline and the ideal by an average of 14% (and up to 26%), without requiring any additional trials. EQUAL can be combined with existing error mitigation schemes to further bridge the difference between the baseline and ideal by an average of 55% (and up to 68%).

2 the acceptor impurities to get the correct description of both valence and conduction bands, due to improvement of compu- tational efforts, has been proposed recently25-27. Despite the fact that, the spin-orbit interactions can be considered negli- gible in most semiconducting materials, its inclusion helps to identify, correctly, the symmetry character of the top valence states and, thus, giving a correct description of the band states, as shown in the recent work done by Bastos and the semiconducting zincblende the improvement of k:pmodelling for wurtzite band structures, initial efforts in this direction were done by Marquardt et al.27. In their new k:pmodel, more bands were evaluated from the Density Functional Theory (DFT) results (16 ones in their new model, but 8 of them were located at the conduction band) than the earlier attempts to obtain However, the symmetry ordering for the bot- tom of the conduction band found in their DFT calculations is one of the possible configurations found in an earlier in- terpretation of resonant Raman experiments done on wurtzite GaAs28. Moreover, there is no mention if the other configu- ration could be theoretically possible to obtain, or which one leads to the correct description of the GaAs ground state, as well as the importance of the internal strains effects together with the small spin-orbit perturbation can affect the conduc- tion band states ordering. Then, the complete picture for the other semiconducting materials, which the wurtzite is an ac- cessible form, is not fulfilled yet. Despite the fact that the electronic structure of extensively studied in the past, we are going to show, in this paper, from a careful analysis of DFT results of the III- Nitrides' band structures, that internal strains effects together with spin orbit effects can change the energetic order of the top valence states and, despite the fact that spin-orbit interactions are negligible in these materials, these effects combined with the internal strains in the wurtzite structure helps us to under- standing the differences found for the p-type doping in AlN, when compared with GaN and InN. Together with a careful symmetry interpretation of bands, our obtained results will help on the improvements and refinements of which will be useful for the theoretical developments of future III-Nitrides based devices. II. THEORETICAL APPROACH A. Computational details The calculations involved in this work were done by using the DFT framework method within the gen- eralized gradient approximation (GGA) proposed and the hybrid by solve the Kohn-Sham equations, we have used the pro- jected augmented wave (PAW) method31, as implemented in the Vienna ab initio simulation package (V ASP, and the PAW projectors were provided by the V ASP package33. To describe the electronic states, the are expanded in plane waves employing a finite cutoff energy, which depends on the calculated properties. We start our work with several convergence tests and we demonstrated that well converged total energies, band struc- tures, and densities of states can be obtained by using a cutoff energy of 600 eV . Stress tensor and elastic constants calcula- tions, however, require a cutoff energy of 800 eV , since they converge slower than other properties. For the Brillouin zone integration, we employed a Monkhorst-Pack G-centered k-mesh of 11 116 to obtain the equilibrium volume, which was performed by several con- secutive optimizations of the equilibrium volume to ensure that the optimized equilibrium volume is consistent with the initial set up of the basis size. On the other hand, to calculate the elastic properties, we employed a k-mesh of 22 2212. B. Strain The hexagonal wurtzite crystal structure has the symmetry defined by the space group P63mcwhich is associated to the point group C6v. From the symmetry group analysis, it is pos- sible to show that it has five non-equivalent elastic constants, being they: applying the uniaxial stress-strain theory to the phase, we have the following conditions for the sxx=syy=0 where tis the stress. In order to describe strain from uniaxial stress, it is nec- essary to determinate the Young module. This parameter is defined here the elastic constants of the wurtzite structural phase. Finally, considering a linear deformation, we can express the deformation energy as: U=Y 2e2 zz (2) III. RESULTS A. Lattice Parameters The most stable crystalline phase for the majority is the zinc-blende structure. However, the III-nitrides (AlN, GaN and InN) prefer to crystallize in the hexagonal wurtzite structure, instead of the zinc-blende one3. The wurtzite structure is a hexagonal close packed lattice with four atoms per unit cell and to completely describe the crys- talline structure, one needs to use three different the lattice parameters a and c, and a parameter, u.3 TABLE I. AlN, GaN and InN structural parameters, compared to experimental measures and other DFT calculations. Parameters a and c are the lattice constants (A) and u is the cell-internal in the wurtzite phase. PBE Other works Experimental a 3.113 3.13a, 3.084b3.110d AlN c 5.013 5.09a, 4.948b4.980d c/a 1.610 1.618a, 1.604b1.601d u 0.386 0.382a, 0.3814b0.382d a 3.201 3.24a, 3.162c3.190d GaN c 5.199 5.24a, 5.142c5.189d c/a 1.624 1.614a, 1.626c1.626d u 0.379 0.384a, 0.377c0.377d a 3.617 3.59a, c 5.825 5.81a, 1.610 1.621a, 1.619b1.61e u 0.386 0.381a, 0.3784b aRef.34 bRef.35 cRef.36 dRef.37 eRef.38 To determine the lattice parameters we used for the ex- change potential the PBE functional and the results are sum- marized in the table I, along with other DFT calculations and experimental values. Starting with the AlN alloy, the calculation shows devia- tions from experimental values of 0.1%, 0.66%, 0.56% and 1.05% for a, c, c/a and u, respectively and are close to the parameters obtained by two other theoretical works using the PBE functional,34, and the LDA functional,35, GaN results show deviations of 0.34%, 0.19% 0.12% and 0.53%, and the comparison with other theoretical results also show results close to parameters determined with PBE34and LDA functionals36. Finally, our InN results show deviations of 2.28%, 2.12% and 0.12%. There is no experimental mea- sure for the u parameter in InN. Finally, the comparison with other theoretical works using PBE34and LDA shows reasonable agreement. B. Elastic constants As experimental setups using III-Nitrides often present strain due to the important difference of lattice when InN is involved20, the determination of elas- tic constants is a necessary step. Five constants have to be determined for these systems: the mod- ulus for the axial compression, i.e., the stress in one a strain in the same direction; stress that induces a strain in the perpendicular directions; and, finally, C44represents the strain across the faces induced by the stress in a direction parallel to it. We determined the elastic constants using two PBE and HSE06 and our results are with other theoretical results and experimental values, in Table II.The constants determined with the PBE an average variation from the literature II. Elastic constants for AlN, GaN and InN calculated with PBE and HSE06 functionals, compared with other theoretical calcu- lations and experimental values. All constants are given in GPa. PBE HSE06 Other 376.28 392.39 410.2 410.5 10b C12 128.63 144.65 142.4 148.5 10b AlN C13 98.23 112.46 110.1 98.9 3.5b C33 356.59 366.67 385.0 388.5 10b C44 112.45 123.82 122.9 124.6 4.5b C11 329.39 361.117 368.6 390c C12 127.83 161.509 131.6 145c GaN C13 95.04 125.058 95.7 106c C33 355.81 391.004 406.2 398c C44 86.06 99.766 101.7 105c C11 203.53 222.796 233.8 225d C12 102.12 127.715 110.0 109d InN C13 84.99 110.851 91.6 108d C33 203.43 215.656 238.3 265d C44 45.39 46.549 55.4 55d aRef.39 bRef.40 cRef.40 dRef.41 values of 8.07%, 13.27% and 15.57%, for AlN, GaN and The deviation of the same constants, when calcu- lated with the HSE06 functional, is reduced to 0.09%, 3.04% and 3.03%. When compared with other HSE06 results show similar deviations. Analysing these devia- tions, one can see that, although the PBE functional may pro- duce values closer to the experiment for C13, HSE06 calcula- tions clearly provide a better set of parameters. C. Band structure As pointed out in the literature, the PBE functional does not provide correctly electronic parameters of the the gap energy for example. In materials with very small gap, such as the InN, there is even an inversion of bands42. In order to reproduce realistic values for the band structure and the energy gap, we used the HSE aparametrization of the HSE functional43,44. In this approach, the mixing parameter of the PBE functional and the exact Fock exchange is defined as different from the usual 25% value, using a linear inter- polation with the PBE functional ( a=0) to approximate the experimental band gap. Even if the computational effort of this approach is even higher than the usual HSE06, since it was successfully used for zinc blend III-V systems26, it was our choice. We assumed values of a: 0.355 for AlN; 0.320 for GaN and 0.305 for InN. Figure 1 presents the band for these overall behaviour of the different systems is very simi- lar. The most characteristic features of these band structures is the direct band gap and the lack of band repulsion at the point A, where the bands of different symmetries cross and also the multiple crossing along the K-M direction, allowed by the symmetry analysis of the C64 vcrystal group. Apart from the4 -6-4-2 0 2 4 6 8 A GK M L Aa)Energy (eV) -8-6-4-2 0 2 4 6 A GK M L Ab)Energy (eV) -6-4-2 0 2 4 A GK M L Ac)Energy (eV) FIG. 1. Band structures: a) AlN; b) GaN; c) InN. overall behavior, very different features may be found when comparing the different band structures. First, there is an in- crease on the splitting energy among the first and second con- duction bands as the atomic number of the cation increase, from AlN to InN, and second, the top of the valence band shows very different behaviors as seen in Figure 2. Looking at the valence band, its most characteristic feature is that the top of the valence band is a fol- lowed by a G9and a A second very char- acteristic feature is the very small separation of the heavy-hole band, HH ( G9), band from the light hole band, LH ( Gk 7), usu- ally associated with the spin-orbit splitting. In wurtzite the third band is called crystal-field splitting hole band, CH ( G? 7). The symmetry breaking fields associated with these 0 A G Ka)Energy 0 A G Kb)Energy 0 A G Kc)Energy (eV)FIG. 2. Band structures close to Gpoint: a) AlN; b) GaN; c) InN. splittings are the spin-orbit field, usually depicted by the Dso term, and the field associated with the spatial break of sym- metry among the planar and perpendicular directions known as the crystalline field, depicted by the term Dcr. It is impor- tant to notice here that, unlike the zincblende (ZB) structures, the LH band is the spin-orbit split-off band and the CH has the combination of components usually associated with the light holes in ZB. Since the symmetry breaks of the crystalline and are non-collinear, we need to use following expression5 to define the relation of the band structure splittings at G,D1 andD2, with the associated parameters, III shows the energy gaps, and parameters for the crystalline and spin-orbit field splittings for the three As expected, since these calculations are done with the gap energies calculated for all systems have small deviations of 1% from the experimental data. As LDA and PBE functionals are known to underestimate the gap and the HSE aapproach adjusts it, the most effective compar- ison of these results with the literature, either theoretical or experimental, may be done comparing the crystal-field split- ting and spin-orbit coupling parameters, DcrandDso. TABLE III. Band gap ( Eg), Spin-orbit coupling (Dso)and (Dcr)compared with measurable parameters for AlN. All energies are given in eV . HSEa Other works Experimental Eg 6.182 5.8a6.11b, 6.2c, 6.28d, 6.1e AlN Dcr -0.271 0.024 0.019f, 0.058g0.013b Eg 3.513 3.474h, 3.455i, 3.5031k, GaN Dcr -0.022 0.042f0.022h, 0.010j, 0.025k, Dso 0.034 0.012e, 0.013f0.011h, 0.018j, 0.017k Eg 0.891 0.9l InN Dcr -0.035 0.041f Dso 0.046 0.001f bRef.47 aRef.48 cRef.49 dRef.50 eRef.51 fRef.52 gRef.53 hRef.54 iRef.55 jRef.56 kRef.45 lRef.38 In AlN both, the spin-orbit and the crystalline field param- eter values obtained in our simulation are close to the exper- imental results obtained by Li et al.48and close in the same range of other theoretical values with different addition, the inversion of bands represented by the negative signs shown at the table is present in both. In GaN and InN the signal of Dcris negative in contradiction with the theoretical work of Li et al.48, although the magnitudes are of the same order. The spin orbit values are consistently bigger in our cal- culations than the values tabulated in their paper, 24 meV to 19 meV in AlN, 34 meV to 13 meV in GaN and 46 meV to 1 meV in InN. Experimental values for AlN and GaN are in the same order of magnitude of the calculations. D. Valence band ordering and strain The common sense, from the modeling of Wurtzite II-VI systems, puts the ordering of the valence bands from top to down being the top of valence band in AlN is known to present an inversion for a long time48. Figure 3 presents a diagram of the valence band ordering of the III-nitrides whose symmetry were determined using the FIG. 3. Band ordering of III-nitrides valence bands in the irvsp57and IrRep, from the Bilbao the energy differences shown in scale. The top valence band in all 3 cases is the usually known as CH or crystal-split hole, ( G? 7), followed by the usually known as HH or heavy hole ( G9) and LH or light hole ( Gk 7). It is interesting to remark, when we consider the results of the presence of the acceptors in that the CH state is the aforementioned hole trap and, its relative position to the HH and LH ones explain why is difficult to have p-type AlN related to the GaN and InN as well. To compare the roles of the crystal and spin-orbit splittings we set the zero of the diagram in the G9(grey) and conse- quently, at G,G? 7(blue) has the value of the energy valued) and Gk 7(red) has the value of the energy separation valued). This or- dering leads to negative values for the crystal field parameter (Dcr). Recent calculations of other III-V alloys in the wurtzite phase show Dcrwith positive values46. To understand the ordering of the states it is useful to use a model based on the strain and on the differences of zincblende and wurtzite systems. The comparison of the WZ and ZB sys- tems is useful here since both of them are repetitions of a layer of sp3hybridized species and the only diference between them is a rotation of 60 degrees in the subsequent layers. While ZB has a repetition unit of one atom of each species, wurtzite has a unit of two atoms of each species, needing 2 to fully describe the structure along the [0001] di- rection, candu, shown in Table I. The parameter u is related to the distance of the separation of atoms of both species in the direction of the repetition and the ideal value, defined as if the distance from one species to other were the same, is 0.375. Assuming that in wurtzite, the divergence of the ideal and actual bonds is due to internal strain, one can determine an associated deformation energy associated with the crystal field splitting. Table IV presents the deformation energy (U) and the Young modulus (Y) calculated with the HSE06 functional. We determined the Young modulus according to equation 1 and with the elastic constants presented at table II. Our results are in good agreement with other works59-61, presenting de- viations of: 0.92% for AlN; 1.12% for GaN and 0.97% for6 TABLE IV . Calculated deformation energy U (meV), compared with eletronic parameters in eV , and Young modulus (Y) in GPa. AlN GaN InN Dso 0.023 0.031 0.042 Dcr 0.261 0.016 0.029 Y 319.570 331.154 145.542 U 177.333 13.764 81.061 InN. The deformation energy was calculated using the results from the Young Modulus and assuming the elongation as the the ratio. ezz=uideal=u. The values obtained for the deforma- tion energy are 177.3 meV , 13.7 meV and 81.1 meV , for AlN, GaN and InN, respectively. Since the spin-orbit and the strain field are not collinear, the comparison with the band structure is not straightforward. However if we use a composition of the states as a baseline, this comparison may be done. As- suming that all the states are equally important, we set the baseline as (EHH+ELH+ECF)=3. Using this value, the pre- dicted ECFis 250 meV , 14 meV and 78 meV , for AlN, GaN and InN, respectively. These values may be directly whose values, determined using HSE06, are 253 meV , 12 meV and 22 meV , for AlN, GaN and InN simple model, based on the uniaxial strain model and applying the ratio uidealu=u, relates the inversion of G9and G? 7bands, previously stated in AlN, and present also in GaN and InN alloys with the relaxation of the crystal, expressed by the differents u. The model provides, through the sum of the deformation energy with the estimated the baseline of the hole states, numerical values compara- ble to the D1parameter that fit extremely well for the AlN and GaN alloys. For InN however, the comparison, although valid, gives a lesser estimation. The InN alloy, however, have a higher spin-orbit splitting that can be estimated by the sepa- ration of the this case) and the estimation of the baseline for it is probably less accurate than in AlAs and GaAs. IV. CONCLUSIONS In summary, we presented our theoretical results for the structural, elastic and electronic properties of wurtzite III- Nitrides in the wurtzite crystalline structure, obtained from DFT calculations within the GGA/PBE, and the HSE ap- proaches for the the obtained structural parameters, our GGA/PBE results show good agreement with the experimen- tal and other theoretical data. For the elastic constants results, we have observed that, by using HSE06 functional there is a significant improvement of our obtained values, showing an excellent agreement with the available ex- perimental the band structures, by tuning ain the HSE functional, better values for the bandgap energies were obtained when compared to the experimen- tal data, keeping the same energetic order of the at both the top of valence band, as well as at the bot- tom of the conduction band, in comparison with the we have shown, from a careful analysis of our DFT results, that internal strains effects together with spin orbit ef- fects play a significant role in the energetic order of the top valence states. As an example, concerning to difficulties to ob- tain p-type AlN related to the GaN and InN, is directly related to the relative position of G? 7state (CH) compared with both theG9(HH) and Gk 7(LH) ones at the top of valence bands.In this case, the CH state behaves as the hole trap mentioned in previous theoretical works13,15. We hope that our results give guidance for future experiments on this acknowledges financial support from CNPq (grants No. 408916/2018-4 and 308806/2018-2) and CAPES (CsF- Grant No. H.W.L.A. is grateful for the financial support given by the FAPEMIG (grant No. CEX APQ 02695-14). A.H.S. acknowledges financial support from CAPES (grant No. Nakamura, S. Pearton, and G. Fasol, The Blue Laser Diode 2000). 2S. C. Jain, M. Willander, J. Narayan, and R. V . Overstraeten, J. Appl. Phys. 87, 965 (2000). 3B. Gil, ed., III-Nitride Semiconductors and their Modern Devices Press, 2013). 4N. Lu and I. Ferguson, Semicond. Sci. Technol. 28, 074023 (2013). 5J. Furthmuller, P. H. Hahn, F. Fuchs, and F. Bechstedt, Phys. Rev. B 72, 205106 (2005). 6P. Rinke, M. Scheffler, A. Qteish, M. Winkelnkemper, D. Bimberg, and J. Neugebauer, Appl. Phys. Lett. 89, 161919 (2006). 7V . V . Buniatyan and V . M. Aroutiounian, J. Phys. D Appl. Phys. 40, 6355 (2007). 8J. S. Kwak, O. H. Nam, and Y . Park, Appl. Phys. Lett. 80, 3554 (2002). 9P. Kozodoy, Y . P. Smorchkova, M. Hansen, H. Xing, S. P. DenBaars, . K. Mishra, A. W. Saxler, R. Perrin, and W. C. Mitchel, Appl. Phys. Lett. 75, 2444-2446 (1999). 10J. Simon, V . Protasenko, C. Lian, and D. Jena, Science 327, 60 (2010). 11Y . Arakawa, K. Ueno, A. Kobayashi, J. Ohta, and H. Fujioka, APL Mater. 4, 086103 (2016). 12H.-T. Chen, C.-Y . Su, C.-G. Tu, Y .-F. Yao, C.-H. Lin, Y .-R. Wu, Y .-W. Kiang, and C.-C. Yang, IEEE Trans. Electron Devices 64, 115 (2017). 13J. L. Lyons and C. G. V . de Walle, Phys. Status Solidi RRL x, Freysoldt, B. Grabowski, T. Hickel, and J. Neugebauer, Rev. Mod. Phys. 86, 253 (2014). 15J. L. Lyons, A. Alkauskas, A. Janotti, and C. G. V . de Walle, Phys. Status Solidi B 252, 900 (2015). 16F. Mireles and S. E. Ulloa, Phys. Rev. B 58, 3879 (1998). 17I. Vurgaftman and J. R. Meyer, J. Appl. Phys. 94, 3675 (2003). 18P. Rinke, M. Winkelnkemper, A. Qteish, D. Bimberg, J. Neugebauer, and M. Scheffler, Phys. Rev. B 77, 075202 (2008). 19Q. Yan, P. Rinke, A. Janotti, M. Scheffler, and C. G. V . de Walle, Phys. Rev. B 90, 125118 (2014). 20H. Morkoc, Handbook of Nitride Semiconductors and Devices, v. 1 (WILEY-VCH Verlag, 2008). 21B. Gil, Physics of Wurtzite Nitrides and Oxides (Springer International Pub- lishing, 2014). 22P. Kackell, B. Wenzien, and F. Bechstedt, Phys. Rev. B 50, 17037 (1994).

2108.10968

Guilherme Sipahi

F\'abio D. Bonani, Anderson H. Siqueira, Hor\'acio W. Leite Alves and Guilherme M. Sipahi

Strain and crystal field splitting inversion in III-Nitrides

7 pages, 3 figures and 4 tables

cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

The wurtzite phase group III-Nitrides (AlN, GaN, InN) have attracted great interest due to their successful applications in the optoelectronics since the 90's. In this paper we perform a comprehensive study of AlN, GaN and InN structural elastic and electronic properties using hybrid and Functional Theory, presenting a comparison of the features of the three compounds. We perform a direct comparison of the features of their including the inversion of the top valence band associated with a negative crystal field splitting and its relation to the challenges on AlN systems. With the determination of elastic constants and the Young modulus we provide a simple model to connect a deformation with the parameter $u$ and the effective crystal-field a direct relation among internal strain and the

2 optimised for Hdecaying exclusively into pair, with an integrated luminosity of 139 atps= 13 TeV LHC by the ATLAS col- laboration [75] has excluded them with masses up to 350 GeV and 230 GeV, respectively, for the pair and associ- ated production modes assuming vt= 0:1 GeV and the mixing between the CP-even scalars to be 104. Patently, the above-cited limits are not betting to the entire parameter space, rather valid only for a con- strained parameter space of the model. For instance, the CMS search in [71] is only valid for
m= 0 and vt<104GeV, whereas the ATLAS search in ref [75] is only valid for
m= 0 andvt>104GeV. Though in a realistic type-II see-saw scenario, the branching fractions of the triplet-like scalars into dierent lepton avours are dictated by the neutrino oscillation parameters, most of the aforecited limits are derived in the context of sim- plied scenarios without reckoning the footprints of the low-energy neutrino parameters. Furthermore, these lim- its are often conservative as these searches do not in- corporate all the production channels for the For instance, the ATLAS search in [75] consid- ered either pair or associated production modes for the doubly charged scalars, but not both at once. As ar- gued in Section III, all the channels entail to be incorpo- rated in the analyses. Moreover, the triplet components in this model are conceivably non-degenerate in mass. For moderate vtand passably large
m, cascade decays quickly dominate over the leptonic and diboson decay modes, see Section IV. Thus, in the non-degenerate sce- nario (
m6= 0), the cascade decays are entitled to play a notable role in the phenomenology, thereby making for the non-degenerate scenario substan- tially contrasting than that for the degenerate one. Bearing the aforesaid discussion in mind, we perform a systematic and comprehensive collider study of this model. Incorporating all the Drell-Yan production modes for the triplet-like scalars and taking into account the complexity of their decays, we derive the most stringent 95% CL lower limit on mHfor a wide range ofvtand
mby implementing already existing direct collider searches by CMS and ATLAS. Then, we forecast future limits on mHby extending the ATLAS search at high-luminsity, and we propose a search strat- egy that yields improved limits on mHfor a part of the parameter space of vtand
m. The rest of this work is structured as follows. In Sec- tion II, we brie y describe the theoretical structure of the type-II see-saw model. Production of the and their decays are discussed in Section III and IV, respectively. In Section V, we discuss the LHC phe- nomenology of this model and obtain stringent limits on mHfor a wide region of parameter space.II. THE TYPE-II SEE-SAW MODEL The scalar sector of the minimal type-II see-saw model employs aSU(2)Ltriplet scalar eld with hypercharge 2,
in addition to the SM Higgs doublet, :
=
+=p 2  = + 0 : The most general renormalizable and scalar potential invloving  and
is given by V(;
) =m2 y + [(Ti2
y) + h:c:] +1(y)Tr(
y
) the mass parameters, andi (i= 1;:::;4) are the independent dimensionless couplings. After the EWSB,  and
acquire the following VEVs h
i=0 0 vt=p 2 of the scalar potential V(;
) allows terms of the VEVs and the quartic and trilinear couplings the rst tadpole equation shortens (2.1) This relation withstands the so-called see-saw spirit. For small, which is justied by the 't Hooft's the triplet mass scale is naturally connected to the physics at the TeV scale (unlike in the original see-saw scenario where the new physics is naturally mo- tivated to be at very large scale), thereby making the present model potentially testable, and hence falsiable at the LHC. The neutral components of  and
can be parametrised as and After the EWSB, the mixing among the scalar elds lead to several Higgs bosons. The gauge ein- genstates can be rotated to obtain the mass eigenstates as in the R() the rotation angles in CP-even and CP-odd Higgs sectors withp 2 tan= tan0=2vt vdand tan 2=2B AC, the Nambu- Goldstone bosons, respectively, eaten by the longitudi- nal modes of WandZ.His the singly-charged Higgs boson with mass the CP-even neutral Higgs with respectively 4B2]: We discern h0as the 125-GeV resonance observed at the LHC. Though h0is not necessarily the lighter of these two eigenstates, we presume it to be the lighter one. A0 is the CP-odd neutral Higgs with mass the doubly-charged Higgs,
(H), mentoined above, m2
andm2 can be enunci- ated in terms of the other scalar potential parame- ters, namely These parameters, in turn, can be framed in terms of the physical the VEVs ( vt;vd) and the rotation angle of the CP-even Higgs sector ( ) 2v2 tm2 H+sin m2 h0+ cos2 m2 h0+ sin2 the masses of the physical Higgs states reduce their mass-squared dierences are given (2.2) For usefulness, we dene the mass-splitting
m=mHmH. Thereby, the masses of all the physical Higgs states can be traded in terms of just two parameters | mHand
m. In addition to the tree-level mass-splitting dictated by the dimensionless quartic coupling 4, dominantly driven by the bosons at one loop, induce mass-splittings among dierently charged triplet scalars: mHmH885 MeV and mHmH0=A0540 MeV [77]. These radiative mass-splittings exclusively are small enough to have considerable eects on the decays of the triplet-like scalars and thereby neglected in the rest of this work [14]. Depending on the value (sign) of 4, and hence
m, three characteristic mass spectra are predicted | (i) 4= (ii) 4>0:mH> mH> mH0=A0and (iii)4<0: mH< mH< mH0=A0. These characteristic mass spectra will be sometimes referred to as and negative scenario, respectively. In what follows, we brie y discuss the relevant constraints on Both the dou- blet and triplet VEVs contribute to the weak gauge bo- son's masses at tree 2v2 t) 4v2 t): The VEVs are subject to the 2v2 t vSM= 246 GeV. The parameter, dened as takes the form = (v2 d+ 2v2 t)=(v2 d+ 4v2 t): The value of the parameter from the data, = 1:00038(20) [78], which is 1 :9above the SM expectation at tree level, (vt= 0) = 1, leads to an upper bound of O(1) GeV on on
mfrom oblique parame- ters: As mentioned earlier, the the electroweak pre- cision data observables, also known as the namely S;T andUparameters. These parameters robustly constrain the mass-splitting
mto be withinj
mj.40 GeV [21, 34, 45, 79].4 Constraint on from Higgs data: The rotation angle in the CP-even Higgs sector ( ) is given by tan this reduces to tan further reduces to tan 125-GeV Higgs data, in particular, the decay rate at the LHC constrains : jsinj.0:3 at 95% CL [45]. This bound is consistent with the above expression of tan 2 for the allowed values ofvt. Neutrino masses: The Yukawa interaction which gen- erates neutrino masses is given h:c: ; (2.4) whereYis a 33 symmetric complex matrix, iandj are the generation indices ( i;j= 1;2;3),L= (L;`L)Tis the left-handed SM lepton doublet and Cis the matrix. After the scalar triplet acquires a VEV, the following neutrino mass matrix is be diagonalised using the matrix usually parametrised by three mixing angles ( 12;23and13), one dirac phase ( ) and two Majorana phases ( and0) = 2 41 0 1 0 cosij(sinij). We simply set ; and0to be zero throughout this work unless stated otherwise as they are either poorly measured or hitherto not measured at all. For normal hierarchy and for inverted hierarchy (IH), m1=p j
m2 32+ where
m2 ij's are the experimen- tally measured mass-squared dierences, and m1(m3) is the lightest neutrino mass for NH(IH). Measurements of large scale structure in the universe by the Planck satellite has put a boundP imi<0:12 eV when combined with baryon acoustic oscillation data [80]. The best t values for the neutrino oscillation parameters used in this work are taken from Ref. [81]. Constraint from lepton avour violating decays: The Yukawa interaction of the scalar triplet with the SMlepton doublet leads to lepton avour violating decays such as`!` at 1-loop and `!`` `at tree level. The relevant branching fractions are given by the electromagnetic ne-structure constant andGFis the Fermi constant. The upper limits of 4 :2 1013[85] and 1:01012[86], respectively, on Br( ! e ) and Br(!e+ee) tightly constrain the space. These limits translates into a lower limit GeV1091 TeV mH for NH(IH) with mi's consistent with the bound relevant parameters: this section, we brie y reckon the relevant parameters. As mentioned earlier, the Yukawa couplings are determined by the neutrino oscil- lation parameters up to vt1, and all the scalar can be framed in terms of the masses of the physical Higgs states, the VEVs and the mixing angle between the CP-even scalars. The mixing angle is deter- mined in terms of the others by Eq. (2.3). Moreover, the masses can be traded in terms of just two parameters | mHand
m. Therefore, the phenomenology of this model, by and large, is governed by three parameters only |mH,
mandvt. III. PRODUCTION OF TRIPLET SCALARS The TeV scale triplet-like scalar bosons are pair pro- duced copiously at the LHC by quark-antiquark annihi- lation via the neutral current and charged current Drell- Yan mechanism, respectively, through of the neutrino oscillation parameters, namely the mass and the CP phases, are either poorly measured or hitherto not measured at all. In this work, we set the phases to be zero for simplicity. However, note that these parameters could substantially change the leptonic decays and thereby of the charged scalars. 2Also, the triplet-like scalars are produced via fusion process [35, 87] and vector [7, 26, 88] with two associated forward jets at the LHC. However, their production through such processes is sub- dominant for the mass range of our interest, and thereby implement the model in SARAH [89, 90] to generate UFO modules, and use MadGraph5aMC@NLO [91, 92] with the NNPDF23 loas0130 qed parton [93, 94] for numerical evaluation of the leading order (LO) production cross-sections of the triplet scalars at the 13 TeV LHC. The left, middle and right plots in Figure 1 shows dierent pair and associated production LO as a function of the doubly charged scalar mass mH, respectively, for degenerate scenario ( i.e.
m= 0), positive scenario with
m= 30 GeV and negativescenario with
m=30 GeV. All the Drell-Yan pro- duction mechanisms are of handsome cross-sections; in particular, production of the doubly charged scalars in association with the singly charged ones, which is some- times precluded by experimental searches, has the for both degenerate and negative scenar- ios. Production of the singly charged scalars in associ- ation with the neutral ones, which is also forsaken by both CMS and ATLAS, has the largest cross-section for both degenerate and positive scenarios. This substanti- ates that all the channels entail to be incorporated into the 200 400 600 800 1000 1200 1400 160013 TeV LHCS0 H0, A0 Degenerate scenario (Dm = 0)cross section [fb] mH+-+- [GeV]H+-+-H, H+-S0 H++H--, 200 400 600 800 1000 1200 1400 160013 TeV LHCS0 H0, A0 Positive scenario (Dm = 30)cross section [fb] mH+-+- 200 400 600 800 1000 1200 1400 160013 TeV LHCS0 H0, A0 Negative scenario (Dm = -30)cross section [fb] mH+-+- 1. Dierent pair and associated production cross-sections of the triplet-like scalars for degenerate scenario (left), with
m= 30 GeV (middle) and negative scenario with
m=30 GeV (right). Refs. [95, 96] have estimated the QCD corrections to the pair and associated productions of the doubly charged scalars at hadron colliders which result in a next- to-leading (NLO) K-factor of 1.2{1.3. Considering that the QCD corrections to the pair and associated produc- tions of the singly charged scalars are, in principle, sim- ilar to those of the doubly charged ones, we apply an overall QCD K-factor of 1:25 to the LO DECAYS OF THE SCALAR BOSONS In this section, we study the decays of the scalar bosons, namely doubly-charged scalars (H), singly-charged scalars ( H), CP-even and CP- odd neutral scalars ( H0andA0). The partial decay widths of the scalars can be found in the literature [14, 18, 97{100]. However, we found few typos in some of those analytical expressions; for the sake of completeness, we provide the same in Appendix A.A. Decays of doubly charged scalar bosons The doubly charged scalar bosons have three possible decay modes | (i) leptonic decay, i.e.``, (ii) gauge boson decay, i.e.WWand (iii) cascade decay, i.e. HW. The latter decay mode kinematically opens up only for
m > 0. Form2 Hm2 W, the ratio of the braching fractions for these decay modes can be esti- mated as (see Appendix
m <O(1) GeV,Hdecays into ``and/or WW. These two decay modes are on a par with each other for vtO(104) GeV, and the former de- cay mode subjugates the other one for vt<104GeV and vice versa. One expects the cascade channel to kick o for
m&O(1) GeV, and become prepotent for larger mass-splitting. For mH= 500 GeV, the cascade decay dominates over the gauge boson one for
m > 55vt 1GeV
2=5, whereas it dominates over the leptonic one for
m> NH neutrino6 mass spectrum with m1= 0:03 eV is assumed.3This has been re ected in the Hdecay phase diagram shown in the left most panel of Figure 2. Figure 2 shows decay phase diagrams of and H0=A0 (right) with mH= 500 GeV segregating the vt{
mparameter space with dierent decay modes' dominance for NH neutrino mass spectrum with m1= 0:03 eV. blue dashed and solid contours, re- spectively, correspond to 99 and 90%, and Black solid contours to 50% branching ratios in dierent decay Decay Leptonic Decay Gauge (GeV)Cascade Decay Leptonic Decay Leptonic 2. Decay phase diagram of and H0=A0(right) with mH= 500 GeV segregating space with dierent decay modes' dominance for NH neutrino mass spectrum with m1= 0:03 eV. blue dashed and solid contours, respectively, correspond to 99 and 90%, and Black solid contours to 50% branching ratios in dierent decay regions. B. Decays of singly charged scalar bosons The singly charged scalar bosons have four types of decay modes | (i) leptonic decay, i.e.`, (ii) hadronic decay, i.e.tb, (iii) diboson decay, cascade decay, These two cascade decay modes kinematically open up, respec- tively, for
m > 0 and
m < 0. ratio of the braching fractions for the modes can be evaluated as (see Appendix the top quark mass. For suciently small mass-splitting, i.e.j
mj<O(1) GeV, Hdecays into The leptonic decay mode dominates for vt<104GeV, 3For NH , 0.03 eV is the maximum possible value for the mass consistent with the bound from the hadronic and diboson decay modes domi- nate forvt>104GeV. Furthermore, the diboson decay modes subjugate the hadronic one for versa. The cascade channels kick o for large i.e.j
mj&O(1) GeV, and quickly dom- inate for larger mass-splitting. Dominance of these dif- ferent decay modes in dierent regions of space has been shown in the Hdecay phase diagram in the middle panel of Figure 2. C. Decays of heavy neutral scalar bosons The CP-odd (CP-even) heavy neutral scalar boson has four types of decay modes | (i) leptonic decay, i.e., (ii) hadronic decay, (iii) diboson decay, i.e.h0Z(WW ,ZZ,h0h0) and (iii) cascade The latter decay mode kinematically opens up only for
m < 0. Form2 H0m2 W, the ratio of the braching fractions for the ,qq,h0h0,WW modes of H0can be estimated as (see Appendix the ratio of the braching fractions for the modes of A0can be <O(1) GeV,A0(H0) decays into neutrinos and hadrons/dibosons respectively, for Furthermore, the dibosons decay modes ofA0(H0) dominate over the hadronic one well as for channel kicks o for
m&O(1) GeV, and swiftly dominates for larger mass-splitting as can be seen from theH0=A0decay phase diagram in the right panel of Figure 2. V. COLLIDER SEARCHES Profuse Drell-Yan production of the triplet-like andWexchanges and their subse- quent prompt decays5to SM particles lead to a variety of nal state signatures at the LHC. Possible nal states includes smoking gun signatures like two pairs of same- sign lepton or two pairs of same-sign W-boson. consequence of the present model at the LHC has been studied extensively in the literature [7{ 47]. Rightfully, central attention of most of those stud- ies pivots around the doubly charged scalars because of their distinct decay signatures. For the very same reason, both the CMS and ATLAS collaborations have carried out a number of collider searches to look for the same at the LHC [66{75]. Hitherto no signicant excess over the SM background expectations has been observed in any of these direct collider seraches. These seraches thereupon have set stringent limits with 95% CL on the masses of the doubly charged scalars in the context of a simplied model. As argued in Section I, these limits are not bet- ting to the entire parameter space, rather valid only for a constrained parameter space of the model. Also, these limits are often conservative as these searches do not in- corporate all the production channels for the Furthermore, most of these limits are derived the decay of A0=H0intottis allowed; thus, the diboson decay modes are to WW. For a region of vt{mHparameter space, H could be long-lived and have displaced vertex signatures at col- lider [42].the context of simplied scenarios without reckoning the footprints of the low-energy neutrino parameters. The quartic scalar interaction of the triplet with the doublet, 4y

y, countenances the to split in mass. For moderate vtand passably large
m, cascade decays quickly domi- nate over the leptonic and diboson decay modes, see Figure 2. Conceivably, the cascade decays play notable role in the pheneomenology, thereby making the pheneomenology for the non-degenerate contrasting than that for the degenerate one. Thus, in the non-degenerate scenario (
m6= 0), the cascade decays are entitled to play a pronounced part in constraining the model parameter space. Not only the mass-splitting overwhelms the decays of the triplet-like scalars, but it also aects their at the LHC, see Figure 1. In what follows, we brie y discuss possible nal state signatures for dif- ferent parts of the parameter space and outline already existing direct collider searches by CMS and ATLAS which are potentially sensitive in constraining the scenario (
m= 0)All the Drell-Yan pair production mechanism of triplet-like scalars except H+Hare of sizeable cross-sections, see the left most plot in Figure 1. For small triplet vev, respectively. Production of to, respectively, three and four light leptons ( e;) in the nal state. Though large enough cross-sections, they fall through to complement the multilepton nal state because of their invisible de- cays. The already existing multilepton searches by CMS and ATLAS such as [71, 72, 101, 102] are expected to constrain this part of the parameter space. For large triplet vev, respectively. All the production channels give rise multiboson nal state leading to Therefore, one anticipates this part of the parameter space to be probed by the existing ATLAS searches [74, 75]. Negative scenario (
m< 0)For small this scenario resembles the degenerate one. However, for passably large mass-splitting and moderatevt, the cascade decays over the other decay modes, thereby enhancing the eective for H. Then, depending on the to either same-sign dilepton pair pair, both leading to in the nal state. Thus, this scenario also can be probed either by the multilepton searches like [71, 72, 101, 102] or multiboson leading to like [74, 75]. Positive scenario (
m > 0)Again, this scenario re-8 sembles the degenerate one for small
m. However, for large enough
mand moderate vt, the cascade over the other decay modes. This enhances the eective pro- duction cross-section for H0andA0. respec- tively. This gives rise to multiboson nal state signa- tures. Therefore, one expects that this part of the param- eter space can probed by the existing ATLAS searches [74, both H0andA0decay invisi- bly into neutrinos. The relevant production soft leptons or jets result- ing from the o-shell W-bosons and neutrinos. Being very soft, these nal state leptons/jets are very dicult to reconstruct at the LHC. Therefore, in this scenario, the most optimistic nal states would be an energetic jet resulting from initial state radiation plus large momentum [103{105] or two/three soft leptons plus missing transverse momentum [106, 107]. For the former nal state signature, the recent monojet search by ATLAS [103] could be potentially sensitive in probing this scenario. Whereas, for the latter, one would expect the soft leptons searches by CMS [106, 107] to be sensi- tive in probing this scenario.7 A. Multilepton nal states search by CMS [101] The CMS collaboration has published a multilepton nal states search [101] with an integrated luminosity of 137:1 fb1of pp collisions atps= 13 TeV. This search targeted the triplet fermions in the type-III see- saw model [108]. However, because of similar multilep- ton nal state signatures, this search is conjectured to be passably sensitive in probing the type-II see-saw model. Hitherto, there is no multilepton search targeting the type-II see-saw model using the full Run-2 dataset by 6We see that for large enough
the existing collider searches [74, 75] fall short in probing the triplet- like scalars. For this part of the parameter space, toh0h0andZZ, andA0decays toh0Z. The leptonic decays ofZgive rise to multilepton nal states. However, the usual searches [74, 75] use a Z-veto to suppress the from the Drell-Yan processes making these in probing this scenario. On the contrary, for hadronic decays ofh0andZ, the signal cross-section is small compared to the overwhelming QCD jets background, and thus probing this scenario is very it turns out that both the monojet search by ATLAS and the soft leptons search by CMS cease to constrain this part of the parameter space. Monojet search usually requires a larger signal cross-section to suppress the huge SM background and hence, by and large, applicable to the strongly produced into soft leptons/jets plus missing particles. Further- more, the soft lepton nal states are suppressed by fractions. That said, inconsiderably small signal cross- section in the present model compared to the SM background makes such a nightmare scenario impossible to probe.CMS or ATLAS. Thereupon, we set forth to implement this search meticulously. We simulate the signal events using MadGraph [91, 92] with the NNPDF23 loas0130 qed parton [93, 94]. The subsequent decays, initial state radiation (ISR), nal state radiation (FSR), shower- ing, fragmentation and hadronisation are simulated with PYTHIA [109]. Hadronized events are passed into Delphes [110] for object reconstruction and signal regions and event selection. In doing so, we rigorously follow the search strategy in Ref. [101]. Lastly, we use a hypothesis tester which uses a library of C++ classes `RooFit' [111] in the ROOT environment to estimate CL. The selected events are categorised into several mu- tually exclusive signal regions (SRs), namely 4LOSSF0, 4LOSSF1 and 4LOSSF2 , based on the multiplicity of light leptons, the multiplicity and mass of opposite-sign same- avour (OSSF) lepton pairs, NOSSF andMOSSF . The names of the SRs are see Ref. [101] for details. The events in the 3LOSSF1 SR are further classied as 3L below-Z, 3L on-Z and3L above-Z whenMOSSF is below, within and above the Z-boson mass window ( MZ15), respectively. All the SRs are further divided into several independent sig- nal bins using a primary kinematic discriminant, in 40 signal bins in total. For 3L on-Z SR, this search uses transverse mass ( MT)8as the whereas for all other SRs, scalar sum of the transverse momenta of all charged leptons ( LT) plus the missing transverse momentum ( pmiss T) is used. These variables, exploiting the relatively high momenta of the triplet fermions' decay products, are found to be nifty the signal from the background. For the detailed description and validation of our this search [101], see Ref. [112, 113]. This enables us to use the distributions of ex- pected SM backgrounds and observed events in Ref. [101] to constrain the type-II see-saw scalars in various vt{
mregions. Figure 3 shows the of the expected SM background events (histograms with black line),9the observed events (big black dots) and the expected signal events corresponding to luminosity data at the 13 TeV LHC for three SRs | 3L above-Z (left), 3OSSF0 (middle) and 4OSSF0 (right). For brevity, we avert to show for the other SRs. The magenta dotted, dark yellow solid and blue dashed histrograms show the expected signal events for three benchmark masses listed 8The transverse mass is dened as MT pmiss T;~ p` T)];where~ p` Tis the trans- verse momentum vector of the lepton which is not a part of the on-Z pair, and
~ pmiss T;~ p` Tis the azimuthal separation between ~ pmiss Tand~ p` T. 9The gray bands represent the total (systematic + on the expected SM background.9 in Table I for vt= 108GeV10and
m= 0 assuming NH neutrino mass spectrum with m1= 0:03 eV. Table I shows three benchmark masses along with their from the afore-described CMS [101]. This shows that BP1 andBP2 are ex- cluded with more than 95% CL signicances, whereas BP3 is / 200 GeV=13 TeV)s (-1137 fb3L Above-Z CMS Obs. Data BP1 (600) Background BP2 (800) Uncertainty BP3 (1000) / 200 GeV=13 TeV)s (-1137 fb3L OSSF0 CMS Obs. Data BP1 (600) Background BP2 (800) Uncertainty BP3 (1000) / 300 GeV=13 TeV)s (-1137 fb4L OSSF0 CMS Obs. Data BP1 (600) Background BP2 (800) Uncertainty BP3 (1000) FIG. of the expected SM background events, the observed events and the expected signal events for3L above-Z (left), 3OSSF0 (middle) and 4OSSF0 (right) SRs. The signal predictions are shown for three benchmark points listed in Table I. Benchmark mHExclusion signicance BP1 600 11.9 BP2 800 4.32 BP3 1000 1.24 TABLE I. Benchmark masses and their exclusion signi- cances from the CMS multilepton search[101]. B. Multiboson leading to multilepton nal states search by ATLAS [75] The ATLAS collaboration has recently published a search for doubly and singly charged Higgs bosons de- caying into vector bosons in multilepton nal states with an integrated luminosity of 139 fb1of pp collisions atps= 13 TeV [75]. As mentioned earlier, this ATLAS search considered either pair or associated production modes for the doubly charged scalars, but not both at once. Also, this search does not incorporate the produc- tion channels for the singly charged and neutral triplet- like scalars. Thus, the limits set by this ATLAS search are conservative. Furthermore, these limits are not be- tting to the entire parameter space, rather valid only for
m= 0 andvt>O(104) GeV. Therefore, we set forth to implement this search incorporating all the pro- duction modes for the triplet-like scalars to constrain the 10The CMS multilepton search in Ref. [101] is designed to probe nal states with hard- pTleptons, and hence sensitive to small vtregion where the scalars directly decay to leptons, and results into hard signal leptons in the nal scalars in various vt{
mregions. For we minutely follow the search strategy in Ref. [75]. After object reconstruction and selection (see Ref. [75]), the events are categorised into three analysis channels, namely same-sign dilepton (2`sc), three leptons (3 `) and four leptons (4 `) chan- nels. The event selection proceeds in two steps | the preselection and the SRs selection. The are based on a number of variables such as the absolute value of the sum of charges of the leptons, their transverse momenta, pmiss T, the jet multiplicity, the b-jet multiplicity, MOSSF,etc. Four signal regions ( SR1, SR2, SR3 and SR4) are dened for each channel. For dening SRs, several other variables such as the invariant mass of all selected leptons, the invariant mass of all the jets, the distance between two same-sign leptons in the-plane, the azimuthal distance between the dilepton system and pmiss T, the smallest distance between any lepton and its closest jet in the -plane, etc. are used (see Ref. [75] for details).11These dierent sets of kinematic variables and selection cuts are used to dene the SRs ( SR1, SR2, SR3 andSR4), they are not mutually exclusive. They are designed by optimising the sensitivity for the Hpair production mode , themH= 200, 300, 400 and 500 GeV mass hypothe- ses. Ref. [75] considers for GeV mass hypothesis. Here, we dier from Ref. [75]; for a given mass hypothesis, we conider all the SRs disjointly, and eventually, choose the most sensitive one. How- ever, for the validation of our implementation, we adhere to the Ref. [75]'s capital out of the boosted decay topology of the triplet-like Higgs bosons as well as the high energy of their decay products, are discerned to the signal and the background. Events in the 2`scSRs are further divided into ee; whereas those in the 3 `SRs are separated into two categories (3 `1 and 3`0) based on whether or not an OSSF lepton pair exists in the event. This ameliorates the sensitivity of this search by lepton- avour composition between signal then, move forward to validate our implementa- tion of this search by reproducing the ATLAS 95% CL bound on the total pair production cross-section times branching fraction for two scenarios corresponding to the pair production HHand the associated Ref. [75]. The left (right) plot in Figure 4 shows the ATLAS observed and expected 95% CL upper limits on the HH(HH) produc- tion cross-section times branching fraction. The green and yellow bands represent the expected exclusion curves within one and two standard deviations, NLO QCD corrected [95] theoretical prediction is shown by the solid red curve. The reproduced 95% CL upper limit is represented by the blue dashed curve. The reproduced result is found to be in passable conformity with the ATLAS one, thereby validating our implemen- tation of this search. Yet again, this entitles us to usethe distributions of expected SM backgrounds and ob- served events in Ref. [75]. Figure 5 shows the expected SM background with black events (big black dots) and the expected signal events corresponding to 139 fb1integrated luminosity data at the 13 TeV LHC for four dierent SRs | SR1, SR2, SR3 andSR4. For each SR, the yields are shown for all the relevant channels, namely ee,e,, 3`0, 3`1 and 4`. The magenta dotted, dark yellow solid and blue dashed histograms show the expected signal events for three benchmark masses listed in Table II for vt= 1 GeV and
m= 0 assuming NH neutrino mass spectrum with m1= 0:03 eV. Table II shows three benchmark masses Benchmark mHExclusion signicance BP1 200 13.6 BP2 350 2.72 BP3 500 0.54 TABLE II. Benchmark masses and their exclusion signi- cances from the ATLAS multiboson leading into with their exclusion signicances from the above- described multiboson leading into multilepton search by ATLAS [75]. This shows that BP1 andBP2 are ex- cluded with more than 95% CL signicances, whereas BP3 is allowed. 0 20 40 60 200 250 300 350 400 450 500 550 600s=13 TeV , 139 fb-1s x Br(H+-+- H - W+-W+-WW) [fb] mH+- +- [GeV]95% expected 68% (NLO QCD) 0 20 40 60 80 100 120 200 250 300 350 400 450 500 550 600s=13 TeV , 139 fb-1s x Br(H+-+- H - W+-W+-WZ) [fb] mH+- +- [GeV]95% expected 68% (NLO QCD) FIG. 4. Left (right): The ATLAS observed and expected 95% CL upper limits on the HH(HH) times branching fraction. The reproduced 95% CL upper limit is represented by the blue dashed curve. 12The gray bands represent the total (systematic + on the expected SM background.C. 95% CL lower limit on mH In view of the observations being consistent with the SM background expectations, it is tempting to derive lim- its onmHusing the LHC searches. In what follows, we present stringent limits with 95% CL on mHfor a wide11 FIG. 5. The expected SM background events, the observed events and the expected signal events for four dierent SRs. For each SR, the yields are shown for all the channels. The signal predictions are shown for three benchmark points listed in Table II. range ofvtand
m. In deriving the present limits, we use the above-described CMS multilepton and ATLAS multiboson leading to multilepton searches. The left plot in Figure 6 shows 95% CL lower limits onmHas a function of vtfor
m= 0 assuming NH neutrino mass spectrum with m1= 0:03 eV. The khaki shaded region (on the left) is excluded from the param- eter measurement from the electroweak precision data, whereas the coral shaded region (on the right) is excluded from the lepton avour violating decay constraints. The dark goldenrod and pink shaded regions are from the CMS multilepton and the ATLAS multiboson leading to multilepton searches. For small vt, the triplet-like scalars with masses below 950 GeV are excluded from the CMS search. This exclusion limit is beyond those from the previous LHC searches [71, 72] by approximately 200{230 GeV. For large vt, the above- described ATLAS search excludes the triplet-like scalarsup to 400 GeV masses which is beyond the ATLAS limit in Ref. [75] by approximately 50 GeV. Given that the CMS and ATLAS searches are , it is reasonable to combine them. That said, as these two searches are primarily designed to target dierent regions in the parameter space, we expect only marginal improvement on the limits while combining them. The purple shaded region shows excluded parameter space when these two searches are combined. We nd that, as expected, improvement on the limit is marginal. The middle (right) plot shows excluded regions from the CMS and ATLAS combined search for
m= 10 and 30 (10 and30) GeV. For very large/small vt, the limits for
m=10;30 GeV are similar to those for
m= 0 case. This is because the cascade decays are yet to kick o for very large/small vt. This makes the non-degenerate sce- nario identical to the degenerate one. For moderate vt12 and large enough
m, the cascade decays kick-o and swiftly dominates over the other decay modes (see Fig- ure 2). In positive scenario, HandHdecay into o-shellWs (which gives rise to soft jets/leptons) and H0=A0(which further decays invisibly into neutrinos on the value of vt). into neutrinos, there are hardly visi- ble objects in the nal state, so much as the monojet search by ATLAS [103] and the soft leptons search by CMS [106, 107] fall short in constraining this part of the parameter space, see the middle plot in Figure 6. On the contrary, for H0=A0decaying into h0h0;ZZ=h0Z, the sig- nal cross-section is small compared to the from either QCD jets or Drell-Yan processes. This makes such a scenario challenging to probe. Note that GeV, the ATLAS search manages to put some bounds in the
m= 30 case, but it fails in the
m= 10 case. This is because for larger
m, some of the leptons from the o-shell W's passthe object reconstruction and selection criteria to con- tribute to the signal yields, whereas the leptons are too soft to do so for smaller
m. As one approaches towards smallvt, the leptonic decays retrieve their dominance over the cascade one, and give rise to multilepton nal states; this occurs at vtO(106) andO(107) for
m= 10 and 30 GeV. This has been re ected in the middle plot. On the contrary, in into o-shell W's and H, thereby enhancing the eective production cross- section forH. Therefore, in such a scenario, the limit gets enhanced compared to the degenerate case, see the right plot in Figure 6. For
m=10(30) GeV, the exclusion limit extends up to 1115(1076) GeV compared to 955 GeV for
m= 0. Note that for a given lighter in the
m=10 GeV case compared to those in the
m=30 GeV case. Thus, the signal cross-section is larger for
m=10 GeV than for
m=30 GeV. This explains the stronger limits for
m=10 GeV than
m=30 GeV. 200 300 400 500 600 700 800 900 1000 10 = 0 ATLAS + CMS CMS ATLAS95 % CL lower limit on m H+-+- [GeV] vt 200 300 400 500 600 700 800 900 1000 10 from lepton flavour violating decaysExcluded from measurement of r parameter 200 300 400 500 600 700 800 900 1000 10 + CMS Dm = 10 Dm = 3095 % CL lower limit on m H+-+- [GeV] vt 200 300 400 500 600 700 800 900 1000 10 from lepton flavour violating decaysExcluded from measurement of r parameter 200 300 400 500 600 700 800 900 1000 1100 10 + CMS Dm = -10 Dm = -3095 % CL lower limit on m H+-+- [GeV] vt 200 300 400 500 600 700 800 900 1000 1100 10 from lepton flavour violating decaysExcluded from measurement of r parameter FIG. 6. Left: 95% CL lower limits on mHas a function of vtfor
m= 0 assuming NH with m1= 0:03 eV. The shaded regions are excluded from dierent searches at the 13 TeV LHC. Middle(Right): Excluded regions from the CMS and ATLAS combined search for
m= 10 and 30 ( 10 and 30) GeV. D. Proposed multilepton nal states search for smallvt For smallvtand
m= 0, the triplet-like scalars up to 950 GeV masses are excluded from the CMS with 139 fb1of data, see Figure 6. Given the inappreciable signal cross-section for mH>1 TeV and comparatively sizeable backgrounds in CMS multilepton search, a similar search at high-luminosity is prophesied not to be sensitive enough in probing the triplet-like scalars much heavier than 1 TeV. Also, note that the said CMS search, which results in the most stringent limits in the small vtregion, is not designed to probe the triplet-like scalars. That said, to probe them, one may envisage a such that the background contributions cease to survive. In this section, we delineate a that is optimised to probe the triplet-like scalars much heavier than 1 TeV in the small vtregion. In what follows, we give a brief description about selection of various objects (jets, leptons, etc.), event selection and classication of selected events into mutually exclusive signal regions (SRs) for our nal states search. Object reconstruction and selection: Dierent physics objects, viz. jets, electrons, muons and miss- ing transverse energy, are reconstructed in Delphes [110]. Jets are reconstructed using the anti-kT algorithm [114] with a distance parameter
R= 0:4 as implemented in the FastJet package[115]. Reconstructed jets are required to have transverse momentum pT>30 GeV within the central pseudorapidity range jj<2:5. Electron with pT>10 GeV andjj<2:5(2:4) are con-13 sidered for further analysis. For the electron barrel (endcap), we demand a maximum relative isolation with
R= 0:4, 13whereas we demand a maximum 15% relative isola- tion with
R= 0:4 for muons. In addition, the fol- lowing set of lepton displacement requirements on the transverse and longitudinal impact parameters, dzand dxy, with respect to the primary vertex are enforced. For electron candidate within barrel (endcap), we demand dz<1 mm anddxy<0:5 mm (dz<2 mm anddxy<1 mm), whereas muon candidates require dz<1 mm and dxy<0:5 mm. Lepton isolation which trims inside the isolation cone along with impact pa- rameter requirements come in handy in vanquishing the reducible backgrounds such as Z+jets andtt+jets, where a jet is misidentied as lepton or additional leptons origi- nate from heavy quark decays. Finally, the missing trans- verse momentum pmiss Tis computed as the magnitude of the negative vector sum of the transverse momenta of particle- ow objects in an event. Overlaps between reconstructed objects resulting in ambiguities among them lead to object double counting. To sidestep that, all selected jets within a cone of
R< 0:4 of a selected lepton are thrown away. In addi- tion, all selected electrons within a cone of
R< 0:05 of a selected muon are discarded as these are likely due interactions of the muon with the inner detector material. Some of the jets, especially those on the tail of the detector response, and single pions could mimic lepton signatures and could be misidentied as leptons. Though the composition of the diers substantially among the without going into the intricacy of modelling the fake-lepton contributions, we the probability of 0.1{0.3% [116] for a jet to be misidentied as a lepton. Furthermore, of the electrons with the inner could lead to charge The radiated photon converts to ee+pair near the primary electron trajectory leading to charge Also, the photon could traverse the inner detector without creating any track. In such a case, the electron usually has a short lever arm on its curva- ture. This could lead to incorrect determination of the electron charge. We adopt the charge from Ref. [117]: P(pT;) is found to be 0.02{0.1 and f() is found to be 0.03{1 such that P(pT;) ranges from 0.02% to 10%. Note that the high- pTelectrons are more likely to be aected by charge misidentication as they have almost straight tracks, thereby making the curvature measure- ment very challenging. Also, the electrons with larger  13The relative isolation is dened as the scalar pTsum, normal- ized to the lepton pT, of photons and hadrons within a cone of
Raround the lepton. For electrons, this is required to be smaller than 0 within barrel (endcap) i.e.,jj<1:479 (jj>1:479) with
R= 0:3.have a larger misidentication probability as they tra- verse through a higher amount of inner detector material. Event selection and signal region with three or more light leptons are considered for this search. Events containing a lepton pair with
R < 0:4 or a same- avour lepton pair with invariant mass below 12 GeV are vetoed. This subdues back- ground contributions from nal-state radiations as well as low-mass resonances | Drell-Yan processes and neutral mesons. Furthermore, events containing a pair with an invariant mass within the nominal Z-boson mass window, i.e.MZ15 GeV are suppresses background contributions from process as well as the WZ pro- duction. Events with exactly three light leptons (3L) in one category and four or more light leptons (4L) in an- other category are considered for further analysis. Noting that the triplet-like scalars, which are to be probed, are heavier than 1 TeV, we persuade to exploit the relatively high momenta of their decay products. Before continuing, let us brie y reckon the processes contributing to the 3L and 4L signal events. For 3L events, the dominant contribution arises either from or from with Therefore, the invariant mass distribution of the same-sign lepton pair is expected to peak at mH. One would expect high- pTleptons, no high- pTjet (except for those coming from ISR and FSR) in the nal states for the former. For the latter, one would expect nal states with high- at least one handsome- pTjet. Then, the dominant contribution to the 4L signal events comes from The invariant mass distributions of both the same-sign lepton pairs are expected to peak at mH. Once again, one would expect high- pTleptons and small pmiss Tin the nal states. SM Backgrounds: A number of SM processes which could mimic the multilepton nal states are considered as relevant backgrounds in this analysis. The relevant back- grounds includes ZZ,WZ,WW ,WWZ ,WZZ ,ZZZ ,ZZh ,WWh Drell-Yan processes. These backgrounds can be classied into two classes | reducible and The reducible backgrounds are from the SM processes like etc., where a jet is misidentied as lepton or additional leptons originate from heavy quark decays. The irreducible ones are from diboson and triboson production and processes like ttW, ttZand Higgs boson production, etc. Note that nal state events with nleptons also contribute to those with 14Note that we have relaxed the opposite charge condition for the same- avour lepton pair to suppress the background contribu- tions due to charge 14 n1 leptons when one of the leptons falls outside the de- tector coverage (in the high rapidity region) or is too soft to pass the object reconstruction and selection criteria or gets misidentied by the detector. All the are generated in association of up to two jets us- ing MadGraph [91, 92] at the leading order using the 5 avour scheme followed by MLM matching in PYTHIA [109], and the corresponding cross-sections are taken at least upto NLO [118{130]. We plot dierent kinematic distributions for 3L events in Figure 7 for a benchmark point BP1 , dened as mH= 1 TeV,vt108GeV and
m= 0. The rst two plots of the top panel show the transverse momen- tum distributions of the leading and subleading lepton in the same-charge lepton pair. The eective mass, de- ned being the scalar sum of transverse momenta of all the jets is shown in the rightmost plot of the same panel. The bottom panel shows distributions of the miss- ing transverse momentum and the invariant mass of the same-sign lepton pair ( msc ``).15These kinematic distri-butions demonstrate that relatively stronger cuts on the same-sign leptons' pTand reduce the relevant backgrounds. For 3L events, we require one same-charge lepton pair. The leading (subleading) lepton in the pair is required to havepT>300(100) GeV. We discard events with me<1500 GeV. To enhance the sensitivity of this search, the selected events are categorised into two mu- tually exclusive SRs, namely 3 L0J16and 3L1J, based on whether or not at least one selected jet exists in the event. 3L1Jevents are further classied as 3 L1J-1 and 3L1J-2 based on whether pmiss Tis larger or smaller than 150 GeV. The 3 L0Jevents with pmiss T<150 GeV or msc ``<800 GeV are thrown away to get rid of the size- able SM backgrounds. Furthermore, we reject 3 L1J-1 events with pmiss T=HT<1:0. The cut on to be remarkably eectual in diminishing the left- over backgrounds. Finally, to supplement the sensitivity of this search, the selected events in 3 L1J-1 and 3L1J-2 SRs are divided into six bins each in the [600:1800] GeV range using msc ``as the primary kinematic 7. Kinematic distributions for 3L events for BP1 . Top panel: the pT-distributions of the leading (leftmost) and sub- leading (middle) same-charge lepton, and the medistribution (rightmost). Bottom panel: the pmiss T(left) and the The events are weighted at 1000 fb1luminosity at the 13 TeV LHC. 15In an ideal scenario, the invariant mass distribution in the bot- tom left plot would have a sharp peak at 1 TeV. However, mo- mentum smearing of the reconstructed objects due to nite res- olution of the detectors results in much broader peaks around 1 TeV.16Three leptons events with no reconstructed jet with pT>30 GeV are considered in the 3 L0JSR. 17The over ow (under ow) events are contained in the last (rst) bin in each signal region.15 FIG. 8. Kinematic distributions for 4L events for BP1 . Left (middle): the pT-distributions of the leading lepton. Right: the LTdistribution. The events are weighted at 1000 fb1luminosity at the 13 TeV LHC. Dierent kinematic distributions for 4L events are plot- ted in Figure 8 for BP1 . The leftmost and middle plot in the top panel shows the transverse momentum dis- tributions of the leading and subleading lepton in lepton pair. The leptons in the negative- charge lepton pair have similar we avert to show them for brevity. The distributions of the scalar pTsum of the leptons is shown in the rightmost plot in the same panel. It is evident from these kinematic dis- tributions that relatively stronger cuts on the in handy in subjugating the Further, a cut on LTturns out to be e- cacious in vanquishing the remaining background. For 4L events, we require two same-charge lepton pairs. The leading (subleading) lepton in both the pairs are required to have pT>300(100) GeV. The events with LT<1500 GeV are vetoed. Further, we require the invariant masses of the same-charge lepton pairs. The last cut ensures correct pairing of the leptons. Number of expected signal and background events in dierent signal regions after passing various selection cuts forBP1 for 1000 fb1of luminosity data at the 13 TeV LHC is are given in Table III and Table IV. The 3 L0J and 4Lsignal regions are free from any some backgrounds remain after all the selection cuts in the other two signal regions. However, these re- maining backgrounds are mostly distributed in the unlike the signal events which are distributed in a narrow central msc ``region, see Figure 9. Figure of signal and background events in 3L1J-1 signal region after passing various selection cuts for BP1 for 1000 fb1of luminosity data at the 13 TeV LHC. For brevity, we avert to show similar distri- bution for the 3 L1J-2 signal region. This simple binning of the selected events enhances sensitivity of these two signal regions.SR Selection cuts Background Signal (3L events)Basic 151444 1277 501 22.7 me>1500 46.0 20.8 3L0JnJ= 0 &pmiss T>150 1.0 6.5 msc ``>800 0.0 6.2 3L1J-1nJ1 &pmiss T>150 22.0 5.1 8.3 3L1J-2nJ1 &pmiss T<150 23.0 3.0 TABLE III. Number of expected background and signal events in dierent 3L signal regions after passing various se- lection cuts for BP1 for 1000 fb1of luminosity data at the 13 TeV LHC. All the dimensionfull cuts are in GeV. Selection cuts Background Signal Basic 350 5.3 2.0 0.7 0.1 5.1 LT>1500 &r<0:1 0.0 4.9 TABLE IV. Number of expected background and signal events in 4L signal region after passing various selection cuts forBP1 for 1000 fb1of luminosity data at the 13 TeV LHC. All the dimensionfull cuts are in GeV. E. Future 95% CL lower limit on mH In this section, we present our forecasted 95% CL lower limits onmHby using the ATLAS search [75] scaled at well as our proposed search de- scribed in Section V D. We simply presume that not only 18Note that the ATLAS search [75] is optimised for probing where the type-II see-saw anchors decay into bosons. Therefore, we use the same search strategy to forecast the future reach of the LHC in probing this part of the 800 1000 1200 1400 1600 / 200 (GeV)=13 TeV)s (11000 fb 3L1J1 of signal and background events in 3L1J-1 signal regions after passing various selection cuts for BP1 for 1000 fb1of luminosity data at the 13 TeV LHC. the detector eciencies and acceptances but also the background uncertainties remain the same while scal- ing the ATLAS search at high luminosity. Given that both statistical and systematic contributions to the back- ground uncertainties are expected to be reduced with in- creasing volume of LHC data, our forecasted future limits are conservative. Also, while estimating signicance for a given signal and background distributions, less than one background event at 3000 fb1 19is replaced by one background event. This renders our estimated limits to be conservative further. For the proposed search, we as- sume an overall 20% total uncertainty on the grey shaded region in Figure 10 is excluded from the existing ATLAS and CMS combined search, see Sec- tion V C. The regions below the green and cyan curves are expected to be excluded from the ATLAS search scaled at 500 and 3000 fb1of luminosity, respectively. Our proposed search is expected to probe the regions below the goldenrod and pink curves, respectively, at 500 and 3000 fb1of luminosity. For small (large) vt, the future reach extends up to 1220 and 1490 (520 and 640) for 500 and 3000 fb1of luminosity. We consider both the NH and IH neutrino mass spec- trum while varying the lightest neutrino mass in accor- dance with the bound from eV. The eect of dierent possible neutrino mass hypotheses on the limits are re ected as bands for small vtregions. 19All the relevant background events are generated in association of up to two jets using MadGraph [91, 92] at the leading order using the 5 avour scheme followed by MLM matching in PYTHIA [109] for an integerated luminosity of 3000 fb1or more, and the corresponding cross-sections are taken at least upto NLO [118{130]. 200 400 600 800 1000 1200 1400 1600 10 = + % CL lower limit on m H+-+- [GeV] vt proxy proxy 200 400 600 800 1000 1200 1400 1600 10 from lepton flavour violating decaysExcluded from measurement of r parameterFIG. 10. Left: 95% CL lower limits on mHas a func- tion ofvtfor
m= 0. The gray shaded region is excluded from the existing CMS and ATLAS combined search at the 13 TeV LHC. The bands resulted from dierent possible neutrino mass hypotheses. See text for details. This is because, for small vt, the triplet-like scalars and these decays are driven by the which, in turn, are determined by the parameters up to vt. For large vt, the triplet- like scalars decay into diboson and hadrons, and these decays are independent of the Yukawa couplings and the neutrino oscillation parameters. The solid curves within the bands correspond to NH with m1= 0:03 eV. The plots in Figure 11 show 95% CL future sensitivity of the LHC to probe as a function of vtassuming NH withm1= 0:03 eV for four dierent values of
m|
m= 10 GeV (top left),
m= 30 GeV (top right),
m=10 GeV (bottom left) and
m=30 GeV (bottom right). The color codings are same as those in Figure 10. For
m=10(30) GeV and moderate vt, the expected reach from our proposed search extends up to 1330(1310) and 1555(1550) GeV, respectively, at 500 and 3000 fb1of luminiosity. F. Summary and outlook The type-II see-saw mechanism based on the annexa- tion of the Standard Model by weak gauge triplet scalar eld proers a natural explanation for the very minute- ness of neutrino masses. Because of the presence of the doubly charged scalar bosons and their illustrious signa- tures, a number of collider searches have been carried out at the LHC by CMS and ATLAS to look for the same. In view of the observations being consistent with the SM background expectations, these searches derived strin- gent limits with 95% CL on mH. Most of these limits are derived in the context of simplied scenarios with- out reckoning the footprints of the low-energy Furthermore, these limits are often conser- vative as these searches do not incorporate all the pro- duction channels for the triplet-like scalars. Above all, in the non-degenerate scenario, the cascade decays are enti- tled to play a notable role in the phenomenology, thereby making the phenomenology for the non-degenerate sce- nario substantially contrasting than that for the degen- erate one. Patently, the aforesaid limits are not betting to the entire parameter space, rather valid only for a con- strained parameter space of the model. Bearing this dis- cussion in mind, we perform a comprehensive study for a wide range of the model parameter space
mandmH. Considering all the Drell-Yan pro-duction mechanisms for the triplet-like scalars and tak- ing into account the all-encompassing complexity of their decays, we derive the most stringent 95% CL lower limit onmHfor a vast range of vt-
mparameter space by implementing already existing direct collider searches by CMS and ATLAS. Further, we forecast future limits by extending the same ATLAS search at we propose a search strategy that yields improved limits for a part of the parameter space. To the extent of our apprehension, such a study of limits for a vast range of parameter space is still lacking. This work is intended to ll this gap. 200 400 600 800 1000 1200 1400 1600 10 = 1 + % CL lower limit on m H+-+- [GeV] vt proxy proxy 200 400 600 800 1000 1200 1400 1600 10 from lepton flavour violating decaysExcluded from measurement of r parameter 200 400 600 800 1000 1200 1400 1600 10 = 3 + % CL lower limit on m H+-+- [GeV] vt proxy proxy 200 400 600 800 1000 1200 1400 1600 10 from lepton flavour violating decaysExcluded from measurement of r parameter 200 400 600 800 1000 1200 1400 1600 10 = -1 + % CL lower limit on m H+-+- [GeV] vt proxy proxy 200 400 600 800 1000 1200 1400 1600 10 from lepton flavour violating decaysExcluded from measurement of r parameter 200 400 600 800 1000 1200 1400 1600 10 = -3 + % CL lower limit on m H+-+- [GeV] vt proxy proxy 200 400 600 800 1000 1200 1400 1600 10 from lepton flavour violating decaysExcluded from measurement of r parameter FIG. 11. 95% CL lower limits on mHas a function of vtassuming NH with m1= 0:03 eV for
m= 10 GeV (top left),
m= 30 GeV (top right),
m=10 GeV (bottom left) and
m=30 GeV (bottom right). See text for details. For large (small) vtand
m= 0, doubly charged scalar up to masses 420(955) GeV are already excluded from the existing ATLAS and CMS combined search; this exclu- sion limit is beyond those from the previous LHC searches [71, 72, 75] by approximately 50(200{230) GeV. For mod-eratevtand
m=10(30) GeV, the present exclusion limit extends up to 1115(1076) GeV. For large (small) vt, the ATLAS (our proposed) search scaled at 3000 is expected to probe them with masses be- low 640(1490) GeV for
m= 0. For
m=10(30)18 GeV and moderate vt, the expected reach from our pro- posed search extends up to 1555(1550) GeV at 3000 fb1 of luminosity. Furthermore, we nd that for moderate vtand large enough positive
m, the LHC searches fail to constrain the triplet-like scalars insomuch as H0=A0 decays invisibly into neutrinos or into on the value of vt. ForH0=A0decaying into neutrinos, there are hardly visible objects in the nal state, so much as the monojet search by ATLAS [103]and the soft leptons search by CMS [106, 107] fall short in constraining this part of the parameter space. On the contrary, for H0=A0decaying into h0h0;ZZ=h0Z, the sig- nal cross-section is small compared to the either from QCD jets or Drell-Yan processes. This makes such a scenario challenging to probe. In clos- ing this section, we mention that eecolliders could have better prospects for probing such a nightmare scenario which we left for future work. Appendix A: Partial decay widths of the triplet-like scalars The decay rates for dierent possible decay modes of the triplet-like physical scalar bosons, given in the following [14, 18, 97{100]. Decay rates of H:The following decay rates for Hcan be (A1) (H!WW) (A3) (H!HW) denotes the charged lepton mass and = 2(1) rates of H:The decay rates for Hcan be obtained as (H!qq0) (A7) (H!WZ) (A8) (H!WZ) (A9) (H!WZ) (A10) (H!^'W) cossinp 2 for ^ the weak mixing angle. Decay rates of H0:The decay rates for H0are given by (H0!ff) m2 f m2 H0!#3 ; (A15) (H0!) (A16) (H0!W+W) (A17) (H0!ZZ) (A18) (H0!WW) (A19) (H0!ZZ) (A21) (H0!HW) (A23) where and Nf cis the color factor with Nq c= 3 andN` c= 1, sin3i . The functions andH(x;y) has the following form: (x;y) = (1xy)24xy ; (A24) (x) =p (x;x) =p 14x ; (A25) F(x) =3(18x+ 4x2) logx ; (A26) G(x;y) =1 12y 2 (1 +x)39 1 +x2
y+ 6 (1 +x)y26 (1 tan1 + (A27) H(x;y) (9y+ 7)x25(1y)2x+ +x)(2 + 2x2+ 6y24x9y+ 6xy) logxo : (A28)20 Decay rates of A0:The decay rates for A0are given by (A0!ff) m2 f m2 A0! ; (A29) (A0!) (A30) (A0!h0Z) (A34) [1] W. Konetschny and W. Kummer, Phys. Lett. B 70, 433-435 (1977) T. P. Cheng and L. F. Li, Phys. Rev. D 22, 2860 G. Lazarides, Q. Sha and C. Wetterich, Nucl. Phys. B 181, 287-300 (1981) J. Schechter and J. W. F. Valle, Phys. Rev. D 22, 2227 (1980) R. N. Mohapatra and G. Senjanovic, Phys. Rev. D 23, 165 (1981) M. Magg and C. Wetterich, Phys. Lett. B 94, 61-64 (1980) K. Huitu, J. Maalampi, A. Pietila and M. Raidal, Nucl. Phys. B 487, 27-42 (1997) [hep-ph]]. [8] S. Chakrabarti, D. Choudhury, R. M. Godbole and B. Mukhopadhyaya, Phys. Lett. B 434, 347-353 (1998) [hep-ph]]. [9] E. J. Chun, K. Y. Lee and S. C. Park, Phys. Lett. B 566, 142-151 (2003) [hep-ph]]. [10] A. G. Akeroyd and M. Aoki, Phys. Rev. D 72, 035011 (2005) [hep-ph]]. [11] J. Garayoa and T. Schwetz, JHEP 03, 009 M. Kadastik, M. Raidal and L. Rebane, Phys. Rev. D77, 115023 (2008) [hep-ph]]. [13] A. G. Akeroyd, M. Aoki and H. Sugiyama, Phys. Rev. D77, 075010 (2008) [hep-ph]]. [14] P. Fileviez Perez, T. Han, G. y. Huang, T. Li and K. Wang, Phys. Rev. D 78, 015018 F. del Aguila and J. A. Nucl. Phys. B 813, 22-90 (2009) [hep-ph]]. [16] A. G. Akeroyd and C. W. Chiang, Phys. Rev. D 80, 113010 (2009) [hep-ph]]. [17] A. Melfo, M. Nemevsek, F. Nesti, G. Senjanovic and Y. Zhang, Phys. Rev. D 85, 055018 M. Aoki, S. Kanemura and K. Yagyu, Phys. Rev. D85, 055007 (2012) [hep-ph]]. [19] A. G. Akeroyd and H. Sugiyama, Phys. Rev. D 84, 035010 (2011) [hep-ph]]. [20] C. W. Chiang, T. Nomura and K. Tsumura, Phys. Rev. D85, 095023 (2012) [hep-ph]]. [21] E. J. Chun, H. M. Lee and P. Sharma, JHEP 11, 106 (2012) A. G. Akeroyd, S. Moretti and H. Sugiyama, Phys. Rev. D85, 055026 (2012) [hep-ph]]. [23] E. J. Chun and P. Sharma, JHEP 08, 162 [arXiv:1206.6278 [hep- ph]]. [24] P. S. Bhupal Dev, D. K. Ghosh, N. Okada and I. Saha, JHEP 03, 150 (2013) [erratum: JHEP 05, 049 [arXiv:1301.3453 [hep- ph]]. [25] S. Banerjee, M. Frank and S. K. Rai, Phys. Rev. D 89, no.7, 075005 (2014) [hep-ph]]. [26] F. del Aguila and M. Chala, JHEP 03, 027 [arXiv:1311.1510 [hep- ph]]. [27] E. J. Chun and P. Sharma, Phys. Lett. B 728, 256-261 (2014) [hep-ph]]. [28] S. Kanemura, K. Yagyu and H. Yokoya, Phys. Lett. B 726, 316-319 (2013) [hep-ph]]. [29] S. Kanemura, M. Kikuchi, K. Yagyu and H. Yokoya, Phys. Rev. D 90, no.11, 115018 S. Kanemura, M. Kikuchi, H. Yokoya and K. Yagyu, PTEP 2015 , 051B02 (2015) [hep-ph]]. [31] Z. Kang, J. Li, T. Li, Y. Liu and G. Z. Ning, Eur. Phys. J. C75, no.12, 574 (2015) [arXiv:1404.5207 [hep-ph]]. [32] Z. L. Han, R. Ding and Y. Liao, Phys. Rev. D 91, 093006 (2015) [hep-ph]]. [33] Z. L. Han, R. Ding and Y. Liao, Phys. Rev. D 92, no.3, 033014 (2015) [hep-ph]]. [34] D. Das and A. Santamaria, Phys. Rev. D 94, no.1, 015015 (2016) [hep-ph]]. [35] K. S. Babu and S. Jana, Phys. Rev. D 95, no.5, 055020 (2017) [hep-ph]]. [36] M. Mitra, S. Niyogi and M. Rev. D 95, no.3, 035042 Y. Cai, T. Han, T. Li and R. Ruiz, Front. in Phys. 6, 40 (2018) D. K. Ghosh, N. Ghosh, I. Saha and A. Shaw, Phys. Rev. D 97, no.11, 115022 A. Crivellin, M. Ghezzi, L. Panizzi, G. M. Pruna and A. Signer, Phys. Rev. D 99, no.3, 035004 Y. Du, A. Dunbrack, M. J. Ramsey-Musolf and J. H. Yu, JHEP 01, 101 (2019) [hep-ph]]. [41] P. S. Bhupal Dev and Y. Zhang, JHEP 10, 199 [hep- ph]]. [42] S. Antusch, O. Fischer, A. Hammad and C. Scherb, JHEP 02, 157 (2019) [hep-ph]]. [43] A. Aboubrahim and P. Nath, Phys. Rev. D 98, no.9, 095024 (2018) [hep-ph]]. [44] T. B. de Melo, F. S. Queiroz and Y. Villamizar, Int. J. Mod. Phys. A 34, no.27, 1950157 R. Primulando, J. Julio and P. Uttayarat, JHEP 08, 024 (2019) [hep-ph]]. [46] R. Padhan, D. Das, M. Mitra and A. Kumar Nayak, Phys. Rev. D 101, no.7, 075050 E. J. Chun, S. Khan, S. Mandal, M. Mitra and S. Shil, Phys. Rev. D 101, no.7, 075008 J. C. Pati and A. Salam, Phys. Rev. D 10, 275-289 (1974) [erratum: Phys. Rev. D 11, 703-703 R. N. Mohapatra and J. C. Pati, Phys. Rev. D 11, 566- 571 (1975) G. Senjanovic and R. N. Mohapatra, Phys. Rev. D 12, 1502 (1975) J. F. Gunion, R. Vega and J. Wudka, Phys. Rev. D 42, 1673-1691 (1990) J. E. Cieza Montalvo, N. V. Cortez, J. Sa Borges and M. D. Tonasse, Nucl. Phys. B 756, 1-15 (2006) [erratum: Nucl. Phys. B 796, 422-423 (2008)] [hep-ph]]. [53] N. Arkani-Hamed, A. G. Cohen, E. Katz, A. E. Nel- son, T. Gregoire and J. G. Wacker, JHEP 08, 021 (2002) [hep-ph]]. [54] N. Arkani-Hamed, A. G. Cohen, E. Katz and A. E. Nelson, JHEP 07, 034 (2002) [hep-ph]]. [55] A. Hektor, M. Kadastik, M. Muntel, M. Raidal and L. Rebane, Nucl. Phys. B 787, 198-210 M. S. Chanowitz and M. Golden, Phys. Lett. B 165, 105-108 (1985) H. Georgi and M. Machacek, Nucl. Phys. B 262, 463- 477 (1985) A. Zee, Nucl. Phys. B 264, 99-110 K. S. Babu, Phys. Lett. B 203, 132-136 K. S. Babu, S. Nandi and Z. Tavartkiladze, Phys. Rev. D80, 071702 (2009) [hep-ph]]. [61] I. Picek and B. Radovcic, Phys. Lett. B 687, 338-341 (2010) [hep-ph]]. [62] K. Kumericki, I. Picek and B. Radovcic, Phys. Rev. D86, 013006 (2012) [hep-ph]]. [63] R. Cepedello, M. Hirsch and J. C. Helo, JHEP 01, 009 (2018) [hep-ph]]. [64] G. Anamiati, O. R. M. Fonseca, J. C. Helo and M. Hirsch, JHEP 12, 066 [hep- ph]]. [65] Avnish and K. Ghosh, [hep-ph]]. [66] G. Aad et al. [ATLAS], Eur. Phys. J. C 72, 2244 S. Chatrchyan et al. [CMS], Eur. Phys. J. C 72, 2189 (2012) [hep-ex]]. [68] G. Aad et al. [ATLAS], JHEP 03, 041 [arXiv:1412.0237 [hep- ex]]. [69] V. Khachatryan et al. [CMS], Phys. Rev. Lett. 114, no.5, 051801 [CMS], [CMS], M. Aaboud et al. [ATLAS], Eur. Phys. J. C 78, no.3, 199 (2018) [hep-ex]]. [73] A. M. Sirunyan et al. [CMS], Phys. Rev. Lett. 120, no.8, 081801 (2018) [hep-ex]]. [74] M. Aaboud et al. [ATLAS], Eur. Phys. J. C 79, no.1, 58 (2019) [hep-ex]]. [75] G. Aad et al. [ATLAS], JHEP 06, 146 [hep- ex]]. [76] A. Arhrib, R. Benbrik, M. Chabab, G. Moultaka, M. C. Peyranere, L. Rahili and J. Ramadan, Phys. Rev. D84, 095005 (2011) [hep-ph]]. [77] M. Cirelli, N. Fornengo and A. Stru- mia, Nucl. Phys. B 753, 178-194 [hep-ph]]. [78] P. A. Zyla et al. [Particle Data Group], PTEP 2020 , no.8, 083C01 (2020) M. Aoki, S. Kanemura, M. Kikuchi and K. Yagyu, Phys. Rev. D 87, no.1, 015012 N. Aghanim et al. [Planck], Astron. Astrophys. 641, A6 (2020) [erratum: Astron. Astrophys. 652, C4 (2021)] I. Esteban, M. C. Gonzalez-Garcia, M. Maltoni, T. Schwetz and A. Zhou, JHEP 09, 178 [hep- ph]]. [82] M. Kakizaki, Y. Ogura and F. Shima, Phys. Lett. B 566, 210-216 (2003) [hep-ph]]. [83] A. G. Akeroyd, M. Aoki and H. Sugiyama, Phys. Rev. D79, 113010 (2009) [hep-ph]]. [84] D. N. Dinh, A. Ibarra, E. Molinaro and S. T. Pet- cov, JHEP 08, 125 (2012) [erratum: JHEP 09, 023 (2013)] A. M. Baldini et al. [MEG], Eur. Phys. J. C 76, no.8, 434 (2016) [hep-ex]]. [86] U. Bellgardt et al. [SINDRUM], Nucl. Phys. B 299, 1-6 (1988) K. Ghosh, S. Jana and S. Nandi, JHEP 03, 180 [hep- ph]].[88] B. Dutta, R. Eusebi, Y. Gao, T. Ghosh and T. Kamon, Phys. Rev. D 90, 055015 F. Staub, Comput. Phys. Commun. 185, 1773-1790 (2014) F. Staub, Adv. High Energy Phys. 2015 , 840780 [hep-ph]]. [91] J. Alwall, M. Herquet, F. Maltoni, O. Mat- telaer and T. Stelzer, JHEP 06, 128 [arXiv:1106.0522 [hep- ph]]. [92] J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H. S. Shao, T. Stelzer, P. Torrielli and M. Zaro, JHEP 07, 079 [arXiv:1405.0301 [hep- ph]]. [93] R. D. Ball et al. [NNPDF], Nucl. Phys. B 877, 290-320 (2013) [hep-ph]]. [94] R. D. Ball et al. [NNPDF], JHEP 04, 040 [arXiv:1410.8849 [hep- ph]]. [95] M. Muhlleitner and M. Spira, Phys. Rev. D 68, 117701 (2003) [hep-ph]]. [96] B. Fuks, M. Nemev sek and R. Ruiz, Phys. Rev. D 101, no.7, 075022 (2020) [hep-ph]]. [97] T. G. Rizzo, Phys. Rev. D 22, 722 W. Y. Keung and W. J. Marciano, Phys. Rev. D 30, 248 (1984) R. N. Cahn, LBL-29789. [100] A. Djouadi, [hep-ph]]. [101] A. M. Sirunyan et al. [CMS], JHEP 03, 051 [ATLAS], G. Aad et al. [ATLAS], Phys. Rev. D 103, no.11, 112006 (2021) [hep-ex]]. [104] M. Aaboud et al. [ATLAS], Phys. Rev. D 94, no.3, 032005 (2016) [hep-ex]]. [105] M. Aaboud et al. [ATLAS], JHEP 01, 126 [CMS], A. M. Sirunyan et al. [CMS], Phys. Lett. B 782, 440-467 (2018) [hep-ex]]. [108] R. Foot, H. Lew, X. G. He and G. C. Joshi, Z. Phys. C 44, 441 (1989) T. Sj ostrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten, S. Mrenna, S. Prestel, C. O. Ras- mussen and P. Z. Skands, Comput. Phys. Com- mun. 191, 159-177 (2015) [hep-ph]]. [110] J. de Favereau et al. [DELPHES 3], JHEP 02, 057 (2014) W. Verkerke and D. P. Kirkby, eConf C0303241 ,

2108.10952

Saiyad Ashanujjaman

Saiyad Ashanujjaman and Kirtiman Ghosh

Revisiting Type-II see-saw: Present Limits and Future Prospects at LHC

23 pages, 11 figures

10.1007/JHEP03(2022)195

hep-ph hep-ex

http://creativecommons.org/licenses/by/4.0/

The type-II see-saw mechanism based on the annexation of the Standard Model by weak gauge triplet scalar field proffers a natural explanation for the very minuteness of neutrino masses. Noting that the phenomenology for triplet Higgs spectrum is substantially contrasting than that for the degenerate one, we perform a comprehensive study for an extensive model parameter space parametrised by the triplet scalar vacuum expectation value (VEV), the mass-splitting between the triplet-like doubly and singly charged scalars and the mass of the doubly charged scalar. Considering all mechanisms for the triplet-like scalars and taking into account complexity of their decays, we derive the most stringent 95% CL lower limits on the mass of the doubly charged scalar for a vast model parameter space by implementing already existing direct collider searches by CMS and ATLAS. These estimated limits are beyond those from the existing LHC searches by approximately 50-230 GeV. However, we also find that a specific region of the parameter space is not constrained by the LHC searches. Then, we forecast future limits by extending an ATLAS search at high-luminosity, and we propose a search strategy that yields improved limits for a part of the parameter space.

Figure 1. An example object (lamp) from datasets used for our project. OWILD, OOWL, and OOWLSeg are part of the ODDS dataset, and are real. Meanwhile, ShapeNet is evaluate the domain adaptation methods, and verify the usefulness incorporating class labels. 2. Related Work 2.1. 3D Single View Object view 3D reconstruction has several immediate ap- plications. For example, it would enable objects to be re- constructed as a 3D model and placed into an augmented (AR) or virtual (VR) environment, so that the user could manipulate them in that environment. Another use case is for robotic grasping - if an object such as a cup could be scanned, it would provide valuable information for a robot trying to pick it up by its handle. As a result, several meth- ods have been proposed in the computer vision literature [28, 16, 6, 20, 25, 9]. They vary in the type of 3D represen- tation used and each have their own trade-offs. For example, voxels are easily adapted to Convolutional Neural Networks (CNNs) but are spatially inefficient; triangular meshes are efficient but suffer from irregularities; point clouds are sim- ple but lack explicit structural information. However, most of the these methods use ShapeNet for training and testing, and do not incorporate techniques used in the domain adap- tation literature to allow for real-world viability. [27] uses Graph Convolutional Neural Net (GCN) to deform a mesh of ellipsoid to obtain an output mesh. The results, however, are not very accurate and it fails to keep the genus of the ground truth. [16] uses 2D convolutional network to gener- ate dense point clouds that shapes the surface of 3D objects in an undiscretized 3D space. The method predicts accurate shapes with higher point density but is problematic when objects contain very thin structures. [6] proposed a that unifies single and multi-view 3D recon- struction into a single framework. The method uses neural networks (3D Recurrent Reconstruc-tion Neural Network) to learn a mapping from observations to their underlying 3D shapes of objects from a large col- lection of training data. It incrementally improves its as it sees more views of an object but is unable to reconstruct many details and struggles with objects having high texture levels. 2.2. Unsupervised Domain approaches to unsupervised domain consist of matching the feature distribution between the source and target domain. Generally these methods can be categorized as either sample re-weighting (eg. [13], [15]) or feature space transformations (eg. [2], [23]). Con- volutional neural networks are also used for this purpose, because of their ability to learn powerful features. These methods, in general, are trained to minimize a classifica- tion loss while maximizing domain confusion. The clas- sification loss is usually computed using a convolutional neural network trained on the labeled data. The domain confusion is usually achieved either by using a discrepancy loss, which reduces the shift between the two domains such as in ([17], [18], [4]) or via an adverserial loss which encourages a common feature space with respect to a discriminator loss, such as in [10], [4], [1]. [24] achieves the domain-confusion by aligning the second-order statis- tics of the learned feature representations. In [26] a loss based on MMD [12] is applied to the final layer representation of a network. [17] uses a sum of mul- tiple MMDs between several layers and [18] continues this line of work by using the joint distribution discrepancy over deep features, instead of their sum. 2Figure 2. The proposed architecture, which is based off the pix2vox framework. We provide two additions: 1) incorporating in the style of DANN through a domain classifier and a gradient reversal layer, and 2) since we have classification labels for both the source and target domain, we also classify the produced voxel. 3. Methods 3.1. 3D Reconstruction Backbone Architecture The backbone for single view 3D reconstruction that was chosen for this project is the pix2vox architecture [28]. It is a convolutional neural network which encodes input im- ages into a latent feature map, which is then decoded into voxel. The encoder uses convolu- tional layers while the decoder uses 3D transpose convolu- tional layers. An additional refiner CNN which is based on the U-Net is also employed to increase performance [21]. Standard techniques are used throughout the network, in- cluding batch normalization [14], ReLU and ELU layers, and ImageNet [8] weights pretrained on VGG [22]. was chosen because it is much more efficient than other competing methods such as PSGN [9], OGN [25], and 3D-R2N2 [6], while being comparable or better in terms of performance. Due to limited computational re- sources, this was a critical factor in our decision. Note that in the original paper, pix2vox utilizes the ShapeNet dataset. 3.2. Maximum Mean Discrepancy domain adaptation is quite we do not have labeled information for the target do- main. Some approaches to the problem are to try to bound the target error by the source error plus a discrepancy met- ric between the source and the target. The Maximum Mean Discrepancy (MMD) is a measure of the difference between two probability distributions from their samples. It is an ef- fective criterion that compares distributions without their density functions. Given two pandqonX, MMD is defined as MMD (F;p;q) a class of functions f:XR . By defining Fas the set of functions of the unit ball in a universal Re- producing Kernel Hilbert Space (RKHS), denoted by H, it was shown thatMMD (F;p;q) = 0 will detect any dis- crepancy between tbe data vectors drawn from distributions DsandDton the data spaceX, respec- tively. Since fis in the unit ball in a universal RKHS, we can rewrite the empirical estimate of MMD as MMD e(xs;xt) :X!H is referred to as the feature space map. 3.3. Deep CORAL Another approach to domain adaptation is by aligning the statistics of the source and target domains. CORAL [23] does this by using a linear transformation to align the co- variances (second order statistic) of the domains. Assuming we have a labeled source domain unlabeled target domain each sample is a ddimensional vector, the CORAL loss is de- fined as: `CORAL =1 4d2kCSCTk2 F (3) Wherek
kFis the Frobenius norm and feature covariance matrices for the the source and target data, respectively. These matrices are given (5) 3Figure 3. Reconstruction results of applying our proposed architecture on the three domains in the ODDS dataset; ShapeNet is used as the source domain, and each are trained on their respective target domains. Figure 4. Learned feature map embeddings visualized with t-SNE, using our proposed model on the three target domains in the ODDS dataset. Purple denotes the source domain (ShapeNet), yellow de- notes the target domain (OWILD, OOWL, and OOWL Seg). Where 1is addimensional vector containing all 1and the data the source and target data, Domain Adversarial Neural Network (DANN) DANN [11] focuses on combining domain adaptation and deep feature learning under one training process. It embeds the domain adaptation method into the process of learning representation to obtain features which are discrim- inative and domain invariant. This is achieved by the underlying features as well as two discrimi- native classifiers operating on these features, the label pre- dictor and domain classifier . The label predictor predicts class labels and is used both during training and at test time. Figure 5. Learned feature map embeddings visualized with t-SNE. Purple denotes the source domain (ShapeNet), yellow denotes the target domain (OOWL). On the left we show the result when do- main adaptation is not used. On the center and right, we show re- sults achieved by applying domain adaptation (DANN or CORAL) to the vanilla pix2vox model. domain classifier discriminates between the source and the target domains during training. The model works to mini- mize the loss of the label classifier and maximize the loss adversarially, thereby encouraging features. 3.5. Voxel Classification Architecture The domain adaptation techniques discussed above can be readily applied to the pix2vox architecture. However, their success may be limited, since the gap between syn- thetic and real domains is large. To help with this process, in addition to using domain adaptation techniques, we per- 4form classification of the output voxel by vectorizing it, and using several fully connected layers of size 100 and 20, with ReLU activations. This is possible since we have the ground truth class labels for both the source and target domain. This proposed architecture can be seen in Figure 2; all losses are trained end-to-end. We utilize the standard cross entropy loss. This idea is inspired by the fact that in general, the output voxels should resemble their respective classes. We found that this additional source of supervision, which ap- plies to both the source and target domain, is highly benefi- cial; further details can be found in Section 4. 4. Experiments 4.1. Relevant Datasets As mentioned previously, ShapeNet is a synthetic dataset which provides many 2D rendered images of objects, as well as their ground truth 3D mesh representations (which can be converted into voxels). As shown in Figure 1, they are CAD models with little detail, no texture, and no background. While the original ShapeNet has 270 classes, pix2vox uses a subset of 13 classes. In addition, we utilize the ODDS dataset, which is a real, multiview, class-organized image dataset with multi- ple domains. It contains 25 classes, 6 of which overlap with the ShapeNet classes used to train the original pix2vox model. The overlapping classes are airplanes, cars, moni- tors, lamps, telephones, and boats, so these are the classes we use when comparing results between datasets. Each ODDS class contains 20 object instances (for example, the monitor class has 20 different types of monitors); each ob- ject instance has 8 images of it taken at 45 degree incre- ments. Please see Figure 1 for some example images. There are 3 domains in the ODDS dataset that we work with. First, OWILD contains images of objects in locations, and pictures are taken with a smart- phone. Second, OOWL is taken with a drone inside a lab setting. As a result, it contains several domain as as camera blur and a lower camera resolution. Fi- nally, OOWLSeg is a segmented version of OOWL. Note that this data captures the real-world input statistics that we would like to work with: they're real objects, taken with smartphone cameras in various real-world locations. How- ever, they also represent a trade-off between ease of collec- tion and realism - this is shown through the arrow in Figure 1 on the right. 4.2. Evaluation Metrics The standard evaluation for 3D reconstruction when us- ing voxels is the intersection over union (IoU) score. For- mally, it )>t)I(gt(i;j;k )>t) +I(gt(i;j;k )2[0;1]is the predicted occupancy proba- bility at voxel location (i;j;k )andgt(i;j;k )2 f0;1gis the ground truth value at voxel location (i;j;k ). Given a predicted reconstruction voxel and ground truth voxel, if IoU= 1, then they are the same voxel. If IoU= 0, then there is no intersection between the two voxels. Thus, a higher IoU score indicates a better reconstruction result. It is important to note that for the case of the ODDS dataset, we do not have ground truth (currently, we are not aware of any publicly available, sufficiently large multiview dataset with 3D ground truth). Therefore, for the scope of this pa- per, it is primarily used as a test dataset to judge qualita- tive reconstruction results. We cannot quantitatively evalu- ate metrics like IoU, due to the lack of 3D ground truth (it is unsupervised in this regard). However, we do try to uti- lize the ground truth class labels that come with OWILD Application of Domain Adaptation on a Vanilla Pix2Vox Model First, we report the results of applying DANN and Deep CORAL domain adaptation to the vanilla pix2vox model for single view 3D reconstruction. Qualitative reconstruc- tion results are shown in Figure 7, under CORAL and DANN. We observed that reconstruction when not using any domain adaptation generally looks like random noise (these reconstruction results are omitted to save space). Also, in our experiments we found that MMD was not ef- fective; therefore, MMD results have also been Deep CORAL and DANN, we can see that in general both help, though results are still far from checking the reconstruction results, we found that DANN performed better than CORAL. We also embed the learned feature maps into 2D, using t-SNE as the dimen- sionality reduction algorithm. This is shown in Figure 5. We can see that the use of DANN and Deep CORAL both helps to make embedded features more domain invariant - the distributions of the source ShapeNet domain (pur- ple) are more matched with the distributions of the target OOWL domain (yellow). We also note that the introduc- tion of domain adaptation also negatively impacts the IoU on the source domain - this is shown in Table 1. makes sense since the network is constrained to only output domain invariant latent representations. This makes training more difficult. In the future, we would like to look into this more and see if we can maintain IoU results on the source domain while performing domain adaptation. 4.4. Reconstruction with a Voxel Classification Loss As mentioned above, we found that regarding the vanilla pix2vox model, DANN is helpful. However, results are still sub optimal. To address this, we utilize our proposed network. Training is performed No DA DANN CORAL Airplane 0.6842 0.6046 0.6377 Car 0.8548 0.8377 0.8485 Monitor 0.5373 0.4845 0.4968 Lamp 0.4430 0.4228 0.4322 Telephone 0.7764 0.7278 0.7555 Boat 0.5946 0.5852 0.5926 Overall 0.7110 0.6762 0.6925 Table 1. IoU results for a model trained on ShapeNet with and without OOWL domain adaptation (DANN, Deep CORAL). Re- sults reported for t= 0:4. Figure 6. Training results on our architecture. On the left, the loss (in blue) is plotted against the IoU. On the right, the domain discrepancy loss (in blue) is plotted against the IoU. and we report training losses as a function of epoch in Fig- ure 6. While the voxel classification loss does decrease, during training we found it difficult to reduce it past epoch 20 - it fluctuates beyond that point. In the future, we plan on trying to look into ways of address this. Meanwhile, the do- main discrepancy loss is maintained at around 0.5. This is expected due to adversarial training induced by the layer. Next, using this trained model we evaluate the differ- ences between the domains in ODDS. This gives us a way to see how large the domain gap is between Shapenet and the target domains OOWL, OOWL Seg, and OWILD. We re- port t-SNE embeddings in Figure 4 and the Figure 3 for the three target domains. We can see that in general, it appears that OWILD is the most We believe that this is because OWILD has com- plex backgrounds, which make it very far from the shapenet domain. On the other hand, we can see that OOWL seg per- forms quite well. We believe that the segmentation makes the images very similar to ShapeNet, which also do not have a background. 5. Conclusion and Future Work In this paper, we have focused on the task of single view voxel reconstruction in the real world. To do this, we ex- tended the pix2vox architecture using domain the supervised synthetic ShapeNet dataset and real ODDS dataset. However, we showed that simply applying domain adaptation is not enough; re- Figure 7. Reconstruction results when domain adaptation is used with OOWL as the target domain. We compare domain adapta- tion using Deep CORAL, DANN, and our proposed + V oxel results are only marginally better. Therefore, we proposed an architecture which also utilizes a voxel clas- sification loss in addition to an adversarial loss, which led to better results. There are several extensions that were not done due to time and computational constraints which we plan on exploring in the future. First, it would be interesting to see if our conclusions in the project hold for other mesh or point clouds). Second, no large dataset exists with real world ground truth 3D data. Perhaps the closest is Pix3D, but it only has 8 classes, and pictures are only from one angle. No class overlaps with Pix3D, OWILD, and ShapeNet. If such a dataset existed, it would be feasible to achieve quantitative, not just qualitative Third, we plan on working towards on OWILD through more experimentation. For example, we want to try data augmentation on ShapeNet by pasting background from the MIT Places datasets [29] and see if that improves results. We also want to further explore the domain gaps between the datasets through methods like domain bridges, which proposes gradually over several intermediate domains which increase in difficulty to the final target domain [7]. Finally, because ODDS is a multiview dataset, it would be natural to generalize results to the multiview

2108.10972

Brandon Leung

Brandon Leung, Siddharth Singh, Arik Horodniceanu

Domain Adaptation for Real-World Single View 3D Reconstruction

cs.CV

http://creativecommons.org/licenses/by/4.0/

Deep learning-based object reconstruction algorithms have shown over classical methods. However, supervised learning based methods perform poorly when the training data and the test data have Indeed, most current works perform satisfactorily on the synthetic ShapeNet dataset, but dramatically fail in when presented with real world images. To address this issue, unsupervised domain adaptation can be used transfer knowledge from the labeled synthetic source domain and learn a classifier for the unlabeled real target domain. To tackle this challenge of single view 3D reconstruction in the real domain, we experiment with a variety of domain adaptation techniques inspired by the maximum mean discrepancy (MMD) loss, Deep CORAL, and the domain adversarial neural network (DANN). From these findings, we additionally propose a novel architecture which takes advantage of the fact that in this setting, target domain data is unsupervised with regards to the 3D model but supervised for class labels. We base our framework off a recent network called pix2vox. Results are performed with ShapeNet as the source domain and domains within the Object Dataset Domain Suite (ODDS) dataset as the target, which is a real world multiview, multidomain image dataset. The domains in ODDS vary in difficulty, allowing us to assess notions of domain gap size. Our results are the first in the multiview reconstruction literature using this dataset.

Shape anisotropy effect on magnetization reversal induced by linear down chirp smaller than anisotropy. The findings of damping depen- dence of magnetization reversal indicate that, for a of DCMP can be lowered by applying direct along with DCMP. Therefore, an of microwave field pulse and current could be applied to achieve faster and power efficient So these findings might be significant to cost magnetic storage device with high speed Analytical Model and Method We consider a spin valve system that consists of a free and fixed ferromagnetic layers by sandwiching a Aandthickness d. direction of the fixed and free layer is i.e., both are perpendicularly magnetized. ofthefreelayercanbetreatedasamacrospin represented by the unit moment mwith total the saturation demagnetization field, which generated by the magne- tization of nanoparticle, can be approximated by an easy- plane shape anisotropy. The demagnetization field field is expressed as the vacuum magnetic perme- ability. Because of the strong uniaxial Hani=Hanimzz, the magnetization of the free layer prefers to stay in two ground states, i.e., mparallel the magnetization dynamics mis governed by (LLG) equation [7, 9, 37, the microwave magnetic field Hmw, the Hk= Hkmzz.hsis the strength of spin transfer the current density, elec- tron charge, the Planck's constant, spin polarization of cur- mf0 0-f(b) 1: (a) Schematic diagram of the system in where mand pdenote the unit vectors of the magnetization of free layer and fixed layer, respectively. A linear down-chirp microwave field and an electric current are applied on a single (b) The frequency profile illustrates the linear down-chirp microwave two components: an uniaxial anisotropy Haniand a shape anisotropy can be expressed as Hk=Hani+Hshape= Thus, the resonant frequency of of energy change is expressed as (3) Since damping is positive, so, the first term is always nega- tive. For time-dependent external microwave field, depend- ing on the angle between the instantaneous the second term can be either positive or neg- ative, in other words, the microwave field pulse can energy absorption or emission. It has been shown that the fast magnetization reversal is achieved by DCMP [42] with the physical picture: the fre- quency of DCMP matches the frequency change of (emission) by (from) the magnetization before (after) it crosses over the energy .t/ is the phase. In this paper, we use a DCMP whose instanta- neous frequency is defined as f.t/ =1 2d dt, which with time at constant rate (in units of ns*2) as shown in Figure 1(b). The phase .t/and the given (5) Z. K. Juthy et al.: Preprint submitted to Elsevier Page 2 of 7Shape anisotropy effect on magnetization reversal induced by linear down chirp /s50/s48 /s51/s48 /s52/s48 /s53/s48 /s54/s48 /s48/s46/s51 /s48/s46/s51 /s48/s46/s54 2: Model parameters of the free layer, s), and= 0:01. (a) Temporal evolution of mzinduced by DCMP with Hmw=0.045 T for the samples of different A. (b) The dependence of switching time tson the chirp rate for different cross-sectional areas while keeping the microwave field The vertical dashed lines indicate the lower and upper limits of for magnetization switching. (c) The dependence of chirp rate on shape anisotropy Hshapefor switching time within 1 /s48 3: (a) The potential along the line = 0 for the samples of different A. Anglesandare defined as The symbols i and s represent the of (b) A0=88nm2and (c)A8=3232 t=0. (3) can be expressed as Therefore, the microwave field pulse can trigger the stimulated energy absorption (for *.t/) which is occurred before crossing the energy barrier and after crossing the energy barrier. For this study, the parameters of Permalloy are chosen from typical experiments on microwave-driven magnetiza- tion reversal as 1310*12J_m and= 0:01. Al- though, the demonstrated strategy and findings would still be valid for other materials. The cell size used throughout this study is .222/ nm3. We solved the LLG circularly polarized microwave field. We con- sider the switching time window of 1ns and least magne- tization reversal mz=*0:9. 3. Numerical Results In a small-scale nanoparticle, demagnetization field account in the investigation of the magnetization rever- sal. In order to approximate demagnetization field Hshape,wechoosecuboidshapedsample d istheheight. d=8 nm is kept fixed. Specifically, we intend to investigate reversal of the samples, A1=1010 ,A4=1616 ,A5=2020 ,A8=3232 nm2. finding the demagnetization factors [45] NzandNx, the shape anisotropy coefficient analytically. Hshapeactually acts along *zi.e., it opposes the intrinsic easy axis anisotropy field Haniand hence reduces the effective anisotropy. Therefore, for in the Table nanoparticles of different A, by keeping to resonant frequency corresponding to Aor Hshape) fixed, we estimate the optimal for each sample or Hshape. dependenceof orHshape. It is noted that there exists a window of denoted by vertical dashed line. For a small range of in the right edge of window, the tsis minimal. It means . findings are useful in device application since it was reported that generating DCMP with higher chirp rate is a Z. K. Juthy et al.: Preprint submitted to Elsevier Page 3 of 7Shape anisotropy effect on magnetization reversal induced by linear down chirp pulse Table 1 Simulated values of the controlling parameters for different samples of cross-sectional area area Shape anisotropy field Natural frequency Simulated minimal frequency A(nm2) hshape(T) f0(GHz) f0(GHz) A1 0.09606 18.3 18.3 A2 0.17718 16 17 A3 0.2465 14.1 14. A4 0.3064 12.4 12.4 A5 0.4049 9.6 10.2 A6 0.4827 7.49 9 A7 0.5457 5.72 7.5 A8 0.5980 4.26 5.2 challenge [42]. Afterward, DCMP (with Hmw= reversals are investigate and thus the time evolu- tions ofmzfor different HshapeorAare shown in the Fig- ure 2(a). It is found that with the increase of Hshape, switching time tsremains almost same until A7= 2828 nm2but forA8g3232nm2, the tsdropsto=0.34 ns)[40]. ultimate demand in device application. To understand the reason why fast reversal is obtained, one can recall that the energy barrier is proportional to Sincethe anisotropy field Haniand thus reduces the height of the energy barrier (energy difference between the initial state and saddle point). This reduction of energy barrier for frequency and magnetization precession matches as expected. To be more explicit, one Figure 3(b) and 3(c). For A0= 88 nm2orHshape= 0, the magnetization takes a longer time because of some additional processions while crossing the higher energy barrier. However, for A8= 3232 nm2 orHshape= 0:598T, the magnetization reversal is barrier is reduced by Hshape. For more A8=3232 nm2. Notethattheangle  remains almost 90oduring magnetization reversal which (emission) by magnetization. Thus the and emission peaks (no distortion) which lead fast reversal. However, Fig 4(b), for A0= 88nm2i.e., for zeroHshape, shows the (red line) absorption and emission peaks (distorted) indicating inefficient i.e., slower magneti- zation reversal (black chirp rate .Hshape/, we obtain the fast /s48/s46/s49 /s48/s46/s50 /s48/s46/s49 /s48/s46/s50 /s48/s46/s51 /s48/s46/s50 /s48/s46/s52 4: (a) ForA8= 3232 nm2, the plot of the time (blue line) and the time dependence of mz(black line). Plots of the energy changing rate of mag- netization against time (red line) and the time dependence of mz(black line) for the cross-sectional area (b) A0=88nm2 and (c)A8=3232 nm2 /s48/s46/s48 /s48/s46/s50 /s48/s46/s52 5: The dependence of initial frequency foand min- imum required field amplitude Hmwas a function of switching time within 1 ns. reversal of different Afor a certain range of f0around time is lowest. Figure 5 shows the simulated These can be explained as the Hshapere- duces the effective anisotropy Hk. For further K. Juthy et al.: Preprint submitted to Elsevier Page 4 of 7Shape anisotropy effect on magnetization reversal induced by linear down chirp /s48/s46/s48/s51 6: (a) Minimal switching time for the samples of dif- ferentAas a function of Gilbert damping . trajectories of A3=1414 nm2for damping (b) = 0.010 and (c) in same Figure 5 and noted that the decreasing trends are consistent as expected. Later on, by keeping the optimal .Hshape/(i.e., This is explained by the same reason of reducing the height of the energy barrier with Hshapeas shown in the Figure 2(b). So even the smaller amplitude Hmwcan drive faster magnetization parameter which has a significant effect on the magne- tization reversal process as well as switching time. In gen- eral, damping has two influences. First, it hinders (cost more energy) during climbing the energy barrier for magnetization reversal [50]. Second, after cross- ing the energy barrier, it dissipates the energy of magne- tization promptly and thus it accelerates the magnetization to reach the ground _stable state [51, 52]. However, in switchingprocess,themagnetizationrequires to climb an energy barrier (originated from the and then goes to a stable state. Ideally, for the () before (after) cross over the energy barrier is better. Therefore, this study emphasizes finding the Accordingly, DCMP-driven magnetization as a func- tion offor different Ais studied. Figure 6(a) shows the Gilbert damping dependence of switching time for different A. It is found that for A0= 88nm2(uniaxial sample), the range of is small but for finite shape anisotropy (biax- ial sample), the range of is large which is useful in Moreover, for the sample of 1414nm2, Tobe more explicit, Figure 6(b) and 6(c) show the trajectories of magnetization reversal for = 0:01and= and observed that for = 0:017, the reversal path is shorter. This is happened because, after cross over the en- /s48/s46/s48 /s48/s46/s53 /s49/s46/s48 /s49/s46/s53 /s50/s46/s48 7: Phase diagram, for magnetization switching within 1 ns, in terms of DCMP field amplitudes Hmwand the samples of different A. ergy barrier, the larger reduces the magnetization preces- sion by dissipating the magnetization energy promptly sample that has an optimal damping situation at which the magnetization reversal might be should be of smaller amplitude Hmw, smaller A=1616 and=40:672ns*2. In order to further reduce Hmw, we simultaneously apply a direct cur- rent such that it assists the magnetization reversal Akept fixed. Figure 7 shows the Hmw*Jphase diagram for the sample of different Aunder the condition of time 1 ns. As an example, for A= 1616 nm2, without DCMP, the required current to obtain fast switch- ing gets very high i.e., J= 2:31012A_m2while point of view. Therefore, with simultaneous applying HmwandJcould be smaller than the above mentioned values (for magnetization reversal strategy. 4. Discussion and dependence of magnetization reversal. The shape anisotropy interestingly assists magnetization reversal since the parameters of DCMP with the the increase of shape anisotropy Hshape. For the sample reduces to 0.346 ns. The minimal parameters of DCMP are estimated as K. Juthy et al.: Preprint submitted to Elsevier Page 5 of 7Shape anisotropy effect on magnetization reversal induced by linear down chirp pulse are useful in device application. These findings can be at- tributed to the shape anisotropy Hshapewhich reduces the effective anisotropy and thus reduces the height of the en- ergybarrier. and emission during magnetization reversal. A recent study [43] reported that the temperature effect also assists the magnetization reversal, i.e., it reduces the parameters of microwave field pulse. Therefore, it is desired that, at further (which is beyond the scope of this study). The result of damping dependence of magnetization rever- sal indicates that for a specific AorHshape, there exists an optimal damping situation. Moreover, it is also shown that the required microwave field amplitude can be lowered by applying the direct current simultaneously. The usage of is suggested to achieve cost efficient and faster switch- ing. So these findings may provide the knowledge to fabri- cate the shape of a single domain nanoparticle to generate the magnetic data storage device. 5. (Grant No. 440 EAS) and the Ministry of Edu- cation (BANBEIS, Grant No. S. Sun, C. B. Murray, D. Weller, L. Folks, A. Moser, nanoparticles and ferromagnetic FePt nanocrystal 287 (5460) (2000) 1989-1992. [2] S. Woods, J. Kirtley, S. Sun, R. Koch, Direct investigation of in Co nanoparticle films, Physical Review Letters 87 (13) (2001) 137205. [3] D. Zitoun, M. Respaud, M.-C. Fromen, M. J. Casanove, P. Lecante, C. Amiens, B. Chaudret, Magnetic enhancement in nanoscale CoRh particles, Physical Review Letters 89 (3) (2002) 037203. [4] B. Hillebrands, K. Ounadjela, Spin dynamics in confined Current-induced magnetization reversal in nanopillars with per- pendicular anisotropy, Nature materials 5 (3) (2006) 210-215. [6] A. Hubert, R. Schafer, Magnetic domains: the analysis of (1998). [7] Z. Sun, X. Wang, Fast magnetization switching of stoner particles: A nonlinear dynamics picture, Physical Review B 71 (17) G. Bertotti, C. Serpico, I. D. Mayergoyz, Nonlinear under circularly polarized field, Physical Review Letters 86 (4) (2001) 724. [9] for stoner particles, Physical Review B 73 (9) (2006) 092416. [10] Physical Review B 74 (10) (2006) 104406. [11] S. Okamoto, N. Kikuchi, O. Kitakami, Magnetization switching be- havior with microwave assistance, Applied Physics Letters 93 (10) (2008) 102506. [12] J.-G. Zhu, Y. Wang, Microwave assisted magnetic recording IEEE Transactions on Magnetics 46 (3) (2010) 751-757. [13] (8) (2003) 524-527. [14] K. Rivkin, J. B. Ketterson, Magnetization reversal in the regime using time-dependent magnetic fields, Applied physics letters 89 (25) (2006) 252507. [15] of Applied Physics 105 (9) (2009) 093903. [16] N. Barros, M. Rassam, H. Jirari, H. Kachkachi, Optimal switching of a nanomagnet assisted by microwaves, Physical Review B 83 (14) (2011) 144418. [17] N. Barros, H. Rassam, H. Kachkachi, Physical Review B 88 (1) (2013) 014421. [18] T. Tanaka, Y. Otsuka, Y. Furomoto, K. Matsuyama, Y. Nozaki, 143908. [19] G. Klughertz, P.-A. Hervieux, G. Manfredi, Autoresonant control of the magnetization switching in single-domain nanoparticles, Journal of Physics D: Applied Physics 47 (34) (2014) 345004. [20] Journal of Magnetism and Magnetic Materials 159 (1) (1996) L1. [21] a current, Physical Review B 54 (13) (1996) 9353. [22] of a magnetic multilayer by an electric current, Physical Review Letters 80 (19) (1998) 4281. [23] J. A. Katine, F. J. Albert, R. A. Buhrman, E. B. Myers, D. C. magnetization reversal and spin-wave excitations in Co_Cu_Co pillars, Phys. Rev. Lett. 84 (2000) 3149-3152. [24] X. Waintal, E. B. Myers, P. W. Brouwer, D. C. Ralph, Role of spin- dependent interface scattering in generating current-induced torques in magnetic multilayers, Phys. Rev. B 62 (2000) J. Z. Sun, Spin-current interaction with a monodomain magnetic body: A model study, Physical Review B 62 (1) (2000) 570. [26] J. Sun, Spintronics gets a magnetic flute, Nature 425 (6956) M. D. Stiles, A. Zangwill, Anatomy of spin-transfer torque, Physical Review B 66 (1) (2002) 014407. [28] 69 (9) (2004) 094421. [29] R. Koch, J. Katine, J. Sun, Time-resolved reversal of in a nanomagnet, Physical Review Letters 92 (8) W. Wetzels, G. E. Bauer, O. N. Jouravlev, Efficient with noisy currents, Physical review letters 96 (12) A. Manchon, S. Zhang, Theory of nonequilibrium intrinsic spin torque in a single nanomagnet, Physical Review B 78 (21) Vogel, P. Gambardella, Current-driven spin torque induced by the rashba effect in a ferromagnetic metal layer, Nature materials 9 (3) (2010) 230-234. [33] switching of a single ferromagnetic layer induced by in- plane current injection, Nature 476 (7359) (2011) 189-193. [34] L. Liu, C.-F. Pai, Y. Li, H. Tseng, D. Ralph, R. Buhrman, Spin- torque switching with the giant spin hall effect of tantalum, Science 336 (6081) (2012) 555-558. [35] J. Grollier, V. Cros, H. Jaffres, A. Hamzic, J.-M. George, G. Faini, J.B.Youssef,H.LeGall,A.Fert,Fielddependenceof by spin transfer, Physical Review B 67 (17) (2003) 174402. [36] H. Morise, S. Nakamura, Stable magnetization states under a spin- polarized current and a magnetic field, Physical Review B 71 (1) Z. K. Juthy et al.: Preprint submitted to Elsevier Page 6 of 7

2108.10965

Torikul Islam Md

Z. K. Juthy, M. A. J. Pikul, M. A. S. Akanda, M. T. Islam

Shape anisotropy effect on magnetization reversal induced by linear down chirp pulse

Physica B: Condensed Matter (2022)

10.1016/j.physb.2022.413671

cond-mat.mes-hall

http://creativecommons.org/licenses/by/4.0/

We investigate the influence of shape anisotropy on the of a single-domain magnetic nanoparticle driven by a linear down-chirp microwave field pulse (DCMP). Based on equation, numerical results show that the parameters of DCMP, namely, microwave amplitude, initial frequency and chirp rate, decrease with the increase of shape anisotropy. For certain shape anisotropy, the reversal time significantly reduces. These findings are related to the competition of shape anisotropy and uniaxial and thus to the height of energy barrier which separates the two stable states. The result of damping dependence of magnetization that for a certain sample shape, there exists an optimal at which magnetization is fastest. Moreover, it is also shown that the required microwave field amplitude can be lowered by applying current simultaneously. The usage of an optimum combination of both microwave field pulse and current is suggested to achieve cost efficiency and faster switching. So these findings may provide the knowledge to fabricate the shape of a single domain nanoparticle for the fast and data storage device.

2 Z DZIARSKI ET AL . the propagation time along the jet, and we correct the expres- sions for jet power. Furthermore, we link the dependencies of energy densities on the distance to the observed hard in- verted spectral index, as well as we use additional data from Rodi et al. (2021). This allows us to obtain constraints based on the full spectrum. MAXI J1820+070 was discovered during its outburst in 2018 (Tucker et al. 2018; Kawamuro et al. 2018). The source is relatively nearby, with a distance of based on a radio parallax (Atri et al. 2020). Then, Wood et al. (2021) determined D3:110:06kpc based on the proper motion of the moving ejecta during the hard- to-soft state transition. The inclination of the radio jet is i 645(Wood et al. 2021), while the inclination of the binary is constrained to et al. 2020). The BH mass is given by et al. 2020). The data presented in T21 were obtained during a observational campaign from radio to X-rays performed during a 7-h period during the hard state on 2018 April 12 (MJD 58220). We also use the simultaneous IR and optical data obtained by Rodi et al. (2021). During the campaign, the source was in a part of the initial hard state which formed a plateau in the X-ray hardness vs. flux dia- gram (Buisson et al. 2019). Our theoretical model is presented in Section 2 and Ap- pendix A. In Section 3, we fit it to the data. In Section 4, we discuss various aspects of our results, and show that our formalism based on time lags between different frequencies of the flat spectrum is equivalent to the formalism based on core shifts. We give our conclusions in Section JETS 2.1. Power-law the theoretical interpretation in T21, we con- sider a continuous, steady-state, jet in the range of its distance from the BH where it has constant both the bulk Lorentz factor and the opening angle, i.e., it is conical. We con- sider its synchrotron emission and self-absorption, and as- sume that the hard, partially self-absorbed, part of the total spectrum results from superposition of spectra from with breaks corresponding to unit depth (Blandford & K onigl 1979). We use the formu- lation of the model of K onigl (1981) (which is an extension of the model of Blandford & K onigl 1979 for cases with radio index different from zero) as developed in Zdziarski et al. (2019), hereafter Z19. In this model, the jet is assumed to be laterally uniform, which is a good ap- proximation for i, where is the jet (half) opening angle. We denote the observed and comoving-frame as and0, respectively, and introduce the di- mensionless distance, 0=(1 +zr) ; 1 (1cosi); the cosmological redshift (equal to null in our case), andare the jet bulk Lorentz factor and the velocityin units of the light speed, respectively, zis the distance from the BH, and z0is the distance at which the jet becomes opti- cally thin to self-absorption at all considered frequencies. As in Konigl (1981), we assume the electron differential n( ;), and the magnetic field strength, B, by power-law =z0tan ; n( ;) =n0a p; B() the jet radius and is the Lorentz factor of the emitting electrons in the jet comoving frame, with min  max. The quantities Randzare measured in the local observer's frame, while nandBare given in the comoving frame (for notational simplicity, we skip the primes). For a conserved electron number along the jet and conserved mag- netic energy flux dominated by toroidal fields, we have a= 2 andb= 1, corresponding to the spectral index of = 0, independent of the value of p(Blandford & K onigl 1979). Here, we define by the energy flux density of F/. If either the electron or magnetic energy is dissipated, a > 2, b > 1, respectively. Then, the emission weakens with the distance and the synchrotron spectrum in the partially self- absorbed frequency range becomes harder than in the con- served case,  > 0. The spectral indices of partially self- absorbed and optically thin synchrotron emission are =5a+ 3b+ 2(b1)p13 2a2 +b(p+ 2);  (K onigl 1981; Z19). Using a delta-function ap- proximation to the single-electron synchrotron spectrum at 21(assumed hereafter), the synchrotron frequency for a given and, and its range emitted by the jet Here z0Mis the distance at which the jet ter- minates,his the Planck constant, Bcr= is the critical magnetic field strength, and meandeis the electron mass and charge, respectively. The spectral density of the synchrotron emission for a single by Equation (2) and for given by (see equation A5 of  (7) HereF0is a constant proportional to the bolometric the synchrotron self-absorption optical depth, 0is the break frequency, see Equation (13) below, andJET PARAMETERS 1. An example of the spatial structure of the jet emission at different frequencies for 0= 2104GHz,F0= 300 mJy, b= 1:1,a= 2b,p= 2. The red dots, blue dashes, magenta dots, cyan dashes, and green dots correspond to = 5:25, 25.9, and 5106GHz, respectively. The black solid curve corresponds to =0. The three lowest and two curves end at min>1and atmax, respectively, beyond which there is no emission at those . These values of min,max were calculated for min= 10 ,max= 107GHz,i= 64, = 3 andB0= 104G (which correspond to from averaging the synchrotron emis- sion and absorption coefficients over the pitch angle, C1(p) ; (8) C2(p) (9) (cf. Jones et al. 1974; Zdziarski et al. 2012), where Eis the Euler Gamma function. In the extragalactic case, Dis the luminosity distance. The lower and upper limits of the integral (6) are, min() = (10) max() = Figure 1 shows an example of the spatial de- pendencies of the emission along the jet at different frequen- cies for min= 10 andmax= 107GHz. For min&30, the emission at all frequencies in this case would be in regime only, cf. Equation (16) below. We note that above we have assumed the single-electron is isotropic in the plasma frame, which is strictly valid for a tangled magnetic field. Then, power-law dependencies assuming min = 1 , max=1in the optically-thick and optically thin cases cm-2s-1DFigure 2. An example of the jet synchrotron spectrum for 0= 2104GHz,F0= 300 mJy,b= 1:1,a= 2b,p= 2 , max= 107GHz,i= 64. This spectrum is virtually the shown range of , as long asmin(). The blue curve shows the accurate spectrum of Equation (6), and the red dashes show the approximation of Equation (12). The gradual high- energy cutoff of the accurate spectrum is due to max decreasing with increasing and reaching unity for max. (cf. equation A11 in +b(p+ 1) 2a2 +b(p+ 2)(=0) 4 +b; +b(p+ 1);  2 shows an example comparison of the accurate spec- trum of Equation (6) with these power-law see they are inaccurate around 0as well as close to max, where they fail to reproduce the gradual cutoff of the accu- rate spectrum. While the power-law asymptotic solutions in- tersect at aslightly different from 0, that frequency has no physical meaning since the actual spectrum in that range does not follow the broken-power law form, see Figure 2. We can define a broken-power law approximation by taking the minimum of the two branches in Equation (12). The optical depth along a line of sight crossing the jet spine can be written as (=0;) = (13) where0is defined by (=0;= 1) = 1 . The place = 1, orz=z0, corresponds to the jet being optically thin for all0. There is no synchrotron emission1atz <z 0, 1We note that since the partially optically-thick emission of a jet at z >z 0 would remain almost unaffected if there were still emission following the scaling of Equation (2) at z < z 0(which would, however, decrease the actual value of z0and increase 0), it is also possible to formulate the structure of the partially optically-thick part without invoking z0and0. Such a formulation is presented in Equations (A1-A4) in Appendix A.4 Z DZIARSKI ET AL . and to the onset of the jet emission. The relationship of 0to the jet parameters is given by equation (A8) of Z19. We express it here as a formula for the normal- ization of the electron the fine-structure constant and Tis the Thomson cross section. From Equation (13), the distance along the jet at which(;) = 1 qp+ 4 2a+bp+ 2b2; 2 andb= 1, we haveq= 1 at anyp. This dis- tance is very close to that at which most of the flux at a givenis emitted, which can be defined by the ln, see Figure 1, and can be calculated using Equation (6). For example, at a= 2:2,b= 1:1, andp= 2, that maximum is at 1:19. The emission around the peak has a broad spatial distribution; the 50% values of the maximum flux are reached at = the Lorentz factor responsible for the bulk of emis- sion (16) which is usually weakly dependent on . While the of Equation (6) is valid for any min, the of Equation (12) require minto be by a factor of at least a few lower than for values of of interest (in the range<0) and maxis required to be a factor of a few larger than . If a high-energy cutoff is observed, an follows from it, see Equation (4). If we knowandthin, we still cannot determine the val- ues However, a likely possibility is that the ratio between the electron and magnetic energy den- sities remains constant, i.e., maintaining the same degree of equipartition along the jet, in which case a= 2b. We define an equipartition parameter as the ratio of the energy p6= 2; ln max min; p = second equality in Equation (17) is at z0,upis the particle energy density, kiaccounts for the energy density in particles other than the power-law electrons, in particular in ions (ex- cluding the rest energy), and p>1has been assumed in theexpression for fN. Fora= 2b,eqis constant along the also constant) at zz0, which + 2p) b(p+ 6)2; q =p+ 4 b(p+ 6)2: (19) Below, we use eqanda= 2bto constrain the jet parameters. We note that the case of a > 2requires that either min decreases or the electrons removed from their power-law dis- tribution move to some low energies below min(with negli- gible emission). Since we assume minto be constant along the jet, the latter has to be the case. We next consider the difference between the arrival times of two photons. The first photon, at 1, is emitted toward the observer at1. The second photon, with 2<1, is emitted at2by the same comoving point of the jet after the time
te, which is further downstream in the jet by c
te. Since the jet moves at an angle iwith respect to the line of sight, the distance of the emitting point to the observer will become shorter during this time by c
tecosi. For an observed dif- ference in the arrival times of
ta, the intrinsic separation be- tween the emission points (measured in the local will +zr):(20) At from Equation (15). Here, we have also taken into account the redshift, making this ex- pression correct also for an extragalactic source. Given an observed
ta, Equations (15) and (20) +zr)=t0c 1 (22) wheret0is the implied lag time between the BH and z0, which can be obtained from time lag data if 0andqare known (from the A provides general solutions for z,B,and nto the equations in this section assuming the validity of Equation (12) for  <  0as functions of as well The jet power The jet power can be calculated using standard expres- sions. Note that it is defined in terms of the proper en- thalpy rather than the energy density, e.g., Levinson (2014). Then, the component of the jet power due to both the relativistic electrons and magnetic fields (assum- ing they are predominantly toroidal at z0) including both the jet and counterjet for a= 2bandki= tan )222b:(23)JET PARAMETERS 5 The usable power in cold ions at any z(calculated at z0), is Pi= the density of both electrons and positrons (whose ratio is assumed to be constant along the jet), and positrons only, respectively, e= 2=(1 +X)is the electron mean molecular weight, X(0:7for the is the H mass fraction, mpis the proton mass, andfNis given by Equation (18). This is the power in ions that has to be supplied to the jet, and then it can be the factor (1). Equation (24) neglects the of background electrons being piled up at < min already atz0. On the other hand, a>2at a constant minre- quires that leptons removed from n( ;)atz>z 0do appear at some low energies below min(with negligible emission). We note that in a steady state, 2n+=ne= 2_N+=_Ne, where 2_N+is the total rate of advection upstream of epairs pro- duced at the jet base, and _Neis the total lepton flow tan )2: (25) This (26) Since the pair production rate at the jet base, is independent of , the condition of Pi0can give a lower limit on . In any jet model, the total usable jet power is approxi- mately constrained by the accretion _Mis the mass accretion rate, Lis the bolometric lu- minosity and e0:1is the accretion efficiency. This then gives an upper limit on . The limit _Mc2can be exceeded if the rotation of the BH is tapped, but only by a factor .1.3 and for a maximally rotating BH, see the references in Sec- tion 2.4 below. Finally, we consider the power lost in synchrotron emis- sion. It equals the synchrotron luminosity emitted by both jets in all directions (which is Lorentz invariant). Since 0 depends on the direction and the partially self-absorbed emis- sion is not isotropic in the comoving frame, we neglect its effect and assume the entire emission is optically thin and isotropic in that frame, which is a good approximation for hard electron distributions with p.2:5or so. This p6= 3; ln max min; p = 3. A sketch of of the pair-producing geometry based on fig. 3.9 of Tchekhovskoy (2015), which shows the result of his 3D GRMHD simulation for accretion on a BH with the spin parameter of a= 0:99. In our case the disk is hot in its inner part (up to the radius Rhot) and surrounded by a cool outer disk. We consider epair production within the jet base (shown in green), which is devoid of matter, with the wavy arrows photons. We denote the characteristic jet radius of the pair-producing region as Rjet. In addition, pairs are produced within the hot disk, but it is magnetically shielded from the jet base. The black solid curves show the poloidal magnetic field. where 2b+a>3is assumed. This PSapproximately equals the intrinsic luminosity of both (29) whereFis for the approaching jet, maxis given by Equa- tion (4) and the transformation law is for a stationary jet emit- ting isotropically in its comoving frame (Sikora et al. 1997). For self-consistency of our equations, PSPjis required. 2.3. Pair production As we see from Equation (26), the jet power in ions (given an observed synchrotron spectrum) strongly depends on the abundance of epairs. In the case of extragalactic jets, there are strong indications that they dominate by number, though most of the rest mass is usually still in ions (Sikora et al. 2020). In the case of jets in microquasars, this is uncertain. An important issue for that is the origin of pairs. A likely mechanism is pair production in photon-photon collisions by photons produced close to the BH. Pairs can be produced within the hot flow, e.g, Svens- son (1987). Since the Larmor radius of either a proton or an electron is orders of magnitude lower than the gravitational radius), the magnetic base6 Z DZIARSKI ET AL . of the jet is shielded from the hot plasma, and pairs produced in the accretion flow cannot enter the jet. On the other hand, pairs can also be produced within the magnetic base of the jet, outside the hot plasma (Sikora et al. 2020). There, photon- photon collisions will create epairs in an environment de- void of background matter, thus strongly reducing the rate of pair annihilation (Beloborodov 1999). A possible geome- try (see also Henri & Pelletier 1991; Ferreira et al. 2006) is shown in Figure 3. From an observed hard X-ray spectrum and a radius, Rhot, of the emitting hot plasma (inferred, e.g., from X-ray spectroscopy; Bambi et al. 2021), we can esti- mate the average photon density within the jet base, which then gives us the rate of pair production per unit We approximate the pair-producing volume as two cylinders with the height Rhotand the characteristic radius of the jet,Rjet, i.e.,V= 2R2 jetRhot. We can then the total lepton production rate (30) where the factor A would follow from detailed calcula- tions. Depending on the equilibrium density of the pairs, some of them would annihilate, and some would be advected to the BH, reducing the effective 2_N+. We address this issue for the case of MAXI J1820+070 in Section 3.1. 2.4. The Blandford-Znajek mechanism We can also estimate the jet power in the framework of the model with extraction of the rotational power of the BH (Blandford & Znajek 1977). The jet power in this case de- pends on the magnetic flux, BH, threading the BH (on one hemisphere), which can be written (31) whereBHis a dimensionless magnetic flux. Its maxi- mum value of50 is obtained in magnetically arrested disks (MAD; Narayan et al. 2003; Bisnovatyi-Kogan & Ruzmaikin 1974), as it was found in GRMHD simulations of MAD ac- cretion (Tchekhovskoy et al. 2011; McKinney et al. 2012; see its more accurate value in Davis & Tchekhovskoy 2020). Then it has been found (32) whereais the BH spin parameter and h0:3is defined by of the disk being Tchekhovskoy 2020). This maximum differs from that of Equation (27) by the factor 1:3h0:3a2 . In the spectrally hard state, the disk is most likely hot, in which case h0:31. We can estimate BHusing the magnetic field far from the BH by using the conservation of the magnetic flux. Specifically, we use the expected equal- ity between the poloidal and toroidal field components at the Alfv 'en surface (in the observer's frame), which radius, for strongly magnetized jets, approaches the light 2010). This implies hB0 i (R=R LC)Bp, whereBpis the poloidal field (which has the same value in the comoving and BH frames) and hB0 iis the average toroidal field strength in the comoving frame, de- noted byBin the remainder of this paper. Then, the toroidal field dominates at and, presum- ably, atzz0. The magnetic flux using this method was determined for a sample of radio loud active galactic nuclei in Zamaninasab et al. (2014) and Z15. We use the as derived in +)1=2 `a; (33) which allows us to estimate BHfor a givenaby setting j=  BH. HereRH= [1 + (1a2 )1=2]Rgis the BH hori- zon radius,`.0:5is the ratio of the field and BH and sis the scaling factor relation between the jet opening angle and the bulk Lorentz factor (Komissarov et al. 2009; Tchekhovskoy et al. 2009), limited by (34) Here,is the magnetization parameter, which is defined as the ratio of the proper magnetic enthalpy to that for the rest the rest-mass density, and 4=3<  < 5=3is the particle adiabatic index. The second equality relates  toeqassuming that the only ions are those associated with the power law electrons (i.e., neglecting the possible pres- ence of ions associated with electrons with < min, e.g., with a quasi-Maxwellian distribution). For p>2and a Then, we TO MAXI J1820+070 Here, we apply the model of Section 2 to the source. We use the VLA fluxes at 5.25, 7.45, 8.5, 11.0, 20.9, 25.9 and the ALMA flux at 343.5 (from table 1 in T21). We also use the IR flux at 1:4105from VLT/HAWK-I, and the optical flux at 3:9105GHz from NTT/ULTRACAM (T21), and the 13 fluxes between 1.37 and 7:00105GHz from the VLT X-shooter and the INTEGRAL /OMC flux at et al. 2021). All of the IR/optical fluxes have been de-reddened with E(BV) = 0:18(Tucker et al. 2018), as assumed in T21. We use the time lags between 25.9 GHz and lower frequencies, 11.0 GHz and lower frequencies, and be- tween 343.5 GHz and lower frequencies. The lags are given in tables 3 and 4 of T21. This gives us 23 spectral measure- ments and 14 time lags. We present analytical and in Sections 3.1 and 3.2, respectively. We assume the ratio between the electron and magnetic energy densities to be constant along the jet, i.e., a= 2b. In our fits below, we use the Markov-Chain Monte Carlo (hereafter MCMC) technique with wide uniform priors, seeJET PARAMETERS 7 b= = 297 = = = 240 320 30 40 50 4. MCMC fit results for the seven quantities, which require only the assumptions of a= 2b, see Section 3.1. Here and in Figure 8 below, the panels show the histograms of the one-dimensional posterior distributions for the model parameters, and the two- parameter correlations, with the best-fitting values of the parameters indicated by green lines/squares. The best-fit results for fitted quantities are taken as the medians of the resulting posterior distributions, and are shown by the middle vertical dashed lines in the distribution panels. The surrounding vertical dashed lines correspond approximately to a 1uncertainty.8 Z DZIARSKI ET AL . Figure 5. The radio-to-optical spectrum from T21 (VLA, NTT/ULTRACAM; red error bars), the 339 MHz measurement (VLITE, magenta error bar; Polisensky et al. 2018), and from the VLT/X-shooter (blue error bars) and the INTE- GRAL /OMC (cyan error bar) as obtained by Rodi et al. (2021), but with the de-reddening correction for E(BV) = 0:18(Tucker et al. 2018). The error bars for the radio and sub-mm measurements of T21 are the square roots of the squares of their statistical and sys- tematic errors, and 10% systematic errors are assumed for the IR and optical measurements. The spectrum above 5 GHz is fitted by the jet model of Equation (6) using the best-fit parameters shown in Figure 4 (cyan dashed curve) and a phenomenological power law approximating the disk component with disk= 0:53(green solid curve). The former is virtually independent of minwithin the ranges obtained in the full fits (Equation 10; Section 3.2), so it can be assumed to be unity. The sum is shown by the solid black curve. The corresponding asymptotic optically thick and optically thin spectra of Equation (12) are shown by the magenta dotted lines. T21 for details. We assume i= 645(Wood et al. 2021) with a Gaussian prior, and D= 2:960:33kpc (Atri et al. 2020) with a Gaussian prior, but truncated at the upper limit ofDmax= 3:11kpc found by Wood et al. (2021). We assume the observed spectrum is from the approaching jet only. Ati64, this assumption is satisfied only roughly. The ratio of the fluxes in the optically- thick part of the spectrum is given by [(1 follows from Zdziarski et al. 2012), which is7 at the fitted p2:21(see below). 3.1. The initial fit and analytical estimates We can solve for only assum- inga= 2b. From the measured fluxes, we obtain Equations (6-7). We find the fitted spectrum is very insensitive to min(), Equation (10), as long as it is low enough. We can just use any low value of it, or just as- Figure 6. The time lags measured by T21 vs. the distance in units of z0between the emission points for the part of the spectrum. The blue and red symbols correspond to the lags between 343.5 GHz and radio fre- quencies (5.25-25.9 GHz), and within the radio frequencies, re- spectively, where we assumed a constant ratio between the elec- tron and magnetic energy densities, = 2:32104GHz. The diagonal line gives the relationship between the two quantities corresponding to t0= 0:79s, see text. sumemin= 1, and check a posteriori the consistency of the choice. Similar to Rodi et al. (2021), we find the presence of an additional hard component beyond max, apparently due to the emission of an accretion disk. Given the limited range of the fitted frequencies, we fit it as a power Then, using the obtained values of b,pand0, we can fit the time lags using Equations (19) and (22). However, with the MCMC technique, we fit the flux and time-lag measurements The fitted parameters and the correlations between them are shown in Figure 4, and the parameters are listed in Table 1. The best-fitting values are given as the median of the resulting posterior distributions, and the lower and upper un- certainties are reported as the range between the median and the 15th percentile, and the 85th percentile and the These uncertainties correspond We use these best-fit values as well as the best- fit values of Dandiin our estimates in this 5 shows the observed average radio-to-optical spec- trum fitted by the above model. The best-fitting spectral in- dices in the optically thick and optically thin regimes are respectively. We show the the- oretical spectrum calculated by integrating Equation (6) and the asymptotic optically thick and thin spectra of Equation (12) for this fit. We then show the time lags in Figure 6, where we plot the values of the measured
taagainst the separation in units, , using Equations (15) and (22). AtJET PARAMETERS 9 Table 1. The basic parameters of the jet in MAXI J1820+070. b p  mJy 104GHz mJy fits are done with the MCMC method assuming a= 2b.Fdiskgives the flux density at 105GHz. 1 mJy 1026erg cm2s1Hz1. the best-fit values of bandp,q0:883, see Equation (19). The actual lags have to follow a single dependence relating the physical separation between the emission points to
ta, which is shown by the diagonal line showing the linear re- lationship between to the best-fit value oft0, see Equation 22. We see a certain offset between the points corresponding to the lags between the sub-mm frequency of 343.5 GHz and 6 radio frequencies (blue error bars), and the lags measured between the radio frequencies (red error bars). This may be related to the different methods used in T21 to determine those. On the other hand, the offset is significantly reduced for q= 0:8, which value of q, how- ever, is not compatible with 0:25. This may indicate that the jet is more complex than we assume, e.g., either ,or bare not constant at zz0. Our formalism assumes the lags correspond to propaga- tion of perturbations between different values of zat the jet speed,. With this assumption, we obtain z0as a function of , see Equation s): (36) We then use the solutions obtained in Appendix A assum- ing min= 3,ki= 0. However, we only know see Equation (5). Since the solutions depend on max relatively weakly, we assume here the best-fit values ofB0= 104G and = 2:2obtained in Section 3.2 for min= 3, which yield max= 125 , which we will use here- after in this subsection. Using Equation (A5), we obtain at the best 1,0:530:67. Next, give at the best Equation (37) shows that we cannot determine both andeven assuming a value ofeq(on which depends very weakly). We can also calculate the Thomson scattering optical depth along the jet radius at zz0, which equals T() =Tn0fNz0tan 7. The locations of the emission at the observed from the break in the power spectra with the assumption of zb=c=f break , shown as their ratio to the locations based on time lags and the slope of the partially self-absorbed spectrum, = 3 andi= 64). Ati= 64and = 2 , 3, 4, we have 0:81, 0.57, 0.43, and, at eq= and T(= respectively. The val- ues ofz0correspond an assumed M= 8M. We find that ,B0andz0depend on for1:5..5. We determine the typical Lorentz factors, , of relativis- tic electrons giving rise to the emission at , which in the partially self-absorbed regime originates mostly from z, see Equation (16). We (41) In order to obtain a power-law emission in that regime, we need minto be a factor of a few smaller. Thus, we the validity of the model. The maximum corresponds to the fitted max, Equation (5). From that, from123 to 147 for increasing from 2 to 4. Combining this with the values of Tfrom Equation (40), we find that the power in the synchrotron is relatively similar to that in the synchrotron can then consider implications of the break in the power spectra for different frequencies mea-10 Z DZIARSKI ET AL . sured by T21. For those power spectra, most of the variabil- ity power per lnfoccurs atffb, with the variability at higher frequencies strongly damped, see figs. 3 and 5 in T21. We define the distance, zb, as that covered by a jet element moving with the jet velocity during the b(): (42) We can compare it to the distance along the jet from the BH up to the location of the peak emission at , 15, 21). In our model, Then, this ratio depends only weakly on (or); ati= only from 1 at 1to 0.56 at1. This implies that this correlation cannot be used to determine the actual bulk Lorentz factor of the jet. Figure 7 shows zb=zvs.zfor = 3 . We see an ap- proximately constant ratio of zb=z1:5-2. Therefore, zb is proportional and close to the travel time along zin all of the cases. A possible explanation of the damping of the vari- ability at frequencies >c=zappears to be a superposition of the contributions to the emission from different parts of the region dominating at a given , which is/z, as shown in Figure 1. The peak of dF=d lnzforp= 2:21is at1:15z and its width defined by dF=d lnzdecreasing to the 50% of the peak is (0:65-3:16)z. Thus, if different parts vary in- dependently, the variability will be damped at f&c=(2z), as the observed radio/IR variability can be driven by the variable power supplied from the vicinity of the BH with a wide range of frequencies (Malzac 2013, 2014) and then transferred upstream. Then, the travel time can act as a low-pass filter, removing most of the variability at fre- quenciesf >c=z. This can happen due to damping of per- turbations along the jet due to some kind of internal viscosity, e.g., collisions between shells within the jet moving with a range of velocities (Jamil et al. 2010). The process would be then analogous to viscous damping in accretion disks, with a period shorter than the signal travel time across the disk are strongly damped (Zdziarski et al. 2009). This picture is also compatible with the integrated of the power spectra (RMS) decreasing with shown in fig. 5 of T21). This means increas- ing the distance travelled along the jet leads to the note that the break frequencies in the power spectra of T21 have been defined by choosing a specific, and not 2T21 assumed z=zbc=f b(), which they used as the final condi- tion determining the jet parameters. Thus, they transformed the frequency to the jet frame ,fb=, and multiplied the resulting time scale,=fb, by the jet velocity in the observer's frame ,c, which does not appear to be correct. We note that in the present case we consider the light curve originating from a fixed region of the jet around z. While the plasma in that region is moving, two adjacent maxima in the observed light curve are emitted from the same region in the frame connected to the BH, which is the same frame as the observer's one (in the absence of a redshift). Thus, a frequency inferred from the variability power spectrum should not be algorithm, as well as the obtained values of fbare close to the minimum frequency at which the power spec- trum is measured for f < 10GHz, which limits the accu- racy of the determination of those fb. Also, while the damp- ing of variability above c=zclearly occurs, details of the physics behind it remain uncertain, and the damping could start of exactly c=z. Summa- rizing, our results are completely compatible with the vari- ability damping at time scales shorter than the light/jet travel time acrossz. However, unlike our previous estimates from the observed spectrum and time lags, which are based on a relatively rigorous and well-understood model, the detailed cause of the connection between the break frequencies and the distance along the jet remains uncertain. We can also consider the prediction of the location of the bulk of the 15 GHz emission, (at = 3 , but weakly dependent on it), with the jet angular size at this frequency from the VLBA observation on 2018 March 16 (MJD 58193), reported in T21 as 0:520:02mas. the deprojected size is The total flux density at 15 GHz was mea- sured However, the VLBA ob- servation was 27 d before the radio/sub-mm ones. On MJD 58220, our best-fit spectral model yields the framework of the continuous conical jet model, we (Equation A2; Zdziarski et al. 2012). Thus, for p= 2:2we predict the size at 15 GHz on MJD 58220 being larger than that on MJD 58193, namely 41013cm. While some- what larger than the above z, this size appears consistent with it since the peak of dF=d lnzforp= 2:21is and that spatial distribution is broad and skewed toward higher distances, see Figure 1 and the discussion of it above. We then estimate the rate of pair production. For MAXI J1820+070, pair production within the hot plasma was cal- culated by Zdziarski et al. (2021) based on the by INTEGRAL in the hard state. That spectrum was measured up to 2 MeV , well above the pair produc- tion threshold of 511 keV , and modelled by was found that an appreciable pair abundance can be ob- tained only provided the hard X-ray source size is as small as severalRg, while the spectroscopy based on the of the fluorescent Fe K line indicates a size of &20Rg. Then, the pair abundance within the is very low. However, as discussed in Section 2.3, pair the jet base can be much more efficient. To calculate it, we adapt the results of Zdziarski et al. (2021). We mod- ify their equation (1) to calculate the photon density above the hot disk, dividing the total rate of the photon both sides). We then use this photon density in equation (3) of that paper for the spectral param- eters of average spectrum (table 2 in Zdziarski et al. 2021). This gives the pair production rate per unit volume. With theJET PARAMETERS 11 assumptions as in Section 2.3, we is then balanced by the sum of the rates of pair annihi- lation and pair advection. Using formulae in Zdziarski et al. (2021), we have found that pair annihilation can be neglected for the advection velocity of &0:1. It appears that such a velocity can be achieved due to the net momentum com- ponent of the pair-producing photons along the zaxis, see Figure 3, and due to pair acceleration by radiation pressure of the disk photons (Beloborodov 1999). Thus, while some of the produced pairs will annihilate (and a small fraction will be advected to the BH), a major fraction of the produced pairs will have a sufficient net bulk velocity to escape upstream. Then, the lepton flow rate through the jet, Equation (25), for min= (44) where Equations (36), (37), (39) have been used and the assumes 1. Comparing with Equation (43), we find _Ne>2_N+at any forRhot= 20Rg,Rjet= 10Rg and min= 3. Thus, at these parameters the plasma is never composed of pure pairs. If we as- sume either Rjet= 15Rgor min= 10 , we find _Ne= 2_N+ at2, which thus represent the minimum possible for these parameters. While the hot disk and jet radii and min are poorly constrained, we consider the fact that the numbers in Equations (43) and (44), obtained with completely differ- ent physical considerations, are of the same order of magni- tude, to be highly remarkable and indicating that indeed the two rates may be similar in this source. Then, the jet can contain a large fractional abundance of pairs, and they can dominate by number over the ions. The pairs produced in the jet base and then advected to large distances will eventually leave the jet and enter the ISM. Our estimated rate, 1040-1041s1, is lower than the ap- proximate estimates of the positron production rates in mi- croquasars of Guessoum et al. (2006). It is also a few or- ders of magnitude below the total rate of pair annihilation in the Galaxy, which has been estimated by Siegert et al. (2016) With the past and present X-ray monitors (ASM, Levine et al. 1996; MAXI, Matsuoka et al. 2009; BAT, Barthelmy et al. 2005), we can detect all of the outbursting accreting BH binaries with luminosities of more than a few percent of the Eddington luminosity. Based on the MAXI data, there is 0.5 source in outburst at given time (the MAXI team, private communication). Thus, the av- erage contribution of such sources to the Galactic to be negligible, contrary to earlier estimates (Gues- soum et al. 2006; Weidenspointner et al. 2008). Next, we calculate the jet power. The power in the rela- tivistic electrons and magnetic fields, Equation (23), s13 + increases very fast with , approximately as /4at 1. Ateq= 1, = 3 , this power is2:21037erg s1. The power associated with the bulk motion of cold (24), (26), s1(1) first term is approximately /3(1). To constrain Pjby the accretion power, we use the es- timate of the hard-state bolometric flux of cm2s1(Shidatsu et al. 2019). This yields L 1:5(D=2:96 kpc)21038erg s1:(47) For the default parameter values, Pj._Mc2implies .3:3. If pair production is efficient enough, we also have a lower limit on from the requirement of Pi>0. The allowed range depends significantly on the assumed parameters, E.g., at min= 10 10Rg,&2:4is required. We can then compare the total jet power, Pj, with the syn- chrotron power. At the low max implied by the max fit- ted to the spectrum, we find PSPjalways. For the maximum allowed = 2 . On the other hand, we weakly depending on ei- theror min. Thus, the synchrotron emission can be entirely accounted for by the power in electrons and mag- netic fields at z0, and most of the decline of PB+Pewith the distance can be due to the synchrotron losses. How- ever, we may see that the decline of than that of the synchrotron power. If the former would be just to the synchrotron emission, we would dPS=d= 0, while the former and the latter terms This implies either some electron re-acceleration at z >z 0at the expense of Pi, or more complexity of the actual physical situation, with the initial energy loss in the flow being faster and followed by a slower one. In the framework of models with the jet dissipation mech- anism being the differential collimation of poloidal the obtained 1indicate the jet low. Using Equation (34), we have (at 1) = (48) Ateq= 1 and assuming s= 0:6(as found as the average value for a large sample of radio-loud AGNs by Pjanka et al. 2017), we obtain This can be com- pared tofrom its definition, Equation (35), which + (49)12 Z DZIARSKI ET AL . G = = (cm) log[z0] (cm) = 10 = (erg s-1) log[Pj] (erg s-1) = (erg 150 180 gmaxgmax= 8. (a) The MCMC fit results for 3 ande= 0:3. The meaning of the panels and lines is the same as in Figure 4. See Section 3.2 for details.JET PARAMETERS 13 G = = (cm) log[z0] (cm) = 10 = (erg s-1) log[Pj] (erg s-1) = 2 3 (erg 240 320 400 gmaxgmax= 8. (b) The MCMC fit results for min= 10 ande= 0:1.14 Z DZIARSKI ET AL . Table 2. The parameters of the jet in MAXI J1820+070 other than those given in Table 1. mine  log10z0B0 log10Pj max cm 104G erg s1 3f 0.3f 0.1f fits are done with the MCMC method for the assumed the fixed (marked by 'f') minimum electron Lorentz factor, min, and the accretion efficiency, e. In the absence of pairs, 0:0078=eq. Comparing the two estimates of , we see it requires &3ateq= 1. However, the actual value of sis uncertain, there could be ions associated with background electrons piled up at < min, and, importantly, eqcould be1. Still, the implied by Equation (48) disfavors the case of strong pair dominance, we can calculate the magnetic fluxes in the model with extraction of the BH rotational power. The jet magnetic flux from Equation (33) with z0B0from Equations (36) and (38) is cm2)s[1 + is5:31021G cm2fora= 1, = 3 ,`= 0:5, s= 0:6,eq= 1. The flux threading the BH, Equation (31) with _Mestimated as above from L, cm2)BH 50D 3 8Mwas assumed for both ( /M). At BH= 50 and the assumed parameters, the two fluxes equal for a0:7. We consider the close agreement of the above two estimates to be very remarkable. They are based on completely different physical Thus, our results are consistent with the jet being pow- ered by the BH rotation and the accretion flow being mag- netically arrested. In this case, the jet power is maximal and given by Equation (32). However, we have found that if pairs dominate in the jet, Pj_Mc2. This requires either a1, the magnetic field in the flow is weaker than that in a MAD, or that some assumption in the model with extraction of the BH rotation power, e.g., the ideal MHD, are not satisfied. 3.2. Numerical estimates In order to solve directly for the physical jet parameters and their uncertainties, we use again the MCMC method. In the fits shown in Figure 4, we fitted the minimum assumption of a= 2b, and, in particular, without the need to specify the value of . Now we fit for all of the parameters. However, since the solution given in Appendix A is given in terms of we fit for the former (which yields maxgiven the values of ,iandB0, see Equation 5). In particular, we determine from Equation (A5), z0from Equation (21) and B0from Equation (A6). That (which is then a free parameter) and min. We fixeq= 1 andki= 0. However, in order to be able to constrain rather than have it entirely free, we include fur- ther constraints, using the pair production rate of Equation (43) and requiring 2_N+=_Ne1in Equation (26) and from the maximum possible jet power, Pj_Mc2, Equations (23- 27). These constraints require specifying RhotandRjet, the bolometric luminosity, L, and the accretion efficiency, e. We then solve simultaneously for all of the parameters, Those parameters have now values similar to those shown in Figure 4, and we thus do not show them again. In the solution, we sample Dandias described at the be- ginning of Section 3. We assume L= 1:51038erg s1, X= 0:7,Rhot= 20Rg,Rjet= 10Rg(forM= 8M). We show the resulting posterior distributions for two cases with ( min= 3,e= 0:3), and with ( min= 10 ,e= 0:1), in Figures 8(a), (b), respectively, and list the fitted parameters in Table 2. We see that the obtained ranges of anddepend on those two sets of assumptions, being larger for for the lat- ter case. The allowed maximum jet power is /1 e, and then it is higher in case (b). On the other hand, the obtained andB0104G depend on those assumptions. For the sake of brevity, we have not shown the effect of changing the values of Rhotand Rjet. For example, for Rhot>20Rg, pair production will be less efficient, which would in turn allow fewer leptons in the flow and lower values of , see Equations (43-44). Thus, we cannot conclusively rule out values of .1:5. Then, values ofhigher than those obtained above would be possible 8(a-b) also show max and the pair The former ir relatively tightly constrained in the 110-150 range. The latter is strongly anticorrelated with the jet power, being low at the maximum Pjand close to unity at the minimum Pj, in agreement with our considerations in Section 3.1. We find the synchrotron power, Equation (28), as in Section 3.1, and thus the jet ra- diative efficiency, PS=Pj, is low. In our fits, we have not used constraints from the break frequencies in the power spectra and from the jet spatial extent measurement, following our discussion in Section The location of the dissipation zone Our results indicate that the jet synchrotron emission, and thus electron acceleration, starts at the distance of z0JET PARAMETERS 15 3104Rgaway from the BH. This is similar to the situa- tion in blazars, in which the so-called blazar zones are found to be at distances between that of the broad-line regions and of the molecular torii, i.e., & Sikora 2016 and references therein). Then, a measured in the blazar BL Lac yields z0of several et al. 2008). The region upstream to the onset of the emission is called an acceleration and collima- tion zone (ACZ). This similarity of z0=Rgin jets in blazars and in binaries accreting onto BHs is consistent with the scale invariance of relativistic jets (Heinz & Sunyaev 2003). of emission closer to the BH is also con- sistent with the jets formed by magnetic processes and be- ing initially Poynting-flux dominated (Blandford & Znajek 1977; Blandford & Payne 1982; Blandford et al. 2019). The onset of the electron acceleration is often associated with the presence of a standing shock. For example, Ceccobello et al. (2018) invoke recollimation shocks at a fast A requirement for a formation of a shock is  < 1, which agrees with our estimates of 1in the region, see Equations (48-49). An alternative ex- planation for the presence of an ACZ in BH X-ray binaries invokes the colliding shell model, in which the dissipation re- gion is associated with the distance at which the shells begin to collide (Malzac 2013, 2014). Our determination of z0is within that found for the jet of the BH X-ray binary MAXI J1836-194 of model of Lucchini et al. (2021). On the other hand, our values ofz0are significantly larger than that of from lags of the IR/optical emission with respect to X-rays measured in the BH X-ray binaries GX 339-4 and V404 Cyg (Gandhi et al. 2008, 2017; Casella et al. 2010). We point out that the analysis of T21 found the the optical light curve ( 3:9105GHz) and that of X-rays (shown in their fig. 8) to be rather complex in MAXI J1820+070. The dominant feature was an on the zero lag, on top of which there was a much weaker positive correlation with the peak at the optical vs. X- rays of 150+500 700ms, and a rather low relative amplitude of the correlation peak of 0.03, in contrast wit the IR/optical rms variability of a few tens of per cent (T21). Given that weak- ness and the very large error bar on the lag, we have not used that constraint in our modelling. The optical/X-ray cross cor- relation in MAXI J1820+070 was later studied in more detail by Paice et al. (2021). They found that the over 2-s segments of the light curves shows the pos- itive correlation to dominate, with the optical lag of 150- 200 ms. Curiously, this lag was almost the same for all six epochs they studied, in spite of a spread of the X-ray flux up to a factor of three (see fig. 1 in Paice et al. 2021). An interpretation of the positive lag of the optical emis- sion with respect to X-rays as due to the signal propation from the BH vicinity to the region of the onset of the jet dis- sipation implies t0150-200 ms. Thus, this region would lie at a significantly lower distance than that found in our study, where see Figure 4. We have found that including this constraint would worsen our fit to the re-maining observables, especially to the 343.5 GHz flux. Given the weakness of the correlation, the complexity of the shape and the unexplained constancy of the pos- itive lag component, we consider that alternative interpreta- tion to be uncertain. A model including that lag would have to also account for the dominant anticorrelation, possibly due to the effect of synchrotron cooling on the X-ray spectrum emitted by the accretion flow (Veledina et al. 2013), which is beyond the scope of the present paper. 4.2. Electron energy losses and have parametrized the electron distribution as a power- law function of the distance, and assume that a constant shape. Such a situation requires the elec- tron energy losses are moderate and satisfying _ / . We compare here the time scale for synchrotron energy (52) with the time Z19). We consider the solution in Section 3.1 for = 3 . At 30, which corresponds to the bulk of the partially self-absorbed emission, tsynis shorter than tadfor .3, and it is3 times shorter at z0. This implies that electrons responsible for the optically-thin part of the syn- chrotron emission have to be re-accelerated above z0. Calculating the electron distribution as a function of the distance as well as accounting for the slope of the spectrum at  <  0is relatively complex, a kinetic equation with both losses and spatial advec- tion (e.g., Z19). This also requires taking into account losses from Compton scattering of synchrotron photons as well as the reduction of the electron energy loss rate due to self- absorption (Ghisellini et al. 1988; Katarzy 'nski et al. 2006). Such a model is beyond the scope of the present work. 4.3. Comparison with other jet models of accreting black holes The main independent study of the hard-state jet of MAXI J1820+070 is that by Rodi et al. (2021). They had at their disposal only the spectral data. They assumed R=z = to  = 5:7, which is much larger than that found by us. They assumed = 2:2following the result of Bright et al. (2020) for the ejection during the hard-to- soft transition, but we note that of the hard-state jet is likely to be different. The jet model of Rodi et al. (2021) is also different from ours, and considers an initial accel- eration event followed by synchrotron cooling assuming no adiabatic losses (following Pe'er & Casella 2009). They do not show the spatial structure of their jet model, and thus we are not able to check whether that model would agree with our time-lag data. Still, they obtain relatively similar16 Z DZIARSKI ET AL . values of the distance of the onset of electron and the magnetic field strength at very long time lags found in T21 unambiguously show that the radio/sub-mm emission originates at size scales sev- eral orders of magnitude higher than Rg. The time lags found to be approximately Knowing the break frequency, 0, above which the entire synchrotron emission is optically thin, we can ex- trapolate this correlation and find the location This is found to be z03104Rg, with the uncer- tainty of a factor of at most a few. This rules out jet mod- els predicting the onset of the synchrotron emission to in an immediate vicinity of the BH, for example that described in Giannios (2005) (based on the model of Reig et al. 2003). The model of Reig et al. (2003) was developed in order to explain time lags of harder X-rays with respect to softer ones by Compton scattering. For that reason, the authors in- voke a rather massive and extended jet, where multiple scat- tering of disk photons place. Consequently, this model (fur- ther developed in a number of subsequent papers of those authors) requires a rather large rate of the electron flow. For the parameters of Giannios (2005) (similar to those in Reig et al. 2003), the base of the jet has a radius of R0= 100Rg with the electron density of 8M), which electrons flow upward with = 0:8. For matter with cosmic composition, this corresponds to the mass flow in the both jets of 2:41020g/s. On the other hand, for MAXI J1820+070 i.e., two or- ders of magnitude less, which rules out this case. We can also consider the case in which the leptons in the flow are pairs. In that case, 2_N+= This is 3-4 orders of magnitude higher than the pair production rate in this source calculated in Equation (43), which rules out this case too. Since the hard state of MAXI J1820+070 is rather similar to that of other X-ray binaries with accreting BHs, we find we can rule out this model in general. 4.4. Other constraints and caveats Our model is based on that of Blandford & K onigl (1979) and K onigl (1981), and it assumes uniform scaling of the emission regions, through the coefficients aandb. As we see in Figure 5, this model does not account for the observed flux at 339 MHz, measured by Polisensky et al. (2018). This hints for the decline of the energy content in the and magnetic field being initially faster (responsi- ble for the emission closer to z0) and then slower (responsible for the emission farther away from z0). This would introduce more complexity in the modelling, and is beyond the scope of this work. On the other hand, the flux at 339 MHz could be due to another component, in particular a pair of radio lobes at the jet ends. An assumption of our model is that the bulk of the emission at a given distance in the partially of the spectrum occurs at a corresponding to 1. As we have found out, this corresponds to the by electrons with 30. If the minimum Lorentzfactor of the electron distribution were higher, min>30, then the emission at a given distance in that part of the spec- trum would be dominated by the electrons at no contribution from assumed the jet is already fully accelerated at z0and then does not decelerate. This may be not the case, and the available data do not exclude that. The jet model of Z19 al- lows for a variable , and we could use some of (z)and refit our data (as done in Zdziarski 2019 for another source). This would, however, introduce more free parameters, and make the resulting fits less constrained than in the present case. We have also considered the steady state, while variability has been observed. However, the was0:3at the sub-mm range and much less than that in the radio regime. Thus, the variability can be considered as a small perturbation of the steady state. We also use a -function approximation to the which is a good approximation for power-law parts of the spectra, but becomes less accurate at cutoffs, given the single-electron synchrotron spectrum is quite broad (Ginzburg & Syrovatskii 1965). We assume the of a single electron is isotropic in the plasma frame, which is valid for a tangled magnetic field, while we assume a toroidal field in some of our equations. Furthermore, we assume a sharp cutoff in the electron distribution at max. While this is not realistic, the actual form of the cutoff de- pends on details of the acceleration process and is Thus, our determination of maxbased on the observed cutoff in the optical range is only we have used our self-consistent set of equations, in which the slope of the partially self-absorbed part of the syn- chrotron spectrum is connected to the rate of decline of the energy density along the jet. The latter determines the rela- tionship between the characteristic emitted frequency and the distance (Equation 15), and thus the time-lag vs. A significant discrepancy between the spectral slope and time lags vs. frequency was found in Cyg X-1 (Tetarenko et al. 2019). In our case, the two are in an approximate mu- tual agreement. We have found that the break frequencies in the are compatible with the origin of the emis- sion atz, which are roughly equal to c=f b()for < 0. However, an increasing fbwith increasing is also observed for the IR and optical data (see fig. 5 of T21), for which  >  0. In our jet model, the emission at  >  0is synchrotron from the entire part of the jet at z >z 0, which implies z> 0=z0. Thus, we expect that the above scaling of fb/z1 no longer holds at  > 0. Then, the IR/optical variability at high Fourier frequencies may be mostly due to electron energy losses and the Section 4.2) rather than due to propagation of some dis- turbances from z<z 0. As shown in fig. 8 of T21, the optical and IR light curves are tightly correlated, with no measurable lag ( spite of a relatively large disk contribution in the optical range ( 3:7105GHz), as shown in Figure 5. This shows the the disk contribution is constant on the studied time scales,JET PARAMETERS 17 which is consistent with the rms variability in the optical range reduced with respect to the IR one, see fig. 5 in T21. As shown in T21, the upper limit on the lag is consistent with the synchrotron energy losses at the magnetic field which agrees with our determination of B0. 4.5. Relationship to core shifts Time lags are closely related to core shifts,
, which are angular displacements of the radio cores, observed at fre- quencies where the synchrotron emission is partially self- absorbed. They are commonly found in radio-loud active galactic nuclei (e.g., Pushkarev et al. 2012). The physical cause of the physical displacement along the jet, z2z1, is the same for both the core shifts and time lags; only the methods to determine it are different. Using equation (4) in Z15 and Equation (20), the relationship of +zr) sini D(1cosi): (54) We can then relate Equations (20- 21) and (A2). We can estimate B0using the core-shift method, but only assuminga= 2,b= 1, which parameters have been as- sumed in published core-shift studies, including Z15. Us- ing equation (7) of Z15 and Equation (54), we obtain B0 1:0104G atp= 2,= 1 andeq= 1, in a good agree- ment with our estimate of Equation (38). We can also obtain B0from equation (8) in Z15 without specifying have based our study on the results of a campaign observing MAXI J1820+070 when it was close to the peak of its luminous hard spectral state, at 15% of the Eddington luminosity. We have used mostly the data published in T21 as well as the IR/optical spectrum from Rodi et al. (2021). Our main conclusions are as follows. A major result of our study is the es- timate of the distances at which the jet emits below the ob- served break frequency, based on the time lags between vari- ous frequencies. These distances are definitely several orders of magnitude higher than Rg. By extrapolating the correlation of the time lags with the the photon wavelengths, the place where that emis- sion begins can be estimated to be at the distance of several tens of thousands of Rgfrom the BH. This value of that dis- tance agrees with the corresponding finding of Rodi et al. (2021), based on spectral modelling alone. We then use the classical model of Blandford & K onigl (1979), as formulated in detail in later works, to determine the parameters of the jet emitting in the The model assumes the hard inverted spectrum is due to the superposition of locally-emitted spectra that are self- absorbed up to some frequency and then are optically thin. Apart from some details, this is the same model as that used by T21. The values of the jet parameters obtained by us up- date those of T21, which suffered from some errors (see Sec- tion 1 and Appendix A). Our analysis is also broader thanthat of T21, utilizing constraints from the break frequency and the optically thin part of the spectrum. By applying the model to the data, we find we cannot uniquely determine the jet bulk Lorentz factor, , which then needs to be specified as a free parameter. However, it can be constrained from above by the requirement of the jet power being less than the accretion power. It can also be constrained from below by estimating the epair production rate in the base of the jet and comparing it to the flux of erequired to account for the observed synchrotron emission. We then use a Bayesian MCMC method to determine all of the jet param- eters, and find the most likely range of 1:7..4:1. We find the jet half-opening angle, constrained to0.6-2. The onset of the emission is at z03-41010cm, where the magnetic field strength is B00:8-1:4104G. The to- tal jet power is between s1. The jet composition is strongly correlated with Pj, being mostly pairs at the lower limit and mostly eand ions at the upper limit. The optical spectral data imply a rather low value of the maximum electron Lorentz factor, of min110-150. In order to explain the possible presence of epairs in the jet, we calculate the rate of pair production in the jet base in immediate vicinity of the hot accretion flow. We use the measurement of a power-law spectral component extending at least to 2 MeV in the same state of MAXI J1820+070. This rate depends on the geometry, see Figure 3, but we find it to be of the same order as the rate of the electron flow through the part of the jet, both We find this coincidence to be a strong argument for the presence of pairs in the hard-state jet of MAXI J1820+070. We have also estimated the contribution of transient accreting BH binaries to the total rate of in the Galaxy and found it to be very small. We also consider the possibility of the jet power being lim- ited by the power from the rotation of the BH in the pres- ence of magnetically arrested accretion flow. To test it, we calculate the magnetic flux of the jet in the emitting region and that threading the BH. We find them to be very similar, 1021G cm2, which remarkable coincidence argues for that scenario. Then, the jet is initially magnetic, slow, and not emitting, and accelerates and dis- sipates its energy only at large distances, in agreement with our finding of the emission being far from the BH. We find the synchrotron power to be only a small fraction, 102, of the total jet power. On the other hand, the syn- chrotron power is very similar to either the electron or mag- netic powers at the onset of the dissipation, showing that de- cline of those powers with the distance necessary to explain the observations can be due to the synchrotron we show the correspondence between the methods to determine the jet parameters based on time lags and ra- dio core shifts. We give a formula relating the lags and the angular displacements of the radio thank M. B ottcher, B. De Marco, P. Gandhi, N. Ky- lafis, P.-O. Petrucci, Th. Siegert and A. Veledina for valu-18 Z DZIARSKI ET AL . able comments and discussions, A. Tchekhovskoy for per- mission to use his plot of GRMHD simulations, and the referee for valuable comments. We acknowledge support from the Polish National Science Centre under the and and from the International Space Science Institute (Bern). Support for this work was provided by NASA through the NASA Hubble Fellowship grant awarded by the Space Telescope Science Institute, which is op- erated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5-26555. This paper makes use of the following ALMA ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), MOST and ASIAA (Taiwan), and KASI (Republic of Korea), in coop- eration with the Republic of Chile. The Joint ALMA Obser- vatory is operated by ESO, AUI/NRAO and NAOJ. The Na- tional Radio Astronomy Observatory is a facility of the Na- tional Science Foundation operated under cooperative agree- ment by Associated Universities, GENERAL SOLUTION We provide here the general solution to the equations pro- viding the jet structure, given in Section 2.1. We first give the solution parametrized by the equipartition parameter, eq, but without utilizing the time-lag constraint. This solution is analogous to that given by equations (28-29) in Zdziarski et al. (2012), which is valid for a= 2,b= 1 only. Here, we assume that equipartition holds along the entire emitting jet, i.e.,a= 2b; otherwise, it would be artificial to impose it only atz0. We note that the solutions below are while observationally we de- terminemax. The relation between the two involves B0, andi, see Equations (1) and (5). As a consequence, in terms of maxwould be rather complicated, and we do not provide them. We determine B0by settingn0from Equation (14) equal to that following from Equation (17), and then use Equation (15), finding B(z) +ki)(fEfN) 2 6+p [2(1 +zr)]4+p 6+p Then,B0=B(z0). We then Equation (A1) in the formula for Fin case, Equation (12), which yields for pc(1 1 13+2 p  (A2) 3 tan 7+p 13+2 p (1 +zr)19+3 p 13+2 from Equation (12) for a= 2b, gbpEb(p+ 5)6 b(p+ 6)2 =(4 +b): (A3)Then,z0=z0, at whichF0= 2F0gbp, see Equation (12)3. We can then substitute the above z(z0)into =3C1(p)gbp(1 2 13+2 p 7+2p 13+2 p29+2p 13+2 pc17+2 p 13+2 pm15+2 p 13+2 pe (1 +zr)15+2 p 13+2 p e3+2p 13+2 p; We see that the above solutions are ob- tained without specifying either 0orz0. Also, the only in the factor gbp, and does not modify the functional dependencies. Equations (A2) and (A4) are equiv- alent to equations (28-29) in Zdziarski et al. (2012), which are fora= 2,b= 1, and differ only in the definition of eq and by factors of the order of unity due to a slightly different way of integrating the emission along the jet. Next, we can use the independent determination of zfrom the time lags,
ta. A single measured lag between the via Equation (20), z2z1. This can be compared to the prediction using zof Equation (A2), which yields a constraint between and. However, a single measurement of
tahas typically a large error. We can combine them by fitting the relationship between
tavs. z2z1. This can be done even when the break frequency, 0, is unknown. However, here it is known, and we find it convenient to define fitted to a number of measured lags. This then implies z0=ct0. We can set it equal to that implied by Equation (A2), and 3Equation (A2) also provides the correct form of equation (5) in Heinz (2006) forp= 2 ,b= 1 . His equation should be multiplied by 1=2 factor, which is due to that factor missing in his equation (1), which should have accounted for the frame transformation from 0to. That incorrect model formulation was used in T21.JET PARAMETERS 19 solve for tan  as a function of , tan  =3 (0t0)13+2 p 7+p(1 +zr)19+3 p 7+p: (A5)Note a relatively strong dependence of can then insert this tan  into Equation (A4) to +ki)(fEfN) +zr)11+p 7+p: We determine n0using the above B0in Equation (17), n0=18+2 +zr)22+2 p 7+p: P., Miller-Jones, J. C. A., Bahramian, A., et al. 2020, MNRAS, 493, L81, doi: C., Brenneman, L. W., Dauser, T., et al. 2021, SSRv, 217, 65, doi: S. D., Barbier, L. M., Cummings, J. R., et al. 2005, SSRv, 120, 143, doi: A. M. 1999, MNRAS, 305, 181, doi: G. S., & Ruzmaikin, A. A. 1974, Ap&SS, 28, 45, doi: R., Meier, D., & Readhead, A. 2019, ARA&A, 57, 467, doi: R. D., & K onigl, A. 1979, ApJ, 232, 34, doi: R. D., & Payne, D. G. 1982, MNRAS, 199, 883, doi: R. D., & Znajek, R. L. 1977, MNRAS, 179, 433, doi: J. S., Fender, R. P., Motta, S. E., et al. 2020, 4, 697, doi: D. J. K., Fabian, A. C., Barret, D., et al. 2019, MNRAS, 490, 1350, doi: P., Maccarone, T. J., O'Brien, K., et al. 2010, MNRAS, 404, L21, doi: C., Cavecchi, Y ., Heemskerk, M. H. M., et al. 2018, MNRAS, 473, 4417, doi: S. W., & Tchekhovskoy, A. 2020, ARA&A, 58, 407, doi: J., Petrucci, P. O., Henri, G., Saug 'e, L., & Pelletier, G. 2006, A&A, 447, 813, doi: P., Makishima, K., Durant, M., et al. 2008, MNRAS, 390, L29, doi: P., Bachetti, M., Dhillon, V . S., et al. 2017, 1, 859, doi: G., Guilbert, P. W., & Svensson, R. 1988, ApJL, 334, L5, doi: G., & Tavecchio, F. 2015, MNRAS, 448, 1060, doi: D. 2005, A&A, 437, 1007, doi: V . L., & Syrovatskii, S. I. 1965, ARA&A, 3, 297, doi: N., Jean, P., & Prantzos, N. 2006, A&A, 457, 753, doi: S. 2006, ApJ, 636, 316, doi: S., & Sunyaev, R. A. 2003, MNRAS, 343, L59, doi: G., & Pelletier, G. 1991, ApJL, 383, L7, doi: O., Fender, R. P., & Kaiser, C. R. 2010, MNRAS, 401, 394, doi: T. W., O'Dell, S. L., & Stein, W. A. 1974, ApJ, 188, 353, doi: S. G., Marscher, A. P., Mattox, J. R., et al. 2001, ApJS, 134, 181, doi: 'nski, K., Ghisellini, G., Svensson, R., & Gracia, J. 2006, A&A, 451, 739, doi: T., Negoro, H., Yoneyama, T., et al. 2018, 11399, 1 Kellermann, K. I., Lister, M. L., Homan, D. C., et al. 2004, ApJ, 609, 539, doi: S. S., Vlahakis, N., K onigl, A., & Barkov, M. V . 2009, MNRAS, 394, 1182, doi: A. 1981, ApJ, 243, 700, doi: 10.1086/158638

2108.10929

Andrzej A. Zdziarski

Andrzej A. Zdziarski, Alexandra J. Tetarenko, Marek Sikora

Jet Parameters in the Black-Hole X-Ray Binary MAXI J1820+070

ApJ, in press

10.3847/1538-4357/ac38a9

astro-ph.HE

http://creativecommons.org/licenses/by/4.0/

We study the jet in the hard state of the accreting black-hole binary MAXI J1820+070. From the available radio-to-optical spectral and variability data, we put strong constraints on the jet parameters. We find while it is not possible to uniquely determine the jet Lorentz factor from the spectral and variability properties alone, we can estimate the jet opening angle the distance at which the jet starts emitting ($sim 3 and the magnetic field strength with relatively low uncertainty, as they depend weakly on the bulk Lorentz factor. We find the breaks in the variability power spectra from radio to sub-mm are consistent with variability damping over the time scale equal to the travel time along the jet at any Lorentz factor. This factor can still be constrained by the pair production rate within the jet base, which we calculate based on the observed X-ray/soft and the jet power, required to be less than the accretion power. The minimum ($sim$1.5) and maximum ($sim$4.5) Lorentz factors correspond to the dominance of pairs and ions, and the minimum and maximum jet We estimate the magnetic flux threading the black hole and find the jet can be powered by the Blandford-Znajek mechanism in flow accretion flow. We point out the similarity of our derived formalism to that of core shifts, observed in extragalactic radio sources.

with. This is because it assumes that what the expert thinks are important attributes, are in fact what the machine finds relevant to make class distinctions. But machines are not people ; the attributes that humans find most salient may not in fact be salient to machines. More broadly, what people find to be obvious class distinctions might appear extremely subtle to the machine and vice versa. Thus, we need a new approach that similarly reduces annotator effort by focusing only on important attributes, but that crucially relies on the machine to define attribute importance. With these considerations in mind, we propose a new learning interface for ZSL learners to learn from experts. For each novel class, the expert first identifies a close cousin that the learner already knows about. The learner then chooses the attributes that it thinks will be most informative for it to learn and actively queries the expert for just these attributes. This learning space brings to the fore the question of how must the learner choose attributes to query that maximize performance (achieve good accuracy) but min- imize expert effort (choose informative attributes)? Similar problems are explored in active learning, where learners must choose unlabeled data points to label. However, where active learning techniques have a priori access to to make a range of measurements (e.g., in our case the learner must choose attributes to query for a completely unseen class . We address this challenge by proposing multiple novel kinds of query strategies to learn from this new active ZSL interface. We design strategies based on a new measure of attribute uncertainty based on class taxonomies and a notion of expected change in model predictions. We also for when we have access to an image of the novel class, thus generalizing to the regime of combined zero- and few-shot learning. Our proposed query strategies can with existing zero-shot learners. We experiment with three datasets, namely CUB, SUN and AWA2, and show significant reduction in annotation costs while keeping the performance on par with the full an- notation models. With only 35% of annotations, we get close to full model performance on SUN and CUB. Our approach also significantly outperforms prior unsupervised ZSL work on CUB and SUN without a single attribute annotation . are: *A new active ZSL approach to collect expert annotations that is more accurate than using descriptions and more time-efficient than an- notating a full attribute query strategies (e.g., based on uncertainty and ex- pected model change), to actively query expert attributes to rapidly train the learner. Our results suggest that thinking about what information to acquire from the annotator, and what interface the expert uses, is a promising direction for building accurate recognition models that are easy to train.2. Related Work Zero-shot learning. In zero-shot learning [ 13,14], the model learns to classify unseen classes without any training images by leveraging side information like attribute descrip- tions. Initial work by Lampert et al. [13] proposed first predicting attributes from images and then classifying im- ages based on the predicted attributes. Recent work has looked at projecting image features into attribute space, and measuring similarity with class descriptions [ recent work has used the attribute description to pro- duce improvements by embedding these descriptions as well as images into a shared feature space [ 39,6,33,20]. The auxilliary losses such as a reconstruction loss can be used to regularize the problem [ 31,39]. Our proposed frame- work builds on these ideas, but leverages sparser but about each class from the expert. Because collecting attributes from experts is have looked at other sources of class methods have been colloquially referred to as Unsu- pervised ZSL as they do not require attribute supervision for test classes. They range from using word embeddings of the class names [ 8,24,1], class and attribute embed- ding [ 3], or textual descriptions (usually from of classes instead of attribute that while these methods are called unsupervised, they still require information such as a large training corpus for word embedding or text articles (written by domain experts). In addition, these sources of information are not for identifying classes visually. While we share the same goal of reducing annotation cost, we propose an alter- native active-learning framework for attribute annotation. Our paper also considers the possibility of going and assuming that the annotator can provide us with one image of the novel class. This notion of combining images with zero-shot attribute information was first intro- duced by Tsai et al. [32]. Schonfeld et al. [27] introduce a V AE model to align the image features and attributes in a latent space to perform combined few-shot and Since a field-guide has images as well as models combining these two are worth learning. Active learning has been extensively ex- plored in the the area of machine learning and computer vision. These methods aim to make judicial use of an an- notation budget by selecting useful unlabeled data to be labelled first. Several methods have been proposed that pick data to be labelled based on objectives such as bringing larger expected model change [ 29,10], increasing the di- versity of labelled set[ 28,40], reducing the uncertainty in predictions [ 17,26,9]. Recently many new adversarial learning to acquire labels for most infor- mative data [ 41,30,42,37]. Our work is inspired by some of these techniques, but applies to a completely different and new problem: that of choosing attributes to label. Since ourExpert Annotator 'Tree Swallow' is similar to 'Cliff Swallow' Expert is the wing color for 'Tree Swallow'?'Tree Swallow' has iridiscent blue wings Initial Attribute Descriptor Attribute unknown to learner Updated 2. Overview of our workflow. (a) An expert annotator introduces the learner to a novel class by first providing a similar base class. (b) The Learner then for values of different attributes using an Acquistion function. (c) The expert annotator provides the values of the attributes and the learner updates its state and class do not have access to unlabelled data, the learning query strategies no longer apply. Neverthe- less, we take inspiration from this prior work to define a family of new techniques more suited to the zero-shot setup. Learning from experts. There are other pipelines that are motivated by the question of learning from experts. Misra et al. [ 18] propose a visual question answering system that learns by interactions. Unfortunately, it is not clear if this system can generalize to the general recognition setting. Ear- lier work leveraged experts to automatically correct part locations [ 4] or define patches for fine-grained classifica- tion [ 22]. However, these models rely explicitly on the part- based design of the recognition models. There has also been work on recognition techniques that query humans during inference for similarity comparisons [ 36] and attribute de- scriptions [ 34,5]. In contrast, we query the human annotator for particular attributes when we are learning a novel class. 3. Method 3.1. Problem Setup Our goal is to produce a learner that can learn new classes from very few interactions with an expert annotator. As in traditional zero-shot learning, we assume that the learner is first trained on some set of base or seen classes B. For each base class y B, the learner knows a vector of attributes A(y)Rd, where dis the number of attributes (we assume real-valued attributes in this work, in line with previous work on ZSL). The learner is also provided with a large labeled training set for the base classes consisting of images xi, i= 1, . . . , n and corresponding labels yi. Once trained and deployed, the learner gets a set of hith- erto unseen classes, Nto recognize. It is here that our problem setup diverges from traditional zero-shot learning. A traditional zero-shot learning system needs the full at- tribute description of all novel classes: {A(y)y N} . In contrast, in our proposed setup, the learner must use veryfew field-guide -like interactions with the expert annotator. In particular, for each new class y N , the annotator first gives the learner the most similar class S(y) Bfrom the set of base classes. Next, the learner incrementally asks the annotator for the values of attributes, one attribute at a time. The goal of the learner is then to learn to recognize the novel class from as few attribute queries as possible. We call this new kind of learning interface and the 2). To design a learner in this interactive setup, two questions must be answered: (a) how should we learn with the incom- plete attribute description?, and (b) how should we choose which attributes to query? We address the first question in the next section and then discuss our strategy for the more challenging question of choosing attributes to annotate. 3.2. Learning from Sparse Attribute Annotation For every novel class y, the learner is told the most similar class S(y), and it queries values for a subset of I(y) {0***d}. For attributes where novel is missing, we impute the attribute descriptor using the values from S(y). A'(y)[i] =( A(y)[i] /I(y)(1) The imputed vector A'(y)is used as the attribute descrip- tor for the novel class yin the zero-shot model. Note that this approach is model-agnostic and can be used with any zero-shot model Strategies for Querying Attributes We now describe how we pick the attributes iteratively to collect the sparse attribute annotation. Suppose the learner is learning about novel class y, with S(y)as its most simi- lar base class. Suppose it has already queried for a subset I(y)of the attributes resulting in an imputed attribute vector A'(y). Given this information, it must now choose a new attribute to query for. The learner will make this choice using a query strategy oracquisition function p; the attribute chosen is p(S(y), I(y), A'(y)). The notion of a query strategy is reminiscent of active learning, where the learner must choose unlabeled data points that it wants labeled. However, in active learning, the learner has access to the unlabeled data, and so it can use its own belief about the unlabeled data points to make the call. For example, it can choose to label data points for which it is most uncertain, or for which an annotation is most likely to change its state significantly. In contrast, in our case, the learner is faced with a completely unseen class . How can the learner identify informative attributes when it has never seen this class before? Below, we present two solutions to this challenge. The first solution uses a taxonomy over the base classes to Oriole Hooded Oriole Scott Oriole Low variance attributes: wings shape, body color, bill shape High variance attribute: nape color Figure 3. While attributes like wing shape, bill shape are com- mon to all orioles, color of the nape has high attributes. The second uses the imputed attribute vector and looks at the change in representation space with respect to the change in a attribute. 3.3.1 Taxonomy-Based Querying Class taxonomies are common in many domains (e.g., birds) and can be defined easily by a domain expert. Siblings in this taxonomy are likely to be similar to each other, sharing several common attributes, and probably differing from each other in only a few attributes. For example, in the domain of birds, all classes of the Oriole supercategory have black wings, a yellow body, and conical bills. Therefore, it is not prudent to query for wing color, body color or bill shape for an Oriole class. Instead we might want to query for the nape color, which varies among Orioles: Hooded Orioles have a yellower nape than Baltimore Orioles. Thus, in general, we want to query attributes that vary a lot among classses in the relevant subtree, and ignore those that don't. We can use this intuition to intelligently pick attributes to query by leveraging the class taxonomy. Let Rzbe the set of sibling classes for a base class zincluding the class zitself. When the annotator provides the similar base class S(y)for the novel class y, we look at the variance of each attribute in sibling classes RS(y). Attributes with lower variance are common to all the sibling classes and will be less informative for a similar novel class y. In contrast, attributes with higher variance vary a lot in this subtree, so querying for their value is prudent. Thus, the variance for each attribute among the sibling categories RS(y)is a measure of which attributes are more informative for this set of classes. We define this measure of variance amongst siblings for novel class y(denoted by (2) where jindexes the attribute, Var denotes the variance of a set of values, and siranges over the siblings of S(y) The most informative attributes are the ones with max- imum Sibling-variance . We therefore choose attributes to label in decreasing order of Sibling-variance : psv(S(y), I(y), A'(y)) = arg max j /I(y)Qsv(y)[j] (3)Attributes sometimes belong to groups. For example, in CUB, many different attributes all correspond to body color: each attribute captures a different color, and together the body color attributes define a multinomial the different colors. For such groups of attributes, we only look at the maximum varying attribute to find . Thus, if gis such a group of attributes, for this group is be defined as, Qsv(y)[g] = select the whole group of attributes with for annotation. Note that this method resembles uncertainty measure- based active learning methods. The learner first picks at- tributes for annotation where the learner is uncertain. But the way we measure this uncertainty is different from these methods, as we have no information about the novel class. 3.3.2 Querying Based on Representation Change Many zero-shot learners encode both the attributes and im- ages in a common latent space, and train classifiers in this space. Attribute encoders in these approaches can learn to perform a variety of useful functions, such as more if they are more identifiable or more the classifier operates in this latent space, anno- tations that significantly change this latent representation are more likely to influence classification decisions. This sug- gests that we should query attributes which when changed cause the largest change in latent space. We call this method of looking at changes in representation denoted by Qrc. Concretely, let Ea:Rd-Rl, be the attribute encoder function that maps an attribute into a latent of dimension l. Suppose that A'(y)is the current imputed attribute representation (note that A'(y)starts as A(S(y))and is gradually filled in with the true attribute val- uesA(y)as the learner queries the annotator). This attribute vector is represented in latent space as Ea(A'(y)). Now we wish to know which attributes when perturbed will lead to the largest change in this representation. Since the are non-linear, we make use of the local as represented by the Jacobian of the encoder to measure these changes. JEa(x) define the of attribute l2-norm of the partial derivative measures how much a small perturbation to that particular attribute changes theencoding in the latent space. The learner then queries the attribute with the maximum . prc(S(y), I(y), A'(y)) = arg max j grouped attributes, similar to Sibling-variance , we measure to be the maximum within the attributes of that group. is similar to the expected active learning methods[ 29,10]. These meth- ods pick entities based on the most change to the model in expectation. Similarly, picks at- tributes that change the encoding the most in latent space. 3.3.3 Querying Using Image Data In practice, an expert can easily provide a single image of the novel class. Only recently have ZSL techniques begun to leverage this information[ 32,27]. Here, we show how we can use this single image to better choose attributes to label, using the following Image-based key intuition here is that the image provides a repre- sentation for the class that must ultimately match the final attribute description. As such, the learner can recognize at- tributes in the image, and attempt to reconcile the recognized attributes and the imputed based on the expert suppose the image available for the novel class yisxy. Suppose also that we have a trained image encoder Eithat maps the image to the latent space, and an attribute decoder Dathat decodes this latent an attribute vector; many recent methods train these [ 27]. Using these modules, the learner can get an attribute vector from the image: ~A(y) =Da(Ei(x)). In general, ~A(y)will not match the imputed Hypothetically replacing the i-th imputed the image-derived counterpart in a new attribute description Ai(y): Ai(y) =( A'(y)[j] strategy then picks the attribute jwhich maximally pushes the embedding of the hypothetical at- tribute vector Aj(y)closest to the novel class image xy)[i] (9) pI(I(y), A'(y), xy) = arg max j /I(y)QI(y|A'(y), xy)[j](10) 4. Results 4.1. Dataset and Implementation Details For all experiments we use 2048 -dimensional features from ResNet-101 [ 11]. We compare our method on three 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Harmonic generalized accuracy 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Unseen accuracy 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Harmonic generalized accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Unseen accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Harmonic generalized w/ 4. Performance of our method in terms of top-1 accuracy vs. number of annotations provided by the annotators during deployment with CADA-V AE as base performs better than the traditional ZSL on all benchmarks and better than unsupervised baselines on CUB and SUN, proving its benchmark datasets: CUB-200-2011 [ 35] (CUB), Animals with Attributes 2 [ 38] (AWA2) and the SUN at- tribute dataset [ 21] (SUN). CUB is annotated with 312 part- based attributes; AWA2 and SUN have 85 and 102 These attributes are labelled for every class. We use the standard train-test split from [ 38] for all annotations of most similar base class: Our prob- lem setup requires that the annotator provide the most similar base class for each novel class. Since CUB is a requiring bird expertise, we worked with a profes- sional bird watcher to annotate this information for every novel class. AWA2 and SUN do not require expert knowl- edge, so we collected the similar class information using 3 annotators. We took a majority vote, and in cases when all 3 disagreed we asked them to come to a consensus. We will release these expert annotations publicly upon : For SUN, the taxonomy is already available along with the dataset. We manually created a taxonomy for AWA2 by looking at the family in biological CUB, we use the general class name as the parent in the taxonomy (Hooded Oriole -Oriole).0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Harmonic generalized w/ 5. Performance of our method with TF-V AEGAN as base model. ZSL-Interactive performs better than the traditional ZSL and unsupervised baselines, proving the generalizability of ZSL- Interactive to other zero-shot models. Base zero-shot learner : To show generalization of our method over different models we experiment with two base zero-shot learners: CADA-V AE[ 27] and TF-V AEGAN [ 19]. CADA-V AE trains two variational auto-encoders on the base classes to learn a common embedding space for attribute de- scriptions and images. It then trains novel class classifiers in this latent space. All the hyperparameters for the architecture and training are kept as prescribed by [ 27]. TF-V AEGAN uses a V AE-GAN [ 15] to generate realistic features from at- tributes and uses these generated features from unseen class to train a classifier. We show the results averaged over 6 different runs of the model. Metrics : We measure the per-class classification accuracy over unseen classes and the harmonic mean of seen and un- seen classification accuracy for generalized zero-shot learn- ing. We plot these metrics as a function of the for each approach. 4.2. Is interactive ZSL accurate and first ask how our proposed interactive setup compares with prior work on ZSL in terms of accuracy and the burden of annotation. We compare our proposed setup against the following ZSL with full attribute annotation. Reduc- ing annotation effort here can only be achieved by using a smaller attribute vocabulary. Given an annotation budget, smaller subset of attributes is chosen uni- formly at random, and a new ZSL model is retrained with the chosen attributes. In deployment, all classes must be described with this reduced vocabulary. *W2V : This unsupervised ZSL approach uses vectors of the classes instead of attribute vec- tors [1]. *CAAP : This approach uses word embeddings for the classes and the attributes to find the attribute vector for un- seen classes [3]. *ZSLNS [23] and GAZSL [43] : use wikipedia articles instead of attributes. Since the original papers use older backbones for ZSL, we show the number for both the original model and using ZSLNS (or GAZSL) extracted descriptorswith newer backbones (CADA-V AE and TF-V AEGAN). While the last three are often deemed unsupervised, they do require either a large corpus of text with attributes and classes to learn word embeddings, or carefully curated text articles typically edited by experts, thus indirectly expert time, which we aim to minimize. Note that reported numbers for unsupervised ZSL ap- proaches typically use a different class split that ensures that each novel class has a closely related base class. Instead, we use the proposed splits in the attribute-based ZSL bench- mark [ 38]. We have used the original code provided by the authors and changed the splits wherever available. We compare our best performing acquisition ), to the above baselines. We plot the accu- racy as a function of the amount of annotation in Figure 4 (CADA-V AE) and Figure 5 (TF-V AEGAN). We find: *The expert annotated closest base class is extremely in- formative . With just that one annotation, one can recover almost two-thirds of the performance of a full zero-shot learn- ing system with access to hundreds of attributes. Note that using a randomly sampled class as the most similar class instead of the expert annotated one yields very poor accuracy (<20%) indicating that the expert-provided information is critically important. *Our approach can dramatically reduce annotator bur- den. On CUB and SUN, our approach with only a third of the annotations (just 10 interactions per class) is as good as traditional ZSL with allof the approach generalizes to other ZSL models . Our method performs better than the baselines even when TF- V AEGAN is used as the base model Figure 5. This shows that as new ZSL methods develop, our method should gener- alize and can be used out-of-the-box with them. See supple- mentary for TF-V AEGAN results on other attributes more informative than unsupervised ZSL . While some methods perform better than our method with less information on AWA2, our method beats all the baselines on CUB and SUN. It is relatively hard to information in corpora like wikipedia for fine- grained categories like those in SUN and CUB. This is the reason attribute-based systems are essential when performing ZSL in fine-grained domains. Just providing most similar class annotation ( i.e., the field guide approach ) turns out to be significantly more useful than using text or word embedding as attributes. 4.3. How do different acquistion functions perform? Our method performs better than baselines primarily for two reasons, the field-guide interface and an intelligent ac- quisition function. We evaluate how different perform with attribute annotation costs. Among the various query strategies we have proposed, in gen- eral, the best performing query strategy is Sibling-variance .0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Harmonic generalized w/ Random w/ w/ Figure 6. Performance of the two acquition functions with CADA- V AE on CUB. Both functions perform better than the ran- dom acquisition function. Sibling-variance performs better than , but the latter does not require See supplementary for results on other datasets. tends to have a weaker performance than Sibling-variance but it does not require the taxonomy to function and hence is still useful. Figure C.1 compares the 2 acquistion function against a random acquistion function as a baseline on CUB with CADA-V AE (see other models and datasets). Both of our proposed query strategies far outperform random attribute selection as well as the traditional ZSL pipeline, no matter what the labeling budget is . This is true both for the unseen-only evaluation as well as the general evaluation. 4.4. Can we do better with images? Figure D.1 shows the performance of our approach when one image is given by the annotator along with the inter- active attributes annotations for CUB. Almost all of our conclusions from the previous section carry over, with the exception that starts weaker than the baseline of choosing attributes to query randomly, but per- forms better than it in the later stages. Additionally, performs on par with the using any additional taxonomy information. This suggests that using the inconsistency between imputed attributes to determine what attributes are use- ful is a viable approach to interact with annotators. for results on other benchmarks. 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Harmonic generalized w/ w/ w/ OSL+ZSL-Interactive w/ ZSL Figure 7. Performance of our method, zero+one-shot setting, where the annotator provides a single image for the novel class along with the interactively queried attributes. All methods perform better than the baselines; Image-based performs on par with requiring an additional taxonomy over the base classes.4.5. Are acquisition functions better than experts? One might question how our interactive method compares to collecting information from an expert. As discussed in section 1 a field-guide provides what the expert finds to be distinctive attribute differences without any interaction. We evaluate how this compares to the learner actively querying. To this end, for 20 of the 50 novel classes the CUB dataset, we additionally ask an expert to identify the 10 most infor- mative attributes that distinguish each novel class from its most similar base class, in order of importance. We use this infomation along with the similar class to construct the expert attribute baseline. 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Harmonic generalized w/ w/ w/ Traditional ZSL Figure 8. Performance of interactive methods against the expert selected attributes for CUB (20 novel classes). The learner performs better than the expert providing the attributes. We find that the interactive methods are significantly bet- ter at asking for useful attributes than what the domain expert gives (Figure 8). This is a surprising result, as it shows that zero-shot learners learn about the domain differently than a human expert. In the set selected by experts, the experts give importance to attributes like bird size and eye color. While this information is useful to classify and birds for a bird watcher, the model finds it difficult to understand the size of the bird as it is a relative property. Other attributes like eye color are very small in images (or obstructed) so the network cannot necessarily utilize that knowledge during representation learning. Hence, providing these attributes to the network is not very informative and it is better to let the learner select informative attributes. 4.6. Is the query selection good? We look at the types of attributes queried by our for different parent categories of CUB. The attributes queried should vary based on the sibling classes. Figure B.1, shows the top-3 attributes queried first by method for Swallows and Cuckoos. It also shows the top-3 attributes picked by measuring variance over all classes. While bill length and overall shape have high variance over all classes, within Swallows (and within Cuckoos) the variance is not big and therefore should not be annotated first. For swallows, it can be seen from the image that throat, crown and forehead color varies within category, and hence should be acquired first. For Cuckoo, Cuckoo Yellow-billed Cuckoo Mangrove underpart color, crown color and belly Swallow Barn Swallow Cliff throat color, crown color and forehead color Total-Variance: bill length, throat color and overall shape Figure 9. Attributes selected by Sibling-variance for a parent class in the taxonomy and the attributes selected by measuring variance over all classes (top). As can be seen from the images within the categories, overall shape and bill length do not change much and measuring variance to pick attributes across all classes is not informative and our approach of looking at the siblings is necessary. crown and belly color varies (See supplementary for other examples and datasets). We also visualize how learning progresses with interac- tive questions and reponses. Figure 10 shows the of 2 CUB classes and the most similar base classes. The full attributes descriptor (bigger dot) for a class is surrounded by images of that class represented by smaller dots of the same color. Dots with red and black edges show the progression of novel class attributes as the learner inter- actively gains more information using the acquisition function respectively. We see yields a faster progression from the similar base class attribute to the novel class. For example, the pro- gression of the attribute descriptor encoding using for Tree Swallow reaches the full attribute de- scriptor quicker, whereas with the random function the at- tribute is still close to the base class Cliff Swallow. More examples are in the Swallow Cliff Orchard Oriole Baltimore Oriole Figure 10. t-SNE visualization for 2 novel CUB classes (first class in legend) and its closest base classes (second) in the latent space of CADA-V AE. Smaller dots represent test images and larger dots rep- resent class attribute embedding. Red edges show the progression of novel class attributes as learners interact using Sibling-variance . Dots with black edges show the progression with random function.4.7. How much does the taxonomy help? As seen from the results in previous sections, lets us select informative attributes. In this section we evaluate if having a taxonomy is a requirement for this to work. Rather than measuring variance within the sibling classes we measure variance over all classes. This variant will always prioritize some types of attributes over of the class, and ignore local variations within siblings. Therefore, we expect it will perform worse than with known taxonomy information. 0 5 10 15 20 25 30 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 30 # Attribute Annotations per Accuracy CUB: Harmonic generalized w/o ZSL Figure 11. Comparing Sibling-variance against a variant where the taxonomy is unknown on CUB. The model loses accuracy if sibling classes are not used to measure Sibling-variance . Figure F.1 compares the method without taxonomy infor- mation against the model where the taxonomy is known. As expected, this method loses performance because the local variation of a class cannot be measured, and hence those attributes cannot be selected. But even without the taxonomy the method performs better than ZSL and is useful for cases when the taxonomy is not known or difficult to acquire ( for performance on other datasets.). Finally, we also show that our methods are not very sensitive to simi- lar base class selection as long as the expert chooses a class that is not wildly different looking (see Conclusions In this work we show that an interactive annotation approach identifies informative and can achieve high performance by an annotation budget. We present different ways to acquire informative sparse annotations from an annotator and show that is is better to let the machine choose the at- tributes to ask for (even when compared to an expert from the domain). Given these promising results, there are many avenues of future work: can one get even more choosing a different number of attributes for based on class confidence? We also need to bridge the gap between the attribute understanding of humans and machines, as our results show that human experts and neural- networks do not find the same attributes equally We thank our funding agen- cies TCS, NSF 1900783, and DARPA LwLL for their Zeynep Akata, Scott Reed, Daniel Walter, Honglak Lee, and Bernt Schiele. Evaluation of output embeddings for fine- grained image classification. In CVPR , 2015. 1, 2, 6 [2]Ziad Al-Halah and Rainer Stiefelhagen. Automatic estimation and learning of semantic attributes for a thousand categories. In CVPR , 2017. 1, 2 [3]Ziad Al-Halah, Makarand Tapaswi, and Rainer the missing link: Predicting class-attribute associ- ations for unsupervised zero-shot learning. In CVPR , 2016. 1, 2, 6 [4]Steve Branson, Pietro Perona, and Serge Belongie. from weak annotation: Interactive training of deformable part models. In ICCV , 2011. 3 [5]Steve Branson, Grant Van Horn, Catherine Wah, Pietro Per- ona, and Serge Belongie. The ignorant led by the blind: A hybrid human-machine vision system for fine-grained IJCV , 2014. 3 [6]Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha. Synthesized classifiers for zero-shot learning. In CVPR , 2016. 2 [7]Mohamed Elhoseiny, Babak Saleh, and Ahmed Elgammal. Write a classifier: Zero-shot learning using purely In ICCV , 2013. 1, 2 [8]Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. De- vise: A deep visual-semantic embedding model. In NeurIPS , 2013. 1, 2 [9]Yarin Gal and Zoubin Ghahramani. Dropout as a Representing model uncertainty in deep learn- ing. In ICML , 2016. 2 [10] Joshua Goodman. Exponential priors for maximum entropy models. In HLT-NAACL , 2004. 2, 5 [11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR , 2016. 5 [12] Elyor Kodirov, Tao Xiang, and Shaogang Gong. for zero-shot learning. In CVPR , 2017. 2 [13] Christoph H Lampert, Hannes Nickisch, and Stefan Harmel- ing. Attribute-based classification for zero-shot visual TPAMI , 36(3):453-465, 2013. 2 [14] Hugo Larochelle, Dumitru Erhan, and Yoshua Bengio. Zero- data learning of new tasks. In AAAI , 2008. 2 [15] Anders Boesen Lindbo Larsen, Soren Kaae Sonderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In ICML , 2016. 6 [16] Jimmy Lei Ba, Kevin Swersky, Sanja Fidler, et al. Predicting deep zero-shot convolutional neural networks using In CVPR , 2015. 1, 2 [17] David D Lewis and William A Gale. A sequential algorithm for training text classifiers. 1994. 2 [18] Ishan Misra, Ross Girshick, Rob Fergus, Martial Hebert, Abhinav Gupta, and Laurens van der Maaten. Learning by Asking Questions. In CVPR , 2018. 3 [19] Sanath Narayan, Akshita Gupta, Fahad Shahbaz Khan, Cees GM Snoek, and Ling Shao. Latent embedding feed-back and discriminative features for zero-shot , 2020. 6 [20] Mohammad Norouzi, Tomas Mikolov, Samy Bengio, Yoram Singer, Jonathon Shlens, Andrea Frome, Greg S Corrado, and Jeffrey Dean. Zero-shot learning by convex combination of semantic embeddings. In ICLR , 2013. 2 [21] Genevieve Patterson and James Hays. Sun attribute annotating, and recognizing scene attributes. In CVPR , 2012. 5 [22] Genevieve Patterson, Grant Van Horn, Serge Belongie, Pietro Perona, and James Hays. Bootstrapping fine-grained classi- fiers: Active learning with a crowd in the loop. In NeurIPS , 2013. 3 [23] Ruizhi Qiao, Lingqiao Liu, Chunhua Shen, and Anton Van Den Hengel. Less is more: zero-shot learning from online textual documents with noise suppression. In CVPR , 2016. 1, 2, 6 [24] Marcus Rohrbach, Michael Stark, Gy orgy Szarvas, Iryna Gurevych, and Bernt Schiele. What helps where-and why? semantic relatedness for knowledge transfer. In CVPR , 2010. 1, 2 [25] Bernardino Romera-Paredes and Philip Torr. An embarrass- ingly simple approach to zero-shot learning. In ICML , 2015. 2 [26] Tobias Scheffer, Christian Decomain, and Stefan Wrobel. Active hidden markov models for information extraction. In ISIDA , 2001. 2 [27] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and via aligned variational autoencoders. In CVPR , pages 8247-8255, 2019. 2, 5, 6 [28] Ozan Sener and Silvio Savarese. Active learning for convo- lutional neural networks: A core-set approach. CoRR , 2017. 2 [29] Burr Settles, Mark Craven, and Soumya Ray. active learning. In NeurIPS , 2008. 2, 5 [30] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Varia- tional adversarial active learning. ICCV , 2019. 2 [31] Richard Socher, Milind Ganjoo, Christopher D Manning, and Andrew Ng. Zero-shot learning through cross-modal , 2013. 2 [32] Yao-Hung Hubert Tsai and Ruslan Salakhutdinov. learning through fusing side information. CoRR , 2017. 2, 5 [33] Vinay Kumar Verma and Piyush Rai. A simple framework for zero-shot learning. In ECML , 2017. 2 [34] Catherine Wah and Serge Belongie. Attribute-based detection of unfamiliar classes with humans in the loop. In CVPR , 2013. 3 [35] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 [36] Catherine Wah, Grant Van Horn, Steve Branson, Subhransu Maji, Pietro Perona, and Serge Belongie. Similarity compar- isons for interactive fine-grained categorization. In CVPR , 2014. 3[37] Shuo Wang, Yuexiang Li, Kai Ma, Ruhui Ma, Haibing Guan, and Yefeng Zheng. Dual adversarial network for deep active learning. 2020. 2 [38] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning-a comprehensive eval- uation of the good, the bad and the ugly. TPAMI , 2018. 5, 6 [39] Yongqin Xian, Saurabh Sharma, Bernt Schiele, and Zeynep Akata. f-vaegan-d2: A feature generating framework for any-shot learning. In CVPR , 2019. 2 [40] Lin Yang, Yizhe Zhang, Jianxu Chen, Siyuan Zhang, and Danny Z Chen. Suggestive annotation: A deep active for biomedical image segmentation. In ICMICCAI , 2017. 2 [41] Aron Yu and Kristen Grauman. Thinking outside the pool: Active training image creation for relative attributes. In CVPR , 2019. 2 [42] Beichen Zhang, Liang Li, Shijie Yang, Shuhui Wang, Zheng- Jun Zha, and Qingming Huang. State-relabeling learning. In CVPR , 2020. 2 [43] Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, and Ahmed Elgammal. A generative adversarial approach for zero-shot learning from noisy texts. In CVPR , 2018. 1, 2, 6Supplementary Material A. Overview In this supplementary material we look at some more results that could not be presented in the main paper. Our code can be found at We show attributes queried by Sibling-variance on SUN and AWA2 in Sec. B. Sec. C shows the performance of the two acquisition functions on AWA2 and SUN. In Sec. D we present results for when the interactive learner uses a single image for SUN and AWA2. In Sec. E we compare the performance of our method with TF-V AEGAN on the AWA2 and SUN. Sec. F shows the effect of not using a taxonomy for AWA2 and SUN. In Sec. G we show the learner's behavior when classes other than the annotators first choice are chosen for SUN and AWA2. Sec. H presents more t-SNE visualization examples of the learning progression for novel class descriptors on all three datasets. We also strongly encourage the reader to refer to the for better visualizations of the t-SNE progression. B. More Qualitative B.1, shows the attributes queried first by method for 2 supercategories of SUN and AWA2. It also shows the attributes picked by over all the classes. For SUN, attributes like en- closed/open area or may help in disam- biguating between very different classes, but within a super- category they do not help. For example, for the sports and leisure, all the classes are closed and man- made. But attributes like competing, spectating are Similarly for indoor workplaces, attributes like using tools and are very patterns can be seen on AWA2. for attributes informative within the superclass. C. Comparison of Acquisition Functions on A WA2 and CUB Figure C.1 shows the performance of the two acquisition functions with the CADA-V AE model on AWA2 and SUN. Our acquisition functions perform sig- nificantly better than a random acquisition function, showing the value of our field-guide the results for fine-grained dataset such as SUN are similar to that on CUB, for AWA2 performs better than Sibling-variance in the later stages. This might be because the AWA2 model is trained for classes with thousands of images and has a better representation and understanding of changing Reading room Cubicle using tools, studying learning, metalIndoor sports and ring Jacuzzi competing, spectating, enclosed/open area, visible horizon and mammals Dolphin Killer whale Blue bulbous shape, colorRodents and small mammals Rat Hamster has tail, nocturnal, size (big/small), aquatic, has paws Figure B.1. Attributes selected by Sibling-variance for a parent class in the taxonomy and the attributes selected by over all classes (top) for SUN and Results for A WA2 and SUN Figure D.1 shows the performance of our approach when one image is given by the annotator along with the annotations for AWA2 and SUN. As for CUB, all the methods perform better than the baselines. For SUN, the Image-based function performs on par with requiring an additional taxonomy. For AWA2, the Image-based function performs better than all the methods we propose and the baselines. AWA2 has more training images and the classes are not very fine-grained. This might be the reason why Image-based acquisition functions work better for AWA2. E. Performance of TF-V AEGAN on A WA2 and SUN Figure E.1 show the performance of our approach when the base model is TF-V AEGAN on SUN and AWA2. Our field-guide way of annotation works better then traditional ZSL baselines for both the dataset, proving the generalization of our method. For fine-grained classes 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Unseen accuracy 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Harmonic generalized w/ Random w/ w/ 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Unseen accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Harmonic generalized w/ Random w/ w/ Figure C.1. Performance of the two acquisition functions with CADA-V AE on AWA2 and SUN. Both functions perform better than the random acquisition function. On SUN, better than , but the latter does not require taxonomy information. Results on AWA2 are different, in the earlier stages Sibling-variance is better than , but in the later stages is better. 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Unseen accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Harmonic generalized w/ w/ w/ OSL+ZSL-Interactive w/ ZSL 0 20 40 60 80 100 # Expert Annotations per Accuracy SUN: Unseen accuracy 0 20 40 60 80 100 # Expert Annotations per Accuracy SUN: Harmonic generalized w/ Random w/ w/ OSL+ZSL-Interactive w/ ZSL Figure D.1. Performance of our method under the when the annotator provides a single image for novel class along with the interactive attribute values for SUN and AWA2. such as SUN and CUB, our method is .0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Unseen accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Harmonic generalized w/ 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Unseen accuracy 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Harmonic generalized w/ E.1. Comparison of our method against unsupervised and traditional ZSL baselines with TF-V AEGAN as the base model. Our method performs better than tradtional ZSL at the same cost for both AWA2 and SUN. Similar to results for CADA-V AE in the main paper, our method works better than the unsupervised baselines for SUN. F. Performance on A WA2 and SUN F.1 compares the method without taxonomy infor- mation against the model where the taxonomy is known. The results follow the conclusion from the main paper. When there is no taxonomy information available, this method loses performance because the local variation of a class cannot be measured, and hence those attributes cannot be se- lected. But even without the taxonomy the method performs better than ZSL and is useful for cases when the taxonomy is not known or difficult to acquire. G. Effect on Performance When Changing Similar Base Class The similar class given by the annotator is certainly more important than each attribute annotation. We look at the effect of choosing another class: either a random class from the full set, or a sibling of the expert selection that is closest to the expert selected sibling in word2vec embedding space. Figure G.1 shows Sibling-variance with these with the ZSL baseline on CUB. The similar class chosen by annotators performs best. When we choose a class that is close to this (sibling), the method performs slightly worse. This shows that although our method performance is affected if a non-optimal nearest class is chosen, it is not very sensitive to it. Both these variants do significantly better than randomly selecting a similar class. This suggests that 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Unseen accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Harmonic generalized w/o ZSL 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Unseen accuracy 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Harmonic generalized w/o ZSLFigure F.1. Comparing Sibling-variance against a variant where the taxonomy is unknown on AWA2 and SUN. The model loses accuracy if sibling classes are not used to measure Sibling-variance . 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Unseen accuracy 0 5 10 15 20 25 # Attribute Annotations per Accuracy CUB: Harmonic generalized & sibling similar & random similar ZSL Figure G.1. Comparing different variants for selecting the CUB dataset. The model is not sensitive to the similar class as long as the selected class is not wildly different. the interactive model will perform well as long as annotators do not provide a wildly different looking similar class. Figure G.2 shows the results for Sibling-variance on AWA2 and SUN. The similar class chosen by similar to when a sibling base class is chosen for AWA2 and SUN. The performance is again not very sensi- tive to choosing the similar class as long as they are not with the sensitivity to the choice of the similar class, we also evaluated our method with sensitivety to at- tribute values. Note that incorrect or noisy attribute values will affect not just our proposed active ZSL but also the traditional non-interactive ZSL. With 10% noise in the when all attributes are provided, both our method and traditional ZSL see a ~3%drop in performance. With partial attribute annotations (5 per class), our proposed anno- tation strategy ( ~1%drop) fares much better than traditional ZSL annotation ( ~5%drop).0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Unseen accuracy 0 20 40 60 80 # Attribute Annotations per Accuracy AWA2: Harmonic generalized & sibling similar & random similar ZSL 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Unseen accuracy 0 20 40 60 80 100 # Attribute Annotations per Accuracy SUN: Harmonic generalized & sibling similar & random similar ZSLFigure G.2. Comparing different variants for selecting the AWA2 and SUN dataset. The model is not sensitive to the similar class as long as the selected class is not wildly different. H. t-SNE Visualizations for More Classes H.1 show the progression of Sibling-variance and random attributes for all 10 AWA2 novel classes. Note that in the standard split of AWA2, the classes are split in a way that sometimes no good similar classes could be found. For example, both seal and walrus are in the test split and hence the annotators chose beaver and walrus as similar no good base class is there for giraffe and bat so the annotators had to chose zebra and squirrel. Nonetheless the faster progression towards novel classes' images and attributes can be seen for the classes when using over random H.2 and H.3 show the progression of and random attributes for all 20 novel classes of CUB (out of 50) and SUN (out of 70). Faster progression can be seen for Sibling-variance over these classes as H.1. t-SNE visualizations. For all 10 AWA2 novel classes and corresponding similar base classes. Smaller dots represent test images and larger dots represent class attribute embeddings. Red edges show the progression of novel class attributes as learners interact . Dots with black edges show the progression with the random function. Both methods start at the base class and aim to reach to the novel class descriptor with as few interactions as possible. In most cases Sibling-variance reaches closer to the novel class descriptor quicker in contrast to random basketball arena loft art auditorium movie theater vault elevator cage outdoor kennel indoor flea market outdoor flea market shop casino bog tree window indoor window depot outdoor bus station natural lake interior frontseat car interior outdoor bank dry Figure H.2. t-SNE visualizations. For 20 SUN novel classes and corresponding similar base classes. Smaller dots represent test images and larger dots represent class attribute embeddings. Red edges show the progression of novel class attributes as learners interact . Dots with black edges show the progression with the random function. Both methods start at the base class and aim to reach the novel class descriptor with as few interactions as possible. In most cases Sibling-variance reaches closer to the novel class descriptor quicker in contrast to random .

2108.10967

Utkarsh Mall

Utkarsh Mall, Bharath Hariharan, and Kavita Bala

Zero-Shot Learning

Accepted to ICCV 2021

cs.CV

http://creativecommons.org/licenses/by/4.0/

Modern recognition systems require large amounts of supervision to Adapting to new domains requires significant data from experts, which is onerous and can become too expensive. Zero-shot learning requires an annotated set of attributes for a novel category. Annotating the full set of attributes for a novel category proves to be a tedious and expensive task in deployment. This is especially the case when the recognition domain is an expert domain. We introduce a new approach to where the learner model interactively asks for the most that define a class. We evaluate our method on with attribute annotations like CUB, SUN, and AWA2 and show that our model achieves the performance of a model with full annotations at the cost of a significantly fewer number of annotations. Since the time of experts is precious, decreasing annotation cost can be very valuable for

dependent observables to regular observables, func- tions of state variables alone [1,6,19]. Assuming a spe- cic state feedback is restrictive and exogenous inputs cannot be involved in this second technique assumes that control uis lin- early involved in all anticausal observable =(x) +Du (1) whereDis a constant matrix to be tuned to data, and (x) is a function of state alone. This is a but yields a signicant error when the input nonlinearity is prominent [24,7,21]. The third method is to assume the eect of the con- trol input on the observables is quasi-periodic in time. This assumption allows singular components of the dynamic mode spectrum to be causal, and is thus pri- marily used as a theoretical stepping stone for model- ing autonomous, periodic dynamical systems with no exogenous input [16]. Finally, the causality problem does not occur if only causal observables are used for lifting [19]. In case physically measured observables are used, however, one must be careful not to include those exogenous inputs. Typically, causal observ- ables are synthesized by articially creating of independent state variables alone. This prohibits the use of physically measured observables. It is desired that the causality problem be solved with- out those restrictive assumptions. In particular, the as- sumption of equation (1), the linear involvement of ex- ogenous inputs, does not hold for many nonlinear control systems, yielding signicant errors. Furthermore, use of synthetic observables alone, although free from causal- ity issues, has two major drawbacks. One is the lack of access to key signals that are often acquired through ac- tual measurements of a nonlinear system. Use of broad physical measurements is much more eective than cre- ating synthetic observables alone. The other drawback is noise. The standard procedure of the Koopman-based uses measured data of state variables and then augments the dataset with a large set of observ- ables created from the data of state variables. In prac- tice, measured state variables are corrupted with noise, resulting in signicantly biased estimates of the linear model. Although the measurement noise of individual state variables is assumed to be uncorrelated to each other, the nonlinear transformations due to are nonlinear functions of state variables, cause complex correlation between the true state variables and the noise-corrupted terms of the observables, leading to signicant errors in the resulting linear model. In con- trast, the bias is small and amenable if both state vari- ables and observables are measured directly. As detailed in Appendix A, the dierence is clear and plateFig. 1. Example system. In this current work, we aim to analyze where for lifting arise in physical control systems, and solve the causality problem at the origin. Causality has a deep root in physics. We will consider physical control systems governed by basic physical laws, and present two methods to remove the causality problem. The remainder of this article is structured as follows: in Section 2, we introduce the basic concepts of causality in lifting linearizations and dual-faceted linearization. Our new approach is then introduced in Sections 3 and 4, and results are shown in Section 5 for a series of nonlinear models with anticausal variables. We conclude with a short summary and possible future work in Section 6. 2 Causality in Lifting Physical Control Systems This section provides background information about causality of physical control systems and how occur when lifting dynamical systems. 2.1 Causal Paths in Dynamical Systems In a physical control system, plant dynamics are gov- erned by physical laws. Consider a consisting of basic elements, such as masses, springs, and dampers, subject to exogenous inputs as shown in Fig. 1. The spring and two dampers are nonlin- ear, and their characteristics are represented with non- linear functions called constitutive laws. The law is given by e =  C(q) whereqis elon- gation, or displacement, of the spring; e is the force generated by the spring; and  C(q) is a nonlinear, dier- entiable function. The constitutive law of each damper is given by a nonlinear, dierentiable, and e Ri=  Ri(fi); i= 1;2, where e Riis the force generated by the ithdamper and f iis the velocity of the damper1. For the rst damper, f 1=dq=dt . 1A damper is an element in the mechani- cal domain. In the electric domain, the el- ement is a resistor with a constitutive law described in terms of voltage and current. In the uid domain, it is described in terms of pressure and ow rate. Generically, we use ef- 2Exogenous Input, Input, 2. Element block diagram of nonlinear system from Fig. 1. The equation of motion of the mechanical system in Fig. 1 is given by dp=dt = eIwherepis momentum of the mass and e Iis the resultant force acting on the mass given by e I= the exogenous input force acting on the mass. From the icon model in Fig. 1, we can also nd the following conditions de- scribing how these elements are connected to each other: ue1= e + e R1and f C= f1f2. Note that the with the mass and the displacement q associated with the spring can be used as variables. This physical system is driven by exogenous input which propagate through the network of the elements and result in a change in the state variables pandq. The propagation of signals can be with an element block diagram, as shown in Fig. 2. Note that the block of the mass has a of signals, that is, from force e Ito momentum pand then to velocity f 2. This direction cannot be re- versed. As rst addressed in Newton's Principia [17], a resultant force acting on a mass determines the motion of the mass, and not the other way around. Along this causal direction, the momentum is determined by inte- gration of the force, and then the velocity is determined by the algebraic relation, f 2=p=m, the constitutive law of the mass. If the direction is reversed, i.e. velocity in- put and force output, a time derivative is involved in the signal transmission. In a sense, derivative uses future in- formation and is not causal. Along the same line, the spring block in Fig. 2 has input velocity f Cand output force e. If the input-output order is inverted, the element block includes a derivative. Therefore, the causal path is from velocity ( ow) to force (eort). The block diagram in Fig. 2 shows causal paths of all the signals originated in the two exogenous inputs [11]. These signal directions also provide a computational procedure that can update state variables, pandq, in response to exogenous inputs through integration and algebraic evaluation. A dynamical system having this causal signal propaga- tion is referred to as a system of integral causality [11]. We assume that the dynamical systems considered in the fort e for force, voltage, and pressure and ow f for and ow rate.current work are of integral causality. Integral causality can be checked conveniently with use of a bond graph. The causality propagation rules of bond graphs can de- termine causal input-output relations of individual ele- ments in such a way that no con ict2and no derivative occur in connecting all the elements. If this is possible, then the system is of integral causality. For readability of the paper, the following argument will be made without use of the bond graph notation. Instead, element block diagrams will be used for manifesting causal paths. 2.2 Dual-Faceted Linearization Linearization (DFL) [7] is a lifting lin- earization method based on a physical model of inte- gral causality. It exploits knowledge of the connectivity of elements, or network structure, as illustrated in the icon model in Fig. 1. Often we do not know the nonlin- ear function of each element's constitutive law, but we know how the mass, spring, and damper are connected to each other. In DFL, knowledge of constitutive laws is not required; experimental data and element connectiv- ity information allow for lifting DFL, the output variable of each nonlinear element is called an auxiliary variable. Given nanonlinear ele- ments involved in a dynamical system, auxiliary vari- ables are collectively denoted as 2<na. In the exam- ple of Fig. 1, the auxiliary variables are = (f 1;e;eR2)|. The dynamic system is lifted by using these to obtain a linear model: _x= Axx+ A+ Bxu _= Hxx+ H+ and Ax;A;Bx;Hx;H, and Hu are constant matrices with consistent dimensions. The rst equation is exact and is determined by the and linear elements, while the second equa- tion is determined by regression. The approximation ac- curacy can be improved by further augmenting the space with observable functions of both independent state and auxiliary variables (x;) [8], and the robustness of mod- eling is shown in [10]. Note that causality plays a critical role in determining the output of each element. As discussed previously, ow (e.g. velocity, current, ow rate, etc.) is the output for an inertial element, and eort (e.g. force, voltage, pres- sure, etc.) is the output of a capacitive element. On the other hand, the output of a resistive element depends on how the elements are connected to each other. Through causal path analysis, the direction of signal transmission is determined and, thereby, the output of a element is the two outputs of two elements is a contra- diction if both elements declare two outputs of the auxiliary variables has a clear physical mean- ing and may be measured with sensors. For variable e can be measured with a force sen- sor attached to the spring in Fig. 1, and the output of the left damper, that is, the auxiliary variable f 1can be measured with a velocity sensor. In physical systems, the connectivity of elements is ba- sically linear if it is governed by Kirchho's Voltage and Current Laws or the Generalized Kirchho's Loop and Node Rules [15]. In an electric circuit, all the voltages along a closed loop sum to zero, which is a linear re- lationship. In a mechanical system, all the forces at a mass, including an inertial force, sum to zero, which is also a linear relationship. The DFL formulation exploits this linearity in element connectivity to represent a state equation as a linear dierential equation of variables and auxiliary variables, which are out- puts of all the nonlinear comes from constitutive laws. Therefore, the output variables of nonlinear elements contain key properties of the nonlinear dynamical systems. It has been reported that lifting linearization using both inde- pendent state variables and auxiliary variables is more ecient for approximating a nonlinear dynamical sys- tem than doing so with independent state variables alone [7,9]. This implies that a richer collection of observables can be created from the augmented space of both state and auxiliary variables: =(x;). A caveat in lifting a nonlinear dynamical system using auxiliary variables is that may be a function of exoge- nous inputs: =(x;u). 2.3 Anticausal Auxiliary Variables This section examines how auxiliary variables can be connected to exogenous inputs. First consider energy- storage elements, i.e. inertial and capacitive elements. For an inertial element, the output of the element is a ow variable. If the element's constitutive law  Iis nonlinear, then the ow, f, is an auxiliary variable. Using this variable for lifting the is the time derivative of a state variable, that is, a state equation. Therefore, no time derivative of exogenous inputs is involved. The output of a is an eort variable with time no exogenous input can be involved. flow - effort subsys .PhR-1e fflow - effort subsys .flow - effort subsys .PhRe feffort - flow subsys . (A) (B)Fig. 3. Two possible congurations of a resistor connected to a loop junction. An anticausal auxiliary variable may be generated only at (resistive) elements. In the element block diagram in Fig. 2, two resistive elements are in- volved. Note that the input-output relations are dier- ent for the two resistive elements. For 1 R1, the input is eort e R1and the output is ow f 1. For  R2, the input is a ow and the output is an eort. These directions are dictated by the propagation of causal signals through the network. Auxiliary variables are determined based on this causality analysis. Namely, ow f 1and eort e R2 are auxiliary variables. Because there exists a causal sig- nal propagation path from exogenous input ue1to aux- iliary variable f 1, this auxiliary variable is a function of the exogenous input: f 1= On the other hand, only state variable pis involved in the causal path of the auxiliary variable e R2. Therefore, this aux- iliary variable is causal: e R2=  R2(I(p)). This auxil- iary variable can be used as an observable for lifting the space. 3 Augmented Lifting Resistors Connected to a Loop Junction We are interested in modeling systems that obey the generalized Kirchho loop and node rules [15] and con- tain resistive elements. As discussed in Section 2, an- ticausal observables result only from elements like resistors and dampers. Such elements can be connected to other elements via a loop or a node. Fig. 3 illustrates the two cases in which a resistor's con- nectivity is governed by Kirchho's Loop Rule: either the output of the resistor is a ow variable, as shown in Fig. 3 (A), or the output of the resistor is an eort vari- able, as shown in Fig. 3 (B). Note that the output e in Fig. 3 (B) comes from the right subsystem, as indicated with a red path. This allows exogenous inputs from that subsystem to be involved in ow variable f. In Fig. 3 (A), the output f comes from both subsystems, providing two paths from which to involve exogenous input. Here, the objective is to intercept the causal paths con- necting the subsystems to the resistor output, any exogenous input coming through the sub- systems from being algebraically involved in the observ- 4flow - effort subsys .PhRp f PhIflow - effort subsys .eeIFig. 4. The system in Fig. 3 augmented with an inertial el- ement. The inertial element produces output f in response to input e Idue to the inertia's inherent causality. This dic- tates the resistor to alter its input-output causal direction. As a result, the new output e is a nonlinear function of mo- mentump, which is an independent state variable, as shown with the red path from pto e. By Kirchho's Loop Rule, a ow variable is common to all the elements connected in the loop. Therefore, the output of the inertial element f is distributed to the resistor and the subsystems. In turn, the input e Iis determined such that all the eorts along the loop sum to zero. able. This can be achieved by adding an inertial to the loop junction. As illustrated in Fig. 4, the inertial element added to the system must output a ow variable f in response to the eort input e Idue to the causality, as addressed previously. The addition of the inertial element dictates that the output of the resis- tor is an eort. As a result, the new auxiliary variable e comes from the state of the inertia, i.e. momentum p, as indicated by the red path, and cannot be directly in u- enced by exogenous input from the subsystems. There- fore, the new auxiliary variable e can now be used as an observable to safely lift the state space. 3.2 Resistors Connected to a Node Fig. 5 illustrates the two cases in which a resis- tor's connectivity is governed by Kirchho's Node Rule, with Fig. 5 (A) showing the general case for eort-output and Fig. 5 (B) for ow output. These represent dual cases for the systems illustrated in Fig. 3. By swapping the ef- fort and ow variables in Fig. 3, we can obtain the block diagrams in Fig. 5. Along the same line, by swapping ef- fort and ow variables in Fig. 4, we can obtain a system with no causality problem. See Fig. 6. Note that, in lieu of the inertial element, a capacitive element is added to effort - flow subsys .PhRe feffort - flow subsys .flow - effort subsys .PhR-1e feffort - flow subsys . (A) (B) Fig. 5. Two possible congurations of a resistor connected to a node junction. effort - flow subsys - flow subsys .eFig. 6. The system in Fig. 5 augmented with a to solve the causality problem. the original system. In all cases, causal paths are retained within the local elements consisting of the resistor and either an inertial or capacitive element added to the system. This pro- hibits the causal paths from adjacent subsystems that may contain exogenous inputs. 3.3 Re-Routing Causal Paths The above method is summarized in the following Consider a dynamical system of where element connections are governed by the generalized Kirchho's loop and node rules. Causal paths from subsystem containing exogenous input can be prohibited from reaching the outputs of resistive ele- ments by either adding an inertial element if the resistor is connected to a loop junction, or a capacitive element if the resistor is connected to a node junction. If a resistor is connected to a loop junction, the output of the aug- mented resistor is given by e =  R(I(p)) where  R(f) and  I(p) are the constitutive laws of the resistor and added inertia, respectively. If the resistor is connected to a node junction, the output of the augmented resistor is given by f = 1 R(C(q)) where 1 R(e) is the law of the resistor and  C(q) is the consti- tutive law of the capacitor. These output variables are causal and can therefore be used to lift the dynamics. Remark 1. The added inertial and capacitive elements can be linear with mass m and capacitance C such that f =p=m and e = q=C. Because these are linear elements, the outputs of these elements are not ; no lifting is necessary for f and e.  Remark 2. Small values should be picked for the mass m and capacitance C so that the impact upon the dy- namical system may be small. To select an initial value, one can use a Bode plot of the linearized system to de- termine the bandwidth of operation. The mass or ca- pacitance can be selected so that it is small enough such that its breakpoint frequency is signicantly higher than those of the other elements of the system. Thereby, the 5dynamics of the system can remain mostly unaected in the low-frequency band useful for modeling.  In the icon model of Fig. 1, the plate to which the spring and damper  R1are attached was assumed massless. This resulted in an anticausal auxiliary variable f 1. Ap- plying the proposition, we add a small inertia to resolve the anticausality problem, eectively modeling the plate in Fig. 1 with a small mass. It is a modeling the mass of the plate is considered or it has a signicant impact on the causality of the lifted dynamics. Remark 3. In the proposition, an energy-storage el- ement, an inertia or a capacitor, is added to the loop or node junction to which the resistor of interest is connected. However, the energy-storage element can be added to other parts of the system so long as it prohibits the causal path from exogenous inputs from reaching the resistor. There may be multiple solutions to a given anticausal system.  Remark 4. If the output of a resistor is directly driven by an exogenous input, e.g. e =  R(uf) or f = of e or f do not re ect any dynamics of the system. Therefore, these auxiliary variables can be omitted from the lifted model.  4 Lifting with Integrated Auxiliary Variables As discussed above, lifting linearization through aug- mentation of the original system can solve the of auxiliary variables. However, the method in- evitably alters the original dynamics, which, although slight, may not be allowed in some applications. This section introduces an alternative method that does not alter the dynamics, but creates a virtual that energy-storage elements do not incur the causality problem in lifting the system. Each energy- storage element contains an integrator that possesses an independent state variable. The time derivatives of these state variables are, by denition, given by dp=dt = e and dq=dt = f where eort and ow variables, f and e, respec- tively, can be expressed as functions of state variables, and exogenous inputs. Time deriva- tives of input are not involved in these expressions. This leads to causal state equations. Based on this observation, now consider the following rick to resolve the causality problem. Suppose that a virtual instrument measures the integral of the out- put of an element. Let the output of a nonlinear, dissipative element that is anticausal for lifting. The virtual instrument Suppose that we use this integrated output of the element for Exogenous Input, usubsys . 7. A dynamic system with an anticausal observable aug- mented with an integrator for causal lifting. lifting the dynamics. Namely, the time derivative of (t) is given Because no input deriva- tive is involved, the dynamics are causal. This method, in a sense, treats the el- ement like an energy storage element, which possesses a state variable. Unlike the rst method, where the original system is altered by adding a small inertial or , this second method does not alter the dynam- ics. The integrator attached to the el- ement is only used for observation via a virtual instru- ment. As shown in Fig. 7, the output of the virtual not in uence the plant dynamics. The above algorithm using Integrated observables for Lifting Linearization (IL2) is summarized below. IL2 Algorithm Use DFL to select naauxiliary vari- ables,(t) =(x(t);u(t)). LetnRube the number of nonlinear resistive elements in which u(t) is involved. All the auxiliary variables associated with these input- dependent resistive elements are placed in a Introduce a new (5) (2) Remove Ru(x(t);u(t)) from the set of augmented state variables (t). (3) If it is not already a state variable, append  Ru;i(t) to the set of state variables x(t), and (4) Append _ Ru;i(t) to the state transition dynamics, f. (5) Proceed to further lift the augmented state using an basis and/or regress a linear dynamic model to evolve the augmented state. Remark 5 In some cases,  Ru(t) is already included in the state. If this happens, then the above simply removing anticausal variables from the lifted a simple massless spring-damper system by the following dynamics: _x(t) = (6) where  Rand  Care nonlinear constitutive laws and the observables are selected using DFL to be Note that the second observable, f, is a nonlinear function of exogenous input u, and there- fore cannot be used for lifting. However, the integration of f(t) based on equation (5) is the same as the state, x: =x. Therefore, the inte- grated auxiliary variable does not make any The IL2 algorithm instructs ignoring f al- together. 5 Numerical Results This section implements the augmented and linearization algorithms on a variety of dynamical systems with anticausal and compares their performance against exist- ing methods. The algorithms, written in SciPy [25], were computed on a laptop running Ubuntu 18.04.5 LTS. The codebase is hosted as a git repository at [20]. We used Adam's real-valued ordinary dier- ential equation solver [5] to perform the integration. We benchmark AL2 and IL2 against the following Linearization (DFL) [7], including the techniques to select observables using outputs of nonlinear elements connected to a lumped parameter model with integral causality. We also implement the linear anticausal lter to attempt to remove the eect of control input from Lifting Linearization (L3) [21] with a neural network with two hidden layers of 256 ReLU neurons trained in batches of 32 using an Adam opti- mizer [12] with = 105,1= 0:9,2= 0:999, and = 108until validation error begins to rise. We im- plemented the machine learning component of L3 in PyTorch [18]. Koopman with Synthetic Observables of State (KSOS) All the observables are synthetically gener- ated as nonlinear functions of state variables alone. No exogenous inputs are involved in the a polynomial basis of dimension 8 is used for generating based on all Measured Quantities (OMQ) Emulating the DMD data-driven frame- work, data are generated based on the same quantities as DFL. All the observables are nonlinear functions of the measured variables, which may include anticausal observables. State and aux- iliary variables are treated as which can be measured, and a polyno- 0.0 0.2 0.4 0.6 0.8 1.0 time 8. Simulation of numerical example from equation (7) given a square wave control input with no noise. In the pres- ence of moderate nonlinearity, AL2 and IL2 perform about as well as state-of-the-art techniques. mial basis of order 8 is used for generating Simple Numerical Example Consider the following dynamical system with states p andqand control input u: d dt p q! =  C, R, and  Iare nonlinear constitutive laws as- sociated with capacitive, resistive, and inertial elements in the system, on DFL, we select the following is a nonlinear function of control input and is therefore anticausal for lifting. For the implementation of AL2, we augment the system by adding a small mass m 0with a new state variable p where and replace (4)with a new observable e R=  R(p=m0). For the implementa- tion of IL2, we replace augment equation (7) with _=  no noise in all measurements, we begin by simulating this system with the following  C(q) = sgn(q)q2, R(eR) = sgn(e R)eR4, and I(p) =p3. The exogenous input given to the system is a square wave. The Sum of Squared Error (SSE) of each method is plotted in Fig. 8. IL2 outperforms all other methods. In DFL and L3, it is assumed that the exoge- nous inputs are linearly involved in all the variables, as shown in equation (1). Since this assumption does not hold due to the nonlinearity in  R, the DFL and L3 models exhibit signicant errors with SSEs of 270 and 17 times that of IL2, respectively. OMQ AL2 IL2SSEFig. 9. Average sums of squared error (SSE) for each algo- rithm modeling the same numerical example from equation (7) with added nonlinearity given ten random control input signals and no noise. Standard error of each SSE is illus- trated as error bars. The AL2 and IL2 model outperforms all other numerical models. in [21], the OMQ model does not track the signal due to the inclusion of anticausal observable (4)without rst ltering out the dependence on u. The KSOS model, which did not include any auxiliary variables but instead lifted only with the state variables, achieved an SSE of 29 times that of IL2. 5.2 Increasing Nonlinearity To further examine the eect of system nonlinearity on the performance of AL2 and IL2, we repeat the same ex- periment using  I(p) = 3p3p3and  R(eR) = 3e R3 3eR, reducing the coecient of linear determination of both constitutive laws by 22%. We repeat the experi- ment ten times using dierent random control input sig- nals and plot the average sums of squared error for each method in Fig. 9. The added nonlinearity breaks the un- derlying linearity assumption behind the anticausal l- ter, so the DFL and L3 models are omitted from this ex- periment. The OMQ model remains unable to track the system, and the KSOS model achieved an SSE double that of IL2. 5.3 Comparison to Koopman KSOS shares fundamental similarities with AL2 and IL2. Neither KSOS nor AL2/IL2 lift the space with an- ticausal observables. However, the proposed algorithms have two key dierences. First, IL2 uses the integral of an anticausal observable as a new causal observable for lifting the space. At the time of writing, the authors are not aware of any Koopman-based technique that uses an integral function, but it is precisely the use of integra- tion that makes IL2 so eective. Second, AL2 and IL2 leverage DFL, which allows for the selection of informa- tive observables so that the order of the linear dynamic model may be kept relatively small. Fig. 10 illustrates the results of the experiment from Fig. 8 with six dierent implementations of KSOS: three with polynomial bases of dimension 8, 16, and 32, and three with Fourier bases [13] of dimension 8, 16, and 32. 0.0 0.2 0.4 0.6 0.8 1.0 time 8 Poly. 16 Poly. 32Fourier 8 Fourier 16 Fourier 32AL2 IL2Fig. 10. Comparison of AL2 and IL2 to KSOS with dierent basis functions. For comparison, the orders of the AL2 and IL2 models are only 7 and 6, respectively. As shown, improving the basis of KSOS without knowledge of the connectivity of elements does little to improve the results. This high- lights the importance of accessing informative signals rather than manipulating basis functions alone. 5.4 Noise in Observables Up to this point, all of the experiments illustrated have been noiseless. The KSOS algorithm, in particular, has beneted from the lack of noise in simulations. In this next experiment, we compare the noise characteristics of two approaches, DFL and KSOS. We repeat the experi- ment modeling the system in (7) with a nonlinear consti- tutive law,  R(eR) = 1=(1 + exp(4eR))0:5). We re- peat the experiment twice: once without noise, and once with zero-mean Gaussian noise with = 0:03 added to each measurement, and compare the results in Fig. 11. It is clear that DFL is considerably more robust in com- parison to simply using the KSOS approach. This sup- ports the argument in Appendix A showing that KSOS is more susceptible to noise as noisy signals are trans- mitted through nonlinear functions. 5.5 Use Case for Inertial Augmentation In the above numerical examples, the integrated models have consistently outperformed the augmented lifting linearizations. Now we consider a case where the auxiliary variable cannot be replaced with DFL DFLSSENoiseless Noisy Fig. 11. Comparison of model performance system from (7) with and without noise. Because it uses synthetic observables, KSOS degrades much faster 12. CDF of SSE for AL2 and IL2 models on the noise- less, massless, nonlinear spring-damper system dened by equation (6). AL2 has a lower median error than IL2, but a greater 90thpercentile SSE. integral, because the integral is already an variable. Let us revisit the situation in equation (6), where the auxiliary variable f replaced with ,R f =x, butxis already an independent state variable. Therefore, IL2 instructs us to ignore important information about the dynamics of the system. In this case, AL2 can outperform IL2. Because the performance of the modeling with the shape of the control input signal, we var- ied it. We simulate the massless spring-damper with con- stitutive equations 1 R(eR) = 3e R(eR+ 1)(e R1) and C(x) =1 R(x) 1000 times with random control in- put signals and plot the Cumulative Distribution Func- tion (CDF) of the SSE for each algorithm in Fig. 12. In most trials, the AL2 algorithm outperformed IL2 (me- dian SSE of 1.0 and 1.6, respectively), but the IL2 algo- rithm had a more consistent performance with a error of 2.1 compared to 2.4 for AL2. 6 a nonlinear control system using measured ob- servables results in an anti-causal lifted model if exoge- nous inputs uare involved in the observables (x;u). Based on physical modeling theory, this paper has ad- dressed (a) how exogenous inputs propagate through a nonlinear dynamical system and can reach outputs of nonlinear elements, a class of observables causing the causality problem, and (b) how the causality problem can be solved. Observables that are outputs of nonlin- ear energy-storage elements never include exogenous in- puts, while elements may include ex- ogenous inputs. Two methods have been presented for resolving the causality problem of energy dissipative el- ements. The rst method augments the system with ei- ther a linear inertial element or a linear capacitive ele- ment, thereby reversing the causality of the energy dissi- pative element and resolving the causality problem. The second method presented replaces input-dependent ob- servables with integrals of linearly related causal observ- ables, which also eliminates the causality problem inher- ent in predicting future values of an exogenous signal. Using these techniques, we synthesize a basis capableof lifting the state space of a nonlinear dynamic system into a more linear regime for modeling. The conrm that these techniques model more accurately than any of the against which we benchmarked. This work contributes to lling the gap between the Koopman- based lifting linearization of nonlinear autonomous sys- tems and the lifted linear model of a nonlinear control system used for control Ian Abraham and Todd D Murphey. Active learning of dynamics for data-driven control using koopman operators. IEEE Transactions on Robotics , 35(5):1071{1083, 2019. [2] Alexander Broad, Todd Murphey, and Brenna Argall. Learning models for shared control of with unknown dynamics. In Proceedings of and Systems , Cambridge, Massachusetts, July 2017. [3] Daniel Bruder, Brent Gillespie, C. David Remy, and Ram Vasudevan. Modeling and control of soft robots using the koopman operator and model In Proceedings of Robotics: Science and Systems Germany, June 2019. [4] Scott T. M. Dawson, M. Hemati, Matthew O. Williams, and C. Rowley. Characterizing and correcting for the eect of sensor noise in the dynamic mode in Fluids , 57:1{19, 2014. [5] Ernst Hairer, Syvert P Nrsett, and Gerhard Wanner. Solving ordinary dierential equations. 1, Nonsti problems 1993. [6] Yiqiang Han, Wenjian Hao, and Umesh Vaidya. Deep learning of koopman representation for control. In 2020 59th IEEE Conference on Decision and Control (CDC) , pages 1890{1895. IEEE, 2020. [7] H Harry Asada and Filippos E Sotiropoulos. Dual of nonlinear dynamical systems based on physical modeling theory. Journal of Dynamic and Control , 141(2), 2018. [8] Yusuke Igarashi, Masaki Yamakita, Jerry Ng, and H Harry Asada. Mpc performances for nonlinear systems using models. In 2020 American Control Conference (ACC) , pages 2426{2431. IEEE, 2020. [9] Yusuke Igarashi, Masaki Yamakita, Jerry Ng, and H Harry Asada. Mpc performances for nonlinear systems using models. In 2020 American Control Conference (ACC) , pages 2426{2431. IEEE, 2020. [10] Yusuke Igarashi, Masaki Yamakita, Jerry Ng, and H Harry Asada. A robust method for dual faceted linearization. , 53(2):6095{6100, 2020. [11] Dean C Karnopp, Donald L Margolis, and Ronald C Rosenberg. System dynamics: modeling, simulation, and control of mechatronic systems . John Wiley & Sons, 2012. [12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on ICLR 2015, San Diego, CA, USA, May 7- 9, 2015, Conference Track Proceedings , 2015. [13] George Konidaris, Sarah Osentoski, and Philip Thomas. Value function approximation in reinforcement learning using the fourier basis. In Twenty-fth AAAI conference on articial intelligence , 2011. 9

2108.10980

Nicholas Selby

Nicholas S. Selby and Filippos E. Sotiropoulos and H. Harry Asada

Physics-Based Causal Lifting Linearization of Nonlinear Control Systems Underpinned by the Koopman Operator

10 pages, 12 figures, submitted to Automatica

eess.SY cs.SY

http://creativecommons.org/licenses/by/4.0/

Methods for constructing causal linear models from nonlinear through lifting linearization underpinned by Koopman operator and physical system modeling theory are presented. Outputs of a nonlinear control system, called observables, may be functions of state and input, input-dependent observables cannot be used for lifting the system because the state equations in the augmented space contain the time derivatives of input and are therefore anticausal. Here, the mechanism of creating is examined, and two methods for solving the causality problem in lifting linearization are presented. The first method is to replace by their integral variables $phi^*$, and lift the dynamics with $phi^*$, so that the time derivative of $phi^*$ does not include the time derivative of input. The other method is to alter the original physical model by adding a small inertial element, or a small capacitive element, so that the system's causal relationship changes. These augmented dynamics alter the signal path from the input to the anticausal observable so that the observables are not dependent on inputs. Numerical simulations validate the effectiveness of the methods.

2 Introduction: In recent years, there has been an intense study of the magnetic textures known as skyrmions, given that their quantized, winding- like spin structure leads to novel behaviors that are attractive from both f undamental and applied viewpoints.1,2 Particularly, the unique spin topology of skyrmions lends them a degree of topological stability and offers avenues towards their efficient motion3 - making them potentially relevant in developing next- generation data storage and novel computation schemes.4 Previous work has shown that while it is possible to stabilize skyrmion phases through an intricate balance of the ferromagnetic exchange, dipolar, and anisotropy energies of a material system5, the Dzyaloshinskii -Moriya interaction (DMI)6,7 - an antis ymmetric exchange interaction that favors the orthogonal canting of neighboring magnetic spins - often plays a crucial role. For example, t he DMI associated with the non- centrosymmetric B 20-type crystal structure was central to the first experimental obser vation of a skyrmion phase.8 The interfacial DMI (iDMI) that develops at the interface between thin ferromagnetic (FM) layers and heavy metals (HM) with significant spin -orbit coupling9 also assists with the formation of skyrmion phases. There have bee n numerous reports on multi -repetition FM/HM heterostructures with perpendicular magnetic anisotropy (PMA), where a large iDMI assists with the formation of skyrmion phases , and the multiple repetitions of the unit structure increases the total volume of m agnetic material, leading to enhanced skyrmion Additionally , there have been several reports of skyrmion phases in single magnetic layer HM/FM/HM systems with a modest iDMI, where the FM layer thickness was chosen to be large enough that the system is close to experiencing a spin reorientation transition and consequentially has a low effective PMA ( Keff= Ku - 0MMss2 2, where K u is the intrinsic anisotropy constant and 0MMss2 2 reflects the shape anisotropy of a thin film) .19,20,21,22,23 The domain wall energy density s in systems with i DMI is often approximated as where Aex and |D| are the exchange stiffness and iDMI energy density, respectively.24 Thus, by lowering Keff while maintaining | D| or vice versa, it is possible to lower s , reducing the energy barrier to forming multidomain states, including the skyrmion phase.25,26 In this work, we explore the approach of reducing A ex to lower the energy of forming domain walls and facilitate transitions to multidomain and skyrmion phases in ultra thin films while maintaining a reasonable K eff and modest D . Given that the domain wall width is typically estimated as =AAeeeeKKeeeeee/ , lowering Aex while maintaining significant K eff to obtain skyrmion phases has the added benefit of maintain ing a lower domain wall width than would be expected from lowering K eff. We report the observation of skyrmion phases at room temperature in several FM/HM including Pt/Co, Pt/Co/NiCu, Pt/Co/Ni, and Pt/Co/Ni/Re , with either one or two repeats of the unit structures listed. While these films are diverse in composit ion, they share several commonalities, including a modest iDMI arising from the asymmetry of the FM/HM interfaces , significant PMA , and a low exchange stiffness . For the Pt/Co, Pt/Co/Ni, and Pt/Co/Ni/Re samples, the exchange stiffness is lowered through the use of ultrathin (< 0.3 nm) Co layers.27,28 Similar properties were obtained in a thicker Pt/Co/NiCu sample by increasing the proportionality of Cu to Ni in the alloy layer.29,30 Concomitant to these changes in the static magnetic properties, we find that as the calculated domain wall energy density diminishes towards negative values , the domain morphology transitions dramatically, permitting the stabilization of labyrinthian stripe domains at remanence and field- induced magnetic skyrmion phase s within a narrow window of environmental conditions ( i.e., temperature or applied magnetic field). By exploiting the spin- orbit torques generated when passing an electrical current through the sample, skyrmion motion including the skyrmion Hall effect is observed. Using micromagnetic and analytic modeling, we de monstrate the strong relationship between the exchange stiffness and the equilibrium domain morphology in thin films with modest iDMI and significant PMA . Additionally, we show how the associated changes in the energy landscape allow for the stabilization of field-induced skyrmion phases within a framework of understanding like that of previous material but through control of an unconventional parameter - the exchange stiffness. 3 Experimental Methods: Samples of the structure [Co ( tCo)/ Pt (1)] 1,2, [Co ( tCo)/ Ni (0.3)/ Re (0.5)/ Pt (0.5)] 2, and [Co (t Co)/ Ni (0.3)/ Pt (1)] 2, and [Co(0.4)/Ni 40Cu60(0.9)] 1 (thicknesses in nm ) are used in this study. In the text, these samples will be referred to as the [Pt/Co] 1,2, [Pt/Co/Ni/Re] 2, [Pt/Co/Ni] 2, and [ Pt/Co /NiCu ]1 samples, respectively . A Ta (2)/ Pt (5) seeding layer and Ta (3) capping layer were used for all samples. The samples were grown using dc magnetron sputtering at ambient temperature onto Si substrates with a 300 nm -thick thermal oxide coating. A 3 mTorr partial pressure of Ar and a sputtering power of 50W were used . For some samples, the Co layers were grown as wedge - type structures using a slit placed on top of the Co sputtering source - allowing for a 14 % thickness gradient to be achieved during a single deposition over a 4 cm -long substrate. Some samples were patterned into 100- um- wide wires using conventional metal lift- off UV Temperature -dependent magnetometry was performed in the out -of-plane (OOP) and in -plane (IP) geometries using vibrating sample magnetometry (VSM). When calculating volumetric parameters from the magnetometry data, we assume that any Pt within 0.2 nm of an interface with a FM layer has become magnetically polarized.31 Magneto -optic Kerr effect ( MOKE ) images and hysteresis loops were ob tained using an Evico Magnetics microscope, in which perpendicular magnetic fields were applied using an air coil. For some MOKE measurements, the sample temperature was varied using a Peltier chip. For t he MOKE images presented within, images of the sampl e when the magnetization was saturated have been subtracted as a background. Micromagnetic modeling was performed using the MuMax3 solver32, employing a 5 um x 5 mm x 1.5 nm geometry discretized into 2 nm x 2 nm x 1.5 nm cells. Periodic boundary conditions were used to account for ten repetitions of the simulation geometry within the film plane . Experimental Results: We begin by discussing experimental results for the [Pt/Co] 2 samples. While the Pt/Co/Pt structure is nominally symmetric, previous work has shown that structural asymmetry between the Pt -Co and Co- Pt interfaces can lead to the development of PMA and a modest iDMI.33,34 From measurements of the domain expansion velocity as a function of in- plane field for a [Pt/Co(0.35 nm)] 2 sample , we observe asymmetries that are typically associated with the left -handed iDMI thought to originate from the lower Pt -Co interface34 and estimate (using techniques described elsewhere24,35) a |D| of ~ 0.14 mJ/m2 (see Supp. Note 1 and Supp. Fig. 136). In agreement with previous studies, the saturation magnetization as a function of temperature M s(T) plots provided in Fig. 1(a) show a strong dependence on t Co, and that TC decreases as t Co is reduced within the ultrathin limit.27,28 In particular, Fig. 1(a) demonstrates that the sample with t Co = 0.32 nm is thermally stable at room temperature, as T C is well above 300 K. However, decreasing t Co to 0.24 nm suppresses the room temperature M s, and the Ms(T) curve demonstrates a TC of ~365 K - indicating a significant lowering of Aex. Further decreasing t Co to 0.2 nm suppresses TC to near room temperature. While it is anticipated that the uniformity of the Co layers may be compromised in the ultrathin limit, previous reports have shown that well-defined FM behavior is observed even with sub- monolayer FM coverage, provided the FM la yer is abutted by materials susceptible to proximity- induced magnetism ( e.g., Pt, Pd).27,28 In Fig. 1(b), we show room temperature polar MOKE hysteresis loops for [Pt/Co] 2 samples with different tCo. As expected for FM/HM heterostructures with PMA, we find that films with t Co >= 0.26 nm have square hysteresis loops. However, r educing tCo to 0.25 nm or 0.24 nm, sheared polar MOKE loops are observed - suggesting a multidomain state at remanence in these samples. Further reducing t Co to a nominal thickness of 0.23 nm, the polar MOKE loop has zero remanent magnetization and a hard axis -like c haracter ; as Fig. 1(a) demonstrated that T C decreases dramatically in this regime of t Co, this may indicate a paramagnetic -like character to this sample . To understand how these changes in the static magnetic properties impact the room temperature domain m orphology, we have collected polar MOKE images at zero field [Figs. 1(c -f)]. Before capturing an image, the following field history protocol was employed: First, the sample magnetization was saturated in the negative perpendicular field direction (-Mz). The magnetic field was th en gradually increased towards each 4 sample's respective positive coercive field (+ Hc) before removing the magnetic field. For samples with t Co >= 0.26 nm, circular magnetic domains were observed at remanence - as is expected for th in films with PMA. In agreement with the polar MOKE hysteresis loops, a labyrinthian stripe domain pattern with a sporadic density of circular domain textures was observed at remanence when t Co was reduced to 0.24 nm. Further reducing t Co to 0.23 nm, the f eatureless remanent state may indicate an in-plane magnetic anisotropy , a paramagnetic- like character to this sample, or that the domain morphology is smaller and/or fluctuating faster than our polar MOKE system can resolve. Overall , Figs. 1(c -f) demonstrate that the domain morphology evolves dramatically as a function of thickness in the limit of ultrathin Co. In Fig s. 2(a,b), we provide more detailed , temperature- dependent magnetometry data for the [Pt/Co(0.24 nm)] 2 sample. From OOP hysteresis loops collected in the temperature range of 294 K to 302 K [Fig. 2( a)], the extreme temperature sensitivity of this sample 's magnetic properties can be seen. For T <= 292 K, square hysteresis loops are observed, as is common for thin films with PMA. As the temperature is increased over the range 294 K <= T <= 302 K, the hysteresis loop become progressively more sheared - suggesting a multidomain remanent state. For T >= 302 K, the hysteresis loops take on a hard axis -like character ( i.e., low remanence, low hysteresis) . Thus, the evolution of the OOP hysteresis loops with increasing temperature is qualitatively similar to the evolution with decreasing Co thickness shown in Fig. 1(b). The corresponding IP hysteresis loops , collected in the same temperature range over which the OOP loops exhibited such dramatic changes, are shown in Fig. 2(b). Surprisingly, we find that the IP saturation field is relatively large, giving an anisotropy field (estimated from the saturation field) of u 0HK 500 mT that does not appreciably change with temperature - suggesting that the temperature dependence of K eff is primarily scaled by M s. The magnetometry data collected at T = 302 K shown in Figs. 2(a) and 2(b) demonstrate that both the OOP and IP hysteresis loops have zero remanence and little to no hysteresis. This evolution in magnetic response with temper ature, including the formation of stripe domain s, is similar in many respects to that observed during the spin reorientation transition (SRT) that occurs in the limit of ultrathin FM layers, when the intrinsic PMA equals the shape anisotropy ( i.e., when K eff = 0).15,16 However, above the SRT temperature, a finite remanence and low saturation field is expected in the IP loops - behavior that is quite different from our observations. A further comparison of our results to the previously studied SRT is provided in the discussion section. In Figs. 1(c-f), we show polar MOKE images of the [Pt/Co(0.24 nm)] 2 sample as the perpendicular magnetic field strength u0Hz was quasi -statically decreased towards negative saturation ( -Mz). As mentioned previously, a labyrinthian domain pattern with a sporadic population of circular features is present i n this sample when u0Hz = 0. As in other systems whose static magnetic properties are extremely sensitive to temperature, we observe that the remanent domain pattern experiences a significant degree of thermal fluctuation when observed over several seconds19, particularly for the [Pt/Co] 1 sample (see Supp. Video 136). When a sufficiently strong u 0Hz is applied, some stripe domains collapse into circular features with a characteristic diameter of ~ 3 mm. Further increasing u0Hz causes all the stripe domains to collapse, forming a disordered ensemble of circular textures. As u0Hz approaches the negativ e saturation field, the circular features begin to disappear, and eventually, the film exhibits a uniform magnetization. A video of this evolution in domain morphology with u 0Hz is provided in Supp. Video 2.36 While this field -induced stripe -to-skyrmion transition can be observed at room temperature in the [Pt/ Co(0.24 nm)] 2 sample, we note that by heating the [Pt/Co(0.26 nm)] 2 sample ( i.e., the sample whose room temperature domain morphology is shown in Fig. 1(c)) s lightly, similar behavior is observed (Supp. Fig 236). To confirm that these field-induced circular features are magnetic skyrmions with chiral domain walls , their response to applied electrical currents was determined . As the samples were seeded and capped with HM materials thought to possess spin- Hall angles of opposing sign (Pt and Ta, a net spin- orbit torque is expected to act on the FM layers when a current is passed through the structure; as such, the skyrmions should move collinear to an applied electrical current (not considering any other dynamic effect s). Furthermore, if the features possess the quantized topology characteristic to skyrmions, they should also experience a Magnus force 5 when driven, leading to a deflection transverse to the applied current axis - a phenomenon known as the skyrmion Hall ef fect.39,40,41 As demonstrated in Fig. 3 (and Supp. Videos 3 and 436), the circular features stabilized by u0Hz in the [Pt/Co(0.24 nm)] 2 sample move in the direction of the conventional current density J ; knowing the spin- Hall angles of the Pt seeding and Ta capping layers, this response is consistent with spin- orbit torques acting on the left -handed chirality observed in other Pt/Co- based As shown in Supp. Video 4, we also observed a dramatic increase in the number of skyrmions present in the wire after an electrical current has been applied - a behavior that may be attributed to the previous ly reported mechanisms by which skyrmions can be generated from stripe domains subject to spin- orbit torques and sporadic pinning sites.47 The skyrmion Hall effect is discerned using a technique similar to that employed in previous works19, whereby skyrmions were ejected from a 100 mm -wide wire (where the skyrmions move with a speed too fast for our system to image) into a 3 mm -wide contact pad where the skyrmions slow, accentuat ing any transverse motion. From Fig. 3 (and Supp. Video 536), it can be seen that the features exhibit a transverse deflection relative to the appli ed current axis and that the direction of this deflection reverses when changing between + Mz and -Mz polarity domains, as would be expected from the skyrmion Hall effect.19,41 Considering these findings in [Pt/Co] samples in the limit of ultrathin t Co, we now examine how modifying the HM and FM layers to increase the structural inversion asymmetry (and hence, the total iDMI) impacts the morphological phases accessible in the ultrathin limit. We highlight two other material systems: [Pt/Co(0.24 nm)/Ni/Re] 2 and [Pt/Co(0.24 nm)/Ni] 2 (thicknesses in nm) - all featuring interfaces (Pt-Co, Ni -Re, and Ni -Pt) thought to contribute to an additive iDMI acting on the FM layers due to their compositional asymmetry ( i.e., in addition to an asymmetry in interfacial quality, as in the [Pt/Co] samples) .48,49 While the Co layers of the two structures mentioned above were grown as wedges, we indicate the thickest approximate tCo at which a u 0Hz- induced skyrmion phase was observed at room temperature. The remanent state of the [Pt/Co/Ni/Re] 2 sample shown in Fig. 4(a) succinctly demonstrates the compositional sensitivity of the domain morphology in the ultrathin limit, as the stripe domain size dr amatically decreases when t Co is varied by as little as ~ 0.05 %. We emphasize that while Fig. 4(a) is reminiscent of previous studies that showed a strong link between film thickness and domain periodicity on account of proximity to a SRT15,16,17,18, the static characterization does not indicate the presence of the SRT in our FM/HM heterostructure samples near room temperature, given the large in- plane saturation field indicating a significant PMA . From the higher -magnification polar MOKE images shown in Figs. 4(b , d), it can be seen that the [Pt/Co/Ni/Re] 2, and [Pt/Co/Ni] 2, samples exhibit a labyrinthian stripe pattern in zero field at room temperature. In line with the [Pt/Co] 2 samples, temperature- dependent magnetometry of these samples indicates TC values below 400 K and OOP hysteresis loops extremely sensitive to variations in temperature in the vicinity of 300 K. Much like the [Pt/Co] 2 samples, the stripe domains formed at remanence in the [Pt/Co/Ni/Re] 2 and [Pt/Co/Ni] 2 samples transform into circular textures when small perpendicular fields are ap plied. By observing these features' response t o spin- orbit torques (using the process previously described for the [Pt/Co] 2 samples), it is possible to classify these features as skyrmions. Figures 4(c, e) show polar MOKE images collected at the u 0Hz value where each sample exhibits its densest skyrmion phase . Thus far, the materials systems discussed have shared the commonalities of an ultrathin (< 0.3 nm) Co layer and a low A ex (inferred from a low T C). However, if a low A ex is responsible for the observed domain morphologies, it should be possible to attain similar results in systems with thicker FM layers. Previous works have shown that alloying Ni with Cu can lower T C relative to that of elemental Ni and that this trend is linear with respect to Cu concentration.29 Based on this, we fabricated a sampl e (thicknesses in nm) to explore the role of exchange stiffness by optimizing the Cu- Ni ratio . From magnetometry data collected near room temperature [Fig. 5(a)], it can be seen that the [Pt/Co/NiCu] 1 sample's OOP hysteresis loops are extremely sensitive to temperature; the square loop present at 285 K becomes more sheared as the 6 temperature is increased to 290 K, and becomes hard axis -like when the temperature is increased to 295 K. These findings are similar to those shown for the [Pt/Co(0.24 nm)] 2 sample shown in Fig. 2(a). Much like the [Pt/Co] 2 samples with ultrathin Co layers, the IP hysteresis loops of the [Pt/Co/NiCu] 1 sample do not change significantly in the vicinity of room temperature [Fig. 5(b)]; given that the sign of K eff does not change over the relevant temperature range , this does not indicate the presence of a SRT near room temperature. The M S(T) curve shown in Fig. 5 (c) (generated from hysteresis loops collected over a wide range of temperatures) indicates a T C of approximately 360 K for the [Pt/Co/NiCu] 1 sample. From polar MOKE images collected at T = 290 K, it can be seen that the [Pt/Co/NiCu] 1 sample exhibits a labyrinthian stripe domain pattern at remanence near room temperature, with a characteristic domain width of ~ 2 mm [Fig. 5(d)]. When a ~ 0.04 mT perpendicular fie ld is applied, the labyrinthian remanent state is transformed into an ensemble of circular textures [Fig. 5(e)] that display the spin -orbit torque induced dynamics emblematic of skyrmions as seen in the previous ly discussed samples with ultrathin Co layers - indicating that the energetic considerations that give rise to the remanent labyrinthian morphology and field- induced skyrmion phase are not contingent on a reduced dimensionality to the FM layers. Discussion: Thus far , the temperature dependence of the OOP hysteresis loops, the suppressed T C values , and the labyrinthian domain morphologies observed at remanence in our samples draw strong parallels to extensive, past studies of the SRT.15,16,18 Particularly, the thickness sensitivity exhibited in Fig. 4(a) is heavily reminiscent of previous studies of a Cu/Fe/Ni structure that was found t o exhibit a SRT when the Fe thickness was reduced to ~ 1.25 ML.18 While the original understanding of the SRT was developed using systems that do not have iDMI, recent reports have shown that the iDMI can assist in raising the temperature at which the SRT occurs and facilitates a direct transition between the ferromagnetic and paramagnetic states for sufficient D .50 However, despite the many similarities between our HM/FM heterostructures and systems that undergo a SRT, our work is differentiated by the fact that K eff does not approach zero when the labyrinthian stripe phase becomes the remanent state in our samples (whether by adjusting the FM layer thickness or the sample temperature) - a key point that rules out a traditional SRT at the thicknesses and temperatures relevant to our samples. Given that a harbinger of a skyrmion phase in our samples has been the obser vation of a labyrinthian stripe phase at remanence and the treatment of the SRT , while related, may not be applicable, further exploration is warranted. To better understand the factors that give rise to skyrmion phases in thin films with significant PMA, modest iDMI, and low Aex, we have performed micromagnetic and analytic modeling. For our models, a MS of 250 kA/m, a first -order uniaxial anisotropy constant K u of 67.5 kJ/m2 (corresponding to K eff = Ku - 0MMss2 2 = 28.25 kJ/m3 in the thin film geometry), and a film thickness t film of 1.5 nm were used to approximate the [Pt/Co/NiCu] 1 sample. While an A ex of ~ 10 pJ/m is often used when modeling Pt/Co and Pt/Co/Ni -based systems, this generally assumes that T C is closer to the bulk value of Co.51,52,53 To reflect that fact that T C of the [Pt/Co/NiCu] 1 sample is ~ 75 % lower than that of bulk Co, we use an A ex of 2.5 pJ/m in our calculations. The iDMI energy density D was varied between 0 and 1 mJ/m2. All simulations were performed at T = 0 K. In the thin film limit, it is well known that the relative balance between the demagnetization energy and the energy penalty associated with forming a domain wall (both of which depend on M S, K, Aex, and D ) determines whether the film will break into domains in zero To understand this energetic competition in the [Pt/Co/NiCu] 1 sample (which can generally be extended to describe the samples with ultrathin Co layers) , we first performed a simple micromagnetic calculation of the difference in energy density D e between the two -domain state with a single domain wall and the uniform magnetic state ( i.e., De = etwo-domain - euniform ) as a function of D . This calculation of D e allows for a micromagnetic estimation that can be compared to the analytical expression for the domain wall energy density A schematic depiction of the magnetic states considered is provided in Fig. 6(a). The e associated with each state was determined after allowing the domain 7 state to relax from the respective initial state to the minimum energy configuration. From the D e(D) plot provided in Fig. 6(b), it can be seen that the uniform state has a lower e state at lower D - indicating that the energetic pena lty associated with breaking into domains/forming a domain wall outweighs the associated minimization in dipolar energy. However, at a relatively modest value of D (~ 0.28 mJ/m2), De(D) becomes negative, suggesting that the energetic penalty to forming a domain wall has been diminished and the energy landscape is more conducive to the formation of multidomain states. If instead, an A ex of 10 pJ/m is used [Fig. 6(b)] - as is often the case when describing Pt/Co- based systems - it can be seen that the root of D e(D) is shifted towards D values which are generally only feasible in material systems optimized to have a large iDMI.49,58 or by reducing the effective anisotropy. Calculations of the domain wall energy density under the same considerations - also shown in Fig. 6(b) - show similar trends as the D e(D) simulations. We note that accounting for the domain wall anisotropy energy density in the net s only slightly modifies the curve shown in Fig. 6(b). While these res ults are most applicable to the [Pt/Co/NiCu] 1 sample , we show how these trends in D e and s evolve with Aex, K, and D with greater detail in Supp. Fig. 336. In keeping with our experimental results, as A ex and D are not expected to change dramatically when varying the temperature by several K59,60,, we argue that the reduction in De/s with temperature in our samples is primarily determined by K eff, which itself is scaled by M s (not m0HK). Indeed, Ms of the [Pt/Co /NiCu] 1 sample changes dramatically for small modifications in temperature within the relevant temperature range [Fig. 5(c)]. Thus, while the shift in relative energy of the uniform magnetic state versus a multidomain state is driven by K eff, a low A ex and modest D permit this transition without Keff approaching zero (i.e., away from the SRT). To understand how these energetic balances impact the equilibrium domain size in the [Pt/Co/NiCu] 1 sample, we have employed the analytic model of Ref. 61. In Fig. 6(c), we show the equilibrium domain width predicted as a function of t film for the [Pt/Co/NiCu] 1 sample (using the same static parameters as those used in the micromagnetic modeling) for several values of D . For the lower D values (0 and 0.26 mJ/m2), a sample with t film = 1.5 nm is predicted to exhibit mm -scale domains, which may be considered as a uniform magnetic state. This calculation agrees with the micromagnetic modeling, which suggested that it was energetically disadvantageous for a sample with t film = 1.5 nm to break into domains at lower D . For D = 0.28 mJ/m2 (i.e., close to the D value where De/s becomes negative), the equilibrium domain size decreases exponentially with the film thickness in the vicinity of the t film values relevant to the [Pt/Co/NiCu] 1 sample . Knowing that t film = 1.5 nm and the experimentally determined domain size is ~ 2 mm, the analytic model indicates that 0.28 mJ/m2 is a reasonable estimation for D in the [Pt/Co/NiCu] 1 film. Taken in concert, the micromagnetic and analytic modeling suggest that a modest D can enable the formation of a labyrinthian stripe phase at remanence in samples with low A ex - behavior that is not typical to thin films with low M s and appreciable PMA. However, the appearance of a labyrinthian stripe phase at zero field does not necessarily guarantee the presence of a field -induced skyrmion phase. To understand the nature of the field-induced morphological transition between stripe domains and skyrmions, we have performed additional micromagnetic simulations, whereby we compare the u 0Hz-dependence of e between the labyrinthian stripe and skyrmion phases at T = 0 K.62 For these calculations, the material parameters were the same as those used when generating Fig. 6(b) and a D = 0.28 mJ/m2. The initial state of the simulation was set to a random magnetization pattern or bubble lattice (for the labyrinthian stripe and skyrmion phases, respectively). At each u 0Hz step in the simulation, the system was relaxed to its minimum total energy configuration. In line with the experimental results, comparing the e(u 0Hz) profiles of these two morphological phases [Fig. 6(d)] demonstrates that the skyrmion phase becomes the ground state when u 0Hz 0.09 mT, which is reasonably close to the experimental findings. While an energetic barrier must be overcome to transition between morphological states, finite - temperature atomistic simulations have shown that bringing the system closer to T C can shallow this barrier.62 Given that all the samples discussed herein have a TC less than 400 K, we argue that the energy barrier separating morphological phases may be easily overcome by the thermal energy present near room 8 We have yet to comment in detail on the magnetic state present whe n the hysteresis loops indicate PMA [Figs. 1(b), 2(a,b)] but polar MOKE contrast has disappear ed and there is little or no hysteresis [Figs. 1(b,f)]. In the simplest scenario , the domain size at such temperatures or FM thicknesses may be too small to resolve with polar MOKE. In the context of the SRT, however, such observations have been interpreted as an indicator of a phase transition, and there has been extensive debate in the literature as to whether this consists of a transition from the ferromagnetic state to a paramagnetic gap state with PMA18,64 or a fluctuating stripe domain phase65,66,67. While our samples do not have the low K eff emblematic of the SRT, the static magnetic properties are similar in many regards ; thus, further investigations with higher spatiotemporal resolution are needed to fully understand the nature of the -induced domain morphology transformations observed in our samples. Regardless of the end result of the transformation, the modeling discussed above permits an understanding of the energetic balances responsible for skyrmion stabilization in our samples with low exchange stiffness, appreciable PMA, and modest iDMI. Conclusion: In summary, we have experimentally characterized the domain morphologies present in a number of thin film heterostructures with perpendicular magnetic anisotropy, modest interfacial Dzyaloshinskii -Moriya interaction energy density, and low exchange stiffness. We find that by lowering the exchange stiffness , the remanent domain morphology transition s from a uniform state to a labyrinthian stripe phase - drawing strong, yet nuanced parallels to previous work on the spi n reorientation transition that occurs in the limit of vanishing effective perpendicular magnetic anisotropy. Furthermore, when a small perpendicular magnetic field is applied at this morphological transition , skyrmion phases become stabilized . Spin- orbit torque -induced motion including the skyrmion Hall effect is observed when an electrical current is passed through the samples. Micromagnetic and analytic modeling demonstrates that in the limit of low exchange stiffness, the presence of a moderate interfac ial Dzyaloshinskii -Moriya interaction modifies the energetic balance of thin films, allowing for the formation of multidomain remanent states and a shallowing of the energy barrier separating the stripe and skyrmion ground states. The authors gratefully acknowledge support from the National Science Foundation, Division of Materials Research (Award #: 2105400). This work was performed in part at the San Diego Nanotechnology Infrastructure (SDNI) of the University of California - San Diego, a member of the National Nanotechnology Coordinateed Infrastructure, which is supported by the National Science Foundation (Grant #: ECCS -1542148). 9 Main Text Figures: Figure 1: (a) Saturation magnetization as a function of temperature for [Pt/Co] 2 samples with varying Co thicknesses ( tCo = 0.32 nm, 0.24 nm, and 0.2 nm), determined using VSM. (b) Polar MOKE loops for [Pt/Co] 2 samples with 0.23 nm <= tCo <= 0.26 nm, collected at room temperature. The MOKE loops have been offset for clarity. Room temperature remanent states of the [Pt/Co] 2 samples with ( c) tCo = 0.26 nm, (d) tCo = 0.25 nm, (e) tCo = 0.24 nm, and (f) tCo = 0.23 nm. The remanent states were attained using the field cycling process discussed in the main text. 10 Figure 2: For the [Pt/Co(0.24 nm)] 2 sample: (a) Out-of-plane and (b) in-plane magnetometry data collected in the temperature range of 294 K <= T <= 302 K. (c-f) Room temperature polar MOKE images collected as the perpendicular magnetic field strength u 0Hz was quasi -statically stepped from u 0Hz = 0 towards negative saturation . 11 Figure 3: (a) and (b) +Mz skyrmions stabilized in a perpendicular magnetic field of - 0.7 mT in a [Pt/Co(0.24 nm)] 2 device patterned into a 100 um -wide wire with 3 mm- wide contact pads (a) before and (b) while an electrical current density of J = 5.5 x 109 A/m2 was applied. (c) and (d) were collected under the same protocol but with u0Hz = + 0.7 mT to obtain -Mz skyrmions. The images shown in (b) and (d) represent the average over images collected at 16 Hz while the current was applied for 1 s. The behavior demonstrated in Fig. 3 is provided as a video in Suppl. Video 4.36 12 Figure 4: (a) Remanent domain morphology of the [Pt/Co/Ni/Re] 2 sample, where the Co layers were grown such that tCo decreases by ~ 0.05 % when moving from left to right across the field of view. The center of the field of view corresponds to an approximate t Co of 0.23 nm. Remanent state and densest u 0Hz-induced skyrmion phases observed in the (b,c) [Pt/Co/Ni/Re] 2 and ( d,e) [Pt/Co/Ni] 2 samples. 13 Figure 5: For the [Pt/Co/NiCu] 1 sample: (a) out-of-plane and (b) in-plane magnetometry data collected in the temperature range of 285 K <= T <= 295 K. (c) Saturation magnetization M s over the temperature range 250 K <= T <= 360 K. Polar MOKE images collected at T = 290 K in perpendicular magnetic fields of 0 and -0.04 mT are shown in (d) and (e), respectively . 14 Figure 6: Modeling results for the [Pt/Co/NiCu] 1 sample: (a) Schematic depiction of the uniform (left) and two - domain (right) states used in the micromagnetic simulations. Blue (red) regions correspond to + Mz (-Mz) magnetic orientations. (b) Micromagnetic calculations of the difference in energy density D e between the uniform and two- domain states (red) and the domain wall energy density s (blue) as a function of iDMI energy density D . For each initial state, D e/s has been normalized to the v alue when D = 0. Data is shown for A ex = 2.5 pJ/m (squares) and Aex = 10 pJ/m (stars). (c) Analytic modeling of the expected domain size as a function of the magnetic film thickness tfilm for selected D values. The dashed line indicates the magnetic layer thickness of the [Pt/Co/NiCu] 1 sample used in our experiments. (d) e as a function of u 0Hz for the labyrinthian stripe and skyrmion phases. The material parameters used in the modeling are stated in the Discussion portion of the main text. 15 Supplemental Information: Supplemental Note 1: To estimate the iDMI energy density D of the [Pt/Co] 2-type samples, we have measured the in -plane of the velocity v of | to | and | to | domain walls ( i.e., the right and left sides (respectively) of the domains shown in Supp. Figs. 1(a,b)) as a pulsed perpendicular magnetic field was applied. As the iDMI promotes chiral Neel -type domain walls (schematically shown in Supp. Fig. 1(b)), an in- plane magnetic field is expected to break the energy symmetry of | to | and | to | domain walls, leading to an asymmetry in the velocity of expansion. As shown in Supp. Fig. 1(c), | to | domains exhibit a minimum in velocity when an in- plane magnetic field m 0Hx is applied in the + xx-direction indica ted in Supp. Fig. 1(b). By determining the m 0Hx at which this minimum in velocity occurs ( u0HDMI), D can be estimated using the expression:24 (S1) The parameters used in Eq. S1 have been previously assigned in the main text. From the v (m0Hx) curves shown in Supp. Fig. 1(c), a D of -0.14 mJ/m2 can be extracted for the [Pt/Co(0.35 nm)] 2 sample. The left - handedness of the iDMI in this sample (denoted by the negative sign of D ) is determined from by comparing the response s of | to | and | to | domain walls to the direction in which u 0Hx is applied . While there is substantial debate as to the appropriateness of using m 0Hx-induced domain growth asymmetry measurements to determine D, previous reports have demonstrated qualitative agreement between D values rendered by the aforementioned technique and other approaches, such as Brillouin light scattering (BLS) 16 Supplemental Figures: Supplemental Figure 1: Polar MOKE images of domain growth of a + Mz domain in a [Pt/Co(0.35 nm)] 2 sample after several out -of-plane field pulses were applied in in -plane fields m 0Hx of (a) 0 mT, and (b) +50 mT. The magenta outlines indicate the initial extent of the nucleated domain. The yellow arrows in (b) schematically depict the left -handed Neel -type domain walls promoted by the Pt -Co interface. (c) The expansion velocity of | to | and | to | domains as a function of m 0Hx. When collecting data to generate (c), the perpendicular field pulse strength was ~ 0.8 mT, and the pulse duration was 4 ms. 17 Supplemental Figure 2: Polar MOKE images of the [Pt/Co(0.26 nm)] 2 sample, collected as the perpendicular magnetic field strength u0Hz was quasi -statically stepped from u 0Hz = 0 towards negative saturation. When imaging, the sample temperature was approximately 315 K. 18 Supplemental Figure 2: D e and s as a function of A ex and K u for different values of D , calculated in the manner discussed in the main text. For all calculations, we took M s as 250 kA/m. The blue (red) circles indicate parameter combinations where the micromagnetic simulations resulted in a positive (negative) D e. The green line demarcates positive and negative s values, as indicated in (b). Note that in (a), the absence of iDMI dictates that the sign of s does not change for any combination of A ex and K u, hence the lack of a s curve. 19 Supplemental Video Captions : Supplemental Video 1: Polar MOKE video of the remanent domain morphology of a [Pt/Co(0.24 nm)/Pt] 1 sample observed for 20 s. Supplemental Video 2: Polar MOKE video showing the evolution of the domain morphology in a [Pt/Co(0.24 nm)/Pt] 2 sample as a perpendicular magnetic field was quasi -statically stepped from 0 to - 0.8 mT. The video was collected at a framerate of 16 Hz, averaging every 4 frames together. Note: The Co layers in this sample were grown as a wedge using the methods outlined in the main text. Supplemental Video 3: Polar MOKE video of the motion of skyrmions in a [Pt/Co(0.24 nm)/Pt] 2 sample as a current density of 3.7 x 109 A/m2 was applied from right to left in an applied magnetic field of - 0.5 mT . The video was collected at a framerate of 16 Hz, averaging every 4 frames together. Supplemental Video 4: Polar MOKE video of the motion of skyrmions in a [Pt/ Co (0.24 nm)] 2 sample as a current density of 7.5 x 109 A/m2 was applied from right to left in an applied magnetic field of -0.6 mT . The video was collected at a framerate of 16 Hz, averaging every 4 frames together. Supplemental Video 5: Polar MOKE video demonstrating the skyrmion Hall effect in a [Pt/ Co (0.24 nm)] 2 sample as a current density of 6.5 x 109 A/m2 was applied for 1 s from left to right in an applied magnetic field of - 0.7 mT. The video was collected at a framerate of 16 Hz, averaging every 16 frames together. 20 References 1 S. Heinze, K. Bergmann, M. Menzel, J. Brede, A. Kubetzka, R. Wiesendanger, G. Bihlmayer, and S. Blugel, Nature Phys. 7, 713 (2011). 2 A. Fert, N. Reyren, V. Cros, Nature Rev. Mat. 2, 17031 (2017). 3 S. Rohart, J. Miltat, and A. Thiaville, Phys. Rev. B 93, 214412 (2016). 4 J. Grollier, D. Querlioz, K.Y. Camsari, K. Everschor -Sitte, S. Fukami, and M.D. Stiles, Nature Electronics, (2020). 5 S.A. Montoya, S. Couture, J.J. Chess, J.C.T. Lee, N. Kent, D. Henze, S.K. Sinha, M.- Y. Im, S.D. Kevan, P. Fisch er, B.J. McMorran, V. Lomakin, S. Roy, and E.E. Fullerton, Phys. Rev. B 95, 024415 (2017). 6 I. Dzyaloshinskii, J. Phys. Chem. Solids 4, 241 (1958). 7 T. Moriya, Phys. Rev. 120, 91 (1961). 8 S. Muhlbauer, B. Binz, F. Jonietz , C. Pfleiderer, A. Rosch, A. Neubauer, R. Goergii, and P. Boni, Science 323, 915 (2009). 9 A. Fert, Mat. Sci. For. 1990, 59- 60, 439. 10 J. Sampaio, V. Cros, S. Rohart, A. Thiaville, and A. Fert, Nat. Nanotech . 8, 839 (2013). 11 A. Soumyanarayanan, N. Reyren, A. Fert, and C. Panagopoulos, Nature 539, 509 (2016). 12 C. Moreau -Luchaire, C. Moutafis, N. Reyren, J. Sampaio, C.A.F. Vaz, N. Van Horne, K. Bouzehouane, K. Garcia, C. Deranlot, P. Warnicke, P. Wohluter, J.- M. George, M. Weigand , J. Raabe, V. Cros, and A. Fert, Nature Nanotech. 11, 444 (2016). 13 S. Woo, K. Litzius, B. Kruger, M.- Y. Im, L. Caretta, K. Richter, M. Mann, A. Krone, R.M. Reeve, M. Weigand, P. Agrawal, I. Lemesh, M.- A. Mawass, P. Fischer, M. Klaui, and G.S.D. Beach, Na ture Mater. 15, 501 (2016). 14 K. Everschor -Sitte, J. Masell, R.M. Reeve, and M. Klaui, J. Appl. Phys. 124, 240901 (2018). 15 R. Allenspach and A. Bischof, Phys. Rev. Lett. 69, 3385 (1992). 16 Z.Q. Qiu, J. Pearson, and S.D. Bader, Phys. Rev. Lett. 70, 1006 (1993). 17 K. Fukumoto, H. Daimon, L. Chelaru, F. Offi, W. Kuch, J. Kirschner, Surf. Sci. 514, 151 (2002). 18 C. Won, Y.Z. Wu, J. Choi, W. Kim, A. Scholl, A. Doran, T. Owens, J. Wu, X. F. Jin, H.W. Zhao, and Z.Q. Qiu, Phys. Rev. B 71, 224429 (2005). 19 R. Tolley, S.A. Montoya, and E.E. Fullerton, Phys. Rev. Mat. 2, 044404 (2018). 20 W. Jiang, P. Upadhyaya, W. Zhang, G. Yu, M.B. Jungfleisch, F.Y. Fradin, J.E. Pearson, Y. Tserkovnyak, K.L. Wang, O. Heinonen, S.G.E. te Velthuis, and A. Hoffmann, Science 349, 283 (2015). 21 K. Gaurav Rana, A. Finco, F. Fabre, S. Chouaieb, A. Haykal, L. D. Buda -Prejbeanu, O. Fruchart, S. Le , Denmat , P. David, M. Belmeguenai, T. Denneulin, R. E. Dunin- Borkowski, G. Gaudin, V. Jacques, and O. Boulle, Phys. Rev. Appl. 13, 044079 (2020). 22 S. Sugimoto, W. Koshibae, S. Kasai, N. Ogawa, Y. Takahashi, N. Nagaosa, and Y. Tokura, Sci. Rep. 10, 1009 (2020). 23 M. He, G. Li, Z. Zhu, Y. Zhang, L. Peng, R. Li, J. Li, H. Wei, T. Zhao, X. -G. Zhang, S. Wang, S.- Z. Lin, L. Gu, G. Yu, J.W. Cai, and B.- G. Shen, Phys. Rev. B 97, 174419 (2018). 24 S.-G. Je, D. -H. Kim, S. -C. Yoo, B.- C. Min, K.- J. Lee, and S.- B. Choe, Phys Rev B 88, 214401 (2013). 25 S. Rohart and A. Thiaville, Phys. Rev. B 88, 184422 (2013). 26 F. Buttner, I. Lemesh, and G.S.D. Beach, Sci. Rep. 8, 4464 (2018). 27 C. Liu and S.D. Bader, J. Appl. Phys. 67, 5758 (1990). 28 D. Li, R. Cuenya , J. Pearson, S.D. Bader, and W. Keune, Phys. Rev. B 64, 144410 (2001). 29 J.B. Sousa, M.R. Chaves, M.F. Pinheiro, and R.S. Pinto, J. Low Temp. Phys. 18, 125 (1974). 21 30 M.W. Keller, K.S. Gerace, M. Arora, E.K. Delczeg -Czirjak, J.M. Shaw, and T.J. Silva, Phys . Rev. B 99, 214411 (2019). 31 A.G. Kozlov, A.G. Kolesnikov, M.E. Stebliy, A.P. Golikov, and A.V. Davydenko, Phys. Rev. B 102, 144411 (2020). 32 A. Vansteenkiste, J. Leliaert, M. Dvornik, M. Helsen, F. Garcia -Sanchez, and B. Van Waeyenberge, AIP Adv. 4, 107133 (2014). 33 S. Bandiera, R.C. Sousa, B. Rodmacq, and B. Dieny, IEEE Mag. Lett. 2, 3000504 (2011). 34 A.W.J. Wells, P.M. Shepley, C.H. Marrows, and T.A. Moore, Phys. Rev. B 95, 054428 (2017). 35 J.A. Brock, M.D. Kitcher, P. Vallobra, R. Medapalli, M.P. Li, M. De Graef, G.A. Riley, H.T. Nembach, S. Mangin, V. Sokalski, and E.E. Fullerton, Adv. Mat., in press (2021). 36 See Supplemental Information for information on how D was estimated for the [Pt/Co(0.35 nm)] 2 sample (Supp. Note 1), measurements of the in -plane field -induced growth asymmetries and iDMI energy density of a [Pt/Co(0.35 nm)] 2 sample (Supp. Fig. 1), images of the field -dependent domain morphology of a [Pt/Co (0.26 nm)] 2 sample at T = 315 K (Supp. Fig. 2), and micromagnetic modeling of the evolution of D e and s with A ex, K, and D over an extended parameter space (Supp. Fig. 3) , videos of thermal fluctuations of the remanent domain morphology in the [Pt/Co(0.24 nm)] 1 sample (Supp. Video 1), the evolution of the [ Pt/ Co(0.24 nm)] 2 sample's domain morphology with perpendicular field strength (Supp. Video 2), motion of skyrmions in response to applied current densities of 3.7 x 109 A/m2 (Supp. Video 3) and 7.5 x 109 A/m2 (Supp. Video 4), and the skyr mion Hall effect observed in response to a current density of 6.5 x 109 A/m2 in a [Pt/Co(0.24 nm)] 2 sample (Supp. Video 5). 37 L. Liu, C.- F. Pai, Y. Li, H.W. Tseng, D.C. Ralph, and R.A. Buhrman, Science 336, 555 (2012). 38 M. Althammer, S. Meyer, H. Nakayama , M. Schreier, S. Altmannshofer, M. Weiler, H. Huebl, S. Geprags, M. Opel, R. Gross, D. Meier, C. Klewe, T. Kuschel, J.- M. Schmalhorst, G. Reiss, L. Shen, A. Gupta, Y.- T. Chen, G.E.W. Bauer, E. Saitoh, and S.T.B. Goennenwein, Phys. Rev. B 87, 224401 (2013). 39 N. Nagaosa and Y. Tokura, Nat. Nanotech. 8, 899 -911 (2013). 40 K. Everschor -Sitte and M. Sitte, J. Appl. Phys. 115 , 172602 (2014). 41 W. Jiang, X. Zhang, G. Yu, W. Zhang, X. Wang, M.B. Jungfleisch, J.E. Pearson, X. Cheng, O. Heinonen, K.L. Wang, Y . Zhou, A. Hoffmann, and S.G.E. te Velthuis, Nat. Phys. 13, 162 (2017). 42 F. Ajejas, V. Krizakova, D. de Souza Chaves, J. Vogel, P. Perna, R. Guerrero, A. Gudin, J. Camareno, and S. Pizzini, Appl. Phys. Lett. 111, 202402 (2017). 43 S. Pizzini, J. Vogel, S . Rohart, L. Buda -Prejbeanu, E. Jue, O. Boulle, I. Miron, C. Safeer, S. Auffret, G. Gaudin, and T. Thiaville, Phys. Rev. Lett. 113, 047203 (2014). 44 M. Belmeguenai, J.- P. Adam, Y. Roussigne, S. Eimer, T. Devolder , J.-V. Kim, S. Cherif, A. Stashkevich, and A. Thiaville, Phys. Rev. B 91, 180405(R) (2015). 45 J. Cho, N.- H. Kim, S. Lee, J.- S. Kim, R. Lavrijsen, A. Solignac, Y. Yin, D. -S. Han, N.J.J. van Hoof, H.J.M. Swagten, B. Koopmans, and C.- Y. You, Nat. Comm. 6, 7635 (2015). 46 N.-H. Kim, J. Jung, J. Cho, D.- S. Han, Y. Yin, J.- S. im, H.J.M. Swagten, and C.- Y. You, Appl. Phys. Lett. 108, 142406 (2016). 47 J.A. Brock, P. Vallobra, R.D. Tolley, S.A. Montoya, S. Mangi n, and E.E. Fullerton, Phys. Rev. B 102, 024443 (2020). 48 P. Jadaun, L.F. Register, and S. K. Banerjee, npj Computational Materials 6, 88 (2020). 49 A. Soumyanarayanan, M. Raju, A.L. Gonzalez Oyarce , A.K.C. Tan, M. -Y. Im, A.P. Petrovic, P. Ho, K.H. Khoo, M. Tran, C. K. Gan, F. Ernult, and C. Panagopoulos, Nature Mater. 16, 898 (2017). 50 H.Y. Kwon and C. Won, J. Magn. Magn. Mater. 351, 8 (2014). 51 T. Eimuller, T.C. Ulbrich, E. Amaladass, I.L. Guhr, T . Tyliszczak, and M. Albrecht, Phys. Rev. B 77, 134415 (2008).

2108.10973

Jeffrey Brock

Jeffrey A. Brock and Eric E. Fullerton

Skyrmion stabilization at the domain morphology transition in metal heterostructures with low exchange stiffness

22 pages, 6 figures, 3 supplementary figures

cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

We report the experimental observation of micron-scale magnetic skyrmions at room temperature in several Pt/Co-based thin film heterostructures designed to possess a low exchange stiffness, perpendicular magnetic anisotropy, and a modest interfacial interaction (iDMI). We find and by micromagnetic and analytic modeling that the combined action of low exchange stiffness and modest iDMI eliminates the associated with forming domain walls in thin film the domain wall energy density approaches negative values, the remanent domain morphology transitions from a uniform state to a labyrinthian stripe phase. A low exchange stiffness, indicated by a reduction in the below 400 K, is achieved in Pt/Co, Pt/Co/Ni, and by reducing the Co thickness to the ultrathin limit (< 0.3 nm). A similar effect occurs in thicker Pt/Co/NixCu1-x structures when the Ni layer is alloyed with Cu. At this transition in domain morphology, skyrmion phases are stabilized when a small (< 1 mT) perpendicular magnetic field is applied skyrmion motion including the skyrmion Hall effect is observed. The temperature and morphological phase transitions observed are similar to the well-studied spin reorientation transition that occurs in the ultrathin limit, but we find that the underlying energy balances modified by the presence of an iDMI.

2 1. Introduction Topological materials are a promising platform for a wide range of next -generation In topological Dirac or Weyl semimetals, symmetry -protected Dirac or Weyl nodes host relativistic fermions with low -energy excitations tha t can be described by Dirac or Weyl equations, In addition , topological nodal -line semimetals exhibiting linear band crossings along a 1D loop have also been discovered.[7] Recently, there is a rapidly growing interest in the ZrSiS topological nodal -line semimetal family, represented by the chemical formula WHM (W=Zr/Hf/rare earth, H= Si/Ge/Sn/Sb, These materials crystalize in a PbFCl -type structure compris ed of 2D square nets of H atoms sandwiched by W -M layers. Two types of Dirac states have been discovered in this material family thus far: (i) a gapless Dirac point state protected by nonsymmorphic symmetry, and (ii) a Dirac nodal -line generated by the glide mirror symmetry, which is gapped by spin -orbit coupling Exotic properties such as electronic correlation enhancement[33] owing to the nodal -line structure[34] and a new surface state due to broken symmetry at surfaces[35] which is robust[36] against surface oxidation[36,37 ] have been reported. In addition to those non -magnetic materials, the spin degree of freedom can be activated in WHM compounds by making W a magnetic lanthanide (Ln) element, H = Sb, and M = Te. Such LnSbTe materials provide a platform to further invest igate exotic quantum states.[38-43] New phenomena such as Kondo effects, charge density waves (CDW), and correlation enhancement have been discovered in various LnSbTe Those quantum phenomena, together with the non -trivial topology generated by the Sb layer in offer great opportunities for exploring top ological physics and for developing topological quantum material -based device s. The study of magnetic LnSbTe compounds is in an early stage with only few mater ials discovered, such as CeSbTe, GdSbTe, NdSbTe, and HoSbTe.[25,26,38 -43,45 ,46] Here , we report the investigation of the previously unexplored SmSbTe member of the family. Our combined magnetization and specific heat measurements revealed and signatures of electronic correlation enhancement. These findings, together with the Dirac states revealed by angle resolved photoemission spectroscopy (ARPES) and electronic band structure calculations, suggest that SmSbTe represents an ideal platform for new exotic quantum phenomena arising from the interplay between magnetism, topology, and electronic correlations. The manipulation of these phenomena would further pave a path for quantum material -based functional devices. 3 2. Result s and Discussion SmSbTe single crystals were grown by chemical vapor transport (CVT) . The sharp (00L) x -ray diffraction peaks ( Figure 1 a) reveal the excellent crystallinity of our single crystals. Single crystal x -ray diffraction further reveals that SmSb Te crystalizes in a tetragonal lattice with nonsymmorphic space group P4/nmm, formed by the stacking of a Te -Sm-Sb-Sm-Te slab in which the Sb square nets are sandwiched by Sm -Te layers ( Figure 1 b). The structural parameters obtained from the refinement are provided in the Supporting Information. Earlier studies on other LnSbTe compounds[38,4 7,48] revealed possible Te substitution and vacancies in the Sb layer, which reportedly cause CDWs. In our Sm SbTe, the composition analysis using energy dispersive x -ray spectroscopy indicates a stoichiometric composition, which is further confirmed by the lack of a CDW order as will be shown later on. The magnetic order induced by rare earth elements in LnSbTe is predicted to tune the topological electronic states via the coupling between magnetism and band topology,[26] hence enabling functional properties in these materials. The magnetic property characterization of SmSbTe is presented in Figure 2 . Similar t o other LnSbTe compounds (Ln = Ce, Nd, Gd, and Ho),[26,40 -43] SmSbTe displays an (AFM) order with a Neel temperature TN ~ 3.7+-0.2 K. The slight variation of 0.2 K is possibly due to tiny variations in compositions of the multiple crystals we measured. The probed TN is independent of magnetic field strength (Figure 2 a, inset) though, a situation in stark contrast with CeSbTe and Nd SbTe, where TN is suppressed upon increasing the magnetic field.[26,39,43] The AFM nature of the magnetic transition can be determined by the absence of irreversibility between zero -field-cooling (ZFC) and field -cooling (FC) measurements in the temperatu re dependence of the susceptibility ( ) (Figure 2 a), and by the linear field dependence at low fields and below TN in isothermal magnetization M(H ) (Figure 2 b). In addition, ( ) is found to follow a modified Curie -Weiss law mol = +C/(T - ) above 20 K where is the temperature -independent part of the susceptibility, C is the Curie temperature, and is the Weiss temperature. The negative ~ - 31.7 K extracted from the fit ( Figure 2 a) implies a strong AFM exchange interaction among Sm momen ts, a behavior analogous to that observed in CeSbTe ( = -10 to -30 K)[26,39] and GdSbTe ( = -19 K to -24K).[25,40,41] From the obtained , the frustration parameter, defined as f =| |/TN, is 8, suggesting possible magnetic which is consistent with our specific heat measurement to be discussed later. The fitted Curie constant , C, yields an effective moment eff=3kBC/NA= 1.05 B, which is slightly larger than the theoretically expected value of 4 eff = 0.857 B for a Sm+3 ion with 4 f 5 configuration. The slightly larger eff may originate from the polarization of conduction electrons, or it may also be due to the reduction of moment density generally seen in frustrated magnetic systems.[49] As mentioned above, the isothermal magnetization M of SmSbTe is linear in H at low field. With increasing magnetic field, a small deviation from linearity can be seen with out -of-plane fields (H//c), as shown in the inset of Figure 2 b. This implies a possi ble metamagnetic transition which has also been observed in other magnetic LnSbTe materials such as CeSbTe, NdSbTe, HoSbTe, and On the other hand, u nder an in -plane magnetic field (H//ab) configuration, a linear M(H) is obse rved for up to 9 T (the highest field of our instrument). Generally, in AFM systems, a large exchange interaction is expected for linear M(H) without polarization up to high field , which appears inconsistent with the low TN measured in SmSbTe. On the other hand, linear M(H) and low TN have indeed been observed in other AFM rare earth compounds such as Sm 2BaPdO 5 and GdFe 0.17Sn2.[49,52 ] Further theoretical and experimental stud ies are needed to fully clarify the magnetism in SmSbTe. Despite of the unambiguous signature of magnetic order in magnetization measurements, the temperature dependence of the resistivity, shown in Figure 2 c, does not display a clear feature at TN (black arrow in Figure 2 c). A Hall effect study reveals an elect ron-type conduction with a carrier density in the order of 1021 cm-3 for SmSbTe ( Figure 2 d), which is comparable to reported values for CeSbTe [39] and NdSbTe.[43] Similarly, no clear signature is observed in the Hall effect across TN (~3.7 K) . These o bservations imply th at electronic transport may not be strongly coupled with magnetic order. Nevertheless, the non -metallic transport with a logarithmic temperature dependence between 10 and 35 K ( Figure 2 c) is suggestive of Kondo physics, which has also been proposed in CeSbTe[39] and NdSbTe.[43] Such a possible Kondo effect, to be discussed later, offers opportunit ies for the realization of new quantum states and for corresponding advanced functional properties. Specific heat measurements provide further insight into the magnetic properties of SmSbTe. The specific heat divided by temperature C(T)/T data displays a broad peak around 3.2 K (Figure 3 a), lower than the magnetic transition temperature probed by magneti c susceptibility in figure 2. This is in sharp contrast with the well -defined specific heat peak at the magnetic phase transition temperature observed in other LnSbTe materials such as NdSbTe,[43] CeSbTe,[26,39] and HoSbTe.[42] Such broad specific heat peak is unlikely to originate from 5 sample quality issues such as inhomogeneity or defects, because it is reproducible in multiple SmSbTe crystals from the same or different growth batches with well characterized compositions, and it does not change after thermal annealing. The profile and position of the broad specific heat peak are essentially field -independent, as shown in the inset of Figure 3 a. Such field -insensitivity was also observed in the magnetization measurements shown i n inset of Figure 2 a. The absence of a clear ordering feature in the specific heat is a common signature of frustrated magnetic materials, where the moments cannot order onto a unique, lowest energy state.[53-55] Such scenario of magnetic frustration agr ees with the sizable frustration parameter derived from magnetization measurements mentioned above. Frustrated magnetism is known to provide opportunities to stem abundant exotic phenomena such as skyrmion s[56] and quantum spin liquid s[57], further enrich ing the SmSbTe material platform with additional functional properties. Geometric frustration is known to exist in trigonal or hexagonal lattices, while the tetragonal structure of SmSbTe and the lack of strong field dependence in specific heat measurements ( Figure 3 , inset) appear inconsistent wit h known geometrically frustrated systems.[58,59] Therefore, frustrated magnetism, if it exist in tetragonal SmSbTe, could most likely be caused by the competition between the nearest -neighbor and next -nearest -neighbor magnetic exchange couplings, which has been proposed[60] and observed in tetragonal FeSe and CoSe.[50,5 8] To clarify the origin of the broad specific heat peak in SmSbTe , which is distinct from all other known LnSbTe materials, more e xperimental and theoretical efforts are needed. The most intriguing feature in our specific heat measurements is the significantly enhanced Sommerfeld coefficient, which can be obtained by separating the electronic specific heat from the total measured v alue. The total specific heat , Ctot, can be expressed by Ctot = Cm + Cele + Cph, where Cele = T is the electronic specific heat , is the Sommerfeld coefficient, Cm and Cph are the magnetic and phonon contributions to the specific heat, respectively. Generally, its phonon contribution can be expressed by Cph = T3 at the low temperature limit where T << Debye temperature Th D, and thus may be separated by a linear extrapolatio n of C/T vs T2 or by a polynomial fitting. However, these approaches may not work for SmSbTe due to a pronounced specific heat upturn at around 10 K in C/T (Figure 3 a). Therefore, to obtain the precise electronic specific heat, we used the isomorphous non -magnetic LaSbTe as a reference sample, and followed the principle of corresponding states to evaluate the phonon contribution Cph, which has been demonstrated to be a justified approach in our previous specific heat studies on NdSbTe[43] and Fe(Te,Se).[61] As shown in Figure 3 a, our analysis yields a Sommerfeld coefficient of 160 mJ/mol, as well as a phonon coefficient of = 0.465 mJ/mol K4 which 6 yields a Debye temperature Th D of 232.14 K. With the obtained electronic specific heat C ele = T and phonon specific heat Cph, the magnetic specific heat C m of SmSbTe can be extracted, from which the magnetic entropy Sm=Cm(T) TdTT 0 is also obtained, as shown in Figure 3 b. With increasing temperature, the entropy quickly rises , saturating to 4.33J/mol K. Assuming a J = 5/2 doublet for Sm3+ in the crystal field of SmSbTe, such entropy is 76% of the expected value of Rln2 where R is gas constant , which implies a possible residual magnetic entropy that is widely observed in frustrated systems.[62,63] The Sommerfeld coefficient , g, in LnSbTe is strongly dependent on the choice of the Ln element. g is as small as 0.51 mJ/mol K2 in non -magnetic LaSbTe,[43] and it increases to 7.6 mJ/mol K2 in GdSbTe,[40] and to 10-40 mJ/mol K2 in CeSbTe.[26,39] In addition to = 160 mJ/mol for SmSbTe as determined in this work, above 100 mJ/mol K2 has also been observed in NdSbTe (115 mJ/mol K2) [43] and HoSbTe (382.2 mJ/mol K2).[42] The drastic enhancement of g in SmSbTe implies a large effective mass which, as will be shown later, might be associated with the presence of flat Sm 4 f bands around the Fermi level (EF). This finding further highlights the importance of electron interaction s in SmSbTe. Its coupling with magnetism and topological bands (shown below) further offers a new route of manipulating quantum phase s in this material. In addition to magnetism and electronic correlation enhancement, the existence of topological relativis tic fermions revealed by our ARPES experiments further enables SmSbTe as a versatile platform with potential to realize advanced functionalities arising from the interplay between various quantum degrees of freedom. Figure 4 a shows the Fermi surface and co nstant energy contour plots in the paramagnetic (PM) state ( T = 78 K) using a photon energy of 30 eV on the (001) plane of the crystal (see Experimental Section for details). A diamond -shaped Fermi surface centered at G is observed, reminiscent of that see n in topological semimetal The constant energy contour plots reveal that no other electronic bands approach the EF down to the 250 meV binding energy, where an extra four -leaf clover -shaped pattern appears around G. These constant energy contours are well reproduced by our density functional theory (DFT) calculations ( Figure 4 b), in which the EF position was adjusted (shifted up) by 0.275 eV in order to reproduce the experimental data, as we will further discuss below. The electronic band structure measured along the M-G-M and X-G-X high symmetry directions ( Figure 4 c, with its related bulk Bril louin zone shown in Figure 4 f) confirms that the 7 diamond -shaped Fermi surface originates from the linearly dispersing Dirac -like bands. The linearly dispersing bands crossing EF are clearly observed along both momentum directions. In particular, Dirac poin ts at 0.275 eV binding energy are observed along G-M, as indicated by the black arrows. The observed electronic band structure agrees well with DFT calculations on a slab geometry (see Experimental Section) as shown in Figure 4 e. The band structure calcula tion of the bulk bands ( Figure 5 a) further revealed that the observed Dirac bands belong to the Dirac nodal -line generated by C 2v symmetry, which are formed mainly by the Sb p - electrons (see Supporting Information). A comparison between the ARPES results and the calculated bands indicates that EF of SmSbTe lies 0.275 eV above the Dirac crossings at the G-M cut. Similar to other compounds in the ZrSiS -family, the nodal -line crossings in Sm SbTe are expected to be slightly (~55 meV) gapped by SOC, as revealed by our bulk band calculations shown in Figure 5 a. Nevertheles s, gap opening at the crossing points is not observed in our ARPES data. An u ndetectable SOC gap is also reported in recent A RPES study on LaSbTe, possibly due to the limitation of the instrument resolution.[64] On the other hand, gapped Dirac crossings have been probed in other LnSbTe The sizable gap has been proposed to originate from the formation of a CDW order in non -stoichiometric compounds, which enlarges the gap at the Dirac Given that the CDW order occurs in the presence of Te substitution and vacancies in the Sb layer in LnSbTe 8] the undetectable gap in our ARPES study confirms the nearly stoichiometric composition and absence of structural distortions in our SmSbTe crystals. Our ARPES observations and DFT calculations (with a shifted EF) indicate that SmSbTe is very similar to ZrSiS, with the Fermi surface of the bulk electronic states formed by the Dirac nodal - line bands only.[17,18,2 1,28] However, the energy dispersion of the nodal -line states in SmSbTe does not depend on the ARPES photon energy , as shown in Figure s 4c,d which present the b and dispersion measured with photon energies of 30 eV (i.e., k z = 0) and 73 eV (corresponding to 17% of the sixth Brillouin zone), respectively. More band structures measured with various photon energies are shown in the Supporting Information. Such observ ation implies that the Fermi surface of SmSbTe consist s of a 2D sheet extending along kz, which is di fferent from that of ZrSiS. A similar 2D -like Fermi surface sheet for bulk electronic states has also been observed in CeSbTe.[44] Another major difference between SmSbTe and ZrSiS is the location of EF, which lies almost at the Dirac crossing in ZrSiS, but is located 0.275 eV above the nodal - line Dirac crossings in SmSbTe as discussed above. Such a shift of EF leads to a clean Fermi surface of relativistic fermions without interference from conventional electronic bands in 8 SmSbTe. The shift of EF has been observed in GdSbTe with non -stoichiometric composition s.[38] In our SmSbTe samples , the composition non -stoichiometry is minima l as discussed above. The origin of such EF shift is unknown, but possibly attributed to very slight composition, or related to the contribution of Sm 4 f bands which cannot be fully captured in our DFT calculation of the energy band struc ture in the high temperature PM state. The calculated electronic band structure for SmSbTe also provides insight into the electronic correlation enhancement probed in specific heat measurements : We performed DFT calculations in the AFM ground state to reveal an AFM ground state that is consistent with the magnetic property characterization described above. Figure 5 b shows the calculated band structure for the AFM bulk SmSbTe with SOC , based on an A -type AFM order with out -of- plane Sm moments aligned parallel within each plane but antiparallel between adjacent planes. Such magnetic structure has been adopted in other LnSbTe materials[45] and experimentally confirmed in CeSbTe.[26] Generally, an A -type AFM structure with in -plane FM order is not compatible with magnetic frustration. Nevertheless, there can be other mechanisms leading to frustration in A-type AFM compound s, such as the electron -electron interactions and ligand distortions in KCuF 3.[65] Distinct from the PM state in which the Sm 4 f bands are located far below the energy range shown in Figure 5 a, we have observed the formation of 4 f flat bands near the EF in the AFM state. As compared with the non -SOC bands (see Figure S 5b in the Supporting Information), the strong SOC of f electrons drastically splits the f-band, leading to the presence of the 4 f flat bands at the calculated E F (Figure 5 b). As stated above, ARPES observations indicate a shift of EF in the PM state. Although further ARPES studies are needed to determine the exact location of EF in the A FM state, the flat band extends onto Dirac nodal - line bands, which opens the possibility of hybridization between the f-orbitals and the conduction nodal -line fermions. Such possible hybridization is reminiscent of a mass renormalization via a Kondo mechan ism in heavy fermions, which agrees with the mass enhancement probed in specific heat measurements. As mentioned above, our transport measurements on SmSbTe probed signatures of Kondo localization, though the carrier density of ~1021 cm-3 extracted from Ha ll effect measurement may to be insufficient to fully screen the localized moments in LnSbTe materials.[44] In principle, SmSbTe in the high temperature PM state is a topological semimetal with a SOC - gapped nodal -line state similar to that reported on ZrSiS. Further, by opening a band gap at the 9 nodal -line and tunning EF into the band gap, SmSbTe could become a weak topological insula tor consisting of stacked 2D topological One practical approach toward stacked 2D topological insulators is composition engineering , which can induce a CDW gap and tune EF [38,44]. In the low temperature AFM state, each layer of SmS bTe possess parallel magnetic moment s for Sm which breaks time -reversal symmetry. Therefore, the bulk AFM SmSbTe can be viewed as the stacking of 2D Chern insulators, which has been proposed in HoSbTe.[45] The rich topological phases in SmSbTe , coupled with charge and magnetic order offers a new playground to study topological physics . 3. Conclusion We have demonstrated that the newly discovered SmSbTe is a promising material platform with a combination of exotic properties, including 4 f magnetism with possible magnetic frustration and Kondo effects, electronic correlation enhancement, and a Fermi surface consisting of Dirac bands only. Such properties convey unique opportunities to study new quantum state s and phenomena arising from the interplay among spin, charge, lattice, and topolog ical degrees of freedom. In addition to the apparent coupling between the topological states and magnetism discussed above, the electronic correlation enhancement further provides opportunities to engineer topological electronic states. For example, the electronic correlations have been found to be associated with the Sm 4 f bands in the AFM state. Thus, the tuning of magnetism via substitution with other rare earth elements could be an efficient way to contr ol electronic correlations. Furthermore, Kondo physics may also enabled by tuning carrier density, and high magnetic fiel ds, which are known to tune the competition between Ruderman -Kittel -Kasuya -Yosida and Kondo effects and lead to variou s quantum states in heavy fermion materials. All of , these knobs may possib ly open a route to realiz e correlated topological quantum phases in SmSbTe. With a unique coexistence of degrees of diverse quantum freedom and their interplay as mentioned above, we conclude that SmSbTe is a promising platform not only for searching new quantum states and phenomena, but also for exploring technology applications exploiting its tunable topological quantum states. 4. Experimental Section SmSbTe single crystals were grown by a two -step chemical vapor transport (CVT) method. First, constituent elements with stoichiometric ratio s were used to grow polycrystals at 700 degC. 10 Second ly, the finely grinded polycrystals were used as a source material for CVT growt h under a temperature gradient of 1000 degC to 900 degC, with Iodine being the transport agent. Square -like crystals with metallic luster can be obtained after 15 days, as shown in the insert of Figure 1 a. The elemental analysis was performed by energy dispe rsive x -ray spectroscopy. The crystal structure at room temperature was determined by single crystal x -ray diffraction at 298 +-2 K. Data collection was performed under a nitrogen gas atmosphere at 90 K using a Bruker SMART APEX -II diffractometer equipped wi th a CCD area detector and graphite -monochromated Mo - Ka radiation (l =0.71073 A). Data reduction and integration, together with global unit cell refinements were performed using the APEX2 software. Multiscan absorption corrections were applied. The atomist ic structure was determined by direct methods and refined by full matrix least-squares methods on F2 using the SHELX package with anisotropic displacement parameters for all atoms. In addition to single crystal x-ray diffraction, powder diffraction was also performed to ensure no impurity phase s were present. For powder diffraction, carefully selected single crystals were ground into powder. A comparison between x-ray diffraction patterns obtained from a single crystal, powdered sample s, and simulations are shown i n the Supporting Information. Magnetization, specific heat, and electronic transport measurements were performed by a physical properties measurement system (PPMS). The electronic band structure was measured by ARPES at the APE -LE b eamline at the Elettra Synchrotron. Reproducible spectra were obtained from multiple SmSbTe crystals. The authos probed the electronic structure by using 26 to 86 eV excitation energy; a comparison with the electronic band structure calculations indicates that the bulk is reached at a 30 eV photon energy. DFT calculations were performed using Vienna ab -initio simulation package (VASP) with projector -augmented wave (PAW) pseudopotentials and the Perdew -Burke -Ernzerhof generalized gradient approximation (PBE -GGA) for the exchange -correlation functional. Calculations were performed with and without SOC for both bulk and slab geometries. For the PM case, the Sm pseudopotential was 4 f-in-core, while for the AFM case it was 4 f-in-valence and we employed a Hubbard -correction in Dudarev's approximation for the f-electrons (U = 6 eV).[66-71] A k-mesh sampling of 12 x 12 x 6 was used for the self -consistent calculations in the bulk , along with a cutoff energy of 300 eV. The slab geometry consisted of five unit cells 11 in the c direction with a vacuum spacing greater than 20 A to prevent interaction between periodic copies. Note added: after the acc eptance of this work, the authors became aware of an AR PES study on the same material system .[72] Supporting Information Supporting Information is available from the Wiley Online Library or from the author. J. H. and S. B .-L. acknowledge the support from US Department of Energy, Office of Science, Basic Energy Sciences program under Award No. DE -SC0019467 for the support of crystal growth, property (magnetic, transport, and thermal), and band structure calculations. J.W.V. ac knowledges funding from the US Department of Energy, Office of Basic Sciences (Award DE -SC0016139). Calculations were performed at Cori at NERSC, a DOE facility funded under contract No. DE -AC02 -05CH11231. R.B. acknowledges the support from the Chancellor' s Innovation and Collaboration Fund at the University of Arkansas. B. D. thanks the support by JSPS KAKENHI Grant Number JP21K14656 and by Grants for Basic Science Research Projects from The Sumitomo Foundation. A. A . and B. G . acknowledge the Science and Engineering Research Board (SERB), Department of Science and Technology (DST) of the government of India for financial support and High -performance computing facility at IIT Kanpur for computational support. This work has been partly performed in the frame work of the nanoscience foundry and fine analysis (NFFA -MUR Italy Progetti Internazionali) facility. Conflict of Interest The authors declare no conflicts of interest. 12 Data Availability Statement The data supporting to the findings of this study is avail able from the corresponding author upon reasonable request. 13 Figure 1 Fig. 1: (a) X-ray diffraction patterns for a SmSbTe single crystal, showing the (00L) reflections. Inset: an image of a single crystal. The square mesh measures 1mm2. (b) Crystal structure of SmSbTe determined by x -ray diffraction. The structur al parameters are provided in Supporting Information. 14 Figure 2 Fig. 2. Magnetization and electronic transport properties of SmSbTe. (a) Temperature dependence of molar susceptibility kh and inverse susceptibility 1/ kh of SmSbTe under an in- plane ( H||ab) magnetic field of 1T. The lack of irreversibility between zero fiel d cooling (ZFC) and field cooling (FC) measurements is apparent . The solid , monotonically increasing red line represent s the modified Curie -Weiss (CW) fit obtained from 1/kh (see text). Inset: magnetic susceptibility under magnetic fields ranging from 0.1 T (uppermost curve) to 9 T (lowermost curve) . (b) Field dependence of magnetization at different temperatures, taken under H||ab. Insets: magnetization for H||c and H||ab at T = 2 K. (c) Temperature dependence of the in-plane resistivity of SmSbTe. A logarithmic temperature dependence can be observed 10 and 35 K. The dashed line is a guide to t he eye. (d) Temperature dependence of Hall coefficient , RH, and the corresponding carrier density (1/ eRH). Inset: field depen dence of the Hall resistivity at 2 K and 300 K. 15 Figure 3 Fig.3 Specific heat of SmSbTe. (a) Temperature dependence of specific heat. The blue solid line represents the fit according to the corresponding state principle using LaSbTe as a reference sample. Data up to 30 K is shown for clarity ; data over a larger temerature range is provided in the Supporting Information. Inset: specific heat measured under various magnetic fields increasing from top to bottom . Data for additional fields are provided in the Supporting Information. (b) Magnetic specific heat divided by temperature , Cm/T, and the corresponding magnetic entropy , Sm. 16 Figure 4 G - X Fig. 4 Electronic structure of SmSbTe in the high temperature PM phase. (a) The Fermi surface and constant energy contours of SmSbTe measured with photon energy of 30 eV at T = 78 K . A four -leave clover feature is seen at (0,0) on the second subplot. (b) Calcula ted Fermi surface and constant energy contours of SmSbTe with kz = 0 corresponding to the measured spectra in (a). The color scale indicate s the surface character , with 0 (light brown ) referring to bulk state and 1 (dark brown ) referring to surface state s; see ref [71]. (c) Band structures of SmSbTe measured with a 30 eV photon energy along the M-G-M and X-G-X directions. Black arrows indicate the linear band crossing points along the M-G-M path. (d) Band structure of SmSbTe measured with a 73 eV photon energy. (e) Band structure of five-unit-cell SmSbTe slabs along the M-G-M and X-G-X directions with SOC in the PM state. The dashed horizontal line is the E F experimentally determined by ARPES, which is shifted by 0.275 eV as compared with th e calculated E F. (f) Bulk and slab Brillouin zone s for SmSbTe. 17 Figure 5 Fig. 5 (a) PM state electronic band structure of SmSbTe with SOC. The dashed lines indicate the shift of the Fermi energy of 0.275 eV from the comparison between the AR PES spectrum (Figure 4) and the calculated band structure in the PM state. (b) AFM state electronic band structure of SmSbTe with SOC. Note that the upper split flat band is beyond the energy range shown in the plot (see Figure S 5 in Supporting Information for an AFM band structure without SOC). 18 References: [1] M. Z. Hasan, C. L. Kane, Rev. Mod. Phys. 2010, 82, 3045. [2] X. -L. Qi, S. -C. Zhang, Rev. Mod. Phys. 2011, 83, 1057. [3] B. Yan, C. Felser, Annu. Rev. Condens. Matter Phys. 2017, 8, 337. [4] N. P. Armitage, E. J. Mele, A. Vishwanath, Rev. Mod. Phys. 2018, 90, 015001. [5] A. Bernevig, H. Weng, Z. Fang, X. Dai, J. Phys. Soc. Jpn 2018, 87, 041001. [6] N. Nagaosa, T. Morimoto, Y. Tokura, Nat. Rev . Mat. 2020, 5, 621. [7] C. Fang, H. Weng, X. Dai, Z. Fang, Chin. Phys. B 2016, 25, 117106. [8] T. Liang, Q. Gibson, M. N. Ali, M. Liu, R. J. Cava, N. P. Ong, Nat. Mat. 2015, 14, 280. [9] J. Xiong, S. K. Kushwaha, T. Liang, J. W. Kriza n, M. Hirschberger, W. Wang, R. J. Cava, N. P. Ong, Sci. 2015, 350, 413. [10] Q. Li, D. E. Kharzeev, C. Zhang, Y. Huang, I. Pletikosic, A. v. Fedorov, R. D. Zhong, J. A. Schneeloch, G. D. Gu, T. Valla, Nat. Phys. 2016, 12, 550. [11] L. Wu, S. P atankar, T. Morimoto, N. L. Nair, E. Thewalt, A. Little, J. G. Analytis, J. E. Moore, J. Orenstein, Nat. Phys. 2017, 13, 350. [12] Q. Ma, S. -Y. Xu, C. -K. Chan, C. -L. Zhang, G. Chang, Y. Lin, W. Xie, T. Palacios, H. Lin, S. Jia, P. A. Lee, P. Jarillo -Herrero, N. Gedik, Nat. Phys. 2017, 13, 842. [13] G. B. Osterhoudt, L. K. Diebel, M. J. Gray, X. Yang, J. Stanco, X. Huang, B. Shen, N. Ni, P. J. W. Moll, Y. Ran, K. S. Burch, Nat. Mat. 2019, 18, 471. [14] Z. Dai, M. Manjappa, Y. Yang, T. C. W. Tan, B. Qiang, S. Han, L. J. Wong, F. Xiu, W. Liu, R. Singh, Adv. Funct. Mat. 2021, 2011011. [15] H. Wang, X. Luo, K. Peng, Z. Sun, M. Shi, D. Ma, N. Wang, T. Wu, J. Ying, Z. Wang, X. Chen, Adv. Funct. Mat. 2019, 29, 1902437. [16] G. D'Olimpio , C. Guo, C. -N. Kuo, R. Edla, C. S. Lue, L. Ottaviano, P. Torelli, L. Wang, D. W. Boukhvalov, A. Politano, Adv. Funct. Mat. 2020, 30, 1906556. 19 [17] Q. Xu, Z. Song, S. Nie, H. Weng, Z. Fang, X. Dai, Phys. Rev. B 2015, 92, 205310. [18] L. M. Scho op, M. N. Ali, C. Strasser, A. Topp, A. Varykhalov, D. Marchenko, V. Duppel, S. S. P. Parkin, B. v. Lotsch, C. R. Ast, Nat. Commun. 2016, 7, 11696. [19] J. Hu, Z. Tang, J. Liu, X. Liu, Y. Zhu, D. Graf, K. Myhro, S. Tran, C. N. Lau, J. Wei, Z. Mao, Phy s. Rev. Lett. 2016, 117, 016602. [20] R. Lou, J. -Z. Ma, Q. -N. Xu, B. -B. Fu, L. -Y. Kong, Y. -G. Shi, P. Richard, H. -M. Weng, Z. Fang, S. -S. Sun, Q. Wang, H. -C. Lei, T. Qian, H. Ding, S. -C. Wang, Phys. Rev. B 2016, 93, 241104. [21] M. Neupane, I. Belopolski, M. M. Hosen, D. S. Sanchez, R. Sankar, M. Szlawska, S. - Y. Xu, K. Dimitri, N. Dhakal, P. Maldonado, P. M. Oppeneer, D. Kaczorowski, F. Chou, M. Z. Hasan, T. Durakiewicz, Phys. Rev. B 2016, 93, 201104. [22] D. Takane, Z. Wang, S. Souma, K. Nakayama, C. X. Trang, T. Sato, T. Takahashi, Y. Ando, Phys. Rev. B 2016, 94, 121108. [23] M. M. Hosen, K. Dimitri, I. Belopolski, P. Maldonado, R. Sankar, N. Dhakal, G. Dhakal, T. Cole, P. M. Oppeneer, D. Kaczorowski, F. Chou, M. Z. Hasan, T. Duraki ewicz, M. Neupane, Phys. Rev. B 2017, 95, 161101. [24] C. Chen, X. Xu, J. Jiang, S. -C. Wu, Y. P. Qi, L. X. Yang, M. X. Wang, Y. Sun, N. B. M. Schroter, H. F. Yang, L. M. Schoop, Y. Y. Lv, J. Zhou, Y. B. Chen, S. H. Yao, M. H. Lu, Y. F. Chen, C. Felse r, B. H. Yan, Z. K. Liu, Y. L. Chen, Phys. Rev. B 2017, 95, 125126. [25] M. M. Hosen, G. Dhakal, K. Dimitri, P. Maldonado, A. Aperis, F. Kabir, C. Sims, P. Riseborough, P. M. Oppeneer, D. Kaczorowski, T. Durakiewicz, M. Neupane, Sci. Rep. 2018, 8, 13 283. [26] L. M. Schoop, A. Topp, J. Lippmann, F. Orlandi, L. Muchler, M. G. Vergniory, Y. Sun, A. W. Rost, V. Duppel, M. Krivenkov, S. Sheoran, P. Manuel, A. Varykhalov, B. Yan, R. K. Kremer, C. R. Ast, B. v. Lotsch, Sci. Adv. 2018, 4, eaar2317. 20 [27] J. Hu, Y. Zhu, X. Gui, D. Graf, Z. Tang, W. Xie, Z. Mao, Phys. Rev. B 2018, 97, 155101. [28] B. -B. Fu, C. -J. Yi, T. -T. Zhang, M. Caputo, J. -Z. Ma, X. Gao, B. Q. Lv, L. -Y. Kong, Y.-B. Huang, P. Richard, M. Shi, V. N. Strocov, C. Fang, H. -M. Wen g, Y. -G. Shi, T. Qian, H. Ding, Sci. Adv. 2019, 5, eaau6459. [29] J. Hu, Z. Tang, J. Liu, Y. Zhu, J. Wei, and Z. Mao, Phys. Rev. B 2017, 96, 045127 . [30] J. Hu, Y. L. Zhu, D. Graf, Z. J. Tang, J. Y. Liu, and Z. Q. Mao, Phys. Rev. B 2017, 95, 205134. [31] M. M. Hosen, K. Dimitri, A. Aperis, P. Maldonado, I. Belopolski, G. Dhakal, F. Kabir, C. Sims, M. Z. Hasan, D. Kaczorowski, T. Durakiewicz, P. M. Oppeneer, M. Neupane, Phys. Rev. B 2018, 97, 121103. [32] A. Topp, J. M. Lippman n, A. Varykhalov, V. Duppel, B. V. Lotsch, C. R. Ast, L. M. Schoop, New J. Phys. 2016, 18, 125014. [33] S. Pezzini, M. R. van Delft, L. M. Schoop, B. v. Lotsch, A. Carrington, M. I. Katsnelson, N. E. Hussey, S. Wiedmann, Nat. Phys. 2018, 14, 178. [34] Y. Shao, A. N. Rudenko, J. Hu, Z. Sun, Y. Zhu, S. Moon, A. J. Millis, S. Yuan, A. I. Lichtenstein, D. Smirnov, Z. Q. Mao, M. I. Katsnelson, D. N. Basov, Nat. Phys. 2020, 16, 636. [35] A. Topp, R. Queiroz, A. Gruneis, L. Muchler, A. W. Rost, A. Varykhalov, D. Marchenko, M. Krivenkov, F. Rodolakis, J. L. McChesney, B. v Lotsch, L. M. Schoop, C. R. Ast, Phys. Rev. X 2017, 7, 041073. [36] X. Liu, C. Yue, S. V. Erohin, Y. Zhu, A. Joshy, J. Liu, M. S. Ana, D. Graf, P. B. Sorokin, Z. Mao, J. H u, J. Wei, Nano letters, in press. DOI: [37] D. W. Boukhvalov, R. Edla, A. Cupolillo, V. Fabio, R. Sankar, Y. Zhu, Z. Mao, J. Hu, P. Torelli, G. Chiarello, L. Ottaviano, A. Politano, Adv. Funct. Mat. 2019, 29, 1900438. 21 [38] S. Lei, S. M. L. Teicher, A. Topp, K. Cai, J. Lin, G. Cheng, T. H. Salters, F. Rodolakis, J. L. McChesney, S. Lapidus, N. Yao, M. Krivenkov, D. Marchenko, A. Varykhalov, C. R. Ast, R. Car, J. Cano, M. G. Vergniory, N. P. Ong, and L. M. Schoop, Adv . Mater. 2021, 33, 2101591 [39] B. Lv, J. Chen, L. Qiao, J. Ma, X. Yang, M. Li, M. Wang, Q. Tao, Z. -A. Xu, J. Phys. Condens. Matter 2019, 31, 355601. [40] R. Sankar, I. P. Muthuselvam, K. R. Babu, G. S. Murugan, K. Rajagopal, R. Kumar, T.-C. W u, C. -Y. Wen, W. -L. Lee, G. -Y. Guo, F. -C. Chou, Inorg. Chem. 2019, 58, 11730. [41] S. Lei, V. Duppel, J. M. Lippmann, J. Nuss, B. v. Lotsch, L. M. Schoop, Adv. Quant. Tech. 2019, 2, 1900045. [42] M. Yang, Y. Qian, D. Yan, Y. Li, Y. Song, Z. Wan g, C. Yi, H. L. Feng, H. Weng, Y. Shi, Phys. Rev. Mat. 2020, 4, 094203. [43] K. Pandey, R. Basnet, A. Wegner, G. Acharya, M. R. U. Nabi, J. Liu, J. Wang, Y. K. Takahashi, B. Da, J. Hu, Phys. Rev. B 2020, 101, 235161. [44] P. Li, B. Lv, Y. Fang, W. Guo, Z. Wu, Y. Wu, D. Shen, Y. Nie, L. Petaccia, C. Cao, Z.-A. Xu, Y. Liu, Sci. Chin. Phys., Mech. & Astro. 2021, 64, 237412. [45] S. Yue, Y. Qian, M. Yang, D. Geng, C. Yi, S. Kumar, K. Shimada, P. Cheng, L. Chen, Z. Wang, H. Weng, Y. Shi, K. Wu , B. Feng, Phys. Rev. B 2020, 102, 155109. [46] Y. Qian, Z. Tan, T. Zhang, J. Gao, Z. Wang, Z. Fang, C. Fang, H.Weng, Sci. China: Phys., Mech. Astron. 2020, 63, 107011. [47] S. Lei, A. Saltzman, L. M. Schoop, Phys. Rev. B, 2021, 103, 134418. [48] E. DiMasi, B. Foran, M. C. Aronson, S. Lee, Phys. Rev. B 1996, 54, 13587. [49] K. Ghosh, C. Mazumdar, R. Ranganathan, S. Mukherjee, Sci. Rep. 2015, 5, 15801. [50] B. Wilfong, X. Zhou, H. Vivanco, D. J. Campbell, K. Wang, D. Graf, J. Paglione, E. Rodriguez, Phys. Rev. B 2018, 97, 104408. 22 [51] R. Sankar, I. P. Muthuselvam, K. Rajagopal, K. Ramesh Babu, G. S. Murugan, K. S. Bayikadi, K. Moovendaran, C. Ting Wu, G. -Y. Guo, Cryst. Growth Des. 2020, 20, 6585. [52] T. Taniguchi, Y. Kawaji, T.C. Ozawa, Y. Nagata, Y. Noro, H. Samata, M.D. Lan, J. Alloy Compd. 2005, 386, 63 -69. [53] S. Pakhira, C. Mazumdar, R. Ranganathan, M. Avdeev, Sci. Rep. 2017, 7, 7367. [54] Y. Okamoto, H. Yoshida, Z. Hiroi, J. Phys. Soc. Jpn. 2009, 78, 033701. [55] C. Lacroix, P. Mendels, and F. Mila, Introduction to Frustrated Magnetism: Materials, Experiments, Theory (Springer Berlin Heidelberg, 2011) [56] T. Kurumaji, T. Nakajima, M. Hirschberger, A. Kikkawa , Y. Yamasaki, H. Sagayama, H. Nakao, Y. Taguchi, T. -h. Arima, and Y. Tokura, Science 2019, 365, 914 [57] L. Balents, Nature 2010, 464, 199 -208 [58] W. -J. Hu, H. -H. Lai, S. -S. Gong, R. Yu, E. Dagotto, Q. Si, Phys. Rev. Res. 2020, 2, 023359. [59] Z. Ma, K. Ran, J. Wang, S. Bao, Z. Cai, S. Li, J. Wen, Chin. Phys. B 2018, 27, 106101. [60] J. Vannimenus, G. Toulouse, J. Phys. C: Solid St. Phys. 1977, 10, L537. [61] J. Hu, T. J. Liu, B. Qian, A. Rotaru, L. Spinu, Z. Q. Mao, Phys. Rev. B 2011, 83, 134521. [62] A. A. Zvyagin, Low Temp. Phys. 2013, 39, 901. [63] J. Xing, L. D. Sanjeewa, J. Kim, G. R. Stewart, M. -H. Du, F. A. Reboredo, R. Custelcean, A. S. Sefat, ACS Mat. Lett. 2020, 2, 71. [64] Y. Wang, Y. Qian, M. Yang, H. Chen, C. Li, Z. Tan, Y. Cai, W. Zhao, S. Gao, Y. Feng, S. Kumar, E. F. Schwier, L. Zhao, H. Weng, Y. Shi, G. Wang, Y. Song, Y. Huang, K. Shimada, Z. Xu, X. J. Zhou, G. Liu, Phys. Rev. B 2021, 103, 125131. [65] J. Lee, S. Yuan, S. Lal, Y. Joe, Y. G an, S. Smadici, K. Finkelstein, Y. Feng, A. Rusydi, P.M. Goldbart, S.L.Cooper, P. Abbamonte, Nat. Phys. 2012, 8, 63 -66.

2108.10983

Krishna Pandey

Krishna Pandey, Debashis Mondal, John William Villanova, Joseph Roll, Rabindra Basnet, Aaron Wegner, Gokul Acharya, Md Rafique Un Nabi, Barun Ghosh, Jun Fujii, Jian Wang, Bo Da, Amit Agarwal, Ivana Vobornik, Antonio Politano, Salvador Barraza-Lopez, Jin Hu

Magnetic Topological Semimetal Phase with Electronic Correlation Enhancement in SmSbTe

23 pages, 5 figures

Adv. Quantum Technol. 2021, 2100063

10.1002/qute.202100063

cond-mat.mtrl-sci cond-mat.mes-hall cond-mat.str-el

http://creativecommons.org/licenses/by/4.0/

The ZrSiS family of compounds hosts various exotic quantum phenomena due to the presence of both topological nonsymmorphic Dirac fermions and In this material family, the LnSbTe (Ln= lanthanide) compounds are particularly interesting owing to the intrinsic magnetism from magnetic Ln which leads to new properties and quantum states. In this work, the authors focus on the previously unexplored compound SmSbTe. The studies reveal a rare combination of a few functional properties in this material, with possible magnetic frustration, electron and Dirac nodal-line fermions. These properties enable SmSbTe as a unique platform to explore exotic quantum phenomena and arising from the interplay between magnetism, topology, and electronic correlations.

opment of the Corvus [opcons] work ow, where Corvus is the work ow engine,[24] and opcons is the target property. In particular, the approach imports the structural and property data from the MP, and then parallelizes and au- tomates the calculations permitting high-throughput cal- culations of optical constants for materials throughout the MP database. Our procedure for theoretical calculations of optical constants over all edges essentially follows that described in detail by Prange et al. ,[20] but has been up- dated for the FEFF10 code. In addition, renements for including vibrational disorder via the correlated Debye model using MP data have been added and the calcula- tions for the UV-VIS range have been simplied. To il- lustrate the approach calculations have been carried out for nearly all elemental solids throughout the periodic ta- ble and explicit examples are given for Cu, Ag, and Au, together with comparisons to experimental data. In ad- dition we show results for a sample compound Al 2O3to validate our approximation in the UV-VIS range. The Corvus [opcons] work ow described here is applicable to any material available in the MP database as dened by a given MP identicatoin label (MP-ID), thus making pos- sible routine calculations of optical constants for over 105 structures. By default, the calculations are carried out with lattice vibration eects included at room tempera- ture, while electronic temperature and thermal expansion eects are neglected. However, corrections for thermal ex- pansion, nite temperature, and pressures other than 1 atm can be added using subsequent calculations that reuse some of the previously computed data. The work ow is naturally parallelized and, for typical settings, requires as little as minutes of wall-clock time per material on mas- sively parallel systems. 2. Optical constants with Corvus and FEFF10 In general, the optical constants are related to the com- plex dielectric constant (!) =1(!) +i2(!) in the long wave-length limit, where !is the frequency of the eld. Since their real and imaginary parts are re- lated by Kramers-Kronig (KK) transformations, they can all be calculated from the imaginary part 2(!) (e.g. Fig. 1). In this work the, the Corvus [opcons] work ow yields the complex dielectric function (!), energy loss function L(!), complex index of refraction n(!), absorption and re ectivity R(!) dened as follows, 0.1 1 10 100 1000 10000 (eV)Expt. (DESY) FEFF 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 2 4 6 8 10 12 14kh(E) k(E) 1: Top: Calculated imaginary part of the dielectric function 2for Cu (blue) along with the atomic background (see text) (red). Bottom: Approximate optical ne structure of the imaginary (blue) part offor Cu vsk=p 2E, along with the same with energies Erelative to the K- and L-edges. Note the similarities between the ne structure found in the optical and that of the K-edge. Note also that the ne structure becomes negligible for largekabove any edges or outside the optical range. terms of2(!), 1(!) = 1 =Im[1(!)] =n(!) +i(!) =(!)1=2(3) (!) = 2! c(!) (4) R(!) + 1]2+(!)2; the principal part of the integral. Note that at high energies dielectric response is weak, Sum-rules for the dielectric properties provide a qualitative check on the reliability of the results.[25] Related properties, such as the local atomic polarizability (!) = the Rayleigh forward scattering amplitudes f(!) the classi- cal radius of the electron, cis the speed of light and n=N=V is the atomic number density), electron energy loss spectra (EELS), inelastic mean-free paths, etc., can also be determined, [20] but are not yet implemented in the present Corvus For simplicity here and below we use atomic unitse= h=m= 1 throughout this paper unless other- wise needed for clarity. The treatment of excitations from the core- and valence- levels generally requires dierent considerations, so it is convenient to separate the calculations the contribution from all levels below a xed core-valence separation energy Ecv, which in FEFF10 is set to -40 eV. Although this separation is approximate due to many-body eects, the contributions from deep core levels (and to a reasonable approximation, respond nearly independently from the valence. 2.1. Core level contributions In this paper calculations of the optical constants are carried out using the real-space Green's function code FEFF10 and an updated version of the approach de- scribed by Prange et al. [20] FEFF uses the same real- space Green's function formalism to calculate a variety and related quantities including XAS, XES, RIXS, EELS, and Compton. Thus the calculations of the optical constants use essentially the same ingredients as those for the well-established calculations of XAS, the work ow design. Brie y, the RSGF approach in FEFF is used to calculate the contribution to 2(!) from each occupied level edge) at each unique site in the system a. For simplicity, the site index will be usually suppressed below. The core level contribution to 2(!) from a given site a consists of a sum of contributions from occupied core levels jii[20] fori<Ecvgiven the core-orbital of interest, iits binding en- ergy, ^dis the transition operator, which in most cases is dipole dominated, and G(!) is the one-electron Finally, (!F+i) is a unit step function that turns on when the photon energy !is sucient to excite the core-electron above the Fermi energy F. If the interaction between the core- and photoelectron levels is ignored, this level of approximation is equivalent to the RPA. However, in this work we use the nal-state to the core hole interaction, in which Green's function is calculated in the pres- ence of a screened core-hole. -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 8980 9000 9020 9040 9060CuXANES (arb. units) Energy 2: Cu K-edge spectrum obtained using the stitching For comparison, the curves calculated using FMS (red) and MSP (blue) are also shown, shifted above and below the stitched curve, respectively. The green curve is the dierence between the FMS and MSP calculated results, and the shaded area shows the region where the stitching algorithm is applied. One of the advantages of the RSGF approach is the separation of the propagator G(!) into the central absorber at a given site and single contributions from neighboring atoms, i.e.,G(!) =Gabs(!) +Gsc(!). As a result the dielec- tric properties exhibit ne structure analogous to that in XAFS, +i(!)]; (8) which we dub optical ne structure (OFS). Typically i(!) is of order a few percent within a range of order 100 eV above each edge; however, i(!) can be large and of order unity in the near-edge regime, within about 10 eV or so of a given edge. Thus the physical interpretation of the OFS is similar to that for EXAFS and XANES, re ecting the local geometrical structure in the vicinity of an ab- sorbing atom. These observations also explain why a local, atomic approximation that neglects the ne structure from neighbor scattering can be a reasonable approximation, ac- curate to within a few percent for the core contributions to the optical constants, except close to an absorption edge. In order to show the ne structure over the entire energy range of our calculations, it is convenient to calculate dened as 2= (2atomic 2 )=atomic 2 , which is shown for Cu in Fig. 1 (bottom), together with the full and atomic approximation to 2(top). A can also be made for any of the other op- tical constants. When core 11, the ne structure in the loss function L(!), absorption and imaginary part of the index of refraction are similar to that in 2. The ne structure in 1is similar, although phase shifted by 90. In order to obtain optical constants over a broad en- ergy range from the UV to hard x-rays, two FEFF cal- 3culations must be performed for each edge ( i.e.each oc- cupied atomic level i) and each unique atomic site ain the presence of the corresponding core-hole: 1) In the near edge regime within about 50 eV of a given (FMS) calculations must be per- formed, and 2) at higher energies (the EXAFS regime), path expansion (MSP) is used. Once these calculations are performed, the results are interpo- lated onto a common energy grid and stitched together to yield the spectrum i 2(!) coming from the core-state jii. The stitching algorithm used is:[20] i 2(k) =FMS;i 2 (k); k<k 0; = (k) + (k); k>k 1(9) wherek0= 3A1,k1= 4A1, andk= the photoelectron wave-number. This is illustrated in Fig. 2 for the K-edge spectrum of fcc Cu. 2.2. Valence contribution from the valence levels is complicated by the extended range and continuum nature of the va- lence wave functions and the large density of states in the valence region. In the full RSGF approach this re- quires sums over multiple sites as in Eq. (32) of Ref. [20]. Moreover, systems with strong excitonic eects often re- quire treatments beyond the RPA such as TDDFT or equation.[9] Several codes are available for these advanced treatments,[10, 11, 13, 15] but they challenging for high throughput calcu- lations. Nevertheless, neglecting these eects is often a reasonable approximation for in this work we use a local approximation at the absorption site to calculate the contribution from the valence electrons. In addition, we replace the with the Dirac-Fock atomic allows us to write Eq. (32) of Ref. [20] as an integral over the local valence density of states val 2(!) is the local atomic spectrum at a given to atomic state i, and ~==N is the normalized, angular momentum projected density of states (LDOS) with symmetry equal to that of the state i, and
=EtopEedgeis a shift to properly position the valence spectra, where Etopis the energy at the top of the gap, if one exists, and is equal to the Fermi energy otherwise. As shown for the energy loss of Al 2O3(top) and Au (bottom) in Fig. 3, which provide good tests for this procedure we nd that the approximation is reasonably accurate when 10 100 1000 Energy Loss Expt. (DESY) FEFF (Corvus)Al2O3 1 10 100 1000 Energy LossExpt. (DESY) FEFF (Corvus) AuFigure 3: Comparison of the experimental[26] and calculated energy loss function for Al 2O3(corundum) (top) and Au (gold) (bottom) with the local approximation (see text). excitonic eects are small. We have found that our in Eq. (10) is signicantly better than a pure atomic approximation alone. One can also dene an optical ne structure valfor the valence although its magnitude is typically large and of order unity (see Fig. 1), as in core-level XANES, over the optical energy range. For this reason, FMS and the local of states are needed to evaluate Eq. (10). In addition, the physical interpretation of val 2is akin to a sum of terms proportional to the joint projected density of states.[20] Even so, the valence ne structure for Cu is rather similar to the K-edge ne structure, as shown in the lower panel of Fig. 1. This makes sense, since the d- to p-state transitions dominate in the optical region, and the DOS is dominated by a single, relatively sharp the total 2(!) is obtained from a double sum over edgesiand unique sites in the system, 2(!) =X ianaia 2(!); a unique site in the unit cell, and 0.1 1 10 100 1000 10000 100000Agu (cm-1) Energy (eV)Expt. (DESY) FEFF (Corvus) Without DrudeFigure 4: Comparison of the theoretical absorption of Ag with and without the Drude contribution, along with experimental results [26]. 1 10 100 1000 10000 100000 Energy (DESY) FEFF 5: Comparison of the experimental[26] and theoretical 2(!) sum rulene(!) Cu. The horizontal dashed line indicates the ideal theoretical limit ( Z= 29). Na=Vis the corresponding number density. 2.3. Debye-Waller factors and Drude corrections A number of other considerations are generally impor- tant in calculations of optical spectra over broad energy ranges. For example, eects of thermal vibrations are tem- perature dependent and strongly damp the ne-structure in the EXAFS regime. For the systems treated here we have approximated the path dependent DW factors exp(22k2) using the default correlated De- bye model in FEFF10. The Debye temperature parameter can be estimated[27] from the bulk and shear elastic mod- uli data available in the MP database. Drude corrections in metals can be added by an additional ad hoc contribution to the lim- iting low frequency behavior of 2(!) given 6: Schematic of the Corvus work ow for optical constants. User input is shown in light green, while code blocks associated with Corvus are shown in blue, and associated with FEFF is shown in black. The output is shown in light blue. Note that Corvus [opcons] requires minimal input from the user (only the CIF le and a request for optical constants), although many are the conductivity, !2 p= 4ne2=m,nis the mean conduction electron density, and the relaxation time, which is typically of order 0 :14:0 femtosecond at Fig. 4 shows a comparison of the absorp- tionof Ag with and without the Drude with expansion can also be treated by calculating the expansion parameter via DFT,[29] or using parame- ters from the literature. Additional broadening due interactions, impurities, and other mech- anisms can also be important. While neglected by default, these eects can be added ex post facto by a of the is obtained by summing both core and va- lence contributions over all sites, the real part of the dielec- tric function is formed via a KK-transform, as in Eq. (1). A check on the completeness of the calculations is provided by the optical constant sum-rules,[25] as illustrated in Fig. 5. 2.4. Corvus [opcons] work ow The Corvus [opcons] work ow is built on the concept of recursive work ow generation, wherein Corvus generates and runs a work ow with the target property opcons , and the optical constants work ow [opcons] in turn generates and runs a [xanes] work ow for each required The work ow tool is described in detail to- gether with various examples in the Corvus approach determines the input les from the data in the MP database, afterwhich all of the individual are set up and run through the Corvus work- ow manager. The Corvus [opcons] work ow then car- ries out the following steps: i) calculating the 2(!) for a given edge; ii) applying the stitching algorithm to combine the near-edge full multiple scattering (FMS) and path (MSP) calculations; iii) convolving with the density of states (Eq. 10); iv) performing the KK- transform to obtain 1; v) adding the Drude found 16 systems Keeping 3 systems after energy screening Found systems : # # System : mp -1279742 Reduced Form .: FeO Cell Comp .: Fe4 O4 Sym . Unique Comp .: Fe1 O1 Total num . of edges : 15 Estimated Debye T: 466.60 ( Pred .) # # System : mp -715262 Reduced Form .: FeO Cell Comp .: Fe2 O2 Sym . Unique Comp .: Fe1 O1 Total num . of edges : 15 Estimated Debye T: 416.52 ( Calc .) # # System : mp -1178232 Reduced Form .: FeO Cell Comp .: Fe4 O4 Sym . Unique Comp .: Fe1 O1 Total num . of edges : 15 Estimated Debye T: 465.00 ( Pred .) # # Keeping 3 systems after symmetry and edge screening Total number of edges to compute in set : 45 Figure 7: Typical output for the showing that the MP query returned 16 possible systems, of which only 3 are left after screening for energy, symmetry, and total number of edges. Also shown are the estimated Debye temperatures, which can be calculated (e.g., from DFT data), or predicted (e.g., from machine learning data). and vi) calculating the other optical constants following the relations in Eqs. (1). In order to facilitate Corvus handles the creation of FEFF10 in- put les from minimal user input, executes FEFF10, and translates the output. Additionally, the Corvus [opcons] work ow is designed to optimize the calculations, for ex- ample re-using output from the XANES results within the EXAFS calculations to bypass the self-consistency loop. Figure 6 shows a schematic of the Corvus [opcons] work- ow for producing the optical constants. 3. Materials Project interface tools 3.1. Global structure of Corvus [opcons] simulations MP optical constants simulations are set up in two steps: 1) The reads the command line inputs, creates a request to the MP, downloads all the structures available matching the query, generates a set of materials to run and calculates the totl number of edges in the set. 2) The runs all the calculations required for the set, distributed in serial or parallel mode according to the resources provided. In the following section we describe their usage in { opcons } usehandlers { Feff } opcons . usesaved { True } cif_input { CIF_symm . cif} feff . scf { 0 30 0.1 0 } feff . fms { 0 0 0.0 0.0 40.0 } feff . debye { 298.0 0 } # The lines below should set the parallel 8: Typical template Corvus [opcons] input generated by the . This input shows the selected Corvus tar- getopcons , the handlers that sould be used (only FEFF is required for these types of runs), the location of the structure in CIF format, and the FEFF SCF and FMS radii. The latter are automatically gen- erated based on the number of shells requests in the command line. Depending on the options used, the calculation can use a Correlated Debye model to add thermal broadening, using a Debye estimated from MP database values of the bulk and shear moduli. Given that Corvus stores the MPI command, number of processors, etc in the input, the actual input is generated at runtime by the crvmprunset, where the required parameters are added at the end of the template input shown here. 3.2. Usage: crv mpmkset The basic input of a typical Corvus [opcons] calcula- tion is set up using the : crvmpmkset [OPTIONS] Set Name where SetName is a text label that will be used through- out to refer to the set of materials. This label is used to name the directory where the calculations will be done and the results stored. crvmpmksetcan be used to query the MP database in three dierent ways (see options below): 1) Query an n arbitrary class of materials selected by for- mula, 2) automatically query some or all of the elemen- tal solids, for which the tools have an internal database of MP-IDs, and nally, 3) an individual material with a given MP-ID. The full set of options can be obtained with crvmpmkset --h . Here we focus on some of the relevant ones. 3.2.1. Required options --k APIKEY API key used to interact with the MP REST interface This option is required for the command to be able to in- teract with the MP database. See section 3.4.4 for details. 3.2.2. System selection options --f FORMULA Use chemical formula FORMULA to dene a set of a set for all the solid elements in the periodic table using the structure for the most stable phase in standard PT ELEMS Same as --pt , but only for a list of specic elements 6--mpid MP-ID Select the individual system with Materials label MP-ID (Default: mp-30) These options are mutually exclusive. If none of them are present the command generates input for a test run on Cu, a traditional choice since FEFF was rst deployed. 3.2.3. Selection ne-tuning options --mxpt MXNAPT Maximum atomic number to use while creating the pe- riodic table with --pt (Default: 99) --enpct EN PCT Keep only ENPCT% of lowest energy systems MXNEDGES Keep only systems with fewer than MXNEDGES edges (De- fault: 250) These options help trim the raw selection generated by the system selection options. 3.2.4. Auxiliary options --nSCF NSCF Number of neighbor shells to use in SCF (Default: 1) --nFMS NFMS Number of neighbor shells to use in FMS (Default: 2) --symprec SYM PREC Symmetry precision (in A) used by symmetrizer (De- fault: 0.01 A) --dw Use the correlated Debye thermal broadening model in the optical constants (in K) used for correlated Debye (Default: 298K) These options control the setup of the optical by dening how many shells to use in the SCF and FMS steps, as well as the precision of the sym- metrization procedure used to reduce the number of unique sites. The shell selection is performed using the Jenks nat- ural breaks classication method[30] as implemented in thejenkspy Python module.[31] Additionally, can also be controlled by requesting that a cor- related Debye model[32] be used. 3.2.5. Calculation control options --dr Dry run, collect and display information without actu- ally creating a set --v Make the command output more verbose 10 100 1000 10000 1e+05 Energy (DESY) FEFF (Corvus, (1,2)) FEFF (Corvus, (3,6)) FEFF (Corvus, (4,8))1000 1200 1400 Energy 10 100 1000 Energy (eV)0.01.02.0 Index of Refraction Expt. (DESY) FEFF (Corvus)AuFigure 9: Comparison of the experimental[26] and theoretical 2for Cu as a function of FEFF Corvus (nSCF,nFMS) shells to (top) and (bottom) index of refraction for Au in the valence region (bottom). 3.2.6. Examples A typical invocation of crvmpmksetis as --k [APIKEY] --f FeO --enpct 20 --nedg 250 --nSCF 2 --nFMS 4 --v --dw --temp 298.0 a set of 3 systems (Figure 7) after screening for en- ergy, symmetry and total number of edges. The command also requests that the Debye temperature be estimated, and that the calculation be done at room the set directory, the structures are organized by reduced formula and MP-ID. A typical Corvus created with for each material, a template Corvus [opcons] input (Fig. 8) and a few auxiliary les with structural information. A set for the Li, Be and B elemental solids can be generated --k [APIKEY] --v --ptel 'Li,Be,B' LiBeB while a single compound (one of the forms of Mg 3N, in this case) can be set up with: 71 - H Hydrogen 3 - Fr Francium4 - Ra Radium21 - Db Dubnium24 - Bh Bohrium26 mp-13 Fe Iron 44 - Hs Hassium27 mp-79 Zn Zinc 48 - Hg Mercury 112 - Uub Ununbium31 - Uut Ununtrium6 mp-84 Sn Tin 82 - N Nitrogen 15 - O Oxygen 16 - F Flourine 17 - Cl Chlorine 35 - Br Bromine 53 - Ne Neon2 - He Helium 18 - Ar Argon 36 - Kr Krypton 54 - Xe Xenon 86 - Rn Radon 118 IA 2 IIA 3 IIIA 4 IVB 5 VB 6 VIB 7 VIIB 8 VIIIB 9 VIIIB 10 VIIIB 11 IB 12 IIB13 IIIA 14 IVA 15 VA 16 VIA 17 VIIA18 VIIIA 57 - Am Americium96 - Cm Curium97 - Bk Berkelium98 - Fm Fermium101 - No Nobelium103 Available Z Table of FEFF Optical ConstantsFigure 10: Periodic table of FEFF optical constants currently available. The index on the top-right corner of each element indicates the MP-ID of the structure used in the calculations. A dynamic version of this table can be found at ,[33] including all optical constants data as well as the input les used to generate it. crvmpmkset --k [APIKEY] --v examples can be found in Ref. [34]. 3.3. Usage: crv mprunset After a set is created, the actual optical constants can be computed with the : crvmprunset [OPTIONS] Set Name where SetName is again the label for the set to run. As for crvmpmkset, the full set of options can be obtained with crvmprunset --h , but here we focus on some of the most computationally relevant ones. 3.3.1. Options to control the number of processors --np NP TOT Total number of processors to be used in the run (De- fault: 1) --ppn PPN Number of processors per node (Default: 1) --nn NNODES Number of nodes with PPNprocessors per node to be used in the run (Default: 1) If more than one of these options are specied, the com- mand will choose max(NP TOT,PPN*NNODES) as the total number of processors to use.3.3.2. Options to estimate computational total runtime up to SER FRAC Fraction of serial code to estimate parallel runtimes (De- fault: 0.5) If the --t runtime option is present, instead of running the set, out estimates of the cost of the calculation from 1 to NPTOTprocessors relative to the cost of a single edge calculations performed in serial mode. 3.3.3. Auxiliary verbose output (timings, etc) --dr Dry run: Do everything but actually run 3.3.4. Examples A typical instance of crvmprunset to run the set IronOxides created in the previous example (Fig. 7) is as --ppn 64 -nn 45 --v -np 2880 --v IronOxides Given that the set contains 45 edges, for maximum ef- ciency we assign one node with 64 processors to each 8edge, or a total of 2880 processors, thus eectively run- ning all materials simultaneously. FEFF10 then computes each individual edge parallelizing over the energy grid with 64 processors. 3.4. Code Corvus[35] and FEFF10[36] are open source and are available on GitHub. Detailed installation instructions can be found in Ref. [34]. 3.4.2. Corvus requirements The Corvus work ow manager is implemented in Python 3. In addition, it requires the Numpy[37] and Scipy[38] packages, as well as a few other packages depend- ing on the required functionality. The Corvus work ows that are possible will be determined by the specic ex- ternal software packages installed on the system. In this case, the optical constants WF only requires the appropri- ate version of FEFF10. 3.4.3. FEFF10 requirements To run eciently, a parallel execution is important. Thus, the FEFF code requires a Fortran90 compiler with some avor of MPI-2 for compilation. FEFF has been tested with a wide variety of compilers and MPI-2 Corvus [opcons] specic requirements The crvmpmksetand crvmprunsettools are also implemented in Python 3 and, in addition to the they need the pymatgen package.[39] In or- der to use the MPRester method in pymatgen to access the Materials Project database directly, users need to have an MP account and a valid API key. For the work pre- sented in this paper, we used pymatgen v2022.0.9, and Performance Analysis The CPU time for a given material scales linearly with the number of edges and unique atom positions in a unit cell and, depending on the SCF and FMS settings, typi- cally takes between a few minutes and 1 =2 hr. Thus, when the available number of processors is equal or greater than the total number of edges, linear parallelization is simply achieved by assigning a processor or more to each edge, up to about 32-64 processors. More processors per edge becomes less ecient at the level of FEFF calculations. In other cases, and for high-throughput runs for many dif- ferent materials, the wall-clock time depends on how the set is structured (e.g., how many edges per material in the set) and how the calculations are distributed in the work ow. In that case, there are no simple rules to deter- mine the optimal parallel distribution. To help with this issue, the crvmprunsettool includes an option ( --t)to estimate how well a given system of interest will paral- lelize for a certain range of number of processors ( --np ) relative to the fully serial calculation. 10 100 Energy Loss540 560 580 Energy (eV)051015Energy Loss 11: Loss function for several Al 2O3oxides listed in the MP. 4. Example Optical constants of the elements As an illustration, we have carried out calculations for the most stable structure (at room temperature and nom- inally 1 atm) of each of the elemental solids listed in the MP with data imported for a given MP-ID. Some of the most important factors limiting the quality of the re- sults are the SCF and FMS radii used in the simulations. The results can be obtained online at [33] Fig. 9 (top) shows a comparison of the results using an increasing num- ber of near-neighbor shells for 2of copper. We nd that 3 and 6 shells in the SCF and FMS modules, converged results above 10-20 eV. Overall, the results are in reasonable agreement tabulations,[26] and illustrate that, except close to an edge or in the optical region, 2can be ap- proximated reasonably by the atomic approximation, ig- noring the ne structure. For example, Fig. 9 shows a comparison to experiment[26] of Cu 2(top) and real in- dex of refraction Au (bottom) computed with converged SCF and FMS radii and including the density of in Eq. 10. These results show that Eq. 10 is only in the visible range. Fig. 10 shows a periodic table of the elements and associated MP- IDs that are currently available and can be freely viewed and downloaded at .[33] 4.2. Optical constants of the Al 2O3materials in the MP In addition to the elemental solids, we have also com- puted several aluminum oxides. For example, Fig. 3 shows 9a comparison to experiment[26] of the energy loss for the full energy range for the corundum form of Al 2O3. In order to demonstrate the set capabilities of the Corvus [opcons] work ow, we calculated results for a sub-set of those listed in the MP with formula Al 2O3. The results for the loss function up to the Al L-edges are shown in the main panel of Fig. 11, while the inset highlights the O K- edge region. The dierences in the near edge structure, as well as those in the optical re ect the dierent in the system. For example, there is a mixture of Al sites in octahedral and tetrahedral symmetry in the rst two materials, which are both monoclinic, while the last material is trigonal, and contains only Al in octahe- dral symmetry. Similarly, the oxygen sites in the rst two systems are a mixture of sites in tetrahedral and These changes in symmetry lead to dierences in electronic structure, which are directly related to the features seen in the near edge spectra. 5. Conclusions We have presented an optical constants ( opcons ) work ow built with the Python-based Corvus work ow manager for high-throughput simulations of full spectrum optical constants from the UV-VIS to hard x-ray wave- lengths for materials dened in the MP data-base. The calculations are based on the ecient RSGF code FEFF10 and an approximate treatment of the valence structural data from the MP database specied by a unique MP materials ID. As for the XAS, the results can be represented in terms of an atomic-like background and an optical ne-structure (OFS) analogous to EXAFS. Except close to an edge where the ne-structure is sub- stantial, the background dominates and can be over the full spectrum. The ne-structure is edge-specic and as in EXAFS, re ects the structure of the environment of a given absorber in the material. The calculations are set up automatically using Corvus auxil- iary tools, and the work ow parallelizes the FEFF compu- tations over edges and sites with an optimal distribution of computational resources. Representative results show that the work ow produces accurate optical properties over a broad range of energies. Although this algorithm paths and Debye-Waller factors, multiplet split- ting, excitonic eects, and long-range contributions to the valence spectra are currently neglected. Consequently, the present results may only be in the UV- VIS regime. 6. Data availability We have computed the optical properties for all el- emental solids in the periodic table for which struc- tures are available in the Materials Project database, using the approach in this work. The data, in- cluding input les and structure, are freely avail- able online at . Complete installation and test- ing instructions as well as examples for the Corvus [opcons] work ow, its associated tools, and the FEFF10 code are available at thank C. Draxl, S. Dwaraknath, L. Hung, L. Rein- ing, and J. Vinson for comments and suggestions. This work was supported primarily by the US Department of Energy, Oce of Basic Energy Sciences, Division of Ma- terials Sciences and Engineering, under Contract No. specically through the Theory Insti- tute for Materials and Energies Spectroscopies (TIMES) program (JJR, JJK, FDV, and CDP), with from NERSC, a DOE Oce of Science User Facil- ity, under Contract No. R. X. Yang and K. A. Persson acknowledge support by the US De- partment of Energy, Oce of Science, Oce of Basic En- ergy Sciences, Materials Sciences and Engineering Divi- sion under Contract No. program E. D. Palik, Handbook of Optical Constants of Solids, Academic Press, Orlando, 1985. [2] S. Adachi, Handbook On Optical Constants Of Metals, The: In Tables And Figures, World Scientic, 2012. [3] B. L. Henke, E. M. Gullikson, J. C. Davis, X-ray interac- tions: photoabsorption, scattering, transmission, and re ection at E=50{30000 eV, Z= 1{92, Atom. Data Nucl. Data Tables 54 (1993) 181. [4] C. T. Chantler, Detailed tabulation of atomic form factors, pho- toelectric absorption and scattering cross section, and mass at- tenuation coecients in the vicinity of absorption edges in the soft x-ray (Z=30{36, Z=60{89, E=0.1 keV{10 keV), issues of earlier work, J. Phys. Chem. Ref. Data 29 (2000) 597. [5] P. Nozi eres, D. Pines, Electron interaction in solids. Phys. Rev. 109 (3) (1958) 741{761. [6] S. L. Adler, Quantum theory of the dielectric constant in real solids, Phys. Rev. 126 (2) (1962) 413{420. [7] N. Wiser, Dielectric constant with local eld eects included, Phys. Rev. 129 (1) (1963) 62{69. [8] G. Onida, L. Reining, A. Rubio, Electronic excitations: versus many-body green's-function approaches, Rev. Mod. Phys. 74 (2) (2002) 601{659. [9] R. M. Martin, L. Reining, D. M. Ceperley, Interacting University Press, 2016. [10] H. M. Lawler, J. J. Rehr, F. Vila, S. D. Dalosto, E. L. Shirley, Z. H. Levine, Optical to UV spectra and birefringence of SiO 2 and TiO 2: First-principles calculations with excitonic eects, Phys. Rev. B 78 (20) (2008) 205108. [11] J. Vinson, J. J. Rehr, J. J. Kas, E. L. Shirley, calculations of core excitation spectra, Phys. Rev. B 83 (2011) 115106. [12] The exc code, . [13] A. Gulans, S. Kontur, C. Meisenbichler, D. Nabok, P. Pavone, S. Rigamonti, S. Sagmeister, U. Werner, C. Draxl, excit- ing: a full-potential all-electron package implementing theory and many-body perturbation theory, J. Phys. Condens. Matter 26 (2014) 363202. 10

2108.10981

Joshua Kas

J. J. Kas, F. D. Vila, C. D. Pemmaraju, M. P. Prange, K. A. Persson, R. X. Yang, J. J. Rehr

Full spectrum optical constant interface to the Materials Project

cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

Optical constants characterize the interaction of materials with light and are important properties in material design. Here we present a workflow for simulations of full spectrum optical constants from the UV-VIS to hard x-ray wavelengths based on the real-space Green's function code FEFF10 and structural data from the Materials Project (MP). The Corvus workflow manager and its associated tools provide an interface to FEFF10 and the MP database. The workflow parallelizes the FEFF computations of optical constants over all absorption edges for each material in the MP database specified by a unique MP-ID. The workflow tools determine the distribution of needed for that case. Similarly, the optical constants for selected sets of materials can be computed in a single-shot. To illustrate the approach, we present results for nearly all elemental solids in the periodic table, as well as a sample compound, and compared with experimental results. As in x-ray absorption spectra, these results are interpreted in terms of an and fine-structure contributions.

2 Melika Golestani et al. the verb of a sentence along with its constellation of arguments, such as subject and object [10]. In a coherent narrative text, events follow a logical order, which makes it possible for a reader to make sense of the text. In NLP tasks, Sentence Ordering models are suggested to learn high-level structure that causes sentences to appear in a specific order in human-authored texts and by learning to order sentences we can model text ordering refers to organizing a given shuffled set of sentences into a co- herent order [25]. Sentence ordering assists in modeling coherence in text, which in turn can be applied in a variety of tasks such as multi-document summarization [2,21], [31], and retrieval based QA systems [25]. A key factor to to solve the problem of sentence ordering is to find a coherent sequence of events appeared in the input sentences. In natural language processing, events are often meant to be the verbs of sentences and their dependencies, such as the subject and the object [29]. A number of studies have proposed methods to order events based on their temporal and causal relations [31], or the semantic roles [9]. The researchers developed a Skip-thought model utilizing an RNN. It takes a sentence and encodes it into vector, followed by previous and subsequent sentences. Therefore, all input is first encrypted using a hidden network vector; then its output is decrypted from that vector. Pichotta & Mooney [34] replaced the RNN with an LSTM, and proposed a sentence-level language model to predict the next sentence for the script inference skip-thought's architecture's simplicity results in a lower quality of creat- ing sentences embeddings compared to many newer methods and networks like BERT [12]. The system uses an RNN-based encoder-decoder architecture, whereas SBERT- WK [39] uses a architecture and an attention mechanism to encode the sentences into to [9], a statistical script system model an event sequence as a proba- bilistic model and infer additional events from a document events. This system teaches the linguistic model of events; despite this, some critical information in the sentence can be ignored because only verbs and their dependents are used. Thus, the clue words (adverbs such as then, before, and after) are omitted. In this paper, we propose a sentence-level language model to solve the problem of sentence ordering. Language models have been popular for years at the word and letter levels [41,19]. LSTMs [16] and RNNs [26] have been vastly used for teaching lan- guage models at the word and letter levels is beneficial [28,17,17,38,5]. These models take a single word or character as input at time t, update the hidden mode vector, and predict the next word or character at time t+1. Kiros used a sentence level language model based on the RNN [18]. Pichotta in [34] used LSTM-based sentence-level mod- els for script inference due to the problem of RNNs, the vanishing gradients. RNNs and LSTMs encrypt information about one sentence and decode information about the following sentence. Despite this, these networks not use a self-attention encoding sentences, and some information is lost. We use a sentence level language model based on the universal transformers (UT) [11]. A generalization of the Transformer model, the UT model is a parallel-in-time re- current sequence model. The UT model combines two aspects of feed-forward and global receptive field. An additional feature is the position-based halt-Using BERT Encoding and Sentence-Level Language Model for Sentence Ordering 3 ing mechanism. Using a self-attention mechanism, the Universal Transformer refines its representations for all sentences in the sequence in parallel. Our model involves three components: the first component encodes sentences; the second one uses a neural network to teach the language model at the sentence level by universal transformers (UT) and to predict the next sentence. The last component uses the cosine similarity and Brute Force search to order sentences. We aim to design a model that can capture the sequence of events by arranging a set of sentences. The model we propose is designed to organize shuffled sentences of short stories. By using this method, we encode sentences by the SBERT-WK model, which causes us to pay attention to every word within a sentence equal to its importance and transform the sentence into a vector. Thus the embedding contains the clue words information. To capture the relationship between Sentences, We use a UT-based language model, which predicts the next sentence of the current sentence. Finally, we use a Brute Force search, using the scoring function based on cosine similarity, which is scored between the candidate sentence and other sentences, to maximize the sum of similarities between two consecutive sentences. As dataset, we use ROCStories [30], which includes around 100K short 5-sentence stories, written by human turkers. The gold data follows the order similar to the original order in the stories. The main applications of sentence ordering can be mentioned as: extractive text summarization [2], retrieval based QA systems [25], text coherence modeling [25], discourse coherence [13], and text generation sys- tems [7]. 2 Related relations between sentences has become increasingly important for vari- ous NLP tasks, such as multi-document summarization [2], text generation [7], and text coherence modeling [25]. Moreover, learning the order of sentences can help in model- ing text coherence [25]. In sentence ordering, the goal is to arrange a set of in a cohesive order. Ordering models aim to identify patterns resulting sen- tences to appear in a specific order in a coherent text. Several previous studies have addressed the task of sentence ordering for a set of data such as news articles methods suggested on the task of sentence ordering fall into two cat- egories: traditional approaches, and deep learning-based approaches. Traditional ap- proaches to coherence modeling and sentence ordering often apply probabilistic models [21]. Barzilay utilized content models to represent topics as HMM states [4], then em- ployed hand-crafted linguistic features, including Entity Grid, to model the [3]. [30] suggested using n-gram overlapping to pick a final sentence from two human-made options of the ROCStories stories. In [35], a similar method is used to arrange the set of unordered sentences of the stories. As the next sentence, they pick the sentence with the most n-gram overlaps with the current sentence. They measured an overlap up to 4-gram using Smoothed-BLEU [24]. In recent years, researchers used neural approaches to solve sentence ordering tasks. [1] proposed SkipThought + Pairwise model. The method involves combining two points identified from the unary embedding of sentences without considering context4 Melika Golestani et al. and a pairwise model of sentences based on their relative composition in context. How- ever, as SkipThought model [18] was applied to map sentences into vector space, some information is missing. Another method proposed in [23] is a productive model that is called Seq2Seq + Pairwise. A Seq2Seq model based on an encoder-decoder architec- ture is used to predict the following sentence, having one. This way, the model learns a probabilities distributions over sentences. Gong employed an end-to-end sentence ordering method, LSTM + PtrNet[14]. Af- ter getting encrypted and decoded by LSTM models, pointer networks are being used to arrange sentences. Another method proposed in [25], called LSTM + Set2Seq, uses LSTM and attention mechanism to encode sentences and learn a representation of con- text. Then a pointer network is applied to ranks sentences. [40] proposed Hierarchi- cal attention networks (HAN). The idea is to capture the information of clue words to learn the dependency between sentences. The system uses a word encoder with a BiLSTM architecture, an attention layer that creates sentence embedding, and an atten- tion mechanism to encode and decode a group of sentences together. Although LSTM and BiLSTM-based models show improvement, they are less efficient comparing to a BERT-based encoder model. This is mostly because of the fact that BiLSTM models do not look at both directions of a word simultaneously. In contrast, BERT is trained to learn both left and right positions at the same time. [27] proposed a CBoW model that generate sentence embedding using the average of the word embedding vectors of the words that compose the sentence. This embedding is used in [35] to suggest a baseline, where cosine similarity between embeddings of pairs of sentences is used to order the pairs. This has two essential defects: it does not include many sentence information such as clue words and there is no way for modelling sentences' presents sentence correlation measure (SCM) for sentence ordering. This mea- sure has three main components: 1. A sentence encoding component based on the pretrained SBERT-WK [39]; 2. A scoring component based on the cosine similarity; 3. A ranking component based on a Brute Force search. None of the above components need access to training data which provides an advan- tage where limited data is available[35]. The method does not work as well as HAN or LSTM + Set2Seq when enough training data is available, since dependencies are not captured. [20] proposed RankTxNet ListMLE, which is a pointer-based model. This method relies on a pre-trained BERT model for encoding sentences and a self-attention network for encoding paragraphs. To predict a relative score for each sen- tence, they use a feed-forward network as a decoder which determines each in a paragraph. Our proposed model has some similarity with the methods presented in [14] and [25] in using neural networks for for predicting sentence ordering in a pairwise man- ner. However, we employ a BERT-based model for sentence encoding, in contrast to the LSTM-based model they used. Moreover, by applying universal transformers we encode a group of sentences and learn a language model.Using BERT Encoding and Sentence-Level Language Model for Sentence Ordering 5 3 Methodology 3.1 Task Formulation The sentence ordering takes the story S as input, which its sentences are probably n (1) Where n is the length of the sequence or the number of sentences in each sequence. It outputs a permutation of the sentences like o so that o is equal to (o 5); (2) whereois to the order of the sentences in the gold data. 3.2 Sentence-Level Language Model for Sentence Ordering (SLM) We proposed Sentence-level Language Model (SLM) for Sentence Ordering compose of Sentence Encoder, Story Encoder, and Sentence Organizer. The architecture of the proposed model is shown in Figure1. Sentence encoder gets sentence as input and en- coded it into a vector using a fine-tuned pre-trained SBERT-WK[39]. The embedding pays more attention to the sentence's crucial parts, such as the verbs and clue stop words are omitted. Then, Story Encoder takes the sentence encoder as an input and learns the sentence level language model using an encoder-decoder ar- chitecture based on universal transformers (UT) [11]. UT couples parallelism with the global receptive field in a feed-forward model. The UT model also includes position- based halting [11]. The UT-based encoder-decoder component corrects vector for each position (sentence) in parallel by combining information from using self-attention and applying a repetitive transfer function [11]. Hence de- pendencies between the sentences are captured. The vector is learned from the hidden state and then decodes the vector and indicates the next sentence's candidate. What matters at runtime is that during training, information are transferred to all nodes at time t+ 1from all nodes at time t. Thus, parallelism is created, which allows all tokens processed at the same time [11]. In the Sentence Organizer, we can rate each sentence and the candidate based on their similarity. We use the cosine similarity as the scoring function. The arrangement is made using a Brute Force search in all state spaces to maximize the sum of the between two consecutive sentences like the sentence ordering component in [35], as = the cosine similarity between s0 iandsj, ands0 iis the can- didate for the next sentence of the i-th sentence, represented with sentence in the order of output, and siis thei-th sentence in the of sentences or the input's order.6 Melika Golestani et al. Fig. 1. Showing the SLM in the abstract. During model training, sentence ordering is given to the network correctly to learn and model the relationships and dependencies between the sen- tences. During testing, input sentences are shuffled. Where wjiis the jth word in the ith the ith sentence in output order. The goal is to output an s(o) equal to the gold data ordering). Our model is trained again using the universal sentence encoder (USE) embeddings [8] to compare them to SBERT-WK embeddings, and additionally, we employ the BiL- STM [36] network to learn the LM. Also, we apply the nearest neighbor search, which is a greedy algorithm. The NN search calculations for sentence ordering are presented in [35]. Consequently, we train, evaluate, and test nine different models, including pretrained SBERT-WK + UT + BFS, -Pretrained SBERT-WK + UT + BFS, -Pretrained SBERT-WK + BiLSTM + BFS, -Pretrained SBERT-WK + UT + NN, -Pretrained SBERT-WK + BiLSTM + NN, -USE + UT + BFS, -USE + BiLSTM + BFS, -USE + UT + NN, -USE + BiLSTM + NN, So we can assess the effect of change on each component.Using BERT Encoding and Sentence-Level Language Model for Sentence Ordering 7 4 Experiment 4.1 Dataset We used ROCStories dataset (a commonsense story dataset) [30]. The dataset contains 98,162 five-sentence stories with an average word count of 50 words. 3,742 of stories have two choices as the fifth sentence or the final sentence. The corpus has been pre- sented for a shared task called LSDSem [32] where models are supposed to choose the correct ending to a four-sentence story. All the stories and options are generated by human. This dataset has some essential characteristics that make it a fit for our task of learning sequences of events: ROCStories contains causal and temporal daily events. This makes it possible to learn the narrative structure of a wide range of events. The dataset contains a collection of daily non-fictional short stories suitable for the training of coherent text models. 4.2 Baselines and competitors The proposed method is compared with five baselines as follows: Sentence n-gram overlap [35], SkipThought + Pairwise [1], Seq2Seq + Pairwise [23], Continues Bag of Words (CBoW) [35], Sentence Correlation Measure (SCM) [35]. The follow as: Hierarchical Attention Networks (HAN) [40], LSTM+PtrNet [25], and RankTxNet ListMLE [20]. 4.3 Metrics The metrics we use to evaluate story ordering outputs are Kendal's Tau and below. 1. Kendall's Tau ( tau) Kendal's Tau measures the quality of arrangements by Equation (4), = 1((2
number of inversions ))=(N
(N1)=2) the sequence length, and numberof inversions is equal to how many binaries' relative order is wrongly predicted. This measure for is correlated with human judgment, according to [22]. The value of this criterion is in the range [1;1]; the lower limit indicates the worst case, and= 1 when the predicted order equals to the order in the gold data are the same. 2. Perfect Match Ratio (PMR) PMR defined as following, calculates the ratio of exactly matching sentence orders without penalizing incorrect ones. PMR = ((# of Correct Pairs ))=(N
(N1)=2) the sequence length. One can see that the PMR is always in the range of 0 and 1, where PMR = 1 indicates the predicted order is exactly the same as the gold order, and a PMR = 0 means that the predicted order is precisely the contrary of the gold order.8 Melika Golestani et al. 4.4 Results and Analysis As mentioned in section 3, SLM has three components. First, a sentence encoder, which is designed based on SBERT-WK. Second, a story encoder, which trains a sentence- level language model and learns the dependencies between the sentences. In this com- ponent, the UT network is used. The third component is sentences' organizer which calculates the cosine similarity between sentence embedding vectors and employs a Brute Force search to maximize the sum of similarities among consecutive sentences. In addition to SLM, we trained and tested nine other models by replacing each com- ponent with other algorithms or architectures to select the best algorithm and compare SLM with them. These nine models were constructed by replacing SBERT-WK encod- ing with USE vectors, the UT network with a BiLSTM one, and the BFS with a nearest neighbor search. The parameters of the SLM are given in Table 1. The BERT vectors have 768 di- mensions, while the USE ones have 512 dimensions. Following previous works [25,40], we randomly split the dataset into training (80%), test (10%), and validation (10%) sets, where they contain 392645, 49080, and 49080 sentences 1. Parameters of the SLM. SLM encodes sentences with SBERT-WK embeddings and learns a language model with the UT network. Also, the USE embeddings with d =512 instead of the SBERT-WK ones are used to allow for a better choice. So hidden layer size of that is 4 512. Method Components Tau Initial Learning Rate ( ) ( ) 105 size of Embedding vectors (d) 768 Hidden Layer size (h) 4768 Based on table 2, SBERT-WK outperforms USE due to its better architecture for sentence encoding. In all cases, UT is superior to BiLSTM, which was expected as UT has an attention mechanism. Due to NN's greedy nature, NN may not always find the global optimal solution and may become stuck in the local optimal, whereas BFS never will. So BFS is better than NN. That is why the best results are happening where we use fine-tuned SBERT-WK embedding, UT to capture dependencies between similarity as a scoring function, and searching the entire state space to maximize the total similarity of consecutive 2 shows the results of the proposed model, SLM, along with the baselines and competitors. SLM has improved by about 3.2% and 4.2% compared to LSTM + PtrNet, and more than 4.3% and 4.8% compared to LSTM + Set2seq in and This performance improvement, is due to the use of a SBERT-WK model sentence encoders and employing an attention mechanism at the story encoder utiliz- ing universal transformers. Taking advantage of the mentioned components, SLM can capture the intradependecy and of sentences very well. of a sentence refers to the relations among each sentence's words, while refers to each sentence's relations with other sentences of the story. OurUsing BERT Encoding and Sentence-Level Language Model for Sentence Ordering 9 Table 2. SLM results by changing each Components Tau PMR SLM (Fine tuned SBERT-WK + UT + Brute Force Search) 0.7547 0.4064 SBERT-WK + UT + Brute Force Search 0.7465 0.3893 USE + UT + Brute Force Search 0.7206 0.373 SBERT-WK + BiLSTM + Brute Force Search 0.7317 0.3762 USE + BiLSTM + Brute Force Search 0.7044 0.3545 SBERT-WK + UT + Nearest Neighbor Search 0.64 0.2755 USE + UT + Nearest Neighbor Search 0.6162 0.267 SBERT-WK + BiLSTM + Nearest Neighbor Search 0.6214 0.2576 USE + BiLSTM + Nearest Neighbor Search 0.5980 0.2511 method is also superior to the HAN network, increasing criterion by more than 2.2 percent and PMR criterion by 1 percent. This could be because of the fact that the sen- tences are encoded using the BERT-based model in the SLM. HAN, however, uses a BiLSM encoder and multi-head attention layer. Besides, SLM performed about 2.6 percent better than RankTxNet in PMR. How- ever, the Kendall's score gives RankTxNet a slight superiority. The difference is not significant, and can happen as a result of random parameters of the network [37,15]. Fig. 2. Kendall's tau ( ) and perfect match ratio (PMR) on test set for ROCStories datasets. The results of the model, compared to the baselines and competitors.10 Melika Golestani et al. 5 Conclusion We presented a framework to solve sentence ordering task. A sequence of shuffled sentences is inputted to our framework, where the output should be a coherent order of the given set. We achieved state-of-the-art performance in PMR scores on ROCStories dataset. We learned that SBERT-WK is a suitable choice for sen- tence encoding. We analyzed how the method is changed by using the USE to We also found that Universal Transformers perform better for story encod- ing and learning the sentence-level language model comparing to BiLSTM we compared Brute Force Search with Nearest Neighbor Search to order sentences. As future work, we plan to develop more robust models to fulfil ordering of instances with longer input Agrawal, H., Chandrasekaran, A., Batra, D., Parikh, D., Bansal, M.: Sort story: Sorting jumbled images and captions into stories. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. pp. 925-931. Association for Com- putational Linguistics, Austin, Texas (Nov 2016). Barzilay, R., Elhadad, N.: Inferring strategies for sentence ordering in multidocument Journal of Artificial Intelligence Research 17, 35-55 (2002) 3. Barzilay, R., Lapata, M.: Modeling local coherence: An entity-based approach. Computa- tional Linguistics 34(1), 1-34 (2008) 4. Barzilay, R., Lee, L.: Catching the drift: Probabilistic content models, with applications to generation and summarization. In: Proceedings of the Human Language Technology Con- ference of the North American Chapter of the Association for Computational 2004. pp. 113-120. Association for Computational Linguistics, Boston, Mas- sachusetts, USA (May 2 - May 7 2004), Bengio, Y ., Ducharme, R., Vincent, P., Janvin, C.: A neural probabilistic language model. The journal of machine learning research 3, 1137-1155 (2003) 6. Bollegala, D., Okazaki, N., Ishizuka, M.: A bottom-up approach to sentence ordering for multi-document summarization. Information processing & management 46(1), 89-109 (2010) 7. Bosselut, A., Celikyilmaz, A., He, X., Gao, J., Huang, P.S., Choi, Y .: rewards for coherent text generation. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, V olume 1 (Long Papers). pp. 173-184. Association for Compu- tational Linguistics, New Orleans, Louisiana (Jun 2018). Cer, D., Yang, Y ., Kong, S.y., Hua, N., Limtiaco, N., St. John, R., Constant, N., M., Yuan, S., Tar, C., Strope, B., Kurzweil, R.: Universal sentence encoder for English. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations. pp. 169-174. Association for Brussels, Belgium (Nov 2018). BERT Encoding and Sentence-Level Language Model for Sentence Ordering 11 9. Chambers, N., Jurafsky, D.: Unsupervised learning of narrative event chains. In: Proceedings of ACL-08: HLT. pp. 789-797 (2008) 10. Chambers, N., Jurafsky, D.: Unsupervised learning of narrative schemas and their partici- pants. In: Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. pp. 602-610 (2009) 11. Dehghani, M., Gouws, S., Vinyals, O., Uszkoreit, J., Kaiser, L.: Universal preprint arXiv:1807.03819 (2018) 12. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, V olume 1 (Long and Short Pa- pers). pp. 4171-4186. Association for Computational Linguistics, Minneapolis, Min- nesota (Jun 2019). Elsner, M., Austerweil, J., Charniak, E.: A unified local and global model for discourse co- herence. In: Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Confer- ence. pp. 436-443 (2007) 14. Gong, J., Chen, X., Qiu, X., Huang, X.: End-to-end neural sentence ordering using pointer network. arXiv preprint arXiv:1611.04953 (2016) 15. Haddad, H., Fadaei, H., Faili, H.: Handling oov words in nmt using unsupervised In: 2018 9th International Symposium on (IST). pp. 569- 574. IEEE (2018) 16. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735- 1780 (1997) 17. Jozefowicz, R., Vinyals, O., Schuster, M., Shazeer, N., Wu, Y .: Exploring the limits of lan- guage modeling. arXiv preprint arXiv:1602.02410 (2016) 18. Kiros, R., Zhu, Y ., Salakhutdinov, R.R., Zemel, R., Urtasun, R., Torralba, A., Fidler, S.: Skip-thought vectors. In: Advances in neural information processing systems. pp. 3294- 3302 (2015) 19. Kozielski, M., Rybach, D., Hahn, S., Schl uter, R., Ney, H.: Open vocabulary using combined word-level and character-level language models. In: 2013 Conference on Acoustics, Speech and Signal Processing. pp. 8257-8261. IEEE (2013) 20. Kumar, P., Brahma, D., Karnick, H., Rai, P.: Deep attentive ranking networks for learning to order sentences. In: AAAI. pp. 8115-8122 (2020) 21. Lapata, M.: Probabilistic text structuring: Experiments with sentence ordering. In: Proceed- ings of the 41st Annual Meeting of the Association for Computational Linguistics. pp. 545- 552 (2003) 22. Lapata, M.: Automatic evaluation of information ordering: Kendall's tau. 32(4), 471-484 (2006) 23. Li, J., Jurafsky, D.: Neural net models of open-domain discourse coherence. In: Pro- ceedings of the 2017 Conference on Empirical Methods in Natural Language Pro- cessing. pp. 198-209. Association for Computational Linguistics, Copenhagen, Den- mark (Sep 2017). Lin, C.Y ., Och, F.J.: Automatic evaluation of machine translation quality using longest com- mon subsequence and skip-bigram statistics. In: Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04). pp. 605-612 (2004)

2108.10986

Melika Golestani

Melika Golestani, Seyedeh Zahra Razavi, Zeinab Borhanifard, Farnaz Tahmasebian, and Hesham Faili

Using BERT Encoding and Sentence-Level Language Model for Sentence Ordering

12 pages, 2 figures, The 24th International Conference of Text, Speech and Dialogue (TSD2021)

cs.CL

http://creativecommons.org/licenses/by/4.0/

Discovering the logical sequence of events is one of the cornerstones in Natural Language Understanding. One approach to learn the sequence of events is to study the order of sentences in a coherent text. Sentence ordering can be applied in various tasks such as retrieval-based Question Answering, storytelling, text generation, and dialogue we can learn to model text coherence by learning how to order a set of shuffled sentences. Previous research has relied on RNN, LSTM, and BiLSTM architecture for learning text language models. However, these networks have performed poorly due to the lack of attention mechanisms. We propose an algorithm for sentence ordering in a corpus of short stories. Our proposed method uses a language model based on Universal Transformers (UT) that dependencies by employing an attention mechanism. Our method improves the previous state-of-the-art in terms of Perfect Match Ratio (PMR) score in the ROCStories dataset, a corpus of nearly 100K short The proposed model includes three components: Sentence Model, and Sentence Arrangement with Brute Force Search. The first component generates sentence embeddings using SBERT-WK pre-trained model fine-tuned on the ROCStories data. Then a Universal Transformer a sentence-level language model. For decoding, the network generates a candidate sentence as the following sentence of the current sentence. We use cosine similarity as a scoring function to assign scores to the and the embeddings of other sentences in the shuffled set. Then a Brute Force Search is employed to maximize the sum of similarities between pairs of consecutive sentences.

2 deterministic model , where all events in the cell cycle are precisely timed. In this model, we will represent the fundamental state of the cell as an age and compute the statistics of cell age in an exponential culture. To make contact with observables, we then apply these results to describe the demographics of the E. coli cell cycle in Sec. II B. In Sec. II C, we consider a stochastic model , where the cell cycle is represented as states j= 1:::m, each with a Although this model cannot fully capture all the complexities of the cell cycle, it is and we can exactly compute expressions for all the same statistics as the deterministic model. The relation between the deterministic and stochastic model statistics is at this point opaque. In Sec. II D, we mean , which is a mean biased toward younger cells that are overabundant in In Sec. II E, we demonstrate that the predictions of the stochastic and deterministic models are in fact identical if the deterministic state ages jare equal to the exponential-mean stochastic state ages j. Finally, in Sec. II E, we consider a number of simple to underline both the mathematical behavior of the exponential mean as well as its In the interest of brevity, we will support for this model elsewhere [13]. II. RESULTS In this section, we will derive the expressions for a large number of statistics relevant for describing an exponential culture. We will first derive expressions for the statistics in the deterministic model and then the stochastic model. A. Deterministic model In the deterministic model, we will consider cells that are born with age = 0 and divide at age=T. By cell age, we mean a continuous cell state variable representing cell cycle progression, not aging in the context of reduced cell fitness over time [14]. 1. Definition of the deterministic model In the deterministic model, cell state is described by a continuous variable, cell age , and therefore the population is described in terms of a number density with respect to age at timet:n(t). Ageis defined on the interval [0;T]with= 0 corresponding to cell birth to cell division. Let the cumulative creation number, N+ (t), be the of cells that have entered state ( i.e.age) and the cumulative annihilation number, N (t), be the cumulative number of cells that have transitioned out ter zT Cell age: L3L/2Distance from old pole: x Creation of 'ori' and 'rep' ori 0 ter z z T- T ori+ T Cell age: from old pole: x Cell of of 'z' Creation of 'z'FIG. 1. Panel A: Schematic for positioning and timing of events during the E. coli cell cycle. In the the duration of the cell cycle as well as the timing of all events in the cell cycle are precise ( we shall assume the positioning of all complexes and the length of the cell are all deterministic as well. The shaded region represents one complete cell cycle. We have also annotated the positioning and timing of a number of cell cycle events: (i) Cell poles appear as the consequence of septation and do not disappear. (ii) Replication of the when the origin ( oriC) is replicated and ends when the terminus is replicated. The origin is replicated before the start of the cell cycle. (iii) New origins are created and move from the quarter-cell positions to the eighth-cell positions (iv) The Z ring, which drives septation at and disassembles at the end of the cell cycle. Panel B: Creation and annihilation. The statistics of different types of quantities will have different statistical quantities undergo both creation and The Z ring and replisome are natural examples of transients since they assemble and then disassemble. In contrast, both cell poles and genetic loci are perpetual quantities that only undergo creation but never annihilation. of state[15]. In the deterministic model, where the cell stateis continuous (not discrete), the and annihilation numbers are identical and are related to the number density by the =N (t) =Zt 0dt1n(t1): (2) It is convenient to define these cumulative numbers in addition to the number density, since they will be a powerful tool for computing some To describe the dynamics, we can write an equation describing the number of cells entering age interval [;+]in the interval [t;t+t]: t _n(t) (3) where _A@tAand the first term on the RHS is the number of cells entering state and the second term represents the cells leaving state +. Eq. 3 can then be =@_N+ (t); (4) except at division, where some care is required. Now consider the process of cell division explicitly: The division process can be understood as the annihilation of a cell in state =Tand the creation of two new-born cells in state = 0: _N+ (t) =( 2nT(t);  = 0 n(t);  > 0(5) _N (t) =n(t): (6) Substituting Eq. 5 into Eq. 4 gives a single piecewise rate equation in terms of the number density n(t): _n(t) =( 2n0 T(t);  = 0 n0 (t);  > 0; (7) whereA0@A. Eq. 7 completely describes the cell cycle dynamics in the deterministic model. The details of the derivation are given in Appendix A 1. 2. Solution to the deterministic model In steady-state growth, we can assume the total number of cells is: N(t) =N0exp(kt); (8) wherekis the growth rate that is determined by solving the rate equation (Eq. 7), as detailed in Appendix A 2. It will often be convenient to rewrite the equations in terms of the doubling time: Tk1ln 2; (9) rather than the growth rate k. Eq. 7 evaluated at = 0 gives a consistency condition between doubling time T and the duration of the cell cycle T: T=T; (10) which is to say that the doubling time is equal to the duration of the cell cycle, as one would na ively expect. In steady-state growth, one can compute the number density of cells, which is: n(t) =n0exp[k(t)]; (11) wherenis the density with respect to cell age and n0is a constant determined by the initial cell number. The details of the derivations for the solution and the consistency condition are given in Appendix A 2.3. Statistics of the deterministic model The solution of Eq. 7 can be used to compute the probability (PDF) and cumulative (CDF) with respect to cell age: f() = 2kek; (12) F() = 2(1ek): (13) The details of the derivation are given in Appendix A 3. Eq. 12 implies that in an exponential culture, there is an enrichment of young cells which decays age. See Fig. 2A. Note that the canonical observable in an is number as a function of time rather than abundances relative to the total number of cells N(t). However, we shall write each expression as the therefore the prefactor can be interpreted as the abundance relative to cell number N(t). The cumulative creation number is: N+ (t) = 2ekN(t): (14) The details of the derivation are given in Appendix A 4. The number of cells younger than age is: N<(t) = 2(1ek)N(t); (15) and the number of cells older than age is: N>(t) = (2ek1)N(t): (16) Finally, the number of cells in a state defined by the age range1< < 2is: N[1;2](t) (17) where the two terms on the right hand side are defined in Eq. 16. B. Application to cell cycle dynamics In this section, we demonstrate how to apply these results in the context of the E. coli cell cycle dynamics shown schematically in Fig. 1A. These formulae can be applied either to predict the numbers from the timing or to infer timing from the observed numbers in an exponential culture. 1. Z ring The Z ring is an ultra-structural complex responsible for the process of bacterial cytokinesis (or septation) in which the cell envelope contracts at midcell forming a septum that closes to form the new poles of the nascent daughter cells [16]. The assembly, dynamics and disassembly of this structure is easily T 2T 3T Time: ring0 0.2 0.4 0.6 0.8 1 Cell Age: / Age PDF: fIo7(A) (B) FIG. 2. Panel A: Cell Age PDF. In an exponential culture, there is an enrichment in young cells relative to old cells. The relative number of cells decays exponentially with cell age . Panel B: Numbers in an exponential culture. The numbers of all quantities grow exponentially with the same growth rate. For perpetual quantities ( e.g.cell number, cell poles, DNA loci) the relative timing of the creation of a quantity can be inferred by the temporal offset of the NX(t)curve relative to the cell number curve N(t). In contrast, transient quantities, like the number of Z rings, also grow exponentially, but their offset cannot directly be interpreted as a time. using a wide range of fluorescent fusions in live cells or in fixed cells [17]. The Z ring is an example of a transient we need to use N(as opposed to it assembles at =Zand disassembles at the end of the cell cycle. The number of Z rings is therefore equal to the number of cells older than Z: NZ(t) =N>Z(t) = (2ekZ1)N(t): (18) It is interesting to consider the limit as ZTZ is small relative to the cell cycle duration Tin order to compare this to our intuitive guess (Eq. 2: (19) Since ln 20:69, this is roughly 30% smaller than our na ive estimate due to depletion of older cells in an exponential culture (Fig. 2A).2. Cell poles Although the number of cell poles is twice the number of cells, it is useful to consider this example more formally. Unlike the Z ring which is transient, the poles are perpetual: Once the state is created, it is (Fig. 1B). In this context, we can use the cumulative creation number N+. Note that we are immediately presented with a conundrum: Are two poles formed at the end of the cell cycle ( =T) or is one pole created at birth ( = 0)? Both approaches give the same number: Npole(t) = 2N+ T=N+ 0= 2N(t); (20) which is twice the number of cells, just as one DNA loci The numbers of DNA loci can be observed by a number of different approaches: Modern deep sequencing methods allow a replication profile ( i.e.the DNA copy number) of all loci to be measured in a single experiment ( e.g.[8]). However, of the relative copy numbers of loci long predate this modern approach [2]. The of loci can also be observed: such as Fluorescence In Situ Fluorescent Repressor Operator Systems (or closely related approaches), can be used to visualize the numbers of segregated loci in single cells [18, 19]. There are multiple equivalent approaches to computing the numbers of genetic locus `. First consider the slow-growth limit where both initiation and termination occur within the current cell cycle [2]. Assume the locus of interest is replicated at time `. The number of copies per cell is one before replication and two after replication. We can therefore write: N`(t) = 1N<`(t) + 2N>`(t); (23) in agreement with previous results [20, 21]. Unlike transient quantities, e.g.the number of Z rings, the form of Eq. 23 implies that the number of genetic loci can be understood as a temporal shift of times, as illustrated in Fig. 2B. The cancelation between the non-exponential Eq. 22 may seem incidental, but from another perspective it is intuitive: The mathematical reason for the non-exponential terms in the prefactors of Eqs. 15-16 is annihilation in the number of cells of a particular age  due to aging. DNA loci correspond to a perpetual state: Once a locus state is created ( i.e.replicated) it does not5 annihilate ( i.e. transition into another state ). To compute the number of genetic loci, we can therefore use the cumulative creation number N+formula (as opposed to Nwhich is reduced by annihilation). This more direct approach yields the same result as Eq. 22: N`(t) =N+ `(t) =ek(`T)N(t); (24) but is applicable for fast growth where before the cell cycle begins ( i.e.`<0), as illustrated in the cell cycle schematic in Fig. 2A. 4. B, C, and D bacterial cell cycle is described by three periods: The B period is defined as the period between birth and replication initiation. The C period is defined as the cell cycle period during which i.e. after replication initiation and [2]. The D period is defined as the period between replication termination and cell division [2]. There are multiple approaches to characterizing of cells by period. A traditional approach is to infer this information from the relative abundance of the origin, terminus and number of cells [2]. However, more recent single-cell approaches can visualize the replication process itself in live cells [22, 23]. The relation between the durations of these periods and the locus number (Eq. 24) are: Bori0 =Tlog22N=N ori; andNterare the number of cells, origins, and termini in the exponential culture (not per cell), which has previously been reported [2, 3]. Note that if replication initiates before the start of the cell let us consider the replisomes and the replication process itself. A traditional to determining the number of replicating cells is to infer it from the relative origin, terminus, and cell abundances. However, many single-cell and live single-cell approaches exist today as well. For fusions to core replisome components that localize during replication can be used to characterize the number of replicating cells [22, 24, 25]. Like the Z ring, replication is a transient state; however, there is a significant subtlety here: Do we count (i) replicating cells, (ii) individual consisting of replisome-pairs, or (iii) individual replisomes?First let us consider the number of the replication process can span the overlap between two successive cell cycles, it is most convenient to use differences in the cumulative creation number N+. In fact, we can express the number concisely in terms of oriC and ter: Nrep(t) =NoriNter; (28) and the number of individual replisomes will be twice the number of pairs. NoriandNterare computed using Eq. 24. For the number of replicating cells, we consider three different cases. First consider a case where the replication cycle is internal to the cell cycle. In this case, we have: Nrep cell (t) oriN> ter (29) which can be evaluated using Eq. 17. If the overlaps by a single cell cycle but do not overlap: Nrep cell (t) =N>T+ori+N< ter; (30) whereoriis negative in this context, as noted in Sec. II B 4. Finally, if the rounds of replication overlap: Nrep cell (t) =N(t); (31) and all cells are replicating in the deterministic model. C. Stochastic model An important complication of a more realistic model for cell cycle dynamics is stochasticity ( the lifetime of the states of the cell cycle. We will incorporate this stochasticity by dividing the cell cycle into mdiscrete states through which the cell must transition sequentially. This model is in Fig. 3. The lifetime of each state j will be described by an arbitrary lifetime PDF, pj(
), for thejth state. It is important to note that stochastic model is not general enough to be an accurate representation of the bacterial cell cycle; however, it is sufficient to explore a number of interesting phenomena and is exactly solvable. 1. Definition of the stochastic model In our analysis, we will use the rate rather than a master equation approach, since we are interested in the steady-state behavior of the model in the large cell number limit where the relative size of the fluctuations are vanishingly the number of cells in state j, the cumulative creation number, N+ j(t), and the states: td1 td1 td2pdj tdm-1 3. Schematic of the stochastic model of the cell cycle. We represent the cell cycle as a series of sequential states j= 1:::m. After the mth state, two daughter cells in state 1 are born. The PDF of state lifetimes, pj, is distinct for each statej. Although most quantities of interest are that the states are created when the cells enter and then annihilated when the cells exit, we also consider that are created but not annihilated ( e.g.DNA number, N j(t), be the total number of cells to have arrived and departed from state jover all The state dynamics is therefore described by the following rate (32) In this model, cells move sequentially through the m states before the final state ( j=m) transitions to the initial state ( j= 1) as two j > 1; 2_N m; j = 1:(33) Each statejhas a PDF of lifetimes pj(t)and therefore the relation between state jarrivals and departures is given (34) where is the (35) Eqs. 32-34 completely specify the stochastic model. 2. Solution to the stochastic model We will work in the steady-state growth limit as before. It is most convenient to work in terms of the Laplace transforms of the rate equations (Eqs. 32-34). The Laplace transform is (36)where the tilde denotes the transformation from time t to Laplace conjugate . The transformed representation is convenient since the ordinary differential equations become in terms of the transform quantities and the convolutions become products of transforms ( e.g.[26]). Of particular importance in what follows is the relation between the PDF of lifetimes of individual state j,pj(t), and the PDF of the age of the cell at the transition out (37) A detailed derivation of the solution to the rate equations (Eqs. 32-34) is given in Appendix A 5. For steady-state exponential growth, the that relates the growth rate kto the PDF of cell cycle durations p(t)pm(t)can be written concisely in terms of the Laplace transform: 1 = 2 ~p(k); (38) an equation that is well known [27, 28]. This is equivalent to Eq. 10 in the although the mathematical equivalence between these two relations is opaque for the moment. 3. The statistics of the stochastic model In an exponential culture, the cell numbers are given by the N+ j(t) = 2~pj1N(t); (39) In statej:Nj(t) = 2[1~pj(k)]N(t); (40) In state>j:N>j(t) = [2~pj(k)1]N(t); (41) In statej:Nj(t) = (42) which have a similar structure to the dynamics of model (Eqs. 14-17), but are dependent on the Laplace transforms of the state lifetime these Laplace transforms give rise to an effective mean time. D. The exponential mean To understand the biological significance of the Laplace transform of the lifetime PDF, consider the generalized f-mean (or Kolmogorov mean) where the random variable tis first transformed by function g, an arithmetic mean is performed, and then the inverse function is applied to generate a generalized (43)7 where Etis the arithmetic expectation over Both the harmonic mean and geometric mean are special cases of this more general formulation. The Laplace transform of the lifetime and age PDFs can be reinterpreted as the expectation of g(t) = we can generate the ~pj(k); ~pj(k); (45) which can be understood as the exponential-mean of the lifetime and age of state returning to our model, we will explore the behavior of the exponential mean. Consider the special case of a distribution that is very narrow relative to the growth rate. In this case: t(k) (46) where the exponential mean is equal to the mean to the order of the variance ( 2 t) times the growth ratek. Details of the derivation are given in Appendix A 6. Short-lived states and states with will therefore have exponential means equal to the mean. More generally, the Jensen guarantees the exponential mean is less than or equal to the mean: t(k)Ett; (47) since the function g(t)is convex [30]. Finally, let us consider the consequences of a very wide distribution of lifetimes. Consider a state jin which fraction of cells arrest ( j! 1 ) while the remaining cells have exponential-mean lifetime j;0. Using Eq. 45, it is straightforward to compute (48) where the second term acts to extend the lifetime by a positive multiple of the doubling time T. Although the arrested cells do lengthen the it remains finite. Eq. 48 is a anytime some fraction of the cells have a lifetime much longer than the doubling time even if all lifetimes are finite. E. Model determine the differences between and stochastic models, we eliminate lifetime PDFs ~pjin favor of lifetimes jusing their definition (Eq. 45). First consider the consistency condition for exponential growth (Eq. 38). The ensures that the exponential-mean lifetimes of successive states add to generate the age of state j 0 0.2 0.4 0.6 0.8 1 Corresponding Deterministic Age: j / age: / Modelt1t2t3.... .... tm tm-1 tm-2 j = 1 j = mFIG. 4. Model correspondence. The deterministic and stochastic models generate identical statistics in once a suitable correspondence is defined between cell statejin the stochastic model and cell age in the deterministic model. State jin the stochastic to age interval (j1;j]in the which is represented by the red line. In the the PDF of age as a function of state jis the age in the deterministic model tracks with the mode of the age in the stochastic model. (e.g.the natural logarithm of Eq. 37). Eq. 38 can now be rewritten as a relation between the duration and the doubling (49) which is now intuitively equivalent to the in the deterministic model (Eq. 10). Details of the derivation are given in Appendix A 7. Now consider the expressions for state number in the stochastic model (Eqs. 39-41). When the evaluated at exponential-mean stochastic age j, the numbers are identical in the two (50) In statej:Nj(t) =N(t)j=j; (51) In state>j:N>j(t) =N>(t)j=j; (52) In statej:Nj(t) =N[j1;j](t): (53) We therefore conclude that the statistics of and stochastic models are identical in an exponential culture for models with the same growth ratek, once a correspondence has been states jand ages. In the deterministic to times This correspondence is in Fig. 4. Since we demonstrated a correspondence between the models, almost all the application discussed in Sec. II B generalize by replacing the deterministic time with the mean [31].8 Strain Doubling Time: TC TW CW Mutant A TW+ ln 2TWCW+ ln 2TW Mutant S I. The effect of mutants on the doubling time Tand C period duration Cof an exponential culture. F. Implications for cell cycle phenomenology To explore the nontrivial consequences of stochasticity in timing, consider an example motivated by replication conflicts [32]: By visualizing the replisome dynamics using single-molecule microscopy, we have recently reported that transcription leads to pervasive replisome instability [24]. To what extent should conflict-induced pauses in replication have been detectable in the classic analyses of unsynchronized a simplified model in which an the difference between the wildtype strain W and two mutant strains. The wildtype W grows C period CWand deterministic cell In mutant strain A(rrest), a small fraction  of cells arrest during replication ( i.e.C period) and never complete the cell cycle, whereas non-arrested cells are identical to wildtype cells. In mutant strain S(low), the replication process is 1+0times slower, but the B and D periods are identical to wildtype. Using Eq. 48, one can compute the C period duration Cand doubling time T. To lowest order in , the inferred cell cycle durations and C period are presented in Tab. I. At an intuitive level, one aspect of the prediction is easy to understand: In both mutants the C period is lengthened, as one might na ively expect since this is the replication period of the cell cycle. Furthermore, the doubling time increases by the same amount as the C period increases. But there is an aspect of this prediction which is perhaps less intuitive: One might naively expect to observe a more dramatic consequence of replication arrest, like a large buildup of C period cells, but the consequences are from a slowdown in an exponential culture. In both mutants, there is a slight lengthening of the inferred C period, even though the slowdown is caused by in the context of the A mutant. Although this prediction is not new in a qualitative sense, it how the statistics of the exponential culture mask two mechanistically distinct phenomena. The statistics of an exponential culture can also generate distinctions where seemingly none exist. Consider a more realistic model in which the duration of the D period is stochastic, has a non-zero width, and is identical for all three strains. The more rapid growth rate of the wildtype strain implies that its effective D period is shorter than for mutants cells: DS;DA>D W; (54)even though the distributions of the D period durations are identical in all three strains. (To understand how this occurs, see the second term on the RHS of Eq. 46.) In most cases this effect should be subtle, but for large changes in growth rate, these changes could be and can clearly complicate the effective period lifetimes in an exponential culture. III. DISCUSSION In this paper we provide a detailed analysis of and stochastic models of the cell cycle. In Sec. II A, we solved the deterministic model in which the cell-aging and division processes are precisely timed and determined the demographics ( an exponential culture. Given a set of Sec. II B provides a detailed road map for how to infer cell cycle state timing in the context of the deterministic model. In Sec. II C, we solved the more realistic stochastic model in which the lifetimes of sequential states are stochastic and again we determined the demographics of an exponential culture. By defining an exponential mean in Sec. II D, we the statistics of the two models were equivalent in Sec. II E. The effective lifetime of states in model is the exponential mean of the lifetimes in the stochastic model. That is to say that lifetimes are the sufficient statistics of the model ( e.g.[33]): Knowledge of only these predicts the demographics of the ; therefore, inference on infers only the exponential means, rather than the underlying lifetime distributions in Sec. II F, we discussed some of the limitations of the exponential-mean lifetimes in resolving the underlying biological mechanisms. A. Applicability of the stochastic model Is the stochastic model sufficiently complex to capture all the relevant cell cycle phenomenology in E. coli and other bacterial systems? Like the before it, the stochastic model is an idealized model that is simple enough to be tractable complex enough to capture some There are a number of shortcomings of this model but perhaps the most significant is that there is no memory beyond the cell state index j. As a consequence, it makes predictions at variance with some observed phenomenology: For instance, the stochastic model must predict that successive cell cycle durations are uncorrelated; however, are observed [4, 34]. (We briefly consider the implications of a more general model in Appendix A 8.) Another important limitation of the stochastic model is that cell divisions are symmetric, which is9 a good approximation in E. coli , but these types of stochastic models can easily be extended to the division case ( e.g.[28]). B. On the applicability of the exponential mean Although the definition of the exponential mean was motivated by the correspondence between the deterministic and stochastic models, it almost certainly has much greater applicability to other more complicated scenarios. For instance, our own using more complex models suggest that the relation between the effective lifetime of the states and the exponential-mean lifetime appears to be more robust than the assumptions of the stochastic model might imply. Since the key mechanism for generating bias toward short times is growth, we expect the exponential mean of wait times to be the determinative statistic in more general models, as demonstrated in Appendix A 8. As such, the exponential-mean lifetime could be a powerful observable to bridge timescales and culture phenomenology in two (i) in experiments probing cell cycle dynamics at the single-cell level and (ii) in complex that are too slow and too memory intensive to simulate in the long time limit [35]. We should note that although we believe of the doubling time as an exponential mean (Eq. 49) is novel, it has already been appreciated in two important respects: (i) From a the formulation (Eq. 38) of Eq. 49 has long been known [27]. (ii) From a qualitative perspective, biologists have long understood the consequences of the exponential-mean lifetime on cell growth rate: I.e.the doubling time Tis an average of the cell cycle duration T; however, a small for whom T! 1 , slows but does not stop growth. There is also physical precedent for this type of mean: Intriguingly, it emerges in context statistical mechanics [36] [37], although what connection this has to our cellular dynamics is opaque. C. On the significance of does stochasticity affect biological we have long known that although the statistics of an exponentially growing population are well described by the deterministic model [2], in cell cycle timing is observed [4, 5]. It is therefore tempting to conclude, based on the literatureand perhaps even our own results, that stochasticity is either small or simply does not significantly function. Our own conclusions are much more our results guarantee that the fits the demographic data just as well as the stochastic model, we that the stochasticity in timing is hidden in plain sight. The distribution of state the exponential means. Therefore, the success of the deterministic model should not be interpreted as evidence against stochasticity or against its importance, but rather it indicates that only state lifetimes are in the model for the demographics of an exponential culture. Perhaps more than anything else, the between the deterministic and stochastic models emphasizes the need for synchronized single-cell measurements: In Sec. II F, we illustrated (i) how similarities in the effective duration of the C period obscures distinct biological mechanisms as well as (ii) how differences in the effective D period could belie an identical mechanism. At a mechanistic level, stochasticity plays a central role in many processes. For instance, the mechanism that restarts replication will prevent the existence of afat tail on the distribution of C periods [24, 32, 38]. Although the existence of the fat tail-- i.e.a small number of cells with very long C break the correspondence with the it does increase the exponential-mean C period, which in turn decreases the growth rate. (E.g. see Tab. I.) Since the growth rate is decreased, there is a strong selective pressure to reduce stochasticity . This argument predicts the existence of to reduce stochasticity, as are already known in many contexts ( e.g.replication restart). In fact, the subtle signature of stochasticity suggests an interesting hypothesis: a significant number of mutants that are currently known to reduce growth rate may in fact generate this phenotype by increasing the level of stochasticity in the cell cycle duration. analysis must play a central role in understanding these acknowledges advice and comments from M. S. Iyer-Biswas, P . Levine, J. Mittler, R. Phillips, M. Transtrum, B. Traxler, I.M. Shelby, and H.K. Choi. This work was supported by NIH grant L. Willis and K. C. Huang, Nat Rev Microbiol 15, 606 (2017).[2] S. Cooper and C. E. Helmstetter, J Mol Biol 31, 519 (1968).10 [3] H. Bremer and G. Churchward, J Theor Biol 69, 645 (1977). [4] P . Wang, L. Robert, J. Pelletier, W. L. Dang, F. Taddei, A. Wright, and S. Jun, Curr Biol 20, 1099 (2010). [5] L. Robert, M. Hoffmann, N. Krell, S. Aymerich, J. Robert, and M. Doumic, BMC Biol 12, 17 (2014). [6] X. Wang, C. Possoz, and D. J. Sherratt, Genes Dev 19, 2367 (2005). [7] H. L. Withers and R. Bernander, J Bacteriol 180, 1624 (1998). [8] C. J. Rudolph, A. L. Upton, A. Stockum, C. A. Nieduszynski, and R. G. Lloyd, Nature 500, 608 (2013). [9] D. Bates, J. Epstein, E. Boye, K. Fahrner, H. Berg, and N. Kleckner, Mol Microbiol 57, 380 (2005). [10] N. J. Kuwada, B. Traxler, and P . A. Wiggins, Mol Microbiol 95, 64 (2015). [11] D. W. Adams and J. Errington, Nat Rev Microbiol 7, 642 (2009). [12] The exact degree to which this is reduced depends on the ratio ofz=Tas discussed below. [13] T. Lo, D. Huang, H. Merrikh, and P . A. Wiggins, E. J. Stewart, R. Madden, G. Paul, and F. Taddei, PLoS Biol 3, e45 (2005). [15] The naming of the cumulative creation and was motivated in relation to the creation and annihilation operators from quantum field theory. [16] S. Du and J. Lutkenhaus, Trends Microbiol 27, 781 (2019). [17] X. Ma, D. W. Ehrhardt, and W. Margolin, Proc Natl Acad Sci U S A 93, 12998 (1996). [18] H. Niki, Y. Yamaichi, and S. Hiraga, Genes Dev 14, 212 (2000). [19] I. F. Lau, S. R. Filipe, B. Soballe, O.-A. Okstad, F.-X. Barre, and D. J. Sherratt, Mol Microbiol 49, 731 (2003). [20] R. E. Bird, J. Louarn, J. Martuscelli, and L. Caro, J Mol Biol 70, 549 (1972). [21] R. H. Pritchard, M. G. Chandler, and J. Collins, Mol Gen Genet 138, 143 (1975). [22] K. P . Lemon and A. D. Grossman, Mol Cell 6, 1321 (2000). [23] M. Wallden, D. Fange, E. G. Lundius, O. Baltekin, and J. Elf, Cell 166, 729 (2016). [24] S. M. Mangiameli, C. N. Merrikh, P . A. Wiggins, and H. Merrikh, Elife 6(2017), S. M. Mangiameli, B. T. Veit, H. Merrikh, and P . A. Wiggins, PLoS Genet 13, e1006582 (2017). [26] P . A. Wiggins, R. Phillips, and P . C. Nelson, Phys Rev E Stat Nonlin Soft Matter Phys 71, 021909 (2005). [27] E. Powell, Microbiology 15(1956). [28] F. Jafarpour, C. S. Wright, H. Gudjonson, J. Riebling, E. Dawson, K. Lo, A. Fiebig, S. Crosson, A. R. Dinner, and S. Iyer-Biswas, Phys. Rev. X 8, 021007 (2018). [29] A. Kolmogorov, Atti Accad. Naz. Lincei 12, 388 (1930). [30] J. L. W. V . Jensen, Acta Mathematica 30, 175 (1906). [31] The exception are the replicating cell statistics Nrep cell . In this case, the implicit assumption that states are sequential cannot always be implemented in the stochastic model. For instance, consider the re-initiation of replication at the quarter cell positions late in the cell cycle in rapidly proliferating cells, as illustrated in Fig. 1A. If these events were modeled as independent, you will first see one replisome initiate at one quarter cell position and then the other. In this case one must compute the exponential means using the order statistics ( e.g.[33]). [32] H. Merrikh, Y. Zhang, A. D. Grossman, and J. D. Wang, Nat Rev Microbiol 10, 449 (2012).[33] D. R. Cox and D. V . Hinkley, Theoretical & Hall, 1974). [34] J. Lin and A. Amir, Cell Systems Personal communication from S. C. Jarzynski, Phys. Rev. Lett. 78, 2690 (1997). [37] Personal communication from S. H. Merrikh, C. Mach 'on, W. H. Grainger, A. D. Grossman, and P . Soultanas, Nature 470, 554 (2011). Appendix A: Supplemental derivations 1. Derivation of the rate equation in the To obtain the cumulative creation and in terms of the number density (Eq. 2), we integrate the number density at fixed age over all timetto obtain the cumulative number of cells that have ever entered (creation) or left (annihilation) age . They are equivalent due to the continuous nature of the deterministic model. If states were discrete, as in the stochastic model, then the cumulative creation and annihilation numbers would differ by the number of cells currently in the state . To obtain Eq. 4 from Eq. 3, we divide both sides of Eq. 3 byt and replace _N +(t)with the which leaves: _n(t) (A1) Taking the limit as goes to 0 and using the definition of a derivative, we are left with: _n(t) =@_N+ (t); (A2) which is Eq. 4 in the main text. Eqs. 5-6 follow from taking the partial time derivative of Eq. 2 and taking into account the consistency condition: n0(t) = 2nT(t): this consistency condition describes how cell division at age Tleads to twice as many daughter cells of age = 0. 2. Derivation of the solution in the deterministic model In the deterministic model, we can assume growth of the population is represented by an exponentially increasing time dependence factor, ekt, with a constant unknown growth rate k. This assumption holds in the long time limit, since only the fastest growing mode remains in exponential growth, while all others (smaller k) are diluted out. We thus stipulate that in the deterministic model, all must grow with this same time dependence. The number density is then a solution of the form: n(t) =n(0)ekt; the initial age distribution at t= 0. Plugging this into Eq. 7 for the  >0case yields: kn(t) =@ @(n(t)): (A5) This can then be integrated to yield the solution: n(t) =n0(t)ek; (A6) =n0ektek; (A7) wheren0is a constant determined by the initial cell number. This equation appears in the main text as Eq. 11. To satisfy the = 0 case of Eq. 7, we must use the consistency condition (Eq. A3): n0(t) = 2nT(t); (A8) n0ekt= 2n0ektekT: (A9) Dividing both sides by n0ektand solving for 2: (A10) Therefore, the doubling time defined in Eq. 9 is equivalent to the cell cycle duration, as one would naively expect. Furthermore, this equation relates the growth rate k, a population measure, to the cell a single-cell measure. 3. Derivation of the PDF and CDF in the To obtain the probability distribution function, f(), with respect to cell age (Eq. 12), we must normalize the number density at any fixed time t: f() (A11) wheren() =n(= 0)ek, which is just Eq. 11 with the fixedtfactor absorbed into n(= 0) . Evaluating the integral and replacing n()with the expanded (A13) Now recall that T=Tk1ln 2, from Eqs. 9-10. Plugging this Tinto Eq. A13 gives Eq. 12: f() = 2kek: (A14) To obtain the CDF, integrate with respect to from 0to : F() 2(1ek): (A16)4. Derivation of the cumulative creation number in the cumulative creation number Eq. 2 evaluated at = 0: N+ 0(t) = 2N(t); (A17) which is double the current number of cells N. The factor of two arises due to the cumulative nature of N+ (t). To understand this intuitively, consider the total number of cells in each = (1 (A18) which is a geometric series and can be summed to 2N, matching Eq. A17. We can now multiply by the -dependence term, ek, to obtain: N+ (t) = 2ekN(t): (A19) More formally, we can use direct integration of Eq. 11: N+ (t) =R n(t0)dt0; (A21) N+ 0(t) =1 kn0ekt+c; (A22) where c is an integration constant. We now allto obtain the total number of cells at time t: N(t) =RT 0n0(t)d0; (A25) Combining this with the consistency condition Eq. A17, we get: N+ 0(t) =1 kn0ekt: (A26) Setting this equal to Eq. A22 allows us to set the integration constant c= 0. Eq. A21 then becomes Eq. A19 from above, which is Eq. 14 from the main text. 5. Derivation of the solution in the stochastic model Clearly, Eqs. 33 and 34 can be combined recursively to generate a relation between the number of cells entering states 1 and j. First let us define the state-transition time PDFpj, describing the total time taken to transition from birth through state (A27) ~p~pm (A28) wherepis the lifetime PDF for the entire cell cycle. We then write an expression of the number arriving in state j: ~N+ j= ~pj1~N+ 1: (A29)12 As before, using this same condition at the end of the cell cycle gives rise to a consistency 2~p~N+ 1: (A30) It follows that in steady-state exponential growth, the growth rate kmust correspond to the solution to the equation: 1 = 2~p(k); (A31) an equation that is well known [27]. LetNjbe the total number of cells in states i= 1:::j. The dynamics of this quantity has a simple form due to the telescoping form of the dynamics equations (Eqs. 32-34) where the number entering the ith state exactly cancel the number leaving the i1th (A32) = (1~pj)~N+ 1: (A33) To determine the overall normalization, we can sum up the cells in all states and set that sum equal to the total number of cells (A34) From ~Nj, we can compute the number in (A35) = (A36) In the long time limit, the fastest growing mode dominates the solution and therefore: N(t) =N(0)ekt; (A37) Nj(t) = (A38) Nj(t) = 2[1~pj(k)]N(t); (A39) N+ j(t) = 2~pj1N(t); (A40) which also appear in the main text. 6. The exponential mean of a very narrow distribution To obtain Eq. 46, we begin with the exponential mean: t(k) (A41) which is obtained by using the function g(t) = exp(kt) in the generalized expectation equation (Eq. 43). We then use a series expansion of the exponential term, ex= 1 +x+1 2x2+:::, which yields: t(k) +:::
:(A43)We then use the series expansion ln(1 +x) only second order terms, since the distribution is very narrow: t(k) the definition of the we obtain Eq. 46: t(k) (A46) 7. Derivation of the consistency condition in the the deterministic model, Eq. 10 is a that describes the na ive expectation that the duration of the cell cycle is equal to the doubling time of the (A47) In the stochastic model, the consistency condition in terms of the Laplace transform is given by Eq. 38: 1 = 2~p(k): (A48) However, the mathematical equivalence is opaque for the moment. To make the equivalence clear, we use the exponential mean Eq. 45: j(k)1 kln ~pj(k); (A49) along with the relation between the PDF of lifetimes of individual state jand the PDF of times taken to transition from birth through state (A50) which is Eq. 37 in the main text. Combining these two equations and letting jbe the final state m, we obtain the exponential mean of the stochastic cell cycle mY i=1~pi! ; ~pi=mX i=1i: (A52) If we use the consistency condition Eq. 38, we can ~p(k); (A53) =k1ln 2 =T; (A54)

2108.10988

Dean Huang

Dean Huang, Teresa Lo, Houra Merrikh, Paul A. Wiggins

Characterizing stochastic cell cycle dynamics in exponential growth

Phys. Rev. E 105, 014420 (2022)

10.1103/PhysRevE.105.014420

physics.bio-ph

http://creativecommons.org/licenses/by/4.0/

Two powerful and complementary experimental approaches are commonly used to study the cell cycle and cell biology: One class of experiments statistics (or demographics) of an unsynchronized while the other captures cell cycle dynamics, either by of full cell cycles or in bulk experiments on synchronized populations. In this paper, we study the subtle relationship between observations in these two distinct experimental approaches. We begin with an existing model: a single-cell deterministic description of cell cycle dynamics where cell states (i.e. periods or phases) have precise lifetimes. We then generalize this description to a stochastic model in which the states have as described by arbitrary probability distribution functions. Our analyses of the demographics of an exponential culture reveal a simple and exact correspondence between the deterministic and stochastic models: state lifetimes in the deterministic model are equal to the exponential mean of the lifetimes in the stochastic model. An is therefore that the demographics of an exponential culture will be well-fit by a deterministic model even if the state timing is we explore the implications of the models in the context of the Escherichia coli cell cycle, we expect both the models as well as the significance of the exponential-mean lifetimes to find many applications in the quantitative analysis of cell cycle dynamics in other biological systems.

1 Introduction The quest for New Physics (NP) phenomena beyond the Standard Model predictions is a principal and interesting issue. Numerous extensions of the Standard Model (SM), including Grand Unified Theories (GUTs) and String Theory derived effective models incorporate novel ingredients in their spectra. The latter could manifest themselves through exotic interactions and their novel predictions. Amongst the most anticipated ones are additional neutral gauge bosons, leptoquark states forming couplings with quarks and leptons, additional neutral states (such as sterile neutrinos) and vector-like experimental data of the Large Hadron Collider (LHC) and elsewhere, on the other hand, provide significant evidence of the existence on possible novel interactions mediated by such exotic states, although nothing is conclusive yet. Some well known persisting LHCb data which are in tension with the SM predictions, for example, are related to various B-meson decay channels. In of the ratio of the branching ratios with the semi-leptonic transitions that lepton flavor universality is violated [1, 2]. Possible explanations of the effect involve leptoquark states, Z0neutral bosons coupled differently to the three fermion families and vector-like generations [3, 4, 5, 6, 7, 8, 9]. In a previous study [10], (see also [11]) we performed a systematic analysis of a class of models with SU(5)U(1)0gauge symmetry obtained from a covering E8gauge group through the (1.1) where U(1)0stands for any linear combination of the four abelian factors incorporated in SU(5)0. In this framework, we have derived all possible solutions of the anomaly-free U(1)0factors and have shown that many of these cases entail non-universal couplings to the three chiral families. Next, we considered the case where the spontaneous breaking of the U(1)0symmetry occurs at a few TeV scale and examined the implications in low energy phenomenology, computing observables of several exotic processes in the effective theory. Despite the rich structure and the variety of the non-universal U(1)0factors, strong lower bounds coming from the K-Ksystem [12] on the mass of the associated Z0boson, far outweigh any in B-meson anomalies and the non universal contributions to completely depleted. It was shown that in the so derived effective F-theory models only the existence of additional vector-like families could interpret the LHCb data [10]. In the present letter we expand previous work [10, 11] on F-theory inspired by including vector-like fermion generations in the low energy spectrum. More precisely, we are interested in models that allow the existence of a complete family of extra fermions in addition with the spectrum of the Minimal Supersymmetric Standard Model (MSSM). To avoid severe constraints for the Kaon system, we look for models that the regular MSSM fermion matter fields acquire universal charges under the additional U(1)0symmetry and are chosen to be different from the corresponding states of the vector- like family. This way the non-universality effects are strictly induced from the considered [13, 14, 15, 16].The paper is organized as follows: in sections 2 and 3 we describe the origin of the gauge sym- metry of the model, the anomaly cancellation and flux constraints and define the content in terms of the flux parameters. In section 3 in particular all models with one vector-like family are sorted out in five classes distinguished by their U(1)0properties. The phenomenological analysis of the models, couplings and mass matrices are presented in section 4. Section 5 deals with the implica- tions on flavor processes and particularly B-meson anomalies. A short discussion is devoted on of R-parity violating terms in section 6. Summary and conclusions are found in section 7. 2 Flux constraints for a spectrum with a complete vector-like family In this section we present a short description of the GUT model, focusing mainly on the basic constraints and characteristics coming from its F-theory embedding. Further technical details can be found in [10]. The (semi-local) F-theory construction in the present work is assumed to originate from an E8singu- larity under the reduction shown in (1.1). The Cartan generators to the four U(1)factors in (1.1), subjected to the condition a5 i=1ti=0, are taken to (2.2) To ensure a tree-level top-quark mass a Z2monodromy t1$t2is imposed, breaking invariant the remaining three abelian factors. In addition, appropriate fluxes [17] can be turned on along the remaining U(1)'s in such a way that some linear combination U(1)0of the abelian factors remains unbroken at low energies. Thus, the gauge symmetry of the effective model under assumed to be left unbroken in the effective model is a linear combination of the symmetries surviving the monodromy action, (2.4) with the coefficients the normalization (2.5) The latter also are subject to anomaly cancellation conditions which have been analysed in detail else- where [11, 10]. After imposing the Z2monodromy, the 10 ;10 and 5 ;5 representations massless fields reside on four matter curves S10j;j=1;2;3;4 and seven S5i;i=1;2;:::; 7 respec- tively mentioned above, also determine the chiralities of the Their effect on the representations of the various matter curves S10j;S5ican be parametrized in terms of n5in5i=Mi; (2.6) while, in order to accommodate the three fermion families, the chirality condition be imposed. Furthermore, turning on a hypercharge flux FY, the SU(5)GUT symmetry is broken down to Parametrizing the hypercharge flux with integers Ni;Njthe of the SM representations are given 5 start with the flux data and the SM content of each matter curve. For details we refer to our previous work [10] and here we present only the properties of the complete spectrum as shown in Table 1. In order to obtain the desired spectrum the following constraints where taken in account. Matter Curve Q0NY M SM N9 M1 M2 M3 M4 1: Matter curves along with their U(1)0charges, flux data and the corresponding SM content. Note that the flux integers satisfy N=N7+N8+N9. The spectrum of a local F-theory model is determined once a set of the above integers -respecting the aforementioned constraints- is chosen. Thus, to start with, we proceed by accommodating the Higgs doublet Huon the S51matter curve. Choosing the associated flux integers to be M1=0;N=1, it can be observed that the Huremains in the massless spectrum and at the same time the down-type color triplet is eliminated. As a consequence of this mechanism [17], proton decay is sufficiently focusing on the S101matter curve, we let m1vary in 0 <m1<3. In addition, thanks to the Z2monodromy [18, 19] already discussed, at least one diagonal tree-level up-quark Yukawa is effectuated in the superpotential W. Furthermore, in order to ensure exactly one extra 3family of vector-like fermions, in addition to the condition ajmj=aiMi=3 which fixes the number of chiral families to three, we also impose the following conditions into the various flux integers [11, of their allowed ranges and values are subjected to conditions, the remaining flux parameters are limited as follows: We restric the m2;3;4flux integers characterizing the number of Q;Qstates in the spectrum, in the range [1;2]. Since S101matter curve always hosts at least two uc's (due to conditions we bound the other ucmultiplicities ( mjNkwith j=2;3;4 and k=7;8;9) to be in the range [1;1]. Similarly, for the multiplicities of the ec;-ecstates we for the same way, for the dc's we set the values of the corresponding multiplicities Mi's, vary in the range [3;1], while for the multiplicities of L's (see Table 1) the relations are set to vary in the range [2;1]. We note here that for the latter, in general we could allow for values in the range [3;1], but this leads to mixing of the vector-like states with the MSSM ones, something that is against our intention to look for models with vector-like U(1)0charges different than the MSSM all the restrictions described above, we receive 1728 flux solutions with one in addition to the three standard chiral families of quarks and leptons. 3 Classification of the Models In order to determine the cicoefficients and consequently the U(1)0charges for each model described by the above set of fluxes, we impose anomaly cancellation conditions. In particular we impose only the mixed MSSM- U(1)0anomalies: The pure U(1)0cubic anomaly ( A111) anomalies ( AG) can be fixed later by taking into account the dynamics of the singlet fields that typically appear in F-theory models. Furthermore, in the quest for we shall confine our search in cases where the three MSSM families have universal U(1)0 charges and only the charges of the vector-like fields will differ. This way, from the resulting 1728 models only 192 of them appear with this property. These 192 models fall into five classes with respect to their Each class contains models that carry the same charges under the extra 4U(1)0and they only differ of how the SM states are distributed among the various matter curves. We present one model for each class in Tables 2 and 3. Model c1 c2 c3 A 1 2 1 -1 0 -1 0 0 -1 -2 1 1 0 0 0p 15 41 4 B 1 2 -1 1 0 0 0 0 -1 -3 1 1 0 0 2 1 1 -1 0 0 0 1 -3 -1 0 0 1 2 -1 1 1 0 -1 0 1 -1 -2 0 0 0 1 -1 2 1 0 0 1 0 0 -1 -3 0 1 0 2: Representative flux solutions along with the corresponding ci' s for the five class of models A, B, C, D and E. Model A Model B Model C Model D Model Q+2uc-2 Q+2uc1/4 2 Q+3uc+ec-3/4 2 Q+3uc+ec9/2 Q+2uc 1/2 2 Q+uc+3ec-2 2 Q+uc+3ec-1/2 Q+uc+ec-1/2 Q+uc+ec 11/2 Q+uc+ec -2 Q+uc+ec-1/2 Q+uc+ec 1/4 Q+2ec7/4 Q+uc+ec9/2 2 Q+uc+3ec -1/2 Q+uc+ec 11/2 Q+uc+ec1/4 Q+uc+ec -3/4 Q+2ec-8 Q+uc+ec -1 Hu 4 Hu -1/2 Hu 3/2 Hu -9 Hu 1 dc+2L -4 L -1/4 L -1/4 dc+2L -1 L -3/2 L -3/2 L 1/2 L 1 L -9 dc 1 L 7/2 L 0 dc 3/2 dc -7/2 L -3/2 dc-3/2 dc-1/4 3 dc+2L 9/4 dc+L 1 L 1 2 dc+L 7/2 3 dc+2L -3/4 dc+L -1/4 2 dc+L -27/2 dc+L 3/2 dc+L -6 dc+L 0 L -1 L -7/2 3 dc+2L Table 3: The particle content of models A, B, C, D and E using the data from Table 2. Table 2 presents the flux data along with the respecting solution1for the coefficients ciof each model. The corresponding U(1)0charges and the spectrum for each model are presented in Table 3. Note that models B and C coincide with the models 5 and 7 respectively, derived in [11]. In addition to the fields presented in Table 3 there are also singlet fields with weights (titj)that appear in the present In the analysis that follows we will denote these singlet fields as qtitj=qi j. 4 Analysis of the Models In the previous section we generated five classes of models which all share a common characteristic. The U(1)0charges of the vector-like states differ from the universal U(1)0charges of the SM chiral families. Models with this feature can explain the observed B-meson anomalies provided there is substantial mix- ing of the SM fermions with the vector-like exotics. At the same time, lepton universality is preserved 1Note that a ' mirror ' solution subject to ci! cialso exists. 2For a detailed definition of the singlet spectrum of the theory see [10]. 5among the three chiral families and severe bounds coming from the Kaon system and other flavour vio- lating processes are not violated. In the following we will analyze the models of Table 3 and derive the mass matrices for each model. 4.1 Model A For this case we have chosen the following set of fluxes corresponding U(1)0charges for the various representations are the 10,5 representations come with the opposite U(1)0charge. We distribute the fermion genera- tions and Higgs fields into matter curves as follows we can write down the superpotential and in particular the various terms contributing to the fermion mass matrices. We start with the up-quark sector. The dominant contributions to the up-type quark masses descend from the following superpotential yi's are coupling constant coefficients and Lis a characteristic high energy scale of the theory. The operators yield the following mass texture y2J2 13vu y1J13vu eytvu y3J14vu y7q51 y1J13vu y1J13vu ytvu y3J14vu y5J2 14vu y8q54 y6q53 y6q53 y7q51 y8q54 (4.2) where j=hqi ji,Ji j=hqi ji=Lande1 is a suppression factor introduced here to capture local effects of Yukawa couplings descending from a common operator [20, 21]. Next we analyse the couplings of the down-quark and charged lepton sectors. In the vector-like part of the model, up and down quark sectors share some common superpotential operators. These are 6given in (4.1) with couplings y6,y7andy8. The remaining dominant terms contributing to the down-type quarks coupling constant coefficients. As regards the charged lepton sector, we start with the common operators between bottom quark and charged leptons which are given in (4.3) with couplings There are also common operators between the up quark and the charged lepton sector which are given in (4.1) with couplings y6 andy8. All the other contributions for the charged lepton mass matrix descend from the lidenotes coupling constant the various singlet fields qi jacquire vacuum expectation values (VEV), hqi ji6=0, they non-zero entries in the mass matrices of quarks and charged leptons. These VEVs, however, are subject to phenomenological requirements. Such an important constraint comes from the m-term which in principle can be materialized through the coupling q155154. Clearly, to avoid decoupling of the Higgs doublets from the light spectrum, we must require hq15i0. Consequently, the mass terms involving q15of the down and charged leptons can be ignored. We obtain the following down quark mass 0 0 k12q43 k2J43vd k8J13vd k7vdk9J14vd k13q53 y6q53 y6q53 y7q51 y8q54 (4.5) The mass texture for the charged leptons has the following form k2J43vd k2J43vd l1vd y6q53 k3J13J45vd k2J43vd k2J43vd l1vd y6q53 k3J13J45vd k2J43vd k2J43vd l1vd y6q53 0 k5vd k5vd Model B The second Model is obtained using the following set of flux parameters corresponding U(1)0for the various matter curves are 4;52: 6: A workable distribution of the fermion generations and Higgs fields into matter curves is as follows here is realized through the coupling q135152, so we require that hq13iis very small compared to the other singlet VEV's. This restriction obligates us to take high order terms in account for some couplings. We write down the various terms that construct the fermion mass matrices starting from the up-quark sector. The dominant contributions to the up-type quarks descend from the following superpotential terms operators yield the following mass texture 0 eytvu y3J15vu ytvu y3J15vu y6q41 0 0 y3J15vu y5J15vu y8q45 y7q43 y7q43 y6q41 y8q45 continue with the bottom sector of the model. There are common operators between top and bottom sector which are given in (4.7) with couplings y6,y7andy8. The remaining dominant to the down-type quarks these operators we obtain the following mass matrix describing the down quark k1J14J2 43vd e2k0vdk5J43J15 k1J14J2 43vd ek0vd k5J43J15 k1J14J2 43vd k0vd k5J43J15 y7q43 y6q41 y8q45 (4.10) Regarding the charged lepton sector, the common operators between bottom sector and charged leptons are those in (4.9) with couplings k1,k5,k6andk7. We also have common operators between the top and charged lepton sector which are given in (4.7) with couplings y7andy8. Pure charged descend from the all the contributions lead to the following mass matrix k1J14J2 43vd k1J14J2 43vd k1J14J2 43vd y8q45 k7q43 k7q43 l5q41 Model C Next, a representative model of the class Cis analyzed. The flux integers along with the are given in Table 2. The resulting U(1)0charges for the various matter curves 0: and we assume the following distribution of the various fermion and Higgs fields into matter here comes through the coupling q145153and so, we require hq14i0. Once again we will consider high order terms for some couplings. Next, we write down the various superpotential terms leading to the the mass matrices for the up, bottom and charged lepton sector. 9The dominant contributions to the up-type quarks descend from the following superpotential terms yield the following mass texture ytvu y2J13vu y5q51 0 y6q53 y7q54 y5q51 y5q51 y6q53 (4.14) where his a small constant parameter describing local Yukawa effects. There are common operators between top and bottom sector. These are given in (4.13) with The remaining operators contributing to the down-type quarks (4.15) Combining all the terms we obtain the following mass matrix describing the down quark 0 0 kJ54vd 0 0 0 k1xvd k1vd k9J13vd k10q13 y7q54 y5xq51 y5q51 y6q53 (4.16) where eandxare small constant parameters describing local Yukawa effects. In the charged lepton sector we have some contributions descending from terms in (4.15). These are the operators with couplings We also have common operators between top and charged lepton sector which are given in (4.13) with couplings y5,y6andy7. All the other descend from the following the mass texture for the charged leptons has the following form 0 0 y7q54 l2J51vd kJ54vd kJ54vd k1vd y5q51 l3J53vd 0 0 l5q53 l5q53 (4.18) 4.4 Model D We now pick out a model belonging to the fourth class. The U(1)0charges for the various matter curves are 1: A promising distribution of the fermion generations and Higgs fields into the various matter curves is as follows this case the m-term is realized through the coupling q145153, and therefore, as in the previous models we require that the singlet VEV is negligibly small, hq14i0. Now, we write down the various superpotential terms of the model that lead to the mass matrices for the top and bottom sector. The dominant contributions to the up-type quarks descend from the following superpotential opera- tors operators yield the following mass texture e3ytvu e2ytvu y2J13J34vu y5q31 y1J15vu e2ytvu eytvu y2J13J34vu y5q31 y1J15vu eytvu ytvu y2J13J34vu y5q31 0 y7q34 y6q35 y5q31 y5q31 y7q34 (4.20) 11In close analogy with the previous cases, the operators with couplings y5,y6andy7in (4.19) con- tribute also in the bottom sector. The remaining dominant terms contributing to the down-type all the terms we obtain the following mass matrix for the down-quark sector of the k5J54vd eybvd k4J13J34vd ybvd k4J13J34vd y5q31 y5q31 y7q34 (4.22) We turn now to the charged lepton sector of the model. The operators with couplings (4.21) contribute also to the charged lepton mass matrix. We also have common operators between top and charged lepton sector which are given in (4.19) with couplings contributions descend from the (4.23) Combining all the contributions we receive the following mass matrix for the charged leptons of k3vd k2J15vd k7J13vd y6q35 k3vd k3vd k2J15vd k7J13vd y6q35 k1J51vd k1J51vd ytvd k6J54vd y5q31 k5J54vd k5J54vd l4q53 Model E For the fifth and final Model we have the following U(1)0charges for the various matter 1;53: order to receive realistic mass hierarchies we choose the following distribution of the fermion genera- tions and Higgs fields into matter this choice the m-term is realized through the coupling q1351-52which implies that hq13i0. With this constraint we write down the various operators for the top and bottom quark sector. We start again with dominant contributions to the up-type quarks. These generate the following mass texture y1J14vueytvu y4J15vu y7q31 y1J14vu y1J14vu ytvu y4J15vu y5J2 15vu y8q35 y6q34 y6q34 y7q31 y8q35 (4.26) The operators in Equation (4.25) with couplings also to the bottom sector of the model. In addition we have the following superpotential terms for the bottom we obtain the following down quark mass k2J15vd k2J15vd ybvd k2J15vd y6q34 y7q31 y8q35 bottom sector shares common operators between with the charged lepton sector of the model. These are given in (4.27) with couplings k1,k2andk8. We also have common operators between top and charged lepton sector which are given in (4.25) with couplings y6andy8. All the other from the the various contributions described so far we end up with the following mass matrix for the charged lepton sector of the model l2J13 l2J13 ytvd l2J13 y8q35 l3q35 l3q35l4q31q45 l5q45 (4.30) 5 Flavor violation the Z0gauge boson couples differently with the vector-like fields, new flavor violation phenomena might emerge and other rare processes could be amplified provided there is sufficient mixing of the vector-like fields with the SM matter ones [13, 14] . In order to examine whether the present models can account for the observed LHCb-anomalies we need to determine the unitary transformations that diagonalize the mass matrices of the models described in the previous section. Due to the complicated form of the various matrices we diagonalize them perturbatively around some small mixing parameter. We perform this procedure for model A while the analysis for the rest of the models is very similar. A detailed phenomenological investigation will follow in a future publication. 5.1 Some phenomenological predictions of model A To proceed with the analysis and discuss some phenomenological implications, first we work out the mass matrices and the mixing for quarks and leptons. Quarks: We start with the quark sector of model A and the matrix for the down quarks. In order to simplify the down quark mass matrix (4.1) we assume that some terms are very small and that approx- imately vanish. In particular, we consider that We further make the following the mass parameter Mcharacterizes the mass scale of the extra vector-like states whilst mvdis related to the electroweak scale. We have also assumed that the small Yukawa parameters are identical (ex). With the above assumptions the matrix receives the simplified 0 qmx q m m x20 0 amx2am m 0 0 bm bm x 0 cmM 0 0 0 M 01 CCCCCCA: (5.1) 14In summary, m;Mandmrepresent mass parameters while a;q;c;bandxare dimensionless coeffi- cients. Keeping terms up to first order in xthe mass matrix MdMT dcan be written q2m2x aq m2x 0 0 q2m2x q2m2aqm22bqm2x 0 aqm2x aq m2a2m2+m2abm2x 0 0 2 bqm2x a bm2x 0 0 cmM M21 CCCCCCA: (5.2) We observe that (5.2) can be cast in the form 0 0 0 0 0q2m2aqm20 0 0aqm2a2m2+m20 0 0 0 0 0 0 cmM q2m2aqm20 0 q2m20 0 2 bqm20 aqm20 0 abm20 0 2 bqm2abm20 0 0 0 0 0 local Yukawa parameter xcouples the electroweak sector with the heavy vector-like part and can be used as a parturbative mixing parameter. The block-diagonal matrix A, is the leading order part of the matrix and can be diagonalized by a unitary matrix For small values of the parameter a the eigenvalues of this matrix are written (5.4) We observe here that the eigenvalues appear with the desired hierarchy. The corresponding unitary matrix which diagonalizes the matrix Aand returns the eigenvalues (5.4), 0 0 0 0 0a2q2 21 aq 0 0 0 aq 1a2q2 20 0 0 0 0 cmM b2m2 1 0 0 0 (5.5) The columns of this matrix are the unperturbed eigenvectors v0 biof the initial matrix. Now we focus on the corrections to the eigenvectors due to the perturbative part xBwhich are given by the relation (5.6) 15where the second term displays the O(x)corrections to the basic eigenvectors of the leading order matrix A. The corrected diagonalizing matrices schematically receive the form similarly for the up quarks and leptons. This way the mixing parameter xenters in the expressions associated with the various flavor violation the eigenvectors using the formula (5.6), the O(x)corrected unitary matrix is 0 0 0 xa2q2 21 aq (5.7) We assume here that the mixing in the top sector is small and that the main mixing descends from the bottom quark sector. Charged Leptons: We turn now to the charged lepton mass matrix (4.6). We notice that some parameters from the top and bottom sector contribute also here so the same assumptions for these pa- rameters will be considered here too. Additionally, we assume that l1=l3=l4=l5=l70 and we make the following the mass parameter Mcharacterizes the vector-like scale and q,care dimensionless parameters. With these approximations the matrix receives the following form 0 qmx3amx a mx20 0 qmx a mx2am 0 0 0 0 0 cmM 0 mq m xq M 01 CCCCCCA: (5.8) We proceed by perturbatively diagonalizing the lepton square mass matrix MeMT e(M2 efor short) using xas the expansion parameter. Keeping up to O(x)terms we write the mass square matrix in the form 0 0 0 0 0 0 0 0 0 0 a2m20 0 0 0 0 c2m2+M2cmM 0 0 0 cmM 0 q2m2x0 0 0 0 0 0 am2xq q2m2x 0 0 0 am2xq 0 0 0 0 0 0 am2xqam2xq0 01 CCCCCCA: (5.10) The eigenvalues of the dominant part the unitary matrix V0 eL, which diagonalizes the dominant matrix 1 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 01cmM m2q2 0 0 0cmM m2q2 11 CCCCCCA: to the eigenvectors due to the perturbative part xBcan be found by applying the relation (5.6). Then for the final unitary matrix we 1 0acmm2xq M3 0 0 x 0 in order to simplify the final result we have assumed the series expansions for small a,qand c keeping only the dominant terms. B-meson anomalies at LHCb In the presence of a fourth generation where the U(1)0- charge assignments of its constituents differ from those of the SM families, many interesting rear flavor processes are expected to be enhanced 17and a detailed consideration will appear in a forthcoming publication. Here we shall focus only on the B-meson anomaly associated with the b!s``decay and in particular the ratio Due to the non-universal coupling of the Z0gauge boson with the the Cmm 9Wilson coefficient which contributes to the flavor violation transition b!sll, is (5.14) where the matrices as (5.15) with q0 fLbeing 55 diagonal matrices of elements in to Cmm 9coefficient can be obtained from (5.15) using the diagonalization matrices VfLcomputed above. We have (5.16) Finally, using the set of values Vtb0:99 and Vts0:0404 we estimate that the mass parameters in TeV units. Using an indicative set of values in eq.(5.17), we (5.18) According to the most recent global fits [24], an explanation of the current experimental data so in this model the Z0gauge coupling-mass ratio should be of the the model to explain the observed RKanomalies. This implies a rather small Z0mass [25] unless g0is associated with some strong coupling regime. Of course the computation of Cmm 9is very sensitive to the mass and mixing details of the representative model chosen in this example and a of whole set of models will determine whether sufficient mixing effects can predict the observed in B-meson decays, however this analysis is beyond the goal of the present work. 3For a discussion on the effects of complex valued contributions to the Wilson coefficients due to large CP-violation effects, see [23]. 186 R-parity violation terms A remarkable observation is that particular R-parity violating (RPV) terms such as l0 i the anomalies related to the b!s``flavor violating process [26, 27, 28, 29]. In this section we look for possible R-parity violating terms (RPV) in the tree-level here WRPV tree) for the models in Table 3 and briefly discuss their conse- quences. We distinguish the RPV terms in those which couple only the MSSM fields and those which share Yukawa couplings with extra vector-like families. If the former are present in WRPV tree, they lead to hard violations of baryon and/or lepton numbers and must be suppressed. In F-theory constructions, such terms can be eliminated either by judicious flux restrictions piercing the various matter curves [30], or by additional (discrete) symmetries emanating from the background geometry of the theory [31, 32, 33, 34]. In section 4 of [32] there are examples how this R-parity can be built. On the other hand, provided and conditions are fulfilled [26], such couplings may contribute to the B-meson anomalies and other interesting effects like the (g2)manomaly [35, 36] without exceeding baryon and/or lepton number violating bounds. Below, for each one of the five classes of models, among all possible superpotential couplings we single out the RPV terms. Taking into account that in all the models presented so far the up-Higgs doublet is isolated at the 5 1matter curve, the possible RPV operators of the 10
5
5 form are (6.1) Next, we discuss each model A . Using Table 3 and taking into account (6.1) we find that the only RPV term of the model (6.2) We observe that R-parity violation occurs with terms that involve the vector-like family and there are no terms which have only the three quark and lepton families of the MSSM. However, as recently shown in [28, 29], the coupling L4Q3dc 2can have significant contributions to the b!smmprocess through photonic penguin diagrams. Model B . Following the same procedure we found that this model contains the following RPV (6.3) The first operator here does not contribute to b!sllprocess due to the absence of the second gener- ation quark in the coupling. On the other hand the term L3L4ec 2which descends from the second operator leads to non-negligible contributions to the anomalous magnetic moment of the muon [35]. Combining this with non-zero Z0contributions may lead to a sufficient explanation of the C, D and E . There are no renormalizable RPV terms in these models. So in this case, an explana- tion of the observed experimental discrepancies is expected from Z0interactions and through the mixing of SM fermions with the extra vector-like states. 197 Conclusions In this article we have expanded our previous work on F-theory motivated models by performing a scan of all the possible constructions predicting a complete family of We use U(1)0hypercharge flux to obtain the symmetry breaking of the non-abelian part and a Z2monodromy to guarantee a tree-level top Yukawa coupling. Moroever, we have imposed restrictions on the various flux parameters in order to obtain exactly three chiral generations and one vector-like complete family of quarks and leptons. In addition, demanding anomaly cancellation, we have found that there exist 192 models with universal U(1)0charges for the MSSM families and non- universal for the vector-like states. These 192 models fall into five distinct classes with respect to classified as A;B;C;D;Ein the analysis. We have presented one for each class, exploring the basic properties, computing the superpotential terms, and construct- ing the fermion mass matrices. For the models derived in the context of class Ain particular, we have exemplified how these types of models can explain the observed RKanomalies through the mixing with the vector-like states without violating other flavor violation observables, by virtue of the universal nature of the three SM families. We also discussed the presence of RPV couplings in these type of models and their possible contribution in the observed experimental deviations from the SM predictions. It is that due to the flux restrictions and the symmetries of the theory only a restricted number of the possible RPV terms appear in the models. This way, with a careful choice of the flux parameters, it is possible to interpret such deviation effects while avoiding significant contributions to dangerous proton decay effects. A detailed account of such new physics phenomena is beyond the present work and is left for a future R. Aaij et al. [LHCb], [hep-ex]]. [2] R. Aaij et al. [LHCb], Phys. Rev. Lett. 122(2019) no.19, 191801 [hep-ex]]. [3] W. Altmannshofer and D. M. Straub, Eur. Phys. J. C 73(2013), 2646 [arXiv:1308.1501 [hep-ph]]. [4] G. D'Amico, M. Nardecchia, P. Panci, F. Sannino, A. Strumia, R. Torre and A. Urbano, JHEP 09 (2017), 010 [hep-ph]]. [5] S. Bifani, S. Descotes-Genon, A. Romero Vidal and M. H. Schune, J. Phys. G 46(2019) no.2, 023001 [hep-ex]]. [6] A. Cerri, V . V . Gligorov, S. Malvezzi, J. Martin Camalich, J. Zupan, S. Akar, J. Alimena, B. C. Al- lanach, W. Altmannshofer and L. Anderlini, et al. CERN Yellow Rep. Monogr. 7(2019), [hep-ph]]. [7] K. Kowalska, D. Kumar and E. M. Sessolo, Eur. Phys. J. C 79(2019) no.10, 840 A. Crivellin, G. D'Ambrosio and J. Heeck, Phys. Rev. Lett. 114(2015), 151801 A. Crivellin, C. A. Manzari, M. Alguero and J. Matias, Phys. Rev. Lett. 127(2021) no.1, [hep-ph]]. [10] A. Karozas, G. K. Leontaris, I. Tavellaris and N. D. Vlachos, Eur. Phys. J. C 81(2021) no.1, [hep-ph]]. [11] M. Crispim Romao, S. F. King and G. K. Leontaris, Phys. Lett. B 782 (2018), [hep-ph]]. [12] P. A. Zyla et al. [Particle Data Group], PTEP 2020 (2020) no.8, 083C01 [13] S. F. King, JHEP 08(2017), 019 [hep-ph]]. [14] S. Raby and A. Trautner, Phys. Rev. D 97(2018) no.9, 095006 [hep-ph]]. [15] P. Arnan, A. Crivellin, M. Fedele and F. Mescia, JHEP 06(2019), 118 [hep-ph]]. [16] F. F. Freitas, J. Goncalves, A. P. Morais and R. Pasechnik, JHEP 01(2021), 076 C. Beasley, J. J. Heckman and C. Vafa, JHEP 01(2009), 059 [arXiv:0806.0102 [hep-th]]. [18] E. Dudas and E. Palti, JHEP 09(2010), 013 [arXiv:1007.1297 [hep-ph]]. [19] S. F. King, G. K. Leontaris and G. G. Ross, Nucl. Phys. B 838(2010), 119-135 S. Cecotti, M. C. N. Cheng, J. J. Heckman and C. Vafa, [arXiv:0910.0477 [hep-th]]. [21] G. K. Leontaris and G. G. Ross, JHEP 02(2011), 108 [arXiv:1009.6000 [hep-th]]. [22] P. Langacker and M. Plumacher, Phys. Rev. D 62(2000), 013006 [hep-ph]]. [23] A. Carvunis, F. Dettori, S. Gangal, D. Guadagnoli and C. Normand, [hep-ph]]. [24] M. Alguero, B. Capdevila, S. Descotes-Genon, J. Matias and M. Novoa-Brunet, S. Dwivedi, D. Kumar Ghosh, A. Falkowski and N. Ghosh, Eur. Phys. J. C 80(2020) no.3, [hep-ph]]. [26] W. Huang and Y . L. Tang, Phys. Rev. D 92(2015) no.9, 094015 [hep-ph]]. [27] S. Trifinopoulos, Eur. Phys. J. C 78(2018) no.10, 803 [hep-ph]]. [28] Q. Y . Hu, Y . D. Yang and M. D. Zheng, Eur. Phys. J. C 80(2020) no.5, 365

2108.10989

Athanasios Karozas

A. Karozas, G. K. Leontaris and I. Tavellaris

models with a vector-like fermion family

23 pages, matches published version, references updated

Universe 2021, 7(10), 356

10.3390/universe7100356

hep-ph

http://creativecommons.org/licenses/by/4.0/

Motivated by experimental measurements indicating deviations from the Standard Model predictions we discuss F-theory inspired models, which, in addition to the three chiral generations contain a vector-like complete fermion family. The analysis takes place in the context of $SU(5) imes U(1)'$ GUT embedded in an $E_8$ covering group which is associated with the singularity of the elliptic fibration. In this context, the $U(1)'$ is a linear combination of four abelian factors subjected to cancellation conditions. Furthermore, we require universal $U(1)'$ charges for the three chiral families and different ones for the of the vector-like representations. Under the we find 192 such models which can be classified into five with respect to their specific GUT properties. We examples for each such class and construct the and the fermion mass matrices. We explore the implications of the vector-like states in low energy phenomenology including the the B-meson anomalies. The r^ole of R-parity violating terms appearing in some particular models of the above construction is also discussed.

give rise to fascinating emergent behavior such as strain stiening, evolving anomalous frictional response [14{24]. These behaviors can potentially aid organisms in balancing multiple complex and often contradictory functions such as locomotion with protection, and softness with stiness. Such possibilities have resulted in numerous studies in the past to understand the na- ture of property enhancements brought about by scales. Early research highlighted and conrmed the outstanding behaviors of 1-dimensional beam-like substrates that were cov- ered uniformly with scales using a combination of analytical and nite element (FE) models [14, 21, 25]. These works established precise relationships in pure bend- ing loads for both smooth and rough sliding between scales, and for both rigid and exible scales [14, 15, 25, 26]. The essential characteristics of bending behavior such as strain sti- ening and locked states were found to be universally valid even when sti scales were not uniformly distributed (e.g. functionally graded [18]) or loaded under non-uniform bending [17]. The locked state is essentially a kinematic conguration beyond with scale motion would result in interpenetration leading to rigid behavior. In reality, somewhat before lock- ing commences, the contact forces on the scales would be large enough to lead to Thus, bending behavior would transition from substrate de ection (soft) to scale leading to a sharp increase in rigidity of the overall structure. Re- cent studies have further conrmed locking and nonlinear strain stiening phenomena under pure torsion, thereby extending their validity even in twisting [21, 22]. However, torsion calculations while underlining the universality also highlighted the strik- ing dierences from bending. For instance, the kinematic locking envelopes turned out to be a complex nonlinear functions of geometry, unlike the bending case where they were linear. In addition, the tilt angle of the scales, Fig. 1 (b-d) had a signicant impact on the nature of locking, with some angles even precluding locking behavior. When Coulomb friction was included between the sliding scales, more dierences with the bending case were evident [15, 22]. For instance, frictional locking was indeed common between bending and twisting, but in twisting, the frictional locking envelopes were highly nonlinear without closed form solution unlike the bending case. Also, the relative energy dissipated from friction during one cycle of loading (start to lock) was found to monotonically increase with the ( ) for twisting case, even though for bending case, increasing did not always increase dissipated work in a cycle [15, 22]. Such dierences are well anticipated as the me- chanical behaviors are dependent on the interplay of structure, load and geometry with no intrinsic guarantees of universality. Thus, it is important to investigate individual canonical load cases carefully and rigorously before declaring either generalities or the mechanics of combined bending and twisting has not been studied. There is virtually no knowledge of the behavior of slender sh scale inspired substrates in 3-D or spatial de ections, which is signicantly more complex but of great practical utility due to the nature of real world loads and possibility of geometrical defects in the leading to cross curvatures (e.g. sagging or intrinsic twist). Prior research outlined above show that existing models may not be universally valid, scale linearly or even hold similar functional forms across geometry-loading combinations. Thus entirely new models are necessary to arrive at mm(a) (b) 20 mm 20 mm (c) (d) Figure 1: (a) Natural sh scales under deformation mode, adapted under CC BY 2.0 [27]. (b) scale metamaterials under bending deformation. (c) Fabricated biomimetic scale twisting deformation. (d) Fabricated biomimetic scale metamaterials under combined bending and twisting deformation. We address this lacuna by developing the mechanics of a frictionless biomimetic scale- covered beam of the geometric form shown in Fig. 1 (b-d). We use a Cosserat rod model to reveal the nature and regimes of nonlinearity, the interplay of bending and between global (substrate) and local (scale) deformations, and locking behavior. In the process, we obtain the relationships. We use FE simulations and prior results in the literature to validate our model. This paper is organized as follows: in Sec. 2, we develop the kinematics of the lament in bending and twisting loads, and derive the moment-curvature relationships, assuming small strains and additive strain energies. In Sec. 3, we brie y describe the nite element model we use to validate our theory. In Sec. 4, we conclude the paper with results and discussion. 2. Contact Mechanics of Scales Engagement We develop a kinematic description of the scale-covered substrate whose schematic is shown Fig. 2 (a). The challenge in describing kinematics of this system is the congurations of the scales and their rotation with respect to the sub- strate. A Cosserat kinematic description, commonly used to model the nite elasticity of rods, is ideal in the present context. Although the structure shown in Fig. 2 has nite 3width, any variation of global strains along the width (i.e., along the x-direction) has been neglected. We therefore model the substrate as an idealized rod. Since Cosserat a natural director basis at every point along the rod, the orientation of the scale relative to the substrate is easy to describe using this basis. In this way, the scales' orienta- tions naturally couple to the bending strain variables of the Cosserat description. of the Cosserat framework is that modeling the mechanics of these straightforward. We add the strain energies of the substrate and the for scales; see Sec. 2.4. This work will assume the substrate's constitutive law to be linearly elastic which assumes small material strains in the substrates. This has been found to be sucient in the practical contexts of relatively thin substrates that tend to lock before very large strains are permitted. More complicated constitutive models can be [28, 29] within our framework. Other simplifying assumptions and the general validity of the energy formulation have been discussed elsewhere[16, 21]. 2.1. Global Cosserat the schematic of the system shown in Fig. 2 (a). We model the substrate as a Cosserat rod [28, 29] whose centroidal curve in its undeformed reference conguration is dened by R(s), wheresrepresents the arc-length along the undeformed conguration. As shown in the gure, the planar undeformed substrate points along the z-direction, i.e., R(s) =se3, where e1,e2, and e3are the standard Cartesian basis vectors along the x, y, andzdirections, respectively. We identify the reference directors of the undeformed rod (Di; i= 1;2;3) with the cartesian basis vectors, i.e., Di=ei. Let r(s) denote the deformed position of the centroidal curve (shown in Fig. 2 (b)), and d1(s),d2(s), and d3, the orthonormal directors moving along the deformed rod. In the Cosserat description, the directors of the deformed conguration are related to those of the undeformed conguration via an orthogonal (rotation) matrix Q(s): di(s) =Q(s)ei;fori= 1;2;3: (1) with respect to sand substituting for eiusing the same, we obtain d0 i(s) =Kdi;fori= 1;2;3; (3) is a skew-symmetric matrix associated with the bending and twisting strains in the rod. We can associate K(with respect to the fd1;d2;d3gbasis) with an axial vector, 01 21 01 A; (4) and write (2) 1;2;3: (5) 4The strain variable 3is interpreted as the twisting strain in the rod, while 1and2as the bending strains for bending about the xandyaxes, respectively. Note that the denition of bending and twisting strains in the Cosserat sense described above is distinct from the actual 3D strains in the substrate. In Cosserat kinematics, the rod is modeled as a 1D curve, and Cosserat strains can be interpreted as the (3D) strains averaged over the cross-section of the substrate [28]. This approximation is justied for thin structures. In the cases of pure bending [14] or pure twisting [21], we identify the Cosserat strains with the twist rates and curvature of the rod, respectively. However, the matrix description of the strains noted in (4) is appropriate in the general case considered here. Henceforth in this paper, we use the term `strains' strictly in this Cosserat sense . We assume spatially homogeneous strain in this model analogous to prior studies [14, 25]. This makes Kindependent of sand helps set up periodicity conditions to extract relationships. For sharp gradients in strains or functionally graded local periodicity [18] or discrete scale by scale approach can be used [17]. The peri- odicity condition allows us to study the kinematics using a representative volume element (RVE) of a pair of scales [14, 25]. The strain variables 1,2, and3are viewed as parameters. Equation (3) can be explicitly integrated to solve for Q(s) and we obtain Q(s) =esK; (6) where without loss of generality we have assumed that the scale at s= 0 is xed and does not change its orientation as the scales deform. That is, Q(0) = (b) Figure 2: (a) Schematic of at reference conguration of the scale-covered rod with arc length variable along the length. (b) Schematic of deformed conguration of the same rod under coupled bend-twist load. Besides bending and twisting, the rod would more generally experience shearing in and stretching in the z-direction (we neglect other stretching and shear modes due to the substrate's thin cross-section). In Cosserat kinematics, strain variables associated 5with the above three modes of deformation are given by 1;2, and3, respectively. These are related to the deformation by r0(s) (7) If we use (1) to express diin terms of Q(s) and ei, and after integrating the we obtain r(s) (8) where we have used the boundary condition r(0) = 0, that is, the rod is xed at the origins= 0. In this work, we assume inextensibility and unshearability of the rod and set1=2= 0 and3= 1. These are reasonable approximations for the current case as soft polymers tend to be incompressible and the cross sections are not known to In the FE simulations presented in Sec. 3, we approximate inextensibility by choosing the Poisson's ratio = 0:49. Numerical simulations presented in Sec. 4 seem to conrm this approximation, as we observe no appreciable length change of the rod under the combined loading. Neither is any signicant change in cross-sectional area observed. Note that in spite of these assumptions, there is little loss of generality. If the model presented were to be applied for design and optimization of compressible substrates, the above easily be relaxed. In that case, the position of the rod must be determined using (8) with appropriate values for parameters 1,2, and3. 2.2. Local Kinematics of Scales We assume that the scales are identical in shape (rectangular with exposed length land width 2b), equally spaced along the rod, and are much stier than the underlying substrates and thus assumed to be rigid. In addition, the assumption of periodicity of contact (i.e., the use of RVEs) is also common [14, 21]. Non-periodic or functionally graded systems have previously for the cases of pure bending and pure twisting and have indicated the salient features of these systems are still preserved under periodicity [17, 18]. In their reference conguration, the scales are oriented parallel to each other with the in terms of the dihedral angles 0and0; see Fig. 2 (a). Here, 0is the angle made by the scale with the z-axis, and0is the scale's tilt angle made by its base with thex-axis. The separation of two adjacent scales measured along the centroidal curve of the undeformed conguration is d. As the rod deforms, the scales rigidly rotate with the rod (maintaining orientations 0and0with the deformed centerline) until contact between adjacent scales occurs. When the scales are under contact, they continue to remain rigid, but the contact induces a change in their orientations, viz., toanddue to scale rotation on the substrate. We now derive conditions for contact between adjacent scales. Let sirepresent the position of the scale ialong the arc-length of the undeformed rod. Let iR3be the set of all points constituting scale iin the reference conguration, and !iR3be the set of 6all points on the same, but in the deformed conguration. These are shown schematically in Figs. 2 (a) and (b), respectively. The midpoint of the base of the scales are located at Ri:=R(si) in the reference conguration and at ri:=r(si) in the deformed that for prescribed strains i(i= 1;2;3),r(si) is given by (8) where si=i
d. It is clear from Fig. 2 that for any given Xi2 iandxi2!i, the vectors iand!i, respectively. Since Nis normal to the scale i(in its reference follows that N
[XiRi] = 0: (9) Since the scales rotate rigidly under deformation, we (10) where Qi:=Q(si) =Q(i
d) is the rotation matrix given by (6) evaluated at s=i
d. Two adjacent scales (say, i= 0 andi= 1) in their deformed congurations intersect if and only if they have at least one common point. That is, if and only if the following equation has at least one (11) Using (10), the previous equation can be equivalently written (12a) where we have used Q0=I, and R0=r0=0to account for the boundary condition of scalei= 0. Since X0andX1lie on (nite) planes 0and 1, respectively, these points must satisfy N
X0= 0;forX02 0; 0;forX12 1: (12c) Note that solutions, X0andX1, of (12) are the 3D coordinates of the points of in- tersection of the scales, pulled back to the reference conguration. From a while it is easy to solve this linear system of equations, verifying if the solutions lie on nite planes 0and 1is not as straightforward. This is because of two of 1and 2, and their complicated 3D orientation in R3. We deal with both of them by employing the following change of i= 0;1; (13) where ^X0and ^X1are rotated coordinate variables and T=0 @cos0 sin 0 1 0 sin0 cos1 A; (14) 7is a rotation matrix. The transformation rule (13) maps ito^ i:=T( i), where the latter's projection on the ^Xi-^Ziplane is shown in gray in Fig. (3). Thus, Trotates the scales such that ^ 0=^ 1=B:= (15) Figure 3: Schematic showing the coordinate system for ^Xi. The dashed lines represent the edges of the substrate and the gray region is the projection of the scale on the XZplane. For each scale i, matrix T is a coordinate transformation from XYZ (12) in terms of ^X0and ^X1and using (13), we obtain the following for the intersection of the two adjacent (16a) NTTT^X0= 0; (16b) NTTT^X1= 0; (16c) where ^X0;^X12Band superscript Tdenotes matrix transpose. Thus, transformed variables ^Xilie in a simple rectangular domain B. Transforming the variables in this manner also resolves the issue of niteness noted above because it is straightforward to check if solutions lie inBby checking the bounds on the variables. 2.3. Contact that (16) is an under-determined linear system with ve equations in six un- knowns. If ^X0and ^X1remain unrestricted (i.e., ^Xi2R3, instead ofB), then we have two possibilities|no solutions, when the two planes containing the scales are parallel, or innitely many solutions with a line of intersection. However, since the scales constitute nite planes, since ^X0and ^X1are restricted toB. When nite planes intersect, there are 8the following possibilities: a) no solution, when the planes are either parallel or do not in- tersect within the domain, b) innitely many solutions when the two planes other with a line-segment intersection c) unique or innitely many solutions (depending on the relative orientation of the scales), but the scales do not interpenetrate each other. The third case is of interest here as they correspond to physically realizable congurations of scales making contact. A schematic for these congurations is shown in Fig. 4 (if any) shown in red. Figure 4: Various possibilities for two nite rectangles to intersect. The left-most gure shows no contact between scales, the next shows interpenetration of scales. The last three are a non-exhaustive sampling of possible modes of contact. Intersections are shown in red. There are various ways in which the third case of intersecting yet non-penetrating scales may be realized. Observe that in any of these congurations there is at least one point of intersection that falls in the following two cases: 1) it is either a corner of one of the scales (corner contact ) or 2) is a point of intersection of two edges of dierent scales ( edge contact ). See last three gures in Fig. 4). To nd the intersection points for the rst case, we prescribe corner coordinates ^Xiand ^Zi(for scalei) and for the second case, any two coordinates but from dierent scales, e.g., ^Xiand ^Zj(i6=j) are prescribed. We use the bounds for the corresponding variables appearing in (15) to x these values. Since ^Yiis unbounded we do not x this coordinate. The two cases are not necessarily mutually exclusive, as some congurations can be both. For instance, when corner contact happens on an edge; see third schematic in Fig. 4. The cases described above do not automatically congurations. For example, the second schematic in Fig. 4 is an edge- contact. In computations, we exclude such cases by explicitly checking for are eight possible corner contacts depending on which one of the eight edges (between the two scales) makes contact with the other. And there are sixteen edge which one of the four edges of rst scale makes contact with the four of the other. Of these twenty four possibilities most correspond to congurations with extreme twists that lead to appreciable material nonlinearities or even cause the scales to debond from the substrate. Therefore, we only consider the following four cases, shown schematically in Fig. 5: 1. Corner contact: Either of the two top corners of scale `0' makes contact with scale `1'. 9(a) (b) (c) (d) Figure 5: Modes of contact for an RVE considered in this work: (a) Corner contact (red) (b) Top-left edge contact (black) (c) Top-right edge (green) (d) Top edge (blue). Contact point is highlighted as a red dot For such a conguration, we set ( ^X0;^Z0) = (b;lcos). This is shown as red in our schematic Fig. 5 (a). In practice, ^X0=bis found only when <0. 2. Top-Left edge contact: This happens when the top edge of scale `0' makes contact with left edge of scale `1'. For this, we set ( ^X1;^Z0) = (b;lcos). This is shown as black in our schematic; see Fig. 5 (b). 3. Top-Right edge contact: This happens when the top edge of scale `0' makes contact with left edge of scale `1'. For this, we set ( ^X1;^Z0) = (b;lcos). This is shown as green in our schematic; see Fig. 5 (c). 4. Top-edge contact: This happens when the top edge of scale `0' makes contact with top edge of scale `1'. For this, we set ( ^Z0;^Z1) = (lcos;lcos). This is shown as blue in our schematic; see Fig. 5 (c). These color codes will be used later in Sec. 4 . All the above cases can be as the (17b) where a0;a1;b0;b1are 3D vectors that select the appropriate coordinates of ^X0and ^X1that we wish to x, and AandBare the bounds on the coordinates. For example, for the corner contact case, i.e., Case 1, noted above with ( ^X0;^Z0) = (b;lcos), (17), we havea0= (1;0;0)a1= (0;0;0)Tandb0= (0;0;1);b1= (0;0;0)T(which pick ^Z0, respectively), and A=b,B=lcos. For Case 2, ( ^X0;^X1) = (b;b), (0;0;0)T,b0= (0;0;0);b1= (1;0;0)T,A=b, andB=b. To obtain an implicit expression for the dependence of K(and therefore, 1,2and3) on, we solve (16) simultaneously with (17). We do this in the following steps. First, we solve (16a) for ^X1, i.e., We then plug the previous (17a) and (17a) to obtain equations only involving ^X0. Taking the together with (16b), we obtain a 3 3 system of linear equations A^X0=c, 1TTa1)T (b0+TQ (18) Plugging the solution ^X0=A1calong with the solution we obtain the following implicit dependence of in terms of K: f(K;) 0: (19) The dependence of fonKandcan be gleaned by observing that Q1andr1depend on K((cf. (6) and (8)), while N,AandBdepend on. Parameters appears in T, andland bappear inA, andB. Equation (19) is a highly nonlinear `bridging' law linking the local to the global kinematics completing the multiscale kinematics description of the system. We solve this equation numerically to obtain the dependence of on the bending strains, i. Solutions corresponding to interpenetrating congurations are discarded. 2.4. Mechanics of Biomimetic Scale Slender Substrate In order to understand the mechanics of this structure, we take recourse to energy balance between the global loads and local deformation. We model the underlying substrate as an Euler elastica whose strain energy is given respectively the bending and twisting rigidities. For beams with circular cross sections, the Young's respective area moments, Gis the shear modulus, and Jis the polar moment. In our case, two complications arise. The embedding of the scales lead to an inclusion eect on the beam, resulting in additional stiness (composite eect). For bending deformation, we postulate a scaling parameter Cfto the bending rigidities, i.e. values of the parameter Cfcan be obtained using FE simulations [21]. For the a non-circular cross section also introduces warping eect. Although warping is not substantial to aect the kinematics of the scales engagement, it does have an eect on the energy. This is addressed using another multiplicative factor Cw, which is in handbooks [30] and often scaled with respect to EI1rather than GJ. Thus, the twisting rigidity takes the form D=CfCwEI1. To include the contribution from the scales, we note that a scale's rotations is captured by the angles and. As the substrate deforms under applied strain, the scales rotate freely until a critical threshold of curvatures is reached when the scales engage. Let eR3 denote the bending strains ( 1;2;3), for which the scales are engaged. This region can be determined by solving (19) for 1,2and3for which> 0and when scales make contact. We determined it numerically. Clearly, the region will depend on the parameters chosen. In Fig. 10 of Sec. 4, we show a plot for a slice of ealong2= 0. Since the scales are built 11into the substrate, under engagement (i.e., when scales make contact), their rotations are resisted by the substrate. We model the substrate resistance using linear torsional springs with the elastic energy stored in the springs for each scale given by: Escale(;) the spring constants of the torsional springs, and Heis the on = 1;if (1;2;2)2e, and zero, otherwise). The total energy per RVE can be additively written as: E(1;2;3) =Ebeam(1;2;3) +Escale(;): (22) Our FE simulations show that the change in from0is minimal even when the scales are engaged. This observation is in agreement with our earlier ndings for the cases of pure bending and twisting [14, 21]. We therefore drop its dependence from Eand henceforth x to0. Note that in light of (19), is itself a function of ( 1;2;3), hence the dependence ofEon the same. Kis related to the Young's modulus of the substrate ( E), scale thickness ( ts), inclusion length (L) and0. As we have shown in [14, 21] the following non-dimensional scaling (23) wherenis constant that we estimate using FEsimulations, CBis a constant, and f(0) is function of angle . For the results presented below, these were CB= 3:62,n= 1:55,f(0)1 [21, 22]. The moments in the three directions are computed by dierentiating (22) with respect to1,2, (24c) These relations can be computed numerically to obtain the moment-curvature relation- ships. Note that the derivative of the indicator function, Heis the (surface) Dirac which is zero everywhere except on the boundary of e. This term does not appear in (24) because it is multiplied by 0, which regularizes to zero precisely on the boundary of e. 123. Finite Element Analysis We build Finite Element (FE) models on ABAQUS/CAE 2017 (Dassault Syst emes) to validate the analytical model developed above. In the simulations, we model the substrate and the scales as 3D deformable solids. The substrate is a rectangular prismatic 200 mm) onto which a row of 19 identical scales are embedded on one side. In this assembly, scales are spaced d= 10 mm apart, oriented with angle of 0= 5with respect to the substrate's top surface, and angle of 0= 30with respect to the cross section. Since our analytical model applies to an RVE, where strains are constant, we chose our substrate to be suciently long so as to avoid edge eects. The substrate was modeled as a linear elastic material with an elastic modulus EB= 2:5 MPa and Poisson's ratio = 0:49. Thus, the shear modulus of the substrate is GB=EB 2(1+)= 0:84 MPa. Scales were modeled to be rigid with respect to the substrate by imposing rigid The contact between the rigid scales is modeled using the 'Surface to Surface Contact' algorithm of the software [31]. This contact option is typically used in literature for biomimetic scale problems of this type [14, 17, 18, 21]. The mechanical loads for bending and twisting were applied quasi-statically to the system and as boundary conditions. In our numerical studies, we set 2= 0. That is, we do not explore bending in the transverse direction to the substrate, which is considered trivial in the current context. To compare the combined eect of bending (with strain 1) and twisting (with strain 3) with the semi-analytical model presented above, we perform the simulation in two static steps with nonlinear geometry option (NLGEOM on) [31]. In the rst step, the bending rotations with the same magnitude but opposite directions were applied to at either ends. The magnitude of bending was increased linearly from 0 1LB=2, whereLBis the length of the substrate. The nal value could be attained only approximately as edge eects precluded a constant bending strain could not be maintained throughout the substrate. To ameliorate these eects we extracted an averaged strain in the beam far from the ends. More details on this procedure will be explained in Appendix A. In the second step, the bending rotations was xed at the nal value (1LB=2) and the twisting rotations were applied to the both end cross-sections with reverse directions, again linearly increasing from 0 to (approximately) 3LB=2 during the step time. A frictionless contact was applied to the scales surfaces. To test the reliability of the numerical results, we carried out a mesh convergence study. Sucient mesh density was found for dierent regions of the model conrming computational accuracy. Mesh convergence led to a total number of about 230,000 elements. Because of the complex geometry of the system particularly due scales inclusions in the substrate, the top layer of substrate was meshed with tetrahedral quadratic elements C3D10 and other regions were meshed with quadratic hexahedral elements C3D20. 4. Results and Discussion In this section we present results exploring cross-curvature eects, with various types of scales contact color-coded dierently in the plots. Without loss of generality, we take the 13initial scale angle 0= 0(grazing scales) since we are more interested in scale A 0>0 would only shift the plots forward till engagement occurs, with the following the same trajectories [14, 21]. We use the following =l=d,=b=d, and ^i=id,i= 1;2;3 to explore the eects. A typical ^3(twist) versus (scale angle) behavior (for a given ^ 1) is shown in Fig. 6. In this particular gure, we have used ^ 1= 0:23,= 3,= 1:25 and= 30. The presence of an initial bending can lead to additional scale contact types, not observed for pure twisting investigated previously [21, 22]. These various contact regimes are color coded in the twist-angle plot as follows: the black curve is the contact type where the top edge of the rst scale makes contact with the left edge of the second, the red curve where the corner of the rst scale makes contact with the face of the second scale, green curve the conguration where the top edge of the rst scale makes contact with the right edge of the second scale, and the blue curve represents the conguration where the top edges of the two scales make contact. We also provide animations of the (see supplementary material) to aid in better visualization of this system. A B 6: ^3vswith dierent modes of contact. Here the xed parameters are: ^ 1= 0:23,= 3,= 1:25 and= 30. Recall that in the case of pure twisting [21] (or pure bending [14]) the kinematic locking is typically identied in the plot at the point where the slope goes to innity. However, in the case of combined load, at locking, the slopes changes sign abruptly from positive to negative. This critical point is indicated by a blue triangle in Fig. 6 14where the curve is predicted to transition from green to red. However, in reality congura- tions beyond this critical point are physically impossible without violating This critical point is thus interpreted as locking in this case. This of locking is a unique artifact of combined load system, not found in pure bend- ing or twisting [14, 21]. In the subsequent analysis, we treat post-locking congurations as forbidden. For the parameter values explored in this work, we nd that the blue curves (which correspond to congurations with top edge contact) always occur as locked states. Hence, they are not explicitly shown in the gures below. In addition, contact regimes that correspond to very high twisting strains are also not considered as they are of little and dicult to achieve without introducing signicant material nonlinearity. We now explore in detail the cross coupling eects of one curvature over the other. In Fig. 7 (a) we show the eect of bending (^ 1) on the twisting (^ 3) versus scale angle ( ). The various values of ^ 1are shown beside their corresponding curves. This gure is from pure bending [14] or twisting [21] kinematics. Here, there are distinct regions of kinematics and they are dependent on the existing bending curvature, with distinct contact regimes emerging throughout engagement. Thus existence of bending can result in entirely new type of kinematic behaviors in twisting. The color coded plots are an indication that there would be kinks in twisting rigidity as contact regimes between scales undergo change as depicted, for instance in Fig. 6, inset C to D. This plot can also be used to note the sensitivity of twisting to existing bending strains. We nd that the sensitivity is quite high and thus any unintended bending either due to sample processing or loading asymmetries can drastically change the overall kinematics. Another remarkable feature is that, locking angle for the scale seems to be insensitive to bending. Hence, the locking envelopes computed from pure twisting would still hold in this case. Overall, bending strains have a dierential eect on the twisting kinematics, aecting the kinematics trajectory substantially while leaving the locked state unchanged. In the same spirit, in Fig. 7 (b), we highlight the eect of twisting on the bending behavior of the biomimetic scale beam. Here, we nd that a twist in the system also has a on the bending behavior. With increasing positive twist, the bending earlier, whereas the opposite occurs with negative twist. This trend re ects the fact that twisting on one side has an `opening' eect on the scale, where they move apart from each other compared to the other side which leads to a 'closing' eect. The slopes also change signicantly indicating a potential eect on the overall stiness of the system. The abrupt changes correspond to sudden changes in contact regimes and shown with dierent colors. Here again, interestingly the presence of twist does not aect the locking angle, analogous to the twisting case. The validation of these analytical results using FE models are discussed next. Fig. 7 (c) and (d) are the counter parts of Fig. 7 (a) and (b), respectively. In Fig. 7 (c) we compare the FE results (shown as dots) with the results of our model (shown as dashed lines) for various values of pre-set bending strains (^ 1) in the substrate. The kinks seen in the curves (close to ^3= 0:05) are precisely the transition from red to green curves shown in Fig. 7 when the mode of contact changes. Analogous results for versus ^1for dierent values of pre-set twisting strains ^ 3are shown in Fig. 7 (d). Note that the pre-set strains for the two 15-0.6 -0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 (b) -0.05 0 0.05 0.1 0 0.05 0.1 (d) Figure 7: (a) versus ^3for dierent values of bending strains (^ 1). (b)versus ^1for dierent values of twisting strains (^ 3). (c) FE simulations of versus ^3for dierent values of bending strains (^ 1) (d) FE simulations of versus ^1for dierent values of twisting strains (^ 1),= 3,= 1:25 and= 30. set of simulations approximate. This is because when the second quasi-static step (see Sec. 3) is performed, the strain values xed in during the rst quasi-static step does not remain xed. That is, as we twist a pre-bent beam, the bending curvatures change mildly under the twisting deformation. In Fig. 7 (c) and (d), we account for this change by the changing values of ^ 1and ^3, respectively. We now use our model to explore the twisting kinematics at a given bending strain, with changing geometric parameters, and. These parameters have shown to be critical in dictating the overall kinematics of pure twisting [21]. For this, we x the bending curvature 16at ^1= 0:1. First, we probe the signicance of overlap ratio on the overall kinematics of the system. In Fig. 8 (a), we plot vs ^3for dierent values of (for= 1:25,= 30and 1= 0:15). Here, we rst note the similarities with pure twisting case. Like pure to steeper slopes in the ^3plots. However, the twist the kinematic trajectories by introducing two distinct contact regimes. The abrupt changes in contact are more apparent for lower overlap ratio but are ameliorated at higher overlaps. Note that on the right side, the plots continue to increase in steepness until the locked states are reached. On the other hand, on the left side (negative twist), the decrease in slope ( atten) till they collapse into a point. This is the point of disengagement ( 3= 3) of the scales beyond which, they no longer remain in contact. A surprising outcome of this plot is that, the disengagement point seems to be independent of the overlap ratio. This can be understood if this point is visualized for an RVE, Fig. 9 (intersection between the two scales is shown in red). Evidently, at this point, the nature of contact is the entire edge of one scale moving over the adjacent one. Thus, at  3,itself is not well dened and the scale angle is purely determined by the dihedral angle of two adjacent intersecting planes that is dependent only on the applied strains. We see here again that in uences the angle of locking. We nd that this parameter has a signicant role to play in determining the overall kinematics of the system. Increasing  leads to an overall stier nonlinearity but at the same time can potentially change the overall contact regime sequence up until locking. Next, In Fig. 8 (b), we plot vs ^3for dierent values of(for= 1:25,= 3 and ^1= 0:1). The eect of the tilt angle is dramatic. For relatively small angles, i.e.  < 20, there is a relatively `sti' response on either direction of twisting. These re ect scales sliding on either direction of twisting. However, there is a marked anisotropy between the directions in both magnitude and contact regimes. More interestingly, at higher tilt angles such bi-directionally distinct stiness indicating the scales do not or only weakly engage in other other direction. The lack of engagement in the other direction can be visualized as `opening' of the scales in one direction vs closing. The eect of on locking is also pronounced. At lower tilt angles, the stiening eect disappears into a lockless behavior. This is due to scales sliding past each other without ever satisfying the contact constraints. Locking behavior emerges again for higher tilt angles. In Fig. 8 (c), we now look at the sensitivity of the bending kinematics in the presence of twist. For these plots the we keep the geometry of the substrates the same and now impose a twist of ^3= 0:1 before bending and keep the tilt angle = 30. Here, we see that the overall impact of higher is to increase the slope of ^1plot, similar to pure bending [14]. However, the presence of twist changes the nature of the ^1curve by introducing dierent contact regimes, very similar to the twisting case discussed earlier (Fig. 8 (a)). Similar to that case, the abruptness of the kinks diminish at higher , disappearing altogether when = 5. Thus, higher scale overlap has a suppressing eect on contact regime transitions. Interestingly, the scales remain engaged even at a negative bending value, due to the `closing' eect on scales due to twisting where scales are pushed closer to each other. However, eventually the scales again lose contact as bending on the other direction increases suciently, once again exhibiting the 'convergence' at the point of disengagement described 0 0.2 0.4 (b) -0.2 0 0.2 0.4 0 0.2 0.4 (d) Figure 8: (a) versus ^3for dierent values of (for= 1:25;  = 30;^1= 0:1). (b)versus 3for dierent values of (= 3; = 1:25,1= 0:1). (c)versus1for dierent values of (for = 1:25; = 30; 3= 0:1). (d)versus1for dierent values of (= 3; = 1:25,3= 0:1). We also investigate the eect of an existing twist on for various tilt anglesbut with xed = 3 in Fig. 8 (d). Here, we nd that at lower tilt angle, the contact regime transition is absent, appearing as tilt angle increases. This shows that the tilt angle is an important parameter that controls the complexity of contact regimes. Overall, the eect of pre-twist in bending is muted at lower tilt angles and higher overlap and more pronounced otherwise. The investigation above, set the stage to investigate the eects of twist- 18Figure 9: Visual depiction of the scales in contact at the point of disengagement as scales 'open' up at 3= 3. (for= 1:25; = 30;^1= 0:1) bend combinations. To put this bend-twist kinematics in more general context, we compute the curvature dependence of scale angles for a biomimetic scale beam of xed This results in a phase plot as shown in Fig. 10. The curvature limits are taken to maintain self consistency with respect to contact regimes studied. For higher twist levels, more contact possibilities would need to be included. The phase plot in Fig. 10 is by solving the kinematic relationships developed earlier connecting scale angles and at various values of bending and twisting strains. Figure 10: Slice of the engagement region, e, as dened in (21) for ^ 2= 0. Points lying inside the shaded region correspond to congurations where scales are in contact. The color bar shows the scale angle (in radians) for the corresponding conguration. Here 0= 5and= 30,= 3, and= 1:25. 19The phase boundary on the left is the limit of engagement, i.e. we do not have any scales engagement beyond this boundary. The phase boundary on the right hand side is the locking boundary. Any (^ 1;^3) combination is forbidden in this region. Thus, the region in between indicates the region of engagement. It is interesting to note that there is no inherent symmetry between bending and twisting. In other words, the eect of bending on twisting is fundamentally dierent from the converse. In addition, one can see that combinations can give rise to same scale angles. This angular degeneracy is quite remarkable and could have signicance for inverse we discuss the mechanical behavior of these systems. In Fig. 11 (a) we plot bending moment vs bending strains and discover the role played by the presence of twist for a given . Higher values of positive twist shift the engagement to earlier parts, whereas higher values of negative twist shift it in the opposite direction. Interestingly, unlike smooth plots seen for pure bending earlier [14], twist eectively changes the contact regime in terms of discontinuities and jumps in the plot indicating sudden changes in bending rigidity. When a twist is added in the positive direction, the neutral position (no bending strain) is fully engaged (due to the engagement brought about by the twist). The FE comparison with these plots is shown in Fig. 11 (b). Notice the validation region is for much smaller values of bending and twisting. This is due to inherent limitation of traditional FE software in simulating this type of system and already well known in literature [14, 21, 32]. The agreements are excellent with respect to the theoretical values. The observed deviations are expected both due to the complex nature of contact, imposing global periodicity, and also the diculties in keeping one curvature constant in an actual numerical simulation due to edge eects. 0 0.1 0.2 0 0.02 0.04 0.06 0.08 (b) Figure 11: Bending moment M1versus ^1for various values of ^ 3: (a) Plots for the analytical model (b) Plots using FE analysis with dashed lines showing theoretical see similar and even more dramatic eect in the twist-torque diagram in Fig. 12 (a). Here again, we see the role of the bending strains as the torque-twists which are smooth non- linear plots in pure twisting case [21] now turn into highly discontinuous plots with disparate twist modulus. Yet again, an existing bending strain can cause the neutral position (zero twist strain) to be pre-engaged due to bending. These highlight the appreciable dierences that can be brought about from cross-coupling eects. Fig. 12 (b) shows the companion FE torque-twist plots that also indicate excellent agreement with the analytical results. As in the case for bending, the deviations are mainly because of the diculties in keeping the cross curvatures xed for the duration of the simulation. The other sources of errors are possibly due to assuming xed curvatures as opposed to spatially varying curvatures arising in a FE simulation. Analogous to bending case, imposing global periodicity in scale engagements also leads to some dierences in behavior. Please note that many of these non ideal eects and sources of discrepancies have been discussed earlier [16, 17]. -0.2 -0.1 0 0.1 0.2 0 0.02 0.04 0.06 0.08 (b) Figure 12: Twisting moment M3versus ^3for various values of ^ 1: (a) Plots for the analytical model (b) Plots using FE analysis with dashed lines showing theoretical predictions. 5. Conclusions In this work, we addressed the cross coupling eects of bending and twisting in a biomimetic scale elastic beam for the rst time. Here the scales were plate like rectan- gular inclusions protruding at an angle from the surface of the elastic substrate. We nd highly intricate and often surprising eect of one over the other across the kinematics and mechanics. We quantied these eects by developing analytical relationships within the framework of Cosserat kinematics and global-local energy balance. This model reduced to the earlier developed model for pure bending and twisting in literature and was also validated 21with FE simulations. This study completes a signicant missing piece in the mechanics of sh scale inspired biomimetic system, which is of great practical importance in from high performance smart skins to soft robotic work was supported by the United States National Science Foundation's and Manufacturing Innovation, CAREER Award B. Wang, W. Yang, V. R. Sherman, M. A. Meyers, Pangolin armor: overlapping, structure, and mechanical properties of the keratinous scales, Acta biomaterialia 41 (2016) 60{74. [2] I. H. Chen, J. H. Kiang, V. Correa, M. I. Lopez, P.-Y. Chen, J. McKittrick, M. A. Meyers, Armadillo armor: mechanical testing and micro-structural evaluation, Journal of the mechanical behavior of biomedical materials 4 (5) (2011) 713{722. [3] C. LaTorre, B. Bhushan, Investigation of scale eects and directionality dependence on friction and adhesion of human hair using afm and macroscale friction test apparatus, Ultramicroscopy 106 (8-9) (2006) 720{734. [4] K. Michielsen, D. Stavenga, Gyroid cuticular structures in butter y wing scales: biological Journal of The Royal Society Interface 5 (18) (2008) 85{94. [5] M. J. Buehler, Nature designs tough collagen: explaining the nanostructure of collagen brils, Proceed- ings of the National Academy of Sciences 103 (33) (2006) 12285{12290. [6] U. G. Wegst, H. Bai, E. Saiz, A. P. Tomsia, R. O. Ritchie, Bioinspired structural materials, Nature materials 14 (1) (2015) 23{36. [7] K. Kert esz, G. Moln ar, Z. V ertesy, A. Ko os, Z. Horv ath, G. M ark, L. Tapaszt o, Z. B alint, I. Tam aska, O. Deparis, et al., Photonic band gap materials in butter y scales: A possible source of Science and Engineering: B 149 (3) (2008) 259{265. [8] J. Long, M. Hale, M. Mchenry, M. Westneat, Functions of sh skin: exural stiness and steady swimming of longnose gar, lepisosteus osseus, Journal of Experimental Biology 199 (10) (1996) 2139{ 2151. [9] J. Song, S. Reichert, I. Kallai, D. Gazit, M. Wund, M. C. Boyce, C. Ortiz, Quantitative of the armor of the marine threespine stickleback (gasterosteus aculeatus), Journal of 171 (3) (2010) 318{331. [10] S. H. Sadati, Y. Noh, S. E. Naghibi, A. Kaspar, T. Nanayakkara, Stiness control of soft for minimally invasive surgery (mis) using scale jamming, in: International Conference on Intelligent Robotics and Applications, Springer, 2015, pp. 141{151. [11] Y. Wei, Y. Chen, T. Ren, Q. Chen, C. Yan, Y. Yang, Y. Li, A novel, variable stiness robotic gripper based on integrated soft actuating and particle jamming, Soft Robotics 3 (3) (2016) 134{143. [12] E. T. Roche, M. A. Horvath, I. Wamala, A. Alazmani, S.-E. Song, W. Whyte, Z. Machaidze, C. J. Payne, J. C. Weaver, G. Fishbein, et al., Soft robotic sleeve supports heart function, Science 9 (373) (2017) eaaf3925. [13] J.-Y. Sire, P. C. Donoghue, M. K. Vickaryous, Origin and evolution of the integumentary skeleton in non-tetrapod vertebrates, Journal of Anatomy 214 (4) (2009) 409{440. [14] R. Ghosh, H. Ebrahimi, A. Vaziri, Contact kinematics of biomimetic scales, Applied Physics Letters 105 (23) (2014) 233701. [15] R. Ghosh, H. Ebrahimi, A. Vaziri, Frictional eects in biomimetic scales engagement, EPL 113 (3) (2016) 34003. [16] R. Ghosh, H. Ebrahimi, A. Vaziri, Non-ideal eects in bending response of soft substrates covered with biomimetic scales, Journal of the mechanical behavior of biomedical materials 72 (2017) 1{5. 22[17] H. Ali, H. Ebrahimi, R. Ghosh, Bending of biomimetic scale covered beams under discrete International Journal of Solids and Structures 166 (2019) 22{31. [18] H. Ali, H. Ebrahimi, R. Ghosh, Tailorable elasticity of cantilever using spatio-angular biomimetic scales, Mechanics of Soft Materials 1 (2019) 10. [19] H. Ali, H. Ebrahimi, R. Ghosh, Frictional damping from biomimetic scales, Scientic Reports 9 (1) (2019) 14628. [20] H. Ali, H. Ebrahimi, J. Stephen, P. Warren, R. Ghosh, Tailorable stiness lightweight soft with architectured exoskeleton, in: AIAA Scitech 2020 Forum, 2020, p. 1551. [21] H. Ebrahimi, H. Ali, R. A. Horton, J. Galvez, A. P. Gordon, R. Ghosh, Tailorable twisting of substrate, EPL (Europhysics Letters) 127 (2) (2019) 24002. [22] H. Ebrahimi, H. Ali, R. Ghosh, Coulomb friction in twisting of biomimetic scale-covered & Biomimetics 15 (5) (2020) 056013. [23] H. Ebrahimi, H. Ali, J. Stephen, S. Dharmavaram, R. Ghosh, Emergent mechanical properties of biomimetic exoskeletal metamaterials, in: Bioinspiration, Biomimetics, and Bioreplication XI, Vol. 11586, International Society for Optics and Photonics, 2021, p. 115860N. [24] H. Ebrahimi, H. Ali, J. Stephen, R. Ghosh, Fish scales: Primitive basis for modern (Europhysics Letters) 133 (6) (2021) 68001. [25] F. J. Vernerey, F. Barthelat, On the mechanics of shscale structures, International Journal of Solids and Structures 47 (17) (2010) 2268{2275. [26] F. J. Vernerey, F. Barthelat, Skin and scales of teleost sh: Simple structure but high performance and multiple functions, Journal of the Mechanics and Physics of Solids 68 (2014) 66{76. [27] J. Kemoole, (accessed September 5, 2020) (2017). [link]. URL S. S. Antman, Nonlinear Problems of Elasticity, Springer, 1995. [29] B. Audoly, Y. Pomeau, Elasticity and geometry: from hair curls to the non-linear response of shells, Oxford university press, 2010. [30] A. C. Ugural, S. K. Fenster, Advanced mechanics of materials and applied elasticity, Pearson V. Abaqus, 6.14 documentation, Dassault Systemes Simulia Corporation 651 (6.2) (2014). [32] A. Shaei, J. W. Pro, R. Martini, F. Barthelat, The very hard and the very soft: Modeling skins using the discrete element method, Journal of the Mechanics and Physics of Solids 146 (2021) 104176. Appendix A. Extracting Cosserat Strains from FE Data To compare the FE simulations with the analytical model presented in Sec. 2, we extract the Cosserat bending strains 1,2, and3from the FE data through the following steps. First, we extract the position vector for the mid-point of the top face of the beam from the numerical simulations. This vector is an estimate for r(s) as given in (8). The information of the directors d1(s) and d2(s) is extracted by subtracting the position vectors of center-line and the right-edge of the beam, respectively, from the estimate for r(s) and normalizing the resulting quantities to produce unit vectors. After validating inextensibility by verifying that the change in length of the rod is small (with typical percentage relative error
LB=LB 1%), we compute the third director using d3=r0(s), estimating the right hand side using nite dierences along the beam. It follows from (1) that rotation matrix mapping of the beam is given by Q(s) = where the directors are taken to be the column vectors of the matrix. The skew-symmetric matrix Kcontaining the bending strains 1(s); 2(s); 3(s) (cf. (4)) along the length of the beam is computed 23

2108.10976

Ranajay Ghosh

Sanjay Dharmavaram, Hossein Ebrahimi, Ranajay Ghosh

Coupled Bend-Twist Mechanics of Biomimetic Scale Substrate

10.1016/j.jmps.2021.104711

cond-mat.soft cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

We develop the mechanics of combined bending and twisting deformation of filamentous structure with protruding stiff embedded at an angle on the surface. We develop Cosserat along with scale contact constraints. This geometrically exact model allows us to bypass the limitations of typical finite inherent in these systems when deflections are large. The relationships reveal for the first time the combined effect of bending and twisting on a slender fish scale inspired substrate. The model subsumes previous models on pure bending and twisting but also shows phenomena that arise due to the coupled effects of these loads. This includes a new interpretation of kinematic locking behavior, multiple contact regimes, asymmetric sensitivities of one curvature over the other, and in the nonlinear moment-curvature and torque-twist the complex scale engagement patterns.

2 Authors Suppressed Due to Excessive Length Fig. 1: Results of a Google image query for dogs from behind . are now almost universally used in computer vision. Many variants on architectures have also been proposed for object detection [5,6,7], action recognition [8,9], image captioning [10], question answering [11], etc. In summary, object recognition breakthroughs have had a surprisingly large mul- tiplier eect for the whole of computer vision. Besides CNNs, much of this progress can be ascribed to the introduction of ImageNet [12]. With 1 million images covering 1,000 object classes, it is one of the largest datasets in the literature. Additionally, the ease with which ImageNet trained networks can be transfered to other tasks suggests that it en- ables the learning of universal visual representations . It could be argued that, at least for lower convolutional layers of modern CNNs, lters learned on Im- ageNet are applicable to most vision tasks. While ne-tuning on specic data may enable gains, these are usually marginal. In many cases, the simple use of ImageNet CNNs as eature extractors is a strong baseline. All of this suggests that ImageNet experiments enable fundamental conclusions about object recog- nition. For example, because modern CNNs achieve smaller error than human annotators on ImageNet, it has been claimed, both in technical [13,14,15] and popular [16,17,18] literature, that CNNs have superhuman object experience has shown that most datasets have biases [19]. Like many recent vision datasets, ImageNet is the product of data collection the wild { images are rst collected on the internet, then annotated on platforms like MTurk. Dataset collection in the wild is a major departure from classic vi- sion practice of collection the lab. Its main advantage is removing dataset biases ensuing from a limited number of object classes, objects per class, light- ing patterns, backgrounds, imaging systems, etc. However, there are at least two signicant limitations. First, images collected on the web have some biases of their own. This is because people who upload do a certain amount of self selec- tion, usually seeking to publish good images. Consequently, it is not easy to collect large quantities of images of objects in the wild under unpopular imag- ing settings , such as poor lighting, ncommon or improper object views, poor focus, motion blur, etc. These unpopular settings vary with the object class. While trees are photographed from all angles and distances, dog pictures tend to depict close-ups of dog faces. Fig. 1 illustrates this point by showing a set of images returned by Google image search for the query dogs from behind . This makes it likely that object recognizers learned in the wild will have much more diculty recognizing dogs from behind than Overcoming Dataset Collection Bias in the Wild 3 (a) (b) Fig. 2: (a) Example images collected per viewing angle with bounding boxes. Green highlights the frontal view. (b) Examples of varying levels of focus captured while hovering. A second important limitation is that the lack of explicit control over the imaging process makes dicult to determine ImageNet's biases. In fact, because the images used to evaluate performance are also collected in the wild, this can be impossible. A lack of dogs from behind in the test set makes it impossi- ble to know if a recognizer fails on such images. Even when the biases can be observed, it is nearly impossible to quantify them. Since Turkers are poor es- timators of variables such as object viewing angle (which we denote as pose), lighting angle, etc., it is dicult to recover these annotations a posteriori. This hampers research in topics such as pose-invariant or light-invariant object recog- nition. We believe that progress along these directions requires bringing back o the lab and enabling explicit control over variables such as object pose, distance, or lighting direction. Several previous works have ad- dressed these problems [20,21,22]. However, many of these eorts predate deep learning, aiming for datasets much smaller and less diverse than what is expected today. Camera domes or turntables were favored, which can be expensive and cumbersome to build and maintain. We believe that more scalable data collec- tion procedures, which can quickly be disseminated and replicated throughout the vision community, are needed to address today's 3: The proposed drone based data collection setup.In this work, we address this problem by propos- ing a new solution to scalable dataset collection the lab. This leverages the fact that drones are cheap, and easy to program, allowing the collection of controlled object datasets to be scaled up. We consider the problem of pose invariant recognition, and propose a drone based setup to collect images over a set of object views. As illustrated in Fig. 3 this consists of having a drone y around the object and at precise pose intervals. The inexpensive drone selected is commercially available, and the setup only requires tape and software. Hence, it can be easily replicated or scaled to dif- ferent environments. Since all code will be made publicly available, we believe that it could enable a community wide eort on in the lab data collection. This is not easy with turntables or camera domes, which are more expensive harder to build and maintain. We also demonstrate the feasibility of the proposed setup by introducing a new dataset of Objects Obtained With fLight (OOWL). Denoted OOWL500, it contains 120 ;000 images of 500 dierent4 Authors Suppressed Due to Excessive Length objects, imaged from 8 viewing angles. The images are labeled for both identity (object and object class), pose, and other metadata that facilitates a number of experiments discussed in the paper. OOWL500 has several unique properties compared to past pose datasets. First, it is larger in regards to its combination of number of classes and number of object instances per class. Second, unlike datasets collected with domes or turntables, there is a small camera shake eect that produces images with statistics dierent than those of ImageNet style datasets. While still to humans, motion blur and somewhat lower resolutions pose ad- ditional diculties to ImageNet trained classiers, beyond the lack of pose ro- bustness. This is an eect similar to adversarial attacks that are now commonly studied in the literature [23,24]. Hence, OOWL500 is a stronger challenge for ImageNet classiers than previous pose datasets. Third, OOWL500 has larger class overlaps with ImageNet than previous datasets and was designed to addi- tionally overlap with a synthetic pose dataset, ModelNet [25]. All of this enables several experiments that were not possible with previous datasets. We leverage all these properties to investigate numerous questions regarding the robustness of current object recognizers. A number of surprising from these experiments. These include the fact that ImageNet classiers have far weaker generalization than humans, that it is quite easy to mount ad- versarial attacks on current classiers by simply manipulating object pose, that pose diversity is as important as instance diversity for training object synthetic data is not very helpful in mitigating the limitations of in the wild, and that the augmentation of ImageNet with OOWL500 style images leads to more robust recognition. While some of these conclusions are not qualitatively surprising, the strength of the eects, i.e. the gains or losses in recognition accuracy, can be surprisingly large. This shows that much more research is needed in object recognition methods with good generalization across poses and higher robustness in general. We thus believe that the combination of our experimental observations, and the introduction of OOWL500, will further research in this area. 2 The OOWL500 dataset In this section, we present the proposed drone-based setup for image collection the lab and the resulting OOWL500 dataset. 2.1 Dataset collection setup Fig. 3 shows the proposed setup for image collection the lab. A drone ies around an object, following ight commands issued by a state machine imple- mented in software which relies on colored markings on the oor for navigation. The drone always takes o from the same 3D location and stops at orange circles on the ground. At these locations, it faces the object to take pictures at angles, producing the images shown in Fig. 2(a). The enclosure and the platform used to place objects upon are surrounded entirely by green recognition algorithms from exploiting background cues. Since theOOWL500: Overcoming Dataset Collection Bias in the Wild 5 objects are ultimately cropped, using the bounding boxes shown in Fig. 2(a), there are no strong background eects. The ight space is illuminated by six LED lamps attached to the approximately homogeneous illumination of the object under all views. As shown in Fig. 2(a), there can still be lighting variations on the object sur- face. This is due to surface curvature or other properties, not the existence of a preferred light direction. To prevent object recognition algorithms from ex- ploiting lighting patterns to ascertain pose, the placement of objects' front face was randomized in regards to which wall it faced. The procedure used to control the drone ight is facilitated by its two cameras. One faces the ground, and is used for navigation. The second faces the object, and is used to collect object views. Images collected by the ground facing camera are processed by standard OpenCV vision operators to detect the oor markings. As the drone approaches the object, a PID controller issues commands to adjust its altitude, center the orange circle in the ground view, and rotate towards the pink line's direction. Once in place, the drone hovers and takes 30 pictures using the frontal camera at a rate of 5 Hz. At this point, the drone uses the ground camera to follow the blue line to the next orange circle. This sequence of commands loops until the drone nishes ying around the object. The software is implemented so that new states can easily be programmed and inserted into the sequence in any order, allowing for many potential ight no further intervention is needed once the drone has been placed on the rst stop. A single button press instructs the drone to calibrate, take o, y around the object, save images, and land autonomously in about 5 minutes. A video of a typical ight is provided as supplemental material, and precise details about the setup as well as all code will be released upon publication of the paper. This will make the infrastructure easily reusable, as a black box, by anyone who intends to generate a dataset or augment OOWL500. Furthermore, the code is written to accommodate ight paths of any reasonable size and shape, as indicated by the oor markings. Because only a laptop, a drone, and colored tape suce to capture objects of many dierent sizes within a wide range of backgrounds and lighting conditions, we hope that these factors will result in a community wide eort to generate a very large dataset. Another interesting facet that we plan to explore in the future is that al- though this is an the lab setup, it can actually be deployed outside the lab. This raises the possibility of a mixed the lab and the wild setting, where the procedure is deployed outdoors against natural backgrounds but retains the ability to control variables like pose and distance to the object. The setup can even be scaled up to very large objects such as entire buildings or the choice to optimize the setup for low-end drones prevents us from doing this. We have so far used a Parrot A.R. Drone quadcopter because, at a cost of about $200, it comes equipped with an ultrasound altimeter, an ac- celerometer, a gyroscope, and two (forward and bottom facing) cameras used to facilitate data collection. This was primarily motivated by the desire of a very low barrier to entry for those interested in contributing to dataset collection.6 Authors Suppressed Due to Excessive (b) Fig. 4: Images from the eapot object class. Left: OOWL500. Right: ImageNet. This is unlike most previous the lab image collection procedures, based on turntables or domes, that are much more expensive and dicult to build. However, the current drone can be challenging to control outdoors. We plan to support other drones in the future. 2.2 Dataset The above setup was used to create the OOWL500 dataset, which contains 500 objects captured at varying angles, for a total of 120 ;000 images. The objects are evenly distributed among 25 classes of everyday objects, such as shoes, and teddy bears. To facilitate the experiments in Section 3, these classes have been chosen to maximize overlap with those of ImageNet [12] and ModelNet [25]. There are 20 objects per class, imaged from the 8 views (poses) of Fig. 3. Since 30 images are taken per view, each object originates 240 images, for a total of 120 ;000 images. As shown in Fig. 2(a), views are acquired in steps of 45deg. The 240 pictures are stored as 640 360 PNGs. Besides the richness of poses, OOWL500 has an adversarial aspect for clas- siers trained on previous datasets, due to the somewhat unorthodox nature of images collected by the drone. While several pictures portray the same object at the same pose, they vary in object position, blurriness, exposure, and noise. Shown in Fig. 2(b), this is due to camera shake that ensues as the drone hovers in place. Some images in the eapot class are shown in Fig. 4(a). Note the dierent eel of these images vs. teapot examples from ImageNet, shown in Fig. 4(b). The latter illustrate an ImageNet bias towards professional quality object shots, which results from the self-selection usually performed for images posted on the web. The camera shake eect of OOWL500 can be thought as a generalization of the standard manipulations, such as pixel-swapping, commonly used in adversarial attacks to CNNs [23,24] or image shifting and rescaling com- monly used to augment training sets. The variation due to hovering images for either attacking existing networks or augmenting training sets to build better ones. This trait is explored further in Section 3.2. Each image in the dataset is accompanied by a CSV le that provides the following annotated information regarding the object: a short description; the pose in degrees relative to other pictures of that object; whether or not the picture depicts the object's frontal face (denoted by a 1 or 0 respectively, or a -1 if the object does not have an easily identiable front face); the object's degree of rotational symmetry (indicated by the next degree increment where the object appears identical); and coordinate pairs specifying the bounding box.OOWL500: Overcoming Dataset Collection Bias in the Wild 7 Dataset Domain Classes Objects per clsImages per the wild Caltech-101 [26] General 102 90 1 9,144 7 Pascal [27] General 20 790-10,129 1 11,540 5 Caltech-256 [28] General 257 119 1 30,607 17 MNIST [29] Digit 10 6,000 1 60,000 0 CIFAR-100 [30] General 100 600 1 60,000 9 MS COCO [31] General 80 11,079 1 123,287 17 ImageNet [12] General 21,841 650 1 14,197,122 23 In the [32] Table Top 3 10 16 480 2 1 NYC3DCars [33] Car 1 3,787 1 1,287 1 1 EPFL Car [34] Car 1 20 115 2,299 1 1 ETH-80 [35] General 8 10 41 3,280 4 1 Multi-PIE [36] Face 1 337 15 5,055 0 0 COIL [20] General 100 1 72 7,200 - - FERET [37] Face 1 1199 1-24 14,051 0 0 PASCAL 3D+ [38] General 12 1158 3 30,899 10 6 NORB [22] Toy 5 10 1944 48,600 1 3 T-LESS [39] Industry 30 1 648-1,296 49,000 0 0 ALOI [21] General 1000 1 72 72,000 - 15 iLab-20M [40] Toy Vehicle 15 25-160 1,320 21,798,480 11 4 HM [41] General 51 6 750 250,000 19 7 OOWL500 General 25 20 240 120,000 23 [25] General 40 100 3D 3D 28 10 ShapeNet-55 [42] General 55 940 3D 3D 46 16 Table 1: Overview of well known object recognition datasets. For datasets with pose annotations, the number of overlapped classes with ILSVRC2012 are listed. This is not applicable for COIL and ALOI however (only individual objects are provided instead of classes). 2.3 Relation to previous datasets In the wild datasets Table 1 compares OOWL500 to previous datasets. The top third of the table summarizes the properties of the wild datasets. These will continue to have a critical role in object recognition; they cover a diversity of imaging scenarios, objects, scenes, and scene congurations not replicable in the lab. They are listed mostly for calibration of what constitutes a large dataset today. Note that, with less than 30 ;000 images, Pascal and the two Caltech datasets are now considered small in comparison to MS-COCO and ImageNet. At 60 ;000 images, MNIST and CIFAR are still widely used in machine learning, but less so in computer vision. This suggests a threshold of 60;000 images, marked by the horizontal lines in the table. The goal of this work is not to replace in the wild datasets. Rather, we seek to complement them by investigating two of their limitations. The rst is that they exhibit some biases resulting from image collection on the web. One example photography bias of Fig. 4(b). Since most people post images to convey a message, they often carefully self-select the images to post. Hence, objects can have standardized poses that makes them more easily recognizable or appealing. This helps explain the pose bias of Fig. 1. This bias is also visible on the ImageNet teapots of Fig. 4(b), which tend to be presented from a side view. The second limitation of the wild datasets is the lack of to study important questions in object recognition, such as invariance to pose, scale, lighting, etc. Consider pose variability. Under mild assumptions on imaging setup (e.g. uniform lighting), a rigid object spans a 3D manifold in8 Authors Suppressed Due to Excessive Length image space as the viewing angle changes. This suggests a decomposition and the pose information of the object. An object in principle, exploit this decomposition to improve its generalization, e.g. by leveraging a model of pose trajectories learned from some objects to better recognize others, for which less training views are available. This is easier if the dataset contains pose trajectories , i.e. multiple views of each object, and even more likely if the dataset is annotated with pose labels . None of the two hold for the wild datasets, which favor instance diversity. For example, 1,000 images of chairs but no dierent views of the same chair. Even if this were the case, pose labels would be dicult to recover on 3D+ NORB* T-LESS ALOI 10 100 1000 10000 100000Objects Per 5: Scatter plot of datasets listed in Table 1. In the wild datasets are shown in red, in the lab in green, and synthetic in purple. An asterisk indicates that turntables were used for the lab datasets The study of these questions has motivated several the lab datasets, anno- tated for factors like pose, distance, or lighting. In fact, as shown in the sec- ond third of Table 1, there is a long history of such datasets. However, most of them are below the 60 ;000 image threshold, and too small to train deep they all have signicant dierences com- pared to OOWL500. First, they exhibit a bias similar to that of in the wild datasets. Sometimes the bias is even stronger, because images are collected in a precisely controlled studio environment; they do not exhibit the camera shake found in OOWL500. Second, objects are often carefully se- lected from ne-grained classes, such as faces or toys. Due to this, many of these datasets have few classes of overlap with in the wild datasets like ImageNet (only HM comes close to OOWL500 in this regard). This prevents studies of robust- ness, such as the ones we pursue in this work, or the use of these datasets to augment ImageNet. Third, beyond total number of images, an object dataset should contain a number of classes suciently large to pose a challenging dis- crimination problem, and a number of object instances large enough to capture the diversity of appearances of each class. Even the largest in the lab datasets can be small in this respect. For example, while iLab-20M contains many images, it only includes 15 classes of toy vehicles. This is a relatively small On the other hand, HM contains 51 classes, but only a few objects (6 on average) per class. This is not enough to capture the appearance diversity of most object classes. Since dataset size is combinatorial on these factors, any dataset must achieve a good compromise between them. Fig. 5 visualizes thisOOWL500: Overcoming Dataset Collection Bias in the Wild 9 trade-o, by showing the number of classes vs. objects per class of each dataset in Table 1. Among in the lab datasets (shown in green), OOWL500 achieves the best trade-o between the two factors. Naturally, in the wild datasets tend to be larger in size. However, the scalability of the drone-based data collection process used by OOWL500, together with the potential for community wide collabora- tive collection, could push it to the red area of the plot. It is hard to see this happening with turntable or dome-based datasets Recently, there has also been interest in large datasets of object shape. These are usually composed of synthetic images, rendered from 3D graphics models. Two popular datasets in this class, ModelNet [25] and ShapeNet [42] are listed in the bottom third of Table 1. While object views Fig. 5 shows that they are not much larger than OOWL500 in terms of object classes or objects per class. Furthermore, these datasets have been primarily used for operations such as shape classication or retrieval on synthetic images. It is not clear whether they can be leveraged to improve real- world object recognition systems. We study this question in Section 3.3. 3 Experiments In this section, we report various experiments using OOWL500. 3.1 Experimental setup Datasets The goal of these experiments were to compare in the wild, in the lab, and synthetic datasets, represented by ImageNet, OOWL500, and Model- Net respectively. Given a 3D shape from ModelNet, 2D images were rendered as in [43]. A preliminary step was used to generate three datasets of compara- ble classes. Since ImageNet is more ne-grained than the others, several of its classes were merged in a larger class with the same number of images. For exam- ple, irliner and warplane were merged into an irplane class. The 1 ;300 training and 50 test images of each original class were randomly sampled (650 training, 25 test images per class) to produce a 1 ;300 (50) image (test) set. Appendix A and B present a list of the merged for each object class also found on OOWL500 and ModelNet, respec- tively. After this fusion, there was an overlap of 23 classes between ImageNet and OOWL500, and 8 between ImageNet, OOWL500, and ModelNet. To max- imize the number of classes per experiment, several new datasets were created. IN(23) contains the ImageNet images of the 23 classes shared with OOWL500, IN(8) of the 8 classes shared by all 3 datasets, and OOWL500 (8) the same for OOWL500. All experiments were performed with AlexNet [30], ResNet [4], and VGG [2]. Models pre-trained on the full ImageNet and ne-tuned on each of the IN datasets are denoted ImageNet classiers (INC) . For example, the classiers ne-tuned on IN(23). Models ne-tuned on OOWL500 and ModelNet are denoted OCand MNC classiers, respectively.10 Authors Suppressed Due to Excessive Length Fig. 6: Recognition rate as a function of .= 0 reports to camera shake, > 0 to an angle change of . Fig. 7: Test set accuracy, on OOWL500, of clas- siers trained on 8 training sets, as a function of the pose angles included in the training set. Frontal Pose Comparing recognition accuracy as a function of pose requires a pose reference, i.e. the 0oview or frontal pose . This is hard to dene for classes like airplanes. The most natural frontal poses (the plane's nose or side) are also uncommon on ImageNet, where airplanes are usually seen at 45 degrees. Because of this, INCs have dierent avorite poses for dierent classes, under the OOWL500 denition of frontal pose. To avoid bias towards OOWL500, we adopted a data-driven denition instead. In all experiments, frontal pose is the view of largest average recognition accuracy by the INC classiers for each class. 3.2 ImageNet bias We now report on experiments testing the usefulness of OOWL500 as a tool to 1) identify biases of the wild data collection and 2) help correct attacks There has recently been much interest on on CNNs [23,24]. These consist of perturbing an image in a way that does not aect human recognition and measuring the drop of recognition accuracy by the CNN. We considered an attack strategy on the INC(23) classiers using OOWL500 images. A random frontal pose of OOWL500, correctly classied by the CNN, was perturbed into another image of either the same or a dierent pose. Images of the same pose dier by camera shake (small perturbation), while those of dierent poses dier by view and camera shake (larger). The dierence of view angles is denoted the angle perturbation . This was repeated for all classes of OOWL500. Fig. 6 shows the recognition rates of the three classiers as a function of. The attacks are quite successful, producing a low recognition rate for most values of. They exhibit the two ImageNet biases' eects illustrated in Fig. 4. The drop to 80 90% recognition rate at = 0, where all perturbations are due to camera shake, is explained by the professional photography bias of ImageNet. The even lower recognition rates for larger are explained by its pose bias. Note that the latter is particularly strong, with a performance drop larger than 50 points between frontal and least favorite pose! Correcting ImageNet biases We next studied the usefulness of the lab datasets such as OOWL500 to correct ImageNet biases. A rst set of whether OOWL500 is a useful complement to ImageNet. These were inspired by previous claims of superhuman performance for the classication of ImageNet object classes. While this has been shown to hold when ImageNetOOWL500: Overcoming Dataset Collection Bias in the Wild 11 classiers are evaluated on the ImageNet test set [13,14,15], it has not been shown how their generalization to a more adversarial set up compares to that of humans. To evaluate this, we measured the recognition accuracy of the INC(23) classiers on IN(23) and OOWL500, and compared them to those of OOWL classiers and human performance. Human performances were evaluated by taking the majority vote of 3 individual Turkers, asked to classify test set images from OOWL500 and ImageNet. Since all classiers behaved similarly, we only report the average across the AlexNet, VGG, and ResNet architectures, shown in Table 2. The rst striking observation is that the INC(23) classiers performed quite poorly on OOWL500. While their performance on the IN(23) test set is only 9 points weaker than that of Turk workers, the dierence on OOWL500 is 46 points! Note that the dierences between OOWL500 and ImageNet pose no diculties to humans, which have similar performance on both datasets. This demonstrates the vulnerability of ImageNet classiers to camera shake and pose bias, the importance of datasets (such as OOWL500) that account for these factors, and the fallacy of claims of superhuman performance, even on ImageNet IN(23) OOWL500 ALOI INC 0.79 0.33 0.55 OC 0.44 0.77 0.71 Both 0.76 0.78 0.72 Human 0.88 0.87 - Table 2: Average recognition accu- racy of AlexNet, VGG, and ne-tunned on IN(23) (de- noted INC(23)), OOWL500 (denoted OC), and both (denoted both), on the test sets of IN(23), OOWL500, and ALOI. Experiments on IN(23) and OOWL500 are based on 23 classes, while those on ALOI are based on 13 classes. Human per- formance was measured by majority vote among three Turk workers.Note that the classier architectures are per- fectly capable of handling OOWL500. When ne-tuned on the latter, the (OC) good performance on this dataset. How- ever, there is a signicant drop of accuracy on IN(23). This is a well known problem, causes the classier to orget the original dataset [44,45]. These results show that OOWL500 is a good testing ground for future research in this area. An interesting perhaps the most surprising of this experiment, is that the OC classiers, ne-tuned to OOWL500, generalized better to IN(23) { 44% accuracy { than the INC(23), ne-tuned to IN(23), generalized to OOWL500 { 33%. Finally, the classiers ne-tuned on both IN(23) and OOWL500 nearly matched the ac- curacy of the classiers ne-tuned to each dataset. While this was expected, it shows that complementing ImageNet with OOWL500 has no degradation on the ImageNet domain, but can substantially increase performance beyond it. Overall, this experiment supports several conclusions. First, claims of super- human performance regarding deep learning classiers trained on ImageNet are widely exaggerated. It is clear that the classiers do not even generalize for a dataset with the same object classes, when the images of the latter are not col- lected on the web. Second, the two datasets are clearly complimentary. When classiers are ne-tuned on both, overall recognition performance is far superior than when they are ne-tuned on only one. Third, the strongest transfer from OOWL500 to IN(23) rather than vice versa suggests that there is more than a simple mismatch of image statistics between the two datasets. Instead, it sug-12 Authors Suppressed Due to Excessive Length gests the existence of an ImageNet pose bias which hurts the ability of to generalize to data with the pose diversity of OOWL500. The last point is somewhat muddled by the fact that the classiers (OC) ne- tuned on OOWL500 were initially pre-trained on ImageNet. Their better general- ization to IN(23) could be due to the pre-training. To test this, the performances of INC and OC were evaluated on a third dataset ALOI [21] which exhibits sig- nicant pose variability and is completely independent of IN and OOWL500. Since it shares 13 classes with IN and OOWL500, IN(13) and OOWL500 (13) were created with the procedure of Section 3.1. OC(13) and IN(13) on these datasets, were nally tested on ALOI. As shown in Table 2, OC(13) generalized signicantly better than INC(13), achieving a gain of 16 points and performance nearly identical to that of the classier ne-tuned on both IN(13) and OOWL500. This conrms the need for in the lab datasets that, like OOWL500, enable training that specically corrects ImageNet biases. 3.3 Guidelines for data collection So far, we have seen that OOWL500 can be used to 1) construct adversarial at- tacks on and 2) improve generalization performance of ImageNet classiers. The nal set of experiments exploited OOWL500 to obtain data collection training strategies Computer generated synthetic datasets pro- vide a simple alternative to the lab datasets. In fact, recent works use synthetic data to improve the generalization of classiers [46] or solve problems such as shape retrieval [43,47]. To evaluate the potential of synthetic data aug- mentation, we ne-tuned ImageNet classiers on the combined training sets of IN(8) and MN and tested their recognition performance on OOWL500(8). To characterize sensitivity to pose, various experiments were performed. Starting from IN(8), all the images with the same pose of MN were gradually added to the ne-tuning set. The rst experiment included only frontal poses, then 45o views were added, and so forth. The recognition performance was recorded after each new pose set was added and is displayed in a cumulative plot in Fig. 7. Various transfer modes were compared. Modes IN and MN are baselines using ImageNet and ModelNet ne-tuning only, respectively. Mode IN+MN used the combination of the two datasets. To minimize the diculty of transfer due to dierences in the overall statistics of OOWL500, ImageNet, and ModelNet, we also considered a few calibrated modes where the ne-tuning set included the frontal pose of OOWL500. Mode Cal used only this set of images, modes IN-Cal and MN-Cal combined the IN and MN sets with the calibration set mode IN-MN-Cal used all ne-tuning modes. All these transfer modes were compared to transfer based on ne-tuning on the OOWL500 training set. Several conclusions can be made. First, synthetic data does nothelp the transfer at all. MN had very poor performance, which in fact decreased as poses were added. The addition of MN to IN also decreased the performance of the latter. Second, the biggest gains came from the calibration modes. The addition of the OOWL500 frontal pose signicantly improved on all the other congu- rations. While this was somewhat expected, the gain is quite signicant { atOOWL500: Overcoming Dataset Collection Bias in the Wild 13 (a) (b) Fig. 8: Recognition accuracy vs. dataset parameters O=C;P=O;E=P for a xed total training set size (D). (a) x axis shows accuracy by training with ( O=C;P=O ) = (16;1), which favors object diversity; y axis by training with (2 ;8), favoring pose diversity. (b) x axis shows accuracy by training with (E=P;P=O ) = (30;1), favoring camera shake; y axis by training with (3 ;8), favoring pose diversity. least 20 points. Again, the addition of MN poses had no noticeable increase in performance. Third, the best performance was obtained with OOWL500 alone. While the addition of IN(8) and MN improved on OOWL500 ne-tuning with the frontal pose only, there were no gains after that. Adding a single pose (45o) of OOWL500 to the ne-tuning set was enough to match the performance of adding the entire IN(8) and MN training sets! Adding one further pose (90o) raised the recognition rate to a level close to the overall best. In summary, show that 1) synthetic augmentation is not the solution to the prob- lem of ImageNet biases and 2) even a small number of in the lab poses guarantee major gains over the ImageNet diversity vs object diversity Object recognition has many dimensions of image variability, such as the diversity of pose or appearance of the objects in each class. For a xed dataset size, it is always necessary to establish trade-os for the coverage of all these dimensions. In the lab datasets such as OOWL500 enable the study of how this coverage impacts the quality of the resulting contains four such dimensions: the number of classes ( C), the number of object instances per class ( O=C), the number of poses per object ( P=O), and the number of examples per pose of each object ( E=P). For a xed dataset many congurations of these factors are possible. For example, ImageNet uses E=P =P=O = 1 and favors object diversity, i.e. large O=C. OOWL500 enables other strategies, such as pose diversity (largerP=O) or camera shake (largerE=P). Two experiments were performed to compare the relative performance of such strategies. In both cases, we xed the number of classes and training set size ( C= 23 andD= 11040 respectively). The rst experiment compared the importance of object vs. pose diversity. For this, we xedE=P = 30 and considered two extreme congurations. The rst, denoted max object diversity , used a ne-tuning set with the largest possible O=C = 16 and smallest P=O = 1. The second, denoted max pose diversity , used the smallest O=C = 2 and the largest P=O = 8. The recognition accuracy of classiers trained with these ne-tuning sets was then compared. The scatter plot of Fig. 8(a) illustrates the average recognition accuracies of the two strategies for all object classes in OOWL500. Points on the blue line14 Authors Suppressed Due to Excessive Length signal classes with equal recognition rate under the two strategies. Points above (below) signal classes that are favored by max pose (object) diversity. Inter- estingly, more than half the classes have no preference or prefer pose diversity. While this was expected for some classes such as Monitor or have relatively low object diversity but whose appearance can vary signif- icantly with pose, it was surprising that it holds for so many. Note that most in the wild datasets only exhibit object diversity, because pose diversity is to control in this setting. The fact that the preference for object or pose diversity varies with the class also suggests that further research is needed on intelligent data collection strategies that can take this into account. In the lab datasets, like OOWL500, enable such research. Pose diversity vs camera shake A second experiment compared the im- portance of pose diversity vs. camera shake. For this, we xed O=C = 16 and considered two strategies to assemble the ne-tuning set. Max pose = 8 andE=P = 3; max camera shake usedP=O = 1 andE=P = 30. The scatter plot in Fig. 8(b) shows the average recognition accuracies results. Points above (below) the blue line signal classes that are favored by max pose diversity (camera shake). Most classes have a clear preference to pose diversity, an eect even stronger than in the previous experiment. Again, while camera shake is usually simulated for (by shifting, cropping, etc.) in classier design, pose diversity is much harder to address with in the wild datasets. 4 Conclusion In this work, we have investigated the hypothesis that dataset collection the wild is not sucient to guarantee unbiased object recognizers. Specically, we have argued that professional photography and pose biases are dicult to eliminate in web-based data collection and proposed a new the lab data collection setup that mitigates these factors. This consists of using a drone to take images as it hovers in various spots around an object. We have shown that this setup has multiple interesting properties. First, by utilizing drone ight, it is possible to collect images of objects at controlled camera angles. This can be used to compensate pose bias. Second, the procedure is inexpensive and Hence, it could lead to a scalable data collection eort by the Third, drone captured images tend to have a certain amount of camera shake that mitigates professional photography bias. The usefulness of the procedure was demonstrated through the collection of the OOWL500 dataset { the largest in the lab dataset recognition dataset in the literature when both number of classes and objects per class are was then shown to enable a number of new insights on object recog- nition. First, a novel adversarial attack strategy was developed, where are dened in terms of semantic properties such as camera shake and pose. Experiments with this strategy have shown that ImageNet indeed has considerable amounts of pose and professional photography bias. Second, it was used to show that the augmentation of in the wild datasets, such as Overcoming Dataset Collection Bias in the Wild 15 with in the lab data, such as OOWL500, can signicantly decrease these biases, leading to object recognizers of improved generalization. Third, it was used to study various questions on est procedures for dataset collection. This has shown that data augmentation with synthetic images does not suce to elimi- nate in the wild datasets biases, and that camera shake and pose diversity play a more important role in object recognizers' robustness than previously Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classication with deep con- volutional neural networks. In: Advances in neural information processing systems. (2012) 1097{1105 2. Simonyan, K., Zisserman, A.: Very deep convolutional networks for recognition. arXiv preprint arXiv:1409.1556 (2014) 3. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A., et al.: Going deeper with convolutions 4. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern 770{778 5. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for ac- curate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. (2014) 580{587 6. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detec- tion with region proposal networks. In: Advances in neural information (2015) 91{99 7. He, K., Gkioxari, G., Doll ar, P., Girshick, R.: Mask r-cnn. In: Computer Vision (ICCV), 2017 IEEE International Conference on, IEEE (2017) 2980{2988 8. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recog- nition in videos. In: Advances in neural information processing systems. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotem- poral features with 3d convolutional networks. In: Computer Vision (ICCV), 2015 IEEE International Conference on, IEEE (2015) 4489{4497 10. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating In: Proceedings of the IEEE conference on computer vision and pattern recognition. (2015) 3128{3137 11. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Lawrence Zitnick, C., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE on Computer Vision. (2015) 2425{2433 12. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A image database. In: Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, IEEE (2009) 248{255 13. Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D., Bagul, A., Langlotz, C., Shpanskaya, K., Lungren, M.P., Ng, A.Y.: pneumonia detection on chest x-rays with deep learning. (2017) 14. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectiers: Surpassing human- level performance on imagenet classication. CoRR abs/1502.01852 (2015)16 Authors Suppressed Due to Excessive Length 15. Shoham, Y., Perrault, R., Brynjolfsson, E., Clark, J., LeGassick, C.: index 2017 annual report (2017) 16. Perry, T.S.: Stanford algorithm can diagnose pneumonia better than Rao, N.: Deep learning and the need for unied tools (2017) 18. Eckel, R.: Microsoft researchers' algorithm sets imagenet challenge Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition. CVPR '11, Washington, DC, USA, IEEE Computer Society (2011) 1521{1528 20. Nene, S.A., Nayar, S.K., Murase, H.: Columbia object image library (coil-100) 21. Geusebroek, J.M., Burghouts, G.J., Smeulders, A.W.: The amsterdam library of object images. International Journal of Computer Vision 61(1) (2005) 103{112 22. LeCun, Y., Huang, F.J., Bottou, L.: Learning methods for generic object recog- nition with invariance to pose and lighting. In: Computer Vision and 2004. CVPR 2004. Proceedings of the 2004 IEEE Computer on. Volume 2., IEEE (2004) II{104 23. Papernot, N., Carlini, N., Goodfellow, I., Feinman, R., Faghri, F., Matyasko, K., Juang, Y.L., Kurakin, A., Sheatsley, R., Garg, A., Lin, Y.C.: cleverhans v2.0.0: an adversarial machine learning library. arXiv (2017) 24. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533 (2016) 25. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. (2015) 1912{1920 26. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object cate- gories. Computer vision and Image understanding 106(1) (2007) 59{70 27. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88(2) (2010) 303{338 28. Grin, G., Holub, A., Perona, P.: Caltech-256 object category dataset. (2007) 29. LeCun, Y., Bottou, L., Bengio, Y., Haner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11) (1998) 2278{2324 30. Krizhevsky, A.: Learning multiple layers of features from tiny images. (2009) 31. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll ar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision, Springer (2014) 740{755 32. Sun, M., Bradski, G., Xu, B.X., Savarese, S.: Depth-encoded hough voting for joint object detection and shape recovery. In: European Conference on Computer Vision, Springer (2010) 658{671 33. Matzen, K., Snavely, N.: Nyc3dcars: A dataset of 3d vehicles in geographic context. In: Computer Vision (ICCV), 2013 IEEE International Conference on, IEEE Ozuysal, M., Lepetit, V., Fua, P.: Pose estimation for category specic multiview object localization. In: Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, IEEE (2009) 778{785 35. Leibe, B., Schiele, B.: Analyzing appearance and contour based methods for In: Computer Vision and Pattern Recognition, 2003. IEEE Computer Society Conference on. Volume 2., IEEE (2003) II{409

2108.10992

Brandon Leung

Brandon Leung, Chih-Hui Ho, Amir Persekian, David Orozco, Yen Chang, Erik Sandstrom, Bo Liu, Nuno Vasconcelos

OOWL500: Overcoming Dataset Collection Bias in the Wild

cs.CV

http://creativecommons.org/licenses/by/4.0/

The hypothesis that image datasets gathered online in the wild can produce biased object recognizers, e.g. preferring professional photography or certain viewing angles, is studied. A new in the lab data collection proposed consisting of a drone which captures images as it circles around objects. Crucially, the control provided by this setup and the natural camera shake inherent to flight mitigate many biases. It's inexpensive and nature may also potentially lead to a scalable data by the vision community. The procedure's usefulness is demonstrated by creating a dataset of Objects Obtained With fLight (OOWL). Denoted as OOWL500, it contains 120,000 images of 500 objects and is the largest in the lab image dataset available when both number of classes and objects per class are considered. Furthermore, it has enabled several of new insights on First, a novel adversarial attack strategy is proposed, where image perturbations are defined in terms of semantic properties such as camera shake and pose. Indeed, experiments have shown that ImageNet has of pose and professional photography bias. Second, it is used to show that the augmentation of in the wild datasets, such as ImageNet, with in the lab data, such as OOWL500, can significantly decrease these biases, leading to object recognizers of improved generalization. Third, the dataset is used to study questions on best procedures for dataset collection. It is revealed that data augmentation with synthetic images does not suffice to eliminate in the wild datasets biases, and that camera shake and pose diversity play a more important role in object recognition robustness than previously thought.

2 ZHONGYANG LI Whenl;dare xed and let m!1 , by Stirling's formula we m;d;l consists of all the ( d;l)-regular hypergraphs with mhyperedges. A random ( d;l)- regular is a random graph uniformly distributed in m;d;l. We shall apply the rst and second moments method; and nd a region for parameters when the exponential growth rate of the second moment is twice of that of the rst moment. Similar approach was used to determine the asymptotic independence number for the dense Erd os-R enyi random graph ([23]); and to show the existence and non-existence of on hypergraphs ([11]). The major dierence between a matching problem and a perfect matching problem lies in the fact that in the matching problem, we need to nd the global maximizer for a function of two real variables, in contrast to the case that in a perfect matching problem, the global maximizer of a single variable function is studied. When the second moment is approximately the square of the rst moment, we use a of the subgraph conditioning technique to show the convergence in probability of the free energy. The subgraph conditioning method was rst used to study on a random graph (see [33]); the major dierence here is that we prove that the asymptotic Poisson distribution of the number of cycles on a random graph conditional on the existence of a specic matching. The combination of the moments method and the subgraph conditioning method were also used in [1, 2, 18]. Note that when l= 2, the model is the same as monomer dimers on a random graph, whose free energy was computed in [3] and the convergence of the density of maximal matching was proved in [15, 9] using the technique of weak convergence (see [4, 12]). When d= 2, the model is the same as the independent set model on a random graph, of which the uctuations of the maximum were studied in [14] when lis suciently large. Free energies of vertex models on a graph with more general constraints were computed in [31, 22, 20, 21]. Here are the main results of the paper. Theorem 1.1. (1) Letbe the unique root LetZbe the total number of matchings on a random d;l-regular graph Gm;d;l. Then when mis suciently + ON RANDOM REGULAR HYPERGRAPHS 3 (2) Let L1: = min( 1 dldl+ 2;1p (d1)(l1) + 1) (1.3) L2: = min( 1 d + =ln+ (l1)(1) ln (1d) one of the following conditions (a)l= 2and2 0;1 d
; that asm!1 probability. (3) Let Z(x) :=m dX k=0Zkxk; (1.6) whereZkis the total number of matchings on Gm;d;l with exactly kpresent one of the following 2, and2 0;1 d
; or (b)d2,l3, such that 0 d;l(L1) + ln(x)0; or (c)d2,l3, such that 0 d;l(L2) + ln(x)0. Then +(x) ln(x): in 1.2. Let t:=t(s) =3l2l+ 1 10l+ 4l24l+ 1 2l2(0;); (1.8)4 ZHONGYANG LI and G(s) := 22d+ 3dl22dl+ 1) + then asm!1 (1.10) in approach to compute the free energy is by computing free energy on trees. For a given spin model on a graph or hypergraph, one can dene a Gibbs measure on an innite, regular (hyper)tree. This Gibbs measure may or may not be unique, depending on the parameters. A Gibbs measure on the tree may or may not exhibit strong spatial mixing (SSM), a strong form of correlation decay. It is known that uniqueness and SSM hold for matchings in d-regular l-uniform hypergraphs when l5 anddis arbitrary (see [35], In particular, since a random regular hypergraph converges locally to the innite tree, if SSM holds then the limiting free energy of the random hypergraph model can be computed from the tree calculation. The example below gives applications of Theorems 1.1 and 1.2, which shows that the a.s. limit of the free energy can be computed beyond the SSM threshold. In Section A, we show that the conditions of Theorem 1.2 can only hold for nitely many ( d;l)'s. Example 1.3. (1) It is straight forward to check that have 0 d;l(L1)<0. Then the free energy of matchings on Gm;d;l converges in probability to d;l()asm!1 by Theorem 1.1(2). (2) Numerical experiments show that assumptions of Theorem 1.2 holds, and we obtain (1.10). However, the assump- tions of Theorem 1.2 does not hold for (d;l) = (100;100). The organization of the paper is as follows. In Section 2, we compute the rst moment of the total number of matchings of a given density, and the rst moment of the totoal number of all matchings on Gm;d;l. In section 3, we compute the second moment of matchings with a given density and nd explicit conditions under which the second moment and the square of the rst moment have the same exponential growth rate. In section 4, we show that conditional on a matching, the distributions of the number of cycles of dierent lengths onGm;d;l converge to independent Poisson random variables as m!1 . In Section 5, we show the convergence in probability of the free energy to an explicit limit. In Section 6, we show the convergence in probability of the weighted energy to an explicit limit when eachMATCHINGS ON RANDOM REGULAR HYPERGRAPHS 5 hyperedge in the matching is given a suciently small weight x > 0 and prove Theorem 1.1(3b). In Section 7, we prove Theorem 1.1(3c). In Section 8, we discuss the implications of our results on maximal LetZmbe the total number of matchings consisting of mhyperedges on a random ( d;l)-regular graphGm;d;l. In this section, we compute the rst moment of EZmand the rst moment of Z, the totoal number of matchings on Gm;d;l. Note that if >1 dthenZm= 0. Now (2.1) Then EZm , and letm!1 , by Stirling's formula we d;lis dened by =ln(l1) ln(1) +lln (1d) (2.4) 00 d;l() (2.5) Then we have Lemma 2.1. = d;l(), whereis given by (1.2). (2) Iff1 d
<0, let0be the unique number in ;1 d
such that d;l(0) = 0: Then (a) If2(0;0), then d;l()>0. (b) If2 0;1 d
, then d;l()<0.6 ZHONGYANG LI Proof. From (2.4) we = +1; =1: By (2.5), the equation 0 d;l() = 0 has exactly one root in (2.5) we see 0 d;lis monotone decreasing in 0;1 d
, 0 d;l>0, and d;l() is monotone 0 d;l<0, and d;l() is monotone decreasing. Then Part (1) of the lemma follows. = (1.5) we = it is straightforward to check Part (2) and (3) of the lemma.  Lemma 2.2. Letn2N. nn+1 2enn!e nn+1 2en Proof. See Exercise 3.1.9 of [24].  Lemma 2.3. Let2 0;1 d
. convergence is uniform in any compact subset Leth2[1;m d1] be a positive integer. By (2.2) and (2.6), we ON RANDOM REGULAR HYPERGRAPHS 7 Then the lemma follows.  Proof of Theorem 1.1(1). Note that Z0= by be a positive integer. We shall divide the sum over hin the right hand side of (2.8) into two parts (1) Part I is the sum over hwhenh mis in aKpmneighborhood of . (2) Part II is the sum over hwhenh mis outside of . More precisely, let Q:= h2N: I + II where I II expansion of  d;l(
) gives ZHONGYANG LI Hence we geometric series !0; asK!1 . + (lddl): Then the lemma follows.  3.Second Moment In this section, we compute the second moment of Zmand nd explicit conditions under which EZ2 mand (EZm)2have the same exponential growth rate. We can also compute the second moment of Zmas follows (1) A double matching conguration is the union of two matchings on the same the partition function of double matching congurations on a ran- dom graph, in which each matching consists of exactly medges. (2) In each double matching conguration M1[M2on a random graph of medges such that each matching consists of exactly medges; all the hyperedges in [ m] are divided into 4 types by the conguration (a) Type (a) hyperedges are in M1M2; assume there are sType (a) edges, where 0sm. ThesType (a) edges splits into lshalf edges, each one of which can not share a vertex with any half edge in M1[M2. A half edge is of Type (a) if and only if it is part of a Type (a) hyperedge. The total number of Type (a) half-edges is ls. (b) Type (b) hyperedges are in M1Mc 2; the total number of Type (b) edges is ms. The (ms) Type (b) edges splits into l(ms) half edges, each one of which cannot share a vertex with any half edge in M1. A half edge is of Type (b) if and only if it is part of a Type (b) edge. The total number of Type (b) half edges is lmls. Depending on whether they share a vertex with a half edge inM2Mc 1, Type (b) half edges can be further divided into 2 sub-types: (i) Each Type (bi) half edge does not share a vertex with any edge in M1[M2. The total number of Type (bi) half edges is lmlst. (ii) Each Type (bii) half edge shares a vertex with exactly one half-edge in M2Mc 1. The total number of Type (bii) half edges is t.. (c) Type (c) hyperedges are in Mc 1M2; the total number of Type (c) edges is ms. The (ms) Type (c) edges splits into l(ms) half edges, each one of which cannot share a vertex with any half edge in M2. A half edge is of Type (c) if and only if it is part of a Type (c) edge. The total number of Type (c)MATCHINGS ON RANDOM REGULAR HYPERGRAPHS 9 half edges is lmls. Depending on whether they share a vertex with a half edge inM1Mc 2, Type (c) half edges can be further divided into 2 sub-types: (i) Each Type (ci) half edge does not share a vertex with any edge in M1[M2. The total number of Type (ci) half edges is lmlst. (ii) Each Type (cii) half edge shares a vertex with exactly one half-edge in M2Mc 1. The total number of Type (cii) half edges is t. (d) Type (d) edges are in Mc 1Mc 2; the total number of Type (d) edges is m 2m+s. A half edge is of Type (d) if and only if it is part of a Type (d) edge. The total number of Type (d) half-edges is (3.1) It is straight forward to check the following lemma Lemma (3.2) Lemma (3.3) Proof. It suces to show thatlmis the total number of half edges in the graph total number of half edges of Type (a) (bi) and (ci); any two of these half edges cannot share a vertex; hence the total number of half edges sharing a vertex with these half (3.4) The total number of Type (bii) and (cii) half edges is 2 t; these half edges are incident and the total number of half edges incident to these tvertices is md (3.5) There are no common half edges in (3.4) and (3.5), therefore their sum is at most ml-the total number of half edges in the graph Gm;d;l. Then the lemma follows.  By (3.3), we ZHONGYANG LI LetKm;;s;t be the total number of possible half edge assignments for the rst (2 Then Km;;s;t we assume 2(0;); (3.9) Whenm!1 , by Stirling's formula we obtain F(m;m )MATCHINGS ON RANDOM REGULAR HYPERGRAPHS ) =lnln l+ (l1)(12+) ln(12+) +ln(d1) we have for all ) @=ln+ (l1) ) @=ln l+ 2 we solve the following system of ) @= 0; @ d;l(;; ) @= 0:(3.13) and is straightforward to check (3.16) is a pair of solutions for (3.14) and by (3.15) we (d1)a() (3.17)12 ZHONGYANG LI Plugging (3.17) to (3.14), we obtain h() := ln(a()) + = 0 Lemma 3.3. Assume one of the following two conditions dened as in (1.7), (1.8) and (1.9) If G(s)>0for alls2(0;c0), then the equation h() = 0 has exactly one solution when 2(0;). Proof. Note that condition (2) implies condition (1); hence it suces to prove the lemma under condition (1). Note that h0() =h1() = 1 the last identity follows from the assumptions that 2(0;),d3 and deduce thata0() a()is strictly decreasing when 2(0;). +1 there exists a unique c02(0;) as dened in (1.8) such if2(0;c0) = 0 if=c0 <0 if>c 0:MATCHINGS ON RANDOM REGULAR HYPERGRAPHS 13 Note thath0()>0 when2(0;c0]. We shall study the zeros of h0() computations show that h0() =l(2+ + 1)>0: IfG(s)>0 for alls2(0;c0); thenh0()>0 for all2(c0;). Hencehis when 2(0;). Then the lemma follows from the fact =1 lim !h() =1:  Lemma 3.4. Under the Assumptions of 3.3, for each xed at (3.16) d;l(;; )achieves its unique global maximum when (;)lie in the following (3.18) Proof. Note that @ d;l(;; ) @ =0= +1;@ d;l(;; ) @ =0= +1; Hence a global maximum can be obtained along none of the vertical line = 0; the 0, the horizontal line = 0 and the boundary 1 2d+d+d l= 0. Then the lemma follows from Lemma 3.3.  Lemma 3.5. Assume (2.1),(3.8) (3.9) hold. Assume d2. Then for any ) @2<0: (3.19) Proof. Note thatl dby (3.9) when 0 < <1 d. By (3.9), (3.3), we ZHONGYANG LI Then (3.19) follows from (3.11).  Dene the Hessian matrix H(;; ) := @2 d;l(;; ) @2! Lemma 3.6. Assume (2.1),(3.8) (3.9) hold. (1) If ld (3.20) Then (3.16) is a local maximizer of d;l(;; )when (;)2Rfor If <1 1 (3.16) is a local maximizer of d;l(;; )when (;)2R. Proof. Since (3.16) satises (3.13), it suce to show that H(;2;l(d1)2) is negative denite when (3.20) holds. By (3.19), (3.12),it suces to show (3.22) when (3.20) holds. Note (3.22) holds if and only if g() :=2(d+ldl)2+ 1>0: Note that when and the identity holds if and only if d=l= 2. Whend=l= 2, 12 >0 given that and at least one of dandlis strictly greater than 2, g()>0 for all2 0;1 d
if and only if g1 d
0. This gives ld.  Lemma 3.7. Assume (2.1),(3.8) (3.9) hold. If l= 2, then for each 2 0;1 d
, the )is negative denite for all ON RANDOM REGULAR HYPERGRAPHS 15 Proof. Again by (3.12) (3.19) it suces to show that det H > 0 whenl= 2. Let A: =1 <0: U: (A+ 2C Note the last inequality follows from (3.9). Hence we have det H > 0, and the lemma follows.  Lemma 3.8. Assume (2.1),(3.8) (3.9) hold. Let l3. For the set R(;) : detH(;; )>0g has at most one connected component in the For simplicity let := l: For each xed 2 0;1 d
, explicit computations show that det H(;;l ) = 0 if and only if = 3l+ ZHONGYANG LI Then for each xed ( d;l; ) such that 2 0;1 d
, the equation () = has at most one solution in (0 ;) given =1 = equation 0() =1 has a unique positive solution given when < 3,F()<0 and when > 3,F()>0. Note that D() = (dlld)2+ (3ld2l+ 22d+dl22dl+ 1)22+: =(12)>0: D() = the unique point in (0 ;) such that D(1) = 0. Whend1 2we have R=f(;) = = (dlld)2+ (d+l+ 22ddl1)+ 22d43d+ the identity holds if and only if = 0 andd=1 2. J() = when2(0;), the curve =() has two ON RANDOM REGULAR HYPERGRAPHS 17 (0,0) > 0 detH < 0 re Figure 3.1. The curve det H= 0 is represented by blue lines. When represented by the triangular region in the ( ;) plane with vertices (;0), (0;0) and (0;). The curve det H= 0 intersects the line +=at a unique point ( 5;5). The curve det H= 0 divide the triangular region into two connected components - the left det H > 0 and the right component satises det H < and lim !1() lim!1+() whenl3 Hence when 2(0;1), the curve =l() does not intersect R. When2(1;), the curve=l() divide the triangular region Rinto two connected components, exactly one of them satisfy det H > 0. It is straightforward to check that the component on the right (resp. left) satises det H < 0 (resp. det H > 0) given that D()<0 for2(5;); see Figure 3.1.  Lemma 3.9. Assume (2.1),(3.8) (3.9) hold. Let l3. For the set R(;) : detH(;; )>0g has at most two connected components in the If 1>d>1 2, then R: = (;) ZHONGYANG LI 0<J(0) (3.28) we can see that the graph of J() is a parabola opening upwards. Therefore there is a unique solution thatJ(2) = 0; and J()<0 for Recall that 1is the unique solution for D() = 0 in (0;). The following cases might We claim that the graph of=() when2(0;2) in this case is always below the straight line = 21 din the (;) plane. By (3.29) we know this is true when = 0. This is also true when therefore 2 0. Therefore if =() intersects = the intersection must be at least two points (counting (3.25) we andd2,F() a parabola opening upwards with F(0)<0. Hence if there exists 0 <  3<  2such 0; andF(3) D(3)<0 for all2(0;3). Then=() intersects = 21 din at most one point when 2(0;2). It follows that all2(0;2). (b) andJ()<0, and there for =() is always below the curve = andJ()<0. HenceJ() D()>0 and lim achieves its minimum in ( 1;) at3. show ON RANDOM REGULAR HYPERGRAPHS 19 Therefore when suciently the derivative shows that have at most two zeros; hence in this case )>0ghas at most two connected components in the (  2: By (3.29) we . Note also that ( ) +1> Then the curve =() must intersect the curve Then somewhere in (0 ;5) we have0()>1. Then we have 3<5. From (3.24) and (3.26) we see that if 3<5, thenl<2. But this is a contradiction to the fact that l3. Then we must have 12. Then the lemma follows; see Figure 3.2.  Lemma 3.10. 0()is strictly increasing when 2(5;). Proof. It suces to show that Note that 00() particular we have 00(3) (3.30) and (3.31) Note that K0() = 3(dldl)L(); =2l(1)22+ (+(12))2 In particular we have L() = 0: SinceLis a parabola opening downwards, with exactly one negative root and one positive root3, we obtain ZHONGYANG LI (0,0) > 0 detH < > 0 detH < 3.2. The top graph represents the case when det H > 0 has a unique component in R; while the bottom graph represents the case when det H > 0 has two components in R.MATCHINGS ON RANDOM REGULAR HYPERGRAPHS (3.32) we is the maximal value of K() when2(5;).By (3.31), we (3.30) we the lemma follows.  Lemma 3.11. that (3.21) holds. Assume R(;) : detH > 0gis connected. ) = (2;l(d1)2)is the unique maximizer for d;l(;; )in the region R(;) If2 0;1 2d , we consider the corresponding region in the ( ;)-plane which is bounded by (a)= 0,2[0;]; and (b)= 0,2[0;]; Figure when the setR(;: detH(;; ))>0ghas exactly one in the ( ;)-plane, its corresponding region in the ( ;)-plane is bounded by (A)= and (B)= Figure 3.2(1).22 ZHONGYANG LI In either case by Lemma 3.6, the point z0:= (;(d1)2) is a local maximizer for d;l(;;l ). If the following conditions hold (i) for any point zin the region or on the boundary of the region R(;l)(;) : detH(;;l ))>0g, the straight line joining z0andzlies in the region R(;l)f(;) : detH(;;l ))>0gexcept the endpoint (in the case that zis on for any point zin the : detH(;;l ))>0g, or on the boundary of the region; let lz0zbe the straight line joining zandz0starting from z0and ending atz; we have d;l(;z) = d;l(;z0) is the directional derivative of d;l(;w) with respect to walong lz0z. Note 0: SinceHis negative denite in the region R(;l)(;) : detH(;;l ))>0g, the second order directional derivative =vTHv<0 for allwin the interior of the line lz0z; where v2R12is the unit directional vector along the linelz0;z. Then we eachwin the interior of the line lz0z; and any point zin the : detH(;;l ))>0g, or on the boundary of the region other than z0. Now it remains to prove condition (i). It suces to check condition (i) for each boundary point on the boundary of the region R(;l)(;) : detH(;;l ))>0g. The straight line in the ( ;)-plane joining z0and (;0) has for conditions (i) holds. Since when 3andd < ON RANDOM REGULAR HYPERGRAPHS 23 it suces to check (3.34) for 2(3;); which is equivalent to (3.33).  Lemma 3.12. Assumed2andl3. If 1 dldl+ 2; (3.35) then (3.33) holds. Proof. Then (3.33) holds if and only gives (3.35).  Lemma 3.13. Assumed2andl3. If (3.35) holds, then R(;) : detH > 0ghas exactly one connected We see that when (3.35) holds, the line segment passing through ( 2;(d1)2) and (;0) in the (;)-plane intersects the curve det H= 0 at a unique point ( ;0), and the line segment lies below the curve, since both of them pass through the point ( ;0); and The slope of the line segment is greater than or equal to the slope of the tangent line of the curve det H= 0 at (;0) in the (;)-plane; and The slope of the line segment is strictly greater than or the slope of the tangent line of the curve det H= 0 at any point other than ( ;0) in the that the line segment passing through ( 2;(d1)2) and (;0) in the lies in the region f(;) : (;l)2Rgsince the region is convex. Hence the line segment does not cross the (possible) boundary = 21 d. Therefore the curve detH= 0 does not cross the boundary = 21 deither. Then the lemma follows.  Lemma 3.14. Assume 3l,2d, and (3.35) holds. ) = 2d;l() Proof. By Lemmas 3.11 and 3.12, we obtain that under the assumptions of the ) = 2d;l(): We now ) When the Hessian determinant is negative, d;l(;; ) as a function of ( ;) cannot achieve a local maximal value at an interior point; it cannot achieve the global maximal value in R along the boundary satisfying det H= 0 either, since this is also the boundary of the region detH > 0 where the function is strictly concave, and the maximal value in det H0 is already achieved at the unique interior point ( 2;l(d1)2). Therefore if a global maximal24 ZHONGYANG LI value of d;l(;; ) inRis achieved when ( ;)2RdetH < 0g, then it must be achieved along the boundary =l(); see Figure 3.1. More precisely, the only for the global maximal value when det H < 0 ()) Along the line =l(), by (3.10) we have Td;l;() : = d;l(;;l ()) (3.36) =ln+ (l1)(12+) ln(12+) +l() ln(1d) Then by ()) =ln+ (l1)(1) ln(1d) = (l1) and the identity holds if and only if =5. monotone increasing when 2[5;). The following cases might in this = both5andare along the boundary of the unique connected component of R such that det H(;; )>0, where d;l(;; ) is strictly concave with a maximal value given by 2 d(); hence we the lemma follows In this case Td;l;() is increasing when 2[5;). Hence the lemma holds as well. MATCHINGS ON RANDOM REGULAR HYPERGRAPHS 25 Lemma 3.15. Assume one of the following conditions holds: l= 2, and2 0;1 d
; or The assumptions of Lemma 3.14 we assume (3.38) we 1(3.39) Proof. LetL1 be a positive integer. By Lemma 2.2 and (3.7), we obtain F(m;m ) =m2em d;l(;; ) = 1 shall split the sum over 1 sm2, two parts Part I is the sum over (s m;t m) in aLpmneighborhood of ( 2;l(d1)2); 1sm2; II is the sum over 1 sm2, 1t lmls1, such thats m;t m
is outside of ( we I + ZHONGYANG the transpose of s;t. Note that when ( s;t) is on the boundary of N, from the Taylor expansion of  d(;
;
) at (2;l(d1)2), we is the maximal eigenvalue of approximately a convergent geometric series. Hence when mis ; whereC00>0 is a constant independent of m. Moreover, I +O1 m Hence we the lemma follows from (3.40), (3.23) and (2.7).  4.Subgraph Conditioning In this section, we show that conditional on a matching , the distributions of the number of cycles of dierent lengths on Gm;d;l converge to independent Poisson random variables as m!1 . Similar results when conditional on a perfect matching was proved in [11]; see ON RANDOM REGULAR HYPERGRAPHS 27 Letk1 be a positive integer. Let [k] ak-cycle onGm;d;l to be an alternating sequence of vertices and that eachviis a vertex for 0ik; eachejis a hyper-edge for 1 jk; v0=vk; and vi6=vj, for and for eachi2[k],eiis incident to both vi1andvi. In particular, a 1-cycle is a hyper-edge, at least two half-edges of which share a vertex. LetCkbe the number of k-cycles onGm;d;l, and letb1 be a positive integer. Forc= eachGm;d;l = (V;E)2 m;d;l, letQGm;d;l;be the set of all the matchings on of exactly mpresent hyperedges. Let M0E,jM0j=mbe a xed set of m hyperedges. Note that for any other set Mofmhyperedges, the probability that M0is matching on a random graph Gm;dis the same as the probability that Mis a matching on a random graphGm;d. More =P(M02QGm;d;l;) from the symmetry of hyperedges we can also see ZHONGYANG 0;1;:::;bk 2cbe the expected number of k-cycles including exactly Then we may compute Etas follows (1) Choose thyperedges from the mhyperedges in M0; there Cycli- cally order these edges, there orders. Choose ordered pairs of half-edges from each of these thyperedges to connect with half-edges not in M0, there are For 1it, letsi1 be the number of edges along thek-cycle between the ( i1)th andith hyperedge in M0along the cycle. Then we need to choose positive integers s1;:::;sk, such that s1+:::+sk=kt. Choose kthyperedges from the mmhyperedges not in M0. Consider all the possible orderings of these hyperedges as well as choices of two ordered half- edges in each of these chosen hyperedge to place them along the cycle, there The factor ( d1)kindicates the possible number of locations of two half edges incident to the same vertex along the k-cycle. (3) In thek2tvertices along the cycle which are not incident to an edge in M0along the cycle, assume that sof them are incident to an half-edge in M0but not along the cycle. Hence we ON RANDOM REGULAR HYPERGRAPHS (4.2) where when t1, E;k(t;s) 4.1. Whenkis xed and m!1 Whenkis xed and m!1 , we have when ZHONGYANG we is a Laurent series of x, fori2Z, let [xi]f(x) be the coecient for xiinf(x). Then we log 1x(1 +x)  =k[xk] log 1x p1+41 2! 1 + 4k +2p1 + 4+ 1k# =(1)k kk+ the lemma follows.  Hence by (4.2) and Lemma 4.1 we (4.3) Lemma 4.2. Letg1be a positive integer; k1;:::;kg1be positive integers, nonnegative integers. Let 2 0;1 d
. ON RANDOM REGULAR HYPERGRAPHS the number of (d;l)-regular hypergraphs with all the following a matching; and The hypergraph contains i.e., we can nd rikicycles in the hypergraph and label them by 1;2;:::;ri; dierent labelings of the same choice of ri k-cycles are counted dierently; and Any two cycles among the . . , rgkgcycles are the number of (d;l)-regular hypergraphs such that M0is a matching. Then (rg;kg), letm!1 The total number of ( d;l)-regular hypergraphs such that M0is a matching is shall prove by induction. First we consider the case when g= 1 andr1= 1. We the number of common hyperedges of the distinguished k1cycle andM0, and sis the number of common vertices of the distinguished kcycle andM0not incident to a common hyperedge of the cycle and M0. Fixk1;d;land letm!1 , by Lemma 4.1 we by (4.3). Note also ZHONGYANG for 1igand 1jri+ 1,ti;jis the number of common hyperedges of kicycle labeled byjandM0, andsi;jis the number of common vertices of the kicycle labeled by jandM0not incident to a common hyperedge of the cycle and letm!1 , we dlm d(1d)lm d(1d)
! (lm(1))! N (r1;k1);(r2;k2);:::;(rg;kg);1((s1;1;t1;1);:::; (sg;rg;tg;rg))dlm d(1d)lm d(1d)
! (lm(1))! (d1)kg(l1)kgtg;rg+1+sg;rg+1(1d)kg2tg;rg+1sg;rg+1(d2)sg;rg+1 2(kgtg;rg+1)(1)kgtg;rg+1 kgtg;rg+1 tg;rg+1kg2tg;rg+1 sg;rg+1 ON RANDOM REGULAR HYPERGRAPHS (sg;rg;tg;rg);(sg+1;1;tg+1;1)) we obtain that for any xed letm!1 for any xed and letm!1 , by (4.4) and Lemma 4.1 the lemma follows.  Lemma 4.3. 0be given as in Lemma 4.2.34 ZHONGYANG the number of (d;l)-regular hypergraphs with mhyperedges satis- fying all the following a matching; and The hypergraph contains precisely for any xed and letm!1 , we 0: Proof. Note that the total number of vertices in a ( d;l)-regular hypergraph with mhyper- edges islm d. It is straightforward to check that for any xed and let m!1 the lemma follows.  Lemma 4.4. (1) Whenm!1 , for in distribution to independent Poisson random variables with means given by E(Ck) (4.5) (2) Conditional on M02QGm;d;l;, whenm!1 , for in distribution to independent Poisson random variables with means given by (4.3) for k2[b]. Proof. Note that Part (1) of the Lemma when l= 2 was given by Lemma 1 of [33]; see also [36]. We only prove Part(2) here; Part (1) can be proved given as in Lemma 4.2. Let the number of ( d;l)-regular hypergraphs with mhyperedges satisfying all the following a matching; and The hypergraph contains and at least two cycles among the . . , rgkgcycles share a vertex. Then by the ON RANDOM REGULAR HYPERGRAPHS 35 By Lemmas 4.2 and 4.3, for any xed and letm!1 , by (4.4) the lemma follows.  Letc= (c1;:::;cb)2Zbbe abtuple of non-negative integers. =c); i.e.,pcis the probability that in a random ( d;l)-regular hypergraph with are exactly ckk-cycles for all k2[b]. Then by Lemma 4.4(1), when m!1 by Lemma 4.4, when mis suciently large E;c (4.6) Recall the following lemma about Poisson 4.5. LetXhas Poisson distritubion with mean . Then P(X < > (1 +))h e(1 See Theorem A.15 of [5].  Lemma 4.6. Leta>0be a positive constant. Then for any positive constant W > 1 (4.7) where (z) :=z(1 +z) log(1 +z): Proof. We useLto denote the left hand side of (4.7). Note that ey a1 2 k 1 (4.8)36 ZHONGYANG LI Consider the following note that log (1 we obtain ez(1+z) the last inequality follows from the fact whereW > 2 is a positive constant: note that 0(z) =log(z+ by (4.8) we > 2: we have z(1 +z) log(1 +z)z(log(1 by (4.8), we 4.7. dene S(y) (4.9)MATCHINGS ON RANDOM REGULAR HYPERGRAPHS 37 Then for all suciently large y, and any constant W > Note that by (4.3), (4.5) we see that ;k=k 1 a Poisson random variable with (1uk) (1 +vk) suciently large. Then by Lemma ZHONGYANG LI By (4.11) and (4.12) and 2 0;1 d
we by Lemma +vk)EXk)evk (1 suciently +vk)vk<0;evk (1 by (4.12) and (4.11) we shall apply (4.7) with a= 2. By (4.10) we since 2 0;1 d
, whenldwe := ON RANDOM REGULAR HYPERGRAPHS (EZm)2 the lemma follows.  Lemma 4.8. LetS(y)be dened as in (4.9); and C1;:::;Cbbe random variables denoting the number of 1-cycles, . . . , b-cycles in a random (d;l)-regular hypergraph with there exists an absolute constant  >0, such that for all suciently large y1 2 kand m, and for any constant W > Recall that as m!1 ,C1;:::;Cbare independent Poisson random variables with parameters 1;:::;b, respectively. When mis suciently large, by Lemma by Lemma (4.7) when a= 1, we the lemma follows.  Lemma 4.9. There exist constants A;B > such that for each c2S(y)and suciently large ZHONGYANG LI Proof. By (4.6), we even 1 0< we convergent. Then the lemma follows.  5.Free Energy In this section, we show the convergence in probability of the free energy to an explicit limit when ( the same exponential growth rate, where is the unique root of (1.2). Lemma 5.1. that one of the following conditions holds (1)l= 2; (5.1)MATCHINGS ON RANDOM REGULAR HYPERGRAPHS 41 in By (2.3), it suces to show 0: (5.2) for all>0. Let C= the random variable denoting the number of 1-cycles, 2-cycles, . . . , b-cycles in a hypergraph with mhyperedges. Let E;Cbe the conditional expectation Chebyshev's inequality, for any Lemmas 4.7 and 3.15, we by Lemma 4.8 we have for any constant W > ZHONGYANG suciently large. By Lemma 4.9, we by (5.3), (5.4) we 1 that when 2 0;1 d
, (1 Lemma 3.15, as y!1 , the right hand side of (5.5) goes to 0. ; ON RANDOM REGULAR HYPERGRAPHS 43 We obtain P7!0 asy!1 by Lemma 4.8 and P8!0 asm!0 by Lemma 4.9. Then the lemma follows.  It is straightforward to check the following lemma. Lemma 5.2. LetL1be given by (3.35). of Theorem 1.1(2) . By Theorem 1.1(1), it suces to show 0: for all>0. Note , where the limit follows from the fact , where the last limit follows from (5.2). Then the lemma follows.  6.Weighted Free Energy In this section, we show the convergence in probability of the weighted energy to an explicit limit when each hyperedge in the matching is given a suciently small Dene the weighted partition Z(x) function for matchings on a ( as in (1.6). When 2 0;1 d
, we ZHONGYANG = d;l(;x) +ln(x): Hence we 0 d;l() + ln(x) + = ln(x) + =1 We obtain that there exists a unique (x)2 0;1 d
, such = moreover (x) is the unique maximizer. Proof of Theorem 1.1(3b) . Note ln(x)]!0; asm!1 , Where the limit follows from the fact = ON RANDOM REGULAR HYPERGRAPHS 45 If 0 d;l(L1) + ln(x)0, then (x)L1: , where the last limit follows from (5.2) and (6.1). Then the lemma follows.  7.Another Criterion to Guarantee Global Maxima Lemma 7.1. 1. For any straight line in the through (2;(d1)2)with slope in s2(1;1], the second order of d;lalong the line is strictly negative when <dl+l22ld+ Lets2(1;1]. It suces to show that under the assumptions of the lemma, Y(s) (3.11) (3.12) and the fact that =lwe obtain Y(s) (;) is along the straight line passing through ( 2;(d1)2) with slope s, they satisfy the following + we it suces to show that Y(1)<0 when2(2;d2) and (7.1) holds.46 ZHONGYANG LI Explicit computations show that Y(1) 1) where U() = +(d2l+ 22dl) + is a parabola opening upwards. The following cases might occur (1)U() = 0 has no real roots, then U()>0 for all2R, in particular this = 0 has two real roots mM. IfM<2, thenU()>0 therefore Y(1)<0. LetTbe the discriminant of U(). In case (1), T <0. In case (2), we need T0 and M=p T[l1 +(d2l+ 22dl) + computations show that (7.2) holds if and only if d2l24+ 2d2l23d2l2+ (2dl2+ 1>0 (7.3) It is straightforward to check that (7.3) holds whenever (7.1) holds, d3 andl2. Lemma 7.2. < 1. Then along the line segment the (;)-plane, there exists 82[2;d2], such decreasing; Let X() suces to show that X()0 when2(2;8); and X()0 (3.11) (3.12) and the fact that =lwe obtain X() 1) where V() =2+ (l2l)+ 22MATCHINGS ON RANDOM REGULAR HYPERGRAPHS 47 Note that V() is a parabola opening upwards with exactly one positive root and one negative root. Moreover, V(0) the lemma follows.  Lemma 7.3. Assume (7.3) and (3.21) hold. (7.4) Then (;) = (2;(d1)l2)is the unique global maximum for d;l(;; )when (;)2R. Proof. Letz= (;) be a point along the line segment joining z0:= (2;(d1)2) and z1:= (d2;0) in the (;)-plane. By Lemma 7.2, we Lemma 3.5, if we consider the region R;2Rin the (;)-plane bounded by =d2; and =; and = 0; and = (d1)2; we ) = ) = the last identity follows from Lemma 7.1 and the fact that all the derivatives at d;l(;; ) at (;) = (2;(d1)l2) vanish. It is straightforward to check that the maximizer in R;2for d;lis unique by the strict negativity of the the region in the ( ;)-plane bounded by =2; and = 0; and =d2; and = 21 dif 21 d>2. Then by Lemma 7.1 and the fact that all the derivatives at d;l(;; ) at (;) = (2;(d 1)l2) vanish, we ) = that the maximizer in R;1for d;l(;;l ) is unique.48 ZHONGYANG (;) : detH(;;l (;) : detH(;;l )<0g: For any point ( ;)2R;3, there exists a line segment joining ( 2;(d1)2) and (;) such that every point along the line segment is in R;3. Note that det H > 0 at each interior point of the line segment, therefore the second order directional derivative of d;l(;;l ) along the line segment is strictly negative at each interior point. Since the rst of d;l(;;l ) at (2;(d1)2) are 0, the rst order directional derivative of d;l(;;l ) along the line segment is strictly negative at each interior point. Hence we consider R;4. When the Hessian determinant is negative, d;l(;; ) as a function of ( ;) cannot achieve a local maximal value at an interior point; it cannot achieve the global maximal value in Ralong the boundary satisfying det H= 0 either, since this is also the boundary of the region det H > 0 where the function is strictly concave, and the maximal value in det H0 is already achieved at the unique interior point ( if a global maximal value of d;l(;; ) inRis achieved when ( ;)2RfdetH < 0g, then it must be achieved along the boundary =l(); see Figure 3.1. More precisely, the only possible candidate for the global maximal value when det H < 0 be dened as in (3.37). As in (3.37), we obtain that the identity holds if and only if =5. monotone increasing when 2[5;). (5;5)2R;3, ( ;0)2R;2, the lemma follows.  Lemma 7.4. Assume (7.3) and (3.21) hold. If 1 d ON RANDOM REGULAR HYPERGRAPHS 49 Then (;) = (2;(d1)l2)is the unique global maximum for d;l(;; )when (;)2R. Proof. It is straightforward to check that (7.5) implies (7.4). Then the lemma follows from Lemma 7.3.  Proof of Theorem 1.1(3c). Theorem 1.1(3c) follows from the similar arguments as the proof of Theorem 1.1(3b).  Proof of Theorem 1.2. Recall thatRis dened as in (3.18). By Lemma 3.4, the assumptions of Lemma 3.3, at ( ;) = (2;l(d1)2), d;l(;; ) achieves its unique global maximum in R. Then when (3.38) holds, we obtain (3.39). By the subgraph conditioning technique, we obtain (5.1) under the assumptions of Lemma 3.3 and (3.38). Then the conclusion follows the rst convergence follows from the denition of , and the second from the assumption that satises the assumptions of Lemma 3.3 and (3.38). In particular, one may check that when d= 2  8.Maximal Matching In this section, we discuss the implications of our results on maximal 8.1. Let2 0;1 d
. By (2.7), we Taylor expansion at 0to approximate  the lemma follows.  Proposition 8.2. Let2 0;1 d
. LetKmbe given as in (8.1). Let ZKm+Cbe the total number of matchings containing at least Km+Chyperedges in a random regular graph Gm;d;l2 m;d;l. 0; where the convergence is uniform for all suciently large mC.50 ZHONGYANG LI Proof. Let Cm:=5 thatCm>0 since  d;l() is strictly decreasing in ;1 d
, (2.7) we since d;l() is strictly decreasing when 2 ;1 d
, we there are at most msummands in W2, we obtain W21 m1 C, sincemC. Again using the Taylor expansion of  d;lat0to approximate  d;l(Km+) we the lemma follows since 0 d;l(0)<0.  Appendix A.Discussions of conditions in Theorem 1.2 In this section, we show that the conditions of Theorem 1.2 can only hold for nitely many then (1.2) can be re-written as (d1)zl+z1 = 0: be the unique solution of (A.1) in (0,1). As d;l!1 , has the ON RANDOM REGULAR HYPERGRAPHS 51 (1) if lim + (1e)
(1 then 1eds> (1)e)
Then whenever  > ewe and therefore large and del. (2) if lim d;l!1lnd l= 0, and lim we 1) lnd ld(1 1) lnd l(1 ZHONGYANG LI Hence G(s) = dl(1dt)s2+ 2(1dlt) +s 2ltd+( + 1) lnd  (1 +o(1)) where 2(0;1), then t=(1 =dls(s)(1 +o(1))<0; for all <1. (3) if lim d;l!1lnd l= 0, and lim d;l!1lnd lnl= +1, we +o(1)) :=dl(1dt)s2+ 2(1dlt) +o(1)) for2(0;1), thent=(1 +o(1)), and G(s) +o(1))<0; for all <1. (4) if lim l!1lnd l= 0, and lim l!1lnd lnl= 0 (asl!1 ,dmay be nite or d!1 ), +o(1)) = 2(1(d1)lt) ON RANDOM REGULAR HYPERGRAPHS 53 Assumes=(1 +o(1)) where 2(0;1), thent=(1 +o(1)), and G(s) +(d1)l] = <1. (5) if lim d!1lnd l= +1, we +l)2 +o(1)) where 0 < +l)i (1 +l)]<0; when Z.L.'s research is supported by NSF grant DMS-1608896 and Simons grant D. Achlioptas and C. Moore. Random k-sat: Two moments suce to cross a sharp threshold. SIAM Journal on Computing , 36, 2006. [2] D. Achlioptas and A. Naor. The two possible values of the chromatic number of a random graph. Annals of Mathematics , 162:1335{1351, 2005. [3] D. Alberici and P. Contucci. Solution of the monomer{dimer model on locally tree-like graphs. Communications in Mathematical Physics , 331:975{1003, 2014. [4] D. Aldous and J.M. Steele. The objective method: probabilistic combinatorial optimization and local weak convergence. Encycl. Math. Sci. , 110:1{72, 2004. [5] N. Alon and J.H. Spencer. The Probabilistic Method . Wiley, New York, 1992.54 ZHONGYANG LI [6] I. Bezakova, A. Galanis, L.A. Goldberg, H. Guo, and D. Stefankovic. Approximation via when strong spatial mixing fails. SIAM Journal on Computing , 48:279{349, 2019. [7] N. Bhatnagar, A. Sly, and P. Tetali. Decay of correlations for the hardcore model on the d-regular random graph. Electron. J. Probab. , 21:1{42, 2016. [8] B. Bollob as. A probabilistic proof of an asymptotic formula for the number of labelled regular graphs. Euro. J. Combinatorics , 1:311{316, 1980. [9] C. Bordenave, M. Lelarge, and J. Salez. Matchings on innite graphs. Probability Theory and Related Fields , 157:183{208, 2013. [10] A. Coja-Oghlan and L. Zdeborova. The condensation transition in random hypergraph 2-coloring. SODA '12, Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms , pages 241{250, 2012. [11] C. Cooper, A Frieze, M. Molloy, and B. Reed. Perfect matchings in random r-regular, s-uniform hyper- graphs. Combinatorics, Probability and Computing , 5:1{14, 1996. [12] A. Dembo and A. Montanari. Ising models on locally tree-like graphs. Ann. Appl. Probab. , A. Dembo, A. Montanari, and N. Sun. Factor models on locally treelike graphs. Ann. Probab. , 41:4162{ 4213, 2013. [14] J. Ding, A. Sly, and N. Sun. Maximum independent sets on random regular graphs. Acta Math. , 217:263{ 340, 2016. [15] G. Elek and G. Lippner. Borel oracles. an analytical approach to constant-time algorithms. Proc. Amer. Math. Soc. , 138:2939{2947, 2010. [16] P. Erd os and A. R enyi. On the existence of a factor of degree one of a connected random graph. Acta Math. Acad. Sci. Hung. , 17:359{368, 1966. [17] A. M. Frieze and S. Janson. Perfect matchings in random s-uniform hypergraphs. Random Structures and Algorithms , 7:41{57, 1995. [18] A. Galanis, D. Stefankovic, and E. Vigoda. of the partition function for the ising and hard-core models. Combinatorics, Probability and Computing , 25, 2016. [19] D. Gamarnik and D. Katz. Sequential cavity method for computing free energy and surface of Statistical Physics , 137:205{232, 2009. [20] G. Grimmett and Z. Li. Critical surface of hexagonal polygon model. Journal of Statistical Physics , 163:733{753, 2016. [21] G. Grimmett and Z. Li. The 1-2 model. Contemporary Mathematics , 696:139{152, 2017. [22] G. Grimmett and Z. Li. Critical surface of the 1-2 model. Int. Math. Res. Not. IMRN , G.R. Grimmett and C.J.H. McDiarmid. On colouring random graphs. Math. Proc. Cambridge Phi- los.Soc. , 77:313{324, 1975. [24] C. Grinstead and J. L. Snell. Introduction to Probability . American Mathematical Society, 1997. [25] E.H. Heilmann, O.J.and Lieb. Monomers and dimers. Phys. Rev. Lett. , 24:1412{1414, 1970. [26] O.J. Heilmann and E.H. Lieb. Theory of monomer{dimer systems. Commun. Math. Phys. , M. Jerrum. Two-dimensional monomer-dimer systems are computationally intractable. Journal of Sta- tistical Physics , 187:121{134, 1987. [28] P.W. Kasteleyn. The statistics of dimers on a lattice. i. the number of dimer arrangements on a Physica , 27:1209{1225, 1961. [29] C. Kenyon, D. Randall, and A. Sinclair. Approximating the number of monomer-dimer coverings of a lattice. Journal of Statistical Physics , 83:637{659, 1996. [30] R. Kenyon. Lectures on dimers. In Statistical Mechanics , pages 191{230. American Mathematical RI, 2009.

2108.11003

Zhongyang Li

Zhongyang Li

Matchings on Random Regular Hypergraphs

math.PR cs.DM

http://creativecommons.org/licenses/by/4.0/

We prove the convergence in probability of free energy for matchings on random regular, uniform hypergraphs and explicitly compute the limit. We also obtain a region on parameters where replica symmetry holds.

Appl. Phys. and iBIB=Ja. As we will discuss later, one of the key parameters is Jb=Ja, which indicates the asymmetry of two JJs. By numerically solving Eqs. (6) and (7) for ex=0,IV curves with step structures are obtained as shown in Figs. 2 (a) and (b). In Fig. 2 (a) for =1:0 and 2. (a)IVcurves with =!0=0:1,=1:0, and iac=0:5, for =1:0 (red upper triangles), =0:8 (blue lower triangles), and =0:6 (purple circles). For clarity, the latter two curves are vertically shifted by iac =0.2 and 0.4, respectively. (b) IVcurves with =!0=0:1 and=1.0 (green circles), iac=0.2 and=1.0 (red iac=0.5 , and=0.2 (black crosses). The first curve (green circles) is s vertically shifted by iac=0.2 for clarity. are plotted for =1.0 (red upper triangles), =0.8 (blue lower triangles), and =0.6 (purple circles). For clarity, the latter two curves are vertically shifted by iac=0.2 and 0.4, re- spectively. The half-integer SS are suppressed by decreasing , which is controlled by changing the ratio of junction areas, WaandWb. Making both Josephson coupling and resistance of two JJs equivalent is a key to observe the half-integer SS. In Fig. 2 (b) for =1,IVcurves are plotted for iac=0.2 and =1.0 (green circles), iac=0.5 and=1.0 (red upper tri- angles), and iac=0.5 and=0.2 (black crosses). The heightof half-integer SS is enhanced by increasing iac, whereas it is suppressed by decreasing . Whenis small by decreas- ing the SQUID loop, the half-integer Shapiro steps can be observed by increasing the iac.is estimated as 1 and L4.7 pH. It satisfies the criteria to overcome the thermal noise, i.e., L must be less than about 20 nH at 4.2 K.3) The half-integer SS can be understood using the In the first order of ,iBis given the second term in Eq.(8) including sin 2  is the origin of the half-integer SS. By ap- plying a voltage (9) where and the kth or- der Bessel function Jk(b). For the-SQUID with ex=0, the first term in Eq. (9) is zero because =0=1/2. When Vsatisfies V=(  0=2)=k0=2, the SS with a and an integer q(k0=2q) appear with a for(0)==4. Meanwhile, in the con- ventional SQUID with ex=0, the second term in Eq. (9) is zero because =0=0. Only integer SS appear at volt- ages of integer multiples of V=(  0=2)=kwith a DC- component, 4[ half-integer SS are explained by the onset of sponta- neous current, leading to a flip-flop between two fluxoid states of the-SQUID synchronized to the alternating field.13, 14)It suggests that the present system can become a qubit. Equation (8) shows that the two potentials with and 2are to the sds-wave JJ.22, 23, 28)The half-integer SS and - qubit are the two sides of the same coin. The potential energy of the-SQUID U(a;b) with Eq. (3) is given (12) where x, and y(=0). The ground state is obtained by minimizing f(x;y) with respect to yfor a fixed which y0is determined as a function of x, i.e., (13) means that f(x;y) is minimized with respect to ex, since we study the -SQUID with two JJs shown in Fig. 1. To avoid ex, the minimization with respect to excan be substituted by another Josephson phase including one more 2Appl. Phys. Express JJ in the-SQUID, as discussed in the previous studies.28, 31) Numerically solving Eq. (13), we find that f(x;y0(x)) has the double minimum with respect to xas shown in Fig. 3. The Fig. 3. The potential energy for =1.0 with=1.0 (red =0.8 (blue lower triangles), and =0.6 (black circles). right minimum corresponds to the current circulating state, whereas the left one has no circulating current. Similar to the previous case,30)the barrier height is suppressed by decreas- ing, which coincides with the suppression of the half-integer SS. When the barrier height was zero, the stable state is x= /2, which may be realized by setting the phase-lock to a= busing Eqs. (3) and (4). In this case, the spontaneous loop current IaIb=0. Then, the SS appear at voltages with in- teger multiples. The two minima correspond to the clockwise and anticlockwise loop currents in addition to the -shift be- cause of the -JJ, meaning that the spontaneous loop current is induced. Then, aandbare not synchronized as discussed in the loop current means the time- evolution of the phase di erence in each junction. It is useful to compare the present model with the previous one, in which a metallic transport is assumed in the -JJ.30) The Josephson current of the metallic junction assigned to the junction- bis given (14) with transmittance case also shows the half-integer SS as shown in Fig. 4, although its small when Tapproaches 1, where the magnitude of the metallic junction becomes large and dominates the cur- rent. Even in such a metallic case, the half-integer SS can be observed by increasing the iac. The potential energy corresponding to Eq. (12) is given ground state is obtained by minimizing g(x;y) with re- Fig. 4. T-dependence of IVcurve with =!0=0.1,=1.0, and= 1.0. Each line is obtained for T=0.2 and iac=0.5 (red upper triangles), T= 0.5 and iac=0.5 (blue lower triangles), and T=0.5 and iac=1.0 to yfor a fixed solving Eq. (16), we find that g(x;y0(x)) ex- hibits the double minimum with respect to xas shown in Fig. 5. Because the magnitude of potential highly depends plotted instead of g(x;y0(x)) for clarity. The rather insulating case with =0.6, which corresponds Fig. 5. Potential energy for =0.6 and=1.0 with T=0.8 (black circles), T=0.5 (blue lower triangles), and T=0.2 (red upper triangles). to the black circles in Fig. 3, shows the shallow minimum, whereas the double minimum becomes clear by increasing T; this would contradict to the suppression of the half-integer SS in Fig. 4, i.e., red upper triangles and blue lower it is caused by a large magnitude of potential in the metallic junction. In fact, the half-integer SS are revived by 3

2108.11006

Michiyasu Mori

Michiyasu Mori and Sadamichi Maekawa

Half-integer Shapiro-steps in superconducting qubit with a $pi$-Josephson junction

4 pages, 6figures

APEX 14, 103001 (2021)

10.35848/1882-0786/ac211d

cond-mat.supr-con cond-mat.mes-hall

http://creativecommons.org/licenses/by/4.0/

A superconducting quantum interference device (SQUID) comprising 0- junctions (JJs), called $pi$-SQUID, is studied by the resistively shunted junction model. The $pi$-SQUID shows (SS) under microwave irradiation at the voltage $V$ (n/2)$, with angular frequency $Omega$ and in addition to integer $n$. We show that the $pi$-SQUID can be a $pi$-qubit with spontaneous loop currents by which the half-integer SS are induced. Making the 0- and $pi$-JJs equivalent is a key for the half-integer SS and realizing the $pi$-qubit.

catastrophic forgetting [30]. Though previous works [1, 21, 29, 40, 48] attempt to address the issue, they are offline and do not work well on structural predic- tion tasks like object detection. What hinders the progress in online continual learning is the lack of realistic datasets and benchmarks. Most of the current research [32, 41, 49] re-purpose existing static datasets such as VOC and COCO to evaluate continual object detection. These approaches use object categories one by one in a sequential manner. These manual splits and artificial setups differ from the sce- narios often encountered by embodied agents, where the emergence of new tasks often follows the trajectories of the agents and the frequencies of the object instances vary from task to task. For example, the agents might observe the in- stances of the same category after a few hours or even days. They may visit some objects more often than others and re- visit previously observed objects. In this paper, we present a new online continual object detection benchmark. Our benchmark consists of a new labeled dataset - OAK (Objects Around Krishna). OAK uses the videos from the KrishnaCam [43] dataset - an ego- centric video dataset collected over nine months of a grad- uate student's life. OAK contains 80 labeled video around 17.5 hours (roughly 1/4 of the raw videos in KrishnaCam) with bounding box annotations of 105 in outdoor scenes. OAK provides a natural and the task emergence following the trajecto- ries of a single person. A few objects frequently appear due to redundancy in daily routines, while new objects and cat- egories constantly appear as they visit various places. This dataset is a realistic playground to study online enables the embodied agent to learn from a hu- man's experience, and provides a unique opportunity for re- searchers to pursue the essence of lifelong learning by ob- serving the same person in a long time span. We introduce several new evaluation metrics in the on- line continual learning setup. In contrast to the previous task incremental or class incremental settings in continual learn- ing, there is no explicit task boundary in our setup and new tasks emerge following the temporal order in the we evaluate the overall performance precision, CAP), transfer trans- fer, BWT/FWT), and forgetting (forgetfulness, F) of the models with an additional temporal dimension. We evalu- ate the models periodically on the frames held out from the same training video frames. The overall performance is ag- gregated from these evaluations and the defined by the time intervals between the appearances of instances from the same tasks. We adapt several typical con- tinual learning algorithms (e.g., iCaRL [35], EWC [21], In- cremental fine-tuning) to object detection and find the per- formance of these approaches mediocre in the new bench- mark, which leaves substantial room for future studies.2. Related Work Continual learning benchmarks. A large body of contin- ual learning algorithms [1, 21, 29, 40, 48, 25, 46, 39, 45, 18, 24] have been developed and evaluated on image classifica- tion benchmarks such as Permuted MNIST [14], CIFAR- 100 [22], and ImageNet [9]. More recently, Lomonaco and Maltoni [28] introduced CORe50, a collection of 50 domes- tic objects belonging to 10 categories, which supports im- age classification at object level (50 classes) or at category level (10 classes) and object detection in a recent update. In contrast to our work, the task splits in CORe50 are created manually and the benchmark is used for offline learning. In the object detection domain, several incremental ob- ject detection algorithms [23, 27, 32, 41] adopt existing object detection datasets such as PASCAL VOC [11] and MS COCO [26] for evaluation. They split the categories and train the object detectors on a pre-defined order of the categories sequentially. Chen et al . [6] study the prob- lem of continual learning by building NEIL with inter- net images. Kuznetsova et al . [23] extend the evaluation from static images to video snippets like Activity Daily Living (ADL) [33] dataset and YouTube Objects (YTO) datasets [34] for incremental domain adaptation. The task splits in these benchmarks are often and the classes in the datasets are carefully bal- anced. Moreover, existing benchmarks are mostly con- structed using static images [9, 14, 22] or samples from short object-centric video snippets ( 15 seconds) [28]. These existing benchmarks may not sufficiently unveil the challenges of building human-like continual learning agents given their artificial settings and static data sources. Online continual learning settings. There is an emerging line of works on online continual learning [13, 2, 4, 3, 36]. Aljundi et al. [3] develop a system that keeps on learning over time in an online fashion, with data distributions grad- ually changing and without the notion of separate tasks. Ren et al. [36] recently extend the standard framework for few-shot learning to an online, continual setting. Similar to Aljundi et al., our online continual learning settings do not have an explicit task boundary. Since the tasks emerge as the person moves around, we introduce new metrics to evaluate the model transfer and forgetting through video recognition. Another related is ego-centric video recognition [38, 12, 20]. Most works focus on developing methods that are suitable for the unique perspective in an ego-centric video using an offline setting [47, 20]. To this sense, the most related work is [10], where they tackle the problem of continual image However, the dataset they use is [28] extremely short and clean, which cannot be sufficient to unveil the challenges in ego-centric video 2: Examples of annotated frames in OAK. OAK captures daily outdoor activities of a single graduate student. The dataset spans a wide variety of environments and life experiences. 3. OAK Dataset We introduce Objects Around Krishna (OAK), which la- bels the objects that appear as Krishna wanders and per- forms his daily routine. OAK is built upon the Krish- naCam [43] dataset, a large egocentric video stream span- ning nine months of a graduate student's life. The origi- nal KrishnaCam dataset contains 7.6 million frames of 460 video snippets with a total length of 70.2 hours. The raw videos have a resolution of 720p and a frame rate of 30 fps. In OAK, we label around 1/4 of the original dataset and re- lease the labeled data in the setup. We consider 105 object categories in outdoor scenes where 16 classes are from the PASCAL VOC dataset [11] and the remaining categories are frequent classes determined by running the LVIS [15] pre-trained Mask R-CNN model [16] on the raw videos. The full list of the categories is provided in the website. We sample 80 videos from KrishnaCam, which uniformly span in time. Each video snippet is about 7 15 minutes long. Objects of the 105 categories are exhaustively labeled at a frame rate of 0.5 fps. In total, OAK contains roughly 326K bounding boxes. Two human annotators are involved in labeling each frame to ensure the label quality. Objects in the frames are exhaustively annotated except for the tiny objects (less than 2020 pixels). In Figure 2, we show some examples of the annotated frames in OAK. Train and evaluation sets. In contrast to the offline set- tings, the model is expected to train while evaluating on the same data stream in an online continual learning setting to evaluate the catastrophic forgetting. Therefore, we hold out one frame every 16 labeled frames to construct an evalua- tion set and the remaining frames are used for training. Thetraining and the evaluation sets cover a similar time range of 9 months with different sample rates. The models will be trained and evaluated in an online manner. Dataset statistics. We show some data statistics in Figure 3. Natural task distribution. In our online continual a new task is defined as recognizing a new object category that was not encountered before. As shown in Fig- ure 3a, the number of seen categories gradually increases and the appearance rate of novel classes decreases over time due to the repeated patterns in daily life. Long-tail distribution. In Figure 3b and 3c, we can see OAK has a long tail distribution both for the number of instances per category and the category counts per image. In Fig- ure 3d, we show the distribution of time intervals between the reappearance of instances from the same category. Skewed box sizes. As shown in Figure 3e, the box sizes in OAK are skewed towards small and medium sizes, which makes it hard for the detector to make correct geo-locations. In Figure 3f, we plot the locations of the video recordings. We can see that places like campus and home are frequently visited while other places are visited. 4. Online Continual Learning Benchmark In this benchmark, we consider two online settings depending on whether the categories of in- terest are known beforehand. In both cases, the training data and labels emerge sequentially following the time stamps and the model is evaluated every N training steps. The main difference is the way that unseen categories are dealt with at the time of evaluation. If the model has a known vocab-(a) The number of categories increase over time but the emergence rate of novel categories decrease over time. (b) The number of instances per cat- egory reveals a long tail existence of many rare classes. (c) Distribution of number of cate- gories per image also reveals a long Distribution of time interval be- tween the reappearances of data points from the same category. (e) Distribution of the bounding box size (pixel space). The distribution is skewed to small / medium size objects. (f) Location distribution. Places in red are often visited and places in blue are occasionally visited. Figure 3: Dataset statistics. Best viewed digitally. ulary of the classes referred as the known setting, we can simply report the average precision (AP) on each category at each evaluation although some categories may not have been trained at the time of evaluation (often leading to a lower evaluation result). The case where the model has an open vocabulary of classes, referred as the unknown setting, is a bit more chal- lenging but also more realistic. We introduce an IDK (I don't know) class for the unseen categories at the current time stamp. For all the objects from the unseen categories in the evaluation set, the model needs to predict IDK for a correct prediction. The average precision of predicting IDK is also part of the evaluation protocol, which indicates the model's ability to identify new classes. For simplicity, we don't require the model to distinguish the exact category among the unseen categories as long as the model predicts IDK for the unseen objects. 4.1. Evaluation Metrics In the online continual learning setting, we focus on three aspects of the learned model: how well does the model perform overall? How well does the model transfer new knowledge? How resistant is the model to catastrophic for- getting? To this end, we introduce five evaluation average precision (CAP) and final average pre- cision (FAP) for overall performance evaluation; forward transfer (FWT) for transfer performance evaluation; back- ward transfer (BWT) and forgetfulness (F) for evaluation. It's worth noting that transfer andforgetfulness are more comparable if two models have sim- ilar CAP. We adopt the commonly used AP50 (i.e., the aver- age precision score with an IoU threshold of 50%) in object detection for measurement. CAP shows the overall performance of the model in the time span of the entire video stream. Inspired by OS- AKA [4], the accuracy of each timestep is evaluated using the current model instead of the final model. At each time t, the model is trained using a small batch ( tb-th train- ing frame to (t+ 1)b-th training frame, where bdenotes batch size) of data from Dtrain. After time t, this small batch of images is no longer allowed to be used. We continue this training process until the entire video stream is covered. Ev- eryNtraining steps, the model is evaluated on the test set Dtest. The reported step) is defined (1) where CAPc tiis the average precision (AP) of the class con the test set. CAP is then defined as the average values across different time stamps. That (2) where Tis the total evaluation times. FAP is the final average precision of the last model when the model finishes training. This is a more fair evaluationmetric when comparing with offline learning models as both models have observed the entire video stream. FWT evaluates the forward transfer ability of new knowl- edge inspired by GEM [29]. FWT shows the influence that learning a scenario (a video clip, denoted as St) has on the performance for future scenarios ( Sk; k > t ). The scenar- ios are a short clip with a fixed interval of 0.94 hour in our benchmark. Positive forward transfer is possible when the model is able to perform zero-shot learning. Specifically, we divide into Tscenarios in temporal order, where each division of as train set and test set of scenario Si. After the model finishes learning from scenario Si, we evaluate its test performance on all Tscenarios. By doing so, we construct the matrix R2RTT, where Ri;jis the test mean average precision (mAP) of the model on scenario Sjafter observing the last frame from Si. Letting bbe the vector of test mAP of an pre-trained object detector for each scenario, we define FWT (3) BWT shows the influence that learning a scenario (a video clip, denoted asSt) has on the performance on (Sk; k < t ). Negative backward transfer is also known as forgetting. Specifically, we define BWT (F) estimates the model forgetting due to the sequential training. For a class c, we sort the to the time interval kbetween evaluation time tiand the last time tikthe model is trained on c. After the CAPc ti is sorted, all CAPc ti(i= 0; : : : ; T ) are divided into Kbins Bkmin; :::; B kmax according to the time interval k. The av- erage CAP (aCAP k) of each bin Bkis defined as the for detecting class cafter the model have not been trained on cforktime stamps. We define of the class cas the weighted sum of the at each kminaCAP k): (5) Thus, the overall forgetting is defined (6) Figure 4: Incremental fine-tuning. The entire object de- tector, including both the feature extractor Fand the box predictor (R&C) are pretrained using PASCAL VOC. In online continual learning, the backbone is fixed, while RPN and box predictor are fine-tuned. 5. Experiments We describe three widely adopted continual learning al- gorithms in Section 5.1 and present their performance on our benchmark in Section 5.2. Although the existing con- tinual learning algorithms generally improve over the non- adaptive model, the CAP value for each of the algorithms is less than 20. This indicates our benchmark is has a large room for future algorithm designs. 5.1. Continual Learning Algorithms We select the three representative continual learning methods according to Parisi et al. [31]. Incremental fine- tuning is intuitive and widely adopted as a baseline for con- tinual learning. iCaRL [35] is a widely used EWC [21] is a representative We first show how each method is deployed to the known setting, and we then show that these methods can easily adapt to the unknown fine-tuning. The first baseline is incremen- tal fine-tuning on the widely used two-stage object de- tector, Faster R-CNN [37], which is pretrained on the PASCAL VOC dataset. As shown in Figure 4, the fea- ture learning components Finclude the backbone (e.g., ResNet [17], VGG16 [42]), the region proposal network (RPN), as well as a two-layer fully-connected (FC) sub- network as a proposal-level feature extractor. There is also a box predictor composed of a box classifier Cto clas- sify the object categories and a box regressor Rto predict the bounding box coordinates. The backbone features, as well as the RPN features, are class-agnostic. However, we find keeping the RPN updated during fine-tuning signifi- cantly improves the performance. Thus, we only keep the backbone fixed during the incremental fine-tuning new data comes in, we fine-tune the RPN, the box classifier, and the regressor on the new data. iCaRL. The second baseline adapts from the iCaRL [35] algorithm proposed by Rebuffi et al. The original iCaRL is designed for image classification and we implement this approach with the Faster R-CNN based detector. As illus- trated in Figure 5, iCaRL additionally includes a memoryFigure 5: Illustraion of iCaRL. The network parameters are updated by minimizing the loss on both streaming data and memorized prioritized store the representative examples of each cate- gory, which are selected and updated randomly during each step. The memory bank has a fixed size (we set it to 5 images per category) and the old examples in the memory bank are replaced with newer data points. For each training step, we would randomly select a sample (image with one object label) from each class of the memory bank to train jointly. Same as incremental fine-tuning, we keep the back- bone fixed and fine-tune the RPN, the box classifier, and the regressor when new data comes in. EWC. The third baseline adapts from the EWC algo- rithm [21], proposed by Kirkpatrick et al. Similar to iCaRL, we apply the EWC algorithm to the box classifiers in Faster R-CNN. EWC does not require access to ground truth labels in the memory bank. The main idea of EWC is to over the gradient updates so that the gradient up- dates on the new examples do not increase the on the old examples as illustrated by Figure 6. Inter- ested readers can refer to the EWC paper [21] for baselines. We provide the performance of the Faster R-CNN pretrained on the PASCAL VOC data, denoted in the result tables. We also provide the model performance using offline training. At each evalu- ation time ti(ithevaluation step), the entire video stream before tiare used as one training set. We then conduct batch training offline and report the CAPtionDtest. This baseline is denoted as offline training in the result tables. Training details. For a fair comparison, we use ResNet- 50 [17] as the backbone of all continual learning algorithms, and the base object detector, Faster R-CNN, is pretrained on PASCAL VOC [11]. OAK share the same 20 categories with the PASCAL VOC dataset and thus the pretrained model can be used as an initial point. In the known setting, the number of categories in the detector is fixed and set to 105. In the unknown setting, the number of categories in- creases through time. We add a column to the box classifier and a box regressor when new data contains previous un- seen categories. At the prediction time, if the confidence of all classes is smaller than a threshold while the confidence of background also maintains a low status, the agent should predict IDK. Figure 6: Illustration of EWC (figure adapted from [29]). When learning to predict a new class B, the gradient up- dates may hurt the performance of the old class A. EWC minimizes the classification loss without incurring a signif- icant loss on the old classes. 5.2. Overall Model class vocabulary. In Table 1, we provide the overall model performance measured by continual ( CAP) and final average precision ( FAP) under theknown setting, where the vocabulary of the classes is known beforehand. The pretrained model on PASCAL VOC (Non-adaptation) only achieves 2.86 points on the new OAK dataset. Even for the classes that overlap with the original PASCAL VOC data (e.g., chair, dining table, etc.), the performance is extremely low. This indicates that OAK has a large domain gap with the existing detection dataset, which uses static Internet images. For the three continual learning algorithms, the memory- based approach iCaRL outperforms the vanilla incremen- tal fine-tuning. This demonstrates that the simple can still play an important role in the new on- line continual learning setting. However, the approach EWC does not help with the overall perfor- mance in our setting, which might be due to the complica- tion of the mixed task orders. It is also worth noticing that both EWC and iCaRL have a significant performance gap compared with the offline training. This indicates that the new online continual learning setting may require more in- novation in the algorithm designs to deal with the new chal- lenges posed by online learning. Additionally, since both of them are not designed for object detection, the model per- formance may be improved if one can design models spe- cialized for object detection. We notice that the model performance of the offline training is only about 49.48 points. It is a relatively low model performance compared with the model performance on other detection datasets such as PASCAL VOC or COCO where the AP50 scores are often above 50 points. This gap is also discovered by other egocentric video chal- lenges [8, 38]. We conjecture the new challenges posed by OAK comes from the intrinsic features of such as motion blur, lots of occlusion and objects due to limited field of view and the long tail distribution. We visualize the CAP for each class during online con-Table 1: Overall performance of existing algorithms on OAK measured by continual average precision (CAP) and final aver- age precision (FAP) with known class vocabulary ( known ). FAP is consistently better than CAP, which shows all from learning from more data. Due to the distribution shift, the non-adaptation model has the lowest performance. The EWC approach performs similar to vanilla incremental fine-tuning. While the memory-based iCaRL approach is significantly better than vanilla incremental fine-tuning. The offline training results show that OAK is a bit chal- lenging to the current object detector even under the offline training setting. Method FAP CAP Top-20 Booth Umbrella Awning Bag Chair Dining Table Fireplug 2.86 2.86 11.72 0.00 0.00 0.00 0.00 23.88 26.38 0.0 12.47 10.47 33.84 0.20 2.46 4.58 18.24 31.70 36.51 39.53 69.68 EWC 12.55 10.52 33.80 0.21 3.02 4.53 17.79 31.38 36.67 38.73 69.71 iCaRL 21.89 16.39 40.92 6.88 6.91 3.59 10.67 28.72 40.59 45.18 67.18 Offline training 49.48 37.11 67.77 20.05 35.86 28.65 42.69 66.51 71.09 64.15 78.80 Figure 7: CAP tichanges on OAK test set, in chronological order. CAP tiincreases as the model is trained on frames where each category exists, but as more time passes between examples, the forgetting effect is observed and the performance falls. Table 2: Overall performance of the existing algorithms on OAK measured by continual average precision (CAP) and final average precision (FAP) with an unknown class vocabulary ( unknown ). In this evaluation, the model needs to predict IDK for the objects in the unseen categories. All the continual learning algorithms considered have a decreased with the known setting. Again Incremental finetuning and EWC achieve similar performance. iCaRL fine-tuning and EWC but there is substantial room for FAP Overall Top-20 Booth Umbrella Awning Bag Chair Dining Table Fireplug Car 2.86 2.86 11.72 0.00 0.00 0.00 0.00 23.88 26.38 0.0 55.90 - Incremental 11.19 9.78 32.10 0.06 1.45 3.87 16.68 29.26 32.81 36.32 68.76 0.32 EWC 10.34 9.57 31.62 0.21 1.63 3.72 16.46 29.76 33.39 33.41 68.76 0.29 iCaRL 17.56 14.70 41.42 3.58 4.83 2.72 10.18 23.81 32.86 39.84 64.02 0.36 Offline training 46.40 36.88 71.35 18.90 31.90 29.36 41.36 64.59 70.57 63.77 78.36 1.21 tinual learning. As shown in Figure 7, the shading indicates when the agent observes new data points from a specific class, while the points on each curve are evaluation steps on OAK. Notice how in each class the performance but as more time passes between examples, the for- getting effect is observed and the performance falls. When the next round of data is seen towards the end, the perfor- mance improves again. Unknown class vocabulary. In Table 2, we provide the overall performance of the algorithms with an unknown class vocabulary measured by CAP and FAP. In this eval-uation, the model needs to predict IDK for the objects in the unseen categories, which is more challenging than the known setting. As we can see from the table, all the con- tinual learning algorithms considered have a decreased per- formance compared with the known setting. Again incremental fine-tuning and EWC but there is substantial room for Transfer and In the first column of Table 4, we compare the forward transferability of three continual learning al-Table 3: Forgetfulness of the continual learning algorithms. The forgetfulness (F) metric indicates the catastrophic forgetting and smaller F scores mean less forgetting on the learned knowledge. iCaRL is better at avoiding catastrophic forgetting than the incremental fine-tuning and EWC. Method Overall Worst-20 Booth Umbrella Awning Bag Chair Dining Table Fireplug Car Incremental 5.77 22.28 0.60 1.10 1.28 -1.95 2.91 -0.29 11.89 5.75 EWC 5.96 23.48 0.43 1.85 1.13 -2.00 2.76 0.02 10.30 5.76 iCaRL 1.78 14.72 -5.65 2.03 0.67 -1.41 1.50 0.68 6.79 3.32 Figure 8: Forgetfulness of sample categories. We provide the evaluation curve on the test set under different time stamps. Method FWT BWT Incremental 15.01 -4.51 EWC 14.84 -3.62 iCaRL 15.73 -5.95 Table 4: Knowledge transferability under the known set- ting. Incremental fine-tuning has better forward and worse backward transferability. This shows that gradient regularization can alleviate catastrophic forgetting at the cost of information gain. gorithms under the known setting. Higher FWT indi- cates a faster learner. Incremental fine-tuning has higher FWT compared with EWC. EWC is more learning from new data points with regularization of the weight change. We don't compare transferability for unknown setting since the tasks keep In Table 3, we present the forgetfulness of the continual learning algorithms under the known setting. The forgetfulness (F) metric indicates the amount of catas- trophic forgetting. Smaller F scores indicate less forgetting on the learned knowledge. iCaRL is significantly better at avoiding catastrophic forgetting than the incremental fine- tuning method and EWC. For more comparison with ex- isting metrics, we also compare the backward of three continual learning algorithms under the known setting (second column of Table 4). Higher BWT indicates less forgetting. EWC has higher BWT than since EWC finds solutions for new tasks with- out incurring significant losses on old tasks. We choose for-getfulness (F) as our main metric to estimate the due to the sequential training. We also visualize the forgetfulness changes of each class during online continual learning. As shown in Figure 8, the forgetfulness increases as the time from seeing the last labeled data increases. Notice that in most cases, the incre- mental fine-tuning baseline performs best. 6. continual learning from continuous data streams in dynamic environments is of an increasing interest in the machine learning and computer vision community. How- ever, realistic datasets and benchmarks, especially for are still missing. In this work, we presented a new online continual object detection benchmark dataset called OAK. OAK uses the videos from the KrishnaCAM dataset, which features an ego-centric video stream collected over a nine-month time span. OAK provides exhaustive annota- tions of 80 video snippets ( 17.5 hours) with 326K bound- ing boxes of 105 object categories in outdoor scenes. Our continual learning benchmark follows the life pattern of a single person, which is more realistic compared to the ex- isting continual learning benchmarks. We introduced new evaluation metrics as well as baseline evaluations for two evaluation setups. We hope this work inspires research in embodied online continual learning of object The authors would like to thank Yi Ru Wang, Samantha Powers, Kenneth Marino, Shubham Tulsiani for fruitful discussion and detailed feedback on the manuscript. Carnegie Mellon Effort has been supported by DARPA MCS, ONR Young Investigator, ONR Rahaf Aljundi, Francesca Babiloni, Mohamed Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 139-154, 2018. 2 [2] Rahaf Aljundi, Lucas Caccia, Eugene Belilovsky, Massimo Caccia, Min Lin, Laurent Charlin, and Tinne Tuytelaars. On- line continual learning with maximally interfered retrieval. arXiv preprint arXiv:1908.04742 , 2019. 2 [3] Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuyte- laars. Task-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and , pages 11254-11263, 2019. 2 [4] Massimo Caccia, Pau Rodriguez, Oleksiy Ostapenko, Fab- rice Normandin, Min Lin, Lucas Caccia, Issam Laradji, Irina Rish, Alexande Lacoste, David Vazquez, et al. Online fast adaptation and knowledge accumulation: a new approach to continual learning. arXiv preprint arXiv:2003.05856 , 2020. 2, 4 [5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European Confer- ence on Computer Vision , pages 213-229. Springer, 2020. 1 [6] Xinlei Chen, Abhinav Shrivastava, and Abhinav Gupta. Neil: Extracting visual knowledge from web data. In on Computer Vision , 2013. 2 [7] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS Journal of Pho- togrammetry and Remote Sensing , 117:11-28, 2016. 1 [8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Pro- ceedings of the European Conference on Computer Vision (ECCV) , pages 720-736, 2018. 1, 6 [9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , pages 248-255. Ieee, 2009. 2 [10] Luca Erculiani, Fausto Giunchiglia, and Andrea egocentric object recognition. arXiv , 2019. 2 [11] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual ob- ject classes challenge: A retrospective. International Journal of Computer Vision , 111(1):98-136, Jan. 2015. 2, 3, 6 [12] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In CVPR 2011 , pages 3281-3288. IEEE, 2011. 2 [13] Enrico Fini, St 'ephane Lathuili `ere, Enver Sangineto, Moin Nabi, and Elisa Ricci. Online continual learning under ex- treme memory constraints. In European Conference on Com- puter Vision , pages 720-735. Springer, 2020. 2 [14] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catas-trophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211 , 2013. 2 [15] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5356-5364, 2019. 1, 3 [16] Kaiming He, Georgia Gkioxari, Piotr Doll 'ar, and Ross Gir- shick. Mask r-cnn. In Proceedings of the IEEE on computer vision , pages 2961-2969, 2017. 1, 3 [17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and , pages 770-778, 2016. 5, 6 [18] Khurram Javed and Martha White. Meta-learning for continual learning. arXiv , 2019. 2 [19] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via In Proceedings of the IEEE/CVF on Computer Vision , pages 8420-8429, 2019. 1 [20] Georgios Kapidis, Ronald Poppe, Elsbeth Van Dam, Lucas Noldus, and Remco Veltkamp. Egocentric hand track and object-based human action recognition. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data of People and Smart City Innovation , pages 922-929. IEEE, 2019. 2 [21] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka et al. Overcoming catastrophic forgetting in neu- ral networks. Proceedings of the national academy of sci- ences , 2017. 2, 5, 6 [22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2 [23] Alina Kuznetsova, Sung Ju Hwang, Bodo Rosenhahn, and Leonid Sigal. Expanding object detector's horizon: Incre- mental learning framework for object detection in of the IEEE Conference on Computer Vision and Pattern Recognition , pages 28-36, 2015. 2 [24] Soochan Lee, Junsoo Ha, Dongsu Zhang, and Gunhee Kim. A neural dirichlet process mixture model for task-free con- tinual learning. arXiv preprint arXiv:2001.00689 , 2020. 2 [25] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelli- gence , 2017. 2 [26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll 'ar, and C Microsoft coco: Common objects in context. In European conference on computer vision , pages 2014. 2 [27] Xialei Liu, Hao Yang, Avinash Ravichandran, Rahul Bhotika, and Stefano Soatto. Multi-task incremental learn- ing for object detection. arXiv preprint arXiv:2002.05347 , 2020. 2

2108.11005

Jianren Wang

Jianren Wang, Xin Wang, Yue Shang-Guan, Abhinav Gupta

Wanderlust: Online Continual Object Detection in the Real World

ICCV 2021

cs.CV cs.LG

http://creativecommons.org/licenses/by/4.0/

Online continual learning from data streams in dynamic environments is a critical direction in the computer vision field. However, realistic benchmarks and fundamental studies in this line are still missing. To bridge the gap, we present a new online continual object detection benchmark with an egocentric video dataset, Objects Around Krishna (OAK). OAK adopts the KrishnaCAM videos, an ego-centric video stream collected over nine months by a graduate student. OAK provides exhaustive bounding box annotations of 80 video snippets (~17.5 hours) for 105 object categories in outdoor scenes. The emergence of new in our benchmark follows a pattern similar to what a single person might see in their day-to-day life. The dataset also captures the shifts as the person travels to different places. These videos provide a realistic playground for continual especially in online embodied settings. We also introduce new evaluation metrics to evaluate the model performance and and provide baseline studies for online continual object detection. We believe this benchmark will pose new exciting challenges for learning data in continual learning. The OAK dataset and the are released at

2 into account the inhomogeneity with respect to the dis- tance from a dopant to a Ce atom. We break the equiv- alence of Ce atoms by placing the dopant close to a spe- cic Ce atom. The electronic structure is investigated by DFT and DFT+DMFT calculations. The DFT results give a hint about the dierent behavior of hole versus electron substitutions, while its limitation is also shown. The small change in the hybridization in the DFT quasi- particle could turn into a totally dierent phase through the DMFT impurity solver. The inhomogeneity of the self-energy and the hybridization functions is shown in the DFT+DMFT results. The dierent evolution of of the heavy quasiparticle band induced by the dopant will be discussed with the calculated coherence temperature. The paper is organized as follows: Section II explains the computational method. In Sec. III we address cal- culated crystal structures, the DFT and Also, the calculated coherence temperature will be provided here. Section IV presents a summary and concluding remarks. II. COMPUTATIONAL DETAILS The charge self-consistent version of DFT+DMFT,29 as implemented in Ref. 30, is based on the full- potential linearized augmented plane-wave (FP-LAPW) band method.31The correlated 4 f-electrons are by the DMFT local self-energy (( !)), while all other delocalized spdelectrons are treated on the DFT level. The charge and spin uctuations considered in DMFT enable the describtion of the Kondo eect cor- rectly. (!) is calculated from the corresponding im- purity problem, in which full atomic interaction matrix is taken into account ( F0= 5.0 eV,F2= 5.41543352601 eV and, F6= temperature of 5 meV is used in the calculations. To solve the impurity problem, we used the one-crossing ap- proximation of the local 7.0 and 3000 k-point mesh are used through the calculation. The spin-orbit coupling is always included in the total charge density and the self-energy through the DFT+DMFT iterations are converged within 105and 104, single-site DMFT solver for the 4 fstate in each unique Ce cite are treated separately. The 4 fstate for the unique Ce sites will be denoted as Ce1 or Ce2. The inter-site correlation beyond the single-site DMFT solver is included through the hybridization function (
). equation between the local Green's func- tion (G impurity energy level, : the self-energy,
: the hybridization function between impurity and bath) and the lattice green function com- puted by the DFT eigenvalue ( HDFT k) andEF, is dened through the projection operator ( P) from the lattice to the impurity and the embedding operator ( E) vice versa. Using and ^E2;k(1:Ce1, 2:Ce2), the self- consistent equations with the diagonal basis are dened FIG. 1. (color online) (a)The schematic crystal structures33 of the 1x1x2 supercell of CeCoIn 5The dierent colors repre- sent the dierent atoms. Indium (In) atoms have the magenta (In1) and purple (In0) colors. For the doped case, the dopant atom replaces In1 to give the concentration of Additionally, the two-types of Ce atoms are distin- guished to Ce1 (close to the dopant) and Ce2 (away from the dopant). (b) The atomic congurations of Cd, In, Sn, and Hg are compared. (c) The primitive unit cell of CeCoIn 5is provided. as ,i;imp, i, and
imean the impurity energy and hybridization function, respectively. i represents the dierent impurity (Ce). The eect is taken into account through
i. III. NUMERICAL RESULTS Figure 1(a) presents the supercell structure of CeCoIn 5 made from the original primitive unit cell as shown Fig. 1(c). We adopt the minimum size of the 1x1x2 supercell of CeCoIn 5, which has two Ce, two Co and ten In atoms for our calculations. Ce(1) at (0,0,0) and Ce(2) at (0,0,0.5) are equivalent in the pristine struc- ture. In the primitive cell, In atoms are distinguished as the ones (In(1)) in the Ce plane and the other four3 FIG. 2. (color online) The density of states for (a) Ce1-4 f, (b) Ce2 -4f, and (c) Co-3 dstates calculated in four with DFT. The labels represent CeCoIn CeCoIn 5), CeCo(In CeCoIn 5), and CeCo(In 5). The red vertical line represents the Fermi level. The blue vertical line is provided as a guideline to compare the shift of the bands with respect to the dopant. In atoms (In(2)) in between the Co plane and the Ce-In plane. For simplicity, we choose the dopant atom (Cd,Sn or Hg) to occur at the In(1) position of the Ce(1) plane. This is consistent with extended X-ray absorption ne structure measurement, which indicate that these dopant atoms preferentially substitute on the In(1) site. force Ce(1) and Ce(2) to become inequiv- alent. In this work, we examine the electronic struc- ture of CeCoIn 5(pristine), CeCo(In CeCoIn 5), CeCo(In CeCoIn 5), and CeCo(In CeCoIn 5). First we report the conventional DFT calculations of the doped CeCoIn 5. Figure 2 shows the density of states (DOS) of (a) Ce(1)-4 fstates, (b) Ce(2)-4 fstates, and (c) Co-3dstates. We note that the DOS for the two Co atoms are identical due to the inversion on the dopant, the Ce-4 fbands are shifted with respect to the pristine case (red line). The two peaks in Fig. 2(a-b) correspond j=5/2 (left peak) and j= 7/2 (right peak) states split by spin-orbit coupling (  0.3 eV). This value is inherited into the impurity solver in the DFT+DMFT calculation. Hereafter we focus on the j=5/2 states, which mainly contribute the states around EF. As Cd and Hg (Sn) have one less (more) electron than In (See Fig. 1(b)), the shift of the chemical potential FIG. 3. (color online) The imaginary part of the lo- cal Green's functions for (a) Ce1 and (b) Ce2 in dif- ferent structures with DFT+DMFT.The labels represent CeCoIn 5(Pristine), CeCo(In 0:9Cd0:1)5, CeCo(In 0:9Hg0:1)5, and CeCo(In 0:9Sn0:1)5. with respect to Ce-4 fstates would be lowered (increased) in the Cd- and Hg-doped (Sn-doped) cases. The trend in Fig. 2(a-b) for Ce(1)-4 fand Ce(2)-4 fin the pristine and doped cases does not follow this simple intuition. In comparison to the j=5/2 peak of the pristine Ce-4 f states, the peak for Cd- and Hg-doped (Sn-doped) cases is shifted toward (away from) the Fermi energy level. Also, the DOS at EFare increased in Cd- and Hg-doped CeCoIn 5systems, whereas that of the Sn-doped CeCoIn 5 is decreased. Those results indicate the eect of doping on the 4fstates could not be described well by the DFT method. On the other hand, the inhomogeneous feature between Ce(1) and Ce(2) sites could be identied. We also explore the doping eects on the 3 d-bands Co. As shown in Fig. 2(c), the band center of Co-3 dstates is shifted toward EFin the Cd- and Hg-doped CeCoIn 5cases, while that of the Sn-doped CeCoIn 5is moved away from EF. Such behavior is in agreement with a rigid band picture. In the remainder, we will discuss how correlations in- uence the electron- and hole-doped cases. Through the DFT+DMFT calculations, Ce(1)-4 fand Ce(2)-4 fstates are treated as correlated impurity states, denoted by Ce1 and Ce2. The DOS ( ImG loc) of Ce1 and Ce2 are pro- vided in Fig. 3. In CeCoIn 5, the peak of j=5/2 states4 TABLE I. Ce-4 felectron occupancy, nf, and probabilities of 4f0, 4f1, andf2for Ce1 in the DFT+DMFT atom) Ce1 Ce2 5 1.0432 1.0432 0.0341 0.9126 0.0532 CeCo(In 0:9Cd0:1)5 1.0299 1.0329 0.0297 0.9267 0.0434 CeCo(In 0:9Hg0:1)5 1.0214 1.0362 0.028 0.9224 0.0495 CeCo(In 0:9Sn0:1)5 1.0822 1.0452 0.064 0.8711 0.0646 TABLE II. Summary of the hybridization strength Im
(0). Im
(0) (eV) Ce1 Ce2 CeCoIn 5 0.08999 0.08999 CeCo(In 0.07345 CeCo(In 0.08064 CeCo(In 0.0982 grows as temperature lowers. Fig. 3 is a snapshot in the middle of the development of the quasiparticle peak. In comparison to Fig. 2 (a,b), the width of the peaks are highly reduced and pushed towards EF. The DOS at EFof the Ce1 in the Sn-doped case is much larger than others. Also no spin-orbit side band at 0:3 eV could be observed. The spin-orbit side band is attributed to the transition between j=5/2 andj=7/2 states. The Sn dopant due to the increased hybridizations pushes the Ce1 away from the Kondo limit. The enhanced DOS of the Ce1 is associated with its closer proximity to the Fermi liquid in CeCo(In 0:9Sn0:1)5. In case of the hole- doped calculations (Cd or Hg), Ce1 shows a reduction of the DOS relative to the pristine case. This is consistent with stabilizing a magnetic state upon hole doping. We examine the 4 felectron occupancy ( nf) for the pristine and doped cases. As summarized in Table I, Sn-doped case is larger than the others. This trend looks contradicting to the intuitive view of va- lence change from 4 f1(Ce3+) tof0(Ce4+) under Therefore, within a simple one would expect that the hole (electron) doping would drive the valence of Ce states toward f1(f0). Here we bring up a new insight: As shown in Fig. 4, although the total number of Ce-4 felectron occupation is larger in the Sn-doped case, under the electron doping, the proba- bility off1conguration decreases while probabilities of increase. We thus conclude that the calculated nfmeasures both count for both low energy and high states. The overall dierences of nfare less than 0.1 electron per Ce atom. However, the enables the ne tuned electron number to bring up the drastic change of its electronic structure. We investigate the calculated hybridization function (Im
) for Ce1 and Ce2 as shown Fig. 5(a,b). Com- pared to Fig. 2(a-b), the DFT+DMFT calculation shows the hybridization strength increases with the number of FIG. 4. (color online) The calculated probabilities of valence states fornf=0 (blue), 1 (green), and 2 (red) of Ce1 in the pristine and doped cases. valence electrons. In energy between 0:1 eV and 0.1 eV, the dopant-dependent hybridization functions are well il- lustrated with respect to the red lines (pristine). The imaginary part of the hybridization ( Im
(!= 0)) at EFof Ce1 in Sn-doped CeCoIn 5is enhanced the other cases (see Table II). This enhancement of hybridization is reduced to be comparable to the pris- tine case for Ce2. This could be understood in terms of the relative distances between the Ce atoms and the Sn atom. The substitution of a hole (Cd, Hg) 0) around EF. Both Cd and Hg dopants give rise to similar decreases of the Im
(!= 0) on both Ce1 and Ce2 with respect to the pristine cases. The be- havior shown in Im
and ImG loccould reveal how the hybridization is changed in each case. The reduction of the hybridization favors a local-moment ground state over the Fermi liquid phase. We conclude that Sn-doped ( Cd- or Hg-doped) CeCoIn 5would be in proximity to a Fermi liquid state (a magnetic ground state). To elaborate more insight into the renormalization of the quasiparticle state, we will examine the imaginary (Im) and real (Re) parts of the calculated self-energy (). Figure 6 shows the imaginary part of the inverse at EFrepresents the inverse of the life-time of quasiparticle state. Im( != 0) of Ce1 in the Sn-doped case is almost zero at T=0.005 eV ( 65 K). The parabolic behavior of Im around != 0 of Ce1 in Sn-doped CeCoIn 5is a typical Fermi liquid The Fermi liquid phase represents the presence of the coherent quasiparticle states. The nite values of of Im (!= 0) indicate the deviation from the Fermi liquid phase in the other cases. This quasiparticle state of the Ce1 in CeCo(In 0:9Sn0:1)5would be consistent with the largely enhanced hybridization function. The renormalization of the band is strongly associated with the mass enhancement under correlation. The slope5 FIG. 5. (color online) The imaginary part of the hy- bridization functions for (a) Ce1 and (b) Ce2 in dier- ent structures with DFT+DMFT. The labels represent CeCoIn 5(Pristine), CeCo(In 0:9Cd0:1)5, CeCo(In 0:9Hg0:1)5, and CeCo(In III. Summary of the quasiparticle weight ( Z) at the Fermi energy at E 5 0.0124 0.0124 CeCo(In 0:9Cd0:1)5 0.0110 0.0110 CeCo(In 0:9Hg0:1)5 0.0121 0.0121 CeCo(In 0:9Sn0:1)5 0.1100 0.0139 of Re atEFis used to represent how large reduces the bandwidth of a quasiparticle state. Fig- ure 7 provides the variation of Re from 0:4 eV to 0.4 eV. The slopes of Re at EFseem to be similar to the red lines of the pristine case except for that of Ce1 in Sn-doped case. The quasiparticle weight Zis computed using 1=(1@Re @!) atEF, and is given in Table III. The Re(!= 0) of the Ce1 shows a very dierent behavior in comparison to the other cases. More strikingly, weight Z(0.11) of the Ce1 in the Sn-doped CeCoIn 5is nearly one order of magnitude larger than that for both the pristine and hole-doped cases. The larger quasiparticle weight represents the larger overlap of the correlate and non-interacting wavefunction at EF. The weight for the Ce2 is around 0.0139, which is slightly FIG. 6. (color online) The imaginary part of the self- energy for (a) Ce1 and (b) Ce2 in dierent struc- tures with DFT+DMFT. The labels represent CeCoIn 5 (Pristine), CeCo(In 0:9Cd0:1)5, CeCo(In 0:9Hg0:1)5, and CeCo(In IV. Calculation of coherence temperature, 5 0.00087 0.00087 CeCo(In 0.000634 CeCo(In 0.000766 CeCo(In 0:9Sn0:1)50.0140 0.001071 larger than the other cases (see Fig. 7(b)). The inhomo- geneity due to the Sn dopant is consistently observed in the DFT+DMFT calculations. On the other hand, weight for the hole-doped CeCoIn 5is and 0.012 (Hg-doped), respectively. This is also consistent with the observation that the bandwidths of the pristine and the hole-doped CeCoIn 5compounds are similarly renormalized (see Fig. 3(a)). The inhomo- geneity due to the hole dopant is relatively suppressed in the DFT+DMFT the calculated Zand Im
, the coherence tem- peratures (Tcoh) could be predicted using the calculated Tcoh is summarized in Table IV. The largest Tcohis 0.0140 eV for Ce1 in CeCo(In 0:9Sn0:1)5, which is one order of mag-6 FIG. 7. (color online) The real part of the self-energy for (a) Ce1 and (b) Ce2 in dierent structures with DFT+DMFT. The labels represent CeCoIn 5(Pristine), CeCo(In 0:9Hg0:1)5, and CeCo(In larger than the used temperature ( T=0.005 eV). However, The Tcohfor Ce2 in CeCo(In 0:9Sn0:1)5is re- duced to 0.001071 eV, which is still higher than all other coherence temperatures in the hole-doped compounds. These results suggest the substantial enhancement eect of the Sn dopant on the Kondo coherence, which pushes the system to be more close to the Fermi liquid phase, and a robust nonequivalence of the Kondo coherence be- tween Ce1 and Ce2 through the DFT+DMFT iterations. The hole-doped case show the decrease in both the Im
andZatEF. Therefore, Tcohin the hole-doped cases is lower than those of the pristine case. Unlike the electron- doped case, Ce1 and Ce2 simultaneously become more incoherent at T=0.005 eV. IV. CONCLUDING REMARKS dopants prefer to be substituted in the Ce plane,13our study here has considered the spatial in-homogeneity in the doped CeCoIn 5system. The much enhanced hybridization
and the non Ce-4 fdensity of statesNFof Ce1 near Sn can drive the system deep into the Fermi liquid state. We have also found through the two single-site DMFT impurity solvers that the 4 f- electron behavior is dramatically dierent on Ce1 and Ce2. Furthermore, the current study has shown a short- ranged electron doping eect. On the contrary, in the case of hole-doped CeCoIn 5, Ce-4fstate would be more incoherent than in the pristine. We further note that our study has been limited to the paramagnetic solu- tions and as such the presence of a G-type conguration ( q= (;; )), as reported in the Cd-doped CeCoIn 5,10is beyond the scope of the present work. However, the dopant induced here, indicating that the state moves away from the Fermi liquid phase, is consistent with the experiment. It is worthwhile to note that while the Sn doping gives a stronger inhomogeneity eect on the local the NMR experiment has observed24that the Cd doping causes a stronger inhomogeneity eect on the magnetic uctuations. In summary, we have investigated the and electron doping eect on CeCoIn 5within the DFT+DMFT framework. We have revealed a clear vari- ation of the hybridization function as well as the bands of Ce-4 fstates. In addition, the provided a good insight into in a doped CeCoIn 5system would be co- herent or incoherent. Specically, we have discovered a highly inhomogeneous in uence of Sn substitution on Ce-4felectron states. This prediction should be acces- sible to scanning tunneling microscopy experiments on a (100)- or (010)-oriented surface of a Sn-doped CeCoIn Choi thanks Soobeom Seo for This work was carried out under the auspices of the U.S. Department of Energy (DOE) National Nu- clear Security Administration (NNSA) under Contract No. It was supported by LANL LDRD Program (H.C.C.), the DOE BES in Narrow Band Systems project (E.D.B, F.R), and NNSA Advanced Simulation and (J.-X.Z.). Additional support has been pro- vided in part by the Center for Integrated a DOE BES user facility, in partnership with the LANL Institutional Computing Program for computa- tional resources. H.C.C. was also supported by Institute for Basic Science in Korea

2108.11008

Hong Chul Choi

Hong Chul Choi, Eric D. Bauer, Filip Ronning, and Jian-Xin Zhu

DFT+DMFT study of dopant effect in a heavy fermion compound CeCoIn5

7 pages, 7 figures

10.1103/PhysRevB.105.115121

LA-UR-21-28472

cond-mat.str-el

http://creativecommons.org/licenses/by/4.0/

We study the dopant-induced inhomogeneity effect on the electronic properties of heavy a combined approach of density functional theory (DFT) and dynamical mean-field theory (DMFT). The inhomogeneity of between Ce-4fand conduction electrons is introduced to impose the inequivalent Ce atoms with respect to the dopant. From the DFT to the we demonstrate a variation of the hybridization strength depending on the hole or electron doping. A drastic asymmetric mass renormalization could be reproduced in the DFT+DMFT calculation. Finally, the calculated reflects the different development of the heavy depending on the dopant.

2 Ueta and Otsuka together with the extinction as a single integrated process . However, what is traditionally exercised in the litera- ture has been to introduce a number of example, ad hocneandTemay be chosen to force a value of extinction to get the subsequent plasma diag- nostics going, or an ad hoc extinction value may even be adopted. In such cases, one should at least between the as the bases for the assumed extinction and the neandTe values as the actual outcomes of the subsequent In other words, the initially assumed neand Tevalues cannot be very dierent from the nal neand Tevalues to assure that these neandTevalues repre- sent the ionized gas in the target object. In practice, the subtlety of such consistency tends to be lost in transla- tion, because the extinction determination and are often dealt with as two separate issues. Hence, consistency between these two sets of neandTeis rarely scrutinized, let alone guaranteed. inconsistencies would usually invite spatially resolved 2-D plasma diagnos- tics have been becoming very relevant lately in many branches of astronomy and astrophysics, especially with the increasing availability of integral eld Walsh & Monreal-Ibero 2020). When measure- ments of extended target sources are made in a manner, both the extinction determination and plasma diagnostics ought to be performed at each detec- tor element. It is to ensure that the spatially extended nature of the target sources is fully appreciated at the end of the the spatial variation of relevant parameters in extended objects has rarely been considered carefully enough in both the extinction determination and Often a single-valued extinction is adopted by assuming uniform neandTeacross the whole ex- tent of an extended object (e.g. Lame & Pogge 1994; Guerrero et al. 1997). Such simplications may be per- missible as long as the aim is to sample of an extended object in an integrated such a deed defeats the purpose of spa- tially resolved observations, by forcing an extended ob- ject with articial uniformity, i.e., imposing an abso- lutely unnecessary source of the present work, therefore, we propose a proce- dure that streamlines both the extinction correction and plasma diagnostics as a single integrated process of it- erative data reduction. By seeking convergence at each spatial element through this iterative process, we can si- multaneously and determine the extinc- tion map as well as the electron ( ne andTe) maps, which allow us to carry out a plethora of analyses. Below, we demonstrate and validate the process in detail using a full suite of the archived Hubble Space Telescope ( HST )/WFC3 nar-rowband images of the Galactic planetary nebula (PN), NGC 6720. First, we brie y describe the adopted data and the line calibration applied, before we detail the iterative proce- dure ( SS2). We then derive the extinction map ( andTe([Nii]) maps ( SS3.2), and line emission maps ( SS3.4), as well as metal abundance distribution maps ( SS3.5), while giving de- tailed discussion of the outcomes and contrasting be- tween the present results and those obtained with typ- ical simplications ( SS3.7). In the end, we summarize and promote one of the most self-consistent and fully 2-D plasma diagnostics ever performed ( WFC3 Narrowband Image Set of NGC 6720 For the present study, we adopt images of the plan- etary nebula, NGC 6720, stored in the Hubble Legacy Archive1. These images were taken with the WFC3 cam- era on 2011 September 19 and 25 as part of the program 12309 (PI: C. R. O'Dell; O'Dell et al. 2013a,b,c). We adopt this data set because this program is one of the few that used the exquisite suite of the WFC3 narrow- band lters most extensively, providing an excellent op- portunity to perform self-consistent plasma diagnostics in full 2-D based on narrowband images and compare the results with those of the previous 1 lists all the lters used in this program, with their average wavelength and rectangular width (accord- ing to pysynphot ; Lim et al. 2015) and ocial de- scription (Dressel 2019). However, the F953N image is not used because of the known severe fringe pattern (Wong 2010; also SS5.4.4 of Dressel 2019). The in a lter name stands for a QUAD lter, which refers to one of the 22 mosaic of four lters providing four dierent bandpasses simultaneously with each band cov- ering a quarter of the nominal eld of view (FoV: Dres- sel 2019). Because of the smaller eld coverage of the QUAD lters, the subsequent analyses are all the NW quadrant of the main ring structure of NGC 6720 (see below; also see the Shared Field of View of All Filters in Fig. 2 of O'Dell et al. 2013a and Figs. 4d,e of Ueta et al. 2019). 2.2. The QP Method Many WFC3 lters can isolate the target emission line reasonably well. However, some lters suer from un- avoidable blending of neighboring lines at (Appendix A.2 of Dressel 2019). The most critical is the F656N and F658N lter pair. Even though the F656N and F658N lters are ocially described as the Hand [N ii] lters, respectively (Table 1), proles cover both the H and [N ii] Plasma Analysis Practice (PPAP) 3 Filter 
 Ocial Description (A) ( A) FQ436N 4367.21 43.35 H 4340 A+[O iii] 4363 A FQ437N 4371.09 29.99 [O iii] 4363 A F469N 4688.14 49.68 He ii4686 A F487N 4871.42 60.40 H 4861 A F502N 5009.70 65.29 [O iii] 5007 A F547M 5451.08 649.14 Str omgren y 5757.87 18.37 [N ii] 5755 A F645N 6453.71 84.22 continuum F656N 6561.45 17.65 H 6563 A F658N 6584.95 27.56 [N ii] 6583 A FQ672N 6716.62 19.37 [S ii] 6717 A F673N 6766.05 117.77 [S ii] 6717/31 A FQ674N 6730.77 17.63 [S ii] 6731 A FQ750N 7502.55 70.43 continuum F953N 9530.92 97.04 [S iii] 9532 A Table 1. HST /WFC3 lters used in the Program 12309. The quoted values are the bandpass average wavelength ( ) and bandpass rectangular width (
) dened in pysynphot (Lim et al. 2015). (Lim et al. 2015). As a result, both of the F656N and F658N images are blends of H and [N ii] lines at dier- ent proportions. To address this line-blending issue of narrowband lter images, we developed a new calibration method based on Quadratic Programming (QP), dubbed the QP line extraction method (Ueta et al. 2019). Using the HST images of NGC 6720 in the F656N and F658N lters in comparison with the ground-based slit-scan spectral imaging data cube around H , we demonstrated that the QP method properly extracted the individual H - only and [N ii]-only line maps from the line-blended raw F656N and F658N images (Ueta et al. 2019). The QP method recognizes narrowband images that are aected by line blending as linear combinations of emission maps of blended lines, each of which is modu- lated by the system throughput at the wavelength of the blended lines. Provided that the underlying continuum is properly subtracted, and assuming that all the lines involved are suciently narrow (i.e. a line can be speci- ed by a single wavelength), the solution for such a set of linear equations can then be sought as a QP problem in a least-squares sense under the presence of appropri- ate constraints. For the present work, we isolate the Hand [N ii] 6548/83 A line maps from the F656N and F658N images, the [S ii] 6717/31 A line maps from the FQ672N, F673N, and FQ674N images, and the [O iii] 4363_ A and H maps from the FQ436N and FQ437N images. As the QP processing itself is already describedfully elsewhere (Ueta et al. 2019), we outline how these images are processed in Appendix A. Here, we emphasize that the QP line is more advantageous than other methods that attempt to remove the fractional contribution of the un- wanted blended lines by just scaling the raw images (e.g. Appendix of O'Dell et al. 2013a). Such scaling methods are strictly dependent on the line emission distribution of the raw line-blended images. That is, the line emis- sion distribution of an isolated target line is always the same as that of the raw line-blended image that mimics the the distribution of the target line. In reality, however, the line emission distribution of the blended lines represents neither that of the target line nor that of the blended line (as it is a mix of the two): the resulting de-blended line emission maps do not have to appear the same as the raw line-blended im- ages. Instead, the QP method can recover the of individual lines as the optimized solution of a QP problem on a pixel-to-pixel basis. This is what makes the QP method unique in comparison with other line calibration methods for narrowband images. 2.3. Iterative Determination of apparent interdependence among the extinction and (ne,Te) may not have been taken into a sucient amount of attention that it deserves for some unknown reason(s). As outlined in SS1, to derive the extinction and ( ne,Te) by breaking the circular logic behind the theories of extinction plasma, both the extinction plasma diagnostics must be performed simultane- ously as one integrated process. In the following subsec- tions, we establish an iterative process through which this interdependence among the extinction and ( ne,Te) is carefully Extinction Correction We start by dening c(), the extinction at some wave- length,, as the base-10 power-law index to describe the reduction of the intrinsic ux, I0(), to the by I() =I0()10c(): (1) Then, we can determine c(H), the extinction at H , using, for example, the H -to-H ux ratio via c(H) has been replaced by is the total extinction at (the extinction on the magnitude basis) and A() = 2:5c() because I() (3)4 Ueta and Otsuka We can thus dene the observed ux at relative to the H ux (4) from which c() can be determined via c() (5) given that we know c(H) and the extinction average extinction at relative to V), which is provided by some extinction law (e.g. Draine 2003; Salim & Narayanan 2020). Then, we can recover the intrinsic line ux map at any from the observed line ux map at via Eq. (1). An obvious caveat here is the choice of the extinc- tion law, and the extinction, RV, which scales the selected extinction curve. Thus, the value of RValong the line of sight to the target object must be evaluated properly before we go any further. For the present study, we opt to adopt the Galactic extinc- tion law of Cardelli et al. (1989, CCM, hereafter), as the present analyses are concerned with data in the op- tical, in which there is little dierence among the ex- isting extinction laws. As for RV, we use the value of 3:130:03 to the direction of NGC 6720, determined for nearby stars within a 0.5radius (Gontcharov 2012; Gontcharov & Mosenkov 2017). We assume that RVis uniform over NGC 6720, as the adopted data by Gontcharov (2012) and Gontcharov & Mosenkov (2017) do not have sucient spatial even more subtle but no less is the choice of the theoretical H -to-Hline ux ratio, which is required to derive c() via Eqs. (2) and (5). Here, we need to remind ourselves that the theo- retical H-to-Hline ux ratio is really a function of ne andTeof the emitting gas of the target (e.g. Storey & Hummer 1995). Hence, the determination of the inter- stellar extinction is actually dependent on neandTeof the target object, which are the very quantities that we seek. Therefore, the present series of processes to deter- minec(H) and (ne,Te) must truly be performed as an integrated iterative procedure seeking the convergence of all of these values.However, such an iteration is hardly exercised in prac- tice: usually neandTeare assumed to be some ypi- cal values. This exercise may be permissible as long as target sources are unresolved and the adopted neand Tevalues are consistent with values derived as a result of the subsequent plasma diagnostics. However, when we expect these quantities to vary spatially in we ought to perform this iterative process rig- orously in each spatial element of the input for the present study, we rst adopt 104K as the initial values . then yield the initial ratio of 2.858 under the Case condition (Storey & Hummer 1995) with the CCM extinction law and RV= 3:13. Here, we stress that the theoretical H iline ratio needed in determining the ex- tinction does depend on neandTe. In the literature, this fundamental dependence is often neglected and the ratio is set to 2.85 or thereabout without any reference to the assumed neandTevalues, which may be dier- ent fromne= 103cm3andTe= 104K. In the present work, however, these values are iteratively updated for convergence (so is the theoretical H iline ratio) to as- certain the optimum consistency between plasma diag- nostics and extinction correction. In addition, we note that extinction toward a target is not necessarily just of interstellar origin. The target object itself may be surrounded by its own obscuring agents (i.e. dust grains) contributing to the on top of what is caused by the (ISM). For evolved stars such as NGC 6720, matter (CSM) most likely causes a signif- icant amount of the CSM extinction. Hence, if the ISM extinction is hastily adopted as the intrinsic a target source with a substantial the corresponding c(H) value toward the source is bound to be meanE(BV) reddening toward NGC 6720 is reported to be 0 :08700:0005 (Schlegel et al. 1998). Under the CCM extinction law with RV= 3:13, translates to c(H) as means that c(H) to be obtained in the present analysis for NGC 6720 has to be greater than 0.1266. Any excess extinction to be detected toward NGC 6720 in the present work, therefore, is the CSM contributed by the circumstellar dust grains, which is not accounted for by any of the ISM (e.g. Schlegel et al. 1998). This is true, of course, 2For example, using the Galactic Dust Reddening and Ex- tinction tool made available via Plasma Analysis Practice (PPAP) 5 only if the obscuring property of the circumstellar mat- ter follows that of the adopted ISM dust grains (CCM for the present case). 2.3.2. Plasma Diagnostics In performing plasma diagnostics, atoms are repre- sented asn-level energy states. Then, neandTeare de- termined by solving a set of equilibrium equations for In these equilibrium equations, the col- lisional excitation coecient has e) dependence (where
Eis the energy dierence between any two levels) and the collisional de-excitation coef- cient has while the radiation de- excitation coecient has dependence on neither nenor Teto the rst order (Osterbrock & Ferland 2006; Prad- han & Nahar 2015). Thus, if we take the ratio between lines whose transi- tion energies are close to each other (i.e.
E0; e.g. the [S ii] 6731-to-6717 A ratio), such ratios can depend mostly onneand only weakly on Te. On the contrary, if we take the ratio between lines whose transition energies are very dierent (e.g. the [N ii] 5755-to-6583 A ratio), such ratios can depend mainly onTeand only weakly on ne. Then, we can nd neandTeas the point in the ne- Tespace at which these two diagnostic line ratio curves intersect (see, also, SS3.7.1). As this point can not be computed analytically, neandTehave to be by iteration. For this work, we use the [S ii] 6731-to-6717 A line ra- tio map and the [N ii] 5755-to-6583 A line ratio map as the input diagnostic line ratio maps. In addition, the present set of data oers the [O iii] 4363-to-5007 A line ratio as another diagnostic for higher excitation regions than those probed via the [N ii] 5755-to-6583 A line ra- tio. However, the HST /WFC3 narrowband lter set does not oer any diagnostic ratio appropriate to prove higher excitation regions with the [O iii] 4363-to-5007 A line ratio. Hence, we will not use the [O iii] 4363-to- 5007 A line ratio (but will come back to this point This forced choice of diagnostic line ratios, however, is not necessarily bad, because the spatial ex- tent of [S ii] and [N ii] is similar to that of H and H, covering the main ring structure of NGC 6720. For the subsequent discussion, we always refer to andTe([Nii]) with the diagnostic line used to explicitly indicate relevant energy regimes. 2.3.3. Iterative Procedure It is not too dicult to imagine that deriving c(H) and (ne,Te) that simultaneously satisfy extinction cor- rection and plasma diagnostics at each spatial element were too cumbersome in the past, especially when com- putational resources were not readily available. be assumed for the and plasma diagnostics were performed sub- sequently only once to yield constant neandTeeven for Figure 1. Schematic representation of PPAP, and are determined by objects, never to be retroactively checked for consistency. This practice appears to have been adopted for years in the literature. In many of the previous inves- tigations of NGC 6720, a preset Tewas adopted to derive ne, and then, a preset ne, instead of the derived ne, was used to derive Te, and the computation was left at that (e.g. Lame & Pogge 1994; Garnett & Dinerstein 2001; O'Dell et al. 2013b), while the subtle intertwined nature of these analyses seems to have been lost in in the present time when decent computa- tional resources are regularly available, there is no rea- son not to perform iterative searches for c(H) and (ne, Te) at each spatial element through both the extinc- tion determination and plasma diagnostics in a manner. Therefore, we propose proper plasma analysis practice (PPAP), a fully procedure of the extinction cor- rection and plasma diagnostics as a streamlined (schematically represented in Fig. 1): (0) Select the extinction law to use and set the RV value toward the target source as well as the Compute the theoretical I0(H)=I0(H) ratio based on the initial neandTevalues following, e.g., Storey & Hummer (1995); (2) Compute c(H) by comparing the and theoretical using the PyNeb setCorr (Eq. (2)), and then, c() using the PyNeb getCorr function (Eq. (5)) with the adopted extinction law;6 Ueta and Otsuka (3) Correct the observed line ux maps for extinction using thec() map obtained in Step (2); (4) Determine ne([Sii]) andTe([Nii]) using [S ii] 6717-to-6731 A and [N ii] 5755-to-6583 A line ratio maps in the function; (5) Compare the old and new ne([Sii]) and terminate the process if the conver- gence is achieved (i.e. the dierence is negligible); (6) If the convergence condition is not met, repeat Steps (1) through (4) with the new ne([Sii]) and Te([Nii]) values. When the convergence is achieved at Step (6), andTe([Nii]) distribution maps are left to us for further analyses. The converged c(H) allows us to produce optimally line emission maps. With the help of the converged maps, line emission maps in turn yield other products of plasma diagnos- tics such as ionic and elemental abundance In implementing PPAP for the present study, we employ PyNeb (Luridiana et al. 2015), a python imple- mentation of a set of tools for analyzing emission lines based on the method used by the IRAF nebular pack- age (Shaw & Dufour 1995; Shaw et al. The Converged c(H)Map 3.1.1. Observed Spatial Variations First and foremost, we examine the converged c(H) map as the basis for all the subsequent analyses. In Fig. 2, we present the c(H) map (top) and the corre- sponding one- percent uncertainty map (bottom) of the NW quadrant of NGC 6720 converged via PPAP based on the observed H -to-H, [Sii] 6717-to-6731 A, and [N ii] 5755-to-6583 A ratio maps. The one- un- certainty is calculated at each pixel by varying each of the three input line ratios by its standard ratio ) and propagating the resulting In Fig. 3, we also present the c(H) radial density distribution. We adopt this format instead of a typical radial prole to better reveal the radial variation of c(H) itself as well as gures indicate that the converged c(H) values largely populate the characteristic main ring structure of NGC 6720 seen between the cen- ter slightly elongated along the position angle (PA; E from N) of120, but not so much in the central and not at all in the region beyond the main ring (beyond 4000). Because the eectiveness of PPAP is dictated by the quality of input line emission maps, the whole process can be bottlenecked at Step Figure 2. Thec(H) distribution map (top) and the cor- responding one- percentage uncertainty map (bottom) of the NW quadrant of NGC 6720, derived through PPAP. Maps are shown in the original pixel scale of the input HST images at 0 :000396 pix1with the relative RA and Dec osets from the position of the central +33:01:44.8831), indicated by the tickmarks. The color wedges show the range of the values presented. (4) by less sensitive line maps (Fig. 1; SS2.3). For the present case, we do not obtain converged results in all pixels within the inner cavity because S/N of [S ii] line emission is unfortunately marginal (especially Fig. we see that c(H) is more or less constant at slightly greater than 0.2 over 20{2500and then in- creases radially to 0.25{0.35 at the outer edge of the main ring (304000), with the mean of 0 :290:09 (cor- responding to49 % attenuation at 4861 A). At many positions near the outer edge of the main ring, especially along the short axis (PA of 30),c(H) is found to beProper Plasma Analysis Practice (PPAP) 7 Figure 3. The radial density distribution map of c(H), with the frequency indicated by the color, showing the c(H) radial prole without suppressing the azimuthal variation. The ISM contribution to c(H) (0:12660:0007) is marked by the dashed line. as great as 0.4 or even greater (more than 60 % atten- uation). On the contrary, the maximum c(H) at the elongated tip of the main ring along the long axis (PA of120) is only marginally high ( 0.25{0.35, corre- sponding to 45{50 % azimuthal trending of c(H) is generally consis- tent with the dust distribution in NGC 6720 revealed by far-IR dust continuum emission maps taken with the Herschel Space Observatory (van Hoof et al. 2010). While the spatial resolution diers quite a bit in the optical and in the far-IR, the dust distribution shows a greater degree of dust concentration around the short axis than around the long axis. As suggested, for exam- ple, by Guerrero et al. (1997) and O'Dell et al. (2013a), the presence of a denser molecular gas of the region (PDR) beyond the main ring may be directing out ows emanating from the central into the low density polar directions that is slightly inclined with respect to the line of sight, eec- tively generating the elongated appearance of the main ring (cf. Fig. 14 by Guerrero et al. 1997; Fig. 11 by O'Dell et al. 2013a). Past kinematic studies indeed revealed an expanding ellipsoidal shell whose long-axis is almost aligned with the line of sight (e.g. Guerrero et al. 1997; O'Dell et al. 2007; Martin et al. 2016). In the central cavity (within 2000), on the other hand, we nd that c(H) is roughly constant at 0 :220:06. The absence of c(H) near the central star (within 500) does not mean the absence of attenuating dust is simply not reliably computed around the central star because the input line maps are aected by the presence of the central star (e.g. imperfect subtrac-tion of the continuum). Again, far-IR dust continuum maps show more or less uniform dust emission in the inner cavity (van Hoof et al. 2010), while it is expected that dust grains exist along the line of sight on the near and far sides of the rareed high-temperature plasma region around the central star. The observed general radial behavior of c(H) is con- sistent with the expected stratication of the nebula. That is,c(H) values tend to be higher in the main ring where the bulk of the nebula material is located than in the inner cavity where the plasma is rareed. Then, the radially increasing c(H) is by the radial decrease of the gas the degree of ionization. The present c(H) map at the exquisite 0 :000396 pix1 scale (Fig. 2) successfully reveals details of its spatial dis- tribution. A comparison between the c(H) map and the observed H -to-Hmap (Fig. 4, top) clearly indicates that the presence of dust grains attenuates the bluer Hline emission more than the redder H line emis- sion, leaving a higher H -to-Hratio at that location. Such highc(H) (high H-to-Hratio) regions appear to form micro-structures that resemble radial and/or Rayleigh-Taylor (RT) instability n- gers as observed in the Helix Nebula (O'Dell & Handron 1996) and the Crab Nebula (Hester 2008), is a greater number of such clumps closer to the outer edge of the main ring. This is probably caused by the inhomogeneous gas distribution in the main ring be- ing eroded by photoevaporating radiation from the cen- tral star. Ionizing radiation does not just simply travel radially in the nebula, but permeates through the inho- mogeneous gas by going preferentially into low density regions, leaving high-density clumps behind. As discussed at the end of SS2.3.1, we expect c(H) of 0:12660:007 from the ISM alone. How the converged c(H) compares with this expectation is in Fig. 3: the ISM component is represented by the horizontal dashed line. We immediately see that the derivedc(H) is greater than the ISM value for the en- tire extent of the nebula where we have result indeed proves that NGC 6720 itself provides the source of self-attenuation in its circumstellar nebula. The circumstellar c(H) is about 0.1 in the inner cavity (20 % reduction), as high as 0.3 at the outer edge in the short axis direction ( 50 % reduction), and 0.12{0.17 at the outer edge in the long axis direction ( 30 % reduc- tion). Hence, the circumstellar contribution to c(H) in NGC 6720 amounts to about the same or even greater than the ISM contribution. Given that NGC 6720 is an object at a moderate Galactic latitude reasonably away from the Bulge (( l;b) = it may not be very appropriate, in general, to adopt the ISM c(H) value indiscriminately as the total c(H) for any target source. The circumstellar c(H) component of0.1 measured in the inner cavity is most likely attributed to dust grains8 Ueta and Otsuka Figure 4. The observed (top) and (bottom) H -to-Hline ratio maps of the NW quadrant of NGC 6720. Image conventions follow those of Fig. 2. The color wedge shows values in the 3range around the in front of the central high-temperature region around the central star along the line of sight. Thus, we expect that there is another c(H)0.1 worth of dust grains behind the central high-temperature region. If so, c(H) in the inner cavity as a whole would be about 0.33 including the ISM component (= 0 :13 + 0:1 + 0:1), which is very much consistent with the amount of c(H) we observe at the outer edge of the main ring. Hence, it appears that the main ring of NGC 6720 contain worth of dust grains in any Comparison with the Previous ResultsOne of the rst spectral imaging investigation of NGC 6720 was performed by Lame & Pogge (1994) with a Fabry-Perot imaging spectrograph. They presented the observed H -to-Hmap (their Fig. 3) and noted a smooth radial increase of the value from 3.68 in the cen- tral cavity to 3.74 in the main ring, with the maximum of3.93 happening at around the 3/4 of the ring (with the median of 3 :740:22). These values compare reason- ably well (within 10 %) with what we see in our QP- processed (observed, but line contamination (Fig. 4, top), in which the ratio rises from 3:360:11 in the central cavity to 3 :610:12 in the main ring with the median of 3 :380:40. Lame & Pogge (1994) went on to correct for extinction by assuming the constant theoretical H -to-Hratio of 2.86 in the main ring, for which neof 102cm3andTe of 104K are assumed via Hummer & Storey 1987 un- der the same CCM extinction law. This prompted the corrected H -to-Hratio of 2.81 in the central cavity, which corresponds to Teof 13,500 K. They to the local inhomogeneity and devi- ation from the Case B condition, and moved on the theoretical H -to-Hratio based on the updated Te, never seeking consistency between the extinction correction and plasma diagnostics (very typi- cal in the literature). So, their c(H) values would have been 0.31{0.32, which are 20 % greater than ours. In our case, on the other hand, c(H),ne, updated for convergence, and out H -to-Hmap (Fig. 4, bottom) yields the mean ratio of 2 :8630:024 in the main ring out of the initial value of 2.858 (and the corresponding andTe([Nii]) maps from the initially as- sumed 103cm3and 104K; see SS3.2), maintaining in every single pixel (Fig. 8a). This means that the 20 % discrepancy in the analyses by Lame & Pogge (1994) arose because (1) a uniform was imposed across the entire nebula and (2) convergence was not sought among iteration, i.e., c(H), via the theoretical H - to-Hratio, was never updated even when same HST /WFC3 data set was previously an- alyzed by O'Dell et al. (2013b). They originally re- ported radially decreasing c(H) along both the long- and short-axis of the nebula. We communicated with them about this discrepancy during the early phase of writing of this manuscript. It turned out that one of the coecients in their ux calibration formula for F656N was in error, and that their updated results now show radially increasing c(H) around 0.3 in the main ring (O'Dell et al. 2021). We note that their derivation is based on the constant H -to-Hratio of 2.87 across the nebula assuming a 104K gas at low density (the value unspecied) under the extinction law by Whitford (1958) with RV= 3:1. Also, their WFC3 image ux calibration does not follow the standard STScI method,Proper Plasma Analysis Practice (PPAP) 9 but is based on their own method established with their own ground-based long-slit spectroscopy data (i.e. not generally previous analyses were typically based on or long-slit spectroscopy (e.g. Hawley & Miller 1977; Barker 1980; Guerrero et al. 1997; Gar- nett & Dinerstein 2001). Each of these studies adopted a certain extinction law (plus RV) and theoretical H -to- Hratio based on their own assumptions of neandTe. The adopted parameters are sometimes explicitly men- tioned, and sometimes not. The reported c(H) (plusne andTe) values are generally consistent with our values, some more so and others less. The direct comparisons to assess details beyond general agreement do not seem in- structive, as none of the previous derivations were done . There is more recent spectral imaging study of NGC 6720 (Martin et al. 2016) based on obser- vations with an imaging Fourier transform et al. 2019). However, their study is focused on the velocity structure of the nebula, and no plasma di- agnostics are discussed. In light of the previous studies smothered with the lesson to be learned here seems rather ob- vious. We ought to eliminate inconsistencies where we can, because any inconsistencies introduced will exacer- bate uncertainties in the outcomes. All things consid- ered, we can reiterate one major advantage of PPAP in comparison with other methods in the literature. PPAP yields the self-consistent c(H) (and hence, ne, based on only the input line ux maps plus the choice of the extinction law to adopt, with no need to adopt any other parameters and ad hoc The Converged ne([Sii]) andTe([Nii]) Maps 3.2.1. Observed Spatial andTe([Nii]) distribution maps con- verged via PPAP are summarized in Fig. 5 with percentage uncertainty maps. The ra- dial density distribution maps of ne([Sii]) andTe([Nii]) are also presented in Fig. 6. In computing ne([Sii]) and Te([Nii]) with PyNeb (Luridiana et al. 2015), we use the transition probabilities (Einstein As) by Rynkun et al. (2019) and Fischer & Tachiev (2004), and colli- sion strengths by Tayal & Zatsarinny (2010) and Tayal (2011), respectively. The ne([Sii]) andTe([Nii]) uncer- tainties are computed in the same way as the As is the case for c(H),ne([Sii]) and Te([Nii]) are reliably determined mostly in the main ring. This is partly because of S/N in the diagnostic [Sii] line ratio map (Fig. 17), but also partly because the [Sii] and [N ii] diagnostics are better suited to probe the moderately ionized region (i.e. the main ring) than the highly ionized wind bubble region (i.e. the inner cavity). The converged ne([Sii]) andTe([Nii]) maps show a lesser degree of spatial variation in the main ring than the converged c(H) map: we do not recognize micro-structures in the ne([Sii]) andTe([Nii]) maps as in the c(H) map. Closer examination suggests that there is some anti-correlation between ne([Sii]) tends to be smaller where c(H) is larger. This is indeed expected because dust grains are more likely found where temperature is low, i.e., the degree of ion- ization is low, and vice versa. The main ring appears to be divided into the follow- ing two parts in terms of ne([Sii]) andTe([Nii]): (1) the inner part (20{2500), in which both ne([Sii]) and Te([Nii]) show a radially increasing trend with uncertainties ( 25 %), and (2) the outer part (be- yond 2500up to 30{4000), wherene([Sii]) decreases and Te([Nii]) is more or less andTe([Nii]) values increase cm3to1500 cm3inne([Sii]) (with the mean of 1 ;400580 cm3) and from7,000 K to 9,300 K inTe([Nii]) (with the mean of 8 ;3901;270 K) in the inner part of the main ring. Their spatial distribu- tion ofne([Sii]) andTe([Nii]) indicates that this inner part is rather circular in projection (extends up to 2500 in all azimuthal directions). Thus, this region is pos- sibly spherical in 3-D, suggesting that this represents the highly-ionized central wind bubble caused by the radiation from the central star, eating its way into the higher density ellipsoidal wind shell (the outer part beyond 2500). Thus, the observed radial rise of ne([Sii]) in the inner part of the main ring can be understood as caused by the snowplowing eect by the PN fast wind emanating from the central star (e.g. Kwok 2000). However, the ob- served radial rise of Te([Nii]) may be radially increasing Te([Nii]) trending is and attributed to the self-shielding eect of the ionizing radiation eld: soft ionizing photons are already absorbed by metals in the inner part and re- maining hard ionizing photons in the outer part would yield more heat per ionization, making Te([Nii]) higher at greater radial distance away from the central star (Kewley et al. 2019). In the outer part of the ring, ne([Sii]) radially de- creased to about 1,000 cm3(with the mean of 1 ;170 540 cm3), whileTe([Nii]) remains about the same at 9,300 K (with the mean of 9 ;2901;060 K). Given that the degree of ionization is more or less the same in this part of the shell as implied from rather the radial fall-o of ne([Sii]) probably re ects the density distribution in this ne([Sii]) andTe([Nii]) change from 1,500 cm3and10,500 K along the short axis to 500 K along the long This azimuthal trending is consistent with what we see in the c(H) distribution (Fig. 2). Gen- erally speaking, this azimuthal variation is what gives the observed spread of the ne([Sii]) andTe([Nii]) val- ues { about 1,000 cm3forne([Sii]) and about 2,000 K forTe([Nii]) { in the density distribution (Fig. 6).10 Ueta and Otsuka Figure 5. Thene([Sii]) andTe([Nii]) distribution maps with the corresponding one- percent uncertainty map derived through PPAP (left and right, respectively). Image conventions follow those of Fig. 2. While the radial density distribution maps do not show any representative ne([Sii]) andTe([Nii]) values in the inner cavity (Fig. 6) as in c(H) (Fig. 2), the mean ne([Sii]) andTe([Nii]) values nevertheless turn out to be 1;340750 cm3and 8;1401;730 K. Here, readers are reminded that these values are based on the [S ii] and [N ii] diagnostics, which are suited to probe a low- excitation gas. In other words, these measurements are obtained from the low-excitation plasma that lies along the line of sight to the inner cavity, i.e., [S ii] and [N ii] are most likely not co-spatial with the central wind bub- ble region. Rather, the valid ne([Sii]) andTe([Nii]) val- ues in this region most likely sample the near-side of the thinned-out bipolar cavities of the ellipsoidal shell of the nebula (cf. Fig. 14 by Guerrero et al. 1997; Fig. 11 by O'Dell et al. 2013a). 3.2.2. Comparison with the Previous ResultsThe previous neestimates yielded 500{700 cm3 (Lame & Pogge 1994), 470{615 cm3(Guerrero et al. 1997), 660{740 cm3(Garnett & Dinerstein 2001), and 350{550 cm3(O'Dell et al. 2013b), based on the as- sumedTevalue of 104K and assumed theoretical H - to-Hratio (for which certain newas assumed). They used the derived nevalues to update Tethat ranged roughly from 10,000 to 13,000 K, but without updating the theoretical H -to-Hratio (hence, extinction cor- rection) according to the updated neandTe. More specically, for example, Lame & Pogge (1994) adopted the theoretical H -to-Hratio of 2.86 uni- formly across the nebula, assuming a uniform 104K. A comparison between the assumed theoretical H -to-Hratio of 2.86 and the ob- served H-to-Hratio of3.74 yielded a certain Then, extinction corrected [S ii] and [N ii] diagnostic line ratios yielded the ne([Sii]) Plasma Analysis Practice (PPAP) 11 Figure 6. The radial density distribution map of with the frequency indicated by the color (top and bottom, respectively). Image conventions follow those of Fig. 3. of 500{700 cm3andTe([Nii]) distribution of 10,000{ 13,500 K. Lame & Pogge (1994) concluded their analyses then, without re-evaluating c(H) via the updated the- oretical H-to-Hratio based on the derived distributions, which dier from the as- sumed uniform neandTe, and following through the rest of the analyses for both of the ne([Sii]) andTe([Nii]) distribu- tions were dierent from the initial uniform assump- tion, the theoretical H -to-Hratio, and hence, the c(H) distribution should have had been recomputed for consistency. Subsequently, the resulting ne([Sii]) and Te([Nii]) distributions should have had been updated again to re ect changes in the diagnostic line ratios be-cause of the c(H) osets. Ideally, this iteration should have had been repeated until all values converged to the nal values for the maximum consistency. Unfor- tunately, consistency among these parameters was not sought iteratively by Lame & Pogge (1994). This is very representative of how analyses were per- formed in the literature. Hence, there does not seem much point in directly comparing the in the literature with our results. The among c(H) and (ne,Te) were not consid- ered carefully enough for consistency in the past. The strangest thing in the literature is that the nal c(H) and (ne,Te) values and their deriva- tives were usually attributed to local though a uniform distribution of the initial param- eters were usually assumed to begin with. It is really this initial assumption of uniformity that imposes in the rst place. There is really no need to assumec(H) and (ne,Te): all can be determined si- multaneously by seeking convergence iteratively. In the present analyses, only a few iterations (at most 5) Tolerable Uncertainties To make PPAP work eectively, it is important to keep self-consistency in the whole of the analyses. How- ever, at some point, the propagated uncertainties from the line uxes themselves would become greater than the uncertainties caused by assuming constant ( ne, Te). Then, what uncertainties are tolerable in follow- ing PPAP? _While this question may sound simple, it is actually not simple at all. This is because the tolerable levels of uncertainties are actually dependent on the ac- tual line ratios used, the ( ne,Te) values themselves, and how varying neandTeare in the actual spatial distri- bution in the target source, among other things. For the case of NGC 6720, the neandTevalues are spread around 1,340 cm3and 8,140 K with a width of 750 cm3and 1,730 K, respectively ( SS3.2.1). Thus, if uncertainties in neandTeexceed these spreads, assum- ing constant ( ne,Te) would be as good as PPAP. A set of experiments with PyNeb reveal that uncertainties of 10{15 % in the [S ii] ratio would cause large enough un- certainties in the resulting neso that considering not be so meaningful. The same is said forTeif uncertainties in the [N ii] ratio are 25{30 %. Hence, uncertainties of roughly 10 % and 25{30 % can be tolerated in the [S ii] and [N ii] line ux While it is dicult to generalize this result for a particular source, uncertainties of 10 % or better appear to be necessary for PPAP to work as a rule of thumb. 3.3. Gas-to-Dust Ratio Map With thec(H) and maps deter- mined, it is possible to evaluate the gas-to-dust mass12 Ueta and Otsuka Figure 7. The gas-to-dust mass ratio map at H obtained from the H , , and c(H) maps by assuming certain dust properties. Image con- ventions follow those of Fig. 2. ratio map at H . This is because dust grains are repre- sented byc(H), while H-emitting ionized gas is repre- sented by the H surface brightness map, ne([Sii]) map, and the Hemissivity map derived from the ( maps. The denition of the extinction allows us to relate c(H) and the gas-to-dust mass ratio as fol- lows. c(H) =AH 2:5= (loge)H = andm[dust;N+]are the col- umn density, mass density, and particle mass of dust grains and H -emitting gas, respectively, and Qext His the extinction eciency coecient under the assumption of grains being spheres of radius a. Hence, the gas-to-dust mass ratio is; H+ dust= the bulk density of dust grain the H surface brightness is expressed in terms of the column density of H -emitting gas as (8) where the integral is done through the target along the line of sight, His the Hemissivity as dened by Storey & Hummer (1995) as a function of ne([Sii]) and Te([Nii]). By combining the above two equations, we can derive the gas-to-dust mass ratio map at Husing the H map (which gives IH), the maps (which give H andne), and the c(H) maps, by assuming spherical dust grains of smoothed astronomical silicate (Wein- gartner & Draine 2001) of radius 0.1 m and bulk den- sity 3 g cm3(as the dust chemistry of NGC 6720 is derived gas-to-dust mass ratio map at H is pre- sented in Fig. 7. The distribution appears similar to the ne([Sii]) map (Fig. 5), with the median of 437 357, which is a few times greater than the ypical value of 100. The derived values need to be considered as the lower limit because we take into account only the H-emitting ionized gas in the main ring, in which some non-negligible amount of atomic H gas component is expected. According to a photoionization model of NGC 6720 by van Hoof et al. (2010), the amount of ion- ized and neutral (both atomic and molecular) gasses is about the same (they quoted log( ne) = 2:62 and log(nH= 2:60)). Thus, we account for only about 50 % of the gas. Also, the derived c(H) includes the which amounts to at most roughly 50 % (Fig. 2). Therefore, the total gas-to-dust mass ratio seems to be about 1600 in the main ring. Clearly, a ypical ratio of 100 cannot be use indiscriminately for the cir- cumstellar dust component, especially when the amount of dust component may be reduced because of local en- vironments. In the literature, it is often practiced by adopting an old empirical ISM relation, (Savage & Mathis 1979). However, this relation implicitly assumes the gas-to-dust mass ratio of 100 for the verage ISM. Thus, as we demonstrate here, it is simply wrong to adopt such an indiscriminately to any object for which some circumsource dust component is expected. 3.4. Line Emission Maps Fig. 8 shows the line maps rela- tive to Hextracted from the present HST /WFC3 dataProper Plasma Analysis Practice (PPAP) 13 Figure 8. Various line maps (rela- tive to H) of the NW quadrant of NGC 6720: (a) [O iii] 4363 A/H, (b) He ii4686 A/H, (c) [O iii] 5007 A/H, (d) [Nii] 5755 A/H, (e) [N ii] 6548 A/H, (f) [N ii] 6583 A/H, (g) [S ii] 6717 A/H, and (h) [S ii] 6731 A/H. The wedge on the right in each panel indicates the adopted linear color scale. Image conventions follow from those of Fig. 2. set. These individual line maps reveal the spatial varia- tion of the relative abundance of specic within extended nebulae. Quick inspection of these maps suggests that there are three general pat- terns of the emission morphology owing to the levels represented by these lines. First, a strong circular emission region is seen only in the inner cavity ( <2000) as in the He ii4686 A map (Fig. 8b), as previously reported by Lame & Pogge (1994). This region can be recognized only by the fastradial decrease of emission around 2000. Most likely this region represents the spherical PN wind bubble at the core of the nebula, in which the ambient tempera- ture is the highest. Second, there is another circular emission region that appears more extended than the He ii4686 A the inner part of the main ring (up to 25{3000), as seen in the [O iii] 4363 and 5007 A maps (Fig. 8a,c). This region is where the H -to-Hratio is high (Fig. 4, bottom), and where c(H),ne, andTeradi- ally increase to the maximum (Figs. 3 and 6), delineating the extent of the ionized region. Third, emission is strong only in the periphery of the main ring (beyond 3000) as in the [N ii] and [S ii] maps (Fig. 8d{h). This region corresponds to where c(H) is high (0.3 and above; Figs. 2, 3). ne([Sii]) andTe([Nii]) are also high in this region (1,200 cm3, Figs. 5;>104K, 6), most likely representing the PDR surrounding the ionized region. The apparent transition of morphology from the in- ner part (radially decreasing) to outer part of the main ring is quite remarkable. The radially increasing trending of Te([Nii]) toward the pe- riphery of the main ring (Fig. 5, right) is promoted by the radially increasing trending of [S ii] and [N ii] line emission (Figs. 8d{h). As discussed earlier, this is inter- preted as the self-shielding eect, in which the majority of the ionizing photons is already consumed in the in- ner cavity and in the inner part of the main ring and the left-over far-reaching high-energy photons heat the outer part of the main ring. Hence, this transition zone at3000from the central star very likely corresponds to the location of the ionization front (IF) that separates the high-excitation emission region (Figs. 8a{c) and region (Figs. between two line maps also provide a wealth of information. H emission is representative of the ionized region, while [N ii] emission is of the PDR. Hence, the [N ii] 6583 A-to-Hline ratio map helps to spatially separate these regions (Fig. 9, top), i.e., to iden- tify where IF is located. This is one of the reasons why blending of H and [N ii] lines in the H band can be a critical issue. IF is most likely located where the [Nii] 6583 A-to-Hgradient becomes the steepest. The radial density distribution map (Fig. 9, that the steepest gradient happens at various radii depending on the azimuthal angle. Around the short axis, the ratio remains low in the inner part of the main ring, and precipitously increases at 2500. (Lame & Pogge 1994) and (O'Dell et al. 2013a) noted that the apparent physical thinness of the [N ii]-and-[S at the periphery of the main ring (roughly a few arcsec width) indicated that at the line of sight in this region IF intersects nearly at parallel with our line of sight. This means that the inclination of the hollow el- lipsoidal shell of the main ring of NGC 6720 (cf. Fig. 14 by Guerrero et al. 1997; Fig. 11 by O'Dell et al. 2013a)14 Ueta and Otsuka Figure 9. [Top] The [N ii] 6583-to-Hline ratio map distin- guishes the ionized region ( <1) from the PDR ( >1). follow those of Fig. 2. [Bottom] The radial den- sity distribution map of the [N ii] 6583-to-Hline ratio shows the IF as the region of the steepest gradient between 2000and 4000. The dashed line is where the ratio is unity. is pivoted around the short axis in the plane of the sky (Guerrero et al. 1997; O'Dell et al. 2007; Martin et al. 2016). This further means that along the long axis we are looking into the inclined inner wall of the shell (i.e. IF), where the [N ii] 6583 A-to-H ratio starts to rise at small radii but increases rather slowly, meandering around 2. This meandering suggests that the IF surface is not necessarily of uniform density and may be sprinkled with micro-structures (e.g. those associated RT/KH instabilities; SS3.1). Hence, the of these may yield useful Figure 10. The [O iii] 4363-to-5007 A (top) and [N ii] 5755- to-6583 A (bottom) line ratio maps both compare the of with that of but reveal opposite spatial the range of ratio is the same as indicated by the color wedge) because of the Te([Nii]) inversion by self- shielding of ionizing photons. Image conventions follow those of Fig. 2. constraints for numerical simulations to determine hy- drodynamical IF interesting comparison is between the [O A and [N ii] 5755-to-6583 A line ratio maps (Fig. 10). These line ratios compare with of the O2+and N+ ions. The former transition probes higher than the latter transition. However, interest- ingly, these line ratios exhibit opposite trending. The [Oiii] 4363-to-5007 A line ratio map (Fig. 10, top) shows larger ratios in the inner region of the main ring thanProper Plasma Analysis Practice (PPAP) 15 in the outer region. This is expected from the general temperature structure of the nebula. On the other hand, the [N ii] 5755-to-6583 A ratio map (Fig. 10, bottom) shows relatively larger values in the outer region of the main ring than in the inner region. This is exactly the same Te([Nii]) inversion in the main ring seen above in SS3.2 (Fig. 5, right) as a conse- quence of self-shielding of low-energy ionizing radiation by metals in the inner part of the main ring (Kewley et al. 2019). This comparison plainly demonstrates the importance of adopting diagnostic lines that arise from the very re- gion of interest. The [S ii] and [N ii] lines probe the same energy regimes, i.e., the resulting ne([Sii]) and Te([Nii]) would be likely co-spatial. However, the [O iii] lines probe higher temperature regimes than the [N ii] lines do (i.e., the [O iii] lines are more suited to probe the inner cavity). Hence, plasma diagnostics done with the [S ii] and [O iii] lines would not make much sense as thene([Sii]) andTe([Oiii]) are not co-spatial, and Te([Oiii]) most likely exhibits opposite trending with respect N+and S+Relative Abundance Maps The nal step of plasma diagnostics is the determi- nation of ionic and elemental abundance for a given set of neandTe, ionic popula- tions for a particular ionic species should be from any line of system. For the present study, it is nat- ural to compute the ionic abundance distribution maps ofn(N+) andn(S+) relative to n(H+) using any of the [Nii] and [S ii] line maps, as our derivation of neandTe is based on the [N ii] and [S ii] lines ( SS3.2; Figs. 5, 6). Here, we use the PyNeb getIonAbundance function with the line maps at [N ii] 5755 and 6583 A and [S ii] 6717 and 6731 A (Figs. 8d,f{h). If the resulting relative N+and S+abundance maps come out to be identical for each of the [N ii] and [S ii] pairs, we can safely say that the robustness of the pro- posed PPAP is guaranteed. The results of this experi- ment are summarized in Fig. 11. The left column maps (Fig. 11a,b) derived from the [Nii] 6583 and 5755 A line maps (Figs. 8d,f), and their dierence in percentage (Fig. 11c). Similarly, the right column presents the same for S+from the [N ii] 6717 and 6731 A lines (Figs. 11g,h). The derived n(N+)=n(H+) andn(S+)=n(H+) vary from about 5105and 106at the inner edge to about 3:2104and 6:3106at the outer edge of the main ring, respectively. These values are consistent with previously determined values at various positions in the nebula presented by Barker (1987) ( n(N+)=n(H+) of 9:5106to 2:21104from multi-position aper- ture spectroscopy) and Liu et al. (2004) ( n(N+)=n(H+) of of 1:23106based on line intensities for the entire nebula obtained by scan- Figure 11. Relative ionic abundance distribution derived from the (a) [N ii] 6583 A and (b) [Nii] 5755 A lines and (c) their percentage dierence map, and and of n(S+)/n(H+) derived from the (d) [S ii] 6717 A and (e) [S ii] 6731 A lines and (f) their percentage dierence maps. These demonstrate that the derived relative ionic abundance maps are identical. The wedge shown in each frame indicates the adopted log color scale from 5 105 to 3:2104forn(N+) and from 106to for the abundance map and the adopted linear color scale from 0:15 % to +0:15 % for the percentage dierence maps. Image conventions follow from those of Fig. 2. ning a long slit across the nebula during exposure), for example. However, readers are reminded that of numerical values do not carry as previous analyses were not fully spatially re- solved and involved is remarkable here is that the spatial dis- tribution of N+and S+ions derived from appears very much identical to each other (Fig. 11a,b for n(N+)=n(H+) and Fig. 11d,e The percentage dierence turns out to be practically nil: 0 :00040:053 % and0:00516 Ueta and Otsuka 0:036 % forn(N+)=n(H+) andn(S+)=n(H+), respec- tively (Fig. 11c,f). We see a greater concentration of N+and S+ions in the outer part of the main ring. This is expected from the fact that N+and S+are species of lower Comparing N+and S+, the are distributed in a physically narrower extent than N+ions at the periphery of the main ring. These ionic abundance distribution maps would certainly al- low empirical analyses of their spatial variations in more depth, providing excellent constraints for and PDR models. In the past, the abundance derivation for a particu- lar ionic species from dierent usually yielded dierent results, and their average was adopted as the nal abundance value. However, it is not too di- cult to understand that one obtains dierent abundance values from dierent , given how anal- yses were typically done, using ne([Sii]) are not necessarily consistent (discrepancies typi- cally at10 % or greater) with c(H) or H-to-H, and hence, not necessarily consistent with n(N+) andn(S+). Such practice is intolerable in performing plasma diag- nostics for extended objects using spectral maps. Here, PPAP is followed diligently, paying particular atten- tion to self-consistency for both the interstellar extinc- tion correction and plasma diagnostics, to yield iden- tical abundances from multiple transitions of the same ionic species. Thus, Fig 11 demonstrates clearly that plasma diagnostics can be performed as rigorous numerical analyses if we adopt PPAP by seeking a converged self-consistent solution concluding this section, we emphasize the im- portance of self-consistency in abundance analyses from a dierent perspective. Fig. 12 illustrates how much dis- crepancy in relative n(N+)/n(H+) abundance can are deviated from the exact so- lution (set to be 103cm3and 104K) for diagnostics with the [N ii] 6583 A (which is known to be tempera- ture diagnostic) and [N ii] 121m (which is known to be density diagnostic) lines. The top panel of Fig. 12 shows that deviations of tens of % from the rue value are possible when Te([Nii]) is altered even with a few hun- dred K in Te. Similarly, the bottom panel of Fig. 12 proves discrepancies in tens of % can arise if neis varied by only a couple of hundred cm3. In the present analyses above, we see that the ex- act solution for ne([Sii]) ranges from500 to1,500 cm3(Fig. 5, left), while Te([Nii]) ranges from8,000 to1,1000 K (Fig. 5, right). The radial density distribution for these values shows at any given radial bin, there is a spread of 1,000 cm3in ne([Sii]) and2,000 K inTe([Nii]) because of the spa- tial variation in the azimuthal direction (Fig. 6). Ac- cording to Fig. 12, the Te([Nii]) spread of2,000 K corresponds to20 to +30 % discrepancy in the re- sulting N+abundance, while the ne([Sii]) spread of Figure 12. Relative % dierences in the when neandTeare deviated from the reference values (of 103cm3and 104K) based on the [N ii] 6583 A (top) and [N ii] 121m (bottom) cm3corresponds to roughly 50 % as soon as we allow simplications and/or round- ing ofneandTein the course of extinction correction and plasma diagnostics, we are destined to be compro- mised by discrepancies at tens of % in the Overall Quantitative Assessment We have thus established that the proposed PPAP (1) allows us to perform both the extinction correction and plasma diagnostics as a streamlined single process, and (2) oers more self-consistent and exact solutions of c(H) and (ne,Te), plus other derivatives, than any of the previous procedures in the literature. The beauty ofProper Plasma Analysis Practice (PPAP) 17 PPAP may rest on its as it is based solely on a set of relevant line emission distribution maps obtained by some spectral mapping observations, with just the initial choice of the extinction law and RVvalue to adopt. There is no need to assume anything else. To reiterate, PPAP is borne out because there is reciprocal, but subtle, dependence of critical parame- ters between the determination of extinction ( c(H)) and plasma diagnostics ( neandTeof the same energy regime). That is, c(H) is necessary to correct observed line emission maps for extinction in order to perform plasma diagnostics, while c(H) cannot be obtained un- less we obtain neandTeby performing plasma diagnos- tics with line emission maps. Thus, none of these quantities can be determined indepen- dently, and hence, by carefully following through these mutual dependence to the end of the analyses via an iterative search for convergence, we can nd c(H) and (ne,Te) that are consistent with each other. To further demonstrate the robustness of PPAP, here we assess how much improvement is oered by PPAP by quantifying discrepancies that arise by not follow- ing PPAP properly. As an example of such a wrong practice, rst we emulate ypical plasma practiced in the literature by adopting the fol- lowing assumptions and No QP line extraction The raw F656N and F658N images are adopted as the Hand [N ii] 6583 A line emission Hence, the adopted H and [N ii] 6583 A line maps are compromised by mutual Uniform theoretical H -to-Hratio of 2.858 Uniform distribution of neat is assumed. Then, it follows that the the- oretical H-to-Hratio is 2.858 (e.g. Hummer & Storey 1987; Storey & Hummer 1995). Hence, the adopted ratio is applicable only when 104K. (3) Simplied ne([Sii]) diagnostic For annediagnostic using the [S ii] 6717-to-6731 A line ratio, a simple analytic form of log is adopted. This ex- pression reproduces the logistic functional the [S ii] line ratio fairly well un- der the assumption of Te= 104K (e.g. Osterbrock & Ferland 2006; O'Dell et al. 2013b). Hence, if from 104K, the relation would not yield correct ne([Sii]). (4) Simplied Te([Nii]) diagnostic For aTediagnostic using the [N ii] the 5755- to-6583 A line ratio, an analytic expression = 5 :17 + 2:5T1=2 e) is adopted. This is an approxi- mated function that relates Teto the [N ii] lineratio under the assumption of Ne= 103cm3 (e.g. Osterbrock & Ferland 2006; Pradhan & Na- har 2015). Hence, if the true nedeviates from 103cm3, the relation would not yield short, in this example, the assumption of of of 104K is forced, and the issue of line contamination is not properly mit- igated. Hence, unless uniform neandTecome out from PPAP as assumed, the results of the analyses have to be regarded as suspect because inconsistencies between the initial assumptions and the outcomes are not resolved. As another example of a wrong practice, we per- form PPAP only partially, by omitting both or either one of the two components of PPAP: the QP line extrac- tion and iterative search for convergence. We can then quantify how much discrepancies can arise in the end by failing to remove line contamination and/or leaving among the reciprocally dependent in the problem. The resulting discrepancies from all of these trial cases are graphically presented in Fig. 13 as relative percent- age dierence distribution maps. For c(H) and (ne, Te), the outcomes from the wrong analyses are com- pared with the results from full PPAP ( SS3.1, SS3.2). For the N+and S+abundances, a comparison is made be- tween the resulting two abundance maps obtained from the adopted two diagnostic lines individually ( SS3.5). The mean and standard deviation of the distribution of the relative percentages are also summarized in Table 2. In general, with the theo- retical H-to-Hline ratio, and hence, c(H). This is rather intuitive as the presence of attenuating dust grains (higher c(H)) suggests more contrasted H -to- Hline ratios and lower degrees of ionization (lower ne andTe). However, when iterative adjustment of c(H) and (ne,Te) are not performed, the initial assumption of (ne,Te), i.e., how they dier from the true ( ne,Te) values, greatly aects the magnitude of inspection of Fig. 13 reveals a great deal of information. The rst row of Fig. 13 and Table 2 between the outcomes from the procedure in the literature and full PPAP. By comparing the absolute value of the median and the standard deviation of the spread, we can say that underestimated, and c(H) to a lesser ex- tent. Plus, the abundance maps derived from dierent lines dier by 10 %, roughly speaking. The major issue of this ypical procedure boils down to the forced uniform The dis- crepancy between the imposed neandTeand as osets in c(H). For the present case, given the spread in the derived ( values (Fig. 5), the forced uncertainties al- ready amount roughly to 10{20 %. As the adopted ne diagnostic curve is very much steeper than the Tecoun-18 Ueta and Otsuka Figure 13. Graphical summary of discrepancies that result from various inconsistencies injected over the course of PPAP. The observed discrepancies are displayed in terms of the relative percentage dierence distribution in c(H) (left-most column from left), and Te(middle column) compared to the full PPAP results ( SS3.1, Fig. 2; SS3.2, Fig. 5) and in n(S+)=n(H+) (second column from right), and n(N+)=n(H+) (right-most column) between the individual results from each of the two diagnostic lines ( SS3.5, Fig. 11). The four dierent permutations of the source of inconsistencies are (1) an simplied method in the literature (top row; panels a{e, marked as TYP), (2) omitting both the QP line extraction and iterative search for convergence (second row; panels f{j, marked as None), (3) executing only the QP line extraction (third row; panels k{o marked as QP), and (4) executing only the itrative search for convergence (fourth row; panels p{t marked as Iter). The color wedge on the right indicates the displayed range of the relative percentage dierence between 75 % to 75 %. The median and standard deviation of the relative percentage dierence for each case are also summarized in Table 2. Other image conventions follow those of Fig. 2. terpart, even slight osets in c(H) (and hence, in di- agnostic line ratios) would in uence nemore than Te. The greatest discrepancies of all at 65 % in nein this procedure reminds us of its shoddiness. Teis o by 7 % on average, but we can still consider that the true value. Similarly, this method is the only one among all that produces abundance maps that are not consistent with each other. Thus, this exercise exem-plies the of such a ypical method of plasma diagnostics. The other rows of Fig. 13 compare dierent permutations of PPAP. The TYP and one is that in one we rigor- ously consider the neandTedual dependence on both of functions. With QP, the mutual contamination between H and [N ii] is ad- dressed, and hence, the resulting c(H) is among theProper Plasma Analysis Practice (PPAP) 19 PPAP
c(H)
ne
Te
(n(S+)=n(H+)) (%) (%) (%) (%) (%) TYP 14116517 7 17 6 1 8 1 None 14112242 7 17 1213 1934 QP 1 43342 7 17 6 21 1933 Iter 18125322 4 5 (101) 6 343 (104) 6 522 (104) Table 2. Quantitative summary of discrepancies that result from various inconsistencies injected over the course of PPAP in terms of the median and standard deviation of the relative percentage dierence distribution for each case shown in Fig. 13. best (Fig. 13f,k,p). This indeed proves the signicance of the QP process where appropriate. neturns out to be a dicult quantity to determine. But, it may be ex- pected from the fact that the nediagnostic curve is a rather steep function of the adopted diagnostic line ra- tio. With Iter, what is striking is the goodness of the match we see in Teand N+and S+abundances: the iterative process alone recovers the correct solutions for these quantities. Nevertheless, it alone cannot It appears that the correctness of is sacriced at the expense of the other quantities during the iterative optimization of the solu- tion. All in all, lessons learned from this exercise are The widely practiced simplications in ypical plasma di- agnostics without considering the extinction would only feed discrepancies. Both the extinction correction and plasma diagnostics must be performed altogether as an integrated process. As for PPAP, it is necessary to perform the QP process (or anything takes care of spatially resolved line calibra- tion) iterative search for converged solutions. If not, im- posed inconsistencies would amplify as relative discrep- ancies at tens of % over the course of the whole analyses in one way or another, aecting various results seemingly at random. In other words, if self-consistency is main- tained without imposing obviously unrealistic in performing plasma the extinction correction, it is possible to ob- tain rather exact solution for c(H) and (ne,Te) as well as other derivatives such as abundances and carry out any subsequent quantitative analyses quite rigorously. In extreme cases where the goal is determining just Te, however, it may be tolerable as long as and plasma diagnostics are performed itera- tively for converged solutions. This is possible when the Tediagnostic curve happens to be only a weak function of the diagnostic line ratios. In that case, it is critical to remember that c(H) andneare compromised (i.e., the remaining discrepancies are absorbed as osets in Other Caveats 3.7.1. Plasma Diagnostics with [O iii] LinesFor the present analyses, we adopt [S ii] and [N ii] line ratios as our primary diagnostics. Naturally, these diag- nostics are suited to probe low-excitation regions in the outer part of the main ring where these emission lines are strong (Fig. 8d{h). The present data set also includes the [O iii] line maps at 4363 and 5007, A (Fig. 8a,c), which provide another diagnostic line ratio for regions in the inner part of the main ring. The dierence in the spatial coverage among these di- agnostics is evident from the line ratio maps (Fig. 10). This spatial anti-correlation is actually a critical point to consider in plasma diagnostics, especially when selecting two diagnostics as a pair. However, this view was often neglected in the past because spectra were rarely taken in a spatially resolved manner. Therefore, it is instruc- tive to closely examine how plasma diagnostics work out taking into account the spatial distribution of especially when one deals with spatially images. Fig. 14 presents the ne-Tediagram at four positions in the main ring. Panels (a) and (b) are of a local peak in [O iii] (Fig. 8e,g) near the inner and outer edge of the main ring along the short axis (i.e. NNW panels (c) and (d) are of a local peak in [N ii] (Fig. 8d{f) and [S ii] (Fig. 8g,h) as well as [O iii] near the inner and outer edge of the main ring along the long axis (i.e. NSW direction), respectively. These show how each of the [S ii], [Nii], and [O iii] diagnostic line ratios behaves dierently at each position because of the local physical conditions (Fig. 14). As reviewed in SS2.3.2, any given diagnostic line ratio can be expressed as a non-linear function of neandTe. Hence, each line ratio takes the form of a curve in the ne-Tespace as shown in Fig. 14. The shaded thickness for each line represents uncertainties propagated from the assumed 20 % uncertainty in the extinction correc- tion. Then, the intersection of two diagnostic curves would specify neandTe, provided that the correspond- ing emission lines arise from the same region of the tar- get nebula. Because these curves are non-linear func- tions, intersections can only be determined is indeed why plasma diagnostics must be done as an iterative process. In addition, for such a numerical neandTesearch to proceed optimally, an nediagnostic should cover a wide range ofTeat nearly constant ne. Similarly, a Tediag-20 Ueta and Otsuka Position (a) Position (b) Position (c) Position (d) Figure 14. Thene-Tediagrams for the range of 102:5<n e<105cm3and 5;000<Te<15;000 K, showing the [S ii], [Nii], and [O iii] diagnostic curves at four local emission peaks in NGC 6720: (a) the inner edge of the main ring along the short axis, (b) the outer edge of the main ring along the short axis, (a) the inner edge of the main ring along the long axis, and (b) the outer edge of the main ring along the long axis. The thickness of each diagnostic curve represents uncertainties propagated from the assumed 20 % uncertainty in the input line ratio. nostic should cover a wide range of neat nearly constant Te. This is because we would like an nediagnostic im- mune to the wrong so that the of the diagnostic can be ignored, and vice versa. Still, for the optimal results, both ne andTeshould be simultaneously determined by through a nested iterative process. As demonstrated by Fig. 14, diagnostic curves inter- sect with each other at multiple positions in the ne-Te space. This exemplies the fact that we cannot choose any pair of diagnostic lines, dierences of a few hundred cm3and a few thousand K can easily cause at least tens of % dierence in the resulting abundances as shown above at the end ofSS3.5. Ideally, we should like to select that are of similar transition energies arising from the same part of the target object. If not, the easily be o by a few thousands K and/or a couple of hundreds of cm3, and hence, several tens of % discrepancies in the resulting Contamination by He iiLines As shown in Fig. 8b, the He ii4686 A line is reason- ably strong: its relative strength to H is as strong as the [S ii] 6717/31 A doublet and roughly 10 % of H. Thus, the eect of the He iiline contamination in the Hand Hmaps needs to be assessed. If it is signif- icant, the resulting c(H), and subsequently, , will be compromised, and will have to be corrected for accordingly. To this end, we perform the iterative derivation of c(H),ne([Sii]), andTe([Nii]) one more time using the H and Hmaps from which their respective contaminating He iilines at 6560 and 4859 Plasma Analysis Practice (PPAP) 21 This consideration actually adds another layer of com- plication in PPAP, because now we have to know the extinction, c(), in order to derive the very at Step (2). More specically, the observed H -to- Hratio that is needed to derive c(H) (Eq. (2)) has to be replaced by the version of the H which He iicontamination is subtracted. How- ever, to calculate the He iiline ux distribution at 6560 and 4859 A (the contaminants to be subtracted from the Hand Hmaps, respectively), the Heii4686 A map needs to be scaled by an line ratio (e.g. Hummer & Storey 1987; Hum- mer & Storey 1998; Storey & Hummer 1995), i.e. I(H)I(He II) 6560 I(H)I(He II) II) II) II) II) II) II) II) II) is the theoretical He iiline ratio be- tween wavelengths at 1and2(which is a andTe([Nii]); e.g. Hummer & Storey 1987; Hummer & Storey 1998; Storey & Hummer 1995) and the subscripted values refer to the line this H -to-Hratio without the He we perform the same iterative processing to obtain the c(H),ne([Sii]), andTe([Nii]) maps. This modication tends to increase the H -to-Hra- tio slightly, and hence, tends to increase the result- ingc(H) slightly. The dierence of the c(H) values before and after applying this He iicontamination re- moval is quantied as the percentage dierence He II )=c(H)no He II , and plotted as a radial prole in Fig. 15. The revised c(H) with the He iicontamination mit- igated comes out slightly greater. Thus, the He turns out greater in H than in Hin the present case. As a result, the percentage dierence ratio comes out to be negative. The dierence is minor in the outer part of the main ring ( >2500; at2:912:7 %), and hence, the revised ne([Sii]) andTe([Nii]) distribu- tions do not show much dierence. On the other hand, the dierence increases progressively toward the inner edge of the main ring ( <2500): the dierence becomes as large as20 % at the inner edge ( 2000). It appears that the relative amount of the He grows greater toward the center of the main ring into the inner cavity. This agrees with the expecta- tion that the eects of the He iicontamination is greater Figure 15. The radial density distribution of the in c(H) compared to the results without the He The original c(H) result (with the He is underestimated especially in the inner half of the main ring interfacing with the high-excitation central cavity (<2500). The 0 % region seen beyond 3500is the no- data part beyond the outer edge of the main ring. where the degree of excitation is higher. For the present case of NGC 6720, it appears that only the inner cavity is where the He iicontamination needs to be as also suggested from its spatial 8b). Thus, we conclude that the He must be properly assessed when determining quan- tities pertaining to high-excitation ionic species (such as Heii, of course). 3.7.3. Contribution by Recombination We should also bear in mind that strong recombina- tion lines may in uence the Tedetermination (e.g. Liu et al. 2000; G omez-Llanos et al. 2020). For the present case, the [N ii] 5755 A line ux may be enhanced by re- combination from the N2+lines to the said line, and that the resulting Te([Nii]) may not solely be of collision. This is, however, not likely for NGC 6720. Zhang et al. (2004) reported that such contribution by re- combination was extremely small in this nebula. Also, based on the far-IR line emission maps obtained with theHerschel Space Observatory as part of the Nebula Survey (Ueta et al. 2014), we can see that the spatial distributions of the [N ii] 122m and [N iii] 57m lines (not shown here; Ueta et al. in preparation ) show similar spatial dierences as we see between the [N ii] 5755/6548/6583 A lines and the [Oiii] 4363/5007 A lines (Fig. 8). When such contribution by recombination is expected to be signicant, one needs to assess and remove the N2+ contribution by recombination. Under PPAP, we can do22 Ueta and Otsuka so by adding another iterative loop, as done to by He iilines (see SS3.7.2). For and [N iii] collisional exci- tation line maps (e.g. at 1750 A or 57m), one can syn- thesise the map from N2+recombination contribution to [N ii] 5755 A by following Eq.(1) of Liu et al. the updated [N ii] 5755 A line ux, which represents purely collisionally excited line ux, is used to compute Te([Nii]) andne([Sii]) and derive c(H), which would then update the extinction corrected line maps for the next round of the have established a proper plasma analysis prac- tice (PPAP), a streamlined iterative procedure that inte- grates the extinction correction and plasma diagnostics for spatially extended targets in one go ( SS2.3; Fig. 1). The major strength of PPAP is that it requires nothing other than just the input spectral images of critical di- agnostic lines plus the choice of the extinction law and RVvalue to adopt. In other words, there is absolutely no need to assume anything. This work is motivated by the recognized but sub- tle dependence of the power-law extinction index, c(), on both the electron density and temperature, ( ne,Te), which are the very quantities that must be known Such circular dependence must be re- solved via an iterative process looking for the optimum c(H) (which isc() at H) and (ne,Te) values for con- vergence. Unfortunately, however, these have rarely been considered altogether as one in the past. With this in mind, using a suite of narrowband l- ter images of NGC 6720 taken with WFC3 on HST , we have derived by following PPAP a solution of c(H) (SS3.1; Fig. 2) and (ne,Te) (SS3.2; Fig. 5) for the object In the present exercise, the obtained to the NW quadrant of the main ring structure in NGC 6720, for which critical diagnostic line maps are available (e.g. the low-excitation [S ii] and [N ii] diagnos- tic line ratio maps). The derived c(H) and maps clearly reveal spatial variations within the nebula, which would not have been uncovered if uniform ( ne,Te) were assumed, in unprecedented detail. Also, the c(H) map correctly accounts for both the ISM and (Fig. 3) to the whole extent of the nebula for which we have measurements. The ionized gas-to-dust mass ratio in the main ring is found to be fairly with the median of 437 357, much greater than the canonical value of 100 ( SS3.3; Fig. 7). The total gas- to-dust mass ratio in the main ring of NGC 6720 is es- timated to be about 1600. Moreover, we have line emission maps ( SS3.4; Fig. 8) as well as the relative ionic abundance distribu-tion maps (N+and S+relative to H+;SS3.5; Fig. 11) as the essential products of plasma diagnostics. We have demonstrated that dierence of only a few hundred cm3inneor a few thousand K in Te(com- monly seen as the amount of spatial variations; Fig. 6) can introduce dierences in tens of % in the result- ing metal abundances (Fig. 12). Therefore, in the era of spectroscopy, we cannot aord to continue this old tradition of assuming constant neand Te(among other things) without paying attention to the interdependence between extinction correction and plasma diagnostics as well as self-consistency among pa- rameters. If we perform plasma diagnostics without PPAP, we will not be able to maintain the integrity of data at each spatial element. This is because results of such analyses will always suer from uncertainties at tens of % that can easily arise from inconsistencies intro- duced by the unnecessary assumptions. This will sim- ply defeats the purpose of conducting spectroscopy in the rst place. We have also evaluated how much discrepancies can arise at each step of the process if PPAP is not ( SS3.6; Fig. 13; Table 2). Our analyses that plasma diagnostics typically prac- ticed in the literature and/or neglecting PPAP even par- tially would sustain uncertainties greater than 10 %. We have also estimated that for PPAP to be eective line uxes need to be determined at un- certainties better than 10 % in general. The strength of securing self-consistent has been vindicated by the fact that andn(S+)/n(H+) distribution maps de- rived from the corresponding [N ii] and [S ii] line maps of distinct transitions come out to be identical. Such a feat has never been accomplished in plasma diagnostics with spectral imaging data in the past. Only by adopt- ing PPAP fully and not incorporating any assumptions that would degrade the observational data in any way (especially spatially), we can obtain robust there are always certain sets of diag- nostic lines that are suited to probe particular parts of target sources, like the [N ii] and [S ii] line ratio pair that probes low-excitation regions. One should always be conscious about which lines to use depending on dierent local physical conditions to be probed ( SS3.7.1). To that end, when probing high- excitation regions for which the [O iii] lines or other di- agnostic lines of the same and higher transition energy regimes work, one should like to remove contamination in H lines by He iilines and consider contribution by strong recombination lines in metal lines for optimum results ( lessons learned from the present exercise are; 1. Plasma diagnostics should really be performed with the extinction correction as an integrated it- erative procedure for the best results.Proper Plasma Analysis Practice (PPAP) 23 2. We should remind ourselves that the (or any H iline ratio) is a never a constant. 3. Imposing ( ne,Te) constancy in plasma diagnos- tics will inject uncertainties at tens of % in the outcomes of the analyses. In the literature, are often attributed to the local uc- tuations of physical conditions. In reality, the ini- tial constancy assumption is what amplies uncer- tainties: PPAP can resolve such local uctuations.4. As long as self-attenuation by the is expected, it is wrong to adopt the ISM extinction value for target sources, as the non- negligible circumsource extinction will be surely missed (and make ( ne,Te) incorrect). 5. By the same token, it is incorrect to adopt the empirical for ISM when some non-negligible amount of cir- cumsource attenuation is expected, as the ISM re- lation implicitly assumes the gas-to-dust mass ra- tio of 100, which is not necessarily true for indi- vidual target LINE EXTRACTION Using the adopted WFC3 images with the QP method (Ueta et al. 2019), we can separate (i) the H map at 6563 A and the [N ii] maps at 6548 and 6583 A from the F656N and F658N image pair, (ii) the [S ii] maps at 6717 and 6731 A from any pair among the FQ672N, F673N, and FQ674N images, and (iii) the H map at 4340 A and the [O iii] map at 4363 A from the FQ436N and FQ437N image pair. The nebular continuum is determined by taking the mean of the three continuum band images (F547M, F645N, and FQ750N) as long as pixel values register with S =N3. Here, the continuum emission is assumed to be dis- tributed more or less similarly across the relevant spec- tral range. Then, to subtract the continuum from each emission band map, an appropriate scaling factor of the averaged continuum map is determined using the inte- grated ADU counts3of about 10 eld stars that ap- pear in the FoV of both the continuum and emission band images. For this work, we implement the QP algo- rithm anew in Python by adopting the Operator Split- ting Quadratic Program (OSQP) solver (Stellato et al. 2020).4 A.1. Hand [N ii] For the Hand [N ii] line separation, we employ the following two conditions, (i) the positivity condition (i.e. derived line uxes cannot be negative) and (ii) the the- oretical line strength ratio of the [N ii] 6583 A line to the [N ii] 6548 A line being 2.96 (Ueta et al. 2019). The second condition is relevant only in separating H and [Nii] lines from the F656N and F658N pair: we use only the positivity condition in separating other lines. 3The native counts of the archived HST data, usually referred to as the Data Numbers (DNs), in e1or e1s1, depending on the instrument (Desjardins & Lucas from required by the QP formulation, all the input im- ages are taken to have been modulated by the Thus, we multiply the input WFC3 im- ages from the archive by the bandpass unit response (which is the value stored under the PHOTFLAM key- word in the FITS image header and converts the original pixel count into the ux density at erg s1cm2A1) and the bandpass equivalent width (which modulates the surface brightness of the bandpass by the cor- responding system throughput) computed via pysyn- phot (Lim et al. 2015). The QP process then extracts line ux distribution maps of H 6563 A, [N ii] 6548 A, [Nii] 6583 A, [S ii] 6717 A, [S ii] 6731 A, H at 4340 A, and [O iii] 4363 A in erg s1cm2pix1as observed (i.e. Figure 16. Summary of the QP extraction of the H and [Nii] 6583 A line maps from the F656N and F658N images. From top-left to bottom-right, (a) F656N, (b) F658N, (c) QP-extracted H , and (d) QP-extracted [N ii] 6583 A line ux maps (in erg s1cm2pix1; log-scaled from and (e) F656N-to-H and (f) F658N-to-[N ii] 6583 A line ux ratio maps (the wedge on the right indicating the range of the displayed ratio). Image conventions follow those for thec(H) map in Fig. 2.24 Ueta and by the system throughput is corrected for, but the extinction is not yet corrected for). Other nar- rowband images that can isolate each of the correspond- ing target lines are multiplied by the bandpass unit re- sponse and the bandpass rectangular width (which is the equivalent width divided by the maximum in the bandpass) to convert the pixel units to ux (erg s1cm2pix1; modulation by the is corrected for, but the extinction is not yet corrected for) for subsequent processes with the QP- extracted line maps. Fig. 16 shows the original WFC3 lter images of F656N (panel a) and F658N (panel b) and the QP- extracted H (panel c) and [N ii] 6583 A (panel d) line ux maps, along with the original-to-QP line ux ratio maps (panels e and f) of NGC 6720. These results are consistent with the previous results (Ueta et al. demonstrating the power of the QP line extraction. For the present work, we display the surface brightness distribution in the petal the main ring, in which the S/N ratio to show that the faint petal emission is largely of (Lame & Pogge 1994; Martin et al. 2016). The F656N-to-H  ux ratio tends to be low in the cen- tral cavity of the main ring (up to 5 %; Fig. 16e). The F658N-to-[N ii] 6583 A ux ratio map shows lower ratios near the periphery of the main ring (a few %; Fig. 16f). This means that if the F656N map is blindly adopted to represent the H emission distribution, the H emission would be underestimated by at least 3 % in the periph- ery of the main ring and up to 5 % in the central if the F658N map is hastily taken to represent the [N ii] 6583 A emission distribution, the [N ii] 6583 A emission would be underestimated by about 1 % in the central cavity and up to 2 % in the main ring. These dierences may be small, but will be compounded in the subsequent analyses to cause greater importantly, we must remind ourselves that the spatial distribution of H and [N ii] is dierent to begin with. H generally represents the ionized re- gion, whereas [N ii] usually corresponds to the PDR. Therefore, if both the H and [N ii] line maps suer from mutual line blending, the line ra- tio map between them tends to be marginalized, i.e., any structures we observe as H -to-[N ii] ratio varia- tions tend to be washed out as we see in the For example, this H -to-[N ii] would blur the location of the ionization front (IF) that presumably exists where the H -to-[N ii] gradient tends to be large. Thus, keeping the spatial consistency is important in investigating spatially resolved line emis- sion. A.2. [Sii] 6717 and 6731 A Fig. 17 displays the QP results for the [S ii] 6717/31 A line pair. For this group, QP processing can be used with any pair of the three maps or all three maps among Figure 17. Same as Fig. 16, but for the [S ii] 6717 and 6731 A line maps from the FQ672N and FQ674N images and log-scaled ux distribution from 1016:5to 1018:5 erg FQ672N (panel a), F673N (not shown), and FQ674N (panel b) lters. For the present analysis, we opt to use the F673N and FQ674N pair because this pair covers the largest extent of the nebula. The FQ674N lter isolates [Sii] 6731 A (panel d) emission fairly well (less than 1 % dierence; panel f), while the FQ672N lter suers from line blending as high as few 10s of % (panel e). We note that the NE and SE edges of the map are aected by the quad-lter edge eect ( SS6.5 of Dressel 2019; panel f). As we will see below, the [S ii] maps at 6717 A (panel c) and 6731 A (panel d) play a critical role in Hence, the extent of [S ii] emission sets the maximum spatial extent where the present full 2- D plasma diagnostics would be valid. Practically, this means that results of the subsequent 2-D plasma diag- nostics are valid only in the NW quadrant of the main ring (see SS3.7). A.3. H and [O iii] 4363 A Fig. 18 presents the original WFC3 quad-lter im- ages of FQ436N (panel a) and FQ437N (panel b) and the QP-extracted line ux maps of H (panel c) and [Oiii] 4363 A (panel d), as well as the line ux ratio maps (panels e{f). The H emission captured by the FQ436N lter is very little as the system throughput at H is only 1.7 %. Hence, the recovered H surface brightness is fairly uncertain (S=N<2 even at the brightest region). Even if a rea- sonably high S/N is achieved in the H map, because the transition energy of H is higher than H and H, the spatial extent covered by H is more restricted, cov- ering the regions than H and H do. For the present work, therefore, we adopt the QP- recovered H 6563 A (Fig. 16e), [N ii] 6583 A (Fig. 16f), [Sii] 6717 A (Fig. 17e), and [S ii] 6731 A (Fig. 17f) maps as well as the original H (= F486N) and [N ii] 5755 AProper Plasma Analysis Practice (PPAP) 25 Figure 18. Same as Fig. 16, but for the H and [Oiii] 4363 A line maps from the FQ436N and FQ437N im- ages and log-scaled ux distribution from 1016:5to 1018:5 erg s1cm2pix1. (= FQ575N; Fig. 8h) maps because of their reasonable S/N and decent spatial coverage. Based on observations made with the NASA/ESA Hub- ble Space Telescope, and obtained from the Hubble Legacy Archive, which is a collaboration between the Space Telescope Science Institute (STScI/NASA), the Space Telescope European Coordinating Facility (ST- ECF/ESA) and the Canadian Astronomy Data research made use of Astropy, a core Python package for Astronomy et al. 2013, 2018) and PyNeb, a to the analysis of emission lines (Luridiana et al. 2015), as well as OSQP, a convex quadratic pro- grams solver (Stellato et al. was supported partially by the National Aero- nautics and Space Administration under Grant No. NNX15AF24G issued through the Mission Directorate and by the Japan Society for the Promotion of Sci- ence (JSPS) through its invitation fellowship program (FY2020, long-term). MO was supported by for Scientic Research(C) thank Dr. Bob O'Dell for having email dis- cussions about the use of the present HST /WFC data set for NGC 6720 and providing valuable comments after reading an early version of the Astropy (Astropy Collaboration et al. 2013, 2018), PyNeb (Luridiana et al. 2015), PySynphot (Lim et al. 2015), OSQP (Stellato et al. Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: Collaboration, Price-Whelan, A. M., Sip} ocz, B. M., et al. 2018, AJ, 156, 123, doi: T. 1980, The Astrophysical Journal, 240, 99, doi: T. 1987, ApJ, 322, 922, doi: J. A., Clayton, G. C., & Mathis, J. S. 1989, ApJ, 345, 245, doi: 10.1086/16790026 Ueta and T., & Lucas, R. 2019, Introduction to the Hubble Space Telescope Data Handbooks, Version 9.0 (Baltimore, MD: STScI) Draine, B. T. 2003, ARA&A, 41, 241, doi: L. 2019, Wide Field Camera 3 Version 12.0 (Baltimore, MD: STScI) Drissen, L., Martin, T., Rousseau-Nepton, L., et al. 2019, Monthly Notices of the Royal Astronomical Society, 485, 3930, doi: C. F., & Tachiev, G. 2004, Atomic Data and Nuclear Data Tables, 87, 1, doi: D. R., & Dinerstein, H. L. 2001, Journal, 558, 145, doi: omez-Llanos, V., Morisset, C., Garc a-Rojas, J., et al. 2020, MNRAS, 498, L82, doi: G. A. 2012, Ast. Lett., 38, 87, doi: G. A., & Mosenkov, A. V. 2017, MNRAS, 472, 3805, doi: M. A., Manchado, A., & Chu, Y. 1997, Journal, 487, 328, doi: S. A., & Miller, J. S. 1977, The 212, 94, doi: J. J. 2008, ARA&A, 46, 127, doi: D. G., & Storey, P. J. 1987, Monthly Notices of the Royal Astronomical Society, 224, 801, doi: D. G., & Storey, P. J. 1998, MNRAS, 297, 1073, doi: L. J., Nicholls, D. C., & Sutherland, R. S. 2019, Annual Review of Astronomy and Astrophysics, 57, 511, doi: S. 2000, Cambridge Astrophysics Series, 33 Lame, N. J., & Pogge, R. W. 1994, AJ, 108, 1860, doi: P. L., Diaz, R. I., & Laidler, V. 2015, PySynphot User's Guide (Baltimore, MD: X. W., Storey, P. J., Barlow, M. J., et al. 2000, MNRAS, 312, 585, doi: Y., Liu, X. W., Barlow, M. J., & Luo, S. G. 2004, MNRAS, 353, 1251, doi: V., Morisset, C., & Shaw, R. A. 2015, A&A, 573, A42, doi: T. B., Prunet, S., & Drissen, L. 2016, Monthly Notices of the Royal Astronomical Society, 463, 4223, doi: D. C., Kewley, L. J., & Sutherland, R. S. of the Astronomical Society of the Pacic, 132, 033001, doi: C. R., Ferland, G. J., Henney, W. J., & Peimbert, M. 2013a, AJ, 145, 92, doi: 2013b, AJ, 145, 93, doi: 2013c, AJ, 145, 170, doi: 2021, AJ, C. R., & Handron, K. D. 1996, AJ, 111, 1630, doi: C. R., Sabbadin, F., & Henney, W. J. 2007, The Astronomical Journal, 134, 1679 Osterbrock, D. E., & Ferland, G. J. 2006, Astrophysics of gaseous nebulae and active galactic nuclei (Sausalito, CA: University Science Books) Peimbert, M., Peimbert, A., & Delgado-Inglada, G. 2017, PASP, 129, 082001, doi: A. K., & Nahar, S. N. 2015, Atomic Astrophysics and Spectroscopy (Cambridge University Press: P., Gaigalas, G., & J onsson, P. 2019, A&A, 623, A155, doi: S., & Narayanan, D. 2020, ARA&A, 58, 529, doi: B. D., & Mathis, J. S. 1979, ARA&A, 17, 73, doi: D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525, doi: R. A., de La Pena, M. D., Katsanis, R. M., & Williams, R. E. 1998, in Astronomical Society of the Pacic Conference Series, Vol. 145, Astronomical Data Analysis Software and Systems VII, ed. R. Albrecht, R. N. Hook, & H. A. Bushouse, 192 Shaw, R. A., & Dufour, R. J. 1995, PASP, 107, 896, doi: B., Banjac, G., Goulart, P., Bemporad, A., & Boyd, S. 2020, Mathematical 12, 637, doi: P. J., & Hummer, D. G. 1995, MNRAS, 272, 41, doi: S. S. 2011, ApJS, 195, 12, doi: S. S., & Zatsarinny, O. 2010, ApJS, 188, 32, doi: T., Mito, H., Otsuka, M., et al. 2019, AJ, 158, 145, doi:

2108.11007

Toshiya Ueta

Toshiya Ueta and Masaaki Otsuka

Proper Plasma Analysis Practice (PPAP), an Integrated Procedure of the Extinction Correction and Plasma Diagnostics: a Demo with an HST/WFC3 Image Set of NGC6720

27 pages, 18 figures, to be published as a Tutorial in the Publication of the Astronomical Society of the Pacific

10.1088/1538-3873/ac20ab

astro-ph.IM astro-ph.GA astro-ph.SR

http://creativecommons.org/licenses/by/4.0/

In this work, we propose a proper plasma analysis practice (PPAP), an of plasma diagnostics in the era of spectroscopy. In particular, we emphasize the importance of performing both of the and the direct method of plasma diagnostics simultaneously as an integrated process. This approach is motivated by the reciprocal critical parameters in these analyses, which can be resolved by iteratively seeking a converged solution. The use of PPAP allows us to eliminate unnecessary assumptions that prevent us from obtaining an exact solution at each element of the spectral imaging data. Using a suite of HST/WFC3 narrowband images of the planetary nebula, NGC 6720, we validate PPAP by (1) simultaneously and deriving the extinction, c(Hb), and electron distribution, (n_e, T_e), maps that are consistent with each other, and (2) obtaining identical metal abundance distribution n(S^+)/n(H^+)), from multiple emission line maps at energies. We also determine that the derived c(Hb) consists both of the ISM and circumsource components and that the mass ratio in the main ring is at least 437 and as high as about 1600. We find that, unless we deliberately seek uncertainties at tens of per cent can easily arise in outcomes, making it impossible to discern actual spatial variations that occurs at the same level, defeating the purpose of conducting spatially resolved spectroscopic observations.

2 AGRAWAL AND HU is a positive definite matrix, we can assume that its eigenvalues are all positive). Then |I+M|= Tr(MTM). Thus we Tr(MTM)|x|. (2.2) To study the above stochastic differential equation, it is common to introduce the Poisson random measure associated with the L' evy process Zj(t). We write the jumps of the process Zj at timetby Zj(t) :=Zj(t)-Zj(t-) if 0j= letB(R0) be the Borel s-algebra generated by the family of all Borel subsetsUR, such that -UR0. For any t >0 and for any U B(R0) we define the Poisson random measure ,Nj: confusion we use the same notation N), associated with the L' evy process Zj(t) by Nj(t,U) j= 1,2,***,d, (2.3) wherekhUis the indicator function of U. The associated L' evy measure nof the L' evy process Zjis given by nj(U) :=E[Nj(1,U)]j= 1,2,***,d. (2.4) We now define the compensated Poisson random measure ~Njassociated with the L' evy process Zj(t) by ~Nj(dt,dz) (2.5) We assume that the process Zj(t) has only bounded negative jumps to guarantee that the solutionS(t) to (2.1) is positive. This means that there is an interval J= [-R,) bounded from the left such that Zj(t) Jfor allt >0 and for all j= 1,2,***d. With these notations, we can write Zj(t) ordZj(t) write ( 2.1) as dSi(t) =phi(t), t[-b,0], i= 1,***,d. (2.6) In fact we can consider a slightly more general version of system of equations than ( 2.6): dSi(t) i= 1,***,d, Si(t) =phi(t), t[-b,0], i= 1,***,d. (2.7)LOGARITHMIC EULER-MARUYAMA SCHEME 3 First, we discuss the existence, uniqueness and positivity of ( 2.7). Theorem 2.1. Suppose that fij:Rd-Randgij:J functions such that there is a constant a0>1satisfying all 1<=i,j<=d, for allz Jand for all xR, whereJ= [-R,)is the common supporting set of the Poisson measures ~Nj(t,dz),j= 1,***,d. If for all allxR, and phi(0)>=0, i= 1,***,d, then, the stochastic differential delay equation (2.7)admits a unique pathwise solution such that Si(t)>=0almost surely for all i= 1,***,dand for all t >0. ProofThe theorem is stated and proved in [ 2, Theorem 1] following the method of [ 3] (where the case of Brownian motion was dealt with). In fact, the existence and uniqueness are routine and easy. The main point is to show the positivity of the solution. The id ea in [2] was to decompose the solution to ( 2.7) as product of some nonnegative processes. Here we give a slightly different decomposition which will prove the positivity and will be very useful in our numerical =fij(S(t-b)) and ~gij(t,z) =gij(z,S(t-b)). LetYi(t) be the solution to the stochastic differential equation dYi(t) initial conditions Yi(0) =phi(0). Since this is a scalar equation for Yi(t), its explicit solution can be the associated L' evy measure for ~Nj(ds,dz). Letpi(t) bethe solutionto the following system of equations dpi(t) p i(0) = 1, i= 1,***,d. Since by the assumption that ~fij(t)>=0 almost surely for all Theorem [ 1, p.173] for all t>=0 almost surely. Now it is easy to check by the It^ o formula that ~Si(t) =pi(t)Yi(t) satisfies ( 2.7) and~Si(t)>=0 almost surely. By the uniqueness of the solution we see that Si(t) =~Si(t) fori= 1,***,d. The theorem is then rate of logarithmic Euler-Maruyama scheme In this section we construct numerical scheme to approximate ( 2.1) by positive value by the proof of Theorem 2.1we shall decompose equation ( 2.1) into the dXi(t) =pi(t)*Xi(t), i= reason is, as in the proof of Theorem 2.1, thatXi(t) andpi(t) are all positive.4 AGRAWAL AND HU Consider a finite time interval [0 ,T] for some fixed T >0 and let pbe a partition of the time interval [0 ,T]: p: 0 =t0< t1<***< tn=T . Let k=tk+1-tkand = max and assume < b. We shall now construct explicit logarithmic Euler-Marauyama recurs ive scheme to numeri- cally solve ( 3.1a)-(3.1c). By the expression ( 2.8) the solution Xon [tk,tk+1] to Equation ( 3.1a) is given by Xi(t) If we denote by F(x) thedxdmatrix whose diagonal elements are all zero and whose off diagonal entries are fij(x), namely, Fij(x) when i=j fij(x) when this notation we can write ( 3.1b) as a matrix p(t) = (3.2) and its solution on the sub-interval [ tk,tk+1] is given by p(t) = t[tk,tk+1], (3.3) where the exponential of a matrix is in the usual sense: the integral of ~F(S(t-b)) we give the following formula for the sake of given by the Formula (see e.g. [ 4], [5]), where Pris the set of all permutations of {1,2,***,r},e(s) is the number of errors in ordering in {s(1),***,s(r)}, [AB] =AB-BAdenotes the commutator of the matrices, and Tr(t) ={0< u1<***< ur< t}. Analogously to [ 2] we propose the following logarithmic scheme to approximate the solu tion: Xp i(t) =phi(0), pp(0) =1, t k<=t<=tk+1, k= introduce step EULER-MARUYAMA SCHEME 5 Using the above step process we can write the continuous interpola tion forXias Xp i(t) = max{k,tk< t}. From (3.5b) we have pp(t) (3.7) We first show that pp(tk)>=0. Lemma 3.1. Ifph(0)>=0a.s., then pp(tk)>=0a.s. with pp(t) =ph(t)for can be seen from ( 3.5b) and by induction. Assume pp(tk)>=0 a.s. Since by our definition of F(Sp(tk-b)) we know all of its components are positive, we see from ( 3.5b) that pp(t)>=0 a.s. for all we will have Lemma 3.2. Ifph(0)>=0a.s., then Xp(t)>=0a.s. , hence Sp(t)>=0a.s. for all sche me (3.5a)-(3.5d), wemake the following initial data phi(0)>0 and it is H older continuous, i.e. there exist constant r >0 andg[1/2,1) such that for i= 1,2,***,d. (3.8) (A2)fijis bounded. fijandgijare global Lipschitz for i,j= 1,2,***,d. This means that there exists a constant r >0 such <=r,xRd. (A3)The support Jof the Poisson random measure Nj(associated with Z) is contained in [-R,) for each j= 1,2,***,dfor some R >0 and there are constants a0>1 and r >0 satisfying allxRdand for all i,j= anyq >1 there is a i= 1,2,***,d. (3.9) Lemma 3.3. Let Assumptions (A1)-(A4) be satisfied. Then, for any q>=1, there exists Kq, independent of the partition p, such AGRAWAL AND HU ProofFrom our definition of Xp iand boundedness of fijfor alli,jwe , (3.10) whereTt= [0,t]xJ. Denote hj= we used Assumption (A4) and the boundedness of gij. Write for k= 1,2,***,d Mk,t:= (Mk,t,0<=t<=T) is an exponential martingale. Now an application of the Cauchy- Schwartz inequality the same way, we can show This completes the proof of the lemma. Lemma 3.4. Assume Assumptions (A1)-(A4). Then for <1, there is a constant K >0, independent of p, such EULER-MARUYAMA SCHEME for some k. We have v2= for which we write in short v2= For any i= Assumption 2 we can bound by lemma ( 3.3) we can We now bound the other two (3.13) By Assumption 2 it is easy to see that for some constant use the expression for Xp i(t), boundedness of fijfor all i,jand use|ex-ey|<= |ex+ey||x-y|to first factor is bounded and now, we want to bound the second f use the same notation Ito denote different quantities in different occasions and this does not cause ambiguity). We write the above sum as an AGRAWAL AND HU By the inequality, we (3.15) Plugging above, ( 3.14), in (3.12) we get for some (3.16) This proves the lemma. Theorem 3.5. Assume that Assumptions (A1)-(A4) are true. Then, there is a constant of psuch we we want to (3.18) From (3.4), we see that when we have a formula for p(t) which is analogous to the one for pp(t) (Equation ( 3.7) ): p(t) can also write pp(t) < s, we have by the Lipschitz <= EULER-MARUYAMA SCHEME 9 We also (3.22) In the same way we <= we have for some C by lemma 3.4we (3.25) We now bound AGRAWAL AND HU and denote for some C1>0 we the Lipschitz condition on =Kn<forj= 1,2***,d, lemma ( 3.4) and Assumption 3 we the inequality we to the bound for I22we EULER-MARUYAMA SCHEME 11 Similar to the bound for I22using assumption (A2) we the bounds for I21,I22,I23with the help of lemma ( 3.3), we get for some (3.27) We write Then we now bound the Lemmas 3.3and3.25we will have assumption 2 we can show that ppis bounded, hence we can write using ( (3.29) Hence we have for some (3.30) Therefore we we anyp>=2. Now, taking r= 2bin (3.31), we this way, we obtain for any positive integer (3.34) Now, since Tis finite, we can choose a ksuch that ( k-1)b < T<=kb. This completes the proof of the theorem.

2108.11020

Nishant Agrawal

Nishant Agrawal, Yaozhong Hu

Logarithmic Euler Maruyama Scheme for Multi Dimensional Stochastic Delay Differential Equation

12 pages

math.NA cs.NA

http://creativecommons.org/licenses/by/4.0/

In this paper, we extend the logarithmic Euler-Maruyama scheme for stochastic delay differential equation in one dimension to the part where we propose a scheme for a system of stochastic delay differential equations. We then show that the scheme always maintains positivity subject to initial conditions. We then show the convergence of the proposed Euler-Maruyama scheme. With this scheme, all the approximate solutions are positive and the rate of convergence of this scheme is 0.5.

2 FIG. 1: Comparison between the electronic descriptor 0and cohesive energy Ecoh[38] of adsorption-site atoms in describing the energy in dissolution. (a) and (b) (111) surface of TMs. (c)-(f) (111) surface of Cu-based NSAs, (100) surface of Ag-based BAs, and (111) surface of RuRhIrPdPt-based HEAs. The symbol numbers 1 -4 of the inset of (a) and (b) denote the adsorption order in dissolution (see details in Figs. S2-S4 and Note S1 [35]). Taking HEAs as example, the red circle and black triangle of the inset of (c) denote the adsorption site and its nearest neighbors. The violet square and cyan rhombus icons for HEAs in (c)-(f) represent the cases that changing only the element of adsorption sites and that changing all the element that is the same as the adsorption site. energy of adsorption-site atoms Ecoh[38] becomes increas- ingly accurate in describing the adsorption energy with the dissolution process going on. Based on our findings on the undissolved surfaces of TMs [33], we propose an expression of the adsorption energy in dissolution to the maximum bondable and actual bonding numbers of the central atom for a given adsor- bate. The generalized coordination number CNis obtained by dividing the summation of the usual coordination number CN of the adsorption sites0nearest neighbors with the CNin bulk [29, 30].is a constant for a given adsorbate and is by the coupling between the spstates of substrates and the states of adsorbates on TMs ( can be easily deter- mined by only doing calculations on one substrate) [33]. The electronic descriptor Dadfor the adsorption energy on TMs in dissolution can be expressed CNuis the usual coordination number of adsorption sites on the undissolved surfaces. The coefficients 25/2 andTABLE I: Comparison between the predicted and Eq. (2) on TMs and HEAs. The calculated values are 1= 0.035 and 2= 0.099 on NSAs and 1= 0.040 and2= 0.120 on BAs. All data are obtained at the top HEAs 121212 CH 4 1 0.060 0.080 0.062 0.081 0.054 0.080 CH2 4 2 0.040 0.120 0.039 0.110 0.035 0.125 CH3 4 3 0.020 0.160 0.016 0.115 0.015 0.120 CO 4 2 0.040 0.120 0.033 0.114 0.033 0.122 CCH 34 1 0.060 0.080 0.064 0.079 NH 3 1 0.050 0.100 0.046 0.104 0.038 0.098 NH2 3 2 0.025 0.150 0.024 0.148 0.025 0.145 OH 2 1 0.033 0.134 0.021 0.156 0.022 0.135 115 of theEcohterm stem from the slope and offset of the fitting relation between 0andEcoh(Fig. S6 [35]). As the dissolution proceeds, there exists a gradual transition from 0 toEcohfor the electronic determinant of adsorption energy on TMs (Figs. S5, S7-S11, S12a and S12b [35]). Eqs. (2) and (3) describe well the adsorption energy of CH x(x= 1-3), CCH 3, CO, NH x(x= 1, 2), and OH at the different adsorp- tion coverage and sites on TMs in dissolution (Fig. 2a and b and Figs. S13-S15 and Note S3 [35]) [39-41], with the fitted mean absolute error (MAE) of 0.16 eV . The fitted also in good agreement with the predictions by3 FIG. 2: Adsorption energies of the different adsorbates as a function of the proposed descriptors in dissolution. (a) and (b) CHx(x= 1-3), CO, NHx(x= 1, 2), and OH on (111) surface of TMs [39]. (c) CH2on (100) and (111) surfaces of Ag-based BAs in the adsorption-site effect of alloying (blue triangle) and the environment effect of alloying (cyan rhombus), and (111) surface of Cu-based NSAs in the of alloying. (d) CH2and CO on (100), (110), (111), (211) and (532) surfaces of RuRhIrPdPt-based HEAs. The blue triangle and violet square represent the adsorption-site effect of alloying, while the rest do the environment effect of alloying. (e) CH, NH and OH on (111) surface of RuRhIrPdPt-based HEAs in the adsorption-site effect of alloying. The data in (a)-(e) are obtained at the top site. (f) CH at the hcp site and CO at the bridge site on the undissolved (211) surface of AgAu, AgPd, IrRu, and PtRh BAs [40], and OH at the bridge site of (111) surface of RuRhIrPdPt-based HEAs in both the adsorption-site and environment effects of alloying. Eq. (2) (Tables I and S1 [35]). Note that the prefactors 1 and2of Eq. (2), determined by the bonding characters can be deduced and rationalized from the first-order approximation of effective medium theory (EMT) and bond-order conserva- tion framework (Note S4 [35]) [33, 42]. We now quantify the adsorption-site effect of alloying on the adsorption energy during dissolution. Fig. 1c-f and Figs. S12c, S12d and S16-S19 [35] show that Ecohof adsorption- site atoms generally performs better than the describing the adsorption energy of CH x(x = 1-3), CO, NH x(x= 1, 2), and OH at the top site of NSAs, BAs, and HEAs in the adsorption-site effect of alloying for each dissolution step. For HEAs, changing only the element of adsorption sites or changing all the element that is the same as the adsorption site follows the same linear relation (Fig. 1c-f). By further including the effect with the electronic gradient descriptor 0, we introduce a uni- versal descriptor to quantitatively describe the electronic con- tribution of alloys to the adsorption energy in dissolution k12 0 1,k225 2Ecoh,115 transforms to 0for TMs, imply- ing that our electronic gradient descriptor of adsorption en-ergy is universal for TMs and alloys. Since the to the effect of adsorption-site atoms, namely the local electronic effect, while the 0term considers the cou- pling between the adsorption site and its surrounding envi- ronments, the prefactors k1andk2can effectively reflect the electronic localization of alloys in adsorption. We find that k1= 1/10 andk2= 9/10 are the optimum values for describ- ing the adsorption-site effect of NSAs, BAs, and HEAs on the adsorption energy. Accordingly, Eqs. (2) and (4) can de- scribe well the adsorption energy of CH x(x= 1-3), CO, NH x (x= 1, 2), and OH at the top site of NSAs, BAs, and HEAs (Fig. 2c-e and Figs. S14b, S20 and S21 [35]), where the fitted MAEs are merely 0.18 eV for NSAs, 0.12 eV for BAs, and 0.14 eV for HEAs. Moreover, the fitted prefactors 1and2 are also consistent with the predictions by Eq. (2) (Tables I, S2 and S3 [35]). Namely, the prefactors 1and2of alloys are also determined by the bonding characters XmandXas those of TMs do, which comply with the first-order approxi- mation of EMT [42] and bond-order conservation that the model, EMT, is suitable for describing the adsorption behavior on both TMs and alloys (Note S4 [35]). We now turn to understand the environment effect of alloy- ing on the adsorption energy. Fig. 2c and d and Fig. S22 [35] show that the adsorption energies of CH x(x= 0-3), CO, NH x(x= 0-2), OH, F and Cl at the top site of Ag-based NSAs [32, 43-45] and in the environment effect of alloying also follow the4 FIG. 3: (a) The d-band center ( d) and d-band upper edge ( u) against the electronic descriptor 0for atoms on the surface of TMs [25] and RuRhIrPdPt-based HEAs. (b) The d-band width ( Wd) for surface sites on the undissolved (111) surface (1 ) and that with three surface atoms dissolved ( 4 ), against the co- hesive energy [38]. (c) Density of states of the sband of Au atoms on the undissolved (111) surface and that with three atoms dissolved. (d) Comparison between the predicted and DFT-calculated adsorp- tion energies of CHx(x= 0-3), CCH 3, CO, NHx(x= 0-2), OH, F and Cl on TMs, NSAs, BAs, and HEAs [17, 32, 39-41, 43-48]. linear relations in the adsorption-site effect of alloying with the fitted MAE 0.06 eV for NSAs, BAs, and HEAs, which fur- ther demonstrates the robustness of our scheme. Notably, Eq. (4) are different upon adsorbing on the different elements of HEAs in the environment effect of alloying, indicating the distinct localization of the elements in determining the adsorption energy. In particular, the localization of RuRhIrPdPt HEAs in the order of Ru !Ir!Pt!Pd!Rh (Note S5 and Table S4 [35]). When both the adsorption site and surrounding environ- ments change simultaneously from one system to the next, Eqs. (2) and (4) can also capture these complicated alloy- ing effects. For instance, the adsorption of OH at the bridge site of HEAs, the adsorption of CH, CO, C and OH at the top, bridge and hollow sites of BAs and the adsorption of OH, H and OH+H at the hollow site of NSAs comply with our model by Eqs. (2) and (4), with the fitted MAE of 0.05 eV for NSAs, 0.17 eV for BAs, and 0.08 eV for HEAs (Fig. 2f and Figs. S23-S26, Tables S2-S6 and Note S6 [35]) [17, 40, 46-48]. Overall, our scheme is universal for capturing the on TMs and alloys in dissolution. Our scheme reveals a novel physical picture for surface ad- sorption. The d-band model by N rskov et al. [20-22, 25], which takes the d-band center or d-band upper edge as the de- scriptors of adsorption and assumes a constant to adsorption for metals, is mainly applied intoTMs but not complex alloys like BAs and HEAs [40]. Our scheme, which is effective in complex alloys, can also be re- lated to the d-band and s-band properties of TMs but in a sig- nificantly different way from the d-band model. 0reflects the atoms0d-band upper-edge ratio between the adsorption site and its (Fig. 3a), whereas the adsorption- site atoms0cohesive energy does their own d-band width and s-band depth [49]. Namely, the d-band upper-edge ratio of active centers and the d-band width and s-band depth of ad- sorption sites together control the alloying effects in adsorp- tion. The dissolution of a surface atom reduces neighbors0d-band width and s-band depth (Fig. 3b and c and Fig. S27 [35]), making their cohesive energy crucial to determining the adsorption energy on metallics in dissolu- tion. Notably, the exclusion of the s-band contribution makes the model lose its prediction ability by increasing the MAE to!0.32 eV (Fig. S28 and Table S7 [35]), demonstrating contribution of sbands of adsorption sites to the adsorption of complex alloys. Clearly, our scheme goes beyond the d-band model in understanding the adsorption on TMs and alloys in dissolution. Now we access the prediction accuracy of our model. En- couragingly, our model, Eqs. (2-4), is convenient to predict the adsorption energy on metallics as all parameters are The MAE of the predicted adsorption energies rel- ative to the DFT results is 0.14 eV for the adsorption of CH x (x= 0-3), CCH 3, CO, NH x(x= 0-2), OH, F and Cl on TMs, NSAs, BAs, and HEAs (Fig. 3d) regardless of the functionals (Note S6 [35]), which is less than the approximate error of (semi-)local functionals, +-0.2 eV . This model is also predictive for the reactivity at surfaces includ- ing the activation energy, reaction energy, and activity for 11 different reactions, and even includ- ing those on HEAs that can break the BEP relation (Note S7 and Figs. S29-S40 [35]). Considering the large amount (2586 adsorption energies and 539 reactivity data) and wide win- dow (11.16 eV for adsorption energy and 6.30 eV for reactiv- ity) of data as well as the easy accessibility of the descriptors, our model exhibits promising prediction power for surface ad- sorption and reaction on TMs and alloys. In summary, we have identified the electronic gradient and cohesive energy of surface sites as new powerful descriptors for understanding surface adsorption and reaction on metallics in dissolution. With these descriptors, a predictive model has been established for the quantitative determination of the ad- sorption energy on TMs, NSAs, BAs, and HEAs. Our scheme is a crucial development of d-band model, revealing that the s- band contribution is and indispensable in the accurate description of the alloying and dissolution effects on surface adsorption and reaction. The electronic corresponding to the d-band upper-edge ra- tio of active centers, is crucial for determining the in both the adsorption-site and environment effects of alloying, whereas the cohesive-energy descriptor, correspond- ing to the d-band width and s-band depth of adsorption sites, plays an increasingly important role in determining the ad-

2108.11017

Wang Gao

Bo Li, Wang Gao and Qing Jiang

Predictive model of surface adsorption in dissolution on transition metals and alloys

5 pages and 3 figures in the main text

cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

Surface adsorption, which is often coupled with surface dissolution, is generally unpredictable on alloys due to the complicated alloying and dissolution effects. Herein, we introduce the electronic gradient and of surface sites to characterize the effects of alloying and dissolution. This enables us to build a predictive model for the of the adsorption energy in dissolution, which holds well for transition metals, near-surface alloys, binary alloys, and high-entropy this model uncovers a synergistic mechanism between the ratio, d-band width and s-band depth in determining the alloying and dissolution effects on adsorption. Our study not only provides insights into surface adsorption on alloys but also offers a long-sought tool for the design of advanced alloy catalysts.

without his/her authorization, though the membership inference result may not have formal guarantees. The second application of membership inference is that an attacker can use it to of the pre-training data when it is For instance, hospitals may collaboratively use contrastive learning to pre-train image encoders that can be shared across hospitals to solve various downstream healthcare tasks, e.g., lung CT image based COVID-19 testing [ 16] and skin disease prediction. In such cases, the pre-training data may include sensitive medical images and one hospital may infer other hospitals' sensitive members of the image membership inference methods [ are mainly designed for classifiers. For example, given the confidence score vector outputted by a classifier for an input, they aim to infer whether the input is in the training dataset of the classifier. The idea of existing membership inference methods [ 42,44] is to of the classifier. For instance, the confidence score vectors of members and non-members of a classifier are Therefore, the confidence score vector outputted by the classifier for an input can capture whether the input is a member of the classifier. However, given an input, an image encoder outputs a feature vector for it. The feature vector itself does not capture the overfitting of the image encoder on the input. As shown by our experimental results in Section 5.2, existing methods for classifiers are close to random guessing when generalized to infer the members of an image encoder. Our work: In this work, we propose EncoderMI, the first member- ship inference method against contrastive learning. Threat model. We call an entity (e.g., a data owner, an attacker) who performs membership inference an inferrer . We assume an in- ferrer has a black-box access to a pre-trained image encoder (called target encoder ), which is the most difficult and general scenario. The inferrer aims to infer whether an input is in the of the target encoder. The pre-training of an image encoder relies on three key dimensions: pre-training data distribution ,en- coder architecture , and training algorithm . In other words, we have three dimensions of background knowledge. The inferrer may or may not know each of them. Therefore, we have eight different types of background knowledge for the inferrer in total. In our methods, we assume the inferrer has a shadow dataset . In partic- ular, the shadow dataset could have the same distribution as the pre-training data distribution if the inferrer knows it. Otherwise, we assume the shadow dataset has a different distribution from the pre-training dataset. Moreover, if the inferrer does not know the encoder architecture (or training algorithm), we consider the inferrer can assume one and perform membership inference based on the assumed one. Our EncoderMI. An important module in contrastive learning isdata augmentation . Roughly speaking, given an input, the module creates another random input (called aug- mented input ) via applying a sequence of random operations (e.g., random grayscale, random resized crop) to the input. We observe that contrastive learning essentially aims to pre-train an image en- coder such that it outputs similar feature vectors for two augmented inputs created from the same input. EncoderMI is based on Specifically, when an image encoder is overfitted to itspre-training dataset, it may output more (or less) similar feature vectors for augmented inputs that are created from an input in (or not in) the pre-training dataset. In EncoderMI, an inferrer builds a binary classifier (called inference classifier ) to predict whether an input is a member of the target encoder. Roughly speaking, our inference classifier predicts an input to be a member of the target encoder if the target encoder produces similar feature vectors for the augmented inputs created from the input. Next, we discuss how to build inference a shadow dataset, we first split it into two subsets, namely, shadow member dataset and shadow non-member dataset . Then, we pre-train an encoder (called shadow encoder ) using the shadow member dataset based on the background knowledge of the inferrer (e.g., the inferrer can adopt the same encoder architecture and training algorithm used to pre-train the target encoder if he/she knows them). Given the shadow encoder and the shadow dataset, we extract membership features for each input in the shadow dataset. In particular, given an input in the shadow dataset, we first create n augmented inputs via the data augmentation module of the used to train the shadow encoder, then use the shadow encoder to produce a feature vector for each augmented input, and finally compute the set of between the nfeature vectors using a similarity metric as the membership features for the input. Given these we construct an inference training dataset via labeling the membership features as member ( or non-member) if they are created from an input that is in the shadow member (or non- member) dataset. Given the inference training dataset, we build an inference classifier to infer the members of the target encoder. We consider three types of classifiers: vector-based classifier ,set- based classifier , and threshold-based classifier . Given an input and a black-box access to the target encoder, we first extract for the input and then use an inference classifier to predict whether the input is a member of the target To evaluate EncoderMI, we first conduct experiments on CIFAR10, STL10, and Tiny-ImageNet datasets via encoders by ourselves. Our experimental results show that EncoderMI can achieve high accuracy ,precision , and recall under all the eight different types of background knowledge. For instance, our vector-based inference classifier can achieve 88.7% - 96.5% accuracy on Tiny-ImageNet under the eight types of Moreover, EncoderMI can achieve higher accuracy as the inferrer has access to more background knowledge. We also apply EncoderMI to infer members of the CLIP's image encoder released by OpenAI [ 38]. In particular, we collect some and ground truth non-members of the CLIP's image en- coder from Google image search and Flickr. Our results show that EncoderMI is effective even if the inferrer does not know the pre- training data distribution, the encoder architecture, and the of the CLIP's image When a data owner uses EncoderMI to audit data misuse, an encoder provider may adopt a EncoderMI to evade auditing. When an attacker uses Enco- derMI to compromise pre-training data privacy, a be adopted to enhance privacy. As EncoderMI exploits the over- fitting of the target encoder on its training data, we can that prevent overfitting. In particular, we gener- alize early stopping [ 47], a state-of-the-art countermeasure against membership inference to classifiers, to mitigate membership inference to pre-trained encoders. the idea of early stopping is to train the target encoder with less number of epochs to prevent overfitting. Our results show that it achieves trade-offs between the accuracy of EncoderMI and the utility of the target encoder. More specifically, it can reduce the accuracy of our EncoderMI, but it also incurs classification accuracy loss of the downstream classifiers built based on the target encoder. In summary, we make the following contributions in this work: *We propose EncoderMI, the first membership inference method against contrastive learning. *We conduct extensive experiments to evaluate our EncoderMI on CIFAR10, STL10, and Tiny-ImageNet datasets. Moreover, we apply EncoderMI to CLIP's image encoder. *We evaluate an early stopping based countermeasure Our results show that it achieves trade-offs be- tween accuracy of EncoderMI and utility of the encoder. 2 BACKGROUND ON a large amount of unlabeled images or (image, text) pairs (called pre-training dataset ), contrastive learning aims to pre-train a neural network (called image encoder ) that can be used as a for many downstream tasks (e.g., image an input image, the pre-trained image encoder outputs a feature vector for it. 2.1 Pre-training an Encoder An essential module in contrastive learning is data augmentation . Given an input image, the data augmentation module can create another random input (called augmented input ) by a sequence of random operations such as random grayscale, random resized crop, etc.. An augmented input and the original input have the same size. Moreover, we can create multiple augmented inputs for each input using the data augmentation module. Roughly speaking, the idea of contrastive learning is to pre-train an image encoder such that it outputs similar (or dissimilar) feature vectors for two augmented inputs created from the same (or different) input(s). formulates such similarity as a contrastive loss , which an image encoder is trained to minimize. Next, we introduce three pop- ular contrastive learning algorithms, i.e., MoCo [ 20], SimCLR [ 10], and CLIP [ 38], to further illustrate the idea of contrastive learning. MoCo [20]: MoCo pre-trains an image encoder on unlabeled im- ages. There are three major modules in MoCo: an image encoder (denoted as h), amomentum encoder (denoted as hm), and a dictio- nary (denoted as G). The image encoder outputs a feature vector for an input or an augmented input. The momentum encoder has the same architecture with the image encoder, but is updated much more slowly compared with the image encoder. Given an input or an augmented input, the momentum encoder also outputs a vec- tor for it. To distinguish with feature vector, we call it key vector . The dictionary module maintains a queue of key vectors outputtedby the momentum encoder for augmented inputs created from in- puts in previous several mini-batches. Moreover, the dictionary is dynamically updated during the pre-training of the image encoder. Given a mini-batch of Ninputs, MoCo creates two augmented inputs for each input in the mini-batch. The two augmented inputs are respectively passed to the image encoder and the For simplicity, we use uiandujto denote these two aug- mented inputs. Given the two augmented inputs, the image encoder h, the momentum encoder hm, and the dictionary G, MoCo defines a contrastive loss as the natural exponential function, simcomputes between two vectors, and represents a The final contrastive loss is summed over the theNaugmented inputs (i.e., ui's) corresponding to theNinputs. MoCo pre-trains the image encoder via minimizing the final contrastive loss. Finally, the Nkey vectors (i.e., by the momentum encoder for the Naugmented are enqueued to the dictionary while the Nkey vectors from the oldest mini-batch are dequeued. SimCLR [10]: Similar to MoCo [ 20], SimCLR also tries to pre- train an image encoder on unlabeled images. Given a SimCLR creates two augmented inputs for each input in the mini-batch via data augmentation. Given (denoted as SimCLR aims to pre-train the image encoder such that it outputs similar (or dissimilar) feature vectors for two augmented inputs that are created from the same (or different) input(s). Formally, given a pair (ui,uj)of augmented inputs created from the same input, the contrastive loss is defined as follows: li Iis an indicator function, expis the natural cosine similarity between two vectors, his the image encoder, gis the projection head, and is a The final contrastive loss is summed over the jof all 2*Npairs of augmented inputs, where each to two pairs of augmented inputs pre-trains the image encoder via minimizing the loss. CLIP [38]: CLIP jointly pre-trains an image encoder and a text encoder on unlabeled (image, text) pairs. In particular, the text en- coder takes a text as input and outputs a feature vector for it. Given a mini-batch of N(image, text) pairs, CLIP creates an augmented input image from each input image. For each augmented input image, CLIP forms a correct (image, text) pair using the augmented input image and the text that originally pairs with the input image from which the augmented input image is created, and CLIP pairs using the augmented input image and Therefore, there are Ncorrect pairs pairs in total. CLIP jointly pre-trains an imageencoder and a text encoder such that, for a correct (or incorrect) pair of (image, text), the feature vector outputted by the image encoder for the augmented input image is similar (or dissimilar) to the feature vector outputted by the text encoder for the We observe that these contrastive learning algo- rithms try to pre-train an image encoder that outputs similar feature vectors for two augmented inputs that are created from the same input. Specifically, we can have this observation for MoCo [ 20] and SimCLR [ 10] based on the definitions of their contrastive losses. For CLIP, given an (image, text) pair, the feature vector outputted by the image encoder for an augmented version of the image is similar to the feature vector outputted by the text encoder for the text. Therefore, the feature vectors outputted by the image encoder for two augmented inputs created from the input image are similar since both of them are similar to the feature vector outputted by the text encoder for the given text. As we will discuss in Section 4, our EncoderMI leverages this observation to infer members of an image encoder's pre-training dataset. 2.2 Training Downstream Classifiers The image encoder can be used as a feature extractor for many downstream tasks. We consider the downstream task to be in this work. In particular, suppose we have a labeled dataset (called downstream dataset ). We first use the image encoder to extract feature vectors for inputs in the downstream dataset. Then, we follow the standard supervised learning to train a downstream classifier ) using the extracted feature vectors as well as the corresponding labels. Given a testing input from the downstream task, we first use the image encoder to extract the feature vector for it and then use the downstream classifier to predict a label for the extracted feature vector. The predicted label is viewed as the prediction result for the testing input. 3 PROBLEM FORMULATION 3.1 Threat Model Inferrer's goal: Given an input image x, an inferrer aims to infer whether it is in the pre-training dataset of an image encoder (called target encoder ). We call an input a member of the target encoder if the input is in its pre-training dataset, otherwise we call the input a non-member . The inferrer aims to achieve high accuracy at inferring the of the target background knowledge: We consider an inferrer has a black-box access to the target encoder. We note that this is the most difficult and general scenario for the inferrer. A typical is that the encoder provider pre-trains an encoder and then provides an API to downstream customers. The pre-training of an encoder depends on three key dimensions, i.e., pre-training ,encoder architecture , and training algorithm (e.g., MoCo, SimCLR). Therefore, we characterize the inferrer's along these three data distribution. This background knowl- edge characterizes whether the inferrer knows the distribu- tion of the pre-training dataset of the target encoder. In partic- ular, if the inferrer knows the distribution, we assume he/she has access to a shadow dataset that has the same distribution as the pre-training dataset. Otherwise, we assume the inferrerhas access to a shadow dataset that has a different the pre-training dataset. Note that, in both cases, we consider the shadow dataset does not have overlap with the pre-training dataset. For simplicity, we use Pto denote this dimension of background architecture. The inferrer may or may not know the architecture of the target encoder. When the inferrer does not know the target-encoder architecture, the inferrer can assume one and perform membership inference based on the assumed one. For instance, when the target encoder uses ResNet architecture, the inferrer may assume VGG architec- ture when performing membership inference. We use Eto denote this dimension of background algorithm. This dimension characterizes whether the inferrer knows the contrastive learning algorithm used to train the target encoder. When the inferrer does not know the training algorithm, the inferrer can perform based on an assumed one. For instance, when the training algorithm of the target encoder is MoCo [ 20], the inferrer may perform membership inference by assuming the training algorithm is SimCLR [ 10]. We useTto denote this dimension of background knowledge. We use a denote the three dimensions of the inferrer's background knowledge. Each dimension in Bcan be yes or no, where a dimension is yes ( or no) when dimension of background knowledge is available (or unavailable) to the inferrer. Therefore, we have eight different types of background knowledge in total. For instance, the inferrer knows the pre-training data distribution, architecture of the target encoder, and/or training algorithm when the encoder provider makes them public to increase transparency and capability: An inferrer can query the target encoder for the feature vector of any input or an augmented input. 3.2 Membership Inference Given the inferrer's goal, background knowledge, and capability, we define our membership inference against contrastive learning as 3.1 (Membership Inference against Contrastive Learn- ing). Given a black-box access to a target encoder, the B=(P,E,T), and an input, membership inference aims to infer whether the input is in the pre-training dataset of the target encoder. 4 OUR METHOD 4.1 Overview Recall that a target encoder is trained to output similar feature vectors for the augmented versions of an input in the . Our EncoderMI is based on this observation. an encoder is overfitted to its pre-training dataset, the encoder may output more (or less) similar feature vectors for augmented inputs created from an input in (or not in) the pre-training our EncoderMI infers an input to be a member of the target encoder if the target encoder produces similar feature vectors for the augmented versions of the input. Specifically, in EncoderMI,an inferrer builds a binary classifier (called inference classifier ), which predicts for an input based on some features we create for the input. To distinguish with the feature vectors produced by the target encoder, we call the features used by the inference classifier membership features . Our of an input are based on the similarity scores between the feature vectors of the input's augmented versions produced by the target encoder. Building our inference classifier requires a training dataset (called inference training dataset ) which consists of known members and non-members. To construct an inference we split the inferrer's shadow dataset into two subsets, which we call shadow member dataset and shadow , respectively. Then, the inferrer pre-trains an encoder (called shadow encoder ) using the shadow member dataset. We construct an inference training dataset based on the shadow encoder and shadow dataset. Specifically, each input in the shadow member (or non-member) dataset is a member (or non-member) of the shadow encoder, and we create membership features for each input in the shadow dataset based on the shadow encoder. After building an inference classifier based on the inference training dataset, we apply it to infer of the target encoder. 4.2 Building Inference Classifiers We first introduce how to train a shadow encoder on a shadow dataset. Then, we discuss how to extract membership features for an input. Finally, we discuss how to construct an inference training dataset based on the shadow encoder and the shadow dataset, and given the constructed inference training dataset, we discuss how to build inference a shadow encoder: The first step of our EncoderMI is to train a shadow encoder whose ground truth are known to the inferrer. For simplicity, we use ~hto denote the shadow encoder. In particular, the inferrer splits its shadow datasetDsinto two non-overlapping subsets: shadow mem- ber dataset (denoted as Dms) and shadow non-member dataset (de- noted asDnms). Then, the inferrer pre-trains a shadow encoder using the shadow member dataset Dms. If the inferrer has access to the target encoder's architecture (or training algorithm), the inferrer uses the same architecture (or training algorithm) for the shadow encoder, otherwise the inferrer assumes an architecture (or training algorithm) for the shadow encoder. We note that each input in the shadow member (or non-member) dataset is a member (or non-member) of the shadow membership features: For each input in the shadow dataset, we extract its membership features based on the shadow en- coder ~h. Our membership features are based on the key observation that an encoder (e.g., target encoder, shadow encoder) pre-trained by contrastive learning produces similar feature vectors for aug- mented versions of an input in the encoder's pre-training given an input x, we first create naugmented inputs using the data augmentation module Aof the training algorithm used to pre-train the shadow encoder. We denote the as x1,x2,***,xn. Then, we use the shadow encoder to pro- duce a feature vector for each augmented input. We denote by ~h(xi) the feature vector produced by the shadow encoder ~hfor the aug- mented input xi, Our membership features forthe input xconsist of the set of pairwise similarity scores vectors. Formally, we our membership features for an input xbased on encoder ~h, the similarity between two feature vectors (e.g., S(*,*)can be cosine similarity). Note that we omit the explicit dependency of M(x,~h)onS,A, andnfor simplicity. scores in M(x,~h)and they tend to be large if the input xis a member of the shadow encoder ~h. Constructing an inference training dataset: Given the shadow member datasetDms, the shadow non-member dataset Dnms, and the shadow encoder ~h, we construct an inference training dataset, which is used to build an inference classifier. In particular, given an input xDms, we extract its membership features M(x,~h) and assign a label 1to it; and given an input xDnms, we extract its membership features M(x,~h)and assign a label 0to it, where the label 1represents member and the label 0represents non- member. Formally, our inference training dataset (denoted as E) is as follows: E={(M( inference classifiers: Given the inference training dataset E, we build a binary inference classifier. We consider three types of classifiers, i.e., vector-based classifier ,set-based classifier , classifier . These classifiers use the membership fea- tures differently. Next, we discuss them one by classifier (EncoderMI-V). In a we transform the set of membership features M(x,~h) of an input into a vector. Specifically, we rank the scores in M(x,~h)in a descending order. We apply the ranking operation to the membership features of each input in the inference training dataset E. Then, we train a vector-based classifier (e.g., a fully connected neural the standard supervised learning procedure. We usefvto denote the vector-based classifier. Moreover, to denote this classifier (EncoderMI-S). In a set-based classi- fier, we directly operate on the set of membership an input. In particular, we train a set-based clas- sifier (e.g., DeepSets [ 53]) based onE. A set-based classifier takes a as input and predicts a label (1 or 0) for it. A set-based classifier needs to be i.e., the predicted label does not rely on the order of the set elements. As a result, set-based classifiers and vector- based classifiers require substantially different neural Moreover, set-based classification is generally harder than vector-based classification. For simplicity, we use fsto denote the set-based classifier and we use EncoderMI-S to denote this classifier (EncoderMI-T). In a threshold- based classifier, we use the average similarity score in M(x,~h) of an input to infer its membership. In particular, our threshold- based classifier predicts an input to be a member if and only if the average similarity score in its membership no smaller than a threshold. The key challenge is toAlgorithm 1 Our Membership Inference Method and x Ensure: Member or iffsthen 6: ifftthen 8: if determine the threshold with which the threshold-based clas- sifier achieves high accuracy at membership inference. Given a threshold, we use()(or()) to denote the number of inputs in the shadow member (or non-member) dataset whose average similarity score in M(x,~h)is smaller (or no smaller) than. The accuracy of our threshold-based classifier with the thresholdfor the shadow dataset is 1-(()+())/|D s|. Our threshold-based classifier uses the optimal threshold * that maximizes such accuracy, i.e., minimizes ()+(). If we plot the probability distribution of the average similarity score for shadow members and shadow non-members as two curves, where the x-axis is average similarity score and y-axis is the probability that a random shadow member (or non- member) has the average similarity score, then the threshold *is the intersection point of the two curves. Yeom et al. [ 52] and Song et al. [ 48] leveraged a strategy for membership inference. Different from us, their methods were designed for classifiers and were based on the confidence scores outputted by a classifier. 4.3 Inferring Membership Given a black-box access to the target encoder hand an input x, we use the inference classifier fv(orfsorft) to predict whether the input xis a member of the target encoder h. Algorithm 1 shows our method . Given the input x, the data augmentation module A, and an integern, the function Augmentation produces. We use the target encoder hto produce a feature vector for each augmented input, and then we compute the set of scores as the membership features M(x,h)for the input x. Finally, we use the inference classifier to infer the of the input xbased on the extracted membership features. The function Ranking ranks the similarity scores in M(x,h)in a descending order, while the function Average computes the average of the similarity scores in M(x,h). 5 EVALUATION We evaluate EncoderMI on image encoders pre-trained on unlabeled images in this section. In Section 6, we apply EncoderMI to CLIP, which was pre-trained on unlabeled (image, text) pairs.5.1 Experimental Setup Datasets: We conduct our experiments on CIFAR10, STL10, [30]. CIFAR10 dataset contains 60,000 colour im- ages from 10 object categories. In particular, the dataset con- tains 50,000 training images and 10,000 testing images. The size of each image is 32x32. *STL10 [12]. STL10 dataset contains 13,000 labeled colour images from 10 classes. Specifically, the dataset is divided into 5,000 training images and 8,000 testing images. We note that STL10 dataset also contains 100,000 unlabeled images. The size of each image is 96x96in this [4]. Tiny-ImageNet dataset contains 100,000 training images and 10,000 testing images from 200 classes. Each class has 500 training images and 50 testing images. Each image has size 64x64. Training target encoders: For CIFAR10 or Tiny-ImageNet, we randomly sample 10,000 images from its training data as the pre- training dataset to train a target encoder; and for STL10, we ran- domly sample 10,000 images from its unlabeled data as the pre- training dataset. By default, we use ResNet18 [ 21] as the archi- tecture for the target encoder. Moreover, we use MoCo [ 20] to pre-train the target encoder on a pre-training dataset. We adopt the publicly available implementation of MoCo v1 [ 5] with the setting when pre-training our target encoders. Unless otherwise mentioned, we train a target encoder for 1,600 epochs. For CIFAR10 or Tiny-ImageNet, we treat its 10,000 testing images as ground truth non-member of the target encoder. For STL-10, we treat its 5,000 training images and the first 5,000 testing images as non-member of the target encoder. Therefore, unless for each target encoder, we have 10,000 ground truth members and 10,000 ground truth shadow encoders: In the scenario where the inferrer knows the pre-training data distribution of the target encoder, we randomly sample 20,000 images from the training or unlabeled data of the corresponding dataset as the shadow dataset. In the scenario where the inferrer does not know the pre-training data distribution, we randomly sample 20,000 images from the training data of CI- FAR10 as the shadow dataset when the pre-training dataset is from STL-10, and we randomly sample 20,000 images from the unlabeled data of STL-10 as the shadow dataset when the pre-training dataset is CIFAR10 or randomly split a shadow dataset into two disjoint sets, i.e., shadow member dataset andshadow non-member dataset , each of which contains 10,000 images. We train a shadow encoder using a shadow member dataset. We adopt the same architecture (i.e., ResNet18) for a shadow encoder if the inferrer knows the archi- tecture of the target encoder and use VGG-11 [ 45] with batch nor- malization otherwise. We adopt the same training algorithm (i.e., MoCo) to pre-train a shadow encoder if the inferrer knows the algo- rithm used to pre-train the target encoder and adopt SimCLR [ 10] otherwise. We use the publicly available implementations [ 5,6] with the default parameter settings for both training algorithms in our experiments. We train each shadow encoder for 1,600 epochs.Building inference classifiers: We build inference classifiers based on a shadow dataset and a shadow encoder. EncoderMI-V uses a vector-based inference classifier. We use a fully connected neu- ral network with two hidden layers as our vector-based classifier. In particular, the number of neurons in both hidden layers are 256. EncoderMI-S uses a set-based inference classifier. We choose DeepSets [ 53] as our set-based inference classifier. Moreover, we adopt the publicly available code for DeepSets [ 2] in our For both the vector-based classifier and the we adopt cross-entropy as loss function and use Adam optimizer with initial learning rate of 0.0001 to train for 300 epochs. Note that EncoderMI-T leverages a threshold-based classifier and does not require metrics: Following previous work [ 42,44], we adopt accuracy ,precision , and recall to evaluate membership Given an evaluation dataset that contains ground truth members and non-members of the target encoder, accuracy of a method is the ratio of the ground truth that are correctly predicted by the method; precision of a method is the fraction of its predicted members that are indeed members; and recall of a method is the fraction of ground truth members that are predicted as members by the method. Compared methods: Existing membership inference methods [ 11, 42,44,46,48] aim to infer members of a classifier or a text embed- ding model. We generalize these methods to the contrastive learning setting as baseline methods. In particular, we compare our methods with the following five baseline methods, where the first three are for downstream classifiers, while the last two are for The target encoder is used to train a for a downstream task. Therefore, in this baseline method, we use the target encoder to train a downstream classifier (called target downstream classifier ) for a downstream task, and then we apply existing membership inference methods [ 42,44] to the tar- get downstream classifier. In particular, we consider CIFAR10 as a downstream task and we randomly sample 10,000 of its as the downstream dataset. The downstream dataset does not have overlap with the pre-training dataset and the shadow dataset. Given a shadow encoder and the downstream dataset, we train a downstream classifier (called shadow downstream classifier ) via using the shadow encoder as a feature extractor. We query the confidence score vector for each input in the shadow member (or non-member) dataset outputted by the shadow and label it as member (or non-member). Given these confidence score vectors as well as the corresponding labels, we train a vector-based inference classifier. For a given input, we first query its confidence score vector outputted by the target down- stream classifier and then use the inference classifier to predict whether it's a member of the target encoder. Note that, work [ 42], we rank the confidence scores for an input, which outperforms unranked confidence Choquette-Choo et al. [ 11] proposed inference to a classifier. Roughly speaking, they con- struct a binary feature vector for an input based on some of the input. An entry of the feature vector is 1 if and only if the corresponding augmented version is predicted correctly by the target classifier. This label-only membership inference methodrequires the ground truth label of an input. The pre-training data are unlabeled in contrastive learning, making the method not appli- cable to infer members of an encoder in practice. However, since the pre-training data CIFAR10 and Tiny-ImageNet have ground truth labels in our experiments, we assume an inferrer knows them and we evaluate the label-only method. Note that we cannot evaluate this method when the pre-training dataset is from STL10 as they are unlabeled. Similar to Baseline-A, we also apply this method to a target downstream classifier to infer members of the target encoder. For each input xin the shadow dataset, we create eaugmented in- puts. Moreover, we use the shadow downstream classifier to predict the label of xand each augmented input. We construct a the membership features for x, if and only if the shadow downstream predicts the label of x(or theith augmented input), We label the membership features of an input as member (or non-member) if the input is in the shadow member (or non-member) dataset. Given membership features and their labels, we train a vector-based inference classifier. Then, we use the inference classifier to infer members of a target encoder via a target downstream classifier. We set e=10in our Song et al. [ 48] developed adversarial example based membership inference methods against classifiers. Specifi- cally, they leverage the confidence scores produced by the for adversarial examples crafted from an input to infer whether the input is in the training dataset of the target classifier. For instance, their targeted adversarial example based method (dis- cussed in Section 3.3.1 of [ 48]) first crafts k-1targeted for an input (one targeted adversarial example for each label that is not the input's ground truth label), then uses the to compute confidence scores for each of them, and the confidence scores as membership features for the input, where kis the number of classes in the target classifier. They train an inference classifier for each class of the target classifier and use the inference classifier corresponding to the ground truth label of an input to infer its example of an input can be viewed as the version. Therefore, we consider these methods in our experiments. We note that these methods require the ground truth label of an input and the ground truth label is one class of the down- stream classifier. However, in contrastive learning, often do not have labels. Moreover, even if the have labels, their labels may not be the same as those of the downstream classifier. Therefore, we adapt the targeted based method to a downstream classifier in our given an input x, we use PGD [ 33] to generate ktar- geted adversarial examples based on a shadow downstream classi- fier, wherekis the number of classes of the shadow . We then obtain kconfidence score vectors outputted by the shadow downstream classifier for the ktargeted , and we concatenate them as membership features for x. Finally, we train one vector-based inference classifier based on the membership features of inputs in the shadow dataset, and we apply it to infer members of the target encoder via the target . Moreover, following [ 48], we set the perturbation budget (i.e.,) to be 8/255when generating targeted adversarial examples.Table 1: Accuracy, precision, and recall (%) of the five base- line methods. - means not applicable. (a) dataset Accuracy Precision Recall CIFAR10 55.1 53.4 73.1 STL10 54.3 53.7 47.3 48.2 68.3 (b) dataset Accuracy Precision Recall CIFAR10 54.6 63.1 58.2 STL10 - - - Tiny-ImageNet 51.8 53.7 47.6 (c) dataset Accuracy Precision Recall CIFAR10 52.8 54.1 43.1 STL10 50.5 50.1 50.2 52.1 42.3 (d) dataset Accuracy Precision Recall CIFAR10 50.7 50.6 51.0 STL10 50.1 49.9 49.5 49.3 49.2 (e) dataset Accuracy Precision Recall CIFAR10 64.5 63.8 67.2 STL10 67.0 65.7 68.6 67.8 70.8 Baseline-D. In this baseline method, we consider that an inferrer treats the target encoder as if it was a classifier. In other words, the inferrer treats the feature vector outputted by the target encoder for an input as if it was a confidence score vector outputted by a classifier. Therefore, we can apply the confidence score vector based methods [ 42,44] to infer members of the target encoder. a shadow encoder, we use it to output a feature vector for each input in the corresponding shadow dataset. Moreover, we assign label member (or non-member) to the feature vector of an input in the shadow member (or non-member) dataset. Then, we train a vector-based inference classifier using the feature vectors and their labels. Given a target encoder and an input, we first obtain the feature vector produced by the target encoder for the input. Then, based on the feature vector, the inference classifier predicts the input to be a member or non-member of the target Song et al. [ 46] studied membership inference to embedding models in the text domain. Their method can be used to infer whether a sentence is in the training dataset of a text em- bedding model. In particular, they use the average cosine the embedding vectors of the center word and each re- maining word in a sentence to infer the membership of the sentence. We extend this method to our setting. Specifically, we could view an image as a sentence and each patch of an image as a word. Then, we can use an image encoder to produce a feature vector for each patch. We view the center patch as the center word and 0.0 0.2 0.4 0.6 0.8 5 10 15 20 25 30 Number of augmented inputs inference (b) Figure 1: (a) Precision-recall trade-off. (b) Impact of non ac- curacy. The dataset is CIFAR10. compute its cosine similarity with each remaining patch. Finally, we use the average similarity score to infer the membership of the original image. Specifically, the image is predicted as a member if the average similarity score is larger than a threshold. Similar to our EncoderMI-T, we use a shadow dataset to determine the i.e., we use the threshold that achieves the largest infer- ence accuracy on the shadow dataset. We evenly divide an 3), 3x3(i.e., 9), or 3x5(i.e., 15) disjoint patches in our experiments. We found 3x3achieves the best performance, so we will show results for 3x3in the main text and defer the settings: We adopt the following default parameters for our method: we set n=10and we adopt cosine similarity as our similarity metric Ssince all contrastive learning algorithms use cosine similarity to measure similarity between two feature vectors. By default, we assume the inferrer knows the distribution, the encoder architecture, and the training algo- rithm of the target encoder. Unless otherwise mentioned, we show results on CIFAR10 as the pre-training dataset. When the inferrer does not know the target encoder's training algorithm, we assume the inferrer uses random resized crop only to obtain of an input when querying the target encoder because we found such data augmentation achieves the best performance. Note that we resize each image in STL10 and Tiny-ImageNet to 32x32 to be consistent with CIFAR10. 5.2 Experimental Results Existing membership inference methods are insufficient: Ta- ble 1 shows the accuracy, precision, and recall of the five Note that we consider an inferrer with the strongest back- ground knowledge in our threat model, i.e., the inferrer knows the pre-training data distribution, the encoder architecture, and the training algorithm. In other words, the shadow encoders are trained in the background knowledge B=(,,). We find that the ac- curacies of Baseline-A, Baseline-B, Baseline-C, and Baseline-D are close to 50%, i.e., their accuracies are close to that of random guess- ing in which an input is predicted as a member or non-member with probability 0.5. The reason is that they were designed to infer members of a classifier instead of an encoder. The confidence score vector can capture whether the classifier is overfitted for the input, while the feature vector itself does not capture whether the encoder is overfitted for the input. As a result, these membership can infer the members of a classifier but not an is better than random guessing. The reason is that aTable 2: Average accuracy, precision, and recall (%) of our methods for the target encoder pre-trained on CIFAR10 dataset. (orx) means the inferrer has (or does not have) access to the corresponding background knowledge of the target encoder. The numbers in parenthesis are standard deviations in 5 Precision 86.2 (2.04) 78.1 (2.21) 82.1 (1.91) 87.8 (1.15) 78.9 (1.69) 80.1 (1.01) 89.3 (2.15) 86.8 (2.44) 87.2 (1.89)xx 86.9 (2.03) 79.6 (2.05) 83.3 (1.75) 88.3 (1.64) 79.8 (1.34) 81.0 (1.12) 90.9 (2.44) 87.4 (2.51) 89.2 (1.63) xx 87.0 (1.21) 79.4 (1.39) 83.5 (1.03) 88.4 (1.37) 79.6 (0.98) 81.6 (0.87) 91.4 (1.17) 87.2 (1.46) 87.9 (0.92) xx86.7 (0.81) 79.2 (1.05) 83.0 (0.77) 88.2 (0.83) 79.9 (1.10) 80.4 (0.81) 91.4 (0.76) 87.1 (1.01) 89.1 (0.79)x 87.2 (2.17) 79.9 (1.88) 83.6 (1.66) 88.4 (1.38) 80.1 (1.03) 81.9 (0.98) 91.5 (1.23) 87.9 (1.32) 87.9 (1.01)x87.6 (0.45) 80.3 (0.47) 84.2 (0.39) 88.7 (0.43) 80.6 (0.44) 83.1 (0.37) 91.7 (0.51) 87.7 (0.49) 88.0 (0.46) x90.2 (0.37) 81.1 (0.43) 85.2 (0.37) 89.5 (0.31) 80.5 (0.39) 85.1 (0.33) 93.3 (0.32) 88.2 (0.48) 88.9 (0.38)91.4 (0.28) 83.1 (0.27) 86.6 (0.28) 90.1 (0.23) 80.8 (0.29) 85.9 (0.27) 93.5 (0.22) 88.3 (0.30) 89.1 (0.25) patch of an input can be viewed as an augmented version of the input, and the similarity scores between patches capture the over- fitting of an image encoder to some extent. However, the accuracy of Baseline-E is still low, compared to our EncoderMI. Our methods are effective: Table 2, 5 (in Appendix), and 6 (in Appendix) show the accuracy, precision, and recall of our methods under the 8 different types of background knowledge for CIFAR10, STL-10, and Tiny-ImageNet datasets, respectively. The results are averaged in five trials. First, our methods are effective under all the 8 different types of background knowledge. For instance, our EncoderMI-V can achieve 88.7% - 96.5% accuracy on the 8 types of background knowledge. Second, we find that EncoderMI-V is more effective than EncoderMI-S and EncoderMI-T in most cases. In particular, EncoderMI-V achieves higher accuracy (or precision or recall) than EncoderMI-S and EncoderMI-T in most cases. We suspect that EncoderMI-V outperforms EncoderMI-S be- cause set-based classification is generally more challenging classification, and thus viewing membership infer- ence as a vector-based classification problem can achieve better inference performance. Our EncoderMI-T can achieve similar ac- curacy, precision, and recall with EncoderMI-S, which means that the average pairwise cosine similarity score for an input image already provides rich information about the input's . Third, our methods achieve higher recall than precision, i.e., our methods predict more inputs as members than the standard deviations tend to be larger when the infer- rer has less background knowledge. This is because is less stable with less background 1a shows the precision-recall trade-off of our three meth- ods under the background knowledge B=(,,). The curves are obtained via tuning the classification thresholds in the three inference classifiers to produce different precisions and recalls. Our results show that precision drops slightly as recall increases up to around 0.9, and then drops sharply as recall further of the inferrer's background knowledge: Based on Ta- ble 2, 5, and 6, we have three major observations about the im- pact of the inferrer's background knowledge on our methods. achieves higher accuracy as the inferrer has access to more background knowledge, and we have the same observa- tion for EncoderMI-S and EncoderMI-T in most cases. For achieves 96.5% accuracy when the inferrer knowsTable 3: Accuracy, precision, and recall (%) of our methods with different similarity metrics. The dataset is Precision 91.5 90.0 93.5 Correlation 89.3 87.9 92.2 Euclidean 88.9 85.3 83.2 80.5 87.9 Correlation 76.3 75.5 78.5 Euclidean 75.8 73.6 86.7 85.7 89.0 Correlation 80.6 79.8 81.6 Euclidean 80.7 80.1 82.5 all the three dimensions of background knowledge while achiev- ing 88.7% accuracy when the inferrer does not know any of them for Tiny-ImageNet dataset. Second, among the three dimensions of background knowledge, training algorithm is the most infor- mative for STL10 and Tiny-ImageNet, while the three equally for CIFAR10. For instance, on achieves 94.1% accuracy when the inferrer only has access to the training algorithm, while EncoderMI-V 89.1% and 93.0% accuracy when the inferrer only has access to the encoder architecture and the pre-training data distri- bution. For CIFAR10, EncoderMI-V achieves around 86.8% accuracy when the inferrer only has access to any of the three there is no clear winner among the encoder architecture and pre-training data distribution. For instance, having access to the pre- training data distribution alone (i.e., B=(,x,x)) achieves higher accuracy than having access to the encoder architecture alone (i.e., B=(x,,x)) for all our three methods on Tiny-ImageNet, but we observe the opposite for EncoderMI-V and EncoderMI-T on STL10. Impact ofn:Figure 1b shows the impact of the number of aug- mented inputs non the accuracy of our methods for CIFAR10, where the inferrer's background knowledge is B=(,,). We observe that, for EncoderMI-V and EncoderMI-T, the accuracy first increases and then saturates as nincreases. However, for EncoderMI-S, the ac- curacy first increases and then decreases as nincreases. We suspect the reason is that as nincreases, the number of pairwise in the membership features increases exponentially, making set-based classification harder.25000 20000 15000 10000 50005000 10000 15000 20000 2500083.1% 87.7% 89.8% 90.2% 96.0% 82.1% 87.1% 89.7% 90.1% 95.9% 81.9% 87.0% 89.7% 90.1% 95.9% 81.0% 86.8% 89.7% 90.1% 95.9% 79.9% 86.7% 89.6% 90.0% 20000 15000 10000 50005000 10000 15000 20000 2500080.1% 83.6% 85.6% 86.5% 92.2% 80.0% 83.6% 85.5% 86.4% 92.2% 79.6% 83.6% 85.5% 86.4% 92.2% 79.2% 83.5% 85.2% 86.4% 92.1% 78.4% 83.1% 85.0% 86.4% (b) 20000 15000 10000 50005000 10000 15000 20000 2500082.8% 86.4% 88.2% 88.9% 94.4% 82.7% 86.2% 88.1% 88.9% 94.3% 82.5% 86.2% 88.1% 88.8% 94.3% 82.5% 85.7% 88.0% 88.8% 94.1% 82.2% 85.6% 87.8% 88.7% (c) 2: Impact of the size of the pre-training dataset (x-axis) and the shadow dataset (y-axis) on the accuracy of . Both the pre-training dataset and shadow dataset are randomly sampled from STL10. Impact of the similarity metric S:Table 3 shows the impact of the similarity metric Son our methods, where Correlation refers to Pearson correlation coefficient. We have two observations from the experimental results. First, the cosine similarity metric achieves the highest accuracy (or precision or recall). The reason is that the cosine similarity metric is also used in the pre-training of the target encoder. Second, our methods still achieve high accuracy (or precision or recall) when using different similarity metrics from the one used in the pre-training of the target encoder. Impact of the size of the pre-training and shadow datasets: Fig- ure 2 shows the impact of the size of the pre-training dataset and the shadow dataset on the accuracy of our three methods, where the inferrer's background knowledge is B=(,,). Both the pre-training dataset and shadow dataset are randomly sampled from the unlabeled data of STL10, but they do not have overlaps. Note that we did not use CIFAR10 in these experiments because its dataset size is small and we cannot sample disjoint pre-training and shadow datasets with large sizes. First, we observe that the accu- racy of our methods increases as the pre-training dataset becomes smaller. The reason is that the target encoder is more overfitted to the pre-training dataset when its size is smaller. Second, our meth- ods are less sensitive to the shadow dataset size. In particular, given a pre-training dataset size, each of our three methods achieves sim- ilar accuracy when the shadow dataset size ranges between 5,000 and 25,000. Our results show that the shadow encoder pre-trained on shadow dataset with various sizes can mimic the behavior of the target encoder in terms of membership of data augmentation: The data augmentation opera- tions an inferrer uses may be different from those used to pre-train the target encoder. In this experiment, we explicitly study the im- pact of data augmentation. We assume the inferrer uses a compre- hensive list of four commonly used data augmentation random grayscale, random resized crop, random horizontal flip, and color jitter. We gradually increase the number of overlapped data augmentation operations between the inferrer's comprehen- sive list and the target encoder. In particular, the target encoder starts from only using data augmentation operation Gaussian blur, which is not in the inferrer's list. Then, we add the inferrer's operations to the target encoder's pre-training mod- ule one by one in the following order: random grayscale, random resized crop, random horizontal flip, and color jitter. We calculate 0 1 2 3 4 #Overlapped data augmentation inference accuracy of downstream 3: Impact of data augmentation, where the method is EncoderMI-V and the dataset is CIFAR10. the membership inference accuracy and the classification accuracy of the downstream classifier for each target encoder, where the downstream classifier is built as in Baseline-A. Figure 3 shows the results. We observe the number of overlapped data between the referrer and the target encoder controls a trade-off between membership inference accuracy and utility of the target encoder, i.e., the target encoder is more resistant inference but the downstream classifier is also less accurate when the target encoder uses less data augmentation op- erations from the referrer's comprehensive list. 6 APPLYING OUR METHOD TO CLIP CLIP [ 38] jointly pre-trains an image encoder and a text encoder on 400 million (image, text) pairs collected from the Internet. OpenAI has made the image encoder and text encoder publicly available. We view CLIP's image encoder with the ViT-B/32 architecture as the target encoder and apply our EncoderMI to infer its given an input image, we aim to use EncoderMI to infer whether it was used by CLIP or not. Next, we first setup and then present experimental results. 6.1 Experimental Setup Potential members and ground truth non-members: To evalu- ate our EncoderMI for CLIP's image encoder, we need an consisting of both ground truth members and the pre-training dataset of CLIP is not released to the public. Therefore, we cannot obtain ground truth members of CLIP.0.70 0.75 0.80 0.85 0.90 0.95 1.00 Average pairwise cosine similarity of imagesPotential members Ground truth non-members(a) Google 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Average pairwise cosine similarity of imagesPotential members Ground truth non-members (b) Flickr Figure 4: Histograms of the average pairwise cosine similar- ity for potential members and ground truth we can collect some images that are potential truth non-members of the CLIP's pre-training according to Radford et al. [ 38], the (image, text) pairs used to pre-train CLIP were collected from the Internet based on a set of 500,000 popular keywords. Therefore, we collect the fol- lowing two evaluation datasets, each of which has 1,000 and 1,000 ground truth We use the class names of CIFAR100 [ 30] as key- words and use Google image search to collect images. Ap- pendix A shows the complete list of class names, e.g., and bus. In particular, we use a publicly crawl images from Google search based on the key- words. We collected 10 images for each keyword, and thus we collected 1,000 images in total. We treat these images as potential members as they were potentially also collected and used by CLIP. To construct ground truth non-members, we further collected 2,000 images from Google search using the keywords. We randomly divided them into 1,000 pairs; and for each pair, we resized the two images to the same size, and we concatenated them to form a new image, which results in 1,000 images in total. We treat these images as ground of CLIP. *Flickr. Similar to the above Google evaluation dataset, we collected an evaluation dataset from Flickr using the 100 key- words and a publicly available tool2. Specifically, we collected 1,000 images as potential members. Moreover, we further col- lected 2,000 images and randomly paired them to be 1,000 images, which we treat as ground truth non-members. We acknowledge that some of the potential members may not be ground truth members of CLIP in both of our evaluation datasets. For each potential member and ground truth non-member, we resize them as the input size of CLIP, which is classifiers: We assume the inferrer does not know the pre-training data distribution, encoder architecture, and of CLIP, which is the most difficult scenario for our En- coderMI. We use the inference classifiers we built in our in Section 5. Specifically, in our previous experiments, for each of our three methods (i.e., EncoderMI-V, EncoderMI-S, and EncoderMI-T) and each of the three shadow datasets (i.e., CIFAR10, STL10, and Tiny-ImageNet), we have built 8 inference to the 8 types of background knowledge. We use the inference classifiers corresponding to the background our previous experiments (i.e., the 4: Accuracy, precision, and recall (%) of EncoderMI for CLIP's image encoder. (a) Google Method Shadow dataset Accuracy Precision 70.0 64.5 88.8 STL10 71.0 65.4 67.6 62.0 74.7 70.2 87.1 STL10 74.6 70.2 73.2 68.1 70.3 64.2 89.7 STL10 71.5 65.6 66.3 62.4 90.1 (b) Flickr Method Shadow dataset Accuracy Precision 74.9 72.0 81.5 STL10 73.5 70.9 74.7 73.1 71.7 69.1 79.2 STL10 72.7 71.4 71.6 69.3 73.9 68.8 80.1 STL10 74.5 70.8 74.3 71.2 79.4 classifiers corresponding to the last rows of Table 2, Table 5, and Table 6) to infer members of CLIP. Given an input image, we create 10 augmented inputs and use the CLIP's image encoder to produce a feature vector for each of them. Then, we compute the 45 pair- wise cosine similarity scores between the 10 feature vectors, which constitute the set of membership features for the input image. Our inference classifiers predict the membership status of the input image based on the membership features. 6.2 Experimental Results Cosine similarity score distribution for potential members and ground truth non-members: Recall that, for each potential member or ground truth non-member, EncoderMI constructs mem- bership features consisting of 45 pairwise cosine similarity scores. We compute the average of the 45 pairwise cosine similarity scores for each potential member or ground truth non-member. Figure 4 shows the histograms of the average pairwise cosine for potential members and ground truth non-members in our two evaluation datasets. We observe that the potential mem- bers and ground truth non-members are statistically respect to the average pairwise cosine similarity scores. In par- ticular, the potential members tend to have larger average pairwise cosine similarity scores than the ground truth of EncoderMI: Table 4 shows the accuracy, preci- sion, and recall of the three variants of EncoderMI based on different shadow datasets when applied to the CLIP's image encoder. The accuracy, precision, and recall are calculated based on the 1,000 potential members and 1,000 ground truth non-members in each evaluation dataset. First, we observe that EncoderMI based on dif- ferent shadow datasets achieves high accuracy, e.g., 0.66 - 0.75. Our results imply that overfitting exists in real-world image encoders800 1200 1600 #Epochs to train target cosine Overfitting of the target encoder 800 1200 1600 #Epochs to train target inference accuracy of downstream classifier (b) tradeoff Figure 5: (a) The target encoder is more overfitted to its pre-training data when it's pre-trained for more epochs. (b) Trade-off between membership inference accuracy and encoder utility for early stopping, where the method is EncoderMI-V and the dataset is CIFAR10. such as CLIP. Second, EncoderMI achieves a higher recall than precision, which means that EncoderMI predicts a large fraction of potential members as members. Third, our EncoderMI achieves higher recall (or lower precision) on Google than Flickr. In other words, our EncoderMI predicts more inputs from Google image search as members. The reason is that, on average, the average pair- wise cosine similarity score for inputs from Google image search is larger than that for inputs from Flickr as shown in Figure 4. 7 DISCUSSION ON overfitting via early stopping: Recall that our Enco- derMI exploits the overfitting of a target encoder on its . Note that the overfitting of a target encoder on its pre- training dataset is different from that of a classifier. For instance, when a classifier is overfitted to its training dataset, it may have different classification accuracies on its training dataset and testing dataset. Moreover, the confidence score vectors outputted by the classifier for its training dataset and testing dataset are also statis- tically distinguishable. Given an input, a target encoder outputs a feature vector for it. However, the feature vector itself does not capture the overfitting of the target encoder on the input. Instead, when the target encoder is overfitted to its pre-training dataset, it may output more similar feature vectors for the augmented versions of an input in the pre-training dataset. We find that a target encoder is more overfitted to its when it is trained for more epochs. Recall that the member- ship features for an input constructed by our EncoderMI consist of 45 pairwise cosine similarity scores. For each member or non- member of the target encoder, we compute the average pairwise cosine similarity scores in its membership features, and we fur- ther compute the average pairwise cosine similarity scores over all members (or non-members). Figure 5a shows the average pair- wise cosine similarity scores for members and non-members of the target encoder as the number of pre-training epochs increases, where the pre-training dataset is based on CIFAR10. We observe the average pairwise cosine similarity increases (or decreases) for members (or non-members) as the number of epochs increases. In other words, the target encoder is more overfitted to its when pre-training it for more epochs. Our observation inspires us to counter EncoderMI via prevent- ing overfitting through early stopping, i.e., pre-training a targetencoder for less number of epochs. We evaluate the early stopping based countermeasure against our EncoderMI. In particular, we pre-train a target encoder on the pre-training dataset based on CI- FAR10. After we pre-train the target encoder for some epochs, we calculate the accuracy of our EncoderMI-V under the B=(,,)and we also calculate the of a downstream classifier built based on the target en- coder. Both the pre-training dataset and the downstream dataset are constructed based on CIFAR10 as we described in Section 5, and the classification accuracy of the downstream classifier is calcu- lated based on the testing dataset of CIFAR10. Figure 5b shows the membership inference accuracy of our EncoderMI-V and the classi- fication accuracy of the downstream classifier as we pre-train the target encoder for more epochs. We observe that the early stopping based defense achieves a trade-off, i.e., it decreases the accuracy but also reduces the classification accuracy of the downstream classifier. We note that Song et al. [ 47] found that early stopping outperforms other against membership inference to classifiers, and they also observed a trade-off between membership inference accuracy and classifier utility for early with differential privacy: Differential privacy [ 7, 15,26,36,43] can provide formal membership privacy guarantees for each input in the training dataset of a machine learning model. Many differentially private learning algorithms have been proposed. These algorithms add noise to the training data [ 14], the [ 25,26], or the gradient computed by (stochastic) gradient descent during the learning process [ 7,26,43]. For instance, Abadi et al. [ 7] proposed differentially private stochastic gradient descent (DP-SGD) which adds random Gaussian noise to the gradient com- puted by stochastic gradient descent. It would be interesting future work to generalize these differentially private learning algorithms to contrastive learning. In particular, when the pre-training dataset changes by one input, the encoder learnt by a differentially learning algorithm does not change much. privacy may also incur large utility loss for the encoder, i.e., the downstream classifiers built based on a differentially private encoder may have much lower classification learning: Adversarial learning based [ 29,34] have been studied to mitigate membership inference to classifiers, which were inspired by adversarial learning against attribute inference attacks [ 27]. For in- stance, Nasr et al. [ 34] proposed to add an adversarial to the loss function when training a target classifier, where the adversarial regularization term models a membership accuracy. Jia et al. [ 29] proposed MemGuard which does not modify the training process, but adds carefully crafted perturba- tion to the confidence score vector outputted by the target classifier for each input. Specifically, the idea is to turn the perturbed confi- dence score vector to be an adversarial example to the which make random membership inference based on the perturbed confidence score vector. It would be interesting future work to extend these countermeasures to pre-trained encoders. For instance, we may capture our EncoderMI's accuracy as an adver- sarial regularization term and add it to the contrastive loss an encoder using contrastive learning; we may also add carefully crafted perturbation to the feature vector outputted byan encoder for each (augmented) input such that the set of member- ship features constructed by our EncoderMI for an input becomes an adversarial example to the inference classifiers, which make random membership inference based on the perturbed . A key challenge is how to find such perturbation to each feature vector since the membership features depend on the pair- wise similarity between a set of feature vectors corresponding to the augmented versions of an input. 8 RELATED WORK Membership inference: In membership inference against ma- chine learning classifiers [ an inferrer aims to infer whether an input is in the training dataset of a classifier (called target classifier ). For instance, in the methods proposed by Shokri et al. [ 44], an infer- rer first trains shadow classifiers to mimic the behaviors of the target classifier. Given the shadow classifiers whose ground truth members and non-members are known to the inferrer, the inferrer trains inference classifiers, which are then applied to infer mem- bers of the target classifier. Salem et al. [ 42] further improved these methods by relaxing the assumptions about the inferrer. Hui et al. [24] proposed blind membership inference methods that do not require training shadow classifiers. Concurrent to our work, He et al. [ 22] also studied membership inference against They assume the pre-training data and downstream data are the same. Specifically, given a labeled training dataset, they first use contrastive learning to pre-train an encoder and then use it to fine-tune a classifier on the labeled training dataset. They try to infer whether an input is in the labeled training dataset via methods [42, 44] to the fine-tuned classifier. Our methods are different from these ones as they were designed to infer members of a classifier while our methods aim to infer members of an encoder pre-trained by contrastive learning. Our experimental results show that these methods achieve accuracy close to random guessing when applied to infer members of an encoder. The reason is that the confidence score vector outputted by a classifier can capture whether the classifier is overfitted for an input, while the feature vector itself outputted by an encoder does not capture whether the encoder is overfitted for an input. The similarity scores between the feature vectors of the of an input capture whether the encoder is overfitted for the input, and our methods leverage such similarity scores to infer the membership status of the input. Existing membership inference methods for pre-trained mod- els [8,46] focused on the natural language domain. For et al. [ 8] proposed membership inference methods for GPT- 2 [39], which is a pre-trained language model, and they the membership inference methods to reconstruct the training data of GPT-2. Specifically, they first reconstructed some candidate texts and then applied a membership inference method to determine the membership status of each candidate text. To the best of our knowledge, no prior work has studied membership inference for encoders in the image domain. Prior work [ 23,54] also studied membership inference against transfer learning. For instance, Hidano et al. [ 23] assume a white- box access to the transferred part of the teacher model while Zouet al. [ 54] leverage the posterior of the teacher model. Our work is different from these, because pre-training an image encoder is different from training a teacher model since the former uses con- trastive learning on unlabeled data while the latter uses the learning on labeled against membership inference: Many [ were proposed to inference for classifiers. The first category of [ 42,44,47] try to prevent overfitting when e.g., standard L2regularization [ 44], dropout [ 42], and early stopping [ 47]. The second category of countermeasures [ 7,43] are based on differential privacy [ 15], which often incur large utility loss for the learnt machine learning classifiers. The third category of countermeasures leverage adversarial learning, e.g., adversar- ial regularization [ 34] and MemGuard [ 29]. We explored an early stopping based countermeasure against our EncoderMI. Our re- sults show that such countermeasure achieves a trade-off inference accuracy and utility of an learning: Contrastive learning [ aims to pre-train image encoders on unlabeled data via ex- ploiting the supervisory signals in the unlabeled data itself. The unlabeled data could be unlabeled images or (image, text) pairs. The pre-trained encoders can be used for many downstream tasks. The key idea of contrastive learning is to pre-train an image encoder such that it outputs similar feature vectors for a pair of augmented inputs created from the same input image and outputs vectors for a pair of augmented inputs created from different input images. Examples of contrastive learning methods include MoCo [ 20], SimCLR [ 10], and CLIP [ 38], which we discussed in Section 2. We note that Jia et al. [ 28] proposed BadEncoder, which embeds backdoors into a pre-trained image encoder such that mul- tiple downstream classifiers built based on the backdoored encoder inherit the backdoor behavior CONCLUSION AND FUTURE WORK In this work, we propose the first membership inference method against image encoders pre-trained by contrastive learning. Our method exploits the overfitting of an image encoder, i.e., it produces more similar feature vectors for two augmented versions of the same input. Our experimental results on image encoders pre-trained on multiple datasets by ourselves as well as a real-world image encoder show that our method can achieve high accuracy, precision, and recall. Moreover, we also find that an early stopping achieves a trade-off between membership and encoder future work includes 1) extending our method to the white-box settings in which the inferrer has access to the parame- ters of the target encoder, 2) extending our method to infer the mem- bership of an (image, text) pair, 3) developing new our method, and 4) exploring other of pre-trained image encoders such as stealing their parame- ters [49] and hyperparameters (e.g., encoder architecture) We thank the anonymous reviewers and our shepherd Reza Shokri for constructive comments. This work was supported by NSF under Grant No. Twitter demands AI company stops 'collecting faces'. (2020). [2] 2021. DeepSets. (2021). [3]2021. FTC settlement with Ever orders data and AIs deleted after facial (2021). [4]2021. MicroImageNet classification challenge. (2021). [5] 2021. MoCo. (2021). [6] 2021. SimCLR. Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. 2016. Deep learning with differential privacy. Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert- Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al.2021. Extracting Training Data from Large Language Models. In USENIX Security Symposium . [9]Qingrong Chen, Chong Xiang, Minhui Xue, Bo Li, Nikita Borisov, Dali Kaarfar, and Haojin Zhu. 2018. Differentially private data generative models. arXiv preprint arXiv:1812.02274 (2018). [10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In ICML . [11] Christopher A Choquette Choo, Florian Tramer, Nicholas Carlini, and 2021. Label-only membership inference attacks. In ICML . [12] Adam Coates, Andrew Ng, and Honglak Lee. 2011. An analysis of in unsupervised feature learning. In AISTATS . [13] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. 2015. Discriminative unsupervised feature learning with exemplar convolutional neural networks. IEEE transactions on pattern analysis and machine intelligence 38, 9 (2015), 1734-1747. [14] John C Duchi, Michael I Jordan, and Martin J Wainwright. 2013. Local privacy and statistical minimax rates. In FOCS . [15] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. 2006. Calibrat- ing noise to sensitivity in private data analysis. In TCC. [16] Daryl LX Fung, Qian Liu, Judah Zammit, Carson Kai-Sang Leung, and Pingzhao Hu. 2021. Self-supervised deep learning model for COVID-19 lung CT highlighting putative causal relationship among age, and COVID-19. Journal of Translational Medicine 19, 1 (2021). [17] Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In CVPR . [18] Inken Hagestedt, Mathias Humbert, Pascal Berrang, Irina Lehmann, Roland Eils, Michael Backes, and Yang Zhang. 2020. Membership inference against DNA methylation databases. In EuroS&P . [19] Inken Hagestedt, Yang Zhang, Mathias Humbert, Pascal Berrang, Haixu Tang, Xi- aoFeng Wang, and Michael Backes. 2019. MBeacon: Beacons for DNA Methylation Data. In NDSS . [20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. In CVPR . [21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep for image recognition. In CVPR . [22] Xinlei He and Yang Zhang. 2021. Quantifying and Mitigating Privacy Risks of Contrastive Learning. arXiv preprint arXiv:2102.04140 (2021). [23] Seira Hidano, Takao Murakami, and Yusuke Kawamoto. 2021. TransMIA: Mem- bership Inference Attacks Using Transfer Shadow Training. In IJCNN . [24] Bo Hui, Yuchen Yang, Haolin Yuan, Philippe Burlina, Neil Zhenqiang Gong, and Yinzhi Cao. 2021. Practical Blind Membership Inference Attack via In NDSS . [25] Roger Iyengar, Joseph P Near, Dawn Song, Om Thakkar, Abhradeep Thakurta, and Lun Wang. 2019. Towards practical differentially private convex S & P . [26] Bargav Jayaraman and David Evans. 2019. Evaluating differentially private machine learning in practice. In USENIX Security Symposium . [27] Jinyuan Jia and Neil Zhenqiang Gong. 2018. AttriGuard: A practical defense against attribute inference attacks via adversarial machine learning. In USENIX Security Symposium . [28] Jinyuan Jia, Yupei Liu, and Neil Zhenqiang Gong. 2022. BadEncoder: Backdoor Attacks to Pre-trained Encoders in Self-Supervised Learning. In IEEE S & P . [29] Jinyuan Jia, Ahmed Salem, Michael Backes, Yang Zhang, and Neil Zhenqiang Gong. 2019. Memguard: Defending against black-box membership via adversarial examples. In CCS. [30] Alex Krizhevsky. 2009. Learning multiple layers of features from tiny images. Tech Report (2009). [31] Jiacheng Li, Ninghui Li, and Bruno Ribeiro. 2021. Membership Inference Attacks and Defenses in Classification Models. In CODASPY . [32] Zheng Li and Yang Zhang. 2021. Membership Leakage in Label-Only Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. 2018. Towards Deep Learning Models Resistant to In ICLR . [34] Milad Nasr, Reza Shokri, and Amir Houmansadr. 2018. Machine learning with membership privacy using adversarial regularization. In CCS. [35] Milad Nasr, Reza Shokri, and Amir Houmansadr. 2019. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks and federated learning. In IEEE S & P . [36] Milad Nasr, Shuang Song, Abhradeep Thakurta, Nicolas Papernot, and 2021. Adversary Instantiation: Lower Bounds for Differentially Private Machine Learning. In IEEE S & P . [37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018). [38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al.2021. Learning transferable visual models from natural language preprint arXiv:2103.00020 (2021). [39] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9. [40] Marc'Aurelio Ranzato, Fu Jie Huang, Y-Lan Boureau, and Yann LeCun. learning of invariant feature hierarchies with applications to In CVPR . [41] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, Yann Ollivier, and Herve Jegou. 2019. White-box vs black-box: Bayes optimal strategies for mem- bership inference. In ICML . [42] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. 2019. Ml-leaks: Model and data independent membership infer- ence attacks and defenses on machine learning models. In NDSS . [43] Reza Shokri and Vitaly Shmatikov. 2015. deep learning. In CCS. [44] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. 2017. Mem- bership inference attacks against machine learning models. In IEEE S & P . [45] Karen Simonyan and Andrew Zisserman. 2015. Very deep convolutional networks for large-scale image recognition. In ICLR . [46] Congzheng Song and Ananth Raghunathan. 2020. Information leakage in em- bedding models. In CCS. [47] Liwei Song and Prateek Mittal. 2021. Systematic evaluation of privacy risks of machine learning models. In USENIX Security Symposium . [48] Liwei Song, Reza Shokri, and Prateek Mittal. 2019. Privacy risks of securing machine learning models against adversarial examples. In CCS. [49] Florian Tramer, Fan Zhang, Ari Juels, Michael K Reiter, and Thomas Stealing machine learning models via prediction apis. In USENIX . [50] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. 2008. Extracting and composing robust features with denoising autoencoders. In ICML . [51] Binghui Wang and Neil Zhenqiang Gong. 2018. Stealing hyperparameters in machine learning. In IEEE S & P . [52] Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. 2018. Privacy risk in machine learning: Analyzing the connection to overfitting. In CSF. [53] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, and Alexander Smola. 2017. Deep sets. In NeurIPS . [54] Yang Zou, Zhikun Zhang, Michael Backes, and Yang Zhang. 2020. Privacy Analysis of Deep Learning in the Wild: Membership Inference Attacks against Transfer Learning. arXiv preprint arXiv:2009.04872 (2020). A KEYWORDS FOR GOOGLE SEARCH AND FLICKR CRALWER The keywords are the 100 class names in CIFAR100: beaver, dolphin, otter, seal, whale, aquarium fish, flatfish, ray, shark, trout, roses, sunflowers, tulips, bottles, bowls, cans, cups, plates, apples, mushrooms, oranges, pears, sweet peppers, clock, com- puter keyboard, lamp, telephone, television, bed, chair, couch, table, wardrobe, bee, beetle, butterfly, caterpillar, cockroach, bear, leopard, lion, tiger, wolf, bridge, castle, house, road, skyscraper, cloud, plain, sea, camel, cattle, chimpanzee, elephant, kangaroo, fox, porcupine, possum, raccoon, skunk, crab, lobster, snail, spi- der, worm, baby, boy, girl, man, woman, crocodile, dinosaur, lizard, snake, turtle, hamster, mouse, rabbit, shrew, squirrel, maple, oak,

2108.11023

Hongbin Liu

Hongbin Liu, Jinyuan Jia, Wenjie Qu, Neil Zhenqiang Gong

EncoderMI: Membership Inference against Pre-trained Encoders in Contrastive Learning

To appear in ACM Conference on Computer and Communications Security (CCS), 2021

cs.CR cs.CV cs.LG

http://creativecommons.org/licenses/by/4.0/

Given a set of unlabeled images or (image, text) pairs, contrastive learning aims to pre-train an image encoder that can be used as a feature extractor for many downstream tasks. In this work, we propose EncoderMI, the first method against image encoders pre-trained by contrastive learning. In particular, given an input and a black-box access to an image aims to infer whether the input is in the training dataset of the image encoder. EncoderMI can be used 1) by a data owner to audit whether its (public) data was used to pre-train an image encoder without its authorization or 2) by an attacker to compromise privacy of the training data when it Our EncoderMI exploits the overfitting of the image encoder towards its training data. In particular, an overfitted image encoder is more likely to output more (or less) similar feature vectors for two of an input in (or not in) its training dataset. We evaluate EncoderMI on image encoders pre-trained on multiple datasets by ourselves as well as the Contrastive Language-Image Pre-training (CLIP) image encoder, which is pre-trained on 400 million (image, text) pairs collected from the Internet and released by OpenAI. Our results show that EncoderMI can achieve high and recall. We also explore a countermeasure against EncoderMI via preventing overfitting through early stopping. Our results show that it achieves trade-offs between accuracy of EncoderMI and utility of the image encoder, i.e., it can reduce the accuracy of EncoderMI, but it also accuracy loss of the downstream classifiers built based on the image encoder.

2 Introduction Ever s ince was discovered in mercury at critical temperature (Tc) of 4.2 K in 1911 [1], research in sup erconductivity has been one of the central topics in Physics . Recently, most excitement in this field comes from hydrogen -rich compounds with Tc above 200 K at high pressures [2-17], in which the superconduct ivity is still the Bardeen -Cooper -Schrieffer (BCS) [18] type and described by the Migdal -Eliashberg theor y [19]. However, these high -Tc hydrides require extremely high pressure to synthesize and usually not quench recoverable at ambient pressure , which limits their applications. Obviously , it is desirable to find new high-Tc superconducting materials that are easy to synthesis and stable at ambient conditions . Encouragingly, some non -hydrogen clathrates , in particular the sp3 bonded ones with light elements, have been suggested to be superconducting at moderate or even ambient pressure . For example, several p-doped silicon clathrates based on Ba 8Si46 have been synthesized below 3 GPa, but the highest observed Tc is mere ly 8 K [20-23]. Many doped (especially p-type) XC6 [24,25] , X(BN)6 [26], XB3C3 [27], and XB3Si3 [28] sodalite -like structures , where X represents different doping elements, as well as n- doped FC34 [29] have been predicted to be good superconductors at ambient pressure with the highest Tc exceeding 100 K. However , most of the predicted materials have not been realized due to the experimental difficulties associated with thermodynami c instability . It is noteworthy that a stable host/gu est clathrate SrB 3C3 was synthesized at around 50 GPa [27]. This clathrate, with guest Sr2+ cations incorporated at the voids in the (B 3C3)3- host framework , is a hole conductor with the estimated Tc of 42 K at ambient pressure [30]. A possible way to further enhance the electronic properties of this material is to replace Sr2+ cations by alkali atoms, which would introduce more holes, but previous calculations shows that when all Sr atoms are replac ed, th e structure becomes dynamically unstable [31]. This leads us to consider the possibility to increase dynamic stability, while still bettering the properties, by a partial replacement of Sr2+ cations in th e clathrate. Preceding effort ha s been devoted to optimiz e the superconducting performance of hydrides through substitution or doping 3 [32-38]; some already show significant enhancement in Tc [32,36,37] . Given that Rb has similar atomic radius and as Sr, it is suitable to be incorporated to the synthesized clathrate SrB 3C3 to manipulate electronic properties and result in the significant modification of . Toward this end , we carried out a theoretical research on Rb-substituted SrB 3C3. The compositional space of Rb xSr1-xB3C3 at 0 and 50 GPa was systematically explored using first -principles calculation in combination with cluster expansion (CE) [39] and CALYPSO structure prediction method [40-42]. Through this we found that the configurations with x <= 0.5 are dy namically stable, and unearthed several Rb xSr1-xB3C3 structures. Furthermore, the electron -phonon coupling ( EPC) calculation and solution of the Eliashberg equation indicate that Rb0.5Sr0.5B3C3 in Pm-3 space group is a phonon - mediated superconductor with an estimated Tc of 73-78 K at ambient pressure. These results provide guidance for future experimental studies and stimulate more exploration on boron -carbon -based high -Tc superconducting materials . Computational details Stabilities of the RbxSr1-xB3C3 alloys are investigated using the CE method [39], as implemented in the Massachusetts institute of technology Ab initio Phase Stability (MAPS) package in Alloy Theoretic Automated Toolkit (ATAT) [43]. In the CE method, alloy is treated by the lattice model where the lattice sites are fixed on the Bravais lattice and a configuration s is defined by specifying the occupation of each lattice site. For quasibinary RbB 3C3-SrB 3C3 system, the alloy configuration is described by pseudospin occupation variable si, which take s the value -1 and +1 when the ith site is occupied by Rb and Sr, respectively. The energy (or enthalpy) for a given configuration s is expressed by the following polynomia l: E()=mJ<i i'> , where a is a cluster and Ja is the effective cluster interaction (ECI) . The multiplicity factor ma indicates the number of equivalent clusters to a by symmetry. The precision 4 of the ECIs can be evaluated by the cross -validation (CV) score, defined as : where n is the number of structures used to obtain the ECIs . Ei is the energy of the ith structure calculated using density functional theory (DFT ) and E(i) is the predicted energy of the same structure obtained from CE method fitting to the energies of remaining (N -1) structures . To ensure a satisfactory convergence of CEs , the CV score should be as small as possible. In the present work, a cubic Pm-3n structure of SrB 3C3 is used as the parent s tructure, and the CV scores are 2.2 and 2.5 meV/ atom at 0 and 50 GPa, respectively. The DFT calculations for fitting the ECI were performed with the Vienna Ab Initio Simulation Package (VASP) [44], using projector augmented waves (PAW) pseudopotential s [45] with Perdew -Burke -Ernzerhof (PBE) generalized gradient approximation (GGA) exchange -correlation functional [46]. A plane -wave energy cutoff of 500 eV and a Monkhorst -Pack Brillouin Zone (BZ) sampling grid of 2p x 0.032 A-1 were employed . The convergence criter ia were set to 10-5 eV per unit cell for energy and 0.02 eV/A per atom for forces. To examine the validity of the CE method, we performed a crystal structure search for RbSrB 6C6 using a heuristic algorithm based on the particle swarm -intelligence approach as implemented in the CALYPSO cod e [40-42]. The structures of RbS rB6C6 were searched using simulation cells with 1 -2 formula units (f.u.) at 50 GPa. The total number of predicted structures is at least ~2000. To determine the dynamic stabilities of predicted structures by the CE metho d, phonon calculations were performed using the finite displacement method implemented in the PHONOPY software [47]. The EPC calculations for superconducting properties were performed with density functional perturbation theory (DFPT) [48] using Quantum ESPRESSO package [49]. GBRV ultrasoft pseudopotentials [50] were used for Rb , Sr, B and C, with a kinetic energy cutoff of 60 Ry, a dense k-point mesh with a 5 spacing of 2 p x 0.008 A-1, and a q-point mesh with a spacing of 2 p x 0.0 34 A-1. The superconducting gap and Tc were obtained by numerically solving the Eliashberg equation [19]. This equation provides a reasonab le estimate of Tc for strong -coupling cases with EPC parameter l above 1.5. Results and discussion s The CE simulations were performed on crystal structures with up to 84 atoms. To ensure a satisfactory convergence , 141 and 113 structures were used for fitting while 6342 and 6370 structures were employed for prediction at 0 and 50 GPa, respectively . The corresponding CV scores are 2.2 and 2.5 meV/at om. Fig. 1 shows the calculated formation enthalpies for various RbxSr1-xB3C3 compositions at 0 and 50 GPa. We found 14 and 11 stable structures of RbxSr1-xB3C3 at 0 and 50 GPa, respecti vely. Among them, a cubic Pm-3 structure of Rb 0.5Sr0.5B3C3 has the lowest energy at both 0 and 50 GPa. In addition , we performed a CALYPSO structure search for fixed Rb0.5Sr0.5B3C3 composition and found the same Pm-3 as the stable phase, indicating the present computational scheme is reliable. Phon on calculation s were carried out to investigate the dynamical stability of the predicted structures . The results show that RbxSr1-xB3C3 structures with x <=0.5 do not have imaginary frequenc ies in the whole BZ at both 0 and 50 GPa (Fig. S1). Therefore, we select candidate structures with x <=0.5 for further analysis and discussion . Fig. 1. Formation enthalpies (in unit of meV/ XB3C3) and corresponding convex hull of the RbxSr1-xB3C3 system at (a) 0 and (b) 50 GPa. The two end points are ternary SrB 3C3 and RbB 3C3, 6 both in Pm-3n structure. Blue crosses denote structures predicted from the CE method . Solid orange circles correspond to the enthalpies of structures computed with DFT, while the red squares represent the enthalpies from the corresponding CE. Purple triangle s indicate the ground state structure s predicted by the CALYPSO code . Fig. 2. Predicted structures of RbxSr1-xB3C3 for (a) x = 0.125, (b) 0. 167, (c) 0. 2, (d) 0. 25, (e) 0. 3, (f) 0. 333, and (g) 0. 5 at 0 GPa . The Rb, Sr, B and C atoms are shown in purple, pink, green and brown, respectively. Detailed information of predicted structures is summarized in Table s S1 and S2 in the Supplemental Material. All predicted structures , as shown in Fig. 2, can be viewed as host -guest structures . The B and C atoms form interconnected host framework, and the guest metals occupy the centers of the cubic units with different Rb /Sr ratios. The B -C cage is composed of 24 alternat ively arranged B and C atoms, of which eight B 3C3 hexagonal faces and six B2C2 rhombic faces are connected to each other by sharing B -C edges to form a B6C6 sodalite -like cage . Each cage accommodates one of the Rb/Sr atoms in the center. Due to the in trinsic difference s between Rb and Sr atoms, slight distortions occur in the B - C cages for the doped compounds. In these cages, t he B-C bond lengths range between 1.71 and 1.75 A, and C -B-C (B-C-B) angles in between 88deg and 92deg (119deg and 121 deg). They differ slightly from the parent structure SrB 3C3, which has a unique bond length (1.73 A) and angle s (90deg and 120deg) [27] at ambient pressure . 7 Fig. 3. Calculated PDOS for various Rb contents in Rb xSr1-xB3C3 at (a) 0 and (b) 50 GPa. Vertical dashed lines indicate the Fermi level . (c) Variation of total DOS at the Fermi level with composition. The electronic properties of the RbxSr1-xB3C3 at different Rb concentrations are revealed by the projected electronic density of states ( PDOS) , as shown in Fig s. 3(a) and 3(b) . The results show that the electrons near the Fermi level are mainly from B and C atoms (cages) , while the contribution from Rb and Sr atoms are negligible. The evolution of total DOS at the Fermi level with substitution concentration x at 0 and 5 0 GPa are displayed in Fig. 3(c). Obviously , the DOS at the Fermi level decreases first and then increases as the Rb substitution rate increases . It is noteworthy that once the Rb content reaches 0.5, the DOS increases sharply to 1. 67 and 1.36 (in units of 8 States/ eV/XB3C3) at 0 and 50 GPa , respectively , which is much higher than the valu e of ~0.9 in SrB 3C3. This is because Rb doping change s the number of electrons transferred from the metal element s to the B-C host framework, a nd therefore the Fermi level shifted down accordingly [see Figs. 3(a) and 3(b)] . A high value of DOS at the Fermi level is critical to raising the possibility of the formation of Cooper pairs in Rb0.5Sr0.5B3C3, thus enhancing electron -phonon coupling strength and . To investigate potential in Rb0.5Sr0.5B3C3, we performed the EPC calculation and estimated the superconducting Tc through solving the Eliashberg equation [19]. Typical Coulomb pseudopotential parame ters m* from 0.1 to 0.1 3 are used in the solution. Fig. 4 shows t he calculated phonon dispersions, projected phonon density of states (PHDOS) to individual atoms, Eliashberg spectral function a2F(o)/o, and integrated EPC strength l for Rb0.5Sr0.5B3C3 at 0 GPa . Obviously, coupling of electrons to phonons in Rb 0.5Sr0.5B3C3 mainly originates from high frequenc y vibrations in the range of 200-800 cm-1, which contribu te 69% to the total l. Intriguingly, these frequencies are dominated by the vibrations of B -C framework. The resulting Tc reaches a maxim um value of 78 K using m* = 0.1 at 0 GPa, which is comparable to the liquid nitrogen temperature (77 K) . Furthermore , the effects of pressure on the Tc were examined for Rb 0.5Sr0.5B3C3. The results show that Tc has a significant drop to 47 K at 50 GPa, driven by crystal lattice hardening and resulting reduction of l from 2.06 at 0 GPa to 0.92 at 50 GPa (See Fig . S2 in the Supplemental Material for results at 50 GPa for comparison with the data shown in Fi g. 4.) As well, t he superconducting gap obtained from solving Eliashberg equation also shows a notable drop at high pressure [Fig. 5(a) ]. 9 Fig. 4. Calculated phonon dispersion relations, PHDOS, and Eliashberg spectral function of Rb0.5Sr0.5B3C3 at 0 GPa. We have also e valuate d the Tcs for other structures with different Rb content , such as SrB 3C3 (Rb0SrB 3C3) and The predicted Tc of SrB 3C3 is consistent with previously published theoretical res ults [30]. Tcs are estimated to be 46 K and 69 K at 0 GPa for SrB 3C3 and respectively . These two species have lower Tc mainly because of a lower l and NEf compared to Rb 0.5Sr0.5B3C3, as sh own in Fig. 5(b). In addition , we have considered other 24 compounds of X0.5Y0.5B3C3 (X = Li, Na, K, Rb, Cs; Y = Be, Mg, Ca, Sr, Ba) , which are isoelectronic as well as isostructural to Phonon calculations show that only K0.5Ca0.5B3C3 and K0.5Sr0.5B3C3 are dynamically stable at ambient pressure . We have subsequently explored the of these structures , in comparison with Rb 0.5Sr0.5B3C3 [Fig. 5(b) ]. According to the isotope effect, system with a lower mass is likely to have a higher Tc. However, Tcs of both K0.5Ca0.5B3C3 (76 K) and K 0.5Sr0.5B3C3 (75 K) do not exceed the Tc of Rb0.5Sr0.5B3C3 at ambient pressure. A greater contribution of high frequency phonon s to the is offset by t he decrease in EPC strength l, 10 resulting in a comparable Tc in K0.5Ca0.5B3C3 and K 0.5Sr0.5B3C3. Fig. 5. (a) Superconducting gap of Rb0.5Sr0.5B3C3 calculated at 0 and 50 GPa . (b) T op to bottom : electronic DOS at the Fermi level ( NEf), EPC parameter l, logarithmic average phonon frequency olog, and Tc of various clathrate structures at 0 GPa. In order to investigate the thermodynamic stability of Rb0.5Sr0.5B3C3, we systematically examined the formation enthalpies as a function of pressure . Fig. 6 shows t he enthalpy differences in several selected reaction paths in the pressure range of 0-120 GPa. It is clearly seen that the Rb0.5Sr0.5B3C3 phase has higher enthalpy than some combinations of elements , binar y or ternar y compounds, indicating that it is a metastable phase. Encour agingly, t he results show that the Rb0.5Sr0.5B3C3 phase would be stable compared with RbB 6, SrC 2 and C above 60 GPa, suggesting that this material can be synthesized for this special reaction route under extreme pressure . The dynamics stability of this phase suggests that it is quenchable to ambient condition s, following the same thermodynamic route as the parent SrB 3C3. 11 Fig. 6. Calculated enthalpy for Rb0.5Sr0.5B3C3 as a function of pressure . Conclusions In summary, we have systematically explored the phase stability of the complete compositional space of Rb xSr1-xB3C3 at 0 and 50 GPa using first -principles method s in combination with cluster expansion and CALYPSO structure prediction methods . A number of Rb xSr1-xB3C3 structures were discovered , among which the structure s with x<=0.5 are dynamically stable . Electron phonon coupling calculations indicate that phase is a promising conventional superconductor with estimated Tc of 78 K at ambient pressure . In addition, comparable Tc values were also predicted in K0.5Ca0.5B3C3 and K 0.5Sr0.5B3C3 compounds, where Tc reach es 76 and 75 K, respectively. Our findings advance the understanding of superconducting materials consisting of low-Z elements, such as boron and c arbon, under ambient conditions and stimulate s future theoretical and experimental studies in this field. Acknowledgments This work is supported by the National Natural Science Foundation of China (Grants No. 52090024 , 12074138 , 11974135, 11874175, and 11874176), the Fundamental Research Funds for the Central Universities (Jilin University, JLU), the Program for J ilin University Science and Technology Innovative Research Team 12 (JLUSTIRT), and the Strategic Priority Research Program of Chinese Academy of Sciences (Grant No. XDB33000000). We used the computing facilities at the High - Performance Computing Centre of Jilin University. Corresponding author s: Hanyu Liu ) and Yansun 13 References [1] H. K. Onnes, Commun. Phys. Lab. Univ. Leiden, b 120 (1911). [2] J. A. Flores -Livas, L. Boeri, A. Sanna, G. Profeta, R. Arita, and M. Eremets, Phys . Rep. 856, 1 (2020). [3] E. Zurek and T. Bi, J . Chem. Phys. 150, 050901 (2019). [4] T. Bi, N. Zarifi, T. Terpstra, and E. Zurek, in Reference Module in Chemistry, Molecular Sciences and Chemical Engineering (Elsevier, Amsterdam, 2019), -0-12-409547 - 2.11435 -0. [5] Y. Sun, H. Liu, and Y. Ma, Acta Phys. Sin. 70, 017407 (2021) . [6] J. Lv, Y . Sun, H. Liu, and Y . Ma, Matter Radiat . at Extremes 5, 068101 (2020). [7] H. Wang, S. T. John, K. Tanaka, T. Iitaka, and Y . Ma, Proc . Natl. Acad . Sci. U.S.A. 109, 6463 (2012). [8] F. Peng, Y . Sun, C. J. Pickard, R. J. Needs, Q. Wu, and Y . Ma, Phys . Rev. Lett. 119, 107001 (2017). [9] H. Liu, I. I. Naumov, R. Hoffmann, N. Ashcroft, and R. J. Hemley, Proc . Natl. Acad . Sci. U.S.A. 114, 6990 (2017). [10] Y . Li, J. Hao, H. Liu, Y . Li, and Y . Ma, J. Chem. Phys. 140, 174712 (2014). [11] D. Duan , Y . Liu, F. Tian, D . Li, X. Huang, Z. Zhao, H. Yu, B. Liu, W. Tian, and T. Cui , Sci. Rep. 4, 6968 (2014). [12] A. Drozdov, M. Eremets, I. Troyan, V . Ksenofontov, and S. I. Shylin, Nature 525, 73 (2015). [13] A. Drozdov et al. , Nature 569, 528 (2019). [14] M. Somayazulu, M. Ahart, A. K. Mishra, Z. M. Geballe, M. Baldini, Y . Meng, V . V . Struzhkin, and R. J. Hemley, Phys . Rev. Lett. 122, 027001 (2019). [15] I. A. Troyan et al. , Adv . Mater . 33, 2006832 (2021). [16] E. Snider, N. Dasenbrock -Gammon, R. McBride, X. Wang, N. Meyers, K. V . Lawler, E. Zurek, A. Salamat, and R. P . Dias, Phys . Rev. Lett. 126, 117003 (2021). [17] E. Snider, N. Dasenbrock -Gammon, R. McBride, M. Debessai, H. Vindana, K. V encatasamy, K. V . Lawler, A. Salamat, and R. P . Dias, Nature 586, 373 (2020). [18] J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Phys . Rev. 108, 1175 (1957). [19] G. Eliashberg, Sov. Phys. JETP 11, 696 (1960). [20] H. Kawaji, H. -o. Horie, S. Yamanaka, and M. Ishikawa, Phys . Rev. Lett. 74, 1427 (1995). [21] S. Yamanaka, E. Enishi, H. Fukuoka, and M. Yasukawa, Inor g. Chem. 39, 56 (2000). [22] D. Connetable et al. , Phys . Rev. Lett. 91, 247001 (2003). [23] Y . Li, J. Garcia, N. Chen, L. Liu, F. Li, Y . Wei, S. Bi, G. Cao, and Z. Feng, J . Appl . Phys. 113, 203908 (2013). [24] S. Lu, H. Liu, I. I. Naumov, S. Meng, Y . Li, S. T. John, B. Y ang, and R. J. Hemley, Phys . Rev. B 93, 104509 (2016). [25] K. Sano, Y . Masuda, and H. Ito, arXiv:2002.02174 (2020). [26] X. Li, X. Yong, M. Wu, S. Lu, H. Liu, S. Men g, J. S. Tse, and Y . Li, J. Phys. Chem. Lett. 10, 2554 (2019). [27] L. Zhu et al. , Sci. Adv. 6, eaay8361 (2020). [28] X. Cui et al. , J. Phys . Chem . C 124, 14826 (2020). [29] F. Zipoli, M. Bernasconi, and G. Benedek, Phys . Rev. B 74, 205408 (2006). [30] L. Zhu, H. Liu, R. Cohen, R. Hoffmann, and T. A. Strobel, arXiv:1708.03483 (2017). [31] J. N. Wang, X. W. Yan, and M. Gao, Phys . Rev. B 103, 144515 (2021).

2108.11026

Hanyu Liu

Peiyu Zhang, Xue Li, Xin Yang, Hui Wang, Hanyu Liu, Yansun Yao

En route to high Tc via Rb substitution of guest metal atoms in SrB3C3 clathrate

6 figures

10.1103/PhysRevB.105.094503

cond-mat.mtrl-sci

http://creativecommons.org/licenses/by/4.0/

Recently, a host/guest clathrate SrB3C3 with sp3-bonded was synthesized at around 50 GPa. On the basis of electron count, the structure is understood as guest Sr2+ cations intercalated in the Previous calculations suggest that SrB3C3 is a hole conductor with an estimated superconducting critical temperature (Tc) of 42 K at If atoms with similar radius, such as Rb, can substitute Sr2+ in the lattice, the electronic as well as properties of this material will be modified significantly. Here, we perform extensive simulations on the stability and physical properties of Rb-Sr-B3C3 system density functional calculation in combination with and CALYPSO structure prediction method. We predict a Rb0.5Sr0.5B3C3 with a remarkably high Tc of 78 K at which is a significant improvement from the estimated value (42 K) in SrB3C3. The current results suggest that substitution of alkali atom in synthesized clathrate SrB3C3 is a viable route toward high-Tc compounds.

Contents 1 Introduction 1 2 Modeling Needs 2 2.1 Simulations 2 2.2 Realistic Models for Virtual Prototyping of Complete Accelerator Systems 3 3 Software Sustainability 4 3.1 Software Toolkits 5 3.2 Software Ecosystem -- Interoperability and Policies 6 3.3 I/O Standardization 7 4 Advanced Concepts and Future Computing 8 4.1 Advanced Concepts 8 4.2 Evolving architectures in standard computing 9 4.3 Adaptively Tuned Simulations as Online Virtual Diagnostics 10 4.4 Getting ready for the arrival of quantum computing age 10 5 Centers for Accelerator and Beam Physics Modeling 11 6 Conclusion 13 1 accelerator simulation is critical to the design, commissioning, operation, and manymillionsto Accelerator complex topic, and much time and effort has been spent in developing simulation in the field of accelerator physics, simulation code development has often been a haphazard affair. Due to developers retiring or moving on to other projects, numerous have been completely abandoned or are seldom used. Examples include: AGS, TURTLE, UAL to name a few[1, 2]. Oftentimes there is a huge impediment to maintaining these programs due to and lack of documentation. Additionally, many of the programs that are available tend to be rigid. not designed to simulate a priori. Adding a new type of lattice element that a particle can be tracked through is one such in developing a program, but also in terms of researchers leveraging existing technology. - 1 -Indeed, a researcher who wants to simulate something that existing programs are unable to, will, due to time and monetary constraints, generally not be able to fully develop a program from scratch as compared to what could have been done if existing software could be with the possible sharing of simulation data between programs only considered, if at all, as an afterthought. The result is that the ability to benchmark software, crosscheck results, perform regression tests, and do quality assurance testing for consistency is hindered. These types of checks are crucial in validating software accuracy and persistence. 2 Modeling Needs As simulation programs become more complex due to the ever-increasing demands placed upon machine performance, the situation will become worse if not addressed. Such demands include the accelerator and beam physics Grand Challenges that have been identified recently by the community [4, 5] *Increasing beam intensities by orders of the beam phase-space density by orders of magnitude, towards the limit. *Complete and highly accurate start-to-end virtual particle accelerators and accurate multi-objective optimization methods to speed up the design of scientists in form of formal education and easy accessibility to the these Grand Challenges will require a community effort to coordinate and mod- ernize the current set of modeling tools, with capabilities that extend far beyond what the complete accelerator systems. 2.1 Simulations simulationsareimportantinanumberofareas. as the CEBAF[7], LCLS-II[8], ANL, etc. Sufficient shielding is required to properly 2 using electromagnetic RF codes for accelerator structures and beam dynamics codes for particle transport in a beamline to characterize the beam profile before it hits the or the have traditionally been used for detector simulation in HEP experiments. However, since these codes and accelerator simulation codes have all been developed without common data from accelerator to radiation codes. Communication in a standardized format such as openPMD[12], which has been adopted in some accelerator codes, would help ensure efficient and error-free field and particle data transfers[13]. Another issue with an integrated simulation is in matching of the geometry of the vacuum chamber surface. The surface geometry in is generally poorly defined if at all. The most comprehensive simulations define the surface using a finite element mesh generated from a CAD model. In contrast, radiation codes generally employ a faceted representation of the CAD model boundary. A converter for Muchtime and effort would be saved if the surface geometry descriptions were standardized so that a single converter module could be used in multiple describe the complex interplay of the various physics phenomena. One particular example being the emission modeling of a high brightness electron photocathode gun. A photocathode gun provides a high brightness electron source for the downstream accelerator beamline where the beam brightness can only be degraded, not improved. Thus, it is essential to understand validated simulations. While Monte-Carlo photo electron emission simulations have been widely employed in studying photocathode performance for dedicated experiments, its potential in integrated simulations has only been explored recently [14]. For such purposes, a tight integration of the micro-physics models into existing gun simulations can be achieved via the best practices and standardization as discussed below. 2.2 Realistic Models for Virtual Prototyping of Complete Accelerator the development of more realistic simulation models. Here, two examples are given. The first example is from single-particle nonlinear dynamics: Many accelerator codes use idealized models of beamline elements such as quadrupoles, sextupoles, bends, etc. The simplest models omit fringe fields. Better models are based on a fringe field that is a step function with hard-edge models typically set these terms to zero[15]. A better approximation is to assume some smooth analytical form for the fringe field. Though this approach is an improvement over the simpler models, it is still an idealization and there is no reason to expect that all of its nonlinear properties will precisely match those of the physical beamline element. - 3 -The precise prediction of single-particle nonlinear dynamics in accelerators can be widespread use in the accelerator community. The main idea of a surface method is to measure or numerically model the fields of a beamline element on a surface near but within the beam pipe. From there, the fields can be extrapolated inward and are represented by so-called that satisfy Maxwell's equations. In the process, measured or computed errors in the fields at the surface are damped, leading to an accurate representation of the generalized gradients in the beam region. The generalized gradients can then be used to compute realistic transfer maps. These methods are now starting to gain popularity modeling space-charge effects in high intensity beams. In fact, there has been in modeling space-charge effects, and 3D parallel Poisson solvers are now ubiquitous. But despite advances in space-charge modeling, the understanding and modeling of 3D code that can accurately and reliably model 3D CSR effects including transient effects with a performance that makes it useful as a beamline design tool. As accelerators push boundaries with higher brightness beams, the ability to accurately model 3D CSR will become a key issue. This is true for accelerators with very high peak current beams such as those in beam-driven and laser-driven plasma accelerators. It is also true for future X-Ray free electron lasers (FELs) the components and phenomena. The development and application of new, more realistic models would enable virtual prototyping of entire accelerator systems including their for collective instabilities, etc. - before a beamline is constructed and reduce the need for magnet shimming, the use of nonlinear correctors, etc. Ultimately such advanced reduce cost, reduce risk, and improve the performance of future accelerator facilities. 3 Software are several aspects that must be addressed to enable the development of the quality which can be defined as [22]: the capacity of the software to endure. In other words, sustainability means are many aspects to software sustainability, as illustrated in Fig 1. Broadly, these as- pects can be grouped into the intrinsic characteristics of the software itself and the extrinsic - 4 -Readability / / / / Reusability Correctness / bility / Openly ExtrinsicFigure 1. There are a number of aspects that make software sustainable. Broadly, they can be grouped into the intrinsic characteristics of the software itself and the extrinsic environment in which the software is developed and in which the software is developed and used. Software sustainability has been stud- ied academically and there is even a Software Sustainability Institute [23], which promotes the advancement of software in research by cultivating better, more sustainable, research software to enable world-class research. As mentioned above, many software packages developed for simulating accelerators, while showing excellent results for their specific application and their era, are not sustainable in the long run. However, software sustainability is extremely important given the limited resources that the accelerator community has for code development in conjunction with the even more for maintaining codes. To meet future needs, it is imperative that there is a community wide effort to promote sustainable practices. 3.1 Software Toolkits One aspect of developing sustainable software is the creation of software toolkits. A software toolkit is an integrated set of modular routines that are used to develop and maintain possible to develop new programs in less time, with less effort, and with fewer bugs. There are many advantages to organizing software via a toolkit. This safety, since modular code provides a firewall. For example, a buggy into the toolkit will not affect programs that do not use therefore get greater scrutiny. *By having modules that can read and write lattice information and data, the sharing of information between programs is made easier. A well-known example is Geant4 [9], which is a toolkit for the simulation of the passage of particles through matter. Geant4 has helped many researchers solve problems that would not be possible if a given researcher had tried to write the simulation code from scratch. - 5 described in more detail in other sections, an important facet to maximizing be to have accelerator community-wide policies and standards [13] that would then enable the meshing of the toolkits into an ecosystem [25], making it easier to develop the that are needed for the next generation of machines programs that can do the common tasks that accelerator physicists routinely do such as Twiss and orbit calculations, nonlinear optimization, lattice design, etc. That is, a researcher to have to do programming when the researcher only wants to do a common task like calculating the closed for specific phenomena. Consider, for example, a Poisson solver library for effects in high intensity/high brightness beams [29]. Subject to different a grid. For an open boundary condition, a FFT-based Green's function method could be used. For a closed boundary condition with regular shape, a finite difference spectral method could be used. multiple times in separate, incompatible and uncoordinated codes, the community would benefit greatly from consolidation in the development of a few selected toolkits for reuse simulation toolkits already exist: Accelerator Toolbox [30], Bmad [31], Cosy Infinity [32], Merlin++ [33], Warp [34], IMPACT[35] and FPP/PTC [36]. Moving forward, one option would be to further develop existing toolkits. Another approach that would sustainability would be to develop a new toolbox based on modern software practices. To not reinvent the wheel, this new toolbox should reuse existing algorithms and code [25]. 3.2 Software Ecosystem -- Interoperability and of accelerator facilities can extend far beyond the accelerator itself. For example, in an XFEL, a comprehensive simulation will go from the creation of the X-ray pulses produced by electrons in an undulator, all the way to the simulation of the X-ray experiment. This includes X-ray transport, photon-sample interaction, signal transport, detector response, and data no single software package or toolkit would cover all of these domains. A software ecosystem is the set of libraries (toolkits) and applications that are independently but with a common set of rules and standards [13, 25]. An the implementing of frameworks for start-to-end workflows that span over SIMEX [39]. Such integration platforms aid in the rapid adoption for new design and analysis needs, using complex tools in automated workflows such as needed for AI/ML and 6 -with a compatible software ecosystem would be the further adoption of such standards as well as abstraction of low-level data layouts, which would foster innovative combinations of toolkits with bespoke software into new applications for tightly coupled models. A large community of devel- opers can focus on domain-specific needs and help improve software capability and projects, while reducing maintenance and development time due to common practices and fundamentals developed by specialists. Fundamental building blocks would also address perfor- mance and portability needs in low-level libraries. This could include support, for example, of, CPUs, GPUs, multi-node communication patterns, and multi-platform compiling. bugs are spotted earlier due to re-use. A progressive path for both adoption and of such components is possible, starting from existing to exchange data among the simulation tools in a more integrated, finer-grained manner that may be at the level of each simulation time step[14, 40, 41]. This may happen when multiple physics effects are to be simultaneously simulated or when coupling multiple solvers, where data exchange can be either in single direction or bi-directional. Such an in-memory or cross-node data transfer can be made to respect physics constraints (conservation, divergence free and is also much more efficient than file I/O based This need can be best served with the common low-level data layouts and interfaces as mentioned above and/or order to advance accelerator modeling further in the direction of a compatible, some general standardization needs, including best practices in software been identified [25]. These are based on community policies that have been established over years by teams of specialists in scientific software development and computing. For instance, the accelerator community could be leveraging on the Interoperable Design of (xSDK) and international collaboration and adoption. For vertical software Cabana [50]) and mesh/particle remapping (e.g., Portage [51]) libraries could be combined as a foundation for the accelerator and beam physics components. 3.3 I/O need to be reproducible and independently (such as new methods, start-to-end modeling, common analysis needs and standards for data and workflows can help increase productivity and Recently, activities such as those supported by the Consortium for Advanced Modeling of - 7 -Particle Accelerators (CAMPA) [53], helped coordinate the standardization of data exchange and simulation control, with the aim of connecting large existing frameworks and enabling Such standardization efforts can also provide the basis for an integration into scientific data portals to curate and re-use modeling data[54-56]. Besides improving reproducibility of simulations and preserving value of collected data, such efforts can also aid meta-studies and tests for new theories. With respect to accelerator simulation data, historically, there exists a variety of the domain-specific data consumed and generated by individual tools. Yet new, existing paradigms, such as POSIX file-I/O, are overtaken by modern approaches such of these low-level file formats for accelerator modeling is needed to utilize the progress in modern storage and data transport technology and overcome bottlenecks arising from of high-fidelity data as well as manual data analysis and curation efforts. The Open Standard for Particle-Mesh Data Files (openPMD)[12] successfully defining compatible meta-data in a file-format agnostic organization for accelerator and beam data is possible, while using scalable, modern file formats from computer science [57-59]. OpenPMD is organized around a written, versioned text document that is supported by of to openPMD, the Standard Input Format for Particle-In-Cell Codes Programming Interface (API), which is an approach similar to successful scripting language suitable for rapid simulation design and backed by a vast, ecosystem. 4 Advanced Concepts and Future Computing 4.1 Advanced Accelerator Concepts (AAC) offer accelerating gradients that go beyond the simulation tools) and can be significantly more challenging computationally, in particular for - 8 -those involving plasmas. The description of the specific needs and challenges of these tools is given in another paper of this issue [62]. Aside from the different physics, numerical methods needs, other challenges and solutions are essentially the same as for with regard to, e.g., software sustainability, standardization, validation, integrated workflows and frameworks, toolkits, ecosystems, centers, evolving and (including quantum computing) that are described in this paper. In terms of numerical approaches, advanced concepts include the need to investigate novel numerical methods and approaches beyond Particle-In-Cell (PIC) methods [35, 63-65] that have shown promise for future incorporation into simulation software as modules. These approaches could potentially tackle different types of advanced simulation problems. Examples include sym- plectic multiparticle space-charge simulations [66, 67], Fast Multipole Method (FMM)-based ap- proaches [68-72], boundary integral solvers [73-75] and hybrid solvers such as [76, 77]. 4.2 Evolving architectures in standard in existing and planned DOE High Performance Computing (HPC) facilities processors offer an attractive boost in computing performance with respect to any of the common denominators such as hardware cost, electricity cost, or wall clock time to accomplish engines acting on localized data at the expense of general purpose computing problems whose algorithms can be cast in this paradigm benefit greatly from this style of computing many accelerator simulation problems are within this class. Propagating many independent particles through beamline components is seemingly custom made for this kind of computing. The issue facing the field then is that programming these devices requires Typically, data for computations should be transferred to co-processor memory and arranged carefully for efficient parallel processing. This transfer is usually slow and should current algorithms on co-processor hardware; new algorithms will need to be on current general purpose processors will continue to be an important part of the landscape. Software should be built to run on either standard hardware or new architectures, of which there are several. Although the most well-established solutions for GPU computing are provided by NVidia, AMD and Intel are also building HPC systems. Some accelerator simulation codes have been implemented in CUDA, the NVidia specific GPU programming language which is specific to NVidia GPUs. Ideally, the accelerator would avoid dependence on a single co-processor supplier. Fortunately, there are several major software efforts to build parallel computing frameworks that can support execution on either CPU or GPU processors with a single high-level code base. No particular framework is clearly superior to the others, and each is targeted for use in a 9 -their algorithms for use in the near and medium future on what will be the dominant Adaptively Tuned Simulations as Online Virtual Diagnostics Some of the most detailed accelerator diagnostics are X-band transverse deflecting cavities. How- ever, deflecting cavities are limited to a resolution of 1 m or 3.3 fs/pixel, and destroy the beam during the measurement process [78]. This measurement floor is of concern since beams at ad- vanced plasma wakefield accelerators and free electron laser (FEL) facilities are starting to exceed those limits. The Facility for Advanced Accelerator Experimental Tests (FACET-II) will provide bunch lengths as low as (1 m or3 fs) at 12 GeV [79], attosecond two-color X-ray pulses with 30 m bunch lengths at 250 GeV [81]. As bunch lengths decrease beyond the resolution of existing diagnostics, simulations will be relied on as virtual diagnostics. approaches are common in other codes with millions of macro-particles because once accelerators are built they do introduced, the electromagnetic fields of components do not perfectly match simulated by disturbances not accounted for. Closely matching simulations to accelerators requires adaptive feedback. Diagnostics can be compared to simulation-based predictions and simulation parameters can be tuned in real-time to achieve a match between measurements and simulation outputs. Once a match is of physics constraints within the simulation, it is likely that other beam properties are uniquely matched. Recently a LiTrack model of FACET was adaptively tuned online to match the simulated beam's energy spread spectrum ^E1xtoto its measured E1xto, minimizing to track the time-varying longitudinal phase space 1zEo of the electron beam [82, 83]. Efforts are also underway to utilize ML tools to map diagnostics back to input beam distributions to be used as the initial conditions of accelerator models of GPUs and field programmable gate arrays it may be possible to use such adaptive models as virtual diagnostics in real-time shot-to-shot for high repetition rate machines. 4.4 Getting ready for the arrival of quantum computing age Studies of collective effects are essential for modern accelerators with high intensity beams. The start-to-end simulation of an accelerator using real beams with billions or more particles and expensive even with state-of-the-art exascale machines. The development of quantum computers provides new opportunities to enhance the particle accelerator abilities. A quantum computer is a device that utilizes the special properties of quantum mechanics to perform computation, which can potentially provide an exponential improvement in efficiency - 10 -for some classes of simulations. The three properties of quantum superposition, interference, and entanglement of a quantum state make quantum computing different from classical computing. Due to quantum superposition, the information stored in a quantum system scales the number of qubits as opposed to the linear scaling with the number of bits in a has been demonstrated on some specific problems [85, 86]. Quantum computing is currently available to the public through cloud services provided by some commercial accelerator modeling relates to solving linear systems [91-95] as well as ordinary (ODEs) and partial differential equations (PDEs) [96-100]. Solving Poisson's the beam-beam interaction, and coherent synchrotron radiation - with quantum platforms, are available for almost all mainstream programming languages, e.g. C/C++, Python, Java, Matlab, etc. There also exist some languages specifically designed for for beam and accelerator physics. Most previous work on quantum algorithms focused on the realization of an algorithm with quantum circuits, but applying the algorithm to solve a practical scientific problem is seldom discussed [105]. Clearly there is a gap between the development of the algorithms and To make a problem suitable for quantum simulations, it has to be described by unitary operators so that the system, which may be classical, is mathematically equivalent to a quantum system. Additionally, in quantum computing, all variables are stored in quantum states. Preparing the initial states and reading out the results accurately can be time intensive. The traditional way of using a large number of particles in simulation and producing all of as output would be classical computer but will probably work together with one. It is thus probably better for now to focus on how to make new quantum simulation tools that collaborate with the existing pool of accelerator modeling programs. Ideally a protocol will be invented through which the quantum packages could be called by the classical programs. Also needed are innovative analyzing tools to process the simulation results, which may happen before the reading-out. To achieve these quantum computing are required. 5 Centers for Accelerator and Beam Physics Modeling It quickly becomes clear that, in order to achieve what is described and proposed in the 11 -Centers for Accelerator and Beam Physics Modeling [106]. Other areas of computer science dispersed computing efforts of the various departments (e.g. MIT's Schwarzman College of Computing [107]), and new centers for Quantum Computing [108, 109] have been built. Energy (DOE) has founded SciDAC [111] to accelerate progress in scientific computing Biological and Environmental Research, Fusion Energy Sciences, High-Energy Physics, and Nuclear industrial partners. They would bring together domain scientists (computational accelerator and beam physicists), applied mathematicians, computer scientists and software engineers centers mentioned above are already supporting accelerator modeling efforts on which the Centers for Accelerator and Beam Physics Modeling could build. Depending on the overall size, the centers could enable part or all of the dedicated, specialized teams [6]. *Collect libraries for field solvers, particle trackers, and other input scripts, output data, lattice description and start-to-end workflows [13]. *Provide compatibility layers to use the same libraries and modules in a number of program- ming and maintenance of End-to-end Virtual Accelerators (EVA) [5]. *User support, high-quality and detailed documentations, online tutorials, and standardized, user interfaces for preparation and analysis of tools for ensemble simulations for optimization with builtin AI/ML regression (e.g., high-order solvers, symplectic multiparticle tracking, Fast Multipole mesh a space to meet (physically or virtually) for the integration of developments into larger codes, such as PhD projects from external groups, organizing devel- opment hackathons, and onboarding. - 12 -*Developing and organizing workshops for developers and users of codes alike. and international (travel/hosting (e.g. tutorials, lectures) on of novel use of machine learning for accelerator modeling, and, further in the future, of quantum computing [112]. Multiple Centers can be organized through a Consortium (e.g., CAMPA [53]). Except for special restrictions such as export control, it would be desirable for the software developed by the Center to be open source, enabling crosschecking, testing and contribution by the community at large, beyond the participants to the Center(s) [113]. 6 future ones: Computer simulations play an indispensable role in all accelerator areas. Currently, there are many simulation programs used for accelerator physics. There is, Moreover there is very little effort currently being made to make these codes generally available to the accelerator community ... -- HEPAP report, millions of dollars were spent, in part due to a flawed simulation. The decision to increase the vacuum chamber bore diameter from 4 cm to 5 cm was based in part upon a faulty simulation using the code SSCTRK[115]. This code was newly developed for the SSC[116] and had not been thoroughly vetted at the time the decision was made. After refinements to the Monte Carlo error model, and after upgrading the SSCTRK code, it was realized that the original 4 cm bore had been adequate. The unnecessary enlargement to 5 cm lead to cost increases along with some turmoil caused by having to redesign vacuum chambers, magnets and other machine this was was a major factor in the demise of the SSC. Communities in the computational sciences face similar challenges, and physics modeling groups would benefit from incorporating research results, best practices, participate actively bodies, and anticipate trends in the broader computer science and community. In sustainable workflows, one wants to avoid heroic efforts relying on few individuals, and instead provide an environment that is inclusive and thrives with various educational backgrounds. Physics groups that integrate early with external applied mathematics and computational physics efforts can share modular cross-domain visions and avoid missing out on solutions developed in related 13 of computational work. Analysis routines, source codes, inputs, and simulation data are often not openly archived in sufficient detail, which hinders reproducibility and adoption of by other groups. This situation poses a significant risk and calls for an advancement of modeling practices that can adequately address the needs of decade-long basic science projects. One part of the problem is that, traditionally, performance metrics of scientific success aim solely on publication numbers. The extra time and effort to make code sustainable is, by this criterion, unproductive. The end result, however, is extensive waste. It is imperative that pay for itself many times Accelerator physics codes. . [2] Alexander Wu Chao, Karl Hubert Mess, et al. Handbook of accelerator physics and engineering . World scientific, 2013. [3] Carole Goble. Better software, better research. IEEE Internet Computing , 18, 2014. [4] S. Nagaitsev, Z. Huang, Power J., J.-L. Vay, P. Piot, L. Spentzouris, J. Rosenzweig, Y Cai, S. Cousineau, M. Conde, M. Hogan, A. Valishev, M. Minty, T. Zolkin, X. Huang, V. Shiltsev, J. Seeman, J. Byrd, and J.R. Patterson. Accelerator and Beam Physics: Grand Challenges and Research Opportunities. Snowmass21 LOI , 2020. URL . [5] Jean-Luc Vay, David Sagan, Axel Huebl, Maxence Thevenet, Remi Lehe, Cho-Kuen Ng, Henri Vincenti, Michael Bussmann, Alexander Debus, Richard Pausch, and Ji Qiang. End-to-End (EVA). Snowmass21 LOI , 2020. URL . [6] Remi Lehe, Axel Huebl, Jean-Luc Vay, Alexander Friedman, Maxence Thevenet, Chad Mitchell, David Bruhwiler, David Grote, Benjamin Cowan, Henri Vincenti, Adi Hanuka, Brigitte Cros, Samuel Yoffe, Rene Widera, Michael Bussmann, and Auralee Linscott Edelen. Embracing modern software tools and user-friendly practices, when distributing scientific codes. Snowmass21 LOI , 2020. URL . [7] C. Hovater, T. Allison, G. Biallas, R. Bachimanchi, E. Daly, M. Drury, A. Freyberger, R. L. Geng, G. Lahti, R. Legg, C. Mounts, R. Nelson, T. Plawski, and T. Powers. Operation of the cebaf 100 mv cryomodules. In 2016 , number28 in Conference, pages 65-67,Geneva, Switzerland, 5 2017. JACoW. ISBN . - 14 -[8] J. R. Lewandowski, R. C. Field, A. S. Fisher, H.-D. Nuhn, and J. J. Welch. Rf gun dark with a transverse deflecting cavity at lcls. In Proc. 37th Int. Free Electron Laser Conf. (FEL'15), pages 583-586. JACoW Publishing, 8 2015. J. Allison, K. Amako, J. Apostolakis, P. Arce, M. Asai, T. Aso, E. Bagli, A. Bagulya, S. Banerjee, G. Barrand, B.R. Beck, A.G. Bogdanov, D. Brandt, J.M.C. Brown, H. Burkhardt, Ph. Canal, D. Cano-Ott, S. Chauvie, K. Cho, G.A.P. Cirrone, G. Cooperman, M.A. Cortes-Giraldo, G. Cosmo, G. Cuttone, G. Depaola, L. Desorgher, X. Dong, A. Dotti, V.D. Elvira, G. Folger, Z. P. Kaitaniemi, N. Karakatsanis, M. Karamitros, M. Kelsey, A. Kimura, T. Koi, H. Kurashige, A. Lechner, S.B. Lee, F. Longo, M. Maire, D. Mancusi, A. Mantero, E. Mendoza, B. Morgan, K. Murakami, T. Nikitina, L. Pandola, P. Paprocki, J. Perl, I. Petrovic, M.G. Pia, W. Pokorski, J.M. Quesada, M. Raine, M.A. Reis, A. Ribon, A. Ristic Fira, F. Romano, G. Russo, G. Santin, T. Sasaki, D. Sawkey, J.I. Shin, I.I. Strakovsky, A. Taborda, S. Tanaka, B. Tome, T. Toshito, H.N. Tran, P.R. Truscott, L. Urban, V. Uzhinsky, J.M. Verbeke, M. Verderi, B.L. Wendt, H. Wenzel, D.H. Wright, D.M. Wright, T. Yamashita, J. Yarba, and H. Yoshida. Recent developments in geant4. Equipment , 835:186 - 225, 2016. [10] Alfredo Ferrari, Paola R. Sala, Alberto Fasso, and Johannes Ranft. FLUKA: A code (Program version 2005). 10 2005. doi: Nikolai V. Mokhov and Catherine C. James. The MARS Code System User's Guide 2 2017. doi: Axel Huebl, Remi Lehe, Jean-Luc Vay, David P. Grote, Ivo Sbalzarini, Stephan Kuschel, David Sagan, Frederic Perez, Fabian Koller, and Michael Bussmann. openpmd: A meta data standard for particle and mesh based data. 2015. doi: . [13] A. Huebl, J.-L. Vay, R. Lehe, M. Thevenet, C. Mayes, D. Sagan, Y.-D. Tsai, J. C. E, F. Tsung, H. Vincenti, A. Ferran Pousa, N. M. Cook, S. J. Gessner, F. Poeschel, M. Bussmann, D. P. Grote, N. A. Murphy, R. Schmitz, C. H. Yoon, D. L. Bruhwiler, K. Cranmer, S. R. Yoffe, B. Cros, A. L. Edelen, and G. Stark. data standards and start-to-end workflows. Snowmass21 LOI, 2020. URL . [14] C.-K. Huang, T.J.T. Kwan, Vitaly Pavlenko, Cho-Kuen Ng, and Erdong Wang. modeling of high brightness beam injectors. Snowmass21 LOI , 2020. . [15] R. Ryne and A. Dragt. Numerical computation of transfer maps using lie algebraic methods. In Proceedings of the 12th IEEE Particle Accelerator Conference. , pages 1081-3, 1996. [16] Robert Ryne, Dan Abell, Martin Berz, David Bruhwiler, Alex Dragt, Kyoko Makino, Chad Ng, Ji Qiang, Jean-Luc Vay, Marco Venturini, and Peter Walstrom. Surface methods Snowmass21LOI ,2020. URL . - 15 -[17] M. Borland, R.R. Lindberg, and R. Soliday. Tools for use of generalized gradient expansions in accelerator simulations. In Proc. of IPAC , page MOPAB059, 2021. [18] Lajos Bojtar. Frequency analysis and dynamic aperture studies in a low energy antiproton ring with realistic 3d magnetic fields. Phys. Rev. Accel. Beams , 23:104002, 10 2020. . [19] Shashikant Manikonda and Martin Berz. Multipole expansion solution of the laplace equation using surface data. Nuclear Instruments and Methods in Physics Research Section A: Detectors and Associated Equipment , 558(1):175-183, 2006. ISSN 0168-9002. . Proceedings of the 8th International Computational Accelerator Physics Conference. [20] Shashikant Manikonda and Martin Berz. An accurate high-order method to solve the value problem for the 3d laplace equation. International Journal of Pure and , 23:365-378, 01 2005. [21] E.L. Saldin, E.A. Schneidmiller, and M.V. Yurkov. Analytical treatment of the radiative interaction of electrons in a bunch passing a bending magnet. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment , 407(1): 112-115, 1998. ISSN 0168-9002. doi: . [22] Daniel S. Katz. Scientific software challenges and community responses, 2015. . Talk given at on 7 December 2015, discussing 12 scientific software challenges and how the scientific software community is responding to them. [23] URL . [24] David Sagan, Axel Huebl, Jean-Luc Vay, David Bruhwiler, Robert Ryne, Cho Ng, and Remi Lehe. Beam Dynamics Toolkit. Snowmass21 LOI , 2020. URL . [25] Jean-Luc Vay, Axel Huebl, David Sagan, David Bruhwiler, Remi Lehe, Cho-Kuen Ng, Ao Liu, Ji Qiang, Robert Ryne, Eric Stern, Alex Friedman, Maxence Thevenet, Henri Vincenti, Spencer Gessner, Benjamin Cowan, Andreas Adelmann, Michael Bussmann, and Sergei Bastrakov. A modular community ecosystem for multiphysics particle accelerator modeling and LOI , 2020. URL . [26] D. Sagan and J.C. Smith. The tao accelerator simulation program. In Proceedings of the 2005 Particle Accelerator Conference , pages 4159-4161, 2005. doi: Christoph Iselin. The mad program. In Winfried Busse and Roman Zelazny, editors, Computing in Accelerator Design and Operation , pages 146-151, Berlin, Heidelberg, 1984. Springer ISBN M. Borland. elegant: A Flexible SDDS-Compliant Code for Accelerator Simulation. In Computational Accelerator Physics Conference (ICAP 2000) , 2000. 16 -[29] Ji Qiang, He Zhang, Jean-Luc Vay, Axel Huebl, David Grote, Kiran Sonnad, and Chad Mitchell. A Parallel Poisson Solver Library for Accelerator Modeling Applications. Snowmass21 LOI , 2020. URL . [30] A. Terebilo. Accelerator modeling with matlab accelerator toolbox. In PACS2001. Proceedings D. Sagan. Bmad: A relativistic charged particle simulation library. Nuclear Instruments and Methods in Physics Research Section A , 558(1):356 - 359, 2006. [32] Martin Berz. New features in cosy infinity. AIP Conference Proceedings , 297(1):267-278, 1993. [33] J. Molson, H. Owen, A. M. Toader, and R. J. Barlow. Advances with merlin - a beam tracking code. IPAC 2010 - 1st International Particle Accelerator Conference , pages 1853-1855, 2010. [34] A. Fiedman, D. P. Grote, D. A. Callahan, B. A. Langdon, and I. Haber. 3d particle simulation of beams using the warp code. Part. Accel. , 37-38:131-139, 1992. [35] Ji Qiang, Robert D. Ryne, Salman Habib, and Viktor Decyk. An object-oriented code for beam dynamics simulation in linear accelerators. Journal of 163(2):434 - 451, 2000. ISSN 0021-9991. doi: E. Forest, Y. Nogiwa, and F. Schmidt. The fpp and ptc libraries. Proceedings of ICAP 2006 , pages 17-21, 2006. [37] C. E. Mayes, P. H. Fuoss, J. R. Garrahan, H. Slepicka, A. Halavanau, J. Krzywinski, A. L. Edelen, W. Lou F. Ji, N. R. Neveu, A. Huebl, R. Lehe, L. Gupta, C. M. Gulliford, D. C. Sagan, J. C. E, and C. Fortmann-Grote. Lightsource unified modeling environment (lume), a start-to-end In Proc. of IPAC , page THPAB217, 2021. [38] The Photon and Neutron Open Science Cloud (PaNOSC). URL . [39] May 2020. URL . [40] D.A. Dimitrov, D.L. Bruhwiler, J.R. Cary, P. Messmer, P. Stoltz, K.L. Jensen, D.W. Feldman, and P.G. O'Shea. Development of advanced models for 3d photocathode pic simulations. In Proceedings of the 2005 Particle Accelerator Conference , pages 2583-2585, 2005. S. Nagaitsev, Z. Huang, J. Power, J. Vay, P. Piot, L. Spentzouris, J. Rosenzweig, Y. Cai, S. Cousineau, M. Conde, M. Hogan, A. Valishev, M. Minty, T. Zolkin, X. Huang, V. Shiltsev, J. Seeman, J. Byrd, Y. Hao, B. Dunham, B. Carlsten, A. Seryi, and R. Patterson. Accelerator and beam physics research goals and opportunities. arXiv: Accelerator Physics , 2021. [42] F. Zhang, Tong Jin, Qian Sun, Melissa Romanus, Hoang Bui, S. Klasky, and M. staging and data-centric task placement for coupled scientific simulation and Computation: Practice and Experience , 29, 2017. [43] Utkarsh Ayachit, Andrew Bauer, Earl P. N. Duque, Greg Eisenhauer, Nicola Ferrier, Junmin Gu, Kenneth E. Jansen, Burlen Loring, Zarija Lukic, Suresh Menon, Dmitriy Morozov, Patrick Matthew Wolf, K. John Wu, and E. Wes Bethel. Performance analysis, and applications of extreme-scale in situ infrastructures. In SC '16: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , pages 921-932, 2016. doi: 17 -[44] Matthew Larsen, James Ahrens, Utkarsh Ayachit, Eric Brugger, Hank Childs, Berk Geveci, and Cyrus Harrison. The alpine in situ infrastructure: Ascending from the ashes of strawman. In Proceedings of the In Situ Infrastructures on Enabling Extreme-Scale Analysis and Visualization , ISAV'17, page 42-46, New York, NY, USA, 2017. Association for Computing Machinery. doi: . [45] IDEAS: Interoperable Design of Extreme-scale Application Software. . [46] xSDK: Extreme-scale Scientific Software Development Kit. URL http://xsdk.info . [47] The HYPRE Team. HYPRE. URL . [48] Stephen Hudson, Jeffrey Larson, Stefan M. Wild, David Bindel, and John-Luke users manual. Technical Report Revision 0.7.0, Argonne National Laboratory, 2020. URL . [49] The AMReX Team. AMReX. URL . [50] The Copa-Cabana Team. Copa-Cabana. URL . [51] The Portage Team. Portage. URL . [52] REANA Reproducible research data analysis platform . URL . [53] CAMPA: Consortium for Advanced Modeling of Particle Accelerators. . [54] CERN Data Portal. URL . [55] FAIR Principles. URL . [56] RECAST framework reinterpretation of an ATLAS Dark Matter Search constraining a model of a dark Higgs boson decaying to two b-quarks. Technical Report CERN, Geneva, 8 2019. URL . [57] William F. Godoy, Norbert Podhorszki, Ruonan Wang, Chuck Atkins, Greg Eisenhauer, Junmin Gu, Philip Davis, Jong Choi, Kai Germaschewski, Kevin Huck, Axel Huebl, Mark Kim, James Kress, Tahsin Kurc, Qing Liu, Jeremy Logan, Kshitij Mehta, George Ostrouchov, Manish Parashar, Franz Poeschel, David Pugmire, Eric Suchyta, Keichi Takahashi, Nick Thompson, Seiji Tsutsumi, Lipeng Wan, Matthew Wolf, Kesheng Wu, and Scott Klasky. Adios 2: The adaptable input output system. a framework for high-performance data management. SoftwareX , 12:100561, 2020. ISSN 2352-7110. doi: . [58] The HDF Group. Hierarchical data format version 5. URL . [59] Axel Huebl, Rene Widera, Felix Schmitt, Alexander Matthes, Norbert Podhorszki, Jong Youl Choi, Scott Klasky, and Michael Bussmann. On the scalability of data reduction techniques in current and upcoming hpc systems from an application perspective. In Julian M. Kunkel, Rio Yokota, Michela Taufer, and John Shalf, editors, High Performance Computing , pages 15-29, Cham, 2017. Publishing. ISBN Curated catalogue of projects supporting openPMD. . - 18 -[61] PICMI: Standard input format for Particle-In-Cell codes. . [62] J.-L. Vay, A. Huebl, N.M. Cook, R. J. England, U. Niedermayer, P. Piot, and Winklehner D. Modeling of Advanced Accelerator Concepts. This Issue , 2021. [63] Alex Friedman, David P. Grote, and Irving Haber. particle simulation of heavy-ion fusion beams. Phys. Fluids B , 4:2203-2210, 1992. doi: R. Hockney and J. Eastwood. Computer simulation using particles. 1988. [65] J. Amundson, P. Spentzouris, J. Qiang, and R. Ryne. Synergia: An accelerator modeling tool with 3-d space charge. Journal of Computational Physics , 211(1):229 - 248, 2006. ISSN 0021-9991. doi: JiQiang. Phys. Rev. Accel. Beams , 20:014203, 1 2017. doi: Ji Qiang. Symplectic particle-in-cell model for space-charge beam dynamics simulation. Phys. Rev. Accel. Beams , 21:054201, 5 2018. doi: L Greengard and V Rokhlin. A fast algorithm for particle simulations. Journal of 73(2):325-348, 1987. ISSN 0021-9991. . [69] Lexing Ying, George Biros, Denis Zorin, and Harper Langston. A new parallel multipole method. In Proceedings of the 2003 ACM/IEEE Conference on Supercomputing , URL . [70] Ilya Lashuk, Aparna Harper Langston, Tuan-Anh Nguyen, Rahul Sampath, Aashay Shringarpure, Richard Vuduc, Lexing Ying, Denis Zorin, and George Biros. A adaptive fast multipole method on heterogeneous architectures. Commun. ACM , 55(5): 101-109, May 2012. ISSN 0001-0782. doi: . [71] He Zhang and Martin Berz. The fast multipole method in the differential algebra Instruments and Methods in Physics Research Section A: Accelerators, and Associated Equipment , 645(1):338 - 344, 2011. ISSN 0168-9002. The Eighth International Conference on Charged Particle Optics. [72] M. Harper Langston, Leslie Greengard, and D. Zorin. A free-space adaptive FMM-based PDE solver in three dimensions. Communications in Applied Mathematics and Computational Science , 6 (1):79-122, 2011. doi: George Biros, Lexing Ying, and Denis Zorin. A fast solver for the stokes equations with in complex geometries. Journal of Computational Physics , 193(1):317 - 348, 2004. ISSN 0021-9991. doi: Andreas Klockner, Alexander Barnett, Leslie Greengard, and Michael O'Neil. Quadrature by expansion: A new method for the evaluation of layer potentials. Journal of Computational Physics , 252:332-349, 2013. ISSN 0021-9991. doi: . - 19 -[75] Journal of Computational Physics , 442:110511, 2021. ISSN 0021-9991. . [76] Felipe Vico, Leslie Greengard, and Miguel Ferrando. Fast convolution with free-space J. Comput. Phys. , 323(C):191-203, October 2016. ISSN 0021-9991. URL . [77] Junyi Zou, Eugenia Kim, and Antoine J. Cerfon. Fft-based free space poisson solvers: should replace 2021. [78] C Behrens, F-J Decker, Y Ding, VA Dolgashev, J Frisch, Z Huang, P Krejcik, H Loos, A Lutman, TJ Maxwell, et al. Few-femtosecond time-resolved measurements of x-ray free-electron lasers. Nature communications , 5(1):1-7, 2014. doi: C Joshi, Erik Adli, W An, CE Clayton, Sebastian Corde, S Gessner, MJ Hogan, M Litos, W Lu, KA Marsh, et al. Plasma wakefield acceleration experiments at facet ii. Plasma Physics and Controlled Fusion , 60(3):034001, 2018. doi: Alexander Malyzhenkov, Yunieski P Arbelo, Paolo Craievich, Philipp Dijkstal, Eugenio Ferrari, Sven Reiche, Thomas Schietinger, Pavle Juranic, and Eduard Prat. Single-and two-color attosecond hard x-ray free-electron laser pulses with nonlinear compression. Physical Review Research , 2(4): 042018, 2020. doi: Hiroaki Aihara, Jonathan Bagger, Philip Bambade, Barry Barish, Ties Behnke, Alain Berggren, James Brau, Martin Breidenbach, Ivanka et al. linear collider. a global project. arXiv preprint arXiv:1901.09829 , 2019. . [82] KLF Bane and P Emma. Litrack: a fast longitudinal phase space tracking code with graphical user interface. In Proceedings of the 2005 Particle Accelerator Conference , pages 4266-4268. IEEE, 2005. doi: Alexander Scheinker and Spencer Gessner. Adaptive method for electron bunch profile Review Special and Beams , 18(10):102801, 2015. [84] Alexander Scheinker, Frederick Cropp, Sergio Paiagua, and Daniele Filippetto. Adaptive deep learning for time-varying systems with hidden parameters: Predicting changing input of compact particle accelerators. arXiv preprint arXiv:2102.10510 , 2021. . [85] Frank Arute, Kunal Arya, Ryan Babbush, Dave Bacon, Joseph C Bardin, Rami Barends, Rupak Biswas, Sergio Boixo, Fernando GSL Brandao, David A Buell, et al. Quantum supremacy using a programmable superconducting processor. Nature, 2019. [86] Han-Sen Zhong, Hui Wang, Yu-Hao Deng, Ming-Cheng Chen, Li-Chao Peng, Yi-Han Luo, Jian Qin, Dian Wu, Xing Ding, Yi Hu, et al. Quantum computational advantage using photons. 2020. [87] Ibm quantum , 2020. [88] , 2020. [89] Amazon quantum solutions lab. , 2020. [90] Azure quantum. , 2020. - 20 -[91] Aram W Harrow, Avinatan Hassidim, and Seth Lloyd. Quantum algorithm for linear systems of equations. Physical review letters , 103(15):150502, 2009. [92] B David Clader, Bryan C Jacobs, and Chad R Sprouse. Preconditioned quantum linear Physical review letters , 110(25):250504, 2013. [93] Andrew M Childs, Robin Kothari, and Rolando D Somma. Quantum algorithm for systems of linear equations with exponentially improved dependence on precision. SIAM Journal on Computing , 2017. [94] quantum linear solver: A hybrid algorithm for linear systems. arXiv , 2019. [95] Yonghae Lee, Jaewoo Joo, and Soojoon Lee. Hybrid quantum linear equation algorithm and its experimental test on ibm quantum experience. Scientific reports , 9(1):4778, 2019. [96] Sarah K Leyton and Tobias J Osborne. A quantum algorithm to solve nonlinear . arXiv preprint arXiv:0812.4423 , 2008. [97] Dominic W Berry. High-order quantum algorithm for solving linear differential equations. Journal of Physics A: Mathematical and Theoretical , 47(10):105301, 2014. [98] Juan Miguel Arrazola, Timjan Kalajdzievski, Christian Weedbrook, and Seth Lloyd. for nonhomogeneous linear partial differential equations. Physical Review A , 100(3): 032306, 2019. [99] Andrew M Childs, Jin-Peng Liu, and Aaron Ostrander. High-precision quantum algorithms for partial differential equations. arXiv preprint arXiv:2002.07868 , 2020. [100] Tao Xin, Shijie Wei, Jianlian Cui, Junxiang Xiao, Inigo Arrazola, Lucas Lamata, Xiangyu Kong, Dawei Lu, Enrique Solano, and Guilu Long. Quantum algorithm for solving linear : Theory and experiment. Physical Review A , 101(3):032307, 2020. [101] Yudong Cao, Anargyros Papageorgiou, Iasonas Petras, Joseph Traub, and Sabre Kais. and circuit design solving the poisson equation. New Journal of Physics , Shengbin Wang, Zhimin Wang, Wendong Li, Lixin Fan, Zhiqiang Wei, and Yongjian Gu. Quantum fast poisson solver: the algorithm and modular circuit design. arXiv preprint arXiv:1910.09756 , 2019. [103] Alexander Engel, Graeme Smith, and Scott E Parker. Quantum algorithm for the vlasov Review A , 100(6):062315, 2019. [104] ,2020. [105] Ilya Y Dodin and Edward A Startsev. On applications of quantum computing to plasma preprint arXiv:2005.14369 , 2020. [106] Jean-Luc Vay, David Bruhwiler, David Sagan, Axel Huebl, Remi Lehe, Cho-Kuen Ng, Ji Qiang, Robert Ryne, Maxence Thevenet, Henri Vincenti, and Alex Thomas. Center(s) for Accelerator and Beam Physics Modeling. Snowmass21 LOI , 2020. URL . [107] MIT schwarzman college of computing. URL . [108] QuICS. URL . [109] Chris Fisher. IBM | quantum computing. URL . - 21 -

2108.11027

David Sagan

D. Sagan, M. Berz, N.M. Cook, Y. Hao, G. Hoffstaetter, A. Huebl, C.-K. Huang, M.H. Langston, C. E. Mayes, C. E. Mitchell, C.-K. Ng, J. Qiang, R. D. Ryne, A. Scheinker, E. Stern, J.-L. Vay, D. Winklehner, H. Zhang

Simulations of Future Particle Accelerators: Issues and Mitigations

21 pages, 1 figure. To be published in JINST

10.1088/1748-0221/16/10/T10002

physics.acc-ph

http://creativecommons.org/licenses/by/4.0/

The ever increasing demands placed upon machine performance have resulted in the need for more comprehensive particle accelerator modeling. are key to the success of particle accelerators. Many aspects of particle accelerators rely on computer modeling at some point, complex simulation tools and massively parallel include the modeling of beams at extreme intensities and the quantum degeneracy limit), and with ultra-fine control (down to the level of individual particles). In the future, adaptively tuned models might also be relied upon to provide beam measurements beyond the resolution of existing diagnostics. Much time and effort has been put into software tools, some of which are highly successful. However, there are also shortcomings such as the general inability of existing software to be easily modified to meet changing simulation needs. In this paper strategies are discussed for issues faced by the as it endeavors to produce better and more comprehensive modeling tools. This includes lack of coordination between code developers, lack of standards to make codes portable and/or reusable, lack of documentation, among others.

from the Defects4J dataset and five well-known spectra for- mulae. The obtained results indicate that for all the the call frequency chain based tie-breaking strategy can improve the localization effectiveness in many ways. For example, it completely eliminated 72-73% of the critical ties over the full dataset. In other cases, it reduced their Ranks of buggy elements improved by two po- sitions on average, the approach achieved positive movement of bug ranks in most rank categories, and in particular, the number of cases where the faulty method became the top ranked element increased by 23-30%. The main contributions in the paper can be summarized as follows: 1) Analysis of rank tie prevalence in the benchmark pro- grams. 2) A new tie-breaking algorithm that successfully breaks critical ties in many cases. 3) The analysis of the impact of tie-breaking on the overall SBFL terms of the concrete research goals, we defined the following Research Questions (RQs) for this paper: RQ1 How prevalent are rank ties when applying a selection of different SBFL formulae? In particular: How common are rank ties in the Defects4J benchmark and what are their sizes? What would be the theoretically achievable if all critical ties were broken? RQ2 What level of tie-breaking can we achieve using the call- frequency based strategy? RQ3 What is the overall effect of the proposed tie-breaking on SBFL effectiveness in terms of global rank improvement? The remainder of the paper is organized as follows. Sec- tion II presents an overview of the related work. Section III describes the tie problem in software fault localization. Sec- tion IV deals with RQ1, while Section V introduces our approach and answers RQ2. Section VI presents the description of our empirical evaluation of RQ3. Section VII reports the potential threats to validity, finally we provide our conclusions in Section VIII. II. R ELATED WORK Software fault localization is a significant research topic in software engineering. Despite having started in the late 1950s, software fault localization research has gained more attention in the last couple decades. This is reflected in the increase in the number of techniques, tools, and publications. The main reason for the increased attention is the dramatic increase in software systems size due to the newly added features they provide. This also has led to an increase in the complexity of these systems. As a result, more faults have also been reported. Here, software fault localization is a good approach to reduce the number of faults and to ensure Many fault localization techniques, in addition to the ones used in this paper, have been proposed and discussed in the literature. There have been several surveys written [5],[7], [8] and various empirical studies [9], [10] performed to compare the effectiveness of various techniques. However, a systematic research work on the problem of addressing ties in the context of fault localization is still modest. The most related publications are presented here. Yu et al. [11] proposed a tie-breaking strategy that firstly sorts program statements based on their suspiciousness and then breaks ties by sorting statements based on applying a confidence metric. The metric is intended to assess the degree of certainty in a given suspiciousness value. For example, when two or more statements are assigned the same level of suspicion, the suspiciousness assigned to the statements with a higher level of certainty is more reliable. As a result, statements are more likely to be faulty. Xu et al. [12] have presented the most systematic analysis of the problem associated with critical ties (ties with where four tie-breaking strategies were considered and evaluated via experimental case studies. Their that some of the strategies can reduce ties without having an adverse impact on fault localization they proposed some other tie-breaking techniques to be studied and evaluated in the future such as to breaking ties. Debroy et al. [6] proposed a grouping-based strategy that employs another influential factor alongside statements' This strategy groups program statements based on the number of failed tests that execute each statement and then sorts the groups that contain statements that have been exe- cuted by more failed tests. Afterwards, it ranks the each group by their suspiciousness to generate the final ranking list. Thus, the statements are examined firstly based on their group order and secondly based on their results show that ranking based on several factors can improve the SBFL effectiveness. Thus, the could be effective in tie-breaking as well. Laghari et al. [13] employed the idea of utilizing method calls to improve the performance of SBFL. In their they combined method calls and their sequences with program slicing to extract spectra patterns from that can be used to effectively locate faults compared to only using the standard SBFL formulae. It can be noted that utilizing method calls to improve the performance of SBFL is not new. However, using method calls frequency for tie-breaking is a novel approach which has not been investigated by other researchers previously. III. S PECTRUM -BASED FAULT LOCALIZATION AND TIES In this section, we present how SBFL techniques are used to locate faults by ranking program elements based on of being faulty and what are the steps to do so. Also, we introduce the problem of ties among program elements in the rankings that these techniques produce. This is achieved by a simple code example that illustrate concepts.A. Fault Localization Formulae SBFL is a dynamic program analysis technique which is performed through program execution. In SBFL, code cover- age information (also called spectra) obtained from executing a set of test cases and test results are used to calculate the probability of each program entity (e.g., statements, blocks, or methods) of being faulty [14]. Code coverage on which program entity has been executed and which one has not during the execution of each test case; while tests results are classified as passed or failed test cases. Passed test cases are executions of a program that output as expected, whereas failed test cases are executions of a program that output as unexpected [15]. To illustrate the work of SBFL, assume a simple Java program, which is adopted from Vancsics et al. [16], that comprises of four main methods ( a,b,f, andg), and its four test cases (t1,t2,t3, andt4) as shown in Figure 1. Suppose that the tests have been executed on the program and the program spectra (the execution information of the four program methods in passed and failed test cases) have been recorded. Table I presents this information. An entry of 1 in the cell corresponding to the method aand the test case t1 means that the method ahas been executed by the test case t1, and 0 otherwise. This is also known as the hit-based SBFL. An entry of 1 in the row labeled results means that test case resulted in failure, and 0 otherwise. For example: t2test case calls the methods a,bandgand it failed because its expected value is 3and not 4. Fig. 1: Running example - program code and test cases The program spectra are then used by a spectra formula to compute the suspiciousness of each program element of being faulty. Often, a spectra formula is expressed in terms of four counters that are calculated from the program spectra set of failed test cases that executed m. b)mep: set of passed test cases that executed m. c)mnf: set of failed test cases that not executed m. d)mnp: set of passed test cases that not executed m. The last four columns of Table I represent these values. For example, efofacontains two tests because failed tests t1and t2are executed by a, and npoffincludes only one test ( t3) because it is not run by f. TABLE I: Coverage hit spectrum (with four basic statistics) t1 t2 t3 t4 ef ep nf np a 1 1 1 1 2 2 0 0 b 1 1 1 1 2 2 0 0 f 1 0 0 1 1 1 1 1 g 1 1 1 1 2 2 0 0 Results 1 1 0 0 Most formulae use these four values to determine the loca- tion of the bugs as accurately as possible. In this paper, we use five popular formulae for quantitative evaluation as presented in Table II. The DStar [17], Ochiai [18] and Tarantula [19] can be seen as the most popular ones. While the Confidence [12] was used to give importance to suspicious program for tie-breaking purposes. Finally, the GP13 [20] is a generated formula by genetic algorithm which is one of the best performing formulae of this kind. TABLE II: SBFL formulae used in the study Name (Conf jmefj
 1 (Och j By applying these formulae on the coverage hit spectra of our Java program example in Table I, we can obtain the suspiciousness scores of each method as presented in Table III. It can be noted that in this example, each SBFL formula produces the same suspiciousness score for more than one method. In other words, SBFL formulae in this case cannot distinguish the methods from each other based on their pure scores. Hence, the tie problem among program elements affects the SBFL effectiveness in this case.TABLE III: Program example scores Method Conf DSt GP13 Och Tar a 0.00 2.00 2.33 0.71 0.50 b 0.00 2.00 2.33 0.71 0.50 f 0.00 0.50 1.33 0.50 0.50 g 0.00 2.00 2.33 0.71 0.50 B. Rank Calculations and Ties When different elements are assigned the same suspicious- ness score, we treat these elements score tied to each other, and we call any such set of code elements rank ties . Clearly, rank ties have at least two elements. Since the output of SBFL algorithms should be a (weakly) monotone list of ranked code elements according to their suspciousness scores, there are various strategies for dealing with rank ties. This is especially important when evaluating the effectiveness of an SBFL method in terms of the location of the actually faulty element in the rank list. In this situation all elements in a rank tie are assigned the same rank value, based on one of these approaches [21]: minimum (MIN) : it refers to the top most position of the elements sharing the same suspiciousness value (op- timistic or the best case) , maximum (MAX) : it refers to the bottom most or the worst case) , or average (MID) : it refers to the medium position of the elements sharing the same suspiciousness value (average case) . As a general way of assessing SBFL effectiveness, we will use the average rank approach (Equation 1), but we will use the other two options as well to examine tie properties in the sections that follow. If there are multiple bugs for a program version, the highest rank of faulty elements is used. MID =S+E - 1 2 (1) Where S is the tie starting position and E is the tie size. Table IV presents the average ranks of the example program based on several fault localization algorithms. Ranks that belong to a tie are marked in gray. It can be stated that two algorithms ( Conf andTar) cannot distinguish the methods from each other at all based on ranks, and the other three approaches result a tie-group that contain 3 methods. TABLE IV: Average ranks of program example Method Conf DSt GP13 Och Tar a 2.5 2 2 2 2.5 b 2.5 2 2 2 2.5 f 2.5 4 4 4 2.5 g 2.5 2 2 2 2.5 Thus, such methods get tied in the ranking and cannot be differentiated from each other in terms of which one has to be examined first. Therefore, tie-breaking strategies are required to break these ties. Tie-breaking strategies arenot only important to measure the effectiveness of a formula, they are also important for designing an efficient algorithm. For example, with an effective strategy, the buggy method can be moved up (i.e. to a better position) in the suspicious list. SBFL formulae that do not deal with the issue of rank ties do not take into account other suspiciousness factors derived from the context in the ranking. It is quite frequent that ties include faulty elements and it is not limited to any technique or target program. As a result, such elements are tied to the same position in the ranking. Also, it gives an indication that the used technique cannot the tied elements in terms of their likelihood of being faulty. Thus, no guidance is provided to developers on what to examine first [12]. In addition, the greater the number of ties involving faulty elements, the more difficult it is to predict at what point the fault will be found during the ties can be divided into two important types: non- critical andcritical . Non-critical ties refer to the case where only non-faulty elements are tied together for the same score in the ranking. Here, if the tied elements have a than the actual faulty element, then every element will be examined before finding the fault, regardless of the ties. On the other hand, if the tied elements have a than the actual faulty element, then the faulty element will be examined before the tied ones. Thus, there is no need to continue examining the ranking. In either case, the internal order in which the non-critical tied elements are examined does not affect the performance of fault localization in terms of the number of elements that must be examined before finding the fault. Critical ties, on the other hand, refer to the case when a faulty element is tied with other non-faulty elements [12]. In this type, the internal order of examination affects the performance. In the case of tie-breaking approaches, critical ties are the main target as it can bring improvement to the efficiency of the SBFL algorithm. But, unfortunately, we do not know which code element is faulty, so all ties have to be dealt with by the tie-breaking strategy. IV. T IESTATISTICS In this section, we analyze the existence of rank ties in a set of benchmark programs. We present the subject programs that were used in our experimental results alongside their proper- ties and the granularity of data collection that was employed as a program spectra for our selected SBFL formulae. Then, we present the properties of ties we obtained before applying our tie-breaking strategy and to which extent they can be reduced. A. Subject Programs An appropriate dataset is required to examine fault local- ization. One of the most popular bug dataset is Defects4J. It is a database of non-trivial real faults which is used to enable reproducible studies in software fault localization for Java programs [22]. Besides, it is the most frequently used benchmark in the fault localization literature [23] as it providesa high-level framework interface to easily access faulty and fixed program versions and their corresponding test suites. The version we used in this study is v1.5.0 and consists of 6 open- source Java programs and 438 real faults which were identified and extracted from the projects' repositories1. However, a few faults were excluded in this study due to or unreliable test results. Thus, a total of 411 faults were included in the final used dataset. Table V presents each program and its main V: Subject testsNumber of methods Chart 25 96 2.2k 5.2k Closure 168 91 7.9k 8.4k Lang 61 22 2.3k 2.4k Math 104 84 4.4k 6.4k Mockito 27 11 1.3k 1.4k Time 26 28 4.0k 3.6k All 411 332 22.1k 27.4k B. Granularity of Data Collection In this paper, method-level granularity was employed as a program spectra or coverage type. Compared to statement- level granularity, the widely used level, it has several ad- vantages [24]: (a) it provides more comprehensive about the program entity under investigation, (b) it can handle (i.e., scales well to) large programs and executions, (c) some studies report that it is a better for the users too [10], [25]. Nevertheless, there is no to investigate lower levels of granularity as well in terms of rank ties in the future. C. Evaluation Baselines In this paper, five standard SBFL formulae, which are presented in Table II, were used as the baselines to evaluate and compare our proposed method against. The reasons behind this are: (a) there is no other proposed tie-breaking approach that works on the method-level granularity as our method does; (b) our goal was to use contextual information from only to break ties, and not as the underlying SBFL formula for all program elements (this was done by Vancsics et al. [16]). The authors in [12] used two confidence formulae to break ties and data-dependency among program statements as well, but these approaches are not directly comparable to ours. D. Basic Statistical Analysis As mentioned earlier, there is no guarantee that SBFL formulae produce unique suspiciousness scores for all the elements of a program under test. As a result, many elements may share the same scores and get tied with each other. Here, we present brief yet informative statistics on the number of ties that the selected five SBFL formulae produce when applied on the Defects4J dataset (see Table VI). It can be noted that all the selected SBFL formulae produce ties across all the target programs. This may indicate different things: (a) ties are in fault localization, (b) ties can be formed regardless of which subject program is under consideration, c) different SBFL formulae are affected. Table VII presents the number of critical ties. An is that the number of ties is not related to program size. For example, smaller programs may have more critical ties than larger programs as in the case of the Lang (22 KLOC) program having more critical ties compared to the Chart (96 KLOC) program. The average number of critical ties per bug is an important indicator, as it means in essence the probability that a buggy element will be tied (assuming a single-bug scenario). TABLE VI: Number of ties: total and average per bug ProjectConf DSt GP13 Och Tar # avg # avg # avg # avg # avg Chart 3656 146.24 508 20.32 512 20.48 506 20.24 490 19.60 Closure 86181 512.98 19069 113.51 19017 113.2 19043 113.35 19109 113.74 Lang 3430 56.23 185 3.03 188 3.08 187 3.07 187 3.07 Math 12856 123.62 844 8.12 846 8.13 854 8.21 851 8.18 Mockito 3381 125.22 779 28.85 780 28.89 779 28.85 789 29.22 Time 5776 222.15 589 22.65 571 21.96 597 22.96 609 23.42 All 115280 280.49 21974 53.46 21914 53.32 21966 53.45 22035 53.61 TABLE VII: Number of critical ties: total and average per bug ProjectConf DSt GP13 Och Tar # avg # avg # avg # avg # avg Chart 14 0.56 17 0.68 14 0.56 15 0.6 16 0.64 Closure 102 0.61 101 0.60 102 0.61 101 0.60 100 0.60 Lang 20 0.33 22 0.36 20 0.33 21 0.34 22 0.36 Math 65 0.62 69 0.66 65 0.62 65 0.62 65 0.62 Mockito 13 0.48 13 0.48 13 0.48 13 0.48 13 0.48 Time 9 0.35 9 0.35 9 0.35 10 0.38 10 0.38 All 223 0.54 231 0.56 223 0.54 225 0.55 226 0.55 We can conclude that more than half of the bugs (54-56%) are within critical ties, i.e. in most cases there is at least one method whose suspiciousness score is the same as the score of the faulty method. The sizes of ties is another important factor when consid- ering the potential improvements by tie-breaking. This can be investigated by looking at the differences between the MIN (best case) and the MID (average case) approaches described in the previous section. Consider Table VIII which shows the number of critical ties for which MIN and MID values are dif- ferent in columns 2 and 3 (essentially, the critical tie numbers as shown above), and also the sum of the corresponding rank differences (column 4), and its average per critical tie (column 5). Put it differently, the double of the average difference is the average critical tie size in the benchmark, which is around 7 methods. The difference between the different formulae is not notable. It also follows that, ideally, the best improvement we could achieve using a tie-breaking technique is these averages. From Table VIII, we can see in how many cases there is any improvement possible, so we can use these numbers as a base- line for evaluating our tie-breaking approach in VIII: Improvement possibilities based on critical tie numbers and average tie sizes MIN != MID (count)MIN != MID (%)Diff. Avg. diff. Conf 223 54.3 758.5 3.40 DSt 231 56.2 795.0 3.44 GP13 223 54.3 799.5 3.59 Och 225 54.7 784.0 3.48 Tar 226 55.0 831.0 3.68 We examined the distribution of the critical tie sizes as well, which is shown in Figure 2. The X-axis represents the number of methods involved in critical ties and the Y-axis represents the percentage of method groups that have the same tie size. As expected, most ties are relatively small (2-4 elements), 67% of the critical ties contain 5 or less methods, and sizes above 15 are rare (the average is 7.8, the median 3 and the maximum 128). Interestingly, there are some outlier cases where the tie sizes are very big, which is the explanation of the relatively large average number. Fig. 2: Distribution of size of critical ties RQ1 : Overall, it can be said that the ties and critical ties are very common (for the bugs in our benchmark). Each of the examined SBFL algorithms created critical ties for more than half of bugs, and on average, the ranks could potentially be improved by around 3.5 positions by eliminating the ties. V. C ALL FREQUENCY -BASED TIE-BREAKING In this section, we present the concepts of our strategy and how it works. Then, we present in reducing critical ties when applied on our Frequency-based Tie Reduction In Section III, we introduced the basic concept of hit-based SBFL. One disadvantage of this approach is that it does not take into account the frequency of executing the in our case methods (also known as the There have been studies that used counts [26], [27], but recent results [28] have shown that these are unable to improve efficiency of the et al. [16] proposed a technique to replace the simple count-based approach that proved to enhance while eliminating the problems of the naive counts. It is based on replacing the value of efin the SBFL formulae with the frequency of different call contexts in the call stack forfailing tests. The basic intuition is that if a method participates in many different calling contexts (both as a caller and as a callee), it will be more suspicious. In other words, the frequency of methods occurring in the unique call stacks belonging to failing test cases can effectively indicate the location of the bugs. In the present research, we will employ this concept for the purpose of tie-breaking. To illustrate the basic concept of frequency-based tie re- duction, first we define the frequency-based SBFL matrix that replaces the traditional hit-based one. In it, each element will get an integer instead of f0;1gindicating the number of occurrences of the particular code element in the unique call stacks (effectively, the different contexts) when executing the given test cases [16]. Table IX shows the for the example. For instance, the call stacks of so the frequency of gwill be 2 IX: Example b f g Results t1 2 1 1 2 Failed t2 1 1 0 2 Failed t3 1 1 0 1 Passed t4 3 1 1 2 Passed  3 2 1 4 In the next step, we define our metric to be used as a discriminating factor for tie-breaking. The corresponds to the frequency-based ef and is calculated by summarizing the corresponding frequency-based values in the matrix for the failing test cases (see Equation 2). The values for our example are shown in the last row of Table IX. (m) =X t2failed 3 shows our tie-breaking process which can be seen as a two-stage process. In the first stage, we compute scores of program methods and their ranks via applying different SBFL formulae on the program spectra (test coverage and test results). The output of this stage is an initial ranking list of program methods including critical and non- critical ties. In the second stage, we trace the execution of program methods to obtain the , i.e. frequency-based ef. This will then be used as a tie-breaker after re-arranging the order of the critical tied methods in the initial ranking list based on the value of for each method. The output of this stage is a final ranking list, where many critical ties either or their sizes were reduced. Our proposed tie-breaking method uses the obtained call frequency values to break the methods sharing the same score, by putting the methods with higher upper in the rank. Thus, the most suspicious one will be the method that was called in more different call stacks from failing test cases.Fig. 3: The proposed tie-breaking process The rationale behind using the rather than other contexts (such as the context of method calls in passing test cases) is the intuition that a method is more suspicious to contain a fault when executed by more failing test cases than passing ones, while non-suspicious when mostly executed by passing tests. However, other different contexts could be considered in the future and investigate their impacts on breaking ties. The ranks without (columns B) and with A) with our approach for the example are presented in Table X. Ties marked in gray were eliminated with the use of call frequency. As a result, we were able to the faulty method ( g) and the other suspicious ones using all of the SBFL formulae (the faulty method got the highest rank in all cases). TABLE X: Ranks Before and After using the DSt GP13 Och Tar B A B A B A B A B A a 3 2.5 2 2 2 2 2 2 2 2.5 2 b 2 2.5 3 2 3 2 3 2 3 2.5 3 f 1 2.5 4 4 4 4 4 4 4 2.5 4 g 4 2.5 1 2 1 2 1 2 1 2.5 1 B. Reduction of the Critical Ties The metric called Tie-reduction was defined by Xu et al. to measure how much a critical tie has been terms of size [12]. Here, size simply means the number of code elements sharing the same score value, and obviously, the minimum tie size is 2. The goal of any tie-breaking strategy is to reduce the size of the tie or completely eliminate it (when the resulting size is 1). We modified the original definition of this metric to better reflect the actual gain in terms of what portion of the superfluous elements in a tie can be eliminated (see Equation = 1size after1 size before1
100% (3)Here, size afteris the size of a critical tie after applying a tie- breaking strategy and size before is the size of a critical tie before applying a tie-breaking strategy. In an ideal case, the critical tie is completely eliminated, in which case the value of the tie- reduction is 100% . If no reduction can be obtained, this metric will be 0%, and in all other cases it will show the percentage of the removed elements that share the same score value as the faulty element. In Figure 4, we visualized the amount of critical tie- reduction on our benchmark using the Tie-Reduction metric. Each dot represents one bug in the dataset and the violin plot offers a more general picture about the distribution of the data points. It can be seen from the shape of the plots that in several cases, reduction was not possible but the majority of the ties was completely eliminated. Similar to the number and size of critical ties, there was no significant difference in this aspect depending on what SBFL formula was used, we obtained very similar results. Fig. 4: Tie-reduction distribution of critical ties Table XI shows some important statistical values for this dataset: mean, median and quartile 1 (the value in the middle between the smallest and the median points). Since the median is 100%, we can state that the critical ties are eliminated by our method in more than half of the cases (72-73%, as detailed below), and the reduction is between 83.9-91.5% for three- quarters of the bugs, while the average rate of reduction is greater than 80% in all cases. TABLE XI: Statistics of tie-reduction (in percentage) Conf DSt GP13 Och Tar Mean 81.8 80.5 81.8 81.5 81.8 Median 100.0 100.0 100.0 100.0 100.0 Q1 91.5 83.9 91.5 90.4 88.9 Table XII presents the number of remaining critical ties for each program and SBLF formula after applying the tie- reduction algorithm. The difference to the previous values (shown in Table VII) is also included. For example, 15 bugs of Chart were in critical ties with the Ochiai formula, but after applying the call frequency-based tie-reduction 11 critical ties are eliminated, which is 73.3% of the initial ties. Overall, we achieved 72-73% improvement in the number of critical ties over the full dataset, the best case being Mockito with over 84.6% and the worst result was 54.5% on Lang using Tarantula andDStar .TABLE XII: Changes in the number of critical ties Closure Lang Math Mockito Time All Confafter 4 24 9 18 2 2 59 diff. (#) 10 78 11 47 11 7 164 diff. (%) 71.4 76.5 55.0 72.3 84.6 77.8 73.5 DStafter 6 24 10 21 2 2 65 diff. (#) 11 77 12 48 11 7 166 diff. (%) 64.7 76.2 54.5 69.6 84.6 77.8 71.9 GP13after 4 25 9 18 2 2 60 diff. (#) 10 77 11 47 11 7 163 diff. (%) 71.4 75.5 55.0 72.3 84.6 77.8 73.1 Ochafter 4 24 9 19 2 2 60 diff. (#) 11 77 12 46 11 8 165 diff. (%) 73.3 76.2 57.1 70.8 84.6 80.0 73.3 Tarafter 5 24 10 19 2 2 62 diff. (#) 11 76 12 46 11 8 164 diff. (%) 68.8 76.0 54.5 70.8 84.6 80.0 72.6 The sizes of the critical ties determine the level of achiev- able improvement after applying a tie-breaking it is also important in which direction in the new ranking the faulty element moved after tie-breaking. Using the terminology from the previous section, moving from the MID position towards MIN means improvement. In the in Table VIII; we presented the maximum that sets a theoretical constraint on SBFL ef- fectiveness after tie-reduction. Table XIII presents what we actually achieved using our proposed algorithm (the meaning of the data is the same as in Table VIII). TABLE XIII: Achieving the minimum ranks MIN != MID (count)MIN != MID (%)Diff. Avg. diff. Conf 59 14.4 53.5 0.9 DSt 65 15.8 61.5 0.9 GP13 60 14.6 54.0 0.9 Och 60 14.6 55.5 0.9 Tar 62 15.1 94.0 1.5 We examined whether our method was able to reduce the number of cases where the MIN (best case) and the MID (average case) approaches give different results. If there were no such cases that would mean that the obtained new ranking after tie-break would always be the best possible, MIN case. Table XIII shows that, after our approach, only around 15% of the bugs contained critical ties (column 3), compared to around 55% before this with the result of Table VIII, we find that in more than 160 cases we managed to achieve the ideal result with our method where the original algorithm was not able to do so. It means that for nearly three quarters of bugs in critical ties (72-73%), the non-optimal result was improved to optimal. If we compare the sum of the rank differences (column 4) and their averages (column 5) in Tables VIII and XIII, it can be seen that our approach was able to reduce the (by 89-93%), and the average by 59-74%. Put it differently, the overall rank positions from the ideal case improved from around 3.5 to 1 in the cases when we achieved optimal result, which essentially means rank 2:2and2:5.RQ2 : Using the call-frequency based tie-breaking strategy, we achieved a significant reduction in both size and number of critical ties in our benchmark. In 72-73% of the cases the ties were completely eliminated, the average reduction rate being more than 80%. In nearly three quarters of the cases (72-73%), the faulty element got the highest rank among the tie-broken code elements, and here it improved its position by 59-74% on average. VI. E FFECT OF TIE-BREAKING ON SBFL P ERFORMANCE In this section, we analyze what is the overall effect of the proposed tie-breaking strategy on SBFL effectiveness in terms of global ranks. For that purpose we use several that were employed in the literature [12], [29], [30]. A. Achieved improvements and the average ranks Average rank is used to rank the program elements that share the same suspiciousness value by considering the av- erage of their positions after they get sorted, in a by the level of their suspiciousness. And, it is calculated using Equation 1. Table XIV presents the average ranks before (column 2) and after (column 3) applying our and it shows the difference between the average ranks before and after tie-reduction (column 4). If the difference is negative then this means that we could achieve improvement with our proposed strategy. TABLE XIV: Average rank of faulty elements before and After Diff. Conf 55.16 53.11 -2.05 DSt 46.86 44.79 -2.07 GP13 68.79 66.68 -2.11 Och 46.95 44.81 -2.14 Tar 50.39 48.33 -2.06 We can see that our strategy achieved improvements with all the selected SBFL formulae: the average rank reduced by more than2in all cases, which corresponds to 3.1-4.1% with respect to the total number of elements. Note, that this average is similar to what we got for RQ2, but it is not the same because for RQ2 we examined only the cases when we achieved the optimal result, while in this section we are interested in the global results. We also examined how many times our tie-breaking strategy changed the rank of bugs (in positive and negative directions) and what was the impact of the changes. Table XV presents the possible changes in several categories, as follows (B means before, A means after applying the faulty method moved to the top of the critical tie (column: best), when BMIN=AMID(this is the case that we discussed using Tables VIII and XIII) 2) it has moved up in the rankings (column: better), BMIN<AMID 3) it remained in the same position (column: same), we worsened the result (column: worse), BMAX>AMID5) it slipped back to the worst place (column: worst), addition, column improve represents improvements in rank modifications (i.e., best+better), while deteriorate is worse+worst. The table also includes the average differences in rank positions for the given categories. The results indicate that in about 3-4 times more cases we achieved improvement than deterioration of the ranking re- sults. Moreover, the improvement differences are much higher than the deterioration differences (compare, for example, the better cases of around -7 to worse cases of around 2). insight is that in the case of best, the difference is relatively small as the size of the ties broken in this category were small as well (they contained 3-4 methods). Looking at the overall result, the average rate of improvement ranged from -3.73 to -3.86, while the deterioration was only between 1.34 and 1.54 rank positions on average. TABLE XV: Comparison of average ranks before and Better Same Worse Worst Improve 85 51 50 17 20 136 37 avg. diff. -1.71 -7.13 0 2.32 0.55 -3.74 1.36 DStcount 85 53 53 18 22 138 40 avg. diff. -1.71 -7.32 0 2.31 0.55 -3.86 1.34 GP13count 85 51 50 17 20 136 37 avg. diff. -1.71 -7.39 0 2.32 0.55 -3.84 1.36 Ochcount 86 52 50 16 21 138 37 avg. diff. -1.70 -7.42 0 2.38 0.62 -3.86 1.38 Tarcount 83 58 47 17 21 141 38 avg. diff. -1.46 -6.97 0 2.68 0.62 -3.73 1.54 The overall rank position improvement might seem modest, but we must consider the fact that the improvement can be achieved only by rearranging the positions in the critical ties. Thus, the sizes of the critical ties serve as a hard constraint (as discussed in the previous section). However, there is a class of improvements which are probably more important than the general case: improvements in the Top-N rank positions, and here the benefits are more pronounced, as presented below. B. Top-N studies [31], [32] report that developers think that inspecting the first 5 program elements in the ranks list produced by a fault localization technique is acceptable and that the first 10 elements are the upper bound for ignoring the ranks list. Hence, we verified the results by focusing on these rank positions only (collectivelly called Top-N ). We will use five cases: where a fault is ranked first (Top-1), it is less or equal to three (Top-3), less or equal to five (Top-5), less or equal to ten (Top-10), and when it is more than ten (Other). We also used a special non-accumulating variant of in which case we counted cases where the bug fell into a non-overlapping intervals of The goal of the evaluation in this part was to see in how many cases our approach moves a bug into a better (for example, from (5;10]to(1;3]) or a worse (for example, from [1]to(1;3]) group. In other words, in how many cases do thebugs get into a higher-rank group (this kind of improvement is also known as enabling improvement [30]) and in how many cases do we downgrade the category. Table XVI presents the number of bugs belonging to Top-N categories (cumulative) with their per- centages, for the whole dataset, before and after applying our tie-breaking strategy, as well as the differences between them. A decrease in the number of cases of the Other category and increase in any Top-N means XVI: Top-N categories Top-1 Top-3 Top-5 Top-10 Other # % # % # % # % # % Conf 75 18.2 169 41.1 203 49.4 246 59.9 165 40.1 After tie-breaking 92 22.4 180 43.8 214 52.1 252 61.3 159 38.7 Diff. 17 22.7 11 6.5 11 5.4 6 2.4 -6 -3.6 DSt 65 15.8 172 41.8 210 51.1 249 60.6 162 39.4 After tie-breaking 84 20.5 186 45.4 222 54.1 257 62.7 153 37.3 Diff. 19 29.2 14 8.1 12 5.7 8 3.2 -8 -4.9 GP13 75 18.2 169 41.1 203 49.4 245 59.6 166 40.4 After tie-breaking 92 22.4 179 43.6 212 51.6 250 60.8 161 39.2 Diff. 17 22.7 10 5.9 9 4.4 5 2.0 -5 -3.0 Och 68 16.5 173 42.1 210 51.1 250 60.8 161 39.2 After tie-breaking 87 21.2 186 45.3 222 54.0 257 62.5 154 37.5 Diff. 19 27.9 13 7.5 12 5.7 7 2.8 -7 -4.3 Tar 65 15.8 166 40.4 203 49.4 244 59.4 167 40.6 After tie-breaking 83 20.2 177 43.1 212 51.6 251 61.1 160 38.9 Diff. 18 27.7 11 6.6 9 4.4 7 2.9 -7 -4.2 It can be clearly seen that our proposed tie-breaking improvements in all categories by moving many bugs to higher ranked categories. On the lower end of the scale (Other category with rank >10), 5-8 bugs were moved into one of the Top-N categories. This is important as it brings a new hope that a bug could be found by the user with the proposed strategy while it was not very probable without it. We can see a quite large number of improvements in higher categories as well, around 18 bugs moved to Top-1, for instance. Note that the percentages of bugs in each category before and after applying the strategy were calculated with respect to the total number of bugs in the dataset. While the difference percentage was calculated with respect to the number of bugs before applying the strategy. To better understand the actual changes between the dif- ferent Top-N categories we should use the of these categories. This shows whether there has been a beneficial change in the rank category. These moves between the Top-N categories are presented by Table XVII. The sign 7indicates the number of changes in the negative result), and 3marks improvement. For example, there were a total of 2 bugs with a rank greater than 1 but less than or equal to 3 before reduction by Tarantula , but our method resulted in a rank value greater than 3 (this is a negative result). In contrast, our method gave a rank of 1 for the faulty method 15 times which was previously greater than 1 but smaller than 3 (using Tarantula ). These numbers clearly show that improvement was dom- inant: degradation by the proposed method was observable only for 2-3 bugs in the dataset, while we observed positive changes for 36-44 bugs.TABLE XVII: Top-N moves [1] (1;3] (1;3] (3;5] (3;5] (5;10] (5;10]Other 7 3 # # # # # # # # 7 7 3 7 3 7 3 3 Conf 0 1 13 1 8 0 11 6 2 38 DSt 0 1 15 1 10 0 11 8 2 44 GP13 0 1 13 1 8 0 10 5 2 36 Och 0 1 15 1 9 1 11 8 3 43 Tar 0 2 15 1 11 0 8 7 3 41 RQ3 : The efficiency of all investigated SBFL formulae could be improved by using the proposed tie-breaking strategy: the average improvement of rank values in the benchmark was about two positions , and about 3-4 times more frequenty we observed improvement than detoriation, such much higher as well. Considering the Top-N improvements could be observed: all Top-N positive results (improvements in 36-44 cases), and at the same time, in only a few (2-3) cases Top-N We were able to increase the number of cases where the faulty method became the top ranked element by 23-30%. VII. T HREATS TO VALIDITY Various threats may affect the validity of in software engineering. In our work, we considered the following actions to avoid or minimize such of evaluation metrics: to ensure that our results and corresponding conclusions are valid, we selected several evaluation metrics that are also used by previ- ous research to ensure all the evaluation metrics employed in this study were reported and described in of implementation: to ensure that our experi- ment implementation is correct and accurate, code review was conducted by the authors. Furthermore, we have run our proposed approach several times to ensure that it is implemented of subject programs: in our experiment, we evaluated the effectiveness of the proposed on fault localization using only six Java Thus, we cannot generalize our findings to other programs in general. However, we believe that the selected subject programs are representative to others as they have real faults, varying in size and complexity, and the benchmark containing them, Defects4J, is used commonly in other studies on software fault of faults: in our experiment, 27 faults (about 6% of the total faults) of the Defects4J dataset were excluded because we could not compute their call due to technical limitations. The issue here is whether other researchers using the same dataset will be able to replicate our findings. This exclusion was in no ways influenced by the results of the used metrics and the excluded faults are distributed in the evenly, so we believe that this risk can be considered of SBFL formulae: to evaluate the effectiveness of our proposed tie-breaking strategy on fault localization,we selected a set of five SBFL formulae in our experi- ment, which is just a fraction of the proposed techniques in literature. The obtained results show all of them. However, we cannot guarantee that the same improvements can be obtained by using other SBFL formulae. To mitigate the effect of this issue, we selected three SBFL formulae that are commonly used in other studies on software fault localization, which we extended with two special kinds, one of which was with tie-breaking in mind. VIII. C ONCLUSION Rank ties in SBFL are very common regardless of the formula employed, and by breaking these ties, improvements to the localization effectiveness can be expected. This paper proposes the use of method call contexts for breaking critical ties in SBFL. We rely on instances of call stack traces, which are useful software artifacts during run-time and can often help developers in debugging. The frequency of the occurrence of methods in different call stack instances determines the position of the code elements within the set of other methods tied together by the same suspiciousness results show that the proposed , using the Defects4J benchmark, (a) completely elim- inated many critical ties with significant reduction of others, and (b) achieved improvements in average rank positions for all investigated SBFL formulae with moving many bugs to the highest Top-N rank positions. However, there are limits to how much improvement one can expect from (we analyzed this limit in the paper and compared to the results achieved). This means that no matter how clever a tie- breaking method is, it cannot rearrange code elements outside of the tied ranking positions. Since ties seem to be prevalent, it could be an interesting further research to devise approaches or modified formulae that minimize ties in the scores and/or break them other future work, we would like to measure the effec- tiveness of the proposed tie-breaking strategy on other levels of granularity such as statement, branch, etc. Employing other SBFL formulae across a much broader range of programs in terms of numbers, types, sizes, and used to capture the ties problem characteristics and identify what factors affect them would be interesting for further investigation. We also would like to tackle the ties problem by employing other contextual factors beyond method calls and to measure their impacts on the SBFL. The results of our experimental study can be found research was supported by the Ministry of Innovation and Technology, NRDI Office, Hungary within the framework of the Artificial Intelligence National Laboratory Program, and by grant of the Ministry for Innovation and Technology. Qusay Idrees Sarhan was supported by the Stipendium Hungaricum scholarship programme.

2108.11031

Qusay Idrees Sarhan

Qusay Idrees Sarhan, Bela Vancsics, Arpad Beszedes

Method Calls Frequency-Based Tie-Breaking Strategy For Software Fault Localization

11 Pages

21st IEEE International Working Conference on Source Code Analysis and Manipulation (SCAM 2021)

cs.SE

http://creativecommons.org/licenses/by/4.0/

In Spectrum-Based Fault Localization (SBFL), a suspiciousness score is assigned to each code element based on test coverage and test outcomes. The scores are then used to rank the code elements relative to each other in order to aid the programmer during the debugging process when seeking the source of a fault. However, probably none of the known SBFL formulae are guaranteed to produce different scores for all the program elements, hence ties emerge between the code elements. Based on our experiments, ties in SBFL are prevalent: in Defects4J, 54-56% of buggy methods are members of ties, i.e., there is at least one other method with the same score in these cases (but typically much more, on average 6), and this inevitably reduces of any SBFL approach. In this work, we present a technique to break ties in such cases based on the so-called method calls frequencies. This counts the number of different contexts of method calls (both as callees and as callers) in failing test cases. The intuition is that if a method appears in many different calling contexts during a failing test case, it will be more suspicious and get a higher rank position compared to other methods with the same scores. This method can be applied to any underlying SBFL formula, and can favourably break the occurring ranks in the ties in many cases. The experimental results show that our novel tie-breaking strategy achieved a significant reduction in both size and number of critical ties in our benchmark. In 72-73% of the cases, the ties were completely eliminated and the average reduction rate was more than 80%.

Preprint in some real-world applications. Teed & Deng (2021) proposed a simultaneous localization and mapping (SLAM) system that was trained only with synthetic data and demonstrated The object detection networks for autonomous driving developed by Tesla was trained with 370 million images generated by simulation (Karpathy, 2021). The performance of syn2real transfer depends on the similarity between synthetic and real data. In general, the more similar they are, the stronger the effect of pre-training will be. On the contrary, if there is a significant gap, increasing the number of synthetic data may be completely useless, in which case we waste time and computational resources. A distinctive feature of syn2real is that we can control the process of generating data by ourselves. If a considerable gap exists, we can try to regenerate the data with a different setting. But how do we know that? More specifically, in a standard learning setting without transfer, a power law-like relationship called a scaling law often holds between data size and generalization errors (Rosenfeld et al., 2019; Kaplan et al., 2020). Is there such a rule for pre-training? In this study, we find that the generalization error on fine-tuning is explained by a simple scaling law, test error'Dn+C; (1) where coefficient D> 0andpre-training rate >0describe the convergence speed of gap C0determines the lower limit of the error. We can predict how large the pre-training data should be to achieve the desired accuracy by estimating the parameters ;C from the empirical results. Additionally, we analyze the dynamics of transfer learning using the results based on the neural tangent kernel (Nitanda & Suzuki, 2021) and confirm that the above law agrees with the theoretical analysis. We empirically validated our scaling law on settings of benchmark tasks, model sizes, and complexities of synthetic images. Our contributions are summarized as follows. *From empirical results and theoretical analysis, we elicit a law that describes how in terms of data sizes on pre-training and fine-tuning. *We confirm that the derived law explains the empirical results for various settings in terms tasks, model size, and data complexity (e.g., Figure 1). Furthermore, we demonstrate that we can use the estimated parameters in our scaling law to assess how much improvement we can expect from the pre-training procedure based on synthetic data. *We theoretically derive a generalization bound for a general transfer learning setting and confirm its agreement with our empirical findings. 2 R ELATED WORK Supervised pre-training for visual tasks Many empirical studies show that the performance at a fine-tuning task scales with pre-training data (and model) size. For example, Huh et al. (2016) studied the scaling behavior on ImageNet pre-trained models. Beyond ImageNet, Sun et al. (2017) studied the effect of pre-training with pseudo-labeled large-scale data and found a logarithmic Similar results were observed by Kolesnikov et al. (2019). Syn2real transfer The utility of synthetic images as supervised data for computer vision tasks has been continuously studied by many researchers (Su et al., 2015; et al., 2016; Georgakis et al., 2017; Tremblay et al., 2018; Hinterstoisser et al., 2019; Borrego et al., 2018; Chen et al., 2021; Newell & Deng, 2020; Devaranjan et al., 2020; Mousavi et al., 2020; Hoda Vn et al., 2019). These studies found positive evidence that using synthetic images is helpful to the fine-tuning task. In addition, they demonstrated how data complexity, induced by e.g., light randomization, affects the final performance. For example, Newell & Deng (2020) investigated how the recent perform well as a pre-training task to improve the performance of downstream tasks. In this paper, following this line of research, we quantify the effects under the lens of the scaling law (1). Neural scaling laws The scaling behavior of generalization error, including some theoretical works (e.g., Amari et al., 1992), has been studied extensively. For modern neural networks, Hestness et al. (2017) empirically observed the power-law behavior of generalization for language, image, and speech domains with respect to the training size. Rosenfeld et al. (2019) constructed a semseg mulclass task # of pretrain imagestest errorfinetune task Figure 1: Empirical results of syn2real transfer for different tasks. We conducted four object detection ( objdet ), semantic segmentation ( semseg ), multi-label ), surface normal estimation ( normal ), and three fine-tuning tasks for object detection for MS-COCO, semantic segmentation for ADE20K, and ( sinclass ) for ImageNet. The y-axis indicates the test error for each fine-tuning task. Dots indicate empirical results and dashed lines indicate the fitted curves of scaling law (1). For more details, see Section 4.2. form for the power-law in terms of data and model sizes. Kaplan et al. (2020) pushed forward this direction in the language domain, describing that the generalization of transformers obeys the power law in terms of a compute budget in addition to data and model sizes. Since then, similar scaling laws have been discovered in other data domains (Henighan et al., 2020). Theoretical analysis was also attempted (Hutter, 2021; Bahri et al., 2021). Hernandez et al. (2021) studied the scaling laws for general transfer learning, which is the most relevant to this study. A key difference is that they focused on fine-tuning data size as a scaling factor, while we focus on pre-training data size. Further, they found scaling laws in terms of the transferred effective data, which is converted data amount necessary to achieve the same performance gain by pre-training. In contrast, Eq. (1)explains the test error with respect to the pre-training data size directly at a fine-tuning task. Other differences include task domains (language vs. vision) (transformer vs. CNN). Theory of transfer learning Theoretical analysis of transfer learning has been dated back to decades ago (Baxter, 2000) and has been pursued extensively. Among others, some recent studies (Maurer et al., 2016; Du et al., 2020; Tripuraneni et al., 2020) derived an error bound of a fine-tuning task in the multi-task scenario based on complexity analysis; the bound takes an additive wherenandsare the data size of pre-training and fine-tuning, coefficients AandB. Neural network regression has been also discussed with this et al., 2020). In the area of hypothesis transfer learning (Fei-Fei et al., 2006; Yang et al., 2007), among many theoretical works, Du et al. (2017) has derived a risk bound for kernel ridge regression with transfer realized as the weights on the training samples. The obtained bound takes a similar form to our scaling law. However, the learning dynamics of neural networks initialized with a pre-trained model has never been explored in this context. 3 S CALING LAWS FOR PRE-TRAINING AND FINE-TUNING The main obstacle in analyzing the test error is that we have to consider interplay between the effects of pre-training and fine-tuning. Let L(n;s)0be the test error of a fine-tuning task with of pretrain imagestest of finetune (a) Figure 2: Scaling curves with different (a) fine-tuning size and (b) pre-training size. data sizenand fine-tuning data size s. As the simplest case, consider a fine-tuning task ( n= 0), which boils the transfer learning down to a standard learning setting. In this case, the prior studies of both classical learning theory and neural scaling laws tell us that the test error decreases the fine-tuning data size s, that is,L(0;s) rate >0and irreducible loss E0. The irreducible loss Eis the inevitable error given by the best possible mapping; it is caused by noise in continuous outputs or labels. Hereafter we assume E= 0for brevity. 3.1 I NDUCTION OF SCALING LAW WITH SMALL EMPIRICAL RESULTS To speculate a scaling law, we conducted preliminary experiments.2We pre-trained ResNet-50 by a synthetic classification task and fine-tuned by ImageNet. Figure 2 (a) presents the log-log plot of error curves with respect to pre-training data size n, where each shape and color indicates a size s. It shows that the pre-training effect diminishes for large n. In contrast, Figure 2 (b) presents the relations between the error and the fine-tuning size swith different n. It indicates the error drops straight down regardless of n, confirming the power-law scaling with respect to s. The above observations and the fact that L(0;s)decays polynomially are summarized as 1. lims!1L(n;s) = 0 . Requirement 2. limn!1L(n;s) = const . Requirement 3. L(0;s) 1 and 3 suggest the dependency of nis embedded in the coefficient B=g(n), i.e., the pre-training and fine-tuning effects interact To satisfy Requirement 2, a for the pre-training effect is g(n) =n+ ; the error decays polynomially with respect to n but has a plateau at . By combining these, we obtain L(n;s) =( +n)s; (2) where; > 0are decay rates for pre-training and fine-tuning, respectively, 0is a constant, and >0is a coefficient. 3.2 T HEORETICAL DEDUCTION OF SCALING LAW Next, we analyze the fine-tuning error from a purely theoretical point of view. To incorporate the effect of pre-training that is given as an initialization, we need to analyze the test error during the training with a given learning algorithm such as SGD. We apply the recent development by Nitanda & Suzuki (2021) to transfer learning. The study successfully analyzes the generalization of neural networks in the dynamics of learning, showing it achieves minmax optimum rate. The analysis uses the framework of the reproducing kernel Hilbert space given by the neural tangent kernel (Jacot et al., 2018). 1For classification with strong low-noise condition, it is known that the decay rate can be exponential (Nitanda & Suzuki, 2019). However, we focus only on the polynomial decay without such strong condition in this paper. 2The results are replicated from Appendix C.2; see the subsection for more theoretical analysis of transfer, it is important to formulate a task similarity between pre-training and fine-tuning. If the tasks were totally irrelevant (e.g., learning MNIST to forecast pre-training would have no benefit. Following Nitanda & Suzuki (2021), for simplicity of analysis, we discuss only a regression problem with square loss. We assume that a vector input xand scalar output pre-training and y=0(x) +1(x)for fine-tuning, where we omit the output noise for brevity; the task types are identical sharing the same input-output form, and task similarity is controlled by 1. We analyze the situation where the effect of pre-training remains in the fine-tuning even for large data size (s!1 ). More specifically, the theoretical analysis assumes a regularization term as the `2-distance between the weights and the initial values, and a smaller learning rate than constant in the fine-tuning. Hence we control how the pre-training effect is preserved through the learning rate. Other assumptions made for theoretical analysis concern the model and a two-layer neural network having Mhidden units with continuous nonlinear activation3 is adopted; for optimization, the averaged SGD (Polyak & Juditsky, 1992), an online algorithm, is used for a technical reason. The following is an informal statement of the theoretical result. See Appendix E for details. We emphasize that our result holds not only for syn2real transfer but also for transfer learning in general. Theorem 1 (Informal) .Let ^fn;s(x)be a model of width Mpre-trained by (xn;yn)and fine-tuned by ssamples (x0 1;y0 1);:::; (x0 s;y0 s)where inputs x;x0p(x) are i.i.d. with the input distribution =0(x0) +1(x0). Then the generalization error of the squared loss L(n;s) bounded from above with high probability (3) MandcMcan be arbitrary small for large M;A0andA1are constants; the exponents and depend on0,1,p(x), and the learning rate of fine-tuning. The above bound (3)shows the correspondence with the empirical derivation of the full scaling law (2). Note that the approximation error Mis omitted in (2). We note that the derived bound takes a multiplicative form in terms of the pre-training and which contrasts with the additive bounds such as et al., 2020). The existing studies consider the situation where a part of a network (e.g., backbone) is frozen Therefore, the error of pre-training is completely preserved after fine-tuning, and both errors appear in an additive way. This means that the effect of pre-training is irreducible by the effect of fine-tuning, and vice versa. In contrast, our analysis deals with the case of re-optimizing the entire network in fine-tuning. In that case, the pre-trained model is used as initial values. As a result, even if the error in pre-training is large, the final error can be reduced to zero by increasing the amount of fine-tuning data. 3.3 I NSIGHTS AND PRACTICAL VALUES The form of the full scaling law (2)suggests that there are two scenarios depending on data is big or small. In big fine-tune regime, pre-training contributes relatively little. By taking logarithm, we can separate the full scaling law (2)into the pre-training part u(n) = log(n+ )and the fine-tuning part v(s) =logs. Consider to increase nby squaring it. Since the pre-training part cannot be reduced below the infinitesimal for large s. Figure 2 (b) confirms this situation. Indeed, prior studies provide the same conclusion that the gain from pre-training can easily vanish (He et al., 2018; Newell & Deng, 2020) or a target task accuracy even degrade (Zoph et al., 2020) if we have large enough fine-tuning data. The above observation, however, does not mean pre-training is futile. Dense prediction tasks such as depth estimation require pixel-level annotations, which critically limits the number of labeled is indispensable in such small fine-tune regime. Based on this, we hereafter analyze the case where the fine-tuning size sis fixed. By eliminating s-dependent terms in (2), we obtain a 3ReLU is not included in this class, but we can generalize this condition; see (Nitanda & Suzuki, law (1)by After several evaluations, these be estimated by the nonlinear least squares method (see also Section 4.1). Log Error Log Pre-training Size Target Error (a) (b) (d) (c) Figure 3: Pre-training scenarios.As a practical benefit, the estimated parameters of the simplified law (1)bring a way to assess syn2real trans- fer. Suppose we want to solve a classification task that requires at least 90% accuracy with limited labels. We generate some number of synthetic images and pre-train with them, and we obtain 70% accuracy as Figure 3 (a). How can we achieve the required accuracy? It depends on the parameters of the scaling law. The best scenario is (b) -- transfer gap Cis low and pre-training rate is high. In this case, increasing synthetic images eventually leads the required accuracy. In contrast, when transfer gap Cis larger than the required accuracy (c), increasing synthetic images does not help to solve the problem. Similarly, for low pre-training rate (d), we may have to generate tremendous amount of synthetic images that are computationally infeasible. In the last two cases, we have to change the rendering settings such as 3D models and light conditions to improve Cand/or, rather than increasing the data size. The estimation of andCrequires to compute multiple fine-tuning processes. However, the tell us whether we should increase data or change the data generation process, which can reduce the total number of trials and errors. 4 E XPERIMENTS 4.1 S ETTINGS For experiments, we employed the following transfer learning protocol. First, we pre-train a model that consists of backbone and head networks from random initialization until convergence, and we select the best model in terms of the validation error of the pre-training task. Then, we extract the backbone and add a new head to fine-tune all the model parameters. For notations, the task names of object detection, semantic segmentation, multi-label classification, single-label surface normal estimation are abbreviated as objdet ,semseg ,mulclass ,sinclass , andnormal , respectively. The settings for transfer learning are denoted by arrows. For indicates that a model is pre-trained by object detection, and fine-tuned by semantic segmentation. The experiments were conducted on an in-house cluster containing NVIDIA V100 GPUs. The total amount of computation was approximately 1700 GPU days (200 for image rendering, 1300 for pre-training, and 200 for fine-tuning). The code is available at . All the results including Figure 1 are shown as log-log We prepared four tasks: mulclass ,objdet ,semseg , and normal . We models, where backbones were ResNet-50, unless otherwise specified, and the head networks were customized for each task. Synthetic images for pre-training were generated by BlenderProc (Denninger et al., 2019), an image renderer that can handle several domain For rendering, we used the setting of the BOP challenge 2020 (Hoda Vn et al., 2020) as our default setting. We used 172 3D models, where ten objects appeared on average for each image. We applied texture randomization for walls and a floor, randomization for area and point lights, and randomization for the camera. In most cases, the models were pre-trained with 64;000images. We trained all models for the same fixed number of iterations depending on pre-training tasks and selected the best models for fine-tuning, which were validated by another 1000 synthetic images generated in the same We evaluated sinclass by ImageNet (Russakovsky et al., 2015), objdet by MS-COCO (Lin et al., 2014), and semseg by ADE20K (Zhou et al., 2016). The number of images used was 1% of each data set (roughly, 12;000for ImageNet, 1000 for COCO, and 200 for ADE20K). We fine-tuned the pre-trained models with these subsets of data for a fixed number of iterations and reported the error metrics for validation sets at the last iteration. The metrics were top-1 classification, mean mAP for MS-COCO, and mean IoU for ADE20K. These metrics take their values from 0 to 1, and we converted them into errors such as 1 - accuracy.4 Curve fitting: After obtaining the empirical errors ^L, we estimated the parameters of (1)by non- linear least squares in the log-log space. We solved the minimization problem a fixed fine-tuning data size sand pre-training data sizes ni= 2i1000 for data point index i= 0;:::; 6. In the experiments, we empirically encountered some We fixedD= 0:48by the median values of D's for all the settings and for each case. We explain this procedure with more details in Appendix D. 4.2 S CALING LAW UNIVERSALLY EXPLAINS DOWNSTREAM PERFORMANCE FOR VARIOUS TASK 1 shows the test errors of each fine-tuning task and fitted learning curves with Eq. (1), which describes the effect of pre-training data size nfor all combinations of pre-training and The scaling law fits with the empirical fine-tuning test errors with high accuracy in most cases. 4.3 B IGGER MODELS REDUCE THE TRANSFER of pretrain imagestest errormodel*r18 r34 r50 r101 r152 * of parameters 4: Effect of model size. Best viewed in color. Left: The scaling curves andobjdet!objdet cases. The meanings of dots and lines are the same as those in Figure 1. Right : The estimated transfer gap C(y-axis) versus the model size (x-axis) in log-log scale. The dots are estimated values, and the lines are linear fittings of them. We compared several ResNet models as backbones in to observe the effects of model size. Figure 4 (left) shows the curves of scaling laws for the pre-training data size nfor different sizes of backbone ResNet- x, The bigger models attain smaller test errors. Figure 4 (right) shows the values of the estimated transfer gap C. The results suggest that there is a roughly between the transfer gap and model size. This agrees with the scaling law with respect to the model size shown by Hernandez et al. (2021). 4.4 S CALING LAW HOLDS WITH MORE PRE -TRAINING IMAGES To check whether the same trend would be observed for larger data, we increased the number of images up to 1.28 million. All the settings were the same of the previous experiments, except we made the learning rate smaller for sinclass fine-tuning. The results (Figure 5) show that the empirical performance still continues to follow the scaling law when increasing the pre-training of pretrain imagestest 5: Results of million-scale pre-training. The models were pre-trained by objdet task and fine-tuned by objdet andsinclass 3000 10000 of pretrain imagestest texture w/o a C obj 6: Effect of synthetic image complexity. Best viewed in color. Left: Scaling curves of different data complexities. Right : Estimated parameters. The error bars represent the standard error of the estimate in least squares. 4.5 D ATA COMPLEXITY AFFECTS BOTH PRE -TRAINING RATE AND TRANSFER GAP We examined how the complexity of synthetic images affects fine-tuning performance. We controlled the following four rendering parameters: Appearance : Number of objects in each image; (max 10 objects). Light : Either an area and point light is random ized or fixed in terms of height, color, and intensity. Background : Either the textures of floor/wall are random texture : Either the 3D objects used for rendering contain texture ( w/) or not ( w/o). Indeed, the data complexity satisfies the following ordered relationships: single<multiple in appearance ,fix<random inlight andbackground , texture5. To quantify the complexity, we computed the negative entropy of the Gaussian distribution fitted to the last activation values of the backbone network. For this purpose, we pre-trained ResNet-50 as a backbone with MS-COCO for 48 epochs and computed the empirical covariance of the last activations for all the synthetic data sets. The estimated parameters are shown in Figure 6, which indicates the following (we discuss the implications of these results further in Section 5.1). *Data complexity controlled by the rendering settings correlates with the negative entropy, implying the negative entropy expresses the actual complexity of pre-training data. 4Although the cross-entropy loss is commonly used, several studies (Sharma & Kaplan, 2020; Bahri et al., 2021) show that the scaling laws also hold for 1 - accuracy. 5The object category of w/o is a subset of w/, andw/has a strictly higher complexity than rate correlates with data complexity. The larger complexity causes slower rates of convergence with respect to the pre-training data size. * Transfer gap Cmostly correlates negatively with data complexity, but not for object texture . As discussed in Section 4.1, we have fixed the value of Dto avoid numerical instability, which might cause some bias to the estimates of . We postulate, however, the value of Ddepends mainly on the fine-tuning task and thus has a fixed value for different pre-training data complexities. This can be inferred from the theoretical analysis in Appendix E.5: the exponent in the main factor s ofDdoes not depend on the pre-training data distribution but only on the fine-tuning task or the pre-training true mapping. Thus, the values of Dshould be similar over the different the correlation of preserves. 5 C ONCLUSION AND DISCUSSION In this paper, we studied how the performance on syn2real transfer depends on pre-training and fine-tuning data sizes. Based on the experimental results, we found a scaling law (1)and (2)that explain the scaling behavior in various settings in terms of tasks, model sizes, and data complexities. Further, we present the theoretical error bound for transfer learning and found our theoretical bound has a good agreement with the scaling law. 5.1 I MPLICATION OF COMPLEXITY RESULTS IN SECTION 4.5 The results of Section 4.5 has two implications. First, data complexity (i.e., the diversity of images) largely affects the pre-training rate . This is reasonable because if we want a network to recognize more diverse images, we need to train it with more examples. Indeed, prior studies (Sharma & Kaplan, 2020; Bahri et al., 2021) observed that is inversely proportional to the intrinsic dimension of the data (e.g., dimension of the data manifold), which is an equivalent concept of data the estimated values of the transfer gap Csuggest that increasing the complexity of data is generally beneficial to decrease C, but not always. Figure 6 (right) shows that increasing complexities in terms of appearance ,light, and background reduces the transfer gap, which implies that these rendering operations are most effective to cover the fine-tuning task that uses real images. However, the additional complexity in object texture works negatively. We suspect that this occurred because of shortcut learning (Geirhos et al., 2020). Namely, adding textures to objects makes the falsely easier because we can identify objects by textures rather than shapes. Because CNNs prefer to recognize objects by textures (Geirhos et al., 2018; Hermann et al., 2019), the may overfit to learn the texture features. Without object textures, pre-trained models have to learn the shape features because there is no other clue to distinguish the objects, and the learned features will be useful for real tasks. 5.2 L ESSONS TO TRANSFER LEARNING AND SYNTHETIC -TO-REAL results suggest the transfer gap Cis the most crucial factor for successful transfer the maximum utility of pre-training. Large-scale pre-training data can be useless when Cis large. In contrast, if Cis negligibly small, the law is reduced essentially to n, which tells that the volume of pre-training data is directly exchanged to the performance of Our empirical results suggest two strategies for reducing C: 1) Use bigger models and 2) fill the domain gap in terms of the decision rule and image distribution. For the latter, existing techniques such as domain randomization (Tobin et al., 2017) would be helpful. 5.3 L IMITATIONS OF THIS STUDY * In the experiments, the scale of data is relatively limited (million-scale, not billion). * We only examined ResNet as a network architecture (no there are various visual tasks, our study only covers a few of them. Extending our observations to other visual tasks such as depth estimation, instance segmentation, and as well as to other data domains such as language is future theoretical results assume several conditions that may contradict the actual setting in the experiments. For example, our theory relies on ASGD instead of vanilla SGD. Also, the task types are assumed to be identical for the pre-training and fine-tuning tasks. *In this study, we focus on finding a general rule of transfer learning, rather than improving on specific tasks. We used popular vision tasks such as classification and settings that is not designed to pre-train for the tasks. We expect to observe more performance gain with other tasks such as optical flow and elaborate in future thank Daisuke Okanohara, Shoichiro Yamaguchi, Takeru Miyato, Katsuhiko Ishiguro for and discussions in the early stage of this study. We also thank Masanori Koyama and Kenta Oono for reading the draft and providing detailed feedback. TS was partially supported by JSPS KAKENHI (18H03201), Japan Digital Design and JST CREST. KF was partially supported by JST CREST Allen-Zhu, Yuanzhi Li, and Yingyu Liang. Learning and generalization in networks, going beyond two layers. CoRR , abs/1811.04918, 2018. URL . Shun-ichi Amari, Naotake Fujita, and Shigeru Shinomoto. Four types of learning curves. , 4(4):605-618, 1992. Sanjeev Arora, Simon Du, Wei Hu, Zhiyuan Li, and Ruosong Wang. Fine-grained analysis of optimization and generalization for two-layer neural networks. In Proceedings of the 36th International Conference on Machine Learning , volume 97, pp. 322-332, 2019. Yasaman Bahri, Ethan Dyer, Jared Kaplan, Jaehoon Lee, and Utkarsh Sharma. Explaining neural scaling laws. arXiv preprint arXiv:2102.06701 , 2021. Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. margin bounds for neural networks. In Advances in Neural Information Processing Systems , volume 30. Inc., 2017. Jonathan Baxter. A model of inductive bias learning. Journal of artificial intelligence research , 12: 149-198, 2000. Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance of the IEEE/CVF International Conference on Computer Vision , pp. Borrego, Atabak Dehban, Rui Figueiredo, Plinio Moreno, Alexandre Bernardino, and Applying domain randomization to synthetic data for object category detection. arXiv preprint arXiv:1807.09834 , 2018. A. Caponnetto and E. De Vito. Optimal rates for regularized least-squares algorithm. Foundations of Computational Mathematics , 7(3):331-368, Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking for semantic image segmentation. arXiv preprint arXiv:1706.05587 , 2017. Wuyang Chen, Zhiding Yu, Shalini De Mello, Sifei Liu, Jose M. Alvarez, Zhangyang Wang, and Anima Anandkumar. Contrastive syn-to-real generalization. arXiv preprint arXiv:2104.02290 Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Ole- fir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. arXiv , Devaranjan, Amlan Kar, and Sanja Fidler. Meta-sim2: Unsupervised learning of scene structure for synthetic data generation. In European Conference on Computer Vision , pp. 2020. Simon Du, Jason Lee, Haochuan Li, Liwei Wang, and Xiyu Zhai. Gradient Descent Finds Global Minima of Deep Neural Networks. In Proceedings of the 36th International Conference on Machine Learning , volume 97, pp. 1675-1685, 2019. Simon S Du, Jayanth Koushik, Aarti Singh, and Barnabas Poczos. Hypothesis Transfer Learning via Transformation Functions. In I Guyon, U V Luxburg, S Bengio, H Wallach, R Fergus, S Vishwanathan, and R Garnett (eds.), Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017. URL . Simon S. Du, Wei Hu, Sham M. Kakade, Jason D. Lee, and Qi Lei. Few-shot learning via learning the representation, provably. arXiv preprint arXiv:2002.09434 , 2020. Li Fei-Fei, R Fergus, and P Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence , 28(4):594-611, 2006. doi: Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias and robustness. arXiv preprint arXiv:1811.12231 , 2018. Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature , 2(11):665-673, 2020. Georgios Georgakis, Arsalan Mousavian, Alexander C. Berg, and Jana Kosecka. data for object detection in indoor scenes. arXiv preprint arXiv:1702.07836 , 2017. Priya Goyal, Piotr Dollar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch SGD: in 1 hour. arXiv preprint arXiv:1706.02677 , 2017. Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imagenet pre-training. arXiv , 2018. Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B. Brown, Prafulla Dhariwal, Scott Gray, Chris Hallacy, Benjamin Mann, Alec Radford, Aditya Ramesh, Nick Ryder, Daniel M. Ziegler, John Schulman, Dario Amodei, and Sam Mc- Candlish. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701 L. Hermann, Ting Chen, and Simon Kornblith. The origins and prevalence of texture bias in convolutional neural networks. arXiv preprint arXiv:1911.09071 , 2019. Danny Hernandez, Jared Kaplan, Tom Henighan, and Sam McCandlish. Scaling laws for transfer. arXiv preprint arXiv:2102.01293 , 2021. Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is arXiv preprint arXiv:1712.00409 , 2017. Stefan Hinterstoisser, Olivier Pauly, Hauke Heibel, Martina Marek, and Martin Bokeloh. An annotation saved is an annotation earned: Using fully synthetic training for object arXiv preprint arXiv:1902.09967 , 2019. Tomas Hoda Vn, Vibhav Vineet, Ran Gal, Emanuel Shalev, Jon Hanzelka, Treb Connell, Pedro Urbina, Sudipta N Sinha, and Brian Guenter. Photorealistic image synthesis for object instance IEEE International Conference on Image Processing (ICIP) , pp. 66-70. IEEE, Hoda Vn, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Ji Vri Matas. BOP challenge 2020 on 6D object localization. on Computer Vision Workshops (ECCVW) , 2020. Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes imagenet good for arXiv preprint arXiv:1608.08614 , 2016. Marcus Hutter. Learning curve theory. arXiv preprint arXiv:2102.04074 , 2021. Arthur Jacot, Franck Gabriel, and Clement Hongler. Neural tangent kernel: Convergence in neural networks. In Advances in Neural Information Processing Systems 31 , pp. 8571-8580. Curran Associates, Inc., 2018. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 , 2020. Andrej Karpathy. Tesla ai day. , 2021. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, An- drei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming in neural networks. Proceedings of the National Academy of Sciences of the United States of America , mar 2017. ISSN 1091-6490. doi: URL . Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. arXiv , 2019. Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In on computer vision , pp. 740-755. Springer, 2014. Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 2117-2125, 2017. Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for In Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 3431-3440, 2015. Andreas Maurer, Massimiliano Pontil, and Bernardino Romera-Paredes. The benefit of learning. Journal of Machine Learning Research , 17(81):1-32, 2016. . Mehdi Mousavi, Aashis Khanal, and Rolando Estrada. Ai playground: Unreal engine-based data ablation tool for deep learning. In International Symposium on Visual Computing , pp. 2020. Yair , Takeo Kanade, and Yaser Sheikh. How useful is photo-realistic rendering for visual learning? arXiv preprint arXiv:1603.08152 , 2016. Alejandro Newell and Jia Deng. How useful is self-supervised pretraining for visual tasks? arXiv preprint arXiv:2003.14323 , 2020. Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. Norm-based capacity control in neural networks. In Proceedings of The 28th Conference on Learning Theory , pp. 1376-1401, 2015. Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro. Exploring Gen- eralization in Deep Learning. In Advances in Neural Information Processing Systems 30 , pp. 5947-5956, Nitanda and Taiji Suzuki. Stochastic gradient descent with exponential convergence rates of expected classification errors. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , volume 89 of Proceedings of Machine Learning Research , pp. 1417-1426, 2019. Atsushi Nitanda and Taiji Suzuki. Optimal rates for averaged stochastic gradient descent under neural tangent kernel regime. In International Conference on Learning Representations , 2021. Atsushi Nitanda, Geoffrey Chinot, and Taiji Suzuki. Gradient descent can learn less two-layer neural networks on classification problems, 2020. Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization , 30(4):838-855, 1992. Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and , 39(6):1137-1149, 2016. Jonathan S Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales. arXiv preprint arXiv:1909.12673 , 2019. Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) , 115 (3):211-252, 2015. doi: Sharma and Jared Kaplan. A neural scaling law from the dimension of the data manifold. arXiv preprint arXiv:2004.10802 , 2020. Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Proceedings of the IEEE on Computer Vision , pp. 2686-2694, 2015. Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting of data in deep learning era. In Proceedings of the IEEE international conference on computer vision , pp. 843-852, 2017. Taiji Suzuki. Fast generalization error bound of deep learning from a kernel perspective. In Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics , volume 84, pp. 1397-1406, 2018. Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. arXiv preprint arXiv:2108.10869 , 2021. Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. for transferring deep neural networks from simulation to the real world. arXiv preprint arXiv:1703.06907 , 2017. Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. arXiv preprint arXiv:1804.06516 , 2018. Nilesh Tripuraneni, Michael I. Jordan, and Chi Jin. On the theory of transfer learning: The importance of task diversity. arXiv preprint arXiv:2006.11650 , 2020. Colin Wei and Tengyu Ma. Improved Sample Complexities for Deep Neural Networks and via an All-Layer Margin. In International Conference on Learning Representations , 2020. URL . Jun Yang, Rong Yan, and Alexander G Hauptmann. Cross-Domain Video Concept Detection Using Adaptive Svms. In Proceedings of the 15th ACM International Conference on Multimedia , MM '07, pp. 188-197, New York, NY , USA, 2007. Association for Computing Machinery. ISBN 9781595937025. doi: URL Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio understanding of scenes through the ade20k dataset. arXiv preprint arXiv:1608.05442 , 2016. Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 633-641, 2017. Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin D. Cubuk, and Quoc V . Le. Rethinking pre-training and self-training. arXiv preprint arXiv:2006.06882 , T RAINING DETAILS A.1 O BJECT DETECTION We used Faster-RCNN (Ren et al., 2016) with FPN (Lin et al., 2017) as object detection models and ResNet (Goyal et al., 2017) as a backbone network of Faster-RCNN. We used the following training procedure: We trained the model using momentum SGD of momentum 0:9with weight decay of 104. The global batch size was set to 64when training ResNet18, and ResNet101. The batch size was set to 32when training ResNet152 to avoid out-of- memory errors. The batch statistics in batch normalization layers were computed across all GPUs. We used a base image size of 640640in the same way as YOLACT training (Bolya et al., 2019). We used mixed16 training to reduce the memory footprint. We also adopted random as data augmentation to images. The learning rate was set to 0:02, and we used the cosine decay with a warmup scheme. The warmup length is 120,000 images (3,750 iterations for ResNet152 and 1,875 iterations for other models). As for evaluation, we followed the standard settings in COCO dataset (Ren et al., 2016). We pre-trained the model with 14,400,000 images (450,000 iterations for ResNet152 and for other models). We used the models that achieved the best mmAP as the initial value used COCO (Lin et al., 2014) as the fine-tuning dataset. We trained the model with 1,440,000 images (45,000 iterations for ResNet152 and 22,500 iterations for other models) during fine-tuning. A.2 S EMANTIC SEGMENTATION We used DeepLabV3 (Chen et al., 2017) with the softmax cross-entropy loss as the model and ResNet50 (Goyal et al., 2017) as its backbone. The model the implementation in torchvision6. It should be noted that DeepLabV3 requires dilated ResNet as the backbone, which is not the case in object detection and classification tasks. Even though, the shapes of weight tensors of dilated ResNet50 exactly match those of non-dilated ResNet50; thus we can use the pre-trained weights of dilated and non-dilated ResNet50 learning procedure is based on the reference torchvision. We added an auxiliary branch based on FCN (Long et al., 2015) which takes conv4 of the backbone as the input. In the computation of loss function, the loss for the auxiliary branch is computed in the same way as for the main branch and is added to the overall loss after multiplying by the factor 0:5. The model was trained using momentum SGD of momentum 0:9with weight decay of 104. The global batch size was set to 32. The batch statistics in batch normalization layers were computed across all GPUs. During training, images were first resized so that the length of the shorter edge becomes an chosen from [5200:5;5202], then horizontally flipped with probability 0:5, finally randomly cropped to 480480. The learning rate (LR) was decayed according to the polynomial LR schedule of rate 0:9and initial LR of 0:02. For the parameters of the auxiliary classifier, the LR was multiplied by 10. The evaluation was performed once every 3;125iterations (almost in full ADE20K). In the evaluation, images were resized so that the length of the shorter edge becomes 520. In pre-training, we trained which roughly equals 200epochs in full ADE20K (Zhou et al., 2017). We used the model that achieved the best mIoU as the initial value of fine-tuning. We pre-trained models using our synthetic datasets. When training with them, backgrounds (points at which no foreground objects were present) were also considered to be a separate class in 6https://github.com/pytorch/vision 7https://github.com/pytorch/vision/tree/master/references/segmentation 15Preprint We used the ADE20K (Zhou et al., 2017) datasets as the fine-tuning target. In fine-tuning, we trained the model for 18,750 iterations, which correspond to 30epochs of full ADE20K. The metric was mIoU score. A.3 M ULTI -LABEL used ResNet (Goyal et al., 2017) with binary cross-entropy used as the loss function in We used the following training procedure: We trained the model using momentum SGD of momentum 0:9with weight decay of 104. The batch size was set as 32per GPU, thus 256in total. We trained the models for 112,500 iterations. The input size of the images was simply resized to 640640. We adopted random horizontal flipping as data augmentation to images. The learning rate was set to 0:1. The cosine decay with a warmup scheme was used. The warmup length was 120,000 images. The evaluation was performed once every 120,000 images. In the evaluation, the image size was the same as that used during training, and data augmentation was not used. We used the mAP score as the metric. A.4 S INGLE -LABEL in multi-label classification, we used ResNet. The softmax cross-entropy was used as the loss function of the single-label classification. The learning procedure is based on (Goyal et al., 2017). However, we used cosine decay for the learning rate scheduling. A.5 S URFACE NORMAL ESTIMATION As in semantic segmentation, we used DeepLabV3 (Chen et al., 2017) as the model for surface normal estimation. The model configuration and the training procedure were exactly the same as in semantic segmentation, except for the following of output channels was changed to 3, each of which corresponds to the 3axes of the normal vector, * Initial LR was changed to 0:04, *Length of pre-training was 200,000 iterations, which corresponds to 100epochs in our Random flipping was not performed during the data augmentation, and *Loss function was the average of the value, 1:0n
^ n, which was computed for each valid pixel where nis the ground-truth normal vector and ^ nis the model output (after L2 S YNTHETIC DATA DETAILS The data generation strategy is based on the on surface sampling setting in the BoP In this setting, the sampled objects will be spawned in a cube-shaped room with one point light and one surface light. As the objects to be spawn, we used all the BoP object sets, i.e., LM, T-LESS, ITODD, HB, YCB-V , RU-APC, IC-BIN, IC-MI, TUD-L, and TYO-L.9. There are 173 objects in total. After generated a random scene (position of objects, lights, etc), we took 10 pictures by 10 different camera poses. This means that, if we have 10K images in total, there are 1K unique scenes, and 9K images are inflated by just changing the camera angle and position. To control the data complexity, we selected four attributes in the generation strategy and prepared two options for each. Appearance controls how many objects are generated in a room in a single scene. For each scene, we randomly select ten objects for the multiple setting and one object for the single setting. Light controls the light sources. In the random setting, the color, height, and strength of lights are randomized. In contrast, in the fixsetting, they are all controls the texture of the room, i.e., floor and walls. In the random setting, we assign a random PBR material from the CC0 and we selected one carpet texture for the texture controls the object set to be used. The BoP object set is consists of several types of object sets as described above. Among them, T-LESS and ITODD consist of objects and they do not have textures and colors.11For the w/osetting, we only used such texture-less objects to be sampled. In contrast, we sample all the 173 objects include T-LESS and ITODD in the w/setting. We generated eight variations of datasets by changing these attributes, which were used in the experiments of Section 4.5. Figure 7 shows the example of generated images with the value of and ITODD contain 30 and 28 objects, texture: w/ appearance: multiple light: random object texture: w/o appearance: multiple light: random object texture: w/ appearance: multiple light: fix object texture: w/ appearance: multiple light: fix background: fix object texture: w/o appearance: multiple light: fix object texture: w/o appearance: multiple light: fix background: fix object texture: w/ appearance: single light: fix background: fix object texture: w/o appearance: single light: fix background: fix Figure 7: Example of generated * *objdet semseg mulclass normala task Figure 8: The estimated values of the pre-training rate and the transfer gap Cin the (as the same as Figure 1). The error bars present the standard error of the estimates in the least of pretrain imagestest errorfinetune ratio of finetune imagestest errorpretrain size (K)* ** ** ** 9: Empirical and fitting results for various pre-train and fine-tune data sizes in . All curves are fitted using the full law (2). Best viewed in color. Left: Effect of pre-training data size (x-axis) for fixed fine-tuning data sizes. Right : Effect of fine-tuning data size (x-axis) for fixed pre-training data sizes. C A DDITIONAL EXPERIMENTS C.1 E STIMATED PARAMETERS IN THE CROSS -TASK SETTING Figure 8 shows the estimated parameters ( ;C) at the experiments described in Section 4.2. Note that the result of atnormal!semseg is omitted because its estimated value is highly unstable (the standard deviation is larger than 1). C.2 F ULL SCALING LAW COLLECTIVELY RELATES PRE -TRAINING AND FINE -TUNING DATA SIZE Next, we verify the validity of the full scaling law (2). In the setting with ResNet-50, we changed the fine-tuning data size from 2% to 64% of the ImageNet.12We then fitted all results by a single equation (2)to estimate the parameters except the irreducible loss E 12ImageNet contains a class imbalance problem. If we use 100% of the ImageNet, we cannot provide the same sample size per class. To eliminate the effect of class imbalance, we made the sampling ratio to keep the balance up to 64% and excluded the case of assumedE= 0from the preliminary results in Figure 2). The results in Figure 9 show that all empirical test errors are explained remarkably well by Eq. (2), which has only four parameters to fit in this case. The estimated parameters are = 0:544,= 0:322, = 0:478, and= 41:8. C.3 L INEARIZED RESULTS The transfer gap Cin(1)causes a plateau of the scaling law. Conversely, if we subtract the estimated Cfrom the results, we must be able to recover the power-law scaling. To confirm this, we subtracted the estimated Cfrom the empirical errors ^Lof the previous results. Figures 10-12 show the modified version of scaling law fittings. Overall, the empirical errors behave linearly along with the term Dn. Note that, in mulclass!semseg andnormal!semseg , a few negative after subtracting C, and these points are not semseg mulclass task # of pretrain imagestest errorfinetune task Figure 10: The linearized version of Figure 1. * ** ** *** ** ** * of pretrain imagestest 11: The linearized version of Figure 4. D E MPIRICAL DEPENDENCY BETWEEN ANDD WhenCis non-zero, the joint estimation of Dandin(1)have an issue of numerical stability due to the small number of observations and noise, which can cause high dependence on each other. Figure 13 (left) shows the curves of (1)withC= 0:5, where the solid red curve is D= 0:5;= 0:4 and the dashed blue line is D= 1;= 0:5. We see that both curves are almost 3000 10000 of pretrain imagestest texture w/o texture Figure 12: The linearized version of Figure 13: Examples of parameter dependency. a largen. Figure 13 (right) shows the actual landscape in terms of andDof the nonlinear least- squares at objdet!objdet , the bright areas indicate the fitting loss is small. We see that there is a quadratic-like trajectory in the landscape, which implies the solutions are somehow landscapes were observed for other tasks (Figure 14). objdet semseg mulclass normalobjdet semseg sinclass 0.0 0.5 1.00.0 0.5 1.00.0 0.5 1.00.0 0.5 task aDfinetune task Figure 14: Loss landscapes of curve avoid this issue, we fixed a common Dfor all the cases and estimated for each. To determine D, we used the following procedure. First, we prepared two global parameters ^;^Dand set 0:5as their initial values. Then, we fitted the curves by two equations, Next, we computed the median of and substituted them into ^. We did the same forDand^D. After a few iterations, we got a converged value of ^D= 0:48. In the experiments, we used the value for Dand fixed it. E D ETAILS OF THEORETICAL ANALYSIS This section gives details of the theoretical discussions given in Section 3.2. For the analysis of learning and generalization bound, we use the techniques developed recently by Nitanda & Suzuki (2021). There are many works on the generalization of neural networks. To list a few, Neyshabur et al. (2015), Neyshabur et al. (2017), Bartlett et al. (2017), Wei & Ma (2020), and Suzuki (2018) analyze the generalization of neural networks based on complexity bounds. These however, do not consider an algorithm of learning, such as stochastic gradient descent (SGD). Recently, learning dynamics of neural networks has been analyzed based on Neural Tangent Kernel (NTK) Jacot et al. (2018) and global convergence of wide neural networks has been revealed Allen- Zhu et al. (2018); Du et al. (2019). Based on the NTK framework, Arora et al. (2019) and Nitanda et al. (2020) showed a generalization bound of the gradient descent learning of neural networks. More recently, Nitanda & Suzuki (2021) focused the functional space given by NTK and showed that the two-layer neural network with averaged SGD achieves the minimax optimal rate with respect to the function class used in the standard theory of function estimation with kernels. We employ the method of Nitanda & Suzuki (2021), which is the most suitable for our analysis of transfer learning: it enables to examine the dependence on the initial parameter in the learning, and avoids the assumption of a positive margin of eigenvalues used in Arora et al. (2019) and Nitanda et al. (2020). E.1 P ROBLEM SETTING In the pre-training, the task is to learn the target function 0withT0training data (~xi;~yi)T0 i=1, where yi=0(xi), while in the fine-tuning phase, the network is initialized by the final parameter learned by the pre-training, and the whole parameter is updated in the training. We assume that the target function in the fine-tuning is given by '(x) =0(x) +1(x); data is given by In this setting, the goal of the fine- tuning phase will be to learn the additional function 1mainly. Note that, for simplicity of analysis, we assume noiseless training data, i.e., we assume the supervised signal yjis given by a ofxj, but extension to more general cases is not difficult as discussed in Nitanda & Suzuki (2021). In the analysis, the data are assumed to satisfy x2Rd,kxk2= 1 andy2[0;1]. The distribution of the input data xis denoted by X, and the same for the pre-training and fine-tuning. For tractable theoretical analysis, we consider a simple scalar-valued two-layer neural network units: g(x) (5) We omit the bias term, but with obvious modification, it is not difficult to include it (see Nitanda & Suzuki (2021)). As in Nitanda & Suzuki (2021), we consider the averaged stochastic gradient descent (ASGD), where one training sample is given at every time step for the stochastic gradient descent as in online learning, and all the parameters in the time course are averaged after the final time step for the inference, that is, after proceeding up to prescribed T(T=T0orT1) time steps, the parameter to be used in the inference is given by (T):=1 T+ 1TX t=0(t): (6) The final network uses this averaged parameter, i.e., the final network is given by parameter is initialized as (0)= For the pre-trainig, each given by the uniform distribution on the unit sphere. As in Nitanda & Suzuki (2021), a(0)are initilized as 1or1so thatg(0)= 0. As explained before, the initial parameter of the fine-tuning is the same as the averaged parameter of the pre-training (T0) pre. The objective function to minimize for the pre-training and fine-tuning is given by the following regularized empirical risk: L() (7) whereis the regularization coefficient, which is a hyperparameter. The values of in the pre-training and fine-tuning can be different, and denoted by 0and1, respectively. Note that the Eq. (7) is not the most common where kk2 2is used for the applied in fine-tuning, however, the above regularization can be interpreted as elastic (Kirkpatrick et al., 2017), which prevents forgetting the pre-trained parameters. We consider online learning, in which at every step t(tT1), one datum xiis sampled from and pre-training) is used to update the parameter according to the gradient descent: (t+1)= (8) whereis a learning rate. More (9) E.2 N EURAL TANGENT KERNEL In the theoretical analysis, the neural tangent kernel (NTK, Jacot et al., 2018) is used for approximat- ing the dynamics of ASGD by a linear functional recursion on the corresponding function space. The NTK of this model is given by k1(x;x0) (10) The positive definite kernel k1naturally defines a reproducing kernel Hilbert space (RKHS), which is denoted byH1. The integral operator 1onL2(X)is defined (11) It is known that 1admits s= s s; (12) where sis an eigenvector with k skL2(X)= 1 and 1 2::: > 0are eigenvalues in descending order. Mercer's theorem tells that k1has an =1X s=1 s s(x) s(x0); where the convergence is understood as in L2(X)for general, and absolutely and uniformly if Xis a uniform distribution on a compact set. E.3 A SSUMPTIONS For theoretical analysis, we make the following assumptions. For an operator , the range of is denoted The activation function is differentiable up to the second order, and there exists C > 0 such andj(u)j1 supp There exist 1=2r0;r11such There exists >1such that `= (`). As in Nitanda & Suzuki (2021), Assumption (A1) assumes that the activation is differentiable in this paper. In Nitanda & Suzuki (2021), however, they have developed a theory on how to extend the results to the case of ReLU by approximating it with a smooth function. It is well known that Assumption (A4) specifies the complexity of the hypothesis class H1(Caponnetto & De Vito, 2007); a faster eigen-decay (large ) implies the small complexity of the class. The assumption (A3) controls the smoothness of the target functions 0;1. In fact, the functions are included inH1, When a function fhas the expansion f=P `a` `, the assumption A function with a larger ris smoother, which is easier to learn. It is known (Caponnetto & De Vito, 2007; Nitanda & Suzuki, 2021) that and rare the two basic parameters to control the convergence rate of generalization attained by for a large sample size. Under the assumptions (A3) and (A4), given Ni.i.d. +iwith additive noise iN(0;2), the kernel ridge regression ^'with the regularization parameter the any function '0with'02R(r), and it is known this rate is optimal. In the sequel, when 2R(r)and= r , we writekrk:=k k. E.4 G ENERALIZATION BOUND The dynamical behavior of pre-training can be discussed exactly in the setting of Nitanda & Suzuki (2021). Letb0be the result of pre-training, i.e., b0:=g(T0) pre, where (T0) preis the averaged parameter by ASGD. By optimizing the regularization parameter 0, Corollary 1 in Nitanda & Suzuki (2021) shows that for sufficiently large T0, with a choice of high probability, where cis a universal constant and Mcan be arbitrarily small for a large M. As discussed in Section E.3, it is known (Caponnetto & De Vito, 2007) that the rate the minimax optimal rate with respect to T0over the class specified by r0and. In fine-tuning, the initial parameter is given by (0)=(T0) pre, and ASGD is applied with (xt;yt)for t= 1;:::;T 1. By extending Theorem 1 in Nitanda & Suzuki (2021), we can derive a generalization bound in the following theorem. Recall that the regularization coefficient and learning rate of fine-tuning are denoted by1and1, 2. Suppose Assumptions (A1)-(A3) hold. After pre-training that gives Eq. (13), fine- tune the network by Eq. (9) with a learning rate 1and regularization coefficient 1that +1)11. Then, for any >0,2(0;1), andT12N, there exists M02Nsuch that for any MM0, the following bound holds with probability at least 1over the random initiailzation of 1 1 +k'k2 H1+ the result of pre-training and ci(i= 0;1;2;3;4)are universal constants. The termis arbitrarily small for a large value of M, i.e., wide network. The proof of Theorem 2 will be given in Section E.6. E.5 A NALYSIS OF CONVERGENCE RATES We consider the rates of the generalization bound for respect to T1. As typical cases, we assume . The dominant terms in Eq. (14) may vary according to the configurations of 1and1with respect to T1. We will show the rates in some settings that are relevant to transfer learning. First, note that under Assumption (A4), the factor given by (Caponnetto & De Vito, (15) By neglecting , the terms in Eq. (14) thus have the following (a1)2r1 1; (b)T1 11 1R0; of constant rate with respect to T1, but explicitly shown for the later the terms (c) and (d) are of smaller rate than (b). Likewise, (f) and (g) are smaller than (e). The candidates of dominant terms are thus (a0), (a1), (b), (e), (g), and (h). E.5.1 L ARGE REGULARIZATION COEFFICIENT In transfer learning, it is reasonable to use strong regularization in fine-tuning, which encourages the parameters to stay close to the initial value that is obtained in the pre-training. In this subsection, we consider the case where 1is larger than T 2r1+1 1 , which would be the optimal rate if the network was trained with random initialization (see Nitanda & Suzuki (2021)). If 1=T 2r1+1 1 was taken, it is easy to see that the influence of pre-training would not appear explicitly in the convergence rate. In the sequel, we write FG(T1)if there area>0andTsuch thatFaG(T1)for allT1T. In this notation, we (17) The is given by equating the rates of (a1) and (h). Therefore, under the assumption of Eq. (17), the rate (a1) is larger than (h), and thus it suffices to consider (a0), (a1), (b) and (e) as the candidates of dominant terms. Note that, if 1!0, the terms (a0) and (a1) decrease, while (b) and (e) increase to infinity. We will discuss below the possible cases of dominant terms under the assumption Eq. (17). In the analysis, although the error of the pre-training R0is regarded as a constant, we yet wish to consider the dependence of the fine-tuning result on R0. We thus set the regularization coefficient 1 dependent on R0, and show that in all the cases, the generalization bound takes the (18) where >0is a constant. As Eq. (13), the factor R 0can be bounded from above largeT0. As a result, we are constants. As we will see, the exponents anddepend (19) accords with the bound in Theorem 1. In the sequel, we use 0for the learning rate such that 1=T 1. Case I: Small learning rate T112 11.In this case, (e)(b). We equate (a0) or (a1) with (e) to obtain1for achieving the best possible upper bound of the two , (a1) is larger than (a0). By equating (a1) and (e), we find that the best choice of further dependence on R0. To determine , we assume that R0is a small value, and consider the rate of (a1) and (e) with respect to R0after plugging the above 1to them. By equating the rates of (a1) and , the best possible rate of R0is attained by = 1=(2r1+ 2) . The dominant rate of Eq. (16) is (20) attained (21) We need to identify the conditions on 1to meet the requirements. The condition 1!0is equivalent to <1. There are two other conditions: 1T 2r1+1 1 andT112 11. GivenR0is of constant rate, the former is equivalent to which results in r1+ 1 2r1+ 1: The latter condition is equivalent to 121 r1+10, which is r1 2r1+ 1: It is not difficult to see r1 2r1+ 1>r1+ 1 2r1+ 1 for>1. As a result, the condition on is r1 2r1+ 1 <1: taken to satisfy this condition, the optimal rate of 1is given by Eq. (21). Finally, the resulting generalization bound is given this case, (a0) is of larger rate than (a1). By a similar argument to (I-A), with (24) The generalization bound is given (25) The condition on is maxnr0 2r0+ 1;(2r1r0)+ 1 2r1+ 1o  <1: <1 1: Generalization bounds in various II): large learning rate T112 11.Next, we consider the case where the learning rate 1 is large so that T112 11, which includes the constant 1. Under this condition, (b) is of larger rate than this case, (a1) is of larger rate than (a0). A similar argument to (I-A) (27) and the generalization bound is given (28) The conditions are 1T 2r1+1 1 andT112 11. The former condition always holds for >1, and the latter is equivalent to r1 2r1+1. The resulting condition on is 0<r1 2r1+ 1: this case, (a0) is of larger rate. (30) the generalization bound is given (31) The conditions 1T 2r1+1 1 andT112 11are respectively we 1; r 0<r1r0+1 2; (32) In summary, the generalization bounds in various conditions are summarized in Table 1. E.6 P ROOF OF THEOREM 2 The proof of Theorem 2 is based on the application of the theory in Nitanda & Suzuki (2021) to the fine-tuning phase, adapting the initialization given by the result of pre-training b0. In the sequel, we focus on the fine-tuning with T1samples with yt='(xt). Recall that '(x) =0(x) R EFERENCE ASGD ONRKHS We use a surrogate sequence of functions in an RKHS for the proof. Let kMbe the random of the TNK k1, i.e., kM(x;x0) (33) where (br)M r=1is i.i.d. random sample from the uniform distribution on the unit sphere Sd1. The associated RKHS is denoted by HM. A reference ASGD is defined by the following update rule of functions in the RKHS HM: g(t+1)= (t= 0;:::;T1) (34) with the initialization given by g(0):=b0. The average is taken at the final 1T1X t=0g(t): (35) By considering continual learning of pre-training and fine-tune, a slight modification of (Nitanda & Suzuki, 2021, Propososion A) derives the following 3. Assume (A1) and (A2). Suppose that 11<1. Then for any T12Nand >0, there that during the fine-tuning (36) holds for any proposition shows that, if we use a very wide network, the learning of ASGD in the parameter space can be approximated by the reference ASGD on the RKHS with negligible error. The generalization bound will be given by the following (37) in which the first term of the right hand side is bounded by Proposition 3 with an arbitrary small value for largeM. The second term will be discussed in the next C ONVERGENCE RATES OF REFERENCE ASGD In this section, we write simplicity. The covariance operators respectively, are defined the adjoint; The regularized target functions (i) M;(i= 0;1) are defined by (i) M;:= (M+I)1Mi (i= 0;1): (39) (i) 1;is defined similarly with 1. Note that'M;:= we decompose

2108.11018

Kohei Hayashi

Hiroaki Mikami, Kenji Fukumizu, Shogo Murai, Shuji Suzuki, Yuta Kikuchi, Taiji Suzuki, Shin-ichi Maeda, Kohei Hayashi

A Scaling Law for Transfer: How Much Is Your Pre-training Effective?

cs.LG cs.CV

http://creativecommons.org/licenses/by/4.0/

transfer learning is a framework in which a dataset is used to pre-train a model to improve its performance on real vision tasks. The most significant advantage of using synthetic images is that the ground-truth labels are automatically available, enabling of the data size without human cost. However, synthetic data may have a huge domain gap, in which case increasing the data size does not improve the performance. How can we know that? In this study, we derive a simple scaling law that predicts the performance from the amount of pre-training data. By estimating the parameters of the law, we can judge whether we should increase the data or change the setting of image synthesis. Further, we analyze the theory of transfer learning by considering learning dynamics and confirm that the derived generalization bound is consistent with our empirical findings. We empirically validated our scaling law on various experimental settings of benchmark tasks, model sizes, and complexities of synthetic images.

approach is an effective way to identify colon oscopy features from EHR -extracted clinical notes . The Bio-Bi-LSTM -CRF model creates new opportunities to identify patients at risk for colon cancer and study their health outcomes. Keywords : Neural Network ; Machine Learning ; Natural Language Processing ( NLP) ; Text Mining ; Sentence Classification ; Colorectal Cancer ; Clinical Information . Introduction The amount of c omplex and varied information ( e.g., structured, semi -structured, and unstructured) has grown dramatically during the past century in numerous sectors , including education [1], media [2], power [3], and healthcare [4]. In healthcare , electronic health records (EHR s) contain information in both unstructured and structured formats on patient health, disease status , and care received , which can be useful for epidemiologic and clinical research [5]. In EHRs , unstructured data is documented without standard content specifications, often recorded as free text , and structured data is generally entered into discrete data fields with standardized responses or parameters (e .g., age, weight) [6]. Compared to the structured data within the EHR, free -text (unstructured data) data often conveys more granular and contex tual information of clinical events and enhances communication between clinical teams [7]. As such, extracting information from these free-text document resources has considerable potential value in supporting a multitude of aims, from clinical decision support and quality care improv ement to the secondar y use of clinical data for research, public health, and activities [8]. Natural language processing (NLP) can be an efficient way of automatically extracting and arranging this information, making NLP a potentially pivotal technology for enabling quality measurement from EHR data. Background Colonoscopy is a well-established procedure used to detect and prevent colorectal cancer (CRC) , the third most common cancer in men and women , which accounts for almost 10% of all cancer - related deaths in the United States [9]. There is evidence that colonoscopy use has risen significantly in recent years, primarily because of increased colonoscopy screening rates [10]. Understand ing the dynamics of the changes in colorectal polyps and cancer involving sizes, shapes, and the number of polyps is critical because these are associated with the efficacy of colorectal screening strategies. Manually extracting the detailed findings of a colonoscopy (e.g., number of polyps, polyp sizes, and polyp locations) from a colonoscopy procedure report and connecting with the pathology report (e.g., the histology of polyps found during colonoscopy) is quite time-consuming [14]. Extracting information entity identification from the clinical free text requires scalable techniques such as NLP, machine learning , and neural networks to convert clinical free text into a structured variable for text mining and further information extraction [15- 17]. Colonosc opy named entity identification has drawn a considerable amount of research in recent years, and various techniques for identifying entities have been suggested [22]. Conventional methods such as Decision Tree, Support Vector Machines (SVM), Hidden Markov Model (HMM), and Conditional Random Field (CRF) are either rule -based or dictionary -based, which usually need to be manually formulated and for the current datasets only. When the dataset is updated or replaced, updating the rules contributes to high machine costs. This will cause a low recall if the original rules or dictionary are used. The latest advanced techniques are data -driven, including the methods of machine learning and deep learning. In particular, the models of Bi -LSTM -CRF (Bi - Long Short -Term Memory and Conditional Random Field) have been used successfully in the biomedical text to achieve better results and represent some of the methods that are most comm only used [23]. Existing work regarding NLP in connection with CRC and colonoscopy is limited in scope and quantity. Adapting a clinical information retrieval system , Named entity recognition (NER) , identified pathology reports consistent with CRC from an electronic medical record system [26]. An application was proposed to automatically identify patients in need of CRC screening by detecting the timing and status of colorectal screening tests mentioned in the electronic clinical narrative documentation [2, 14, 26, 27] . Neither of these studies assesses the colonoscopy procedure's detailed findings, such as the number, shape, and size of po lys, due to the problem of entity boundary extraction with different entity lengths and name d entity recognition. Understanding the dynamics of the changes in colorectal polyps and cancer, involving sizes, shapes, and the number of polyps, is critical because these are associated with the efficacy of colorectal screening strategies [28]. Named entity recognition (NER), also known as entity identification, entity chunking, and entity extraction, is a process of information ext raction that helps to locate and classify named entities mentioned in unstructured and structured text into pre -defined categories such as medical codes, person names, etc. NER in the medical domain is more complicated for two key reasons [24]. First, because of non -standard abbreviations, acronyms, terminology, and several iterations of the same entities, several entities seldom or sometimes fail to appear in the training dataset. Second, text in clinical notes is noisy due to shorter and incomplete sentences and grammat ical and typographical errors. Additionally, clinical NER for multiple facilities is more difficult due to potential inter - facility variation compared with single facility data. On the one side, clinical reports involve a paragraph border , incomplete sent ences , rendering it impossible to identify complex characters , and sentence boundary detection [25]. A dependency of NLP is t he sentence boundary detection (SBD) task is to identify the sen tence segments within a text [19]. Good segmentation of sentences will not only improve translation quality but also reduce latency in colonoscopy reports. Most of the existing such as SBD systems SATZ - An Adaptive Sentence Segmentation System [20], proposed by Palmer and Hearst (1997), Disambiguation of Punctuation Marks ( DPM ), and Automated Speech Recognition (ASR ) system are highly accurate on standard and high -quality text, but extracting sentences or detecting the boundary of sentences from a clinical text in colonoscopy reports is a challenging task because clinical language is characterized, in addition to peculiarities like misspellings, punctuation errors and incomplete sentences, by an abundance of acronyms and abbreviations [21]. Neural emb edding methods such as word2vec [29] and GloVE [30] have been widely explored over the past five years to construct low -dimensional vector representations of words and passages of text and entity boundary extraction . Neural embeddings are a common group of low-dimen sional vector representation method s, phrases, or text (50 -500 dimensions) whose values are assigned by a neural network functioning as a black box . However, it is n't easy to view these dimensions in a meaningful way in clinical notes . It takes extensive training and tuning of several parameters and hyperparameters to construct neural embeddings. This paper describes and evaluates a method for identifying clinical entities and entity boundary extraction from the CRC clinical report. Given a sequence of words, our model represents each word, phrase , or text using a concatenated low -dimensional vector of its corresponding word and character sequence embedding. A near -comprehensive vector representation was built of words and selected bigrams, trigrams , and abbreviations. The word vector and dictionary features are then projected into dense representations of the vectors. These are then fed into the Bidirectional Long Short -Term Memory (Bi -LSTM) to capture contextual features. Lastly, a Conditional Random Field (CRF) is used to capture dependencies betw een adjacent tags. With the ability of Bio-Bi-LSTM -CRF, we capture detailed colonoscopy -related information within colonoscopy records like indicators, the number of polyps, the size of polyps, and the polyp's location, and the procedure and polyp histolog y like adenoma and tubulovillous and adenocarcinoma. Section Methods explains the proposed Bio-Bi-LSTM -CRF in depth. Section Model Evaluation explains about strategies for model evaluation Section Results presents the experiments and results. Section Discussion discusses several important issues of the proposed model. Section Strengths and Limitations explains about str engths and limitations of the model and f inally, Section conclusion concludes the current paper and points out potential future work. Method s Dataset and Annotation A total of 44,405 de-identified colonoscopy reports , which represent all colonoscopies between 2002 and 2012 on 36,537 subjects, were extracted from four Veterans Affair Medical Centers (Figure 1). These colon oscopy report s contain sections such as indications and findings that contain text input by a colonoscopist. The exact sections can vary between facilities. Only sections that contain text with information on the indication o f polyps found during the colonoscopy were retained for analysis. Pathology reports include sections such as impressions and findings with text input by a pathologist . Only sections that contain text with information on polyp pathology were retained for analysis. Manual annotation of colonoscopy reports provided a gold standar d to benchmark th e automated methods. A total of 4,000 colonoscopy reports were randomly selected (1,000 from each facility ) to annotate sentences. We used 80 percent of the reports as training data , 10 percent as data validation set for hyperparameter tuning , and the remaining 10 percent as test data. Figure 1. Use of colonoscopy reports for training, validating , and testing automated text extraction methods. The annotation processes are described in detail [31]. Briefly, two annotators manually marked the sentences devised a list of lexical units (trigger phrases) and corresponding elements (attributes) for each sentence. Based on the existing literature, three common phenotyping tasks were selected. The fi nal list is shown in Table 1. Table 1. List of the variable s of interest from colonoscopy reports Lexical Units Description The main components of the entity extraction pipeline are depicted in Figure 2 . The pipeline includes (1 ) data cleaning and pre-processing , (2) sentence boundary detection , (3) vector representation , and (4) the entity extraction model . We used python 3.7 and the open -source packages NLKT, sci-kit-learn , and Tensor Flow that provide various NLP methods, machine learning metho ds, and neural network module s as described in depth in the sections below. Indicators Reason or indications for colonoscopy examination : Abnormal Computed Tomography; Anemia; Diarrhea; Dyspepsia; Endoscopy; Foreign body removal; Hematochezia; History of colon cancer; Melena; Modification of bowel habits; Neoplasia; Personal history of colonic polyps; polypectomy site; Presence of fecal occult blood; Screening; Screenin g and surveillance for colon neoplasia; Sigmoidoscopy; Surveillance of patients with neoplastic polyps; Treatment of bleeding from such lesions as vascular; malformation; Ulceration; and Weight Loss Polyp Descriptions The d iameter of the polyp : Small (1-4mm ), Medium (5-9mm ), or Large (>10mm ) Morphology of polyp : sessile, pedunculated, mass The n umber of polyps : either a numerical value or qualitative description such as many, several, numerous, etc. Location Proximal : Ileocecal valve, Cecum, Hepatic flexure, Ascending, Transverse, and Splenic flexure Distal location : Descending, Sigmoid, Rectosigmoid , and Rectum Polyp histology Adenoma , Tubulovillous, Villous , High -grade dysplasia, Adenocarcinoma Figure 2. Model architecture. There are four parts to our model : (1) sentence boundary detection, (2) vector representation , (3) entity extraction model, and (4) rule update. Data cleaning and Preprocessing We pre-processed the text reports to remove all special characters (e.g., comma s in a list). Misspellings in our data set included typographical errors (e .g., cacel instead of cecal ), phonetic errors that could be associated with lack of familiarity with medical terms ( e.g., byopsi and methastasis), and everyday language errors (e .g., hooooooot snare). The correction of spelling error s from clinical entries using the MetaMap and auto spell checker by Google 's query suggestion service. Sentence Boundary Detection Sentence segmentation is no t a simple task in clinical data. Automatically segmenting clinical free text is an essential first step to information extraction. There are many sentence boundary detection (SBD) packages that mainly relied on punctuation marks and wide spaces. In the most text, sentence boundaries are indicated by a period , but colonoscopy reports frequently contain inter- sentence periods (e.g., Dr. Patil or size of 3.5 cm or No. of polyps: ) contrasting with typical biomedical text that usually had concepts from standard terminologies. Theoretically, boundary failures can result from standard medical terminologies. To solve this problem , we applied a machine learning approach to identify the sentence boundaries. 1 To perform sentence boundary detection, it is essential to represent the input sentences with suitable tags. S .F. (S=beginning of a sentence, F= ending of the sentence) format is used tagging in sent ence boundary detection. We used the features and keywords listed in Table4 to identify and tag the sentence boundaries. We used a decision tree for sentence boundary detection classification. This approach is constructed by applying the C4.5 algorithm. The e xample is shown in Textbox 1 . Textbox 1. Example of clinical note. The clinical note extracted sentences containing essential lexical units. To avoid duplicate sentences and improve sentence diversity, the sent ences were sorted by TF -IDF cosine distance. The sentences were manually de -identified and checked by automatic sentence boundary detection . Vector Representation Its complexity characterizes information on clinical notes in vocabulary and morphological richness. Therefore, using the recognition of a named entity to capture morphological information from complex medical terms can avoid vocabulary problems and compensate for traditional embedding words . The name entity reco gnition of c linical colonoscopy notes is usually regarded as a sequence labeling task. Given a sequence of words in a clinical notes sentence, we label each word in the sentence with a BIEOS ( Begin, Inside, End, Outside, Single) tag scheme. For example, in Textbox 2 below, The mass was circumferential and Estimated blood loss was minimal sentences are not significant. The widely used annotation modes for named entity recognition training data are BIO, BIEO, and BIESO, where B is the begin ning of an entity, I reflect the center of an entity, E is the end of an entity, S is a single entity , and O is not an entity. Textbox 2. Example output after sequence labeling task. In this study, we create a new vector representation by integrating dictionary function vector due to the uncertainty in the boundary of medical vocabulary terms and the complexity of representing uncommon and unknown entities. Each word 's vector representation ei consists of the word vector wi and the dictionary feature vector di as ei = wi + diWhere + is the concatenation operator , as shown by the diagram, the word vector is constructed as a combination of word embedding and character [S]Six sessile polyps were found in the sigmoid colon and in the proximal descending colon[F]. [S]The polyps were 4 to 15 mm in size. All polyps were completely removed by cold snare polypectomy [F].[S] Resection and retrieval were small and large - mouthed diverticulitis were found in the sigmoid colon[F]. [S]A sessile polyp was found in the descending colon [F]. [S] The polyp was 8 mm in size [F]. [S] The polyp was removed with a cold snare [F]. [S] Resection and retrieval were complete [F] .[S] [O] Estimated blood loss was minimal [O] [F] . [S] A fungating partially obstructing mass that could be easily traversed was found in the sigmoid colon [F] .[S] [O] The mas s was circumferential [O] [F] .[S] The mass measured five cm in length from 25 to 30 cm [F] . 2 embedding. To represent character embedding, we use bidirectional LSTM to extract character features and implement certain character features. Dictionary feature to vector We use the n -gram models to divide the original sentence into text segments based on the word's context . After that, combined with the domain dictionary (colonoscopy) C, we produce a binary value dependent on whether or not the text segments are in C. Dicti onary feature vectors are built in various dimensions, based on the number of entity categories in the dictionary. Bi-LSTM converts the final dictionary function vector from the dictionary element. Figure 3 . Word vector as a combination of character embedding and word embedding. Each word trains a word vector to the model. Entity Extraction Model The entity extraction utilizing Bi -LSTM -CRF[32] used word embedding as inputs ; we employ embedding character and word embedding. It allows the model to get trained with more feature information. The combined character words and dictionary feature vectors are feed into the model. This step has the following advantages:(1) more detail s about the features; and (2) better identification of the entity boundary problem. After that, we feed this into a Bi -LSTM model . Instead of splitting the two feature vectors as inputs, three separate Bi -LSTM versions of the proposed model are seen in Fig ure 3. To obtain a vector of the final character representation, we implemented a Bi -LSTM neural network and added the forward and backward sequences' output vector s. The key aim of utilizing Bi - LSTM is that it functions best to take on long -term dependencies in a phrase, and the bidirectional sequential design provides further advantages by remembering both a word's previous and potential meaning. 3 CRF Layer: The colonoscopy reports for a sentence , the measurement of the from BI -LSTM neural networ k obtains a word representation, ignoring neighboring marker dependencies. For example, the tag I (Inside) cannot be followed with the tag B (Begin) or E (End). We use conditional random field (CRF) to determine labels from a sequence of background representations, rather than using it independently for tagging decisions can produce greater tagging accuracy in general. Rules for Spec ific Task As we are dealing with unstructured colonoscopy clinical note s from multiple facilities , we use manual rules to fix this problem . The aim of defining rules is to identify sentences that contain certain entities from the text. These rules will be dynamical ly updated for each facility. We expla ined the main four rules in Table 2. Table 2. Description of the six major rules Task Rules Sentence boundary detection Keywords ( e.g., Number of polyps, jar , etc. ), period, semicolon, , and Polyps size Keywords (e.g., small, medium , large , etc.) , number preceeding CM, or number preceeding MM Location of polyps Keywords (e.g., rectal, sigmoid , etc.) or distance from rectum given in CM. Word processing Abbreviation s (e.g., SIG for sigmoid, REC for rectal, etc.) and special word processing (e.g., Recto -sigmoid is a combination of rectal and sigmoid , Tubulovillous is a combination of Tubular and Villous , etc.) Number of polyps Keywords (polyps, number, numerics , and sessile) Dataset and Experimental Settings As previous ly mentioned , there are 4,000 manually annotated colonoscopy reports and five kinds of medical named entities of interest: a number of polyps, size of the polyp, location of the polyp, the procedure of removal of polyps , and status of removal as shown in Table 3. There are several sentences to every instance. These sentences are further separated into clauses by using S as the beginning of the sent ence and F as mentioned in section 4.3 . We calculate micro -averaged precision, recall, and F1-score for exact match ing word s and classify sentences accordingly to conceptual meaning from colon oscopy report s. We compare our model Bi-LSTM -CRF with improved NER and sentence boundary detection with the Bi-LSTM -CRF. Table 3. Statistics of entity and sentence. Type Number of Entities Number of Sentences Indicat ion 4,215 - Number of polyps 9,542 4,687 4 Location of polyps 12,647 5,321 Size of polyps 8,654 3,549 Polyp removal procedure 3,657 2,648 Type of colorectal cancer 6,548 4,598 We used Tensor Flow to complete the CRF module. We first run a bidirectional LSTM -CRF model forward pass for each batch , including the forward pass for both the forward and backward states of LSTM. We divide the learning rate by five if the validation accuracy decreases and used mini-batches of size 100 , and training is stopped when the learning rate goes under the threshold of 10-8. For classifier s, we set hidden layers size to 300; because neural network architectures are usually over - parameterized and vulnerable to overfitting to solve it, we used regularization strategies that have been implemented: Elastic Net regularization has been added to each LSTM hidden layer. The parameters predomi nantly include tag indices and batch size, where tag indices indicate the number of actual tags, and the batch size reflects the number of batch samples. The parameter setting s for the pr oposed mod el are shown in Table 4. Table 4. Parameter setting of the proposed method. Parameter Value Word vector embedding size 200 Dictionary feature vector embedding size 100 # hidden neurons for each hidden layer 300 Batch size 100 Tag Indices 4 Learning Rate 0.005 Number of epochs 10 Optimizer Adam optimizer Model evaluation The proposed Bio-Bi-LSTM -CRF method and the state-of-the-art model B i-LSTM -CRF were trained on the training subset , and the validation subset was used for tuning the hyper -parameters. Ultimately, the trained models were tested on the test subset , and the following outcomes were reported as the models ' performance. We used precision, recall, accuracy , and F1 score to ev aluate the Bio-Bi-LSTM - CRF model and Bi -LSTM -CRF model's performance compared with the overall accuracy. Evaluations of precision, recall , accuracy , and F1 score were performed for each report . For example, all of the polyp numbers, sizes, and locations in a findings report must be correct for it to be considered accurate regardless of the number of polyps. Also , the precision, recall, accuracy, and F1 score were perfo rmed at the entity -level (e.g., each polyp size , number, and location) to find the models' performance for each facility. An exact match with the gold standard of manually annotated records 5 is required for each entity versus the record -level where each record is considered separately. Results The proposed Bio-Bi-LSTM -CRF outperformed Bi -LSTM -CRF and achieved higher accuracy by 5.6%, 6.6% , and 1 9.1% percentage points in the indication, findings, and pathology reports, respectively , as shown in Table 5. The Bio-Bi-LSTM -CRF model's absolute accura cy was relatively high for all three reports: 88 .0% for findings reports , 93.5% for indication reports , and 96 .5% for pathology reports . The proposed Bio-Bi-LSTM -CRF thus reports a 15.3% increase in precision, a 14.8% increase in recall , and a 10.3% increase in F1 over the state-of-the-art model Bi -LSTM -CRF. The hierarchical influence structure of Bio-Bi-LSTM -CRF aided in reducing noisy text from c linical colonoscopy notes. Table 5. Comparative results by report between Bi -LSTM -CRF and Bio -Bi-LSTM -CRF model Reports Method Accuracy Recall Precision F1 Indication Reports Bi-LSTM -CRF Bio-Bi-LSTM -CRF 88.5% (354/400 ) 93.5% (374/400) 82.5% 93.0% 82.5% 93.0% 83.2% 93.4% Findings Report Bi-LSTM -CRF Bio-Bi-LSTM -CRF 82.5% (300/400) 88.0% (352/400) 83.5% 88.7% 84.0% 89.0% 82.7% 88.1% Pathology Report Bi-LSTM -CRF Bio-Bi-LSTM -CRF 81.0% (324/400) 96.5% (386/400) 77.6% 96.0% 75.0% 96.0% 79.7% 96.4% Both the Bi -LSTM -CRF and Bio-Bi-LSTM -CRF models performed better in entity identification indications reports , and pathology reports achieved an average of 84.75 % and 95% accuracy between facilities , as shown in Table 6. However, they struggled in organiz ing the findings reports that mentioned characteristics of number polyps and locations of polyp in body and procedure of removal of a polyp, e.g., two diminutive mass polyp, and reports where the status of the location of the polyp was defined, e.g., A sing le small 3 -5mm polyp away from rectum and sigmoid colon ' by long -range dependencies and words. In indication reports and pathology reports , there are no long dependency sentences and multiple sentences. Whereas findings have multiples sentences representin g the information of polyps. To understand polyps' characteristic s, the model needed to look further back in the sentence and relate polyp or sessile . Table 6. Comparative results by facility between Bi -LSTM -CRF and Bio -Bi-LSTM -CRF model Facility Methods Accuracy Recall Precision F1 6 Albany Bi-LSTM -CRF Bio-Bi-LSTM -CRF 83.0% (332/400 ) 92.0% (368/400) 79.0% 91.5% 81.0% 91.0% 78.0% 90.0% Ann Arbor Bi-LSTM -CRF Bio-Bi-LSTM -CRF 86.0% (344/400 ) 94.0% (376/400) 84.0% 92.0% 82.0% 93.0% 85.0% 93.5% Detroit Bi-LSTM -CRF Bio-Bi-LSTM -CRF 83.0% (332/400) 91.0% (364/400) 83.0% 91.0% 78.0% 90.5% 81.0% 91.0% Indianapolis Bi-LSTM -CRF Bio-Bi-LSTM -CRF 80.0% (320/400 ) 93.0% (372/400) 81.0% 95.5% 82.0% 95.5% 84.0% 95.5% On the other hand, Bio-Bi-LSTM -CRF models entity identification on findings reports correctly. Bi - LSTM -CRF models achieved an average of 92.6% recall , as shown in Table 8. The Bio -Bi-LSTM - CRF model only made a few of entities where the polyp location was represented in the form of distance from the rectum (130 cm). T he number of polyps was described in an atypical fashion , including uncommon terms such as semi -mass of large 10mm polyp , which only appeared in one and a few notes respectively in the dataset. The Bio-Bi-LSTM -CRF model's overall accuracy was between 8.0% and 13.0% points higher than the Bi -LSTM -CRF method in entity identification . The accuracy of the Bio-Bi-LSTM -CRF methods between facilities ranged from 92.0% to 94.0% , as shown in Table 7. The F1 -scores of Albany facility, Ann arbor facility, Detr oit facility, and Indianapolis facility are 93.7%, 96.3%, and 87. 1%; as shown in Table 8 The Bio-Bi-LSTM model achieved the highest overall results for all three facilities. The proposed Bio-Bi-LSTM -CRF thus reports a 12.8% increase in the F1 score. Likely, our Bio-Bi-LSTM - CRF model adapted well to other physicians' reporting style and language to achieve comparable performance in other facilities. For the polyp removal procedure, the performance of the two models of about the same level (at around 90 .0%). In comparison, the Bio-Bi-LSTM -CRF model shows an improvement of 20.9% F1 score over the Bi - LSTM -CRF in a type of CRC information from pathology reports. Bi -LSTM -CRF is struggling to extract entities from long dependences terms and multiple sentences. The weakest performance ar ea was for location information of polyp. There is a large variety of location representation in the dataset. These are also relative ly specific and unlikely to occur in the other facility data. Table 7. Comparative results by performance evaluation betwe en Bi -LSTM -CRF and Bio -Bi-LSTM - CRF model Category Methods Accuracy Recall Precision F1 7 Indicator Bi-LSTM -CRF Bio-Bi-LSTM -CRF 83.5% (334/400) 93.5% (374/400) 82.5% 93.1% 82.5% 93.0% 83.2% 93.4% Number of polyps Bi-LSTM -CRF Bio-Bi-LSTM -CRF 80.0% (320/400) 85.0% (340/400) 76.0% 88.0% 86.0% 89.0% 77.0% 89.0% Size Bi-LSTM -CRF Bio-Bi-LSTM -CRF 78.0% (312/400 ) 81.0% ( 324/400) 78.0% 88.0% 80.5% 89.0% 79.0% 90.0% Location of polyp Bi-LSTM -CRF Bio-Bi-LSTM -CRF 82.0% (328/400) 94.0% (376/400) 83.0% 86.0% 78.0% 87.0% 82.0% 85.0% Polyp removal procedure Bi-LSTM -CRF Bio-Bi-LSTM -CRF 89.0% (356/400 ) 92.0% ( 368/400) 94.0% 94.0% 93.0% 93.0% 84.0% 92.0% Type of CRC Bi-LSTM -CRF Bio-Bi-LSTM -CRF 81.0% (324/400) 96.5% (386/400) 77.6% 96.0% 75.0% 96.0% 79.7% 96.4% Discussion This study proposed a Bio-Bi-LSTM -CRF model for efficient and accurate entity extraction from colonoscopy report notes and free -text medical narratives . The novel Bio-Bi-LSTM -CRF accurately analyzed a large sample of colonoscopy reports. Our findings demonstrate a clear improvement in accuracy, precision, and recall within a highly regarded academic healthcare system. Bio-LSTM -CRF model consistently obtained better entity extraction accuracy than a Bi-LSTM -CRF model with identical feature sets. Over experiments, it was found that the addition of sentence boundary detection, rules, and domain dictionary in the tasks could significantly boost the accuracy, thus proving the efficacy of merging practic es and domain dictionary in entity extraction tasks. Our results h ighlight several advantages of using Bio-Bi-LSTM -CRF in routine quality measurement using data in EHRs. The critical advantage of Bio-Bi-LSTM -CRF is that it is economically feasible. It would be expensive and time -consuming to review tens of thousands of colonoscopies reports manually . Another advantage is tha t Bio-Bi-LSTM -CRF allows provider s to continue to use natural narrative when describing patient care. There has been criticism that structured note systems in current EHRs force providers to create unnatural and overly structured notes , which take extra ti me to develop and impede communication because these notes are difficult to read [33]. Compared to the direct and implicit term metrics, the neural embeddings were observed to have the inferior performance on the biomedical term similarity /relatedness benchmarks, as assessed both by 8 our tests (against several different implementations of wor d2vec that other groups have developed for similar tasks) and those reported by others (see Results). However, we acknowledge that that word2vec performance can be optimized for specific tasks by adjusting different choices of parameters and hyperparameter s, different neural network architectures, and other ways of handling multi -word phrases [34, 35] , and we did not test all possible implementations. Thus, we do not claim that our novel metrics necessarily perform better than neural embeddings in general. We decided to include only the most essential , informative words and phrases in our vocabulary in contrast to the word2vec based embeddings that encompass a much larger vocabulary . We feel that the relatively prevalent words and phrases are likely to be the most valuable for repres enting biomedical articles. However, it is worth exploring the effects of different vocabulary restrictions, especially when applying our vectors to other corpora, domains , and tasks. Strengths and Limitations In many aspects, our research is novel. Firs t, to our knowledge, this is only the research to examine the validity of NLP for multiple center data to investigat e indications, findings , and pathology laboratory results of dictated consultation notes for this purpose [36, 37] . Secondly, NLP research in CRC clinical notes concentrates exclusively on a single clinical condition, such as tumor presence [26]. Our analysis is significantly mo re comprehensive than other more widely reported studies , looking at four different medical centers. There are multiple drawbacks to our research . In our initial training dataset (n=4 ,000), our choice to randomly sample resulted in CRC cases' under -samplin g. As a result, the model was never trained on this feature and subsequently performed poorly during the final test set for this feature, possibly underestimating a properly trained model's effectiveness . This suggests that a real -world application of this technology may require a more purposive sampling strategy than our random sampling approach. Although the inclusion of many bigrams, trigrams, and abbreviations should reduce the problem of term ambiguity to some extent, the vector representations of word s represent an overall average across word instances and different word senses. The implicit term metrics have relatively few parameters, which we have attempted to set at near -optimal values. However, the choice of 300 -dimensions for the vector representations, and the particular weighting schemes for the similarity of vectors, may be regarded as best guesses and might be tuned further for optimal performance in specific tasks. However, th is study's goal was to establish the feasibility of using a deep learning model to extract clinical features from dictated consult notes and inform the approach to more extensive future studies. Conclusion Our proposed Bio -BI-LSTM -CRF improved the perception of unknown entities, which means that the model should have a better capability to deal with emerging medical concepts and multicenter data without extra training resources. The pre -processing of the word -level input of the model with extern al word embeddings allowable to improve performance further and achieve state -of-the-art for the colonoscopy clinical notes entity extraction. This idea could not only be applied in medical entity extraction but also other medical named entity recognition applications such as drug names and adverse drug reactions, as well as named entity recognition tasks in different fields. 9 References 1. Alblawi, A.S. and A.A. Alhamed. Big data and learning analytics in higher education: Demystifying variety, acquisition, storage, NLP and analytics. in 2017 IEEE Conference on Big Data and Analytics (ICBDA). 2017. IEEE. 2. Denecke, K. Extracting medical concepts from medical social media with clinical NLP tools: a qualitative study. in Proceedings of th e fourth workshop on building and evaluation resources for health and biomedical text processing. 2014. Citeseer. 3. Strubell, E., A. Ganesh, and A. McCallum, Energy and policy considerations for deep learning in NLP. arXiv preprint 2019. 4. Jiang, F., et al., Artificial intelligence in healthcare: past, present and future. Stroke and vascular neurology, 2017. 2(4). 5. Dinov, I.D., Methodological challenges and analytic opportunities for modeling and interpreting Big Healthcare Data. Gigascience, 2016. 5(1): p. s13742 -016-0117 -6. 6. Mishuris, R.G. and J.A. Linder, Electronic health records and the increasing complexity of medical practice: it never gets easier, you just go faster . Journal of general internal medicine, 2013. 28(4): p. 490 -492. 7. Polnaszek, B., et al., Overcoming the Challenges of Unstructured Data in Multi -site, Electronic Medical Record -based Abstraction. Medical care, 2016. 54(10): p. e65. 8. Meystre, S., Clinical Data Reuse or Secondary Use: Current Status and Potential Future Progress (2017). 9. Jemal, A., et al., Cancer statistics, 2010. CA: a cancer journal for clinicians, 2010. 60(5): p. 277 - 300. 10. Center, M.M., A. Jemal, and E. Ward, International trends in colorectal cancer incidence rates. Cancer Epidemiology and Prevention Biomarkers, 2009. 18(6): p. 1688 -1694. 11. Levin, B., et al., Screening and surveillance for the early detect ion of colorectal cancer and adenomatous polyps, 2008: a joint guideline from the American Cancer Society, the US Multi - Society Task Force on Colorectal Cancer, and the American College of Radiology. 2008. 134(5): p. 1570 -1595. 12. Philli ps, K.A., et al., Trends in colonoscopy for colorectal cancer screening. Medical care, 2007: p. 160 -167. 13. Winawer, S.J., et al., prevention of colorectal cancer by colonoscopic polypectomy. New England Journal of Medicine, 1993. 329(27): p. 1977 -1981. 14. Denny, J.C., et al., Extracting timing and status descriptors for colonoscopy testing from electronic medical records. Journal of the American Medical Informatics Association, 2010. 17(4): p. 383 -388. 15. Goldberg, Y., Neural network methods for natural language processing. Synthesis lectures on human language technologies, 2017. 10(1): p. 1 -309. 16. Jagannatha, A.N. and H. Yu. Structured prediction models for RNN based sequence labeling in clinical text. in Proceedings of the conference on empirical met hods in natural language processing. conference on empirical methods in natural language processing. 2016. NIH Public Access. 17. Pons, E., et al., Natural language processing in radiology: a systematic review. Radiology, 10 2016. 279(2): p. 329 -343. 18. Camb ria, E. and B. White, Jumping NLP curves: A review of natural language processing research. IEEE Computational intelligence magazine, 2014. 9(2): p. 48 -57. 19. Wong, D.F., L.S. Chao, and X. Zeng, isentenizer -: Multilingual sentence boundary detection model . The Scientific World Journal, 2014. 2014 . 20. Palmer, D.D., SATZ -an adaptive sentence segmentation system. arXiv preprint cmp - lg/9503019, 1995. 21. Kreuzthaler, M. and S. Schulz. Detection of sentence boundaries and abbreviations in clinical narratives. in BMC medical informatics and decision making. 2015. BioMed Central. 22. Imler, T.D., et al., Natural language processing accurately categorizes findings from colonoscopy and pathology reports. Clinical Gastroenterology and Hepatology, 2013. 11(6): p. 689 -694. 23. Li, F., et al., Recognizing irregular entities in biomedical text via deep neural networks. Pattern Recognition Letters, 2018. 105: p. 105 -113. 24. Ratinov, L. and D. Roth. Design challenges and misconceptions in named entity recognition. in Proc eedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL -2009). 2009. 25. Zhang, J., et al., Enhancing HMM -based biomedical named entity recognition by studying special phenomena. Journal of biomedical informatics, 2004. 37(6): p. 411 -422. 26. Denny, J.C., et al., Natural language processing improves identification of colorectal cancer testing in the electronic medical record. Medical Decision Making, 2012. 32(1): p. 188 -197. 27. D'Avolio, L.W., et al., Evaluation of a generali zable approach to clinical information retrieval using the automated retrieval console (ARC). Journal of the American Medical Informatics Association, 2010. 17(4): p. 375 -382. 28. Halligan, S., et al., C .T. colonography in the detection of colorectal polyp s and cancer: systematic review, meta -analysis, and proposed minimum data set for study level reporting. Radiology, 2005. 237(3): p. 893 -904. 29. Goldberg, Y. and O. Levy, word2vec Explained: deriving Mikolov et al.'s negative -sampling word -embedding metho d. arXiv preprint arXiv:1402.3722, 2014. 30. Pennington, J., R. Socher, and C.D. Manning. Glove: Global vectors for word representation. in Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 2014. 31. Si, Y. and K. Roberts. A frame -based NLP system for cancer -related information extraction. in AMIA Annual Symposium Proceedings. 2018. American Medical Informatics Association. 32. Huang, Z., W. Xu, and K. Yu, Bidirectional LSTM -CRF models for sequence tagging. arXi v preprint 2015. 33. Walsh, S.H., The clinician's perspective on electronic health records and how they can affect patient care. Bmj, 2004. 328(7449): p. 1184 -1187. 34. Al-Sadi, A.M., et al., Genetic analysis reveals diversity and genetic relationship among Trichoderma isolate s from potting media, cultivated soil and uncultivated soil. BMC microbiology, 2015. 15(1): p. 1 -11. 35. Henry, S., C. Cuffy, and B.T. McInnes, Vector representations of multi -word terms for semantic relatedness. Journal of biomedical informatics, 2018. 77: p. 111 -119. 36. Imler, T.D., et al., Multicenter colonoscopy quality measurement utilizing natural language

2108.11034

Shashank Reddy Vadyala

Shashank Reddy Vadyala, Eric A. Sherer

Natural Language Processing Accurately Categorizes Indications, Findings and Pathology Reports from Multicenter Colonoscopy

18 pages, 3 figures

cs.CL cs.LG

http://creativecommons.org/licenses/by-sa/4.0/

Colonoscopy is used for colorectal cancer (CRC) screening. Extracting details of the colonoscopy findings from free text in electronic health records (EHRs) can be used to determine patient risk for CRC and colorectal We developed and evaluated the accuracy of a deep learning model framework to extract information for the clinical decision support system to interpret relevant free-text reports, including indications, pathology, and findings notes. The Bio-Bi-LSTM-CRF framework was developed using Short-term Memory (Bi-LSTM) and Conditional Random Fields (CRF) to extract several clinical features from these free-text reports including indications for the colonoscopy, findings during the colonoscopy, and pathology of We trained the Bio-Bi-LSTM-CRF and existing Bi-LSTM-CRF models on 80% of 4,000 manually annotated notes from 3,867 patients. These clinical notes were from a group of patients over 40 years of age enrolled in four Veterans Affairs Medical Centers. A total of 10% of the remaining annotated notes were used to train hyperparameter and the remaining 10% were used to evaluate the accuracy of our model Bio-Bi-LSTM-CRF and compare to Bi-LSTM-CRF.

2 Wang et al. et al. 2018; Teske et al. 2018). For example, the occurrence rate of Earth-size planets could be overestimated by tens of percents if not accounting for the binarity of KOI host star (Bouma et al. 2018). Given the ubiquity of these two kinds of FPs, separating FPs from genuine (and correctly classied) planets is crucial for transit survey like Kepler , whose main objects are Earth-size planets in habitable zones. One way to separate these FPs from genuine planets is to measure the mass of transiting body through radial velocity (RV) measurements. However, RV observations are primarily based on several premium ground-based facilities that are of high demand and are best targeted for planet candidates with high mass from bright host stars, which are not typical for the ensemble of Kepler candidates (Borucki et al. 2011; Batalha et al. 2010). Moreover, directly conducting RV observations to identify BEBs and CTPs could be problematic in that (1) the spectrum of background eclipsing binary can be concealed by the contamination of foreground star and (2) the RV residuals of stellar companions cannot be easily corrected without knowing the binarity of system. Thus, the conventional RV measurements are not the optimal way to comprehensively vet Kepler of directly measuring the planet's mass, statistical validation is proposed to validate the transit validation aims to argue that the transit signal is more likely to be created by planet transits than any FP scenarios. The prototype of this procedure, BLENDER , was rst embodied and employed by Kepler team (Torres et al. 2005, 2011). More recently, Morton (2012), and D az et al. (2014) develop more ecient tools to evaluate the candidates' false positive probability (FPP) under Bayesian framework. Most notably, the VESPA has over a thousand of Kepler candidates (Morton et al. 2016). A commonality of these methods is that any available ancillary information of the candidate (e.g. contrast curve, RVs, multi-color transit photometry, etc.) can be utilized to further constrain and clear the stellar parameter space of FPs that fail to reproduce the observed light curves. When the FPP is below the dened threshold, the candidate can be safely validated. Using multi-color photometry to vet transit candidates is based on the fact that blends containing stars with dierent colors will project color-dependent depth signature on dierent observing bandpasses (Rosenblatt 1971; Tingley 2004). As the color-dierence between dierent components increases, the dierence between mutil-band transit depths will increase, too. While the apparent depths of the non-emitting planet transiting single star will remain constant in dierent bandpasses (neglecting the depth-changing eect caused by stellar limb darkening). Based on the chromatic depth of blending objects, O'Donovan et al. (2006) rst utilized the multi-color photometry to identify and rule out blending eclipsing binaries in the wide-eld transit survey. Later, several Kepler follow-up observations eliminated KOIs that displayed chromatic transit depths by ground-based high-precision photometry (Col on & Ford 2011; Col on et al. 2012; Tingley et al. 2014; Col on et al. 2015) and by parallel Spitzer observations (Torres et al. 2011; Fressin et al. 2012; D esert et al. 2015). All these successful practices have demonstrated that multi-color photometry is eective in revealing the blends, thereby sparing resources of follow-up observation for other candidates with more this paper, we study how many false positives can be identied by combining an additional reference bandpass in a Kepler -like survey. This double-band photometry observation strategy applies to observation and a follow-up photometry campaign, assuming the reference photometry will accumulate time baseline comparable to Kepler 's (4 year). We consider three reference photometric bands: SDSS z1, 2MASSKs(Cohen et al. 2003), and TESS (Ricker et al. 2014). As the color-dependent apparent depth is a strong indicator of stellar blends, we use it as the discriminator to identify and rule out the blending signals, and estimate the overall rule-out fraction of false positives achieved by Kepler and chosen reference photometry combinations. In addition, we also evaluate how an extra photometry observation would improve the FPPs of candidates compared to those derived from Kepler providing an alternative metric for false positive false positives considered in this paper are background eclipsing binaries and companion transiting planets, which are most pervasive in transit surveys and may remain undetermined in more sophisticated vetting procedures. The paper is structured as follows. In Section 2 we dene the FPs in consideration and introduce the In Section 3 we compute the fraction of FPs that can be identied with variable transit depth and study how dierent stellar properties aect the detectability of depth variations. For those signals which cannot be instantly identied due to insignicant depth variations, we adopt a Bayesian FPP computation framework double-band FPP results for typical cases in Section 4. We compare these two approaches and discuss some issues in Section 5, and summarize our work in Section 6. 2.SIMULATING THE FALSE POSITIVE SIGNAL ON DIFFERENT PHOTOMETRIC BAND As the objective of this paper, we investigate chromatic transit depth originating from two typical false background eclipsing binary (BEB) and companion transiting planet (CTP). For clarity, we dene as follows: *The background eclipsing binary consists of a luminous foreground star ( target star ) and a pair of faint eclipsing binary in the background ( blended companion ). *The companion transiting planet consists of a binary system, with a planet transiting the lower-mass stellar object ( blended companion ), and the signal is largely diluted by the brighter primary star ( target star ). In both cases, the objects that produce the signal are not the target stars themselves, but rather are the companion . We do not consider higher multiplicity systems in our simulation since their occurrence rates are lower than the binary star systems (Raghavan et al. 2010). 2.1. Transit depth calculation The observed depth of BEB and CTP on certain wavelength ,o;, is only a fraction of the true depth is the uxes from blending objects and target stars measured on observed wavelength. The ratio between the ux of the blending object and the total ux is denoted as the dilution factor d. If the observed signal comes from a single and isolated system, the dilution factor will be unity. The dierence between the observed depth in the Kepler (KP) and the reference band (RF) can be expressed (2) For a planet transit, the true depth is xed at the planet-star surface ratio. Therefore the observed CTPs is merely the dierence of dilution factors on dierent photometry. For eclipsing binary, the eclipse depths will also vary across wavelengths when the binary components have internal tem- perature dierence, and so does the dilution factor, which will further change the depth of the signal. 2.2. False positive simulation To produce the representative population for the target stars and the blended components dened at the beginning of this section, we generate synthetic stellar populations that are not only physics-based but also consistent with constraints provided by Kepler and other surveys. Due to the main purpose of this paper, we need accurate stellar photometry to calculate the apparent depths, which is mainly associated with the stellar masses. We generate these stellar properties with the python package isochrones (Morton 2015), which perform in parameter space for Mesa Isochrones and Stellar Tracks (MIST; Choi et (2016)). isochrones also predicts the stellar photometry using the bolometric correction grid2provided by MIST. The reference catalog of our target star population is Kepler Input Catalog (KIC, Brown et al. (2011)), and we use this star population to assign line-of-sight background eclipsing binaries or bounded stellar companions, which will be described in following sections. We list the simulation parameters of target star population in Table 1. We compare the distributions of mass, apparent Kepler magnitude, and eective temperature of the target star population, along with those of KIC stars, in Figure 1. The agreement between KIC and our target star population is satisfying, though our simulation contains more hotter stars. We will adress this dierence and its impact in Section 5.3. To infer the stellar properties of background stars and their spatial gradients in background elds, we use TRILEGAL to simulate background stars (Girardi et al. 2005). TRILEGAL is a web-based Milky Way star population Wang et al. Table 1. Input parameters of target star mass function (IMF) Chabrier IMF [Fe/H] Double-Gaussian in (Casagrande et al. 2011) Age Uniform(0,10) Gyr Ava0.1-1.6 dNd2; d2[0;1800] pc Magnitude limit ( Kepler ) 16 a. Drawn from calculator, depending on the observing location. 0.5 1.0 1.5 2.0 Mass targets Kepler Input Catalog 10 12 14 16 Kepler 6000 1. The cumulative density function of target stars' mass, Kepler magnitude and eective temperature, compared to the stars in Kepler Input Catalog. that considers dierent galactic components. The request can be called by setting the coordinate of desired elds and galactic structures (we use defaults of structure parameters). We use the stellar mass and distance provided by TRILEGAL to generate eclipsing binaries for BEB simulation. We provide detailed descriptions about how we construct the appropriate BEBs and CTPs in following Section 2.2.1 and Section 2.2.2. 2.2.1. Simulating background eclipsing binaries We downloaded background stars in the centers at each of the 21 modules in Kepler eld of view via TRILEGAL. For the sake of consistency in deriving stellar properties as we generate target stats populations, we extract the returned parameters, e.g., initial mass, metallicity, age, distance, and extinction factor, from TRILEGAL and pass them to ischrones to obtain the multi-band photometry and radius of the background primaries. Next, according to the at mass-ratio and the binary fraction as the function of primary's spectral type reported in Raghavan et al. (2010), we generate secondary stars to selected primaries. The eclipse of a binary system only occurs when the impact parameter meets the criterion (Winn +R2 R1(3) where stands for the presence of primary eclipse and for secondary eclipse, bfor the impact parameter of the eclipse, afor the semi-major axis, R1; R2for the radius of primary and secondary star, efor eccentricity, i for inclination and !for argument of periastron. We assume the orbital elements of each binary pair follow and eccentricity distribution from Multiple Star Catalog (Tokovinin 2018), and truncate the period up to 300 days.5 *Uniform distribution between 0 360for!and a cosine distribution for orbital the systems that are in each other's Roche Lobe (Equation 2 in Eggleton (1983)). For each background eclipsing binary, we assign a foreground star from the target star population, and then we calculate the diluted eclipse according to Equation 1. Next, we calculate the diluted ratio (SNR) according to SNR the observed depth on Kepler band,e;KPis the eective photometric precision, which is based on the target star's 6-hour Combined Dierential Photometric Precision (CDPP) inferred from, the orbital period P, and decide whether they are detectable on Kepler band. To model the signal detection condition of Kepler , we reject BEBs whose both of the eclipses' SNR KP<7 (Jenkins et al. 2010), which is the nominal detection threshold of Kepler pipeline to avoid misidentication due to statistical uctuation in the ensemble of all target stars. Central for a BEB to mimic planet-transit is it must not show distinguishable dierence between primary and secondary eclipse. Based on the calculated SNR in Kepler bandpass, we classify the BEBs that have passed the secondary eclipse test into two 1: the primary and secondary eclipse are both detectable, but are in depths (within 3 ), which are taken as candidates with periods of half of the actual eclipse period, and the number of will be doubled. Thus, the detection threshold of this type of BEB will be7p 25. *Type 2: one eclipse is `missing' in the observation, which may result from high-eccentric orbits, or the eclipse is diluted to be undetectable (SNR pri=sec<3) by the foreground star while the other eclipse is still detectable. The two groups of BEBs are well separated and distinctive on the SNR KPdiagram of eclipse. An example of the SNR KPdiagram is shown in Figure 3. Type 1 BEBs lie in the diagonal of the diagram, with a median orbital eccentricity of 0.01 and a mass ratio of 0.89. Type 2 BEBs cluster along the `two wings' symmetric to the diagonal, with a median orbital eccentricity of 0.39 and mass-ratio of 0.52. 2.2.2. Simulating companion transiting planets The CTP model consists of a planet transiting the secondary star of the binary system. To construct the we select stars from the target population as the primary stars, in which we account for the reported binary fraction's dependence on primary mass. Then, we assign secondary companions to the primaries with a uniform mass ratio between 0.1 and 1 (Morton & Johnson 2011; Raghavan et al. 2010). We assume the binary members are co-located and coeval, that is, the secondary stars have the same metallicity, age and distance as primaries. Next, we assign a planet to each secondary star. The planets are parameterized by their radii and orbital periods. Orbital periods are randomly selected from KOI lists (Batalha et al. 2013). The assignment of planet radii, however, is not so straightforward. We will address this issue in detail in Section 3.2. 2.3. Reference bands and noise model In our work, we have chosen three photometric bands to be the reference band of Kepler , and each of them has a distinctive eective wavelength and coverage: SDSS z(800-900 nm), 2MASS Ks(1.9-2.3m), and TESS (600-900 nm) band. SDSS zband is a narrow-band lter commonly used in ground-based photometry. TESS band is is employed by the spacecraft TESS especially designed for transit photometry. 2MASS Ksis a near- infrared lter band and may represent future space-based infrared photometry. Our preference for partly stems from the assumption about the considered false positives: we assume the Kepler targets are often blended with a lower-mass object, which is a lot redder in colors (and interstellar extinction will further redden the color of background binaries in BEB). Hence, a combination of photometry bands with concentration at visible light (such as Kepler band) and infrared (and even near-infrared) would properly exploit and reveal the color-dierence6 Wang et al. in BEBs and CTPs. The aforementioned assumption is reasonable for most of the stubborn cases of FPs that are of relatively small sizes and low brightness. Target stars blended with higher mass (and thus brighter) objects may have already been compiled in the KIC catalog or identied by other techniques, such as centroid analysis (Bryson et al. 2013) or ephemeris matching (Coughlin et al. 2014). To quantify the uncertainties in transit depth measurements and the signicance of the depth variations in it is necessary to assign photometric precision to each simulated FP instance. The Combined Precision (CDPP) measured in Kepler stars provides the best template for our purpose. Here we assume the CDPP is only correlated to the Kepler magnitudes mKP. To derive an empirical relation, we take median 6- hour CDPP evaluated in 1-magnitude bins between mKP= 9 and 16. We assign Kepler noise to each FP instance in terms of its apparent mKPby linearly interpolating the CDPP- mKPrelation explained above. For the reference bands, we multiply Kepler 's CDPP-mKPcurve by factors of one, three, and ten, interpolating the scaled noise- mRF relation based on the apparent mRF. By adopting dierent noise templates can we provide a ducial estimate for the photometry precision for the reference band: when RF=KP, the reference photometry attain the best precision so far, as have been examplied by Kepler . IfRF= 3KP, this is another typical precision for space-based telescope with smaller aperture, like TESS (30cm). In the case of RF= 10KP, this is where the photometry can achieve. Examples of Kepler and reference band noise model are presented in Figure 2. Figure 2. The adopted photometric precision relation to evaluate the observed depth uncertainties for our simulated false positives (red, blue and yellow dashed curves). The photometric uncertainty on Kepler band is represented by the red curve, which is the median 6-hr CDPPs (black dots) in 9 bins among all stars in Kepler Input Catalog with mKPbetween 10 and 16. The photometric precision on reference bands ( z; K s;TESS) is evaluated by the apparent reference magnitude mRFaccording to the three example precision curves, which are the scaled version of Kepler precision with factors of one, three, and ten. 3.RULING OUT FALSE POSITIVES WITH Background eclipsing binaries 3.1.1. Identify BEBs with two photometric band This section demonstrates that the Kepler 's candidates mimicked by background eclipsing binaries can be identied and excluded by joining additional reference photometry. In the previous section, we dene those BEBs have passed the secondary eclipse test on Kepler band will be from genuine planet transits, i.e., on the Kepler band, the primary's and secondary eclipse have similar depths, or only eclipse is detectable. Since a joint photometry is observing with Kepler , we may also check whether these BEBs can also pass the secondary eclipse test on the reference band. We then identify the BEBs for which the SNR of eclipse dierence on reference band, SNR prisec;RF, is larger than 5. This quantity is expressed in Equation 5. SNR the diluted primary and secondary depth of BEB, and pri;RF; sec;RFis the eective noise of eclipses. Some BEBs may pass the secondary eclipse test on both photometry bands. To eliminate those stubborn cases, we proceed to compare the apparent depth on two photometry bands. To estimate the number of simulated BEBs that show detectable depth variations between Kepler and the reference band, we calculate the SNR of depth-dierence, SNR RFKP, and identify those BEBs whose SNR RFKPis larger than 5. This quantity can be expressed as: SNR (6) whereRF; KPare the apparent transit depth observed in reference band and Kepler . With either of these two features observed, one can condently identify the blending nature of the transit signals. We present an example of the SNR of dierence and double-band depth variations for the BEBs using Kepler -TESS pair photometry, which is displayed in the left panel of Figure 3, and we mark the BEBs if they show any of the two characteristics in the right panel. The majority of Type 1 BEB can only show double-band depth variations since most of them contain equal-mass EBs so that the primary and secondary eclipse will remain nearly identical across all photometry. The detectable dierence is preferentially displayed by BEBs mass-ratio, primarily Type 2 BEB. Figure 4 compares these two identication methods at dierent reference bands and precision. Generally, more BEBs will show identiable depth variations rather than dierence. On the Ksband, most of the BEBs will show both of the features, and this fraction only diers between 80-95% under the considered reference In contrast, using zand TESS band can at most identify 83% and 69% of the BEBs, respectively. And the fractional number will drop sharply to only around 12% and 6.8% when RF=KP= 10. Moreover, the fraction of BEBs identied via dierence is nearly negligible under relatively large noise ( RF=KP= 3;10) forzand TESS band. 3.1.2. Factors aecting the the amplitude of depth to Figure 4, most BEBs in our simulation show signicant double-band depth dierence rather than dierence. We may examine how the amplitude of depth-dierence is correlated with BEB We present the SNR RFKPof all BEBs in Figure 5, along with the apparent depths and Kepler by the systems. We nd that BEBs preferentially produce and mimick Earths and small-Neptune candidates. This pattern agrees well with the simulation of Fressin et al. (2013). Besides, BEBs with large apparent Kepler depth tend to show large SNR RFKP, even though they are produced by faint foreground target stars. However, even for small planet candidates, at least half of them can display detectable depth variations, assuming the reference bands can achieve a Kepler -like precision. The color-dierence between dierent stellar components in BEB congurations will also in uence the amplitude of the depth feature: (i). the color-dierence (temperature dierence) between the foreground target star and background system, in which foreground stars' dilution level will dier between bandpasses and (ii). the internal color-dierence (or equivalently the eective temperature dierence) between the will result in a chromatic eclipse depth. With such a view, we may proceed to explore the which stellar components would impact SNR RFKPmore signicantly. To better illustrate this, we plot BEBs' SNR RFKPagainst the between background binary com- ponents and stars separately in Figure 6. The eect of (i) is evidenced by the positive trend between the SNR RFKPand star temperature ratio (black dots) presented in Figure 6. The eect of (ii), on the contrary, show a negative correlation (red dots) with SNR RFKP, as shown in Figure 6. This is because the eective temperature not only correlates to the color of star, but also with star's size and luminosity. As the secondary star gets a lot cooler than the primary, it will become smaller and less luminous, therefore the largest possible decline in brightness during the eclipse will drop, too. The decrease of the true eclipse depth will reduce the magnitude of apparent diluted depth and the amplitude of depth variations. Thus, BEBs that consist of members with distinct spectral types tend to show minor amplitude in depth variations, even if the is large. 3.1.3. Ruling out BEBs with dierent apparent depths8 Wang et al. Figure 3. Left: The SNR KPof the primary and secondary eclipse of simulated background eclipsing binaries (BEBs), color- coded by the mass-ratio qof background binary. Type 1 BEBs has primary and secondary eclipses in They lie in the diagonal and consist of mainly equal mass ( q0:88) and circular-orbit ( e0:01) . Type 2 BEBs have one undetectable eclipse. They cluster in the upper left and bottom right corner of the plot, and they are mainly composed of low mass-ratio ( q0:43) and eccentric-orbit ( e0:26) binaries; Right : the same as the left gure, but the BEBs are marked in terms of the detectable double-band photometry feature. Light blue circles are the BEBs that only show distinguishable depth variations (SNR RFKP>5), which are mainly Type 1 BEBs and low qType 2 BEBs; deep blue circles are those BEBs that can only be revealed by their identiable eclipse dierence on the reference band, which are mainly binaries with intermediate q; red lled circles are the BEBs that both of the signatures, and the grey circles show neither depth variation nor dierence. The instances whose SNR KPof primary or secondary eclipse is below 1 are plotted on the edge of the gure, for visualization. The reference band used to plot this gure is TESS, assigned with a noise estimate identical with Kepler 's. 1 3 rule-out fraction / %z 1 3 10 sK/sKPKs 1 3 pri-sec,RF>5 SNR 4. The fraction of BEBs showing detectable depth variations or detectable dierence or both, as a function of the photometric precision on reference band. We learn that dierent types of BEBs can be identied through dierent approaches, and the detectability is correlated to the stellar properties in the system. We may predict how much BEBs in Kepler can be identied overall. The mass distribution of background star have gradient along the galactic latitude covered by Kepler Field of View (FoV); thus such mass variance may aect the BEB's rule-out fraction in dierent modules in the entire Kepler eld. To estimate the overall rule-out fraction, we simulate BEB population in the centers of 21 modules of Kepler FoV via TRILEGAL, while the pool of foreground stars are the same as introduced in Section 2.2.1. We calculate the and in each eld. We present the average fraction in Figure 7, sorted by dierent reference bands, the size of planet candidate they mimic (based on apparent Kepler depth), and the assumed reference photometry 5. BEBs' SNR RFKPas a function of the apparent Kepler depth, color-coded by the Kepler magnitude of foreground star. The black dashed line separate the detectable depth-dierence (SNR RFKP>5) from those undetectable. BEBs with fainter foreground stars tend to produce deeper signals with larger SNR RFKP. Red and blue points represent the median SNR value of BEB subsets with hot and cold foreground star (respectively) in each apparent depth bin. Error bars correspond to 25th and 75th percentile. Hotter foreground stars may have larger color-dierence relative to the background binaries. Therefore they can be more easily distinguished by double-band 6. BEBs' SNR RFKPplotted against the eective temperature ratio between dierent stellar components. Black dots are the temperature ratio between background primaries and foreground stars; red dots are the temperature ratio secondaries and primaries. The median SNR RFKPcalculated in each Te-ratio bin is represented by big dots. The gure suggests that the larger color-dierence between foreground and background stars will enhance the detectability of depth variation. However, the color-dierence between background binary's negatively impacts the SNR RFKP. The opposite trend results from the decreasing eclipse depth when the secondary star is much smaller and fainter in low Te-ratio region. The Kepler -Ksphotometry pair can eciently rule out all BEBs with any depths. Even if the depth comparison is done with the largest noise on the reference band, the rule-out rate can still attain nearly 80% for our BEB samples. The rule-out fractions achieved by Kepler -TESS photometry pair strongly depend the apparent depths that the BEBs display. Although 80% of BEBs that mimics giant planet candidates can be eectively ruled out if TESS can achieve a Kepler -like precision, only half of the BEBs with Earth-size depth can be identied. Since most BEBs' apparent depth is shallow, the overall rule-out fraction is only around 69%, still below that of the the largest noise. Moreover, the exact level of TESS will signicantly aect the rule out rate of Earth-like BEBs, too. When TESS is three times of Kepler precision, the rule-out rate for Earth-like BEBs drops to 12%. In the worst precision, nearly none of Earth-like BEBs can be identied, and only those mimicking giant planet candidates are identiable. The performance of Kepler -zphotometry pair better than TESS. If zband can achieve a Kepler -like precision, it can rule out 80% of the BEBs with any apparent depths. However, shallow-depth BEBs get harder to identify as the noise on zband gets larger, and thus the total rule-out rate decreases. 3.2. Companion transiting planets10 Wang et al. Earth Super Earth Small Neptune Big Neptune Giant 4.76%+-0.38% 15.60%+-1.17% 30.18%+-2.38% 47.00%+-2.63% 47.41%+-1.11% 56.15%+-1.91% 63.54%+-2.88% 74.61%+-1.85% 85.96%+-0.47% 82.71%+-1.19% 80.94%+-1.80% 86.34%+-1.07% -z Earth Super Earth Small Neptune Big Neptune Giant 81.90%+-0.48% 78.19%+-1.34% 76.99%+-1.76% 83.06%+-1.02% 94.48%+-0.21% 91.52%+-0.61% 90.17%+-0.87% 92.90%+-0.57% 97.17%+-0.17% 94.59%+-0.48% 93.35%+-0.72% 95.07%+-0.47% -Ks Earth Super Earth Small Neptune Big Neptune Giant 0.88%+-0.10% 7.00%+-0.68% 20.30%+-1.91% 39.06%+-2.56% 23.52%+-1.04% 39.80%+-2.11% 54.04%+-3.23% 68.87%+-2.16% 69.67%+-0.78% 72.33%+-1.39% 74.46%+-2.27% 82.10%+-1.37% -TESS 0 20 40 60 80 100 Fraction / % Figure 7. This gure sums up the BEBs' rule-out fraction, classied by dierent apparent planet radii on Kepler band. The number on left column denotes the adopted photometric precision on the reference band. The numbers in each cell are the mean value of rule-out rate, averaged over the simulation from 21 elds centered on Kepler 's CCD modules. The 95% condence interval. Our planet radii classication is the same as that in Fressin et al. (2013). Ksband identify the BEBs mimicking all sizes of planet radii. TESS's and zband's successful identication prefers giant planet candidates mimicked by BEBs, though zis generally better than TESS. The search of planet transit is often carried out without knowing whether the target stars are single are binarity of the target star will bring uncertainty to the planet radius inferred from the observed signal. If the planet is transiting the primary star, the inferred planet radius is underrated byp 2 at most (in case of equal- mass binary) and is thus not seriously aected. Also, there may well be a chance that this is actually a large an unseen stellar companion, and the transit depth is diluted to appear much smaller due to primary's In this case, the radius correction XR, the ratio of the planet's true radius to apparent radius, can be up to six (Ciardi et al. 2015). Multi-color photometry can help identify the binarity of the system by identifying the depth variation in transit light curves.11 Part of the objective of this paper is to estimate the exclusion rate of Kepler -detected false positives via However, deriving the exclusion rate in the case of companion transiting planets is not so background eclipsing binaries. This is because currently, we lack the specic radius distribution of planets that are transiting the secondary star in a binary system, PCTP(Rp). Simply extrapolating the radius distribution from other work based on completely dierent contexts would undermine the reliability of our result. Thus, we refrain from employing any functional forms of the planet radius distributions in our simulation, nor will we derive a denitive value of the exclusion rate. Instead, in this section, we aim to derive the specic exclusion rate as the function of the true planet radius, i.e., the probability that a CTP system can show detectable depth variations if the hosting planet has a radius of Rp. These formalisms can accommodate any forms of PCTP(Rp) to derive the overall exclusion rate. We will rst discuss how dierent spectral-type binary would in uence the specic exclusion rate in Section 3.2.1, then we present the nal formalism in Section 3.2.2. 3.2.1. Depth variation detection probabilty for CTPs with dierent spectral types According to Equation 2, the amplitude of depth variation of CTP model comes from the color-dierence between the binary members and the radius ratio between the transiting planet and secondary star. In this regard, we may expect the amplitude of CTP's depth variation is correlated to the spectral types of binary members and the true radius of the planet. However, the detectability of depth variations not only depends on the amplitude of variations but also on the planet orbital periods and apparent Kepler magnitude, which are needed to calculate the precision. To start with, we simulate a simplied population of CTPs, which all locate at the same distance (500 pc) and uniformly have a Jupiter-size planet. We calculate the SNR TESSKPof one-transit signal, so that the SNR TESSKP is mostly determined by binary's spectral type composition. In the left panel of Figure 8, we present these CTPs' SNR TESSKPas well as the eective temperatures of binary components. It can be seen that CTP instances with similar SNR TESSKPform a distinctive U-shape pattern, which suggests these CTPs, despite that they are made up of dierent spectral types of stars, have a similar amplitude of depth variation. For a given primary star, the SNR TESSKPincreases as the from one, in which the binary components have more discrepant color and the amplitude of depth variation subsequently increases, too. Then SNR TESSKPdecreases as keeps dropping after some point. When the primary star is much more massive than the secondary star, its ux contamination is so heavy that the apparent transit depth is attenuated on both photometry band, and will be small. The primary's dilution eect on SNR TESSKPis most extreme at the bottom of the gure, where the binaries are mainly an M dwarf bound to an F/G dwarf, rendering the depth variations to be U-shape pattern no longer holds if we add in distance and planet orbital distributions to CTPs (the right panel of Figure 8), which introduce uctuations to the eective precision of individual CTP system. However, the of CTP depth variation is determined by the system's spectral type composition. Therefore we may expect there are certain populations of CTPs that are most amenable to display identiable depth variation. To evaluate the probability that the depth-dierence signal arising from a binary system with certain spectral types can attain sucient SNR RFKPin a Kepler -like survey. We redo the simulation in the right panel of Figure 8; in each trial we change the binaries' distances and planets' orbital periods. We calculate the fraction of systems whose SNR RFKPis larger than 5 in gridded Finally we average the value over ten trials. We present the depth variation detection probabilities of CTP systems for each reference bands in Figure 9. Contours are the radius correction factor XRfor CTPs with dierent spectral types. Most notably, the Kepler -Ksphotometry can resolve the depth variation of CTPs with wide ranges of spectral Although CTPs comprising stars with nearly identical spectral types ( are hard to show detectable depth variations, the planet radius estimate from this CTP population is close to the true value so they do not plague the planet classications as severely as the other CTPs do. Usingzand TESS band as the reference band is not as eective in resolving the depth variations. The CTP system involving a K dwarf primary ( 4500K) and an M dwarf secondary ( 3500K) is most ameanble to display resolvable depth variations for Kepler -zand -TESS photoemtry, even in the case where z;TESS=KP= 10. The adequate and the small size of the secondary star would allow larger amplitude of depth variations and render this CTP population the optimal targets for the search of Wang et 3, depth variations can be resolved for more than 40% of the CTPs with of 0.6-0.9. CTPs containing stars with very similar spectral types (i.e. two G dwarfs or two F dwarfs, etc.) or with very dierent spectral types (i.e. an F/G dwarf paired with an M dwarf) are still dicult to display detectable which is due to the lack of color-dierence and the heavy ux contamination of primaries (respectively), as explained above. When TESS and zband can achieve a Kepler -like photometric precision, one can expect to observe the depth variations of most CTPs, but only comparable to the most noisy Ksphotometry ( K=KP= 10). Figure 8. Left: an example of the SNR of equal-distance ( 500 pc) CTP simulations plotted against the primary star eective temperature Te;priversus the temperature ratio Te;sec=Te;pri. CTPs depth variations (black crosses) are mainly binary systems having identical primary and secondary stars. Therefore there is little color-dierence in between. The other part of the diagram shows a U-shape distribution at a xed value of SNR TESSKP, which is the result of primary dilution eect and stellar color-dierence; Right : the same as the left gure, but the CTPs are spatially distributed far out to 1800 pc. The spatial distribution will randomly decrease the photometric noise of the system 3.2.2. Depth variation detection probability of CTPs with dierent planet radii We present a framework to calculate the overall exclusion rate of Kepler -detected CTPs. We list the following key ingredients needed to compute the overall exclusion rate of Kepler -detected false positives in double-band photometry. We start with the probability that the planet in CTP conguration having a radius of Rp,PCTP(Rp). Only a fraction of CTPs with given planet radii can be still detectable on Kepler band after the dilution. To obtain the modied planet radius distribution of Kepler -detected CTPs, we derive the Kepler detection probability for CTPs This quantity is plotted as the black thick line in Figure 10. More than 70 % of giant can still be detected by Kepler band. However, the detection probability will decrease sharply for small planets as their apparent transit depths are susceptible to get diluted to be undetectable by the primary star. For those Earth-sized planets that occur in CTP systems, there is nearly zero chance for their diluted transits to be detected by Kepler band. Next we calculate the specic exclusion rate for a Kepler -detected CTP with given This quantity can be obtained by integrating the detection probability shown in Figure 9 over the whole The expected specic exclusion rates computed with dierent reference bands and photometric precisions are displayed as the colored lines in Figure 10. The specic exclusion rate increases with the radius of the resident planet since the amplitude of depth dierence scales with planet size. Earth-size planets' specic exclusion rate is nearly zero for all reference bands. Ksband has a specic exclusion rate larger than 80% for all planets larger than 0 :5Rjup. The reference band photometric precision has a small impact on the Ksband's specic exclusion rate while aecting the rate ofzand TESS more seriously. The specic exclusion rate for noisy ( K=KP= 10)Ksphotometry is comparable to that ofzand TESS with best photometry precision ( z;TESS=KP= 1, 3, 10, 1, 3, 10, Rp=Rjup Ks 4000 4500 5000 5500 6000 6500 7000 Teff,pri/ 1, 4500 5000 5500 6000 6500 7000 Teff,pri/ 3, 4500 5000 5500 6000 6500 7000 Teff,pri/ 10, rate / rate / rate / rate / rate / rate / rate / rate / rate / % Figure 9. The probability that a CTP system show detectable depth variations, in terms of the eective temperature of primaries and secondaries. The probability is evaluated in a grid dened in averaged over 10 trials for each reference band and noise. Blocked area is the region where the mass of secondary star is below 0.1 M. Contours are the planet radius correction factor dened as the same in Ziegler et al. (2016). Based on all the quantities, the overall exclusion rate, fCTP ex, can be computed. With selected planet radius function in CTPs, the probability density function of a Kepler -detected CTP having a planet with radius Rp,PCTP KP(Rp), can be written as PCTP KP(Rp) (7) N0is the nomalization factor. The overall exclusion rate can be computed by integrating the specic exclusion rate with Equation 7 over Rp fCTP ex (8)14 Wang et al. 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 Planet true radius / sz/sKP= 1 z, sz/sKP= 3 z, sz/sKP= 10 Ks, sKs/sKP= 1 Ks, sKs/sKP= 3 Ks, sKs/sKP= 10 TESS, sTESS/sKP= 1 TESS, sTESS/sKP= 3 TESS, sTESS/sKP= 10 PKP Figure 10. This gure presents the probability of a CTP system will show detectable double-band depth variation (colored lines) and the probability of a CTP system can show detectable transit signals on Kepler band (black thick line), as a function of the true radius of resident THE FALSE POSITIVE PROBABILITY WITH DOUBLE-BAND PHOTOMETRY In the previous section, we take a clear cut in depth variation SNR and use this as the sole metric to identify BEBs and CTPs, and estimate their rule-out fraction. While a depth-variation detection limit of ve is acceptable, it is desirable to evaluate the signal's status of being a false positive quantitatively. As the apparent depth behaves dierently between false positive and genuine planet transit, we could incorporate the into the false positive probability (FPP) computation, yielding a more improved and reliable result. This section aims to demonstrate the capability of multi-color photometry in constraining the FPP of binaries and companion transiting planets. The calculation of FPP makes use of the details of the curve and quantify how well they agree with the expectance of each false positive scenario. We parameterize the double-band transit signals into only two observables: the apparent ransit depths on Kepler band and reference band. Rather than including all conceivable false positive scenarios in the FPP estimate, which is beyond the scope of this work, we present examples where only a certain FP scenario is most probable. As a result, the FPP calculation will be simplied while the results remain reasonable in showing the FPP's improvement realized by Adopting a Bayesian probabilistic approach, the FPP of a transit signal can be expressed as: FPP the a priori of the hypothesis, and Lis the likelihood of the hypothesis evaluated by the observed ransit depths in two bands, describing how well the depths obtained from double band photometry match those predicted from the corresponding hypothetical model. The subscript TP denotes the true positive hypothesis, which is usually the case of a genuine planet transit with the right radius estimate, and FP denotes the false positive . We will point out the denition of each hypothesis in the following case study. 4.1. Background eclipsing binaries mimicking Earth-sized planets False positives involving an eclipsing binary can be categorized into three groups: undiluted eclipsing binaries triples (HTs), and blended background eclipsing binaries (BEBs), and each of them tend to produce dierent orders of eclipse depths, with EBs mimicking Jupiter-size transits, HTs mimicking Neptune-size transits and BEBs mimicking the sub-Neptune size and smaller planet transits (Morton 2012; Morton et al. 2016; Fressin et al.15 2013). We note that this statement is not denitive but empirical in the sense that heavy dilution will attenuate the eclipse depth shallow. Driven by this inference, we evaluate the impact of multi-color photometry on the FPP by the example: an earth- sized planet transiting signal (100 ppm) found from a bright ( mKP= 11) G dwarf (one solar mass) located in low galactic latitude region in Kepler FoV (b7 degree). The reason that we assign shallow depth and low-latitude region to the signal is that, under such circumstance and among all the possible false positive scenarios, background eclipsing binary is most likely to represent the false positive hypothesis in the FPP calculation (Fressin et al. 2013; Batalha et al. 2010). The true positive scenario, in this case, is a genuine Earth-size planet transiting the target star. 4.1.1. FPP to equation 9, we determined the prior and likelihood of BEB and planet hypothesis in the of BEB ||The prior probability of BEB BEBshould be the product of (1) the probability of a background star being blended within the aperture of the target and (2) the probability of the blended star being an The rst term is set to be 0.37, derived from Figure 1 of Morton & Johnson (2011); the second term is set to be 0.79%, which is the occurrence rate of short-period eclipsing binaries discovered in Kepler eld (Slawson et al. 2011). We have pBEB= 0:003. Prior of planet ||To obtain the prior probability of the signal being a genuine transiting Earth-size planet, we rely on the frequency of detected Earth-size planets in entire Kepler survey. We count the number of planets and candidates from KOI lists, then divide the number by the number of KIC stars, which of BEB ||To create the likelihood space of BEB, we generate background EBs from the following the same formula described in 2.2.1, except that all foreground stars are subject to be solar-like G dwarf stars (a gaussian mass distribution of 1 0:05M). Then we compute the diluted primary eclipse depths inKepler band and reference bands. The sample's logarithmic Kepler depth logKP, and double-band depth be used to dene the two-dimensional probability density function (PDF) of BEB hypothesis. The likelihood is then LBEB=Z PDF sim()PDF obs()d (10) where is the vector of (log KP; RF=KP), andPDF simis the posterior PDF mapped by BEB simulation, PDF obs is the posterior PDF of observed signal, which is a gaussian distribution centered on , with depth uncertainty as its standard deviation. We dene the observed KPat 10020 ppm, and leave the RFas a varying value. We set the depth uncertainties on the reference band, RF, to be 20 or 50 ppm. Likelihood of planet ||For the planet hypothesis, we compute the planet radius based on the observed and the simulated foreground star radius. Then we compute the apparent transit depth on reference bands with limb-darkened transit model (Mandel & Agol 2002), with appropriate quadratic limb darkening coecients taken from Claret & Bloemen (2011) and Claret (2017). The likelihoods of planets is evaluated in the same way as for the BEBs. Figure 11 shows an example of the PDFs of BEB and planet hypothesis. The expected depth ratio distribution of each hypothesis is quite dierent. BEBs tend to produce deeper the background binaries are generally of low masses and appear consequently red in color, while the foreground G dwarf is less luminous in Thus the ux contamination of foreground stars will be less signicant on the redder reference bands. The depth ratio expected for planet transits is smaller than 1, which is due to the stellar limb darkening eect, which states that light with longer wavelengths is more evenly distributed on stellar disk than light with shorter the planet will block more blue light than red light, and the RFwill be smaller than KP. Moreover, the most probable depth ratio expected for each hypothesis, represented by the area with the deepest color in Figure 11, is also unique when adopting dierent reference bands. Kepler -Kspair has the most for two hypotheses, with the most probable ratio value of 1.5 for BEB hypothesis and 0.88 for planet transits. Kepler -TESS photometry has the least separable depth ratios for two hypotheses. The BEB's and the planet's likelihood distributions results from the photometry pair's capability to reveal the color change during the eclipse/transit, which is consistent with the pattern we demonstrate in Figure 4 and 7.16 Wang et 11. Likelihood space for BEB and planet scenarios described in Section 4.1. The area with deeper color means a that our hypothesis simulations have a certain value of (log KP; RF=KP) in the region. Note that the depth ratio for two hypotheses is dierent. BEB tends to show a larger apparent depth in the redder reference band compared to Kepler . Planet transits show a smaller depth due to stellar limb darkening. The observed depth ratio for dierent photometry pair is separated at dierent levels, with Kepler -Kspair showing the most distinctive depth ratio distribution for BEB and while the Kepler -TESS shows a similar depth ratio between the two hypotheses, in both cases the depth ratios are close to 1. 4.1.2. Improved FPP from double photometry We present the improved FPP results in Figure 12, and compare to the single-band FPP. Besides selecting three dierent photometric bands as the reference bands, we also provide two cases when the depth uncertainty is by solid and dashed line). The general trend is that as the observed depth ratio increases, the FPP goes higher. When the observed depth ratio is close to unity, the FPP will decrease, supporting the planet hypothesis. It is also shown from the gure that the order of FPP increment or decrement varies across reference bands. Ksband shows the best performance in discriminating the planets and FP scenarios. If Ks=KP= 1, the FPP will decrease by an order. Also, it takes a smaller depth ratio for the Kepler -Kspair to fully conrm the false positive origin of the signal than the other two photometry pairs (1.3 and 1.7 for the two cases). This is because Kepler -Kspair has the most discrepant for planet and BEB hypothesis (see Figure 11), which makes it easier to distinguish the underlying hypothesized scenario that creates the observed depth ratio. Secondly, the comparison between two cases with dierent depth uncertainties suggests that the actual value of depth uncertainty will aect the double-band FPP more signicantly than the dierent choice of reference band. For example, if a depth ratio of 1.3 is observed, the corresponding FPPs derived from RF= 50 ppm are generally ten times lower than the case of 20 ppm, while FPPs are close to each other for all three photometry pairs at a given RF. A result of this eect is that larger uncertainties in depth measurement would require a larger observed depth ratio to falsify a signal. The critical depth ratio, which makes the FPPs increase to nearly 100%, is 1.3-1.4 and 1.7-1.8 for RF= 20 ppm and RF= 50 ppm, we may compare the FPP results with the depth-dierence identication method adopted in Section 3. If we calculate the SNR RFKPfor these critical depth-ratios, the value is only around 1.4, which is well below the threshold of 5 we set in Section 3. This comparison suggests that incorporating the double-band photometry in FPP calculation can more easily identify false positives. 0.95 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 Depth ratio BEBz:sRF= 20 ppm z:sRF= 50 ppm K:sRF= 20 ppm K:sRF= 50 ppm TESS:sRF= 20 ppm TESS:sRF= 50 ppm Figure 12. The false positive probability that an Earth-sized candidate is caused by a background eclipsing binary, evaluated by the observed double-band depth-ratio. The black dotted line is the single-band FPP derived from Kepler photometry alone. The solid and dashed lines are cases when the depth observed by reference band is of dierent uncertainties. See Section 4.1 for more discussion. 4.2. Which is the planet-host the planet-host star in a binary system is crucial in correcting the dilution of observed transit the double-band photometry FPP calculation to the model of companion transiting planets provides a natural way to distinguish the planet-host star in a binary system. We illustrate this by providing a specic observation event. Suppose a Super-earth-size transit candidate ( KP200 ppm) is observed from a solar-like G dwarf, but it is unknown whether the target star is single or not. To predict the multi-band transit depths for each hypothesis, we simulate binary populations with a primary mass of around one solar mass and a varying secondary mass, according to the procedure described in Section 2.2.1. We assume that, before the transit on reference band is obtained, the apparent depth observed by Kepler alone provides no clues about the host star of the planet; that is, the planet is likely to18 Wang et al. orbit either star in the binary with equal probability, hence TP=FP= 50%. Here we dene the true positive as the planet orbiting the brighter target star, while false positive as the planet orbiting the fainter companion of the target star. To determine the likelihood of each hypothesis, we rst draw the apparent KPfrom a Gaussian distribution with a mean value of 200 ppm and a standard deviation of 20 ppm; the latter quantity is the depth uncertainty KPwe assigned to the Kepler photometry. Then we recover the true planet radius from the apparent depth KPfor each instance, assuming the star is the planet-host star. We reject those instances with a recovered planet radius larger than 22 Rto avoid overestimation from systems that unlikely exists. Then we calculate the observed depth on reference band RFand establish the PDFs. Finally, we evaluate the likelihoods in the similar manner described in Section 4.1.1. We present the likelihood distribution of the two hypotheses in Figure 13. If the primary star is the host star, the apparent radius of the planet is close to the true value, and the depth observed on dierent photometry band will not dier a lot. This is because the primary star usually makes up most total photometry. Therefore the variance of contamination on dierent photometry bands would be relatively small. The situation can be reversed when the secondary star is the host star, which is shown in the right panel of Figure 13 that the apparent transit depth observed in the reference bands can be up to 5-10 times of that observed on Kepler band. This is because the true planet radius is usually much larger than the apparent radius so that the depth variation will increase, too. The apparent depth on the Ksband can vary in a wide range, from 170 ppm to 2000 ppm. On the other hand, TESS and zband expect a much narrower possible RF=KPrange. We calculate the probability of planet host star based on a set of transit depth ratio values and present the result in Figure 14. We also provide two cases with dierent RF. As expected, the depth ratios close to unity will decrease the FPP, which suggests that the target star is indeed the planet host star. FPPs increase with the non-unity value of depth ratios, which alerts that the planet may be transiting the secondary companion of the target star. Larger depth uncertainties from the reference bands will lead to a loosely constrained FPP value. We nd the SNR RFKPof the critical depth ratios to be 0.7 and 1.2, for two cases, 0.90 0.95 (Primary 5.0 (Secondary host) Depth ratio dRF/dKp Figure 13. The likelihood distribution of the transit depth if an apparent super-earth-size ( KP200ppm) candidate is observed on three reference bands, under the assumptions that the planet is transiting the primary star (left) or the secondary star (right).19 0.95 1.0 1.1 1.2 1.3 1.4 Depth ratio CTPz:sRF= 20 ppm z:sRF= 50 ppm K:sRF= 20 ppm K:sRF= 50 ppm TESS:sRF= 20 ppm TESS:sRF= 50 ppm Figure 14. The false positive probability that a super-earth-size planet candidate is transiting the unseen stellar companion of the solar-like target star, evaluated by double-band depth-ratio. The black dotted line represents the uninformative that the planet is equally likely to orbit either of the stars in the binary system inferred from one-band = 50% for each star). The solid and dashed lines are cases when the depth observed by the reference band is of See Section 4.2 for more Comparing two multi-color approaches to identifying false positives In this work, we demonstrate two methods by which we can use the color-dependence feature of transit candidates to reveal and rule out false positives. The rst method is illustrated in section 3, where we measure the SNR of the dierence between observed depths and identify those false positives whose SNR of depth-variation is large enough to condently conrm that the observed transit depth in two photometry is statistically inconsistent, thereby pointing out the possible presence of blending stars or the chance-alignment of eclipsing binary. In section 4 we adopt the second method. We compare the observed depth ratio with a comprehensive set of depth-ratios simulated under on the observed transit signal's origin, and we adopt a Bayesian approach to calculate the false of the signal based on double-band photometry. The underlying assumptions for these two methods are a bit dierent. The basic assumption for the rst method is that the apparent depth of false positives should be dierent. To identify this dierence, we set the detection threshold SNR RFKPas 5 to evaluate whether the dierence is real or due to statistical uctuation. With this said, satisfying this detection criterion (SNR RFKP>5 and SNR KP>7) will require the signal to be observed with sucient high SNR on reference band (SNR RF>12), which is not only determined by the color-dierence between the blends and targets, but also by the reference band's relative eective wavelength compared to the main photometry, and the thereof. The approach is widely adopted in the early ground-based wide-eld survey to reject transit et al. 2006; Col on et al. 2012), but is mainly subjected to vetting the transit signals' with deep depths (1%) for which sucient SNR can be attained. Small planet transit candidates may not be the optimal targets for multi-color vetting; as suggested by Figure 7, the detectability of small planet candidate's depth variation is most susceptible to the choice of reference band and precision. Thus, although we could identify blends and exclude them based only on one such metric, the denitive threshold will implicitly demand high SNR on the signal detection by reference photometry. The expected yield of false positive identication is strongly limited by the specics of reference band and noise systematics. The basic assumption for the FPP calculation is that the double-band depth ratio of planet transit is dierent from that of false positives. In simpler terms, we could falsify a transit signal as long as we can tell it does not resemble the planet transit's light curves. This is why we nd in Section 4 that the critical depth ratio's corresponding SNR RFKP is usually much smaller than 5 since the observed depth ratio cannot be reproduced by most of the systems. However, since we adopt a Bayesian approach to achieve this goal, the value of double-band20 Wang et al. FPP will indeed vary as we change each hypothesis's a priori probability and as we change the way we likelihood PDFs. If the a priori probability of the planet hypothesis is larger, the corresponding FPP will be generally lower than we note that this is the common feature of Bayesian probability, which strongly depends on the models and assumptions. One advantage of double-band FPP, which we did not explicitly address in this work, is that it can work in parallel with other observational constraints to explore and narrow down the possible blending systems that can produce the double-band light curves, e.g., high-resolution imaging can reveal both physical companions in wider orbits and line-of-sight objects. The dilution information derived from magnitude dierence helps to locate the star that harbours the observed signal (Furlan et al. 2017; Ziegler et al. 2016). For targets including several stars, which are hardly resolved, joint analysis of multi-color transit light curves and SED of the targets may provide of the multiplicity and conguration of the system (Miyakawa et al. 2021). 5.2. The role of limb darkening Limb darkening will also cause the planet transit's depth to vary on dierent bands. This is because uxes with dierent wavelengths distribute on stellar disk dierently, with the center of the disk appear bluer and the limb appears redder. As a consequence, the transit depths observed on bluer photometry will be deeper than redder projects RF=KP<1 since our considered reference photometry are all redder than KP. To this end, one may question that whether using SNR RFKP>5 as the sole metric to reject blends exclude the cases of limb-darkened planet transits signals. However, for the false positives considered in this paper the apparent depth ratio would be RF=KP>1. According to our denition of BEB and CTP in 2, the blending components , which generate the are often redder in color compared to their from target population ) and thus the apparent depth would be deeper on red reference photometry. Therefore, we could discriminate blends from limb-darkened transits by comparing the apparent depth ratio. One exception would be grazing giant planets for b > 0:8 (Tingley 2004), which only pass through the limb of star and will block more blue light than red light. In this case the apparent depth ratio would be the chromaticity of stellar blends. However, transits are rather rare cases and we do not consider it when using color-dependent feature to reject blends. 5.3. Impact of dierent stellar properties from KIC Though our target star population mainly consists of G-type stars as KIC included, we were unable to fully recreate the KIC's stellar mass and eective temperature distributions in our simulation (as shown in Figure 1). These as demonstrated in Section 3, are the major stellar properties that determine the detectability of double- band depth variations. The deviation between the simulation and KIC stars partly stems from the fact that we employed a more physical model (see Section 2.2) to construct the stellar population, while KIC is a specially curated catalog following some selection rules. As for this deviation's impact on the nal results. Qualitatively speaking, our target star population is on average hotter than KIC stars. In the BEB model, hotter foreground stars tend to produce earth-like transit signals, because they are brighter so the dilution is heavy, therefore the rule-out rate of BEB signals (summarized in Figure 7) is aected the most for earth-size BEBs, the other sizes of BEB signals only being minorly aected. The excess of BEBs with hotter foreground stars leads to an overestimation of earth-like BEB rule-out rate. As for the overall rule-out rate (summarized in Figure 4&7), the case for RF=KPis aected the most, while for other reference band noise levels, the identiable earth-like BEBs only make up a small fraction of the total rule-out rate and thus are not CTP model, we demonstrate that CTPs with hot primaries are the hardest to rule out at low reference band precision, espeacially for zand TESS band (see Figure 9). Thus, we expect the excess of hotter stars in our sample will lead to an underestimation on the rule-out probabilities given planet radius (summarized in Figure 10) for zand TESS band in RF=KP= 3;10. However, we do manage to construct a KIC-like target star simulation to quantitatively measure the uncertainties in nal results, even though we adopt some unphysical input distributions. The rule-out probabilities in both BEB and CTP models vary within 10%, even for the most severely aected this work, we study how many false positives can be identied in terms of their chromatic depth feature if band dierent from Kepler band were added and observing with Kepler . We mainly focus on two types of false positives. One is that when the target star is blended with a pair background eclipsing binary (BEB); another is when the planet is transiting the unresolved stellar companion to the target star (we term this as companion or as CTP). We build up physics-based models for these two false positives. To systematically estimate what fraction of BEBs and CTPs can be identied through their color-dependent feature, we simulate representative BEB and CTP populations according to Kepler observations and galactic star population models. We derive their multi- band photometry based on isochrone interpolation and bolometric correction. We calculate the signal-to-noise ratio of the dierence of observed transit depth, SNR RFKP, observed between Kepler and reference bands, which include z,Ks, and TESS. We identify the subset of false positives whose SNR RFKPis larger than ve as the fraction of false positives that can be readily ruled out and show how this fraction varies with dierent levels of photometric precision we assigned to the reference bands. Some false positives cannot gain sucient high SNR RFKP. However, since the depth variation predicted by false positive hypothesis and planet hypothesis is quantitatively dierent, our enabled us to calculate the transit signals' false positive probability based on double-band photometry. We summarize the conclusion of this work as eclipsing binaries (BEBs) can be eectively ruled out by combined identication of dierence (Equation 2) and primary and secondary dierence (Equation 5); We list the overall rule- out fraction via identifying the depth-dierence of BEB in Figure 7. Ks,z, and TESS, when adopted as the reference band to Kepler , can rule out 95%, 80%, and 65% of Kepler candidates mimicked by BEBs if both of the photometry can attain a precision like Kepler 's. Assuming a larger photometry precision of the reference band, the overall rule-out fraction of TESS and zband will drop drastically towards around 10%. The sharp decrease is mainly due to the undetectable depth variation of shallow-depth candidates that comprise most BEB signals, and only those signals with deep depths can show detectable depth variations. However, KeplerKs dual photometry is only slightly aected even if it is assigned with the largest noise, i.e. ten times of reaching 70% which is generally better than the TESS and zband. * *The amplitude of the depth variations of CTPs comes from the true planet radius and the color-dierence between the binary stars the planet resides. We calculate the detection probability for Kepler -detected CTPs with certain spectral type binary members (Figure 9). The CTPs that consist of the smallest stars, such as a K dwarf and an M dwarf, often show detectable depth variations, while CTPs are the hardest to identify for all the considered reference bands. Next, we present the detection probability of CTPs with given true planet radius, averaged over all possible CTP binary masses (Figure 10). Using the provided detection one can obtain the overall exclusion rate of CTPs expected in a Kepler -like survey by integrating the radius distribution of planets transiting the secondary star in a binary system. *We adopt a Bayesian approach to calculate the false positive probability (FPP) based on the observed depth ratio between two photometry observations. We apply the framework to identify whether an Earth-size candidate is a diluted signal from a background eclipsing binary. For given transit SNR on the reference band, Ksband can better constrain the FPP of signal better than zand TESS band. However, the photometric precision of the reference band would in uence the degree of FPP improvement more than the specic choice of reference using double-band photometry to determine the FPP and reveal blends usually takes less demand on photometry precision compared to identifying the blends via SNR RFKP>5 criterion. (Figure 12). *Applying the FPP calculation to our CTP model can assist in guring out whether an Earth-size transit signal is a larger planet transiting the bound companion of the target star, or equivalently identifying the planet host star in the binary system. For this purpose, assuming that the planet is equal-likely to orbit either of the stars inferred from initial single-band observation, the probability of the primary host can be up to 99% if the observed depth ratio is nearly unity. The conrmation of the secondary host will need the observed SNR RFKPto be at least 0.8 (Figure 14). With a more exible treatment of input assumptions about eld stars' properties, the simulation procedure we have developed can have broader applications to other surveys as long as we have the priors and models specic to the surveyed area, whether physical or empirical. One possible application is to use in the follow-up of TESS, as the pixel22 Wang et al. scale of TESS (21) is much larger than that of Kepler (4); thus, a considerable amount of blended stars will be imaged. Merited for its economical cost in a single observation, multi-color photometry can extensively search for clues of false positive in Kepler if future missions employ it, and the valuable time of advanced telescopes would be work is supported by the National Natural Science Foundation of China (Grant No. 11973028, and the National Key R&D Program of China We also acknowledge the science research grants from the China Manned Space Project with Astropy (Astropy Collaboration et al. 2018), Matplotlib (Hunter 2007), Pandas (pandas development team 2020), isochrones, (Morton 2015), scipy (McKinney Collaboration, Price-Whelan, A. M., Sip} ocz, B. M., et al. 2018, AJ, 156, 123, doi: N. M., Rowe, J. F., Gilliland, R. L., et al. Journal Letters, 713, doi: N. M., Rowe, J. F., Bryson, S. T., et al. Journal, Supplement Series, 204, doi: W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977, doi: W. J., Koch, D. G., Basri, G., et al. Journal, 736, doi: L. G., Masuda, K., & Winn, J. N. 2018, arXiv, 155, 244, doi: T. M. 2003, The Astrophysical Journal, 593, L125, doi: T. M., Latham, D. W., Everett, M. E., & Esquerdo, G. A. 2011, AJ, 142, 112, doi: S. T., Tenenbaum, P., Jenkins, J. M., et al. Journal Letters, 713, 1, doi: S. T., Jenkins, J. M., Gilliland, R. L., et al. of the Astronomical Society of the Pacic, 125, 889, doi: L., Sch onrich, R., Asplund, M., et al. 2011, Astronomy and Astrophysics, 530, doi: J., Dotter, A., Conroy, C., et al. 2016, Journal, 823, 1, doi: D. R., Beichman, C. A., Horch, E. P., & Howell, S. B. 2015, Astrophysical Journal, 805, 1, doi: A. 2017, A&A, 600, A30, doi: A., & Bloemen, S. 2011, A&A, 529, A75, doi: M., Wheaton, W. A., & Megeath, S. T. 2003, The Astronomical Journal, 126, 1090, doi: on, K. D., & Ford, E. B. 2011, Publications of the Astronomical Society of the Pacic, 123, 1391, doi: on, K. D., Ford, E. B., & Morehead, R. C. 2012, Monthly Notices of the Royal Astronomical Society, 426, 342, doi: on, K. D., Morehead, R. C., & Ford, E. B. 2015, Monthly Notices of the Royal Astronomical Society, 452, 3001, doi: J. L., Thompson, S. E., Bryson, S. T., et al. 2014, Astronomical Journal, 147, doi: esert, J. M., Charbonneau, D., Torres, G., et al. Journal, 804, doi: az, R. F., Almenara, J. M., Santerne, A., et al. 2014, Monthly Notices of the Royal Astronomical Society, 441, 983, doi: A. 2016, The Astrophysical Journal 222, 8, doi: P. P. 1983, ApJ, 268, 368, doi: F., Torres, G., Pont, F., et al. 2012, 745, doi: F., Torres, G., Charbonneau, D., et al. Journal, 766, doi: E., Ciardi, D. R., Everett, M. E., et al. 2017, The Astronomical Journal, 153, 71, doi:

2108.11028

Mutian Wang

Mu-Tian Wang, Hui-Gen Liu, Jiapeng Zhu and Ji-Lin Zhou

Follow-up photometry in another band benefits reducing emph{Kepler}'s false positive rates

Revised under first referee report from AJ; 23 pages, 14 figures

10.5281/zenodo.5126472

astro-ph.EP astro-ph.SR

http://creativecommons.org/licenses/by/4.0/

Kepler Mission's single-band photometry suffers from astrophysical false positives, the most common of background eclipsing binaries (BEBs) and companion transiting planets (CTPs). Multi-color photometry can reveal depth feature of false positives and thus exclude them. In this work, we aim to estimate the fraction of false positives that are unable to be classified by Kepler alone but can be identified with their feature if a reference band (z, Ks and TESS) were adopted in We build up physics-based blend models to simulate of false positives. Nearly 65-95% of the BEBs and more than 80% of the CTPs that host a Jupiter-size planet will show detectable depth variations if the reference band can achieve a Kepler-like precision. Ks band is most effective in eliminating BEBs exhibiting any depth sizes, while z and TESS band prefer to identify giant candidates and their identification rates are more sensitive to photometric precision. Provided the radius distribution of the secondary star in binary systems, we derive formalism to calculate the overall identification rate for CTPs. By comparing the of the double-band depth ratio for BEB and planet models, we calculate the false positive probability (FPP) for typical Kepler we show that the FPP calculation helps distinguish the host star in an unresolved binary system. The analysis framework of this paper can be easily adapted to predict the multi-color photometry yield for other transit surveys, especially for TESS.

2 Shirasaki et al. 0 2 4 6 8 1014h-1M Tinker et al. et al. (2011) Watson et al. (2013) Despali et al. (2016) This study Figure 1. Top-hat mass variances 2covered by several N-body simulations. The red open region represents the parameter space explored in this paper. Our paper aims at calibrating the halo mass function at lower-mass regimes ( previous studies at redshifts of z<7. For comparison, the shaded, orange lled, purple lled, and green lled regions show the coverage in the 1-zplane in Tinker et al. (2008), Bhattacharya et al. (2011), Watson et al. (2013), and Despali et al. (2016), respectively. Note that black dashed lines provide the linear-theory predictions at M= 1014h1M, 108h1Mfrom top to bottom. mass density. The variance of the linear density eld smoothed by a top-hat lter with R, denoted as 2(R), is then used to characterize the mass fraction of dark mat- ter halos of M. Under simple but the analytic approaches predict that any dependence of halo masses, redshifts, and models in the halo mass function can be determined by the variance (R) alone. One can factor out a puredependence in the halo mass function with the analytic approaches. This dependence is known as the multiplicity function f(), expecting that it is a universal function for dierent halo masses, redshifts, and cosmological simulations have validated the universality of the multiplicity function so far. Jenkins et al. the redshift- and of the multiplicity function to be less than a 15% level when the halo mass is dened by the algorithm (Davis et al. 1985) with some FoF link- ing lengths. Dierent denitions of halo masses can in- troduce a systematic non-universality of the in simulations (White 2002; Tinker et al. 2008;Diemer 2020). As increasing particle resolutions, several groups have found that the multiplicity function for the FoF halos depends on cosmological models with a 10% level (e.g. Warren et al. 2006; Bhattacharya et al. 2011) but its redshift dependence is weak (e.g. Watson et al. 2013). Recently, Despali et al. (2016) have claimed the universality of the multiplicity function in their simula- tion sets when dening the halo mass with a virial spher- ical overdensity and expressing the multiplicity function in terms of a re-scaled . In this paper, we extend previous measurements of the multiplicity function toward lower halo masses. Figure 1 summarizes the coverage of halo masses and redshifts in our paper. In the gure, we convert the halo mass scaleMto the top-hat variance 2(R) using the linear matter power spectrum in the best-t CDM in Planck Collaboration et al. (2016, Planck16). The red shaded region in Figure 1 shows our coverage, while other shaded and hatched regions represent ones in some of previous studies. Our measurements of f() in- clude the range of wide redshifts of z7, which is not explored in the liter-Halo mass function in Planck cosmology 3 ature. We examine the universality of the multiplic- ity function in the Planck16 CDM cosmology from gaseous mini-halos (e.g. Ben tez-Llambay et al. & Frenk 2020) to massive galaxy clus- ters. To do so, we analyze dark-matter-only with dierent resolutions at 40 dierent red- shifts in Ishiyama et al. (2015); Ishiyama & Ando (2020), allowing to study possible redshift evolution of the mul- tiplicity function in details. The paper is organized as follows. In Section 2, we de- scribe the simulation data of dark matter halos used in this paper. In Section 3, we present an overview of the halo mass function, introduce how to estimate the multi- plicity function from simulated halos as well as statistic errors in our measurements. The analysis pipeline to nd the best-t model to our measurements is provided in Section 4. The results are presented in Section 5, while we summarize some limitations of our results in Section 6. Finally, the conclusions and discussions are provided in Section 7. Throughout this paper, we as- sume the cosmological parameters in Planck16. To be specic, we adopt the cosmic mass density m0= 0:31, the baryon density b0= 0:048, the cosmological con- stant = 1 m0= 0:69, the present-day Hubble parameter H0= 100hkm s1Mpc1withh= 0:68, the spectral index of primordial curvature 0:96, and the linear mass variance smoothed over 8h1Mpc,8= 0:83. We also refer to log as the loga- rithm with base 10, while ln represents the natural SIMULATIONS AND HALO CATALOGS To study the abundance of dark matter halos, we use a set of halo catalogs based on high-resolution simulations with various combinations of mass resolutions and volumes (Ishiyama et al. 2015; Ishiyama & Ando 2020). In this paper, we use the halo catalogs based on three dierent runs of 2GC-L (L), 2GC-H2 (H2), and phi1.1Table 1 summarizes speci- cations of our simulation sets. The simulations were performed by a massive parallel TreePM code of GreeM2(Ishiyama et al. 2009, 2012) on the K computer at the RIKEN Advanced Institute Science, and Aterui super-computer at Center for Computational Astrophysics (CfCA) of Na- tional Astronomical Observatory of Japan. The authors 1The halo catalogs are available at Note that 2GC stands for new numerical galaxy c 1. A summary of simulation sets analyzed in this paper. The total number of N-body particles ( Np), the sim- ulation box size on each side ( Lbox), the mass of each N-body particle (mp), and the softening length ( ) are provided for three dierent runs. Note that the softening length is set to be 3% of the mean free path of particles in each 819231120 2:201084:27 H2 2048370 2048332 the initial conditions for the L and H2 runs by a publicly available code, 2LPTic ,3while another public code MUSIC4(Hahn & Abel 2011) has been adopted to generate the initial conditions for the phi1 run. Note that either public code uses second-order theory (e.g. Crocce et al. 2006). All sim- ulations began at z= 127. The linear matter power spectrum at the initial redshift has been computed with the online version of CAMB5(Lewis et al. 2000). In the simulations, the Planck16 cosmological model has halo catalogs in this paper have been produced with the ROCKSTAR halo nder6(Behroozi et al. 2013). We focus on parent halos identied by the ROCKSTAR al- gorithm and exclude any subhalos in the following anal- yses. We keep the halos with their mass greater than 1000 times mp, wherempis the particle mass in the N-body simulations. Throughout this paper, the halo mass is dened by a spherical virial overdensity (Bryan & Norman 1998). We analyze the halo catalogs at 40 dierent redshifts below: z=0.00, 0.03, 0.07, 0.13, 0.19, 0.24, 0.30, 0.36, 0.42, 0.48, 0.55, 0.61, 0.69, 0.76, 0.84, 0.92, 1.01, 1.10, 1.20, 1.29, 1.39, 1.49, 1.60, 1.70, 1.83, 1.97, 2.12, 2.28, 2.44, 2.60, 2.77, 2.95, 3.14, 3.37, 3.80, 4.04, 4.29, 4.58, 5.98 and 7.00. It is worth noting that we adopt the default halo mass denition in the ROCKSTAR nder. This mass denition does not include unbound particles. Because around a given dark matter halo can contribute to the spherical halo mass, the halo mass function with- out unbound particles may contain additional system- atic uncertainties. In Appendix A, we examine the im- pact of unbound particles in our measurements using a dierent set of N-body simulations. We conrmed Shirasaki et al. unbound particles did not introduce systematic errors in the calibration beyond statistic MASS FUNCTION 3.1. Model In spite of a strong dependence on the matter power spectrum, there exist successful analytical the number density of dark matter halos (e.g. Press & Schechter 1974; Bond et al. 1991; Sheth & Tor- men 2002). Those approaches commonly relate the halos with their mass Mto the linear density eld smoothed at some scale R= (3M=4m)1=3. To develop an ana- lytic formula of the halo abundance, Press & Schechter (1974) assumed that the fraction of mass in halos of mass greater than Mat redshiftzis set to be twice the probability that smoothed Gaussian density elds ex- ceed the critical threshold for spherical collapse, c. In this ansatz, the number density of halos can be ln1 dMf(); (1) whereis the root mean square (RMS) uctuations of the linear density eld smoothed with a lter encom- passing this mass M. The RMS is usually dened with a spherical top-hat lter, 2(M;z) (2) wherePL(k;z) is the linear matter power spectrum as a function of wavenumber kand redshift z, andWTHis the Fourier transform of the real-space top-hat window function of radius R. To be specic, the top-hat win- dow function is given by WTH(x) = & Schechter (1974) found that the function f, re- ferred to as the multiplicity function, is expressed (3) The multiplicity function in Press & Schechter (1974) has been revised later by introducing excursion set the- ory (Bond et al. 1991) and adopting the ellipsoidal col- lapse (Sheth & Tormen 2002). Note that previous an- alytic models predict that the multiplicity function is a universal function and any dependence on redshifts and cosmological models can be encapsulated in by these analytic predictions, we assume that the multiplicity function can be described by a four- parameter model, f(;z) =A(z)r 2 exp + andq(z) are free parameters in the model. In Eq. (4), we introduce the critical overdensity for spherical collapse, c;z. In a at CDM cosmology, this quantity is as (Kitayama & Suto 1996) c;z=3 (12)2=3 20[1 + 0:123 log m(z)]; (5) m(z) = m0(1 +z)3 m0(1 +z)3+ (1 m0): (6) 3.2. Estimator of the multiplicity function To nd a best-t model of f(;z) to the simulation data, we need to construct the multiplicity function from the halo catalogs. We start with a binned halo mass function, which is directly observable from the simula- tions. Let
nbinbe the comoving number density of dark matter halos in a bin size of
log Mwith mass ranges [M1;M2]. Using Eq. (1), one nds
nbin= ln 10ZlogM2 logM1d logMm Md log1 d logMf():(7) In the limit of
log M!0, we logM d log1 d a center of the binned mass. In Eq (8), we require the logarithmic derivative of logwith respect to the halo mass M. This derivative is known to be aected by the size of simulation boxes (e.g. Reed et al. 2007; Luki c et al. 2007), because Fourier modes with scales beyond the box size are missed in the simulations. To account for the nite volume eect on the estimate of f(), we follow an approach in Reed et al. (2007). The mass variance in a nite-volume not equal to the global value in Eq. we assume that the halo mass function in the nite-volume simulation, (d n=dM)loc, can be writ- ten (9) This ansatz is motivated by the consideration in Sheth & Tormen (2002). Here, we dene the halo mass function in the nite-volume simulation (10)Halo mass function in Planck cosmology 127 Theory w/o w/ correction for phi1 Theory w/ correction for 2. Finite-volume eects in computing the linear top-hat mass variance TH. The gray dashed line shows the prediction by linear perturbation theory (Eq. 2), while the blue circles and orange diamonds are the mass variance mea- sured inN-box boxes of phi1 ( Lbox= 32h1Mpc) and H2 runs (Lbox= 70h1Mpc) atz= 127. The blue and orange lines are the mass variances with a correction by Eq. (11). Using Eqs. (9) and (10), we can predict the global mass function with f() =floc() once we calibrate the functional form of flocas a function of loc. It would be worth noting that Eq (8) provides an esti- mate offloc(loc) (notf(loc)) in practice. To eval- uate we directly compute the variance of smoothed density elds at the initial con- ditions of our simulations as varying smoothing scale of R= (3M=4m)1=3. To do so, we grid N-body par- ticles onto meshes with 5123cells using the cloud-in- cell assignment scheme and apply the Fourier Transform (FFT). Figure 2 shows the nite- volume eect of the mass variance measured in the phi1 and H2 runs at z= 127. We nd that the can be approximated as loc(M) =(M)M M0 ; (11) whereM0= give a rea- sonable t to the phi1 run, while M0= can explain locin the H2 run. For the L run, we assume no nite-volume eects on the mass variance and set loc=. To estimate fsim, we rst measure the comoving num- ber density of halos with 160 logarithmic bins in the range ofM= [108;1016]. The bin size
log Mis set to0.05. We then compute the multiplicity function using Eq. (8) with !locas in Eq (11). 3.3. Statistical errors For the calibration of our model with numerical sim- ulations, we need a robust estimate of statistical errors in halo mass functions. In this paper, we adopt an ana- lytic model of the sample variance of d n=dMdeveloped in Hu & Kravtsov (2003). Apart from a simple Pois- son noise, the model takes into account the uctuation of the number density of dark matter halos in a that the uctuation in d n=dMis caused by underlying linear density modes at their scales compa- rable to the simulation box size, one Err[
nbin] represents the statistical error in Eq. [3Vsim=(4)]1=3, and bL(M;z) is the linear bias at the halo mass of Mand redshiftz. The rst term in the right hand side in Eq. (12) corresponds to the Poisson error, while the term ofS(M;z) represents the sample variance. To compute Eq. (13), we use the model of bLin Tinker et al. (2010). We validate the model of Eq. (12) with the L run atz= 0. In the validation, we divide the L run compute the mass function in each subvolume, and then estimate the standard deviation of the mass function over the Nsubsubvolumes. We set each subvolume so that it has an equal volume. Figure 3 summarizes our validation. In the gure, we show the fractional error of the binned mass function and up-turns at higher masses indicate that the sample variance term becomes dominant. Because of Eq. (8), the fractional error offsimis given by Eq. OF THE MODEL PARAMETERS To calibrate our model of the multiplicity function (Eq. 4) with the simulation, we introduce a at a given =2 L() +2 H2() +2 phi1(); represents the multiplicity in the run (= L;H2;phi1) at the i-th bin6 Shirasaki et 0 Nsub= 43 Nsub= 83 Nsub= 163 Nsub= 323 Figure 3. Sample variance of the mass function in a nite volume. The colored points show the fractional error of the mass function in subvolumes in the L run at z= 0, while the error bars represent the Gaussian error. In this gure, we set the simulation results with the bin width being
log M= 0:25 for visualization. From top to bottom, we show the results when dividing the L run into Nsub= 43;83;163, and 323pieces, respectively. Solid lines correspond to our based on Eq. (12), showing a reasonable t to the simulation results for dierent is given by Eq. (4) with the parameters being Erriis the statistical error of the multiplicity function at possible systematic errors at =iin our simulations (we set sys;ilater). To nd we minimize Eq. (14) with the algorithm implemented in the open et al. 2001{). When computing Eq. (14), we reduce the number of bins inby setting the bin width of
(1 =) = 0:05 for the L run and
(1 =) = 0:1 for other two runs. This re-binning of allows to reduce scatters among bins in the analysis and provide a reasonable goodness-of-t for our best-t model. In our calibration, we do not scatters among bins (Comparat et al. 2017), while we expect that our re-binning would make the correlation between nearest bins less important. Also, we remove the data with
nbinbeing smaller than 30 to avoid signicant Poisson uctuations in our tting. After these post-processes, we found 40 data points available at z4:58, while 31 and 25 data points atz= 5:98 and 7:00, respectively. Because our model consists of four parameters, we expect that our simulation data give sucient information to nd a good t. There exist several factors to introduce systematic ef- fects on computing the mass function in N-body sim- ulations (e.g. Heitmann et al. 2005; Crocce et al. 2006; Luki c et al. 2007; Knebe et al. 2011, 2013; Ludlow et al. 2019). Because we set the minimum halo mass to be 1000mpwith the particle mass of mpin the simulation, the nite force resolution does not aect our measure- ment of the mass function beyond a 1 Poisson error (Ludlow et al. 2019). Although our simulations have a sucient mass and force resolution and we properly correct the nite box eect in the measurement of mass functions (see Section 3.2), the identication of sub- structures in single halos can cause a systematic eect in our measurement. There is no unique way to in N-body simulations (e.g. Onions et al. 2012), and over-merging may take place even in the lat- estN-body simulations (e.g. van den Bosch & Diemer (2021) found that the denition of boundary radii of host halos can change the by a factor of 2, implying that our halo catalogs contain a non-negligible mislabeling of subha- los. The mislabeled subhalos are mostly located at out- skirts of host halos (Diemer 2021). Note that we use the virial radius to identify subhalos in the simulation, but the virial radius does not correspond to a gravita- tional boundary radius in general (e.g. More et al. 2015; Diemer 2017). Because more subhalos will be found as halo-centric radii increase, we expect that the dierence of the multiplicity function with and without subhalos can provide a reasonable estimate of systematic errors in our measurement. In our simulation, we nd that the multiplicity function with subhalos is dierent from one without subhalos in a systematic way. The dier- ence is less sensitive to the redshift and it can be (  <1) 0:025 (3) (13) 0 (  the multiplicity function with subhalos, fd is the counterpart without subhalos (our ducial Using Eq. (16), we set 10. Note that our estimate of sys should be overestimated because we assume that all sub- halos are subject to mislabeling. Hence, our analysis is surely mass function in Planck cosmology 0) Figure 4. Non-universality of virial halo mass functions in the2GC simulations. The top panel shows the a function of =c;z=at dierent redshifts. In the top, blue circles, orange diamonds, and green the simulation results at z= 0:00, 1.01, and The blue dashed line in the top panel is the best-t model of fatz= 0:00. In the bottom, we show the dierence of log fbetween the simulation results at the three redshifts and the best-t model at z= 0:00. If the multi- plicity function is universal across redshifts, all the symbols should locate at y= 0 in the bottom panel. This gure claries that the virial halo mass function exhibits a strong redshift A simple check of showing the results of our calibration, we per- form a sanity check to see the redshift dependence of the multiplicity function in the simulation. Figure 4 shows the result of our sanity test. In this gure, we rst nd a best-t model for the multiplicity function atz= 0, denoted as fBest(;z= 0). We then com- pare the multiplicity function in the simulation at 0). If the function f(;z) is universal across dierent redshifts, we should nd a residual be- tween the simulation results at z>0 andfBest(;z= 0) as small as the case for z= 0. In the top panel in Fig-ure 4, the blue dashed line represents the best-t model fBest(;z= 0), while blue circles show the atz= 0. Comparing between the blue dashed line and blue circles, our tting at z= 0 provides a representative model of the simulation result. The or- ange squares and green diamonds in the gure show the simulation results at z= 1:01 andz= 4:04, respec- tively. The bottom panel shows the residual between the simulation results and the model of fBest(;z= 0), highlighting prominent redshift evolution of the multi- plicity function at c;z=>3 fromz= 0 to 4. In this gure, the error bars show the statistical we account for the sample variance caused by nite box eects in our simulation (see Section 3.3). 5.2. Fitting results We here show the main result in this paper. Figure 10 in Appendix B summarizes the residual between the multiplicity functions in our simulations and the best- t model at dierent redshifts. We nd that our tting works well across a wide range of redshifts (0 z7). A typical dierence between the simulation results and our best-t model is an of order of 0.02 dex except for bins at c;z=(z)>23. Although the bins at c;z=(z)>2 suer from statistical uctuations induced by the sample variance, we nd that the residual is still within 0.06 dex even for such rare objects. for our best-t models ranges from 0.1 to 1.1. It would be worth noting that we include possi- ble systematic errors due to modeling of subhalos in our tting. Such systematic errors can reduce the score of 2for the best-t model, making the best-t 2smaller than the number of degrees of freedom. The redshift evolution in best-t parameters is shown in Figure 5. The gray shaded region in each of a given parameter, inferred from the Jacobian of Eq. (14). In each panel, the circles at z4:57 represent averages of the best-t parameter within coarse bins of z, while the ones at z= 5:98 and 7.00 show the best-t parameters. To compute the av- erage, we set seven coarse redshift bins. The edge of coarse bins is set to [0 :00;0:30), [0:30;0:61), [1:49;2:44), [2:44;3:36), and [3:36;4:57]. We choose this binning of redshifts so that the dynamical time (i.e. the ratio of the virial radius and virial circu- lar velocity for halos) can be comparable to the Hubble time at bin-centered redshifts. The blue solid lines in Figure 5 present our We nd that the following form can explain the redshift evolution of the best-t parameters and return a similar level of the residual as in Figure 10 when used8 Shirasaki et study n2GC Tormen et al. et al. (2016) Figure 5. Redshift evolution of parameters for virial halo mass functions in the 2GC simulations. In each panel, the shaded region shows 1error bars inferred by our tting, while the circles are best-t parameters at representative coarse redshift bins (see the main text). The blue line represents our calibrated model. For comparison, the red dashed line is the in Bhattacharya et al. (2011), while the orange solid and purple dotted lines are the models in Sheth & Tormen (1999) and Despali et al. (2016), respectively. in Eq. (4): A(z) = 0:3250:017z; (17) a(z) = 0:940 (1 (18) p(z) = 0:692; (19) q(z) = 1:611 (1 +z)0:12; (20) wherec;zis given by Eq. (6). 5.3. Comparison with previous studiesOur model (Eqs. 17-20) can be compared with previ- ous models in the literature. The most popular model in Sheth & Tormen (1999) predicts a universal multi- plicity function f() withA= 0:322,a= 0:707,p= 0:3 andq= 1:0. It would be worth noting that the pa- rameter of Ain Sheth & Tormen (1999) is xed by lnf() = 1; (21) where it means that all dark matter particles reside in halos. Because our model has been calibrated by the data with a nite range of , it is not necessary to satisfy the condition of Eq. (21). The integral ofHalo mass function in Planck cosmology 9 Eq. (21) for our model can be well approximated as 1:37 (1 within a 0.5%-level accuracy. In reality, the upper limit in the integral of Eq. (21) may be set by the free streaming scale of dark matter particles. If dark matter consists of Weakly Interacting Massive Particle (WIMP) with a particle mass being 100 GeV, the minimum halo mass is estimated to be Hofmann et al. 2001; Berezinsky et al. 2003; Green et al. 2004; Loeb & Zaldarriaga 2005; Bertschinger 2006; Profumo et al. 2006; Diamanti et al. 2015). When we set the up- per limit to be upp(z) =(M= 106h1M;z), the fraction of mass in eld halos can be approximated the approximation is valid at z7 within a 0.4%- level accuracy. Eq. (22) is less sensitive to the choice of the minimum halo mass as long as we vary the minimum halo mass in the range of 1012103h1M. Our model predicts that about 72% of the mass density in the present-day universe resides in dark matter halos. For the parameter a, our model (Eq. 18) shows a modest redshift evolution with a level of 20% from z= 7 toz= 0. Note that an eective critical less dependent on zand evolves only by 23% in the range of 0 z7. The of ais mostly consistent with the model in Bhattacharya et al. (2011), but the overall amplitude diers by20%. Note that the model in Bhattacharya et al. (2011) has been calibrated for the halo mass func- tion when the mass is dened by the algorithm. Because the FoF mass is expected to strongly depend on inner density proles and substruc- tures (More et al. 2011), our model needs not match the one in Bhattacharya et al. (2011). The ofa,p, andqare in good agreement with the result in Comparat et al. (2017), which presented the calibration of the halo mass function at z= 0 with the MultiDark simulation (Prada et al. 2012; Klypin et al. 2016). Note that Comparat et al. (2017) adopted same halo nder, mass denition and cosmology as ours. Despali et al. (2016) argued that the multiplicity func- tion can be expressed as a universal function once one includes the redshift dependence of the spherical Although we adopt the density as well, we nd that the parameters in the multiplicity function depend on redshifts (also see Figure 4). The main dierence between the analysis in Despali et al. (2016) and ours is halo identications inN-body simulations. We dene halos by the FoF al- gorithm in six-dimensional phase-space, while Despali et al. (2016) adopted a algorithm with a smoothing scale being the distance to the tenth nearest neighbor for a given N-body particle. Note that halo nders based on particle positions may not dis- tinguish two merging halos. The ROCKSTAR the velocity information of N-body particles, al- lowing to very eciently determine particle-halo mem- berships even in major mergers. We next compare our model of the halo mass functions with previous models in the literature. For compari- son, we consider six representative models summarized in Table 2. Three of them assume a universal functional form of the multiplicity function, while others include some redshift evolution. Among the previous studies, Despali et al. (2016) have investigated the mass func- tion when using the virial halo mass, and their results can be directly compared to ours. Tinker et al. (2008) and Watson et al. (2013) studied mass functions for var- ious spherical overdensity masses. We here use their t- ting formula which takes into account the dependence of spherical overdensity parameters, while the calibration for the virial halo mass has not been done in Tinker et al. (2008) and Watson et al. (2013). Bhattacharya et al. (2011) have calibrated the halo mass function when the mass is dened by the FoF algorithm, indicating that a direct comparison with our results may not be appro- priate (e.g. see More et al. 2011, for details). The left panel in Figure 6 shows the comparison of t- ting formulas for the halo mass function at z= 0, while the right presents the case at z= 6. Our model is shown by blue solid lines in both panels and the gray shaded region in the gure represents the range of halo masses not explored by our simulations. At z= 0, our model is in good agreement with most of previous models in the range of 10913h1M, but there are at the high mass end ( gure implies that previous tting formulas are not sucient to predict of the mass function for cluster-sized halos at z= 0 even within a concordance CDM cosmology. A by Tinker et al. (2008) may have a systematic error to predict the cluster abundance with a level of 2030%, but the exact amount of systematic errors should depend on the denition of spherical overden- sity halo masses. We need larger N-body see Ishiyama et al. 2020, for a relevant example) to precisely calibrate the mass function at high mass ends and make a robust conclusion about systematic errors in cluster cosmology (e.g. McClintock et al. 2019; Boc-10 Shirasaki et al. Table 2. Models of the multiplicity function f(;z) selected in this paper. Note that the functional form of fis provided in Eq. (4) except for the models in Tinker et al. (2008, T08) and Watson et al. (2013, W13). The T08/W13 model assumes f(;z) =A0[(=b0)a0+ 1] exp( c0=2) parameters. In this table, the second column shows which cosmological model is used to calibrate the functional form of f. WMAP1, WMAP3, WMAP5 and WMAP7 represent the rst-, third-, fth- and seventh-year WMAP constraints, respectively (Spergel et al. 2003, 2007; Komatsu et al. 2009, 2011). Planck14 means the best-t cosmological model derived in Planck Collaboration et al. (2014), while Planck16 refers to the model in Planck Collaboration et al. (2016). The fourth column summarizes the mass denition of dark matter halos in simulations, while the fth column shows if the parameters in fare dependent on redshifts or not. Name Cosmology Calibrated ranges of halo masses and redshifts Halo mass paper Planck16 0 z7ROCKSTAR Yes Press & Schechter (1974) { { { No Sheth & Tormen (1999) { { { No Tinker et al. (2008) WMAP1, WMAP3 0 z2 SO Yes Bhattacharya et al. (2011) WMAP5 0 z2 FoF Yes Watson et al. (2013) WMAP5 SO 8<z26 Despali et al. (2016) WMAP7, Planck14 0 z5 SO 0 This study Sheth & Tormen (1999) Tinker et al. et al. (2011) Watson et al. (2013) Despali et al. (2016) Press & Schechter Diff Figure 6. Comparison of tting formulas for halo mass functions at z= 0 and 6. The left and right panels show the comparison atz= 0 and 6, respectively. Note that there are no available data for the gray shaded regime in our simulations. In each panel, the blue thick solid line presents our result. Other predictions in Press & Schechter (1974), Sheth & Tormen (1999), Tinker et al. (2008), Bhattacharya et al. (2011), Watson et al. (2013) and Despali et al. (2016) are also shown by pink thin solid, green thick dashed, red thin dashed, purple thick dotted, brown thin dotted lines, mass function in Planck cosmology 11 1 2 3 4 5 WDM mass mWDM 1s 2s 3sUV cumulative luminosity function phobsatz= 6 This study Sheth & Tormen (1999) Press & Schechter (1974) Figure 7. Lower limits of warm dark matter (WDM) masses obtained from UV luminosity function of galaxies atz= 6. The dashed horizontal lines represent the lower bounds of the UV luminosity function at z= 6 in the Hub- ble Frontier Fields as estimated in Menci et al. (2016) (1 , 2 and 3levels from top to bottom). The solid line shows the maximum number density of dark matter halos as a function of WDM masses based on our calibrated halo mass functions and the correction proposed in Lovell (2020). Assuming that an one-to-one correspondence of dark matter halos and faint galaxies, one can exclude the WDM model if the maximum number density of dark matter halos becomes smaller than the observed galaxy abundance. Our calibrated mass func- tion places the lower limit of mWDM>2:71 keV at the 2  condence level. This constraint is degraded with a level of 28% and 16% when one uses the commonly-adopted mass functions by Press & Schechter (1974) and Sheth & Tormen (1999), et al. 2020; Klypin et al. 2021). We leave of cluster mass functions for future studies. Atz= 6, our model shows an oset from the uni- versal models by Sheth & Tormen (1999) and Despali et al. (2016). The dierence reaches a 15-20% level atM= 10910h1Mand becomes larger at higher masses. This is mainly caused by the redshift evolu- tion of the amplitude in the multiplicity function, A(z). Our model predicts that the amplitude in the multiplic- ity function decreases as going higher redshifts, and this trend is clearly found in the simulation results (see Fig- ure 4). 5.4. Implications An important implication of our calibration is that modeling of the halo mass function for CDM cos- mologies can aect constraints of nature of dark mat- ter particles by high-redshift galaxy number counts (e.g. Pacucci et al. 2013; Schultz et al. 2014; Menci et al. 2016;Corasaniti et al. 2017). Warm dark matter (WDM) is an alternative candidate of cosmic dark matter with free streaming due to its thermal motion. Some extensions of the Standard model predict the existence of WDM with masses in keV range, such as sterile neutrino (e.g. Adhikari et al. 2017; Boyarsky et al. 2019). Structure formation with WDM particles can be suppressed at scales below free-streaming lengths fs, while the bottom-up formation of dark matter halos re- mains as in the standard CDM paradigm at scales larger thanfs. A characteristic mass scale for fshas been es- timated as1010Mfor WDM with a particle mass of O(1) keV (e.g. Bode et al. 2001). In a hierarchical structure formation, less massive galaxies form at higher redshifts. At a xed cosmic mean density, WDM particles with larger masses are less ef- fective at suppressing the growth of low-mass halos (e.g. Schneider et al. 2012). Assuming that high- zgalaxies only form in collapsed halos, the observed abundance of high-zgalaxies can thus provide a lower limit to the particle mass of WDM. Recent high-resolution N-body simulations for WDM have indicated that the correction of the halo abundance due to free streaming can be expressed as a universal form = 1 +Mhm M# ; (24) where dn=dMjWDM is the halo mass function for WDM cosmologies, d n=dMjCDM is the counterpart of CDM, Lovell (2020) found that = 2:3,= 0:8 and = 1:0 provide a reasonable t to simulation results. In Eq. (24),Mhmis so-called half-mode mass, which is dened as the mass scale that corresponds to the power spectrum wave number at which the square root of the ratio of the WDM and CDM power spectra is 0.5 (Schneider et al. 2012). The mass Mhmdepends on the particle mass of WDM mWDM = 0:153 (26) where= 1:12 and WDM is the dimensionless of WDM. According to Eq. (23), an of mass function for CDM cosmologies is12 Shirasaki et al. essential to predict the counterpart of WDM. We here caution that Eqs. (23) and (24) have been validated at z= 0 andz= 2 in Lovell (2020). We assume that these equations are valid at z6. We leave a validation of Eqs. (23) and (24) at z>2 for future studies. Figure 7 demonstrates the importance of the calibra- tion of CDM halo mass function when one constrains WDM masses with high-redshift galaxy number counts. For given limiting magnitude and redshift, the cumu- lative galaxy number density should be smaller than the whole halo mass function within WDM the luminosity corresponding to the lim- iting magnitude, d ngal=dLis the galaxy and we set Mmin= Note that our lowest mass Mmin = 1h1Mis much smaller than the half-mode keV. For the UV lu- minosity function at z= 6 in the Hubble Frontier Fields with the limiting AB magnitude of 12:5 (Livermore et al. 2017), the lower bound of obshas been estimated as at a 2condence level (Menci et al. 2016). Assuming the best-t cosmologi- cal parameters in Planck Collaboration et al. (2016) and WDM is made of the whole abundance of dark matter, our model of the halo mass function with Eq. (24) allows to reject WDM with their mass smaller than 2.71 keV at the 2level. This lower limit is degraded to be 2.27 keV and 1.96 keV for the commonly-adopted models by Sheth & Tormen (1999) and Press & Schechter This simple example highlights that cali- bration of the mass function for CDM cosmologies is essential to accurate predictions of the mass function for WDM cosmologies. Our lower limit of the WDM particle mass can be compared to other methods. For example, Baur et al. (2016) found a lower limit of 2.96 keV from observations of the Lyman-alpha forest, while al. (2020) placed a lower limit of 5.3 keV (a similar limit is found by Ir si c et al. (2017)). Chatterjee et al. (2019) used high-redshift 21-cm data from EDGES to rule out WDM with mWDM<3 keV. Note that our limit is less sensitive to details of baryonic physics than the others, because our analysis relies on the of dark matter halos. For a conservative analysis, we consider that a sig- nicant small halos of M= 1h1Mcan be respon- sible to the observed galaxy abundance at high red- shifts. To further tighten the limit of WDM particlemass, it would be interesting to discuss more realistic halo mass scales to the faintest galaxy at z6. In the Planck CDM cosmology, we nd that the minimum halo mass the cumulative halo abundance of 1 :35 Mpc3, which is close to the observed galaxy abudnance at z= 6. When setting the minimum halo mass to 107h1Min Eq. (27), we nd a ofmWDM>14:1 keV. However, this stringent limit is very sensitive to the choice of the minimum halo mass. For the minimum halo mass of 106h1M, the limit changes to mWDM>6:23 keV. This simple analy- sis implies that a more detailed modeling of at would be worth pursuing in future summarize the major limitations in our model of virial halo mass functions. All of the following issues will be addressed in forthcoming studies. 6.1. Cosmological dependence Our model of halo mass functions is calibrated against N-body simulations in the CDM cosmology consis- tent with Planck16. In terms of studies of m0and8are the primary parameters and the simulations in this paper adopt m0= 0:31 and 8= 0:83. Therefore, our calibration of Eqs. (17)-(20) may be subject to an overtting to the specic cos- mological model. To examine the dependence of our model on cosmological models, we use another halo cat- alog fromN-body simulations with a dierent CDM model. For this purpose, we use the Bolshoi simulation in Klypin et al. (2011) and the rst performed in Prada et al. (2012). The Bol- shoi simulation consists of 20483particles in a volume assumes the cosmological parame- ters of m0= 0:27, b0= 0:0469, = 1 m0= 0:73, h= 0:70,ns= 0:95, and8= 0:82. These are consistent with the ve-year observation of the cosmic obtained by the WMAP satellite (Komatsu et al. 2009) and we refer to them as the WMAP5 cos- mology. The MultiDark-Run1 simulation adopted the same cosmological model and the number of particles as in the Bolshoi simulation, while the volume is set to 1 [h1Gpc]3. We use the ROCKSTAR halo catalog at z= and 4:07 from the Bolshoi simulations.8To compute our model 8The halo catalogs at dierent redshifts are pub- licly available at Hlists Rockstar/ and Catalogs/.Halo mass function in Planck cosmology 8. Comparison of the multiplicity function fmea- sured in the MultiDark-Run1 and Bolshoi (MB) et al. 2011; Klypin et al. 2011; Prada et al. 2012) and predictions by our tting formula. Dierent symbols in each panel show the simulation results at various redshifts. The dashed lines in the top panel show our predictions. Note that we introduce an arbitrary shift in f=2for visibility in the top panel. In the bottom, we show the dierence of logfbetween the simulation results and our that the MB simulations adopt the cosmological model with m0= 0:27, while our tting formula of frelies on the simulation with m0= 0:31. Other CDM parameters are almost same between the two. The bottom panel shows that our tting formula of fcan have a systematic error with a level of 0.05-0.2 dex if one changes m0by 13%. prediction for the WMAP5 cosmology, we x the func- tional form of Eq. (4) and parameters in Eqs. (17)-(20) but include the of the the top-hat mass variance , 8 summarizes the multiplicity function in the Bolshoi and MultiDark-Run1 simulations. In this g- ure, dashed lines show the predictions by our model for the WMAP5 cosmology, while dierent symbols repre- sent the simulation results. We nd that our model can reproduce simulation results within a 10% level even for the WMAP5 cosmology at 0 :7<c;z=<2:5. At high mass ends ( c;z=>2:5), our model tends to underes- timate the halo abundance by 3040% in a wide range of redshifts. It is worth noting that the residualbetween our model and the WMAP5-based simulation is less dependent on redshifts. In fact, we found a better matching between our models and the when reducing the overall amplitude in by 45%. In summary, our model can not predict the simulation results for the WMAP5 cos- mology with the same level as in the Planck cosmology. The 10%-level dierence in m0can cause in our model predictions with a level of 10% except for high mass ends. Future studies would need to calibrate the cosmological dependence of a(z) for precision cosmology based on galaxy clusters. At low masses and high redshifts, our model can provide a rea- sonable t to the simulations adopting the WMAP5 cos- mology. According to this fact, we examine how much the WDM limit in Section 5.4 is aected by the choice of underlying cosmology. For the WMAP5 cosmology, we nd a 2 limit ofmWDM>2:75 keV, which diers from our ducial limit by only 1:4%. 6.2. Baryonic eects Our calibration of halo mass functions relies on N-body simulations and ignores eects. Baryonic eects on halo mass func- tions have been studied with a set of (e.g. Stanek et al. 2009; Cui et al. 2012; Sawala et al. 2013; Cui et al. 2014; Cusworth et al. 2014; Martizzi et al. 2014; Bocquet et al. 2016; Beltz- Mohrmann & Berlind 2021). The evolution of cosmic baryons is governed not only by gravity but also associated with galaxy formation. include gas cooling, star formation and en- ergy feedback from supernovae (SN) and Active Galac- tic Nuclei (AGN). Adiabatic gas heated only by gravi- tational processes aects the halo mass function with a level of<23% (Cui et al. 2012), while radiative cool- ing, star formation and SN feedback increase individual halo masses due to condensation of baryonic mass at the halo center, changing the halo mass function (Stanek et al. 2009; Cui et al. 2012). Ecient SN feedback can decrease the halo mass function at (Sawala et al. 2013). The mass be aected by AGN feed- back, while current hydrodynamical simulations adopt a sub-grid model to include the AGN feedback. Because a variety of sub-grid models has been proposed, the im- pact of the AGN feedback on the mass function is still uncertain (Cui et al. 2014; Cusworth et al. 2014; Mar- tizzi et al. 2014; Bocquet et al. 2016). To account for the baryonic eects on the halo mass function, it is important to correct individual halo masses according to baryonic processes. Baryons do not14 Shirasaki et al. change the abundance of dark matter halos, but aect internal structures and spherical masses of halos. Hence, one may be able to model the halo mass function in the presence of baryons by abundance matching between the gravity-only and hydrodyamical simulations for a given denition of the halo mass (e.g Beltz-Mohrmann & Berlind 2021). In this sense, our calibration of the halo mass function provides a baseline model and still plays an important role in understanding the baryonic eects on large-scale AND CONCLUSION In this paper, we have studied mass functions in the concordance  cold dark matter (CDM) model inferred from the measurement of cosmic s by the Planck satellite (Planck Collabora- tion et al. 2016). We have calibrated the abundance of dark matter halos in a set of N-body et al. 2015; Ishiyama & Ando 2020) cover- ing a wide range of redshifts and halo masses. For a virial spherical we have employed least-square analyses to nd best-t models to our simulation results in the range redshifts zrange from 0 to 7. Our calibrated models are able to reproduce the simu- lation results with a 5%-level precision over all in this paper, but except for high-mass ends. We found that the multiplicity function fdened in Eq. (1) exhibits some redshift dependence, contradict- ing the commonly-adopted analytic model as in Sheth & Tormen (1999, ST99). The redshift evolution of the multiplicity function is prominent in our simulation data even at high redshifts z>3. Our calibrated halo mass function is in good agreement with previous models in the literature within a level of<15% atz= 0, while our model predicts that the halo mass function in the range ofM= 6 can be smaller than the ST99 prediction by 20 30%. If cosmic dark matter consists of Weakly Interact- ing Massive Particle (WIMP) with a particle mass of 100 GeV, the minimum halo mass would be of an or- der of Hofmann et al. 2001; Berezinsky et al. 2003; Green et al. 2004; Loeb & Zal- darriaga 2005; Bertschinger 2006; Profumo et al. 2006). An extrapolation of our calibrated halo mass function to such minimum halo masses allows us to predict the fraction of cosmic mass density containing in halos. We found that the fraction can be well approximated as 0:724 0z7. This implies that about 72% of the mass density at present is conned in objects. If cosmicdark matter consists of warm dark matter (WDM) with a particle mass of 1 keV such as sterile neutrino (e.g. Adhikari et al. 2017; Boyarsky et al. 2019), our model with a WDM correction (Lovell 2020) provides a powerful test of WDM scenarios by compar- ing with galaxy number counts at high redshifts. We found that WDM with a particle mass smaller than 2.71 keV is incompatible to the UV luminosity function at z= 6 in the Hubble Frontier Fields (Livermore et al. 2017) at a 2 condence level. It would be worth not- ing that this upper limit can be degraded by 16% when one adopts the ST99 prediction. This highlights that the calibration of halo mass functions in CDM cos- mologies is required to place cosmological constraints of WDM especially when using high-redshift observables. Our tting formula of the virial halo mass function is based on dark-matter-only N-body simulations for a specic cosmology. We found a 10%-level dierence in the average cosmic mass density can cause in our model predictions with a<10% level. Baryonic eects such as gas cooling, star forma- tion, and some feedback processes can aect of dark matter halos. Although it is still dif- cult to account for the baryonic eects in our model, re- cent simulations indicate that abundance matching be- tween the gravity-only and hydrodynamical be promising (e.g. Beltz-Mohrmann & Berlind 2021). This implies that our calibrated model is still meaningful as a baseline prediction before eects, while more detailed analysis with hy- drodynamical simulations is demanded. Our can be applied to any N-body simulations based on non-CDM, e.g. allowing to validate a universal sup- pression of the halo abundance (see Eq. 24) in work is in part supported by MEXT KAK- ENHI Grant Number (19K14767, 19KK0344, 21H01122). Numerical computations were in part carried out on Cray XC50 at Center for Compu- tational Astrophysics, National Astronomical Observa- tory of Japan. TI has been supported by MEXT as Program for Promoting Researches on the Supercom- puter Fugaku (Toward a unied view of the Universe: from large scale structures to planets, proposal numbers hp200124 and hp210164), and JICFuS. The 2GC and Phi-1 simulations were run on the K computer at the RIKEN Advanced Institute for Computational Science and the Aterui supercomputer at Center for Computa- tional Astrophysics, of National Astronomical Observa- tory of Japan.Halo mass function in Planck cosmology 15 Software: GreeM (Ishiyama et al. 2009, 2012), 2LP- Tic (Crocce et al. 2006), MUSIC (Hahn & Abel 2011),CAMB (Lewis et al. 2000), ROCKSTAR (Behroozi et al. 2013), Scipy (Jones et al. 9. Comparison of the multiplicity function fwith and without unbound particles. Dierent symbols in the top panel show the simulation results including unbound parti- cles at various redshifts, while the dashed lines in the top panel show the best-t models for our ducial halo unbound particles). In the bottom, we show the dierence of log fbetween the two and the error bars show statistical OF UNBOUND PARTICLES IN HALO MASS DEFINITION We here examine possible eects of unbound parti- cles around dark matter halos on our calibration of the halo mass function. For this purpose, we use with 40963particles and the box length on a side being 560 h1Mpc, referred to as the 2GC-M run in Ishiyama et al. (2015). We prepare two dierent halo catalogs at redshifts of z= and 4:04. One is the catalog with the default option for the ROCKSTAR nder and does not include unbound particles in the virial mass for individual halos, while another im- poses the option of STRICT SOMASSES=1 to account for unbound particles. Figure 9 compares the measured in the two halo catalogs. We nd that our tting of the multiplicity function with the de- fault halo catalogs (our ducial runs) is not aected by the inclusion of unbound particles beyond the for the 2GC-M runs. B.SUMMARY OF OUR FITTING RESULTS In this Appendix, we provide a summary of our t- ting results as in Figure 10. Note that Figure 10 shows the residual between simulation results and the best- t model. It is non-trivial to reproduce the s with a similar level to Figure 10 when we inter- polate the model parameters as in Eq. (17)-(20). The comparison with the simulation results and the model with Eq. (17)-(20) are summarized in Figure 11. The g- ure represents the performance of our calibrated model and we nd a 5%-level precision in our model for a wide range of masses and R., Agostini, M., Ky, N. A., et al. 2017, JCAP, 2017, 025, doi: S. W., Evrard, A. E., & Mantz, A. B. 2011, ARA&A, 49, 409, doi: J., N., Y eche, C., Magneville, C., & Viel, M. 2016, JCAP, 2016, 012, doi: P. S., Wechsler, R. H., & Wu, H.-Y. 2013, ApJ, 762, 109, doi: G. D., & Berlind, A. A. 2021, arXiv e-prints, A., & Frenk, C. 2020, MNRAS, 498, 4887, doi: tez-Llambay, A., Navarro, J. F., Frenk, C. S., et al. 2017, MNRAS, 465, 3913, doi: V., Dokuchaev, V., & Eroshenko, Y. 2003, PhRvD, 68, 103003, doi: E. 2006, PhRvD, 74, 063509, doi: Shirasaki et 0.2 0.4 0.6 0.8 z=2.12 0.2 0.4 0.6 0.8 z=2.28 0.2 0.4 0.6 0.8 z=2.44 0.2 0.4 0.6 0.8 z=2.60 0.2 0.4 0.6 0.2 0.4 0.6 0.8 z=2.95 0.2 0.4 0.6 0.8 z=3.14 0.2 0.4 0.6 0.8 z=3.36 0.2 0.4 0.6 0.8 z=3.80 0.2 0.4 0.6 0.2 0.4 0.6 0.2 0.4 0.6 0.2 0.4 0.6 0.2 0.4 0.6 10. A summary of our ts to virial halo mass functions in the 2GC simulations. Each panel shows the residual of the multiplicity function fin logarithmic space at dierent redshifts (0 z7). The colored symbols in each panel represent the residual between the simulation result and its best-tted model and the black line shows no dierence between the two. The blue circles, orange diamonds, and green squares are the results in the L, H2, and phi1 runs, mass function in Planck cosmology 0.2 0.4 0.6 0.8 z=2.12 0.2 0.4 0.6 0.8 z=2.28 0.2 0.4 0.6 0.8 z=2.44 0.2 0.4 0.6 0.8 z=2.60 0.2 0.4 0.6 0.2 0.4 0.6 0.8 z=2.95 0.2 0.4 0.6 0.8 z=3.14 0.2 0.4 0.6 0.8 z=3.36 0.2 0.4 0.6 0.8 z=3.80 0.2 0.4 0.6 0.2 0.4 0.6 0.2 0.4 0.6 0.2 0.4 0.6 0.2 0.4 0.6 11. Similar to Figure 10, but we show the residual between the simulation results and the model with Eq. S., Heitmann, K., White, M., et al. 2011, ApJ, 732, 122, doi: S., Heitmann, K., Habib, S., et al. 2020, ApJ, 901, 5, doi: Shirasaki et al. Bocquet, S., Saro, A., Dolag, K., & Mohr, J. J. 2016, MNRAS, 456, 2361, doi: P., Ostriker, J. P., & Turok, N. 2001, ApJ, 556, 93, doi: J. R., Cole, S., Efstathiou, G., & Kaiser, N. 1991, ApJ, 379, 440, doi: A., Drewes, M., Lasserre, T., Mertens, S., & Ruchayskiy, O. 2019, Progress in Particle and Nuclear Physics, 104, 1, doi: G. L., & Norman, M. L. 1998, ApJ, 495, 80, doi: A., Dayal, P., Choudhury, T. R., & Hutter, A. 2019, MNRAS, 487, 3560, doi: J., Prada, F., Yepes, G., & Klypin, A. 2017, MNRAS, 469, 4157, doi: P. S., Agarwal, S., Marsh, D. J. E., & Das, S. 2017, PhRvD, 95, 083512, doi: M., Pueblas, S., & Scoccimarro, R. 2006, MNRAS, 373, 369, doi: W., Borgani, S., Dolag, K., Murante, G., & Tornatore, L. 2012, MNRAS, 423, 2279, doi: W., Borgani, S., & Murante, G. 2014, MNRAS, 441, 1769, doi: S. J., Kay, S. T., Battye, R. A., & Thomas, P. A. 2014, MNRAS, 439, 2485, doi: M., Efstathiou, G., Frenk, C. S., & White, S. D. M. 1985, ApJ, 292, 371, doi: G., Giocoli, C., Angulo, R. E., et al. 2016, MNRAS, 456, 2486, doi: R., Catalan, M. E. C., & Ando, S. 2015, PhRvD, 92, 065029, doi: B. 2017, ApJS, 231, 5, doi: 2020, ApJ, 903, 87, doi: 2021, ApJ, 909, 112, doi: A. M., Hofmann, S., & Schwarz, D. J. 2004, MNRAS, 353, L23, doi: O., & Abel, T. 2011, MNRAS, 415, 2101, doi: K., Ricker, P. M., Warren, M. S., & Habib, S. 2005, ApJS, 160, 28, doi: S., Schwarz, D. J., & St ocker, H. 2001, PhRvD, 64, 083507, doi: W., & Kravtsov, A. V. 2003, ApJ, 584, 702, doi: si c, V., Viel, M., Haehnelt, M. G., et al. 2017, PhRvD, 96, 023522, doi: T., & Ando, S. 2020, MNRAS, 492, 3662, doi: T., Enoki, M., Kobayashi, M. A. R., et al. 2015, PASJ, 67, 61, doi: T., Fukushige, T., & Makino, J. 2009, PASJ, 61, 1319, doi: T., Nitadori, K., & Makino, J. 2012, arXiv e-prints, T., Prada, F., Klypin, A. A., et al. 2020, arXiv e-prints, A., Frenk, C. S., White, S. D. M., et al. 2001, MNRAS, 321, 372, doi: E., Oliphant, T., Peterson, P., et al. 2001{, SciPy: Open source scientic tools for T., & Suto, Y. 1996, ApJ, 469, 480, doi: A., Yepes, G., Gottl ober, S., Prada, F., & He, S. 2016, MNRAS, 457, 4340, doi: A., Poulin, V., Prada, F., et al. 2021, MNRAS, 504, 769, doi: A. A., Trujillo-Gomez, S., & Primack, J. 2011, ApJ, 740, 102, doi: A., Knollmann, S. R., Muldrew, S. I., et al. 2011, MNRAS, 415, 2293, doi: A., Pearce, F. R., Lux, H., et al. 2013, MNRAS, 435, 1618, doi: E., Dunkley, J., Nolta, M. R., et al. 2009, ApJS, 180, 330, doi: E., Smith, K. M., Dunkley, J., et al. 2011, ApJS, 192, 18, doi: A., Challinor, A., & Lasenby, A. 2000, ApJ, 538, 473, doi: R. C., Finkelstein, S. L., & Lotz, J. M. 2017, ApJ, 835, 113, doi: A., & Zaldarriaga, M. 2005, PhRvD, 71, 103520, doi: M. R. 2020, ApJ, 897, 147, doi: A. D., Schaye, J., & Bower, R. 2019, MNRAS, 488, 3663, doi: c, Z., Heitmann, K., Habib, S., Bashinsky, S., & Ricker, P. M. 2007, ApJ, 671, 1160, doi: D., Mohammed, I., Teyssier, R., & Moore, B. 2014, MNRAS, 440, 2290, doi: T., Rozo, E., Becker, M. R., et al. 2019, ApJ, 872, 53, doi:

2108.11038

Masato Shirasaki

Masato Shirasaki, Tomoaki Ishiyama, Shin'ichiro Ando

Virial halo mass function in the ${it Planck}$ cosmology

19 pages, 11 figures, 2 tables. Accepted for publication in ApJ

10.3847/1538-4357/ac214b

astro-ph.CO astro-ph.GA hep-ph

http://creativecommons.org/licenses/by/4.0/

We study halo mass functions with high-resolution $N$-body simulations under a $Lambda$CDM cosmology. Our simulations adopt the cosmological model that is consistent with recent measurements of the cosmic microwave backgrounds with the ${it Planck}$ satellite. We calibrate the halo mass functions for $10^{8.5} uildrel < over sim ;$} M_mathrm{vir} uildrel < over sim ;$} 10^{15.0 - 0.45 , z}$, where $M_mathrm{vir}$ is the virial spherical overdensity mass and redshift $z$ ranges from $0$ to $7$. The halo mass function in our simulations can be fitted by a four-parameter model over a wide range of halo masses and redshifts, while we require some redshift evolution of the fitting parameters. Our new fitting formula of the mass function has a 5%-level precision except for the highest masses at $zle 7$. Our model predicts that the in Sheth $&$ Tormen would overestimate the halo abundance at $z=6$ with $M_mathrm{vir} = 10^{8.5-10}, h^{-1}M_odot$ by $20-30%$. Our calibrated halo mass function provides a baseline model to constrain warm dark matter (WDM) by high-$z$ galaxy number counts. We compare a function of galaxies at $z=6$ with the total halo abundance based on our model and a recently proposed WDM correction. We find that WDM with its mass lighter than $2.71, mathrm{keV}$ is incompatible with the observed galaxy number density at a $2sigma$ confidence level.

2 I. the inception of quantum theory, it is believed that physical observables are Hermitian and having real eigen- values. After the seminal work of Bender [1{3], the researchers understood that there are few non-hermitian systems also possessing real eigenvalues and generate nonunitary time evolution. Due to rapid development in both theoretical and experimental facts, the class of general non-Hermitian systems acts as an eective tool in the domain of theory to describe the dynamics of the quantum system which interacts with the environment [4, 5]. For instance, non-Hermitian systems have been used to explain the well-known dissipation and absorption phenomena like radioactive decay [6] and interaction between neutrons and nuclei [7]. Characterizing the nonlocal attributes of a quantum system is a formidable task and a key to implement the various quantum information processing [8]. To manifest the nonlocality, various measures with dierent notions have been introduced. A few popular measures are coherence [9], entanglement [10, 11], Bell inequality [12], steering [13], etc,. In general, nonlocality of a state witnessed or demonstrated through the violation of Bell inequality. For any pure state, the entanglement and Bell inequality are the complete manifestation of nonlocality. On the other hand, the nonlocality of the mixed state is not Indeed, the entanglement and Bell inequality are the incomplete manifestation of nonlocality. Due to this incompleteness, we are in need of identifying new quantum correlation measures, beyond entanglement. To characterize the complete nonlocal attributes of the quantum system, in the past two decades, various measures have been identied and explored in detail [14{19]. In geometric perspectives, based on the locally invariant projective measurement, a new measure of quantum correlation has been introduced which is known as nonlocality (MIN) [15]. Originally, MIN is dened via maximal Hilbert-Schmidt norm, and considered as a more general manifestation of nonlocality than the Bell inequality and entanglement. The dual of this quantity is also dened using Hilbert-Schmidt norm [17]. Even though such measures were realizable in terms of local observable [20{22], Piani indicated that the bipartite correlation measures based on the Hilbert-Schmidt distance are not a faithful measure of quantum correlation [23]. In order to circumvent the local ancilla issue, MIN has redened using dierent distance measures [24{27] and entropic quantities [28{30]. In recent times, these quantities were studied in various physical systems such as Heisenberg spins [31], optical systems [32], and diamond spin chain [33]. In general, the interaction between a quantum system and the environment is unavoidable and the system loses its unique quantum mechanical property. For an ecient information processing application, the quantum resources like quantum correlation should be protected against the environmental eect or decoherence. As a matter of fact, the dynamics of quantum correlation in Hermitian systems under decoherence and intrinsic decoherence is quite well studied [32, 34, 35]. The dynamical behaviors of quantum correlation under the non-Hermitian system are still address the above, in this article, we study the nonunitary evolution of quantum correlations quantied by maximal Bell function and MIN. It is evident that the Bell inequality and entanglement are of nonlocality for the non-Hermitian system also. The nonunitary evolution causes sudden death in entanglement and Bell function, whereas the MINs are more robust against decoherence and exhibits the article is prepared in the following manner: First, we provide the overview of dierent measures of in Sec. II. Then, we present the considered system in Sec. III. Further, we study the quantum for the pure state in Sec. IV A. Then, we discuss the correlation measures of the mixed state in Sec. IV B. Finally, we conclude the results in Sec.V. II. NONLOCAL CORRELATION QUANTIFIERS In this section, we review the quantum correlation measures examined in this Letbe a bipartite quantum state shared by the marginal states aandb. The degree of entanglement between the subsystems is quantied by the concurrence and is dened as [36] C() = 2 (1) whereiare the square root of eigenvalues of matrix R=p~pand arranged in decreasing order i.e., 123 4. Here ~= (y y)(y y) is the spin ipped density operator and the symbol denotes the usual in the computational basis. The well-known fact is that the concurrence varies from 0 to 1 with minimum and maximum values correspond to unentangled and maximally entangled states, nonlocality: The Bloch representation of the bipartite density operator is introduced by Fano. The general expression of a two-qubit density operator acting in the separable composite Hilbert space Ha Hbis =1 4 (2) wherej(j= 1;2;3) are the Pauli operators. The vectors xandyare real, their components being xj= the 22 unit operator acting on the subsystem a(b) andyj= Tr((1 j)),tmn= Tr((m n)) being real matrix elements of correlation matrix Tand Tr denote trace of a matrix. Without loss of generality, the canonical form of the Fano parametrization of the density operator (2) is written (3) withcj= captures the maximal nonlocal eects of bipartite state due to locally invariant projective measurements, is dened as the maximal distance between the quantum state of our consideration and the corresponding state after performing a local measurement on one of the subsystems, say ai.e.,[15] N2() (4) wherekOk2= Tr(OOy) is Hilbert-Schmidt norm of operator Oand the maximum is taken over the locally measurements on subsystem awhich does not change the state. The post-measurement is dened as a() with the projective measurements on the subsystem a, which do not change the marginal state alocally i.e., a(a) =a. Ifais non-degenerate, then the maximization is not required. For a given state Eq. (2), the MIN has a closed formula as N2() 0; Tr(TTt)T3 ifx= 0;(5) whereT3is the least eigenvalue of matrix TTt, the superscripts tstands for the transpose of a matrix. Trace distance-based MIN : It is a well-known fact that the MIN based on the Hilbert-Schmidt norm is not a bonade measure in capturing nonlocal attributes of a quantum state due to the local ancilla problem [23]. A natural way to circumvent this issue is dening MIN in terms of contractive distance measure. Another alternate form of MIN is based on trace distance [25], namely, trace MIN (T-MIN) which resolves the local ancilla problem [23]. It is dened as [25] N1() (6) wherekOk 1= Trp OyOis the trace norm of operator O. Here also, the maximum is taken over all locally measurements. For the considered system, the closed formula of trace MIN N1() is given as N1() the absolute values of ciand the summation runs over cyclic permutation of f1;2;3g. Maximal Bell function: To manifest another form of bipartite nonlocality, we employ the Bell maximal function B() as a another measure of nonclassical correlation. In 1964, J. Bell demonstrated that violation of the Bell inequality by a bipartite state using a local hidden variable [12]. In general, the violation of Bell inequality witnesses the of the quantum system. The Bell function is dened as [37, 38] B() (8) whereTi(i= 1;2;3) are the eigenvalues of the matrix TtT. The functionB()>2 quanties the amount of Bell inequality violation and the nonlocal correlation of two-qubit quantum states.4 FIG. 1. Time evolution of quantum correlations for the pure state j under non-Hermitian system for dierent values of ~ : ~ = 0:1 (magenta solid line), ~ = 0:25 (lightblue solid line), ~ = 0:5 (dark blue solid line), and ~ = 1:5 (orange solid line). (a) shows Concurrence, (b) shows Bell inequalities, (c) shows MIN, and (d) shows Trace MIN. III. MODEL AND EVOLUTION In order to understand the dynamical properties of dierent quantum correlation measures under we consider the Hamiltonian of a single qubit non-Hermitian system, it can be written as (9) where
and are the real parameters, xandzare usual Pauli spin operators. The rst term in the above the decay rate operator and the second term implies the qubit which undergoes coherent Rabi
represents the coupling strength, and characterize the non-Hermiticity of the Hamiltonian. The eigenvalues of the Hamiltonian given in Eq. (9) are E1;2=p
2 2. The eigenvalues are real when
< and turn to be imaginary when value exceeds
. The time evolution of arbitrary quantum state (0) can also be represented as (t) introducing a time-evolution operator U(t) =eiHt: Note that(t) is not a trace preserving here due to the non-Hermiticity of Hamiltonian, while observers have to measure the quantum system in a conventional quantum world. Thus, (t) has to be characterize the bipartite correlation, we consider the two-qubit initial state (0), Then the evolved state will be: ~(t) is the normalization parameter. For any two-qubit initial state, we obtain the time evolved state as ~(t) (11) IV. RESULTS AND DISCUSSIONS A. Pure state In what follows, we focus on the dynamics of bipartite quantum correlation quantied by the entanglement, nonlocality (MIN), and Bell nonlocality. For this purpose, we consider a pair of initially corre- lated qubits as initial state (0) the complex coecients satises the 2. Fig. (a) and (b) combinely shows the MIN (purple dashed line), Trace MIN (darkgreen solid line), Bell solid line) and Concurrence (cyan solid line) for two dierent values of ~ = 0:5and1:5 for an initally pure state j 1. The matrix elements of the time evolved state (12) where,=
t, ~ =
, and=p ~ 21. The normalization parameter is M analyze the evolution of quantum correlations under nonunitary evolution, we set a=b= 1=p 2 and the initial state is a maximally entangled state. Fig. 1 depicts the dynamical behaviors of quantum correlation measures for the exemplary values of ~ (= 0:1;0:25;0:5;and 1:5). Att= 0, the concurrence C(),B() and nonlocality (MIN) are maximum, it is obvious that the initial state is maximally entangled (correlated). Further, we observe that the trace MIN and concurrence are identical in capturing quantum correlation. A similar observation is made in Heisenberg spin (Hermitian) system under intrinsic decoherence as well [35]. For the lower values of ~ , we observe that the MINs, concurrence, and Bell function oscillate with time. At a given instant of time, the function B()>2, implies the existence of nonlocal aspects in the considered quantum system. Further, we observe that the considered measures are similar in quantifying nonclassical correlation. The increment in ~ factor accelerates the two-qubit oscillating correlation to reach the steady state value rapidly compared to the lower values of ~ . The measuresC(), andN2() are vanishing after a nite time, whereas B() saturates at 2 and this fact refers to the sudden death of entanglement. Further, the trace MIN N1() saturates at 0.5, implying that the trace MIN is more pronounced than the other companion quantities. In general, the entangled state witnessed through the violation of Bell inequality. From Fig. 1(b), we observe that after some nite time the Bell function is always equal to 2 which indicates that the quantum state has no nonclassical correlation and the corresponding quantum state obeys the Bell inequality. This observation is also re ected in the dynamics of entanglement. Whereas the nonzero trace MIN indicates the presence of quantum correlation in the state. Hence, trace MIN is a good measure of quantumness compared to entanglement measure. Next, we consider the above pure state with another initial condition Substituting the values ofaandbin Eq. (11), one can obtain the matrix elements of time evolved state and characterize the of quantum correlation. We have plotted the correlation measures in Fig. 2 for the xed values of ~ . Here also, we observe that all the correlation measures oscillating with time for the value of ~ = 0:5 as shown in Fig. 2(a).6 FIG. 3. Time evolution of quantum correlations for the pure state j under non-Hermitian system for dierent values of ~ : ~ = 0:1 (magenta solid line), ~ = 0:25 (lightblue solid line), ~ = 0:5 (dark blue solid line), and ~ = 1:5 (orange solid line). (a) shows Concurrence, (b) shows Bell inequalities, (c) shows MIN, and (d) shows Trace MIN. FIG. 4. Fig. (a) and (b) combinely shows the MIN (purple dashed line), Trace MIN (darkgreen solid line), Bell solid line) and Concurrence (cyan solid line) for two dierent values of ~ = 0:5;1:5 for an initally pure state j higher values of ~ say ~ = 1:5, the functionsN2() andC() decreases monotonically, then reaches to zero. As shown in the red curves of Fig. 2(b), the maximal Bell function B() saturates at 2 after a nite time and fails to manifest the nonlocality. Further, we observe that the rapid decaying behavior of N1() and the frozen dynamics after a shorter time. The above observation emphasizes that the absence of entanglement (obey the Bell inequality) not necessarily indicates the absence of nonlocality of a quantum system. To assert the pure state dynamics of quantum correlation, next, we consider another form of pure state j i(0) complex numbers and a2+b2= 1. The corresponding density matrix elements (13) At initial time t= 0, all the correlation measures for the initial conditions a=b= 1=p 2 of the statej i(0) = aj01i+bj10iare maximum, which are illustrated in Fig 3. As time increases, all the measures are exhibiting similar periodic patterns which are observed in the previous case and implying that the measures are identical in correlations. Here also, we observe that the higher values of ~ accelerate the decay rate of s from the maximal value to steady state. The quantum correlation is quantied by the concurrence and MIN vanishes after a nite time, whereas the trace MIN reaches a non-zero steady state value. This nonzero steady state correlation has practical implications in quantum information processing. Again, the Bell inequality saturates at 2 and re ects the same observation from that of understand further, we consider the initial conditions Also, in this case, the functions C() andN2() are oscillatory with nearly equal amplitudes in time for ~ = 0:5. For larger values of ~ , the measures C() andN2() vanishes after a nite time. As shown in the red curves of Fig. 4(a) and 4(b), the values of B() is maximum at t= 0, implying that the considered quantum states have a correlation between the marginal states. But for larger values of ~ (say ~ = 1:5), the scenario is dierent where the red curve decays to a constant value as depicted in Fig. 4(b). In a similar manner, the Trace MIN N1() oscillates for lower values of ~ and reaches a steady state for higher values of ~ . B. Mixed state As we mentioned earlier, the characterization of mixed state from the perspectives of quantum correlation is still not understood and entanglement is believed to be an incomplete manifestation of nonlocality. In order to study the complete nonlocal behaviors of mixed state, we consider an entangled mixed state (0) =rj (0)ih (0)j+ (1r 4)14; the purity of the state, j i(0) =1p 2(j00i+j11i) and 14is a 44 identity matrix. If r= 1, the state is maximally entangled and r= 0, the state is maximally mixed. The evolved density matrix has the structure is given in Eq. (11). The diagonal elements are, 11=(1 +r)2cosh2() + (1 + ~ 2+r(~ 21)) sinh2()(1 +r) sinh2() + +r) sinh2() + (1r)(cosh() + +r)2cosh2() + (1 + ~ 2+r(~ 21)) sinh2() + (1 (15) and the o-diagonal elements (16) To understand the role of factor r, we plot the MIN and Bell nonlocality of the considered system in Fig. 5 for xed values of ~ = 0:5;1:5. We observe that the Bell function B() is always less than 2, implying that there is no between its local constituents. On the other hand, the companion quantities (Concurrence, MIN, and T- MIN) are nonzero while the state obeys the Bell inequality. In the realm of MIN, the separable (uncorrelated) state also possesses the nonlocal correlation and quantum advantages. This nonclassical correlation of local states is quantied by the MINs. This also reveals the fact that the Bell nonlocality is an incomplete manifestation of nonlocality. Here again, for higher values of ~ , we observe similar decaying behavior and frozen dynamics of correlation measure andN2() tends to be zero. One can observe the similar eects for lower values of r. Next, we consider a initial mixed state rj (0)ih (0)j+ (1r 4)I4. Here, ifj , then the matrix8 FIG. 5. Fig. (a) and (b) shows the comparison of MIN (purple dashed line), Trace MIN (darkgreen solid line), Bell solid line) and Concurrence (cyan solid line) for two dierent values of ~ = 0:5;1:5 withr= 0:5 for a mixed state j . FIG. 6. Fig. (a) and (b) shows the comparison of MIN (purple dashed line), Trace MIN (darkgreen solid line), Bell solid line) and Concurrence (cyan solid line) for two dierent values of ~ = 0:5;1:5 withr= 0:5 for a mixed pure state j . elements are given as 11=(1 +r) sinh2() + +r)2cosh2() + (1 + ~ 2+r(~ 21)) sinh2()(1 23=32=rsinh2() 22; 44=(1 +r) sinh2() + (1r)(cosh() + +r)2cosh2() + (1 + ~ 2+r(~ 21)) sinh2() + (1 (17) Fig. 6 shows the comparative dynamics of concurrence ( C()), MIN (N2()), TMIN (N1()) and for the mixed state with two dierent values of ~ . As discussed in the earlier case of mixed state, here also, the correlation measures exhibit periodic oscillations with respect to time. For the values of ~ <1, the andN1() clearly depict the presence of quantum correlations whereas B() does not capture the presence of quantum correlations as reported in Fig. 6(a). Again, the function B() is less than 2 for any value of ~ weakly denied to show the correlations. As seen in Fig. 6(b), the other two functions C(), andN2() turns to be zero indicates the absence of quantum correlations for ~ >1. Nevertheless, the quantum correlation is clearly quantied by the N1()9 for any given value of ~ >1 as shown in Fig. 6(b). The above observation clearly indicates that the N1() acts as an eective quantier of bipartite quantum correlation and indicators of nonlocal attributes of a quantum system. V. CONCLUSIONS In conclusion, we have examined the dynamical behaviors of quantum correlations quantied by the entanglement, nonlocality, and Bell nonlocality under local nonunitary evolution. For the pure state, at lower values of ~ , the correlation measures show the periodic oscillation and the higher values of ~ causes rapid decay of quantum correlation to a steady state value. The trace distance MIN reaches its steady state value after some nite time while MIN and entanglement decay to zero. Further, the Bell inequality saturates at 2. We observe that the presence of nonlocality in terms of trace MIN, whereas the Bell inequality and entanglement fail to capture the quantum correlation. For the initially mixed state, we observe that Bell inequality is completely failed to capture the nonlocality even when the qubits are entangled. This also implies a fact that the presence of entanglement does not necessarily indicate the nonlocal attributes of the quantum system. It is observed that the MINs are more robust against nonunitary evolution and pronounced than the Bell inequality. Our results also emphasize that the trace distance based nonlocality oers more resistance to nonunitary evolution than the Bell inequality and entanglement. Further, we observe that the time evolution of Hermitian and non-Hermitian systems exhibit similar thanks the Department of Science and Technology, Government of India, for providing an INSPIRE Fellowship No. DST/INSPIRE RM acknowledge the nancial support from the Council of Scientic and Industrial Research (CSIR), Government of India, under Grant No. The work of VKC forms part of the research projects sponsored by SERB-DST-MATRICS Grant No. MTR/2018/000676 and CSIR Project Grant No. JRP, RM, and VKC wish to thank DST, New Delhi for under the DST-FIST programme (SR/FST/PS- 1/2020/135) to the Department of Physics. [1] C. M. Bender and Boettcher, Phys. Rev. Lett. 80, 5243-5246 (1998). [2] C. M Bender and Boettcher, and S. Meisinger, J. Math. Phys, 40, 2201-2229 (1999). [3] C. M. Bender, D. C. Brody, and H. F. Jones, Phys. rev. Lett. 89, 270401 (2002). [4] A. Sergi, and K.G. Zloshchastiev, Int. J. Mod. Phys. B 27, 1345053 (2013). [5] A. Sergi, and K.G. Zloshchastiev, Phys. Rev. A 91, 062108. [6] C. M. Bender, Rep. Prog. Phys. 70, 947-1018 (2007). [7] H. Feshbach, C. E. Porter, and V. F. Weisskopf, Phys. Rev. 96, 448-464 (1954). [8] M. Nielsen, and I. Chuang.: Quantum Computation and Quantum Information, Cambridge University Press, T. Baumgratz, M. Cramer, and M.B. Plenio, Phys. Rev. Lett. 113140401 (2014). [10] A. Einstein, B. Podolsky, and N. Rosen, Phys. Rev. 47, 777 (1935). [11] E. Schrodinger, Proc. Cambridge Philos. Soc. 31, 555 (1935). [12] J. S. Bell, Physics 1, 195 (1964). [13] J. Schneeloch, C.J. Broadbent, S.P. Walborn, E.G. Cavalcanti, and J.C. Howell, Phys. Rev. A 87, 062103 (2013). [14] H. Ollivier, and W.H. Zurek, Phys. Rev. Lett. 88, 017901 (2001). [15] S. Luo, and S. Fu, Phys. Rev. Lett. 106, 120401 (2011). [16] S. Luo, Phys. Rev. A 77, 022301 (2008). [17] B. V. Daki c, and V. Brukner, Phys. Rev. Lett. 105, 190502 (2010). [18] S. Luo, and S. Fu, Phys. Rev. A 82, 034302 (2010). [19] S. Xiong, W. J. Zhang, C. S. Yu, and H-S. Song Phys. Lett. A 378, 344 (2014). [20] J.-S. Jin, F.-Y. Zhang, C.-S. Yu, and H-S. Song, J. Phys. A: Math. Theor. 45, 115308 (2012). [21] G. Passante, O. Moussa, and R. La amme, Phys. Rev. A 85, 032325 (2012). [22] D.Girolami, and G. Adesso, Phys. Rev. Lett. 108, 150403 (2012). [23] M. Piani, Phys. Rev. A 86, 034101 (2012). [24] R. Muthuganesan, and R. Sankaranarayanan Phys. Lett. A 381, 3028 (2017). [25] M-L. Hu, and H. Fan, New J. Phys. 17, 033004 (2015). [26] R. Muthuganesan, and V. K. Chandrasekar, Commun. Theor. Phys. 72, 075103 (2020).

2108.11041

V.Kuppusamy Chandrasekar

J. Ramya Parkavi, R. Muthuganesan, and V. K. Chandrasekar

Facets of nonlocal correlation under non-Hermitian system

10 pages, 6figures

quant-ph

http://creativecommons.org/licenses/by/4.0/

In this article, we investigate the dynamics of a bipartite system under the action of a local non-Hermitian system. We study the quantum correlation of the bipartite system quantified by the entanglement, (MIN) based on Hilbert-Schmidt norm, trace distance, and Bell inequality. We find that the quantum correlations of the system depend on the initial conditions and system parameters. We observe that the states with nonzero quantum correlation obey the Bell inequality even in the absence of entanglement. Moreover, the Bell inequality completely fails to manifest the nonlocality for the mixed quantum state. However, we have identified the nonlocal attributes of the mixed quantum state in terms of MIN and trace distance MIN. Our results show that the trace distance-based correlation is more robust against the nonunitary evolution compared to the other quantifiers.

identify influential spreaders [22-26]. However these prog resses are mostly in single layer networks, while some real complex systems are better represented as multilayer netwo rks [27]. For example an individual may have others in different ways, either being friends, colleagues, schoolmates , or doing business, which is a [28]. Integrating these di fferent types of interactions into a single network may lose so me critical information of the system. The multilayer network approach is adopted in understanding the robustness of infrastructures [29, 30], coevolving spreading dynamics of information and epidemic s [31], evolutionary games [32], functions of brain [33] and stability of economical and financial systems [34]. When an epidemic outbreaks in the contact population, infor mation on the epidemic is easily transmitted through kinds of channels, such as the online social platform and mob ile communication network. The information on epi- demic disease promotes people to adopt self-protection mea sures to reduce their risks of being infected, thus helping to suppress the diffusion of epidemic. Meanwhile, the wide spread of epidemic fu rther enhances the spreading of information [35]. These can be considered as two processes o n a two-layer multiplex network where the informa- tion diffuses on the communication layer and the disease spreads on th e physical contact layer, and nodes are the individuals [31]. It is found that the asymmetrical interpl ay between layers has significant impact on the on top of the multiplex networks, such as changing t he epidemic threshold and suppressing the in the stationary state [36-39]. As for searchin g for the influential spreaders, it is natural to consider tha t the spreading influence of a node in the physical contact laye r will be reduced due to the information spreading on layer, but the extent of reduction depends on the structure and dynamics on the communication layer. To address the issue of identifying the most important nodes in multiplex networks, a lot of centralities and methods are proposed, such as the multiplex PageRank [40, 41 ], multiplex eigenvector centrality [42, 43], [44], based methods [45 ], and methods based on the local or global structure of the networks [46-48]. These methods provide more accurate rank ings than approaches based on the aggregated networks or single-layer networks. However, these methods mainly fo cus on the multilayer structure, while neglecting the dynamical couplings between layers. Researches have point ed out that the influence of a node in the spreading process is a result of the interplay between dynamics and network str ucture [49, 50], and measures are more efficient in identifying super-spreaders than structural meas ures in single networks and interconnected networks [51, 52]. In this paper, we work on identifying the most influential spr eaders in the coupled dynam- ics. By considering three dynamical and structural couplin gs between two layers, which are the two-layer rate, the inter-layer coupling strength and the i nter-layer degree correlation, we propose a measure centrality (CS) to identify super-spr eaders in the coevolving dynamics on multiplex networks. Si m- ulation results show that the CS centrality is not only much m ore accurate than the centralities of degree, k-shell index and PageRank in the contact layer , but is also more accurate than two typical multiplex cen- tralities, which are the multiplex PageRank and multiplex e igenvector centrality. While being applied to a variety of real-world multiplex networks, the proposed measure als o works well, implying that considering the layers is crucial in finding key nodes in multilayer s ystems. The rest of the paper is organized as follows. In section 2, we give the preliminaries. In section 3, we propose the centrality. In section 4, we repres ent the effectiveness of the proposed centrality in and real-world networks. Finally we give a conclus ion in section 5. 2. Preliminaries In this part, we first describe the cou pled spreading model on a two-layer multiplex network. Then we give a brief description of the benchmark centraliti es of degree, eigenvector centrality, k-shell index and PageRank on single layer network, and the multiplex PageRan k and multiplex eigenvector centrality on which are used as competitors to our measure to iden tify super-spreaders. Finally we give the definition of the imprecision function and Kendall's tau correlation coe fficient, which are two methods to evaluate the performance of different measures in identifying The coupled model on multiplex n etwork To describe the coevolving dynamics , we use the SIR-SIRV spreading model in ref. [38]. In this model, a multiplex network is composed of two layers whe re the nodes represent individuals and links 1: The schematic representation of the information- epidemic coupled dynamics on a multiplex network. Layer A is the where information spreads, and layer B is the physic al contact network where disease spreads. (a) At t=0, a node is informed and infected and all other nodes are in susceptible states in both layers. The informed and infected node transmits information and di sease to its neighbors in layer A and B respectively with rate lAandlB. (b) The informed node in layer A makes its counterpart in lay er B be vaccinated with rate lAB, and the infected node in layer B makes its counterpart in layer A b e informed with rate lBA. their interactions. The information spreads on the upper co mmunication layer A and the disease spreads on the bottom physical contact layer B. In the communication layer, the cl assical (SIR) model is us ed to describe the information spreading process. In the SIR mo del, a node can be in one of the three states: susceptible (S), infected or informed in information spreading (I) and r ecovered (R). At each time step, an I-state node infects or informs all its susceptible neighbors with rate bAand then recovers with rate uA. The spreading stops until there is no I-state node in the network. In the physical contact lay er where disease spreads, the same SIR dynamics are adopted with an additional state vaccinated (V), where node s of V-state will neither be infected nor transmit disease. The disease transmission rate is bBand the recovering rate of infected nodes is uB. The effective transmission rates for layer A and B are respectively dynamical coupling of the two layers is as follows. For an informed node iAin layer A, if its counterpart iB in layer B is in S-state, then iBtransfers to state V with immunization rate lAB. This immunization rate represents the willingness or capability of the informed individuals to ge t vaccinated. The larger lABis, the more counterparts of the informed nodes will get vaccinated and the suppression e ffect from information layer is more significant. For an infected node jBin layer B, if its counterpart jAin layer A is in S-state, then jAtransfers to state I with rate lBA. the probability that an infected individual is a ware of the epidemic or is willing to transmit information on it. The largerlBAis, the more nodes in layer A become aware of the epidemic and t he information spreads more widely. A schematic representation of the SIR-SIRV model is shown in Fig. 1. In simulations, each layer of the multiplex network is a scal e-free network with power law degree We generate layer A from the uncorrelated configuration mod el (UCM) with the number of nodes N=10000, the power exponent g=2.6 and the average degree <k>=6. The minimal degree is set as kmin=3 and the maximal degree is set as kmax= N. Then we generate layer B by keeping the same node sets of laye r A. The degree sequence of layer B is generated by copying the deg ree sequence of layer A first and then exchanging the degrees of randomly selected pairs of nodes until the specifi ed inter-layer degree correlation is reached. The multiple x network is thus a set of nodes with two di fferent types of connections represented by two layers. The in ter-layer degree correlation of the two layers is quantified by the Spea rman rank correlation coe fficient 2.1. Inter-layer degree correlation. The inter-layer degree co rrelation is defined (1) where Nis the number of nodes, iis the rank difference of node iin two ordered sequences ranked by degree in layer A and layer B respectively. msranges in [-1,1]. If the degrees of nodes in two layers are pos itively correlated, ms-1. If the degrees of nodes in two layers are negatively correlat ed,ms-- 1. If the degrees of nodes in two layers ms-0. The degree correlation of a node in two layers has a potenti al impact on the spreading influence of the node in the contact layer. From the centrality's persp ective, a node with larger degree has a greater in the network. For nodes with the same degree in the contact layer B, their degrees in the A imply different suppression e ffects of information on epidemic spreading, making the centr ality in the contact layer be unable to predict the spreading influence of nodes in disease spreading. In the coevolving dynamics, initially a seed node is infecte d on layer B and its counterpart in layer A is informed. The information and disease spread in layer A and layer B resp ectively until a final state is reached. As we focus on disease spreading, we take the proportion Riof recovered nodes in the contact layer in the steady state as the spreading influence of the seed node i. The disease transmission rate lBis set to be 0.13 for the synthesized network which is three times of the epidemic threshold of layer B. As i ndicated in ref. [54], the transmission rate will not impact the relative ranking of node influence if it is above th e epidemic threshold and within a few times of The recovery rates are set as uA=uB=1 for simplicity. The relative spreading rate of layer A and l ayer B isgl AB=lA/lB, which reflects the dynamical coupling of the two layers. If gl AB=lA/lBis large enough, we can consider that the information spreads first and the disea se spreads on the residual network after vaccination. All results in simulations are obtained by averaging over 100 ru ns. 2.2. Competitor Centralities on single-layer network Consider a network G(V,E), where Vis the set of nodes, Eis the set of edges, and N=|V|is the number of nodes. The adjacent matrix of GisAN*N=ai j, where ai j=1 if there is an edge between node iandj, otherwise ai j=0. The degree kquantifies how many direct neighbors a node has, which is j. Degree is the simplest in quantifying node 2.2. Eigenvector centrality. The eigenvector centrality of nod eiis defined jej, (2) which givesle=Aein matrix notation. Here eis the right leading eigenvector corresponding to the large st eigenvalue lof the adjacent matrix A. The eigenvector centrality takes i nto account both the quantity and quality of neighbors in determining the importance of a node. Definition 2.3. K-shell index. The k-shell index ksis obtained in the k-shell decomposition process. Initiall y, nodes with degree k=1 are removed from the network together with their links. Aft er removing all nodes with k=1, some nodes initially with degree more than one may have only one li nk left. Continue to remove them until no node with degree one is left. The removed nodes are assigned a k-shell v alue ks=1. Then nodes with degree k<=2 are removed in the same way and given a value ks=2. The pruning process continues until all nodes are removed and given a ks value. Nodes with large ksare considered to be in the core of the network and are super-s 2.4. PageRank Centrality. The PageRank centrality is famous for its success in the web ranking technology used by Google corporation. It is defined jPR (3) where kout jis the out-degree of a neighbor node j, and cis a free parameter to represent the random jump of users not visiting along the links. The algorithm iterates until a ste ady-state and the PRifor each node is Centralities on multiplex network First we introduce the Functional Multiplex PageRank (FMPR ) defined in ref. [41], which is a generation of PageRank centrality to the multiplex networks. This centra lity considers the weights of di fferent types of two nodes. Suppose a multiplex network- is composed of a set VofNnodes and M layers. Each layer is Ga=(V,Ea), where Eais the set of edges in layer aanda=1,2,...,M. The adjacent matrix element aa i j=1 if node iand jhas a link in layer a, otherwise aa i j=0. To characterize the multiplex network with overlap, a vector called multilink- -mi jis defined, where- -mi j=(m1,m2,...,mM) has elements ma=0,1. The number of possible types of multilink is 2M. A pair of nodes iand jare connected by a multilink- -mi jif and only if - -mi j=(a[1] i j,a[2] i j,...,a[M] i j). Then the multiadjacency matrices A- -mdefines the connection of nodes with multilink- -mi j, with elements A- -m i j=1 if the node pair iandjis connected by a multilink- -mi j, otherwise A- -m i j=0. Therefore A- -m i jcan be expressed as A- -m i j)]. (4) For the two-layer multiplex network the multi-adjacent mat rices are A(1,0) i j=a[1] i j(1-a[2] i j) A(0,1) i j=(1-a[1] i j)a[2] i j A(1,1) i j=(a[1] i ja[2] i 2.5. The Functional Multiplex PageRank centrality Xi(z) of node iis defined -mi j i jz- -mi j1 kjXj+bvi, the possibility that in the steady state a random walker ju mps from node jto a neighbor i, otherwise it jumps to a randomly connected node in the multiplex network. zis a tensor with elements z- -m>=0 associated to every type of multilink- -mrepresenting its influence, and z- -0=0. A random walker jumps to a neighbor along the multilink- -mwith probability proportional to z. In Eq. -mi j i jz- -mi -mi j i jz- -mi -mi j i jz- -mi -mi j i jz- -mi -mi j i jz- -mi -mji jiz- -mi the Kronecker delta and th(x) is the Heaviside step function. The Functional Multiplex P ageRank centrality Xi(z) of node idepends on the values of z. In our calculations for the two layer network, we use a represent an equal importance of each type of the multilin ks. The second multiplex centrality used in this manuscript is t he Global Heterogenous Eigenvector-like centrality (GHEC ) proposed in ref. [42]. This centrality takes into account th e contribution of neighbors from all layers and the mutual influence between layers. Consider an M*Minfluence matrix Wwhere the non-negative elements wabrepresents the influence of the layer bon layera, which is W=w11*** wM,M, (8) 5and a N*NM matrix Then Ais the Khatri-Rao product of the matrices WandA, which wMMAM. (9) The Global Heterogenous Eigenvector-like centrality of th e multiplex network- -Gis the positive and normalized eigen- vector cRNMof the matrix A. If one introduces the vectors c 1,...,c mRN, then the Global Heterogenous Eigenvector-like centralit y matrix of- -Gis given by The information contained in the vectorial-type centrali ty should be aggregated to associate a number to each node, which is where CRN*1and the ith-row Ciis the Global centrality of node iin the multiplex network. 2.3. Evaluation methods To evaluate the performance of measures in identifying supe r-spreaders, we use the imprecision function [6] and the Kendall's tau correlation coe fficient [55]. The imprecision function quantifies how close to the optimal spreading is the average spreading of pNnodes with the highest 2.6. Imprecision function. The imprecision function is defined a s e(p)=1-M(p) Me f f(p), (11) where pis the fraction of nodes considered ( p[0,1]).M(p) is the average spreading influence of pNnodes with the highest centrality, and Me f f(p) is the average spreading influence of pNnodes with the highest spreading influence. A smallerevalue indicates a more accurate centrality to identify supe r-s 2.7. Kendall's tau correlation coe fficient. The Kendall's tau correlation coe fficient is used to between two rankings which is defined (12) where K(r1,r2) is the number of node pairs that appear in the concordant ord ering in the ranking lists r1andr2, and K'(r1,r2) is the number of node pairs that appear in the reverse orderi ng in r1andr2.nis the number of nodes in each ranking list. In our applications, nodes are ranked by centr ality measure in ranking list r1and nodes in ranking list r2 are ranked by spreading influence, which are obtained by comp uter simulations. A large tindicates that the and the node spreading influence are highly correlat ed. 3. The proposed centrality In this part, we first show the two-area phenomena of the centr alities to predict the spreading influence of nodes in the contact layer. Then we propose the e centrality (CS) in multiplex network to predict influence of nodes in the coevolving dyna mics. 6Figure 2: The scatter plots of nodes with their disease-spre ading influence Rand degree centrality kB(a)-(c) or eigenvector centrality eB(d)-(f) in the contact layer B. The color of plots represents centralit ies of nodes in layer A. It can be seen that the spreading influe nce of nodes with in layer A is obviously suppressed. The paramete rs are set as: (a) and (d) degree correlation ms=0.3, relative spreading rate ratelBA=0.1, and immunization rate lAB=1.0; (b) and (e) (c) and (f) We take these parameter values for example to reflect di fferent dynamical and structural coupling strengths and disc uss their impacts later in detail. 3.1. The two-area phenomena of centrality in predicting dis ease-spreading influence The centrality measures are heuristic ways to predict the sp reading influence of nodes in the network. The idea is that the more central a nodes is, the more influential it is in s preading, where the centrality is positively related with the node spreading influence [19]. Figure 2 displays the scat ter plots of nodes with their influence Rand degree centrality kBor eigenvector centrality eBon the contact layer B under di fferent parameters. We use three groups of parameter values for the relative spreading rate gl AB, the inter-layer dynamic coupling strengths lABandlBA and the inter-layer degree correlation ms, where the dynamical coupling strength and structural coup ling strength vary. It can be seen that with the increase of kBandeB, the spreading influence Rincreases in general. But there are some nodes where their spreading influences are lower than th at of others with the same centrality of kBoreB, and a two-area phenomena appears. The suppressed spreading influ ence of these nodes is due to the coevolving dynamics between layers. Let's consider two nodes iBand jBwith the same centrality in layer B and their counterparts in layer A are If iAhas a larger centrality than jA, then the spreading of information originating from iAis supposed to be wider than that originating from jA, making more nodes in layer B be vaccinated. The influence of node iBin layer B is thus smaller than that of jB. It can be seen from Fig. 2 that for nodes with relatively high centrality in layer A (nodes with color approaching red), their influence s are obviously lower than that of nodes with lower centrality in layer A under di fferent parameters. This that the centralities on the contact layer are not ad equate to identify the super-spreaders in the on the multiplex networks. 3.2. The centrality on multiplex netwo rk We take into account the dynamical and structural couplings between layers and propose a (CS) centrality to identify super-spreaders on multiplex netwo rks. Definition 3.1. centrality. The c entrality of node iis defined the benchmark centrality of node iin layer B andthA iis the centrality of its counterpart node in layer A. th can be any centrality measure such as degree, eigenvector ce ntrality and k-shell index. lAandlBare the in layer A and B respectively. lABis the immunization rate and lBAis the informing rate. 7In the definition of CS centrality, the first term on the right s ide of Eq. (13) is the node centrality in the contact layer, which is taken as the baseline. The second term repres ents the impact of information spreading on where the suppression e ffect of node idepends on the transmission rate lAin layer A and the strengthlABfrom layer A to layer B. thA iandthB iare the centralities of node iin layer A and B the spreading abilities of node iin layer A and B respectively. Existing researches have indi cated that the degree centrality is positively correlated with many other centralities, such as the eigenvector centrality and k-she ll centrality, Eq. (13) reflect the inter-layer degree correlation. As th e suppresses the , the spreading influence of node iin layer A has a negative impact on the baseline centrality thB i in predicting the , thus the second term h as a minus sign. On the other hand, the disease spreading in layer B has an impact on layer A through inter-layer coupli ng, which depends on the transmission rate lBand the inter-layer coupling strength lBA. The more nodes informed by infected nodes in layer B, the les s impact of the spreading origin in layer A on information spreading. Thus w e add the third term representing the reduced impact of node centrality in layer A. The algorithm to calculate the co upling-sensitive centrality is shown in Algorithm 1. Algorithm 1 Algorithm to obtain A multiplex network- -G=(G1,G2), where G1=(V,E1) and G2=(V,E2)//The multiplex network consists of a set of nodes V and two types of edges E1andE2which are layer A and layer B The centrality of each node CSth i; 1:Set the values of parameters to|V|do 3: Calculate the centrality thB iof node iin layer B; 4:end for 5:fori=1 to|V|do 6: Calculate the centrality thA iof node iin layer A; 7:end for 8:fori=1 to|V|do 9: Calculate the centrality CSth iof node iin the multiplex network using Eq. (13) 10: for 4. Experimental results We evaluate the performance of the CS centrality and compare it with that of the benchmark centralities in syn- thesized networks and real-world networks by large simulat ions. Results show that the proposed CS centrality outper- forms the corresponding benchmark single-layer centralit y and two typical centralities defined on multiplex networks in identifying super-spreaders in co evolving process. 4.1. Performance comparison of centra lity and single-layer centrality We first compare the performance of CS centrality with that of the single-layer centrality. We use degree centrality kB, eigenvector centrality eB, k-shell index ksBand PageRank centrality PRBin layer B as the benchmark and calculate the corresponding CS centralit ies, which are The imprecisions of these centralities are shown in Fig. 3. It can be seen from F igs. 3 (a)-(d) that the CS centralities are more accurate to identify super-spreaders in disease spreading than thei r corresponding benchmark centralities in layer B. Similar results are obtained under a di fferent set of parameters as shown in Figs. 3 (e)-(f). We use the relatively emphasize on the cases when the spreading influences of nod es in epidemic are strongly impacted by the information spreading. 80 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 3: Comparison of the imprecisions of benchmark centr alities and centralities as a function ofp.pis the proportion of nodes considered. (a)-(d) The parameters are set as (e)-(h) The parameters are set as The smaller imprecision, the more accurate the measure is in predicting the spreading influence. 4.2. Performance comparison of centra lity and multiplex centrality Next we explore and compare the performance of CS centrality with two centralities defined on multiplex net- works. We use the eigenvector centrality as the benchmark ce ntrality to calculate the centrality CSe, and compare with the Functional Multiplex PageRank [41] an d the Global Heterogeneous Eigenvector-like cen- trality [42]. To validate the robustness of the proposed CS c entrality, we vary the structural and dynamical which are represented by the immunization rate lAB, the informing rate lBA, the relative spreading rate gl ABand the inter-layer degree correlation ms. Results indicate that the CS centrality is in general robus t under all considered parameters, as shown in Fig. 4-Fig.7. Figure 4 displays the imprecisions of centralities when the immunization rate lABvaries. WhenlAB<=0.5, the imprecision of CSeis a little higher than or close to that of the multiplex centr ality FMPR. When the other two centralities. This is because whe n the immunization rate lABis relatively small, according to the definition of CS centrality, the second term is relatively small, and the CS centrality is more close to the centrality of node in layer B. When lABbecomes large, the CS centrality is more impacted by the cent ralitythA i of node iin layer A. In this case the CS centrality can reflect the contr ibutions of both layers, thus is more e ffective. While the informing rate lBA, the relative spreading rate gl ABand degree correlation msvary respectively, as shown in Fig. 5-Fig.7, the imprecision of CSeis the smallest, which indicates that the coupling-sensiti ve centrality is the best measure to identifying the super-spreaders in the coevolvi ng we calculate the Kendall's tau correlation co efficient of the proposed CS centrality with the of nodes, and compare with that of the multiplex cen tralities. It be can be seen from Fig. 8 that the CS centrality has the largest correlation with the spreading i nfluences of the influential spreaders. 4.3. Application in real-world networks Many real-world networks can be represented as multilayer n etworks, such as the biological networks, the social networks and the cooperation networks. In this part, we appl y the centrality in eight networks and compare its performance with other two multiplex centralities. For the multilayer networks with more than two layers, we choose two layers and use the mut ually connected giant component in which the nodes are connected to the largest connected component in both lay ers for study. The real-world networks studied are: (1) SacchPomb (gene and protein interaction network. Layer A is a synthetic genetic interaction network and layer B is a physical association network); (2) Drosophila (protei n-protein network with di fferent nature of interactions as layers. Layer A is suppressive genetic interaction and laye r B is additive genetic interaction); (3) IPv4 IPv6 of Autonomous System (AS). Layer A is IPv4 AS netw ork and layer B is IPv6 AS network); (4) Human 90 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 4: Imprecisions as a function of pfor the multiplex eigenvector centrality GHEC, multiplex P ageRank FMPR and centrality CSebased on eigenvector centrality in identifying super-spre aders when the immunization rate lABvaries. 0 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 5: Imprecisions as a function of pfor the multiplex eigenvector centrality GHEC, multiplex P ageRank FMPR and centrality CSein identifying super-spreaders when the informing rate lBAvaries. 100 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 6: Imprecisions as a function of pfor the multiplex eigenvector centrality GHEC, multiplex P ageRank FMPR and centrality CSein identifying super-spreaders when the relative spreadin g 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 7: Imprecisions as a function of pfor the multiplex eigenvector centrality GHEC, multiplex P ageRank FMPR and centrality CSein identifying super-spreaders when the inter-layer degre e correlation msvaries. 110.1 0.2 0.3 0.4 0.2 0.3 0.4 0.2 0.3 0.4 0.2 0.3 0.4 8: Kendall's tau correlation of the centrality measu re and the spreading influence of nodes under di fferent parameters. The centralities used are the multiplex eigenvector centrality GHEC, multiplex P ageRank FMPR and centrality CSe.pis the proportion of (a) (b) (c) (d) (structural and functional networks where nodes are b rain regions. Layer A is the structural network and layer B is the functional network ); (5) Physicians (di fferent types of relationships among physicians in four US tow ns. Layer A is a discussion network and layer B is a friendship network) ; (6) arXiv (co-authorship of scientists in di fferent research categories. Layer A is a co-authorship network in d ata analysis, statistics and probability and layer B is a co-authorship network in physics and society); (7) Air train (airports network and train stations network in India . Layer A is the train network and layer B is the airport network ); (8) Pardus (friendship relationship and relationship between individuals in an onli ne game virtual society. Layer A is the friendship network and layer B is the message communication network). Data for t he first seven networks are collected from ref. [56] and there are more detailed descriptions of the data and their or igins, and data for the last network are collected from the massive multiplayer online [28]. The characteristics of the studied ne tworks are listed in Table 1. Table 1: Characteristics of the real-world multiplex netwo rks studied in this work. These characteristics include the number of nodes N, the number of edges EAin layer A and EBin layer B, the inter-layer degree correlation of two layers ms, the infection rate lBused in the SIR spreading in layer B, and the type of N E A EB mslB Type SacchPomb 426 2236 1678 0.27 0.28 449 2656 2172 0.65 0.21 Biological IPv4 IPv6 4710 48026 25366 0.55 0.018 brain 74 426 396 0.24 0.517 106 460 362 0.47 0.818 Social arXiv 2252 15926 14570 0.94 0.24 66 634 354 0.44 0.185 2501 16119 20860 0.65 0.052 Social It can be seen from Fig. 9 that in all studied networks, the cou pling-sensitive centrality outperforms the other two multiplex centralities. This implies that considering the structural and dynamical couplings between layers is significantly meaningful to identify the most important nod es in the multilayer networks. 120 0.05 0.1 0.15 FMPR CSe 0 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 Brain 0 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 0.05 0.1 0.15 9: Imprecisions of the centralities on real-world mu ltiplex networks as a function of p. The centralities used are the multiplex GHEC, multiplex PageRank FMPR and coupling-sen sitive centrality CSe. The parameters are set as each network is three times of the epidemic threshold of l ayer B. 5. Conclusion and the most influential spreaders is an important s tep to make use of available resource and control the spreading process. In this paper, we work on identifying sup er-spreaders in the coupled spread- ing dynamics on multiplex networks. We find that the centrali ties on the contact layer are no longer e ffective to identify the most influential which is d ue to the suppression e ffects from the information layer. By considering the centralities of nodes in both layers and the structural and dynamical couplings between layers, we propose a new measure called centralit y on multiplex networks. Simulation results on and real-world networks indicate that the CS centr ality is not only much more e ffective than the centralities on the single contact layer, but also more e ffective than two typical multiplex centralities without con sidering real-world multiplex networks studied in our simulatio ns are not necessarily the networks where information and disease spread. But we can consider that the dynamic coup lings in our measure have their specific meanings in different contexts of the various real-world networks. For exam ple, the immunization rate lABmay represent the proportion of autonomous systems (ASs) routing IPv4 packet s that can also route IPv6 packets in the IPv4 IPv6 multiplex network, and the informing rate lBAmay represent the proportion of ASs routing IPv6 packets tha t can also route IPv4 packets. While in the air train network, the immunization rate lABmay represent the possibility that the travellers planning to travel by airplane switch to trav el by train, and the informing rate lBAmay represent the possibility of travellers changing the transportation mod e from train to airplane. The e ffectiveness of the centrality implies that when identifying the mos t important nodes in the multilayer systems, considering th eir structural and dynamical couplings are very necessary. Our work gives a possible way to synthesize the s into one measure which may find its applications in various multilayer systems. The computation of CS centrality contains the spreading rat es of two layers, which are the immunization rate and the informing rate. Their values vary depending on the sp ecific real-world scenarios of information spreading and epidemic spreading. For example the basic reproduction number R0, which is the average number of caused by a primary case, is used to estimate the s preading parameters, where R0=l/u[20]. Our that in a reasonable parameter range, for exam ple the information spreading rate is larger than the epidemic spreading rate corresponding to gl AB>1, the CS centrality is e ffective and robust under di fferent values of the In addition, in the definition of the ce ntrality, there are two represent the mutual influence of two layers. If we add tw o weighting factors a andbbefore each item, adjusting them may come to an optimal expre ssion of the centrality. 13We work on the two-layer multiplex network where the informa tion spreading and epidemic spreading coevolves on each layer. For networks with more than two layers the iden tification of critical nodes will be more complex due to the coupling and mutual influences between layers. The study of more general index in multilayer networks requires further explore in the work is supported by the National Natural Science Found ation of China (Grant Nos. 61802321, 11975099), the Sichuan Science and Technology Program (No. 2020YJ0125 ), the Natural Science Foundation of Shanghai (Grant No. 18ZR1412200) and the Southwest Petroleum University In novation Base S. P. Borgatti. Identifying sets of key players in a socia l network, Comp. Math Organiz. Theor. 12 (1) (2006) 21-34 . [2] S. Pei, F. Morone, H. A. Makse, Theories for influencer ide ntification in complex networks, In Lehmann S., Ahn YY . (eds) Complex in Social Systems, Springer, Cham (2018) 125-148 . [3] D. J. Watts, P. S. Dodds, Influentials, networks, and publ ic opinion formation, Res. Consumer 34 (4) (2007) 441-458. [4] G. Del Ferraro, A. Moreno, B. Min, et al., Finding influent ial nodes for integration in brain networks using optimal pe rcolation theory, Nat. Commun. 9 (1) (2018) 2274. [5] D. Kempe, J. Kleinberg, E. Tardos, Maximizing the spread of influence through a social network, In Proc. 9th ACM SIGKDD Int. Conf. on Know. Dis. and Data Mining (2003) 137-146. [6] M. Kitsak, L. K. Gallos, S.Havlin, et al., Identification of influential spreaders in complex networks, Nat. Phys. 6 (1 1) (2010) 888-893. [7] F. Morone, H. A. Makse, Influence maximization in complex networks through optimal percolation, Nature 524 (7563) (2 015) 65-68. [8] A. Braunstein, L. Dall'Asta, G. Semerjian, L. Zdeborov' a, Network dismantling, Proc. Nat. Acad. Sci, USA 113 (44) (2 016) 12368-12373. [9] L. Tian, A. Bashan, D.N. Shi, Y . Y . Liu, Articulation poin ts in complex networks, Nat. Commun. 8 (1) (2017) 1-9. [10] F. Morone, G. Del Ferraro, H. A. Makse, The k-core as a pre dictor of structural collapse in mutualistic ecosystems, N at. Phys. 15 (1) L. C. Freeman, Centrality in social networks conceptua l clarification, Soc. Netw. 1 (3) (1978) 215-239. [12] P. Bonacich, P. Lloyd, Eigenvector-like measures of ce ntrality for asymmetric relations, Soc. Netw. 23 (3) (2001) 191-201. [13] G. Sabidussi, The centrality index of a graph, Psychome trika 31 (4) (1966) 581-603. [14] L. C. Freeman, A set of measures of centrality based on be tweenness, Sociometry 40 (1) (1977) 35-41. [15] S. Brin, L. Page, The anatomy of a large-scale hypertext ual Web search engine, Comp. Netw. 30 (1-7) (1998) 107-117. [16] B. Bolob' as, Graph Theory and Combinatorics: Proc. Cam bridge Combinatorial Conf. in honor of P. Erd os, Academic P ress, NY , 1984. [17] T. Martin, X. Zhang, M. E. J. Newman, Localization and ce ntrality in networks, Phys. Rev. E 90 (5) (2014) 052808. [18] L. L u, D. Chen, X. L. Ren, Q. M. Zhang, Y . C. Zhang, T. Zhou , Vital nodes identification in complex networks, Phys. Rep. 650 (2016) 1-63. [19] M. Jalili, M. Perc, Information cascades in complex net works, J. Comp. Net. 5 (5) (2017) 665-693. [20] R. M. Anderson, R. M. May, Infectious diseases of humans : dynamics and control, Oxford university press (1992). [21] J. Zhang, M. Litvinova, Y . Liang, Y . Wang, W. Wang, S. Zha o, Q. Wu, S. Merler, C. Viboud, A. Vespignani, et al.,Changes in contact patterns shape the dynamics of the COVID-19 outbreak in china, Scienc e 368 (6498) (2020) 1481-1486. [22] F. Radicchi, C. Castellano, Leveraging percolation th eory to single out influential spreaders in networks, Phys. R ev. E 93 (6) (2016) 062314. [23] Y . Liu, M. Tang, Y . Do, P. M. Hui, Accurate ranking of influ ential spreaders in networks based on dynamically asymmetr ic link weights, Phys. Rev. E 96 (2) (2017) 022323. [24] B. Min, Identifying an influential spreader from a singl e seed in complex networks via a message-passing approach, E ur. Phys. J. B 91 (1) (2018) 18. [25] T. Wen, D. Pelusi, Y . Deng. Vital spreaders identificati on in complex networks with multi-local dimension, Know.-B . Sys. 195 (2020) 105717. [26] A. Zareie, A. Sheikhahmadi, M. Jalili, M. S. K. Fasaei. F inding influential nodes in social networks based on neighbo rhood Know.-B. Sys. 194 (2020) 10580. [27] M. Kivel a, A. Arenas, M. Barthelemy, J. P. Gleeson, Y . M oreno, M. A. Porter, Multilayer networks, Journal of Comple x Networks 2 (3) (2014) 203-271. [28] M. Szell, R. Lambiotte, S. Thurner, Multirelational or ganization of large-scale social networks in an online worl d, Proc. Nat. Acad. Sci, USA 107 (31) (2010) S. V . Buldyrev, R. Parshani, G. Paul, H. E. Stanley, S. Ha vlin, Catastrophic cascade of failures in interdependent n etworks, Nature 464 (7291) (2010) 1025-1028. [30] J. Gao, S. V . Buldyrev, H. E. Stanley, S. Havlin, Network s formed from interdependent networks, Nat. Phys. 8 (1) (201 2) 40-48. [31] C. Granell, S. Gomez, A. Arenas, Dynamical interplay be tween awareness and epidemic spreading in multiplex networ ks, Phys. Rev. Lett. 111 (12) (2013) 128701. [32] M. Perc, J. G' omez-Garde~ nes, A. Szolnoki, L. M. Flor' i a, Y . Moreno, Evolutionary dynamics of group interactions o n structured populations: a review, Jour. Roy. Soc. Interf. 10 (80) (2013) 20120997. [33] F. Morone, K. Roth, B. Min, H. E. Stanley, H. A. Makse, Mod el of brain activation predicts the neural collective influe nce map of the brain, Proc. Nat. Acad. Sci, USA 114 (15) (2017) 3849-3854. [34] S. Poledna, J. L. Molina-Borboa, S. Mart' inez-Jaramil lo, M. Van Der Leij, S. Thurner, The multi-layer network natu re of systemic risk and its implications for the costs of financial crises, J. Fina. Stab . 20 (2015) 70-81. 14

2108.11043

Ying Liu

Qi Zeng, Ying Liu, Ming Tang, Jie Gong

Identifying super-spreaders in coevolving dynamics on multiplex networks

physics.soc-ph

http://creativecommons.org/licenses/by/4.0/

Identifying super-spreaders in epidemics is important to suppress the spreading of disease especially when the medical resource is limited.In the modern society, the information on epidemics transmits swiftly through channels which contributes much to the suppression of epidemics. Here we study on the identification of super-spreaders in coupled spreading dynamics. Firstly, we find that the centralities in physical contact layer are no longer effective to in epidemics, which is due to the suppression effects from the information spreading. Then by considering the structural and between the communication layer and physical contact layer, we propose a centrality measure called centrality to in disease spreading. Simulation results on synthesized and real-world multiplex networks show that the proposed measure is not only much more accurate than centralities on the single network, but also outperforms two typical multilayer centralities in identifying super-spreaders. These findings imply that considering the structural and dynamical couplings between layers is very necessary in identifying the key roles in the coupled multilayer systems.

[2] or a UAV becoming damaged in an urban environment [3], [4] our goal is to the unknown system's set of reachable states [5], [6] while assuming about the system dynamics. We call such a set the guaranteed reachable set (GRS) . The primary contribution of this paper is to provide a meaningful the GRS of a control-ane system. We assume the only available information at the time of computation consists of (i) local dynamics at a single point, which can be obtained with an arbitrarily small error from applying test control inputs over a short period of time [7], and (ii) Lipschitz bounds on the rate of change of the system's dynamics provided by prior knowledge of the system design and physical laws. The reachable set of an unknown system is impossible to compute. Prior work determining the reachable set often focused on [7],[8]. Without discussing reachable sets, the work in [7], operating under similar assumptions as our paper, focused on optimistic reachability , i.e., attempting to reach a particular objective while there exists any chance of reaching it. Conversely, this paper computes states that are guaranteed to be reachable using admissible control signals. Apart from [7], work on reachability under uncertainty considered computation of reach- able sets with dynamics generated by a nite number of uncertain parameters [9], [10] or having bounded disturbances [11], [12]. Work in adaptive and robust control [8], [13], [14] assumes more knowledge on the magnitude or structure of the system dynamics, and clas- sical data-driven learning methods [2], [15], [16] collect data through repeated system runs. In contrast, this paper contains substantially fewer assumptions and focuses on deriving as much information as possible for the GRS based on one system run. Our approach relies on the interpretation of a control system as a dierential inclusion [17], [18], [19], [20] whose right hand side equals the set of velocities that the system can achieve at every state in the state space. While, for an unknown system, exact velocity sets may not be available anywhere, we can determine the family of all velocity sets that are consistent with prior knowledge of the local dynamics at a single point and Lipschitz bounds on the rate of change of the system's dynamics. The intersection of all elements of such a family is dened as the guaranteed velocity set (GVS) . Such a set is dicult to express in closed form, however we can analytically derive its to compute an of the GRS. The outline of the paper is as follows: we discuss the problem statement in greater detail in Section II. We then proceed to derive two simply expressible sets in Section III, with one set being a ball, and the other being a more complex convex set, later utilized to derive a polygon such that both the ball and polygon are contained within the GVS. Next, in Section IV, we use these sets to derive two classes of simple control-ane systems whose reachable sets are contained in the GRS. Lastly, we present numerical examples in Section V with a brief discussion on the implementation of our method. I-A Notation The set of all matrices with nrows andmcolumns is denoted by Rnm. For any vector v, kvkdenotes its Euclidean norm and kvk1denotes its 1-norm. For any matrix M,MTdenotes its transpose and kMkdenotes its Euclidean norm: kMk= maxkvk= 1kMvk. wherei(M) represents the i-th singular value of M. We also let Mydenote 2the Moore-Penrose pseudoinverse, Im( M) denote the image (range space) of M, and Ker(M) denote the kernel (null space) of M. For matrices MandN, we will say M2Imm(N) if M=NP for some matrix P. Notation Bn(a;b) denotes a closed ball in Rncentered at a2Rnwith radius b0 under the Euclidean norm. Set CL(Rn;Rm) denotes the set of a Lipschitz constant L, i.e., the set of all functions fthat allx,y2Rn. and XRmdenotes the set a+BX=fa+Bxjx2Xg . II Problem the paper, we attempt to meaningfully underapproximate the reachable set of a nonlinear control-ane system M(f;G) dened by _x(t) =f(x(t)) +G(x(t))u(t); x (0) =x0; (1) where allt0,x(t)2Rn, admissible inputs u(t)2U =Bm(0; 1), which is a common setting in reachability analysis [21], [22], and functions f:Rn!Rn, andG:Rn!Rnm are globally Lipschitz continuous with Lipschitz constants Lf0 andLG0, Cases where Lf= 0 orLG= 0 are simple, thus for the remainder of the paper we assume Lf>0 andLG>0. Noting that any function is globally Lipschitz continuous on any compact set, the theory developed in this paper can also be applied to a system whose states are guaranteed to be bounded, which is naturally true for a large class of systems [3], [4]. Without loss of generality, we assumex0= 0. II-A Assumptions and Technical Requirements In order to approximate the reachable set of (1), the technical work of [1] requires full actuation at x0= 0, i.e.,m=nwithG(0) being full rank. We relax these requirements, so the systemM(f;G) need not be fully actuated at x0= 0, that is, mdoes not Instead of assuming full actuation, we use the following 1 : Functions fandGare of the form f(x) =Rr(x) andG(x) andH(x)2Rmmsuch thatH(0) is invertible. The case of full actuation in [1] corresponds the case where R=Inin Assumption 1. Motivated by the online learning technique introduced in [7], we make the about the knowledge regarding the system 2 : BoundsLfandLGare known, as well as values f(0) andG(0) such that G(0)6= 0. Note we are not assuming any knowledge about matrix R; we only assume such an R exists. We only consider the case where G(0)6= 0 because it is otherwise impossible to determine anything about guaranteed velocities at states x6= 0. It is easily shown that Im( G(0)) = Im(R) under the conditions of Assumption 1. For our future results, it is important to determine the set of xsuch that Im( G(0)) = we show that the images of G(x) andG(0) are equal in some neighborhood of 0. Lemma 1 : Under Assumptions 1 and 2, if then Im(G(x)) = Weyl's inequality for singular values dictates that singular values as functions on matrices are uniformly Lipschitz with respect to the operator norm [23]: such that 1srwithr= rank(G(0)). theorem [24] along with the singular value decomposition of G(0) show thatr(G(0)) =kG(0)yk1is the smallest non-zero singular value of G(0). Forxthat satises we thus rank( dened G(x) =RH(x), so Im(G(x))Im(R). Since Im( G(0)) = Im( R), Knowing rank( we conclude Im( G(x)) = Im(G(0)). II-B Guaranteed Reachable Set Let us denote a set as the set of all pairs fand Gconsistent with Assumptions 1 and 2. We want to underapproximate the set of reach- able states given solely the knowledge of Dcon. We rst dene the (forward) reachable set R^f;^G(T;x 0) denotes the con- trolled trajectory of the system M(^f;^G) with control signal = 0. LetT0. We describe the guaranteed reachable set (GRS) as: RG(T;x 0) 0): (2) The GRS describes the set of all states that are reachable by any system consistent with our knowledge of the system 1: Determine or underapproximate the GRS. To solve Problem 1, we rst represent ordinary dierential equations with control inputs as an ordinary dierential inclusion (ODI) . We discuss doing so in Section III. Given the assumed knowledge of the system dynamics, we develop s to the right hand side of this inclusion. In Section IV, we use these s to derive two control-ane systems whose reachable sets are subsets of RG(T;0). III Guaranteed Velocities We follow the classical approach of interpreting ordinary dierential equations with control inputs as inclusions [17], [18], [19], [20]. In this section, we formally dene the set of an unknown control system and determine analytically computable of such a set. III-A Guaranteed Velocity Set We dene the available velocity set of the systemM(f;G) at statexbyVx=f(x) +G(x)U, and introduce the following ODI: 4_x2Vx=f(x) +G(x)U; x (0) =x0: (3) If a trajectory (
;x0) satises (3), then it obviously serves as a solution to the control system (1) for an admissible control input, and vice versa. Given Assumption 2, set Vx0=V0 is known. The goal of this section is to provide an for set Vxusing sets DconandV0. We rst dene the guaranteed velocity set (GVS) +^G(x)UVx: (4) The GVSVG xis the set of all velocities that can be taken by all systems consistent with the assumed knowledge of the dynamics. Let us consider the following ODI: _x2VG x; x 0= 0: we will consider by convention that the trajectory of (5) ceases to exist at timeT. The following proposition then holds directly from (2) and (4). Proposition 1: LetT0. If a trajectory : (5) at all times tT, then(T)2RG(T;x 0). Proposition 1 implies that the reachable set of (5) is a subset of RG(T;x 0). As brie y described in [1], these sets are not necessarily equal; establishing conditions for the equality of the reachable set of (5) and RG(T;x 0) is an open problem for future work. III-B Ball 1 motivates us to underapproximate RG(T;x 0) by determining the reachable set of (5). We start by examining the geometry of the ODI. Given x2Rn, our show thatU=Bm(0; 1),VG xis an intersection of innitely many ellipsoids An intersection of innitely many ellipsoids is generally not a geometrically simple object [25]. Thus, we will determine an of VG x. Our approach will be to implicitly ofVG xand underapproximate the distance of the boundary of set VG xfrom 0 in every direction. In Theorem 1, we calculate one such . Theorem 1 : LetU,Lf, andLGbe dened as above. Let x2Rnsatisfy (7) Then, On the other hand, where ~Dconis dened the same as before, just with the assumption that f(0) = 0. Thus, we can assume without loss of generality that f(0) = 0. Letd2Im(G(0)) = Im(R) such thatkdk= 1. We will prove that if then +^G(x)u; (8) where ( ^f;^G)2Dcon, admits a solution u2U=Bm(0; 1). We subtract ^f(x) from both sides of (8). Since ^f(x)2Im(R) by Assumption 1 and kd2Im(R) by denition, then kd^f(x)2Im(R). Also, Im( ^G(x)) = Im(R) by Lemma 1. Hence, there exists a vector  u2Rmsuch thatkd^f(x) =^G(x)u. Now, through the rank- nullity theorem [24], we can write  andu22Ker( ^G(x)). Thus, =^G(x)u; hence,kd^f(x) =^G(x)u. We multiply both sides of kd^f(x) =^G(x)uon the left by ^G(x)y, resulting in = ^G(x)y^G(x)u. The term in the projection of uonto the Im( ^G(x)T) [24]. Given that u2Im(^G(x)T), by denition of a projection, ^G(x)y(kd^f(x)) = ^G(x)y^G(x)u= u. Thus, if we prove (9) we will havekuk1, i.e.,u2U. Utilizingkdk= 1 along with the product and for matrices, we arrive at (10) and (11) From (11) it follows that the set of all kthat is a subset of all kthat satisfy (9). In other words, if: jkjk (9). We note k^G(x)yk6= 0 from the denition of the Moore-Penrose pseu- doinverse and because G(x)6= 0 from Lemma 1. By Weyl's inequality for singular values [23] and Assumption 2, we obtain the Thus, since we (13) it satises (8). We slightly generalize Theorem 1 by also considering all xwhich satisfy ( 1 : Let us take xsuch a sequence x1;x2;:::such . Theorem 1 shows that for all i, there exists a = ^f(xi) +^G(xi)ui. Thus, becauseUis a compact set, there exists a converges to some u2U [26]. Since xpi!xandupi!u, we have f(0) = ^f(xpi) +^G(x)u. Therefore, f(0) = ^f(x) +^G(x)u. III-C Advanced Convex allxthat we now have a set VG xguaranteed to be a subset of VG x. Such a set is a projection of a ball onto Im( G(0)). However, there are instances, singular values of G(0) are far apart, where a ball may be a poor VG x. Consequently, we derive a new setVG xin Theorem 2. Theorem 2 : LetU,Lf, andLGbe dened as above. Let = 1 if rank( G(0)) =m=n, =p 2 if rank(G(0)) = min( m;n) and letxsatisfy If VG x=ff(0) +kdjkdk= As in Theorem 1, we will show that for kanddgiven in (14), equation (8) admits a solutionu2U. Like in the proof of Theorem 1, without loss of generality, we set f(0) = 0. From inequality (10), it follows that the set of all kthat (15) is a subset of all kthat satisfy (9). The term kG(x)ykis the inverse of the smallest nonzero singular value of G(x). Therefore, using Weyl's inequality for singular values [23], we By Assumption 2, we also have k^f(x)kLfkxk. We conclude that the set of all kthat (16) is a subset of all kthat satisfy (9). Next, we the product and triangle inequalities for Lemma 1, Im( G(0)) = Im( ^G(x)). Therefore, we can apply the inequality 3.3 in [27]). We can now rewrite the upper bound (18) We again usek^G(x)yk According to Weyl's inequality for singular Therefore, allk^G(x)ydkthat satisfy (18) will (19) By plugging (19) into (16), we obtain all kthat +kG(0)ykLGkxk; (20) also satisfy (9). By rewriting the denominator of K(d) in Theorem 2 as we can easily see that it cannot be negative since 1 We now compare the derived sets VG xandVG xin Corollary 2. Corollary 2 : For invertible matrices Given the numerator of K(d) in (14) is identical to the radius of VG xderived in Theorem 1, VG xis contained withinVG xif and only (21) Inequality (21) is obviously equivalent (22) Dividing results (23) We see in the case of invertible G(0) where= 1, (23) reduces to holds true for all d such that kdk= 1. Therefore, for invertible G(0),VG xis 1: Velocity sets plotted for a system with f(0) =0 0 ,G(0) =10 3 2 7 ,kxk= 1, and Lf=LG= 1. The blue curves represent the boundaries of the available velocity sets for all systems such that ( ^f;^G)2Dcon. The intersection of these sets produces the setVG x(white). VG x(bounded in red) in black) are both contained in VG x. Our calculations result in two derived sets. Corollary 2 proves is invertible. Hence, Theorem 2 provides a better than the one obtained in [1], under the assumptions present in that paper. Figure 1 illustrates a simple example of the GVSVG xand its approximations obtained by Theorems 1 and 2. Consistent with the results of Corollary 2, since G(0) is invertible, general cases, VG x6VG x. Thus, we can take the union of both of sets to determine a larger set of guaranteed velocities. We denote such a set Section IV, we show how ^VG xcan be used to identify a polygon that can generate a control system with solutions that satisfy (1). IV Reachable Set In this section, we aim to utilize the derived sets VG xand^VG xto determine a set of to satisfy (1). First, we use ^VG xto identify a polygon contained within VG x. Lemma 2 : LetS(x) be any nite set of points on the boundary of ^VG x. LetP(S(x)) be a convex hull ofS(x). Every solution to _ x2P(S(x)) is a solution to (1). Proof. The guaranteed velocity set is the intersection of an innitely many ellipsoids; such a set is convex [25]. Given ^VG xVG x, we Lemma 2 permits us to chose any set of points on the boundary of ^VG xasS(x), in the remainder of the paper, we will choose the points along the singular vectors of G(0). We do so by noting from inequality (20) that the smaller the magnitude of kG(0)ydk, the 9larger the corresponding K(d). The smallest magnitude of kG(0)ydkforkdk= 1 will be obtained when dis the singular vector corresponding to the largest singular value of G(0) [24]. We choose other points to be along other singular vectors because of the orthogonality of singular vectors. Theorem 1 shows that the reachable set of _x2VG x; x (0) =x0; (24) is a subset ofRG x(T;x 0), while Lemma 2 shows the same for the reachable set of _x2P(S(x)); x (0) =x0: (25) IV-A Control System { Ball Analogous to the interpretation of dynamics (1) as an ODI (3), inclusion (24) can be inter- preted as a control x (0) =x0; 1)Im(G(0)T), and where g(s) We thus obtain the 3 : Let R(T;x 0) be dened as the reachable set of (26) at time T. Then, R(T;x 0)RG(T;x 0). Proof. Proposition 1 and Theorem 1 show that R(T;x 0)RG(T;x 0). We can expand on Theorem 3 using control system (26) to help determine the of R(T;x 0). Corollary 3 : Consider a control system of the form _ x= (bckxk)udened on some ball BRn, such that x2Rn,b;c2R, andu2U. Then, the reachable set R(T;x 0) is a ball inB. Proof. Letz=Rxsuch thatR2Rnnis any orthonormal matrix. Obviously, _ Note thus we have _ z= (b+ckzk)Ru. Therefore, _ z= (bckzk)vsuch thatv2U, so the reachable set of _ x= (bckxk)uis invariant to all rotations. Any rotation of any point on any trajectory of the original system is thus itself on some other trajectory of the original system. Hence, as the trajectories are continuous, the reachable set of the original system is a ball. IV-B Control System { to Theorem 3, we can dene ^R(T;x 0) as the reachable set of (25) at time T, and again ^R(T;x 0)RG(T;x 0). We follow the method above and interpret inclusion (25) as a control system dened in Theorem 4. 10Theorem 4 : Lets2Randg(s) dened in Theorem 2. Let UVTbe the singular value decomposition of G(0) whereU= [1;:::;n]. Letr= rank(G(0)); we dene ( s) = diag(i(s)) such that i(s) = 1;:::;r andi(s) = 0 elsewhere. The reachable set of (25) equals the reachable set of the control x (0) =x0; witha=f(0) If ^R(T;x 0) denotes the reachable set of (27), then ^R(T;x 0)RG(T;x 0). Proof. By Theorem 1, the boundary of VG xlies at a mag- nitude ofg(kxk) along any direction in Im( G(0)) fromf(0). By Theorem 2, the at a magnitude of along the direction ifrom f(0), for allir. Since the vertex ofS(x) in the direction ithus has a mag- nitude ofi(kxk), forir. In the direction of ifori>r , the distance of boundary points for both equals 0, as both of these sets are contained in Im(G(0)). By its construction in Lemma 2, polygon P(S(x)) is thus given by f(0) +U(kxk)Q, note that the dynamics of (26) and (27) are entirely known . Finding R(T;x 0) and ^R(T;x 0) becomes a standard problem of determining the reachable set of a nonlinear control system [28]. In order to exploit previous work on computing reachable sets in Rn, we can follow methods outlined in [1], where (24) and (25) can be continuously extended to all Rn by dening =ff(0)gfor allxsuch Although there is no method to analytically determine the exact reachable set for all nonlinear level set methods nd the reachable set by determining the viscosity solution equations [11], [29]. Additional methods create an of the true reachable set by utilizing trajectory piecewise linearized models [30] or set [31], [32], [33]. For simplicity, in our numerical examples, we approximate the true reachable set using a Monte Carlo method by solving ODEs (26) and (27) with inputs. V Numerical Examples We consider two examples. The rst shows novel theory applied to an academic nonlinear system with initial conditions similar to the system brie y discussed in Section III, shown in Figure 1. This illustrates an example where Theorem 1 results in a poorer of the true reachable set, while utilizing ^R(T;x 0) generates a signicantly better . The second example is a control system with decoupled quadrocoptor dynamics motivated by the scenario of landing a damaged UAV safely [3], [4]. The goal is to determine a reachable set of pitch and roll velocities in order to help stabilize the UAV for landing. 11As mentioned at the end of Section IV, we calculate the reachable sets by all ordinary dierential equations using the standard ode45 function in MATLAB which implements a Runge-Kutta method [34] with a variable time step for ecient com- putation. Doing so avoids limitations of current numerical solvers such as CORA [35] that often rely on set propagation methods to calculate the reachable set and face issues as kxk is not dierentiable at 0. Numerical results are supported through analytical means and theoretical results derived in Sections II, III, and IV. V-A Nonlinear System Figure 2: True reachable set (blue) with the s R(T;0) (red) and ^R(T;0) (green) numerically calculated consider a system with dynamics _x(t) =f(x(t)) +G(x(t))u(t) =2 410u1+ 7u2 (2:5 +x3)u33 5; =2 40 0 03 5; G (x(t)) =2 410 3x2 0 2x1 7 0 0 0 2 :5 +x33 5; with the primary interest of nding the reachable set of x1andx2. Taking the Jacobian of f(x) andG(x) yieldsLf=LG= 1 to be acceptable Lipschitz constants. For simplicity, we setx0= 0. We remind the reader that fandGare assumed to be unknown, and only Lf andLGare known. Results from Figure 2, showing the projection of the reachable sets to the rst two coordinates, illustrate that (i) R(T;0) and ^R(T;0) are indeed s of the true reachable set, (ii) ^R(T;0) produces a better than R(T;0), and (iii) the accuracy of the s increases as T!0. Phenomenon (i) validates the results from Theorems 3 and 4. The singular value decomposition of G(0) and3= 2:5. Because of the large dierence between 1and3, the true reachable set is not accurately represented by the ball R(T;0). On the other hand, in Theorem 4 we 12derive control system (27) using ^VG x, which more accurately represents the complex shape of the GVS. Thus, the resulting set ^R(T;0) is larger along singular vectors pertaining to larger singular values, contributing to result (ii). Lastly, Figure 2 illustrates how the s R(T;0) and ^R(T;0) perfect as T!0. AsT!1 , although comparatively worse, these yield progressively larger reachable sets. In the next example, we apply this novel theory to decoupled quadrocopter dynamics to help safely land a damaged UAV. In this case, we will show that, in contrast to the above example, the polygonal by Theorem 4 produces worse results than the ball approximation derived from Theorem 3. V-B Decoupled Quadrocopter Dymamics Figure 3: True reachable set (blue) with the s R(T;0) (red) and ^R(T;0) (green) numerically calculated show that the novel theory can be applied to a real system by example of We consider the objective of adjusting the quadrocopter's pitch and roll velocities. For a safe landing, a UAV ideally needs these velocities to equal 0 [36]. Given the of a standard UAV, the yaw rate is inconsequential due to the symmetrical shape of the quadrocopter. We consider the scenario where a UAV collides with an obstacle, which would result in unwanted high velocity rotations. This scenario translates to a problem of reachability: given initial conditions of roll and pitch rates, we aim to determine if it is possible to reach p=q= 0 without knowing the system's dynamics after the collision. The dynamics of a standard UAV are modeled in [37]. The model is comprised of a solid sphere with mass M= 1kg and radius R= 0:1m, which represents the central frame; it is connected to four point masses m= 0:1kg, each representing one of four propellers at an equidistant length of l= 0:5m away from the central sphere. The dynamics are shown 3 5; (29) where states p,q, andrcorrespond to velocities pertaining to the roll rate, pitch rate, and yaw rate, respectively, and 13Jx=2MR2 5+ 2l2m; Jy=Jx; Jz=2MR2 5+ and pertain to the applied torque that directly aects the roll, pitch, and yaw velocities. As Jx=Jy, the yaw rate can directly be changed by increasing the control action to without aecting the roll or pitch rates. Including additional states, such as the roll, pitch, and yaw angles and translational position, would move the model beyond the requirements of Assumption 1. Applying the theory to a dynamic structure such as this is a subject for future work. We note the system dynamics in (29) are not globally Lipschitz continuous. However, as previously discussed, given the yaw rate can be directly changed by increasing control action to without aecting other states, we can trivially reduce _ rto 0, causing rto be some constant; we arbitrarily set r0==2. Let the initial conditions after collision be p0= 15, q0= 10 radians per second. Since novel theoretical results are derived under the assumption x0= 0, we perform a simple coordinate transformation; let  p=p15 and q=q10 such that p0= q0= 0. Obviously, _p= _pand _q= _q, thus resulting in the new (30) with initial conditions f(0) G (0) thatJx=Jy= 0:009, The new dynamics (30) are we overapproximate the Lipschitz bounds to be Lf= 1 andLG= 1 to account for overly conservative knowledge about the rate of change of the dynamics. As in the previous sections, we bound inputs u2B2(0; 1) and denote x= system states. The blue shape in Figure 3 illustrates the true reachable set of the to the roll and pitch velocities, in ( p;q) coordinates, plotted at Given the structure of control systems (26) and (27), it is trivial to see that if point (p0;q0) is contained within either R(T;0) or ^R(T;0), then we know there exists a control uwhich guarantees velocities pandqcan be reduced to 0, regardless of the true we apply novel theory to solve for R(T;0), resulting in the set shown in red in Figure 3, which is proven in Sections III and IV to be a guaranteed of the GRS. We begin by calculating =B2(f(0); system (26) thus 0 0 111 :112kxk u (31) such thatu2B2(0; 1). Its reachable set is R(T;0). Lastly, we consider control system (27) to determine ^R(T;0). SinceG(0) is invertible, we know from Corollary 2 that VG xVG x. Thus, when determining the diagonal entries of 14(kxk), we know from Theorem 4 that i(kxk) SinceG(0) is diagonal, we know the matrix of left singular vectors Uequals identity. System (27) therefore thatkuk11. The same methods for determining the previous two reachable sets numerically can be applied to nd ^R(T;0). Figure 3 displays ^R(T;0) for the roll and pitch velocities  pand q. Notice thatkG(0)yikis identical for all i. Thus,i(kxk) = 111:112kxkfor alli, which is identical to the diagonal terms in control system (31). Since system (27) considers kuk11 and system (26) considers u2B2(0; 1), for this particular control system, nal result is a reachable set ^R(T;0) denoted in green in Figure 3. It naturally resembles a polyhedron as T!0, with edges that curve as time to the numerical results in Figure 3, roll and pitch velocities with initial condi- tionsp0= 15,q0= 10 radians per second can be reduced to 0 in no more than 0 :25 seconds because (15;10) is contained in R(0:25;0). In fact, for any points that lie within R(T;0), we can guarantee there exists a control signal which reaches these states within time T. For timeT= 0:05, we see that neither the true reachable set, nor the s reach (15;10), with R(T;0) becoming asymptotically perfect as T!0. Conversely, at timeT= 1, the theory provides larger, generally worse s where the true reachable set, along with R(T;0) and ^R(T;0), clearly include ( 15;10). Unlike in the rst numerical example, set R(T;0) is larger than ^R(T;0). Naturally, in general, computing R(T;0)[^R(T;0) provides the best approximation of RG(T;0) available from our theory. VI Conclusion and Future Work This paper provides a novel approach to the reachable set of a sys- tem with unknown dynamics. By assuming the nonlinear control-ane system structure and exploiting solely the knowledge of system dynamics at a single point and | | Lipschitz bounds on the rate of change, we are able to determine two R(T;x 0) and ^R(T;x 0) that are guaranteed to be contained within the guaranteed reachable set RG(T;x 0). Both s rely on an intermediate ap- proximation of the GRS by an ODI _ x2VG x, where the right hand side is a set of for the unknown nonlinear control-ane system. The two s dier by the shape of the right hand side set, i.e., the approximation of VG xis either de- termined by balls VG xor a dierent shapeVG xthat more closely resembles the shape of an intersection of an innite many ellipsoids. A natural area of future work is to focus on creating a larger set the GVS. One possibility is to consider approximations in norms other than the spectral norm considered in this paper. For example, potentially by bounding the perturbations of the unknown system's dynamics with the Frobenius norm instead of the spectral norm, we could utilize the Mirsky Inequality [23] to produce new s of the GVS. However, there is currently no guarantee that such a bound would produce a more 15accurate of the GVS. Another possibility includes the utilization of semi-innite programming [38, 39] to determine the maximal extent of VG xin every rst- and second-order optimality conditions, the problem could potentially be reduced locally to one with nitely many constraints; similarly, given the convex structure of the GRS, additional optimization techniques such as duality may help further simplify the problem. Another approach to obtaining a larger of the GVS is to increase the knowledge of the system dynamics. In other words, making additional assumptions on the structure of the dynamics could also help determine a larger of the GVS by reducing the size of the set of systems consistent with prior knowledge about system dynamics. One option is to utilize knowledge from multiple system runs instead of solely dynamics at a single point. Expanding on theoretical results derived from this paper by incorporated additional knowledge consistent with a large class of systems could in substantial progress for the development of sophisticated safety critical M. Ornik, Guaranteed reachability for systems with unknown dynamics, in 59th IEEE Conference on Decision and Control , 2020, pp. 2756{2761. [2] N. Nguyen, K. Krishnakumar, J. Kaneshige, and P. Nespeca, Flight dynamics and hybrid adaptive control of damaged aircraft, Journal of Guidance, Control, and Dy- namics , vol. 31, no. 3, pp. 751{764, 2008. [3] G. Chowdhary, E. N. Johnson, R. Chandramohan, M. S. Kimbrell, and A. Calise, Guid- ance and control of airplanes under actuator failures and severe structural damage, Journal of Guidance, Control, and Dynamics , vol. 36, no. 4, pp. 1093{1104, 2013. [4] D. Jourdan, M. Piedmonte, V. Gavrilets, D. Vos, and J. McCormick, Enhancing through damage tolerant control, in AIAA Guidance, Navigation, and Control Conference , 2010, pp. 7548{7569. [5] R. W. Brockett, onlinear systems and dierential geometry, Proceedings of the IEEE , vol. 64, no. 1, pp. 61{72, 1976. [6] A. Isidori, Nonlinear Control Systems . Springer Science & Business Media, 2013. [7] M. Ornik, S. Carr, A. Israel, and U. Topcu, learning on the y, on Automatic Control , vol. 65, no. 11, pp. 4800{4807, 2019. [8] I. M. Mitchell and C. J. Tomlin, reachable sets by Journal of Scientic Computing , vol. 19, no. 1, pp. 323{346, 2003. [9] T. F. Filippova, Estimates of reachable sets of a nonlinear dynamical system with impulsive vector control and uncertainty, in 14th International Conference: Stability and Oscillations of Nonlinear Control Systems . IEEE, 2018, pp. 1{4. 16[10] M. Rungger and M. Zamani, ccurate reachability analysis of uncertain nonlinear sys- tems, in 21st International Conference on Hybrid Systems: Computation and Control , 2018, pp. 61{70. [11] I. M. Mitchell, A. M. Bayen, and C. J. Tomlin, time-dependent of reachable sets for continuous dynamic games, IEEE Transactions on Automatic Control , vol. 50, no. 7, pp. 947{957, 2005. [12] B. Zhang, J. Lam, and S. Xu, Reachable set estimation and controller design for distributed delay systems with bounded disturbances, Journal of the Franklin Institute , vol. 351, no. 6, pp. 3068{3088, 2014. [13] G. E. Dullerud and F. Paganini, A Course in Robust Control Theory: A Convex Ap- proach . Springer Science & Business Media, 2013. [14] P. A. Ioannou and J. Sun, Robust Adaptive Control . Courier Corporation, 2012. [15] S. L. Brunton, J. L. Proctor, and J. N. Kutz, Discovering governing equations from data by sparse identication of nonlinear dynamical systems, Proceedings of the National Academy of Sciences , vol. 113, no. 15, pp. 3932{3937, 2016. [16] Y. Chen, H. Peng, J. Grizzle, and N. Ozay, Data-driven computation of minimal robust control invariant set, in IEEE Conference on Decision and Control . IEEE, 2018, J.-P. Aubin and A. Cellina, Dierential Inclusions: Set-Valued Maps and Viability The- ory. Springer Science & Business Media, 2012. [18] A. Bressan and B. Piccoli, Introduction to the Mathematical Theory of Control , 2007. [19] G. V. Smirnov, Introduction to the Theory of Dierential Inclusions . American Math- ematical Society, 2002. [20] A. B. Kurzhanski and P. Varaiya, Ellipsoidal techniques for reachability analysis, in International Workshop on Hybrid Systems: Computation and Control . Springer, 2000, pp. 202{214. [21] M. Margaliot, the reachable set of nonlinear control systems with a nilpotent lie algebra, in European Control Conference . IEEE, 2007, pp. 4261{4267. [22] R. Vinter, characterization of the reachable set for nonlinear control systems, SIAM Journal on Control and Optimization , vol. 18, no. 6, pp. 599{610, 1980. [23] G. W. Stewart, Perturbation theory for the singular value decomposition, Tech. Rep., 1998. [24] G. Strang, Introduction to Linear Algebra . Press, 2016. [25] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex Optimization . Cambridge 2004. 17

2108.11045

Taha Shafa

Taha Shafa and Melkior Ornik

Reachability of Nonlinear Systems with Unknown Dynamics

math.OC math.DS

http://creativecommons.org/licenses/by/4.0/

Determining the reachable set for a given nonlinear control system is crucial for system control and planning. However, computing such a set is impossible if the system's dynamics are not fully known. This paper is motivated by a scenario where a system suffers an adverse event mid-operation, resulting in a substantial change to the system's dynamics, rendering them largely unknown. Our objective is to conservatively approximate the system's reachable set solely from its local dynamics at a single point and the bounds on the rate of change of its dynamics. We translate this knowledge about the system dynamics into an ordinary differential inclusion. We then derive a of the velocities available to the system at every system state. An inclusion using this approximation can be interpreted as a control system; the trajectories of the derived control system are guaranteed to be trajectories of the unknown system. To illustrate the practical consequences of our work, we apply our algorithm to a simplified model of an unmanned aerial vehicle.

2 which provides a system of first -order hyperbolic Partial Differential Equations (PDEs) . Many papers study stabilization problems for the linearization of this system around an equilibrium point, employing either the backstepping methodology (see [ 11,12 ]) or the Control Lyapuno v Functional (CLF) methodology (providing local stabilization results for the original nonlinear system in many cases; see [2,3,4,8,9,10, 21]). The present work studies the problem of stabilization of the motion of a tank that contains an incompressible, Newtonian liquid. The problem is not new and has also been studied in [ 24] and we consider stabilization of both the state of the liquid (liquid level and velocity) and the state of the tank (tank position and velocity) by manipulating the acceleration of the tank. However, in contrast with [ 24], we do not consider the inviscid Sa int-Venant model but the viscous Saint -Venant model (i.e., the model that ignores friction forces and surface tension and takes instead into account gravity forces and viscous stresses). From a ma thematical point of view this makes a huge difference: the resulting nonlinear model consists of two Ordinary Differential Equations (ODEs) and two PDEs which are no longer first -order hyperbolic PDEs but a system of one hyperbolic first -order PDE and one parabolic PDE. From a physical point of view the replacement of the inviscid Saint -Venant model with friction by the viscous, frictionless Saint -Venant model is complete ly justified: viscous stresses are negligible compared to friction forces when the liquid velocity is relatively large (e.g. rivers) but in tanks w here the liquid velocity is small, viscous stresses become dominant. T he viscosity term is nonlinear and is similar to the viscosity term appearing in compressible fluid flow models (see [ 18,22,26,27 ]). In addition , we assume that the tank walls have a spe cific height and we require that the movement be spill-free and slosh -free (in the sense that the liquid is transferred without residual end point sloshing; see [13,32] for the descri ption of the slosh -free motion problem) . In other words, we require that the liquid level in the tank does not exceed the height of the tank walls. To our knowledge this is the first paper that studies the spill -free movement problem: all other works (see [24,25]) do not take into account the limited height of the tank walls. We exploit the similarity with compressible fluid models in order to apply the CLF methodology for the viscous Saint -Venant system. The CLF methodology was first used for global stabilization of nonlinear parabolic PDEs in [19] and wa s subsequently st udied in [16,17,18] (but see also [8 ] for the presentation of the CLF methodology in finite -dimensional and infinite -dimensional systems). Recently, it was used in [ 18] for the stabilization of a nonlinear Navier -Stokes model of compressible fluid. As in [ 18], the CLF plays also the role of a barrier function and guarantees a positive lower bound for the liquid level. However, here the CLF also plays the role of a barrier function for the additional requirement of spill -free movement by providing an upper b ound for the liquid level. The CLF is constructed by combining the mechanical energy of the system and the use of a specific transformation that has been used in the literature of co mpressible fluids (see [18,22,26,27 ]). It is important to note here that t he nonlinearity of the control problem hinders the application of standard feedback design methodologies like backstepping (see [28,29,31 ]). The CLF allows the determination of a specific set X of initial conditions for which spill-free movement of the tank is possible (by applying an appropriate input): this is the state space of our problem. We achieve semi -global stabilization of the liquid and the tank by means of a simple feedback law. The term semi -global here refers to the state space X : for every closed subset of the state space X we are in a position to find appropriate controller gains, so that every classical solution of the closed -loop system initiated from the gi ven closed subset satisfies specific stability estimates. More specifically, an exponential convergence rate is achieved for the closed -loop system. The stabilizing feedback laws do not require measurement of the liquid level and velocity spatial profiles in the tank and simply require measurements of : (i) the tank position error and tank velocity, (ii) the total momentum of the liquid, and (iii) the liquid levels at the tank walls (boundary). To our knowledge, this is the first paper in the literature that achieves stabilization of the nonlinear viscous Saint -Venant system. In other words, 1) the nonlinear viscous Saint -Venant model is studied (instead of its linearization) , 2) a simple feedback law (with minimum measurement requirements) is obtained that guaran tees exponent ial convergence to the desired equilibrium point , 3 3) a spill-free movement of the liquid in the tank is guaranteed. The structure of the paper is as follows. Section 2 is devoted to the presentation of the control problem. Section 3 conta ins the statements of the main result of the paper (Theorem 1) . Moreover, in Section 3 we also provide insights for the proof of the main result : we explain the construction of the CLF, the selection of the state space and we provide some auxiliary res ults that are used in the proof . Section 4 provides an algorithmic solution of the problem of the spill -free movement and slosh -free settle ment of water in a glass , by a robot , to a pre -specified glass position. We show that no matter how full the glass is, the robot can transfer the glass without spilling out water and without residual end point sloshing . The proofs of all results are provided in Section 5 . Finally, Section 6 gives the concluding remarks of the present work. Notation . Throughout this paper , we adopt the following notation. [0, ) denotes the set of non -negative real numbers. Let nS be an open set and let nA be a set that satisfies () S A cl S . By 0( ; )CA , we denote the class of continuous functions on A , which take values in m . By ( ; )kCA , where 1k is an integer, we denote the class of functions on nA , which takes values in m and has continuous derivatives of order k . In other words, the functions of class ( ; )kCA are the functions which have c ontinuous derivatives of order k in int( )SA that can be continued continuously to all points in SA . When then we write 0()CA or ()kCA . When I is an interval and 1() G C I is a function of a single variable, ()Gh denotes the derivative with respect to hI . Let I be an interval, let ab be given constants and let : [ , ]u I a b be a given function . We use the notation []ut to denote the profile at certain tI , i.e., ( [ ])( ) ( , )u t x u t x for all [ , ]x a b . When ( , )u t x is (twice) differentiable with respect to [ , ]x a b , we use the notation ( , )xu t x ( ( , )xxu t x ) for the (second) derivative of u with respect to [ , ]x a b , i.e., ( , ) ( , )xuu t x t xx ( 2 2( , ) ( , )xxuu t x t x x ). When ( , )u t x is differentiable with respect to t , we use the notation ( , )tu t x for the deriva tive of u with respect to t , i.e., ( , ) ( , )tuu t x t xt . Given a set nU , U denotes the characteristic function of U , i.e. the function defined by ( ) : 1Ux for all xU and ( ) : 0Ux for all xU . The sign function sgn : is the function defined by the relations sgn( ) 1x for 0x , sgn(0) 0 and sgn( ) 1x for 0x . Let ab be given constants. For [1, )p , ( , )pL a b is the set of equivalence classes of Lebesgue measurable functions : ( , )u a b with 1/ : ( )pb p p au u x dx . ( , )L a b is the set of equivalence classes of Lebesgue measurable functions : ( , )u a b with ( , ): sup ( ) x a bu ess u x . For an integer 1k , ( , )kH a b denotes t he Sobolev space of functions in 2( , )L a b with all its weak derivatives up to order 1k in 2( , )L a b . 4 2. Description of the Problem We consider a one -dimensional model for the m otion of a tank. The tank contains a viscous, Newtonian, incompressible liquid. The tank is subject to a force that can be manipulated. We assume that the liquid pressure is hydrostatic and consequently, the liquid is model ed by the one- dimensional (1 -D) viscous Saint -Venant equations , whereas the tank obey s Newton's second law. The control objective is to drive asymptotically the tank to a specifie d position without liquid spilling out and having both the tank and the liquid within the tank at rest. Figure 1 shows a picture of the problem. Figure 1: The control problem . Let the position of the left side of the tank at time 0t be ()at and let the length of the tank be 0L (a constant). The equations describing the motion of the liquid within the tank are ( ) 0tzH HV , for 0t , ( ), ( ) z a t a t L (1) 221()2tz z zHV HV gH HV , for 0t , ( ), ( ) z a t a t L (2) where ( , ) 0H t z , ( , )V t z are the liquid level and the liquid velocity, respectively, at time 0t and position ( ), ( ) z a t a t L , while ,0g (constants) are the acceleration of gravity and the kinematic viscosity of the liquid, respectively . The liquid velo cities at the walls of the tank coincide with the tank velocity, i.e., we have: ( , ( )) ( , ( ) ) ( )V t a t V t a t L w t , for 0t (3) where ( ) ( )w t a t is the velocity of the tank at time 0t . Moreover, applying Newton's second law for the tank we get 5 ( ) ( )a t f t , for 0t (4) where ()ft , the con trol input to the problem, is the force exerted on the tank at time 0t divided by the total mass of the tank. The condi tions for avoiding the liquid spilling out of the tank are: max max( , ( )) ( , ( ) )H t a t H H t a t L H (5) where max 0 H is the height of the tank walls. The variable ()ft is the control input. Applying the transformation ( , ) ( , ( ) ) ( ) ( , ) ( , ( ) ) ( ) ( )v t x V t a t x w t h t x H t a t x t a t a (6) where a is the specified position (a constant) where we want to have the left side o f the tank, we obtain the model ,w w f , for 0t (7) ( ) 0txh hv , for 0t , 0,xL (8) 221()2tx x xhv hv gh hv hf , for 0t , 0,xL (9) ( , 0) ( , ) 0v t v t L , for 0t (10) where the control input f appears additively in the second equation of (7) and multiplic atively in (9). Moreover, t he conditions (5) for avoiding the liquid spilling out of the tank become : max max ( ,0), ( , ) h t h t L H , for 0t --- Condition for no spilling out (11) We consider classical solutions fo r the PDE -ODE system (7) -(10), i.e., we consider functions 12(0, ) CC , 01(0, ) w C C , 01[0, ) [0, ] (0, ) [0, ] v C L C L , 12[0, ) [0, ];(0, ) (0, ) (0, ) h C L C L , with 2[ ] (0, )v t C L for each 0t that satisfy equations (7) -(10) for a given input 0fC . Using (8) and (10), we can prove that for every solution of (7) -(10) it holds that 0( , ) 0Ldh t x dxdt for all 0t . Therefore, the total mass of the liquid 0m is constant. Therefore, without loss of generality, we assume that every solution of (7) -(10) satisfies the equation 0( , )L h t x dx m (12) The open -loop sys tem (7) -(10), (12), i.e., system (7) -(10), (12) with ( ) 0ft , allows a continuum of equilibrium points, namely the points 6 ()h x h, ( ) 0vx , for 0,xL (13) , 0w (14) where / h m L . We assume that the equilibri um points satisfy the condition for no s pilling out (11), i.e., max hH . Our objective is to design a feedback law of the form ( ) [ ], [ ], ( ), ( )f t F h t v t t w t , for 0t , (15) which achieves stabilization of the equilibrium point with 0 . Moreover, we additionally require that the spill -free condition (11) holds for every 0t . The existence of a continuum of equilibrium points for the open -loop system given by (13), (14) means that the desired equilibrium point is not asymptotically stable for the open -loop system. Moreover, experience shows that there are smooth initial conditions 21( (0), (0), [0], [0]) [0, ];(0, ) [0, ] w h v C L C L with ( [0])(0) ( [0])( ) 0v v L for which it is not possible to av oid liquid spilling out of the tank -no matter what the applied input f is. Thus, the described control problem is far from trivial. 3. Construction of the Feedback Law 3.1. The Control Lyapunov Functional (CLF) Let ,0kq be position error and velocity gains (to be selected). Define the set 220[0, ] S C L : 01 0 0 2[0, ];(0, ) (0, ) [0, ] ( , , , ) () ( , ) , (0) ( ) 0Lh C L H L v C L w h v S h x dx m w v v L (16) We define the following functionals for all ( , , , )w h v S : 2 2 2( , , , ) : ( , ) ( , )22qk qV w h v W h v E h v w k (17) 22 0011( , ) : ( ) ( ) ( )22LL E h v h x v x dx g h x h dx (18) 2 2 1 0011( , ) : ( ) ( ) ( ) ( ) ( )22LL x W h v h x h x v x h x dx g h x h dx (19) where / h m L . We notice that: 7 the functional E is the mechanical energy of the liquid within the tank . Indeed, notice that E is the sum of the potential energy ( 2 01()2L g h x h dx ) and the kinetic energy ( 2 01( ) ( )2L h x v x dx ) of the liquid , the functional W is a kind of mechanical energy of the liquid within the tank and has been used extensively in the literature of isentropic, compressible liquid flow (see [ 18,22,26,27 ]). We intend to use the functional ( , , , )V w h v defined by (17) as a CLF for the system. However, the functional ( , , , )V w h v can also be used for the derivation of useful bounds for the function h . This is guaranteed by the following lemma. Lemma 1: Define the increasing 1C function 4: 0, ,3G h h by means of the formula 24( ) : sgn 233G h h h h h h h h h (20) Let 1 4: , 0,3G h h be the inverse function of G and define 1:c g (21) Then for every ( , , , )w h v S with 4( , , , )3V w h v h gh , the following inequality holds: 11( , , , ) ( ) ( , , , ) G cV w h v h x G cV w h v , for all [0, ]xL (22) 3.2. The state space In order to be able to satisfy the condition for no spilling out (11) we need to restrict the state space. This becomes clear from the fact that the set S contains st ates that violate the condition for no spilling out (11). Define : ( , , , ) : ( , , , )X w h v S V w h v R (23) where max max2: 2 min 3 ,03gR h h H H h (24) Notice that defi nitions (23), (24), the fact that max hH and Lemma 1 imply that for all ( , , , )w h v X it holds that 11 max 0 ( , , , ) ( ) ( , , , )G cV w h v h x G cV w h v H , for all [0, ]xL (25) Therefore, the condition for no spilling out (11) is automatically satisfied when ( , , , )w h v X . However, it should be noticed that requiring the state to be in X in order to satisfy the condition for no spilling out (11) is a conservative approach: we actually require that the liquid level ()hx is 8 below the height of the walls maxH everywhere in the tank (i.e., for all [0, ]xL ; recall (25)) and not simply at the walls (i.e., at 0x and 1x ) as the condition for no spilling out (11) requires . The state space of system (7) -(10), (12) is considered to be the set X . More specifically, we consider as state space the set 2 1 20, 0, X H L L L endowed with the norm of the underlying normed linear space 2 1 20, 0,H L L L , i.e., 1/22 2 2 22 2 2 2( , , , )Xw h v w h h v (26) 3.3. Main result Now we are in a position to state the first main result of the present work. Theorem 1: For every [0, )rR , where 0R is the constant defined by (24), and for every , , 0qk with 1 1G crkq b G cr (27) where 0c is defined by (21), :g gL , 2 max 24:mL Hb , there exist constants ,0M with the following property: (P) Every classical solution of the PDE -ODE system (7) -(10), (12) and 0( ) 2 ( , ) ( , ) ( , ) ( ,0) ( ) ( )L f t h t x v t x dx h t L h t q w t k t , for 0t (28) with ( (0), (0), [0], [0])V w h v r , satisfies ( ( ), ( ), [ ], [ ])t w t h t v t X and the following estimate for all 0t : [0, ] [0, ] ( ( ), ( ), [ ] , [ ]) exp ( (0), (0), [0] , [0])LLXXt w t h t h v t M t w h h v (29) Theorem 1 guarantees semi -global exponential stabilization of the state in the norm of 2 1 20, 0,H L L L by means of the nonlinear feedback law (28). Here the term semi -global refers to the set X for which spill -free transfer can be guaranteed: the constant rR can be arbitrarily close to R . It should be noticed that the feedback law (28) does not require the measurement of the whole liquid level and liquid velocity profile and requires the measurement of only four quantities: the tank position ()t and the tank velocity ()wt , the total liquid momentum 0( , ) ( , )L h t x v t x dx , and the liquid level differenc e at the tank walls ( , ) ( , 0)h t L h t . The proof of Theorem 1 shows that as rR we have 0k , 0 and M . This is expected because as rR we may have initial conditions satisfying ( (0), (0), [0], [0])V w h v r 9 for which the liquid in the container is agitated. Therefore, in such cases the movement of the container may cause additional agitation of the liquid which can make the liquid spi ll-out unavoidable. 3.4. Auxiliary results For the proof of Theorem 1, we need some auxiliary lemmas. The first auxiliary lemma provides formulas for the time derivatives of the energy functionals defined by (18), (19). Lemma 2: For every classical solution of the PDE -ODE system (7) -(10), (12) the following equations hold for all 0t : 2 00[ ], [ ] ( , ) ( , ) ( ) ( , ) ( , )LL xdE h t v t h t x v t x dx f t h t x v t x dxdt (30) 2 00[ ], [ ] ( , ) ( ) ( , ) ( , ) ( , )LL xxdW h t v t g h t x dx f t h t x v t x h t x dxdt (31) where ,EW are the functionals defined by (18), (19), respectively. The two next auxiliary lemmas provide useful inequalities for the CLF V defined by (17). Lemma 3: Let ,0qk be given. Then there exists a non-decre asing function :[0, ) (0, )R , where 0R is defined by (24), such that for every ( , , , )w h v X with 1(0, ) v H L , the following inequality holds: 2 2 2 2 00( , , , ) ( ( , , , )) ( ) ( ) ( )LL xx V w h v V w h v h x dx h x v x dx w k (32) Lemma 4: Let ,0qk be given. Then there exist non -decreasing functions :[0, ) (0, )iGR , 1, 2i , where 0R is defined by (24), such that for every ( , , , )w h v X , the f ollowing inequality holds: 2 [0, ] 1 2( , , , )( , , , ) ( , , , ) ( , , , )( , , , )LXV w h vw h h v V w h v G V w h vG V w h v (33) where X is defined by (26) . Inequality (32) provides an estimate of the dissipation rate of the Lyapunov functional for the closed -loop system (7) -(10), (12) with (28). On the other hand, inequalities (33) provide estimates of the Lyapunov functional in terms of the norm of the state space. 10 4. Can a Robot Move a Glass of Water Without Spilling Out W ater? We next consider the problem of the movem ent of a glass of water by means of a robotic arm , from near-rest to near -rest in finite (but not prespecified) time . The glass of water starting from an almost at rest state (both the gla ss and the water in the glass), is formulated as [0, ] (0, (0), [0] , [0])LXw h h v (34) where 0 is a small number (tolerance) and X is the norm defined by (26). The robotic arm should move the glass to a pre -specified position witho ut spilling out water of the glass and having the water in the glass almost still at the final time , i.e., at the final time 0T we must have [0, ] ( ( ), ( ), [ ] , [ ])LXT w T h T h v T (35) In other words, we require the spill -free and slosh -free motion of the glass. The initial condition is (0), (0), [0], [0]w h v S , where (0) 0 (recall the definition of S (16)) . The problem ca n be solved by using Theorem 1 w hen the tolerance 0 is sufficiently small. More specifically, we require that 0 is small so that 122 max3max , ,2Hh L g R and max min , h H h L (36) where max max2: 2 min 3 ,03gR h h H H h . If inequalities (36 ) hold then w e can follow the following algorithm: Step 1: Pick number s (0, )rR , ,0q with 12 max3max , ,2Hq h L g and 122 max3max , ,2Hh L g r (this is possible due to (36)) . Step 2: Select 0k so that (27) holds and so that 122 max3max , , ,2 2 3 (0)Hr h L g q kq (37) Inequality (37) guarantees the inequality ( (0), (0), [0], [0])V w h v r . Indeed, this fact is a direct consequence of (34), (36) , (37) and the following proposition. Proposition 1 : Let ,0qk be given. Then for every ( , , , )w h v S satisfying the inequality [0, ] (0, , , )LXw h h v for some 0 with max min , h H h L , the following inequality holds: 11 21 222 max [0, ]3 3( , , , ) max , , , (0, , , )22LXH qkV w h v h L g q w h h v (38) where X is defined by (26) . Let ,0M be the constants involved in (29) and correspond to the selected parameters (0, )rR , ,0q and 0k . Step 3: Set (0) 1lnMMT and apply the feedback law (28) for (0, ]tT . Inequalit y (29) implies that estimate (35 ) holds. The application of t he above algorithm can guarantee that the robotic arm will move the glass to the pre-specified position without spilling out water of the glass and without residual end point sloshing no matter how small the difference max 0 Hh is. Of cou rse, the (minimization of the ) final time 0T depends on the (appropriate ) selection of the parameters (0, )rR , ,0q and 0k . 5. Proofs We start by providing the pr oof of Lemma 1. Proof of Lemma 1: The fact that the function 4: 0, ,3G h h defined by (20) is a 1C increasing function follows from the fact that the equation ()h hrh G h dr r holds for all 0h . Let ( , , , )w h v S with 4( , , , )3V w h v h gh be given. Using the inequality 2 2 2 2 21()2xx hv h h h v and definition (19), we obtain the following estimate: 221 2 2 0 0 011( , ) ( ) ( ) ( ) ( ) ( )4 2 2L L L x W h v h x h x dx h x v x dx g h x h dx (39) Using definitions (17), (18) and estimate (39), we obtain: 2212 00( ) ( ) ( ) ( , ) ( , ) ( , , , )4LL x h x h x dx g h x h dx W h v E h v V w h v (40) Let arbitrary , [0, ]x y L be given. Using the Cauchy -Schwarz inequality and the fact that ()hh Gh h for all 0h (a consequence of definition (20)), we get: 12 max( , ) min( , ) 1/2 1/2max( , ) max( , ) 2 12 min( , ) min( , ) 1/2 2 12 00( ( )) ( ( )) ( ( )) ( ) ( ( )) ( ) ( ) ( ) ( ) ( ( )) ( ) ( ) ( ) ( ( ))xy x ss y x y x y x y s x y x y LL sG h x G h y G h s h s ds G h s h s ds h s h s ds h s G h s ds h s h s ds h s G h s ds 1/2 1/2 1/2 212 00( ) ( ) ( )LL s h s h s ds h s h ds (41) Since 2 22max :4Vab a gb V g for all 0V , we obtain from (40), (41) and definition (21): ( ( )) ( ( )) ( , , , )G h x G h y cV w h v , for all , [0, ]x y L (42) Since 0()L h x dx m (recall definition (16)) and since / h m L , it follows from continuity of h and the mean value theorem that there exists [0, ]xL such that ()h x h . Moreover, since ( ) 0Gh (a consequence of definition (20)), we get from (42) (with yx ): ( , , , ) ( ( )) ( , , , )cV w h v G h x cV w h v , for all [0, ]xL (43) Inequality (22) is a direct consequence of (43), the fact that the inverse function of G is increasing and the fact that 4, , ,3cV w h v h h . The proof is complete. We continue with the proof of Lemma 2. Proof of Lemma 2: Using (18) we obtain for all 0t : 2 0 001( [ ], [ ]) ( , ) ( , )2 ( , ) ( , ) ( , ) ( , ) ( , )L t LL ttdE h t v t h t x v t x dxdt h t x v t x v t x dx g h t x h h t x dx (44) Using (8), (9) we obtain: 1 t x x x xv vv gh h hv f , for 0t , 0,xL (45) Combining (8), (44) and (45) we get for all 0t : 13 2 00 2 00 001( [ ], [ ]) ( , ) ( , ) ( , ) ( , ) ( , ) ( , )2 ( ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , )LL x xx LL x LL x xdE h t v t h t x v t x v t x dx v t x h t x v t x dxdt f t h t x v t x dx h t x v t x v t x dx g h t x v t x h t x dx g h t x h h t x v t x dx (46) Integrating by parts and using (10), we get for all 0t : 2 00 00 22 00( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) 2 ( , ) ( , ) ( , )LL xx x LL x x LL x xv t x h t x v t x dx h t x v t x dx h t x h h t x v t x dx h t x v t x h t x dx h t x v t x v t x dx h t x v t x v t x dx (47) Equation (30) is a direct consequence of (46) and (47). Next define for all 0t and [0, ]xL : ( , ) ( , ) ( , ) ( , )x t x h t x v t x h t x (48) Using (8), (9) and definition (48) we conclude that the following equation holds for all 0t and (0, )xL : 2 1( , ) ( , ) ( ) ( , ) ( , ) ( , )2t xt x h t x f t t x v t x gh t x (49) Using (19) and definition (48), we obtain for all 0t : 22 0 1 001( [ ], [ ]) ( , ) ( , ) ( , )2 ( , ) ( , ) ( , ) ( ) ( , )L t LL ttdW h t v t h t x h t x t x dxdt h t x t x t x dx g h x h h t x dx (50) Combining (B12) and (8), (49), we get for all 0t : 22 0 12 00 01( [ ], [ ]) ( , ) ( , ) ( , ) ( , )2 1( ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , )2 ( ) ( , ) ( , )L x LL x L xdW h t v t h t x t x h t x v t x dxdt f t t x dx h t x t x t x v t x gh t x dx g h x h h t x v t x dx (51) Integrating by parts and using (10), we get for all 0t : 14 22 0 2 2 1 00( , ) ( , ) ( , ) ( , ) 2 ( , ) ( , ) ( , ) ( , ) 2 ( , ) ( , ) ( , ) ( , )L x LL xxh t x t x h t x v t x dx h t x h t x t x v t x dx h t x t x t x v t x dx (52) Combining (51), (52) and (47), we obtain for all 0t : [ ], [ ]) ( , ) ( , ) ( , ) ( , ) 2 ( , ) ( , ) ( , ) ( , ) ( ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , ) ( , h t v t h t x h t x t x v t x dxdt h t x t x t x v t x dx f t t x dx h t x t x v t x dx g h t x v t x t x h t x dx (53) Equation (53) in conjunction with definition (48) gives for all 0t : 21 0 2 00( [ ], [ ]) ( , ) ( , ) ( , ) ( ) ( , ) ( , ) ( , ) ( , )L x LL xxdW h t v t t x h t x v t x dxdt f t h t x v t x h t x dx g h t x dx (54) Equation (31) is a direct consequence of (54) and (10). The proof is complete. Next, we provide the proofs of Lemma 3 and Lemma 4. Proof of Lemma 3: Let arbitrary ( , , , )w h v X with 1(0, ) v H L be given. Since 0()L h x dx m (recall definitions (16), (23)) and since / h m L , it follows from continuity of h and the mean value theorem that there exists [0, ]xL such that ()h x h . Using the Cauchy -Schwarz inequality, we get for all [0, ]xL : 1/2max( , ) 2 00 min( , )( ) ( ) ( ) ( ) ( ) ( )xx x L L s s s s x x xh x h h s ds h x h h s ds h s ds L h s ds (55) Therefore, using (55), we obtain: 222 00( ) ( )LL x h x h dx L h x dx (56) Using the inequality 2 2 2 2 2( ( ) ( ) ( )) 2 ( ) ( ) 2 ( )xx h x v x h x h x v x h x and (25), we obtain: 15 1 2 2 2 1 2 0 0 0 2 22 1 00( )( ( ) ( ) ( )) 2 ( ) ( ) 2 ( ) ( ) 22 ( ) ( ) ( ) ( , , , )L L L xx LL xh x h x v x h x dx h x v x dx h x h x dx h x v x dx h x dx G cV w h v (57) where G is defined by (20). Since (0) ( ) 0v v L (recall definitions (16), (23)), by virtue of Wirtinger's inequality we have: 2 22 2 00( ) ( )LL xLv x dx v x dx (58) Combining (58) and (25) we get: 2 1 2 00 21 2 1 2 2 2 2 1 00( ) ( ) ( , , , ) ( ) ( , , , )( , , , ) ( ) ( ) ( ) ( , , , )LL LL xxh x v x dx G cV w h v v x dx L G cV w h v LG cV w h v v x dx h x v x dx G cV w h v (59) Using (17), (18), (19) and (56), (57), (59), we obtain: 212 00 22 22 0 2 2 2 2 1 00 22 2 22 001, , , ( )( ( ) ( ) ( )) ( )2 1( ) ( )2 2 2 3( ) ( ) ( )2 ( , , , ) 22 ( , , , ) ( ) ( ) ( w h v h x h x v x h x dx g h x h dx qk qh x v x dx w k h x v x dx gL h x dx G cV w h v qk qwk V w h v h x v x dx h x dx 2 2wk (60) where 21 22 2 2 1 13: max , , ,22 2L G cs qk qs gL G cs G cs , for [0, )sR (61) The proof is complete. Proof of Lemma 4: Let arbitrary ( , , , )w h v X be given. Using definitions (17), (18), (19), the inequalities 2 2 2 2 2( ( ) ( ) ( )) 2 ( ) ( ) 2 ( )xx h x v x h x h x v x h x , 2 2 2 222 w k w k and (25), we obtain: 16 212 00 22 22 0 222 2 1 2 2 2 0 0 0 2 22 max 1 001, , , ( )( ( ) ( ) ( )) ( )2 1( ) ( )2 2 2 33( ) ( ) ( ) ( ) ( )22 3( ) ( )2 ( , , , )LL x L L L L x LL xV w h v h x h x v x h x dx g h x h dx qk qh x v x dx w k qkh x v x dx h x h x dx g h x h dx qw H v x dx h x dx g h G cV w h v h dx qkqw (62) Inequality (62) implies the left inequality (33) with 22 2 max 133: max , , , ,22qkG s H q g G cs for [0, )sR . Using definitions (17), (18), (19), the inequalities 2 1 2 3( ) ( ) ( ) ( ) ( ) ( )43xx v x h x h x v x h x h x , 22 31 43kwwk and (25), we obtain: 212 00 22 22 0 2222 1 2 2 2 0 0 0 2 1 2 2 max 001, , , ( )( ( ) ( ) ( )) ( )2 1( ) ( )2 2 2 1( ) ( ) ( ) ( ) ( )4 6 4 6 1, , , ( ) ( )46LL x L L L L x LL xV w h v h x h x v x h x dx g h x h dx qk qh x v x dx w k qk qh x v x dx h x h x dx g h x h dx w G cV w h v v x dx h x dx g hH 2 0 2 22() 46L x h dx qk qw (63) Inequality (63) implies the right inequality (33) with max 11 2 2 max max max max12: min 3 , 2 ,3 , 2 ,12HGs H G cs H qk H q H g for [0, )sR . The proof is complete. We next give the proof of Theorem 1. Proof of Theorem 1: Let [0, )rR be given (arbitrary) and let constants , , 0qk for which (27) holds (but otherwise arbitrary) be given. Consider a classical solution of t he PDE -ODE system (7) -(10), (12) with (28) that satisfies ( (0), (0), [0], [0])V w h v r . Then we obtain from Lemma 2 and (7), (17), (28) for all 0t : 17 2 2 3 2 00 2 0( ), ( ), [ ], [ ] ( , ) ( , ) ( , ) ( ) ( ) ( ) ( ) 2 ( , ) ( , ) ( , ) ( ) ( )LL xx L xdV t w t h t v t g h t x dx h t x v t x dx qk tdt qk w t k t f t h t x v t x h t x dx q w t k t (64) Using (28) and (64), we obtain for all 0t : 2 0 2 2 3 2 0 2 0( ), ( ), [ ], [ ] ( , ) ( , ) ( , ) ( ) ( ) ( ) 2 ( , ) ( , ) ( , ) ( ) ( )L x L x L xdV t w t h t v t g h t x dxdt h t x v t x dx qk t qk w t k t h t x v t x h t x dx q w t k t (65) Using the inequality 0 2 2 1 02 ( ) ( ) 2 ( , ) ( , ) ( , ) ( ) ( ) 2 ( , ) ( , ) ( , )L x L xw t k t h t x v t x h t x dx w t k t h t x v t x h t x dx that holds for every 0 , we obtain from (65) the following estimate for all 0t , (0, )q : 22 00 2 32 2 1 0( ), ( ), [ ], [ ] ( , ) ( , ) ( , ) ( ) ( ) ( ) ( ) 1 2 ( , ) ( , ) ( , )LL xx L xdV t w t h t v t g h t x dx h t x v t x dxdt qk t q R k w t k t q h t x v t x h t x dx (66) Using the inequality 00 22 12 002 2 ( , ) ( , ) ( , ) 4 ( , ) ( , ) ( , )LL x LL xh t x v t x dx h t x dx h t x v t x dx h t x dx that holds for every 0 , we obtain from (66) the following estimate for all 0t , (0, )q , 0 : 18 22 00 2 32 2 1 0 2 2 1 1 0( ), ( ), [ ], [ ] ( , ) ( , ) ( , ) ( ) ( ) ( ) ( ) 4 1 1 ( , ) ( , ) 1 1 ( , )LL xx L L xdV t w t h t v t g h t x dx h t x v t x dxdt qk t q q k w t k t q h t x v t x dx q h t x dx (67) The Cauchy -Schwarz inequality in conjunction with (12) implies the following inequalities: 2 2 00 2 2 00( , ) ( , ) ( , ) ( , ) ( , ) ( , )LL xx LLh t x dx L h t x dx h t x v t x dx m h t x v t x dx Thus, we obtain from (67) the followin g estimate for all 0t , (0, )q , 0 : 1 1 2 0 2 2 3 2 0 12 0( ), ( ), [ ], [ ] 1 1 ( , ) ( , ) ( , ) ( ) ( ) ( ) ( ) 4 1 1 ( , ) ( , )L x L x LdV t w t h t v t g L q h t x dxdt h t x v t x dx qk t q q k w t k t m q h t x v t x dx (68) Define for each 0t : max0( ) : max ( , ) xLh t h t x , min0( ) : min ( , ) 0 xLh t h t x (69) The fact that min( ) 0ht follows from the fact that 1[0, ) [0, ];(0, ) h C L (which implies that for every 0t , []ht is a continuous positive function on [0, ]L ). Definition (69) in conjunction with (68) implies the following estimate for all 0t , (0, )q , 0 : 1 1 2 0 2 2 3 2 0 12 max 0( ), ( ), [ ], [ ] 1 1 ( , ) ( , ) ( , ) ( ) ( ) ( ) ( ) 4 1 1 ( ) ( , )L x L x LdV t w t h t v t g L q h t x dxdt h t x v t x dx qk t q q k w t k t m q h t v t x dx (70) Since ( , 0) ( , ) 0v t v t L (recall (10)), by virtue of Wirtinger's inequality and definition (69), we have for 0t : 19 22 2 2 2 22 min 0 0 0( , ) ( , ) ( , ) ( , ) ()L L L xxLLv t x dx v t x dx h t x v t x dx ht (71) Combining (70) and (71), we obtain the following estimate for all 0t , (0, )q , 0 : 1 1 2 0 2 12 max 2 min 0 2 32( ), ( ), [ ], [ ] 1 1 ( , ) ()4 1 1 ( , ) ( , )() ( ) ( ) ( ) ( )L x L xdV t w t h t v t g L q h t x dxdt L h tm q h t x v t x dxht qk t q q k w t k t (72) Setting 1 1 1G crqqb G cr (notice that since g gL it is automatically guaranteed that (0, )q ) and 1 1( 1)LG crgL Lb G crgL where 1 11:12 22LG crgL Lb G crgL (notice that 0 ), it follows from (72) and the facts that g gL , 2 max 24mL Hb that the following estimate holds for all 0t : 2 0 1 2 max max 1 max min 0 1 2 3 2 2 1( ), ( ), [ ], [ ] 1 ( , ) ()( , ) ( , )() ( ) 1 ( ) ( )L x L xdV t w t h t v t g h t x dxdt G cr H h th t x v t x dxH G cr h t G crqk t q w t k tb G cr (73) where 1 1:k b G cr q G cr and 1:1 ( 1)b Lb G crgL (74) Notice that (27) implies that (0,1) . Inequality (73) in conjunction with the fact s that (0,1) , (0,1) shows that the following implication holds: If 0t and max max 1 min() ()h t H h t G cr then ( ), ( ), [ ], [ ] 0dV t w t h t v tdt (75) The fact that 1[0, ) [0, ];(0, ) h C L implies that both mappings max() t h t , min( ) 0 t h t , defined by (69) are continuous (see Proposition 2.9 on page 21 in [14 ]). Consequently, the mapping is continuous with 1 max 1 min(0) () (0) ()h G cr h G cr . The latter inequality is a consequence of 20 the fact that ( (0), (0), [0], [0])V w h v r and (25). Since 1 max ()G cr H and 1 it follows that max max 1 min(0) (0) ()hH h G cr . Therefore, by continuity there exists 0T such that max max 1 min() ()h t H h t G cr for all [0, )tT . We next prove by contradiction that max max 1 min() () ()h t H ht G cr for all 0t . Assume the contrary, i.e. that there exists 0t such that max max 1 min() () ()h t H ht G cr . Therefore, the set max max 1 min(): 0:() ()h t HAtht G cr is non -empty and bounded from below. Thus we can define inf( )tA . By definition, it holds that 0 tT and max max 1 min() () ()h t H ht G cr for all [0, )tt . By continuity of the mapping we obtain that max max 1 min() ( ) ( )h t H h t G cr . Moreover, since max max 1 min() () ()h t H ht G cr for all [0, )tt , it follows from implication (75) that ( ), ( ), [ ], [ ] 0dV t w t h t v tdt for all (0, )tt . By continuity of the mapping ( ), ( ), [ ], [ ] t V t w t h t v t , we obtain that ( ), ( ), [ ], [ ] (0), (0), [0], [0]V t w t h t v t V w h v r . On the other hand, the previous inequality and (25) imply that 1 max max 11 min() () ( ) ( ) ( )h t H G cr h t G cr G cr which contradicts the equation max max 1 min() ( ) ( )h t H h t G cr . Since max max 1 min() () ()h t H ht G cr for all 0t , we conclude from implication (75) that ( ), ( ), [ ], [ ] 0dV t w t h t v tdt for all 0t . By continuity of the mapping ( ), ( ), [ ], [ ] t V t w t h t v t , we obtain that ( ), ( ), [ ], [ ] (0), (0), [0], [0]V t w t h t v t V w h v r R for all 0t Hence, ( ( ), ( ), [ ], [ ])t w t h t v t X for all 0t (recall definitions (16), (23)). Moreover, (25) implies that 1 max 1 min() () () ()ht G cr ht G cr for all 0t . The previous inequality in conjunction with (73) gives for all 0t : 2 2 2 2 00( ), ( ), [ ], [ ] ( , ) ( , ) ( , ) ( ) ( ) ( )LL xxdV t w t h t v t h t x dx h t x v t x dx t w t k tdt (76) where 21 1 1 3 2 1 max: min 1 , 1 , , 1G crg G cr qk qH b G cr (77) It follows from Lemma 3 and (76) that the following estimate holds for all 0t : ( ), ( ), [ ], [ ] ( ), ( ), [ ], [ ]( ), ( ), [ ], [ ]dV t w t h t v t V t w t h t v tdt V t w t h t v t (78) where :[0, ) (0, )R is the non -decreasing function involved in (32). Since :[0, ) (0, )R is non -decreasing and since ( ), ( ), [ ], [ ]V t w t h t v t r for all 0t , we obtain from (78) the following estimate for all 0t : ( ), ( ), [ ], [ ] ( ), ( ), [ ], [ ]dV t w t h t v t V t w t h t v td t r (79) By continuity of the mapping ( ), ( ), [ ], [ ] t V t w t h t v t , the differential inequality (79) implies the following estimate for all 0t : ( ), ( ), [ ], [ ] exp (0), (0), [0], [0]tV t w t h t v t V w h vr (80) Estimate (80) in conjunction with Lemma 4 implies the following estimate for all 0t : 2 [0, ] 2 1 2 [0, ]( ( ), ( ), [ ] , [ ]) exp ( ), ( ), [ ], [ ] (0), (0), [0], [0] ( (0), (0), [0] , [ 0])LX LXtt w t h t h v tr G V t w t h t v t G V w h v w h h v (81) where :[0, ) (0, )iGR , 1, 2i , are the non -decreasing functions involved in (33). Since :[0, ) (0, )iGR , 1, 2i , are non -decreasing functions and since ( ), ( ), [ ], [ ]V t w t h t v t r for all 0t , we obtain from (81) the following estimate for all 0t : 22 [0, ] 1 2 [0, ] ( ( ), ( ), [ ] , [ ]) exp ( (0), (0), [0] , [0])LLXXtt w t h t h v t G r G r w h h vr (82) Estimate (29) is a direct consequence of estimate (82). The proof is complete. We end this section by providing the proof of Proposition 1. Proof of Proposition 1: Let arbitrary ( , , , )w h v S that satisfies (34) with max min , 0h H h L be given. Using defi nitions (17), (18), (19), the inequalities 2 2 2 2 2( ( ) ( ) ( )) 2 ( ) ( ) 2 ( )xx h x v x h x h x v x h x , 2 2 2 222 w k w k and (25), we obtain: 22 212 00 22 22 0 222 2 1 2 2 2 0 0 01, , , ( )( ( ) ( ) ( )) ( )2 1( ) ( )2 2 2 33( ) ( ) ( ) ( ) ( )22LL x L L L L xV w h v h x h x v x h x dx g h x h dx qk qh x v x dx w k qkh x v x dx h x h x dx g h x h dx qw (83) Since 0()L h x dx m (recall definition (16)) and since / h m L , it follows from continuity of h and the mean value theorem that there exists [0, ]xL such that ()h x h . Using the Cauchy - Schwarz inequality, we get (55) for all [0, ]xL . Consequently, we obtain for all [0, ]xL from (55), (26), (34) and the fact that max min , 0h H h L : max 220 ( )h L h h x h L h h L H (84) Therefore, we obtain from (83) and (84): 21 22 2 2 2 2 max 0 0 03 3, , , ( ) ( ) ( )22L L L xH qkV w h v v x dx h L h x dx g h x h dx qw (85) Inequality (38) is a direct consequence of (85) and definition (26). The proof is complete. 6. Concluding Remarks By applying the CLF methodology we have managed to achieve semi -global stabilization results for the viscous Saint -Venant liquid -tank system. As mentioned in the Introduction, to our knowledge, this is the first paper in the literature that achieves stabilization of the nonlinear viscous Saint - Venant system. Moreover, this is the first paper that guarant ees a spill -free condition for the movement of the tank. The obtained results leave some open problems which will be the topic of future research : 1) The results were applied to classical solution s. It is of interest to relax this to weak solutions . It is an open problem to show that the proposed feedback laws preserve their strong stabilizing role in the case of weak solutions. Moreover, it is an open problem to show of (weak) solutions for the closed -loop system. To this p urpose, ideas utilized in [30 ] will be employed. 2) The construction of CLFs which can allow the derivation of stability estimates in strong er spatial norms for the liquid level and velocity profiles. Another more demanding problem that can be studied in the future is the spill -free, slosh -free and smash -free movement of a glass of water. This problem arises when we want to move the glass of water to a position which is close to a wall. In this case, we need to have control on the overshoot of the glass posit ion error in order to avoid smashing the glass on the wall. 23 References [1] Barre de Saint -Venant, A. J. C., Theorie du Mouvement non Permanent des Eaux, avec Application aux Crues des Rivieres et a l'I ntrodu ction de Marees dans Leurs Lits , Comptes Rendus de l'Academie des Sciences , 73, 1871, 147-154 and 237 -240. [2] Bastin, G. , J.-M. Coron and B. d'Andrea Novel, On Lyapunov Stability of Linearised Saint - Venant Equations for a Sloping C hannel , Networks and Heterogeneous Media , 4, 2009, 177 - 187. [3] Bastin, G. and J. -M. Coron, Stability and Boundary Stabiliz ation of 1 -D Hyperbolic Systems , 2016, Birkhauser . [4] Bastin , G., J. -M. Coron and A. Hayat , Feedforward Boundary Control of 2x2 Nonlinear Hyperbolic Systems with Application to Saint -Venant Equations, European Journal of Control , 57, 2021, 41 -53. [5] Bresch, D. and P. Noble , Mathematical Justification of a Shallow Water Model , Methods and Applications of Analysis , 14, 2007, 87-118. [6] Chin -Bing S. A., P. M. Jordan and A. Warn -Varnas, A Note on the Viscous, 1D Shallow Water Equations: Traveling Wave Phenomena, Mechanics Research Communications , 38, 2011, 382 - 387. [7] Coron , J. M. , Local Controllability of a 1 -D Tank C ontaining a Fluid Modeled by the Shallow Water E quations , ESAIM: Control Optimisation and Calculus of Varia tions , 8, 2002, 513-554. [8] Coron, J. -M., Control and Nonlinearity , Mathematical Surveys and Monographs, Volume 136, American Mathematical Society, 2007. [9] Coron, J .-M., B. d'Andrea Novel and G. Bastin, A Strict Lyapunov Function for Boundary Control o f Hyperbolic Systems of Conservation L aws, IEEE Transactions on Automatic Control , 52, 2007, 2 -11. [10] de Halleux, J. and G. Bastin, Stabilization of Saint -Venant Equations Using Riemann Invarian ts: Application to Waterways with Mobile Spillways, IFAC Proceedings , 35, 2002, 131-136. [11] Diagne, A., M. Diagne, S. Tang and M. Krstic, Backstepping Stabilization of the Linearized Saint -Venant -Exner Model, Automatica , 76, 2017, 345 -354. [12] Diagne, M., S. -X. Tang, A. Diagne and M. Krstic, Control of Sha llow Waves of two Unmixed Fluids by Backstepping, Annual Reviews in Control , 44, 2017, 211 -225. [13] Feddema J. T., C. R. Dohrmann, G. G. Parker, R. D. Robinett, V. J. Romero and D. J. Schmitt, Control for Slosh -Free M otion of an Open C ontainer , IEEE Control Systems Magazine , 17, 1997, 29-36. [14] Freeman , R. A. and P. V. Kokotovic, Robust Nonlinear Control Design , 1996, Birkhauser. [15] Gerbeau , J.-F. and B. Perthame, Derivation of Viscous Saint -Venant System for Laminar Shallow -Water: Numerical Validation , Discrete and Continuous Dynamical Systems, Series B , 1, 2001, 89-102. [16] Karafyllis, I. and M. Krstic, Global Stabilization of a Class of Nonlinear Reaction -Diffusi on PDEs by Boundary Feedback, SIAM Journal on Control and Optimization , 57, 2019, 3723 - 3748 . [17] Karafyllis, I., Lyapunov -Based Boundary Feedback Design For Parabolic PDEs, International Journal of Control , 94(5), 2021, 1247 -1260 . [18] Karafyllis , I. and M. Krstic , Global Stabilization of Compressible Flow Between Two Moving Pistons , su bmitted to SIAM Journal on Control and Optimization (see also arXiv:2104.09153 [math.OC] ). [19] Krstic, M., On Global Stabilization of Burgers' Equat ion by Boundary Control , Systems & Control Letters , 37, 1999, 123 -142. [20] Lannes , D., The Water Waves Problem . Mathematical Analysis and Asymptotics , 2013, American Mathematical Society. [21] Litrico, X. and V. Fromion, Boundary C ontrol of Linearized Saint -Venant Equations Oscillating Modes, Automatica , 42, 2006, 967 -972.

2108.11052

Iasson Karafyllis

Iasson Karafyllis and Miroslav Krstic

Spill-Free Transfer and Stabilization of Viscous Liquid

24 pages,1 figure, submitted to IEEE TAC for possible publication

math.OC cs.SY eess.SY math.AP physics.flu-dyn

http://creativecommons.org/licenses/by/4.0/

This paper studies the feedback stabilization problem of the motion of a tank that contains an incompressible, Newtonian, viscous liquid. The control input is the force applied on the tank and the overall system consists of two nonlinear Partial Differential Equations and two Ordinary Moreover, a spill-free condition is required to hold. By applying the Control Lyapunov Functional methodology, a set of initial conditions (state space) is determined for which spill-free motion of the liquid is possible by applying an appropriate control input. Semi-global stabilization of the liquid and the tank by means of a simple feedback law is achieved, in the sense that for every closed subset of the state space, it is possible to find gains, so that every solution of the closed-loop system initiated from the given closed subset satisfies specific stability estimates. The closed-loop system exhibits an exponential convergence rate to the point. The proposed stabilizing feedback law does not of the liquid level and velocity profiles inside the tank and simply requires measurements of: (i) the tank position error and tank velocity, (ii) the total momentum of the liquid, and (iii) the liquid levels at the tank walls. The obtained results allow an algorithmic solution of the problem of the spill-free movement and slosh-free settlement of a liquid in a vessel of limited height (such as water in a glass) by a robot to a no matter how full the vessel is.