![Arxiv:2007.15477V1 [Cond-Mat.Stat-Mech] 29 Jul 2020 Lwdnmc Ntemnecrosmlto.Swend- Simulation](http://data.docslib.org/img/af220c82a08f116776f5fbc2a1130e7c-1.webp)
arXiv:2007.15477v1 [cond-mat.stat-mech] 29 Jul 2020 lwdnmc nteMneCrosmlto.Swend- simulation. Carlo Monte the in dynamics slow were phase, topological classified. ordered type successfully the BKT the and with disordered along The phases, studied. was [4–7] tran- sition transitions (BKT) Berezinskii-Kosterlitz-Thouless first-order the the parameter. also and order but second-order vector multi-component the a only with the Not systems including the and models provided was systems spin treatment similar various a so, configura- to doing spin By the of itself. instead tion spa- treated long-range was a correlation of tial configuration The transitions. phase oe n nIigltieguetheory. gauge lattice Ising Ising (2D) an two-dimensional neural and the of model of using study use classification the the for image demonstrated networks to They similar ferromag- learning. was machine low-temperature It a high-temperature phase. netic and a configura- identified phase spin and of paramagnetic classified sets data they large tions, using By interacting systems. studying of complemen- spin approach is conventional that the paradigm to a tary propose to learning vised [1]. Statistical disciplines scientific such interest. fundamen- of increasing same one for is gained the learning physics has At machine research of tal etc. potential includ- the healthcare, industry which in time, driving, decade, areas past many automated on the ing impact in an given made have been have techniques ∗ § ‡ † [email protected] [email protected] [email protected] [email protected] lse loihs[,9 aebe sdt overcome to used been have 9] [8, algorithms Cluster Shiina super- of technique a used [2] Melko and Carrasquilla based machine-learning of developments Remarkable tal. et 4 h unu Ymdlcnb efre yuigtetann da training the using by performed be approach ph can machine-learning BKT model the the XY using classify quantum We model the XY lattice. quantum square the the of on qu model the Berezinskii-Kosterlitz XY to the analyze quantum classi applied We also of algorithm. is phases loop method The the the models. and config spin estimators, improved correlation of ordinary classification the phase to the alternative an as uration ahn-erigSuyuigIpoe orlto Configu Correlation Improved using Study Machine-Learning colo optn,Ntoa nvriyo igpr,1 C 13 Singapore, of University National Computing, of School eueteFrunKseenrpeetto ae improve based representation Fortuin-Kasteleyn the use We 3 eotdamcielann td on study machine-learning a reported [3] 2 eateto hsc,TkoMtooia nvriy Hac University, Metropolitan Tokyo Physics, of Department uueTomita Yusuke 5 3 1 igpr y eerhIsiue(EI,1 hr Hospital Third 11 (SERI), Institute Research Eye Singapore iifraisIsiue gnyfrSine Technology Science, for Agency Institute, Technology, Bioinformatics of Institute Shibaura Engineering, of College n plcto oQatmMneCroSimulation Carlo Monte Quantum to Application and 6 mg n evsv cesLbrtr IA) Fusionopol 1 (IPAL), Laboratory Access Pervasive and Image 1 0Booi tet 0-1Mti,187,Singapore 138671, Matrix, #07-01 Street, Biopolis 30 2-1Cnei SuhTwr,183,Singapore 138632, Tower), (South Connexis #21-01 , ∗ et Shiina Kenta Dtd uy3,2020) 31, July (Dated: 2 , 3 , † uaaOkabe Yutaka eut ftesi / Ymdlo h qaelattice. square [20]. the transition on BKT the model the show XY exhibits we 1/2 model example, spin This an the As of results systems. technique spin this quantum apply to we Then, Shiina configuration. of correlation method representa- machine-learning the cluster al. use the We in tion. configuration correlation the be can By quantities clusters. non-diagonal measured. non-local algorithm, some loop of configura- the units which using in in updated update, simulation are global tions the Carlo to Monte due largely quantum was the improvements in on The systems accomplished spin [17–19]. quantum representation The to worldline Con- satisfied. the applied was automatically flipped. algorithm are are loop vertices bonds the of at treating paths straints In closed algorithm. models, loop vertex called is which algorithm, ter lse loih,asl-dpe loih otn the tune to [14]. automatically algorithm point self-adapted critical a probability-changing the algorithm, in a cluster utilized in consid- correlations is spin be configuration the should spin manifesting of cluster feature FK spin The same the ered. the only to correlations, improved belonging spin statisti- pair so-called calculating the In reduce to to errors. per- lead designed cal are be they which can [13] that clus- of estimators flip is advantage spin further algorithms A cluster ter cluster. freedom the FK of the Then, with degrees ran- formed Ising up. a onto the picked spin vector, are vector vec- unit a treat of chosen projecting to idea domly By 13] the 12, models. proposed spin [9, tor formalism also was Wolff cluster algorithm embedded [9]. the cluster Wolff the by spins. of proposed of variant Fortuin-Kasteleyn clusters identify single-cluster the to The applied representation 11] [8] [10, (SW) (FK) Wang and sen nti td,w osdrteipoe siao for estimator improved the consider we study, this In Evertz 3 o h lsicto fpae sn h improved the using phases of classification the for [3] Tols BT rniino h pn1/2 spin the of transition (BKT) -Thouless rto ntemcielann td of study machine-learning the in uration a pnmdl r lsie sn the using classified are models spin cal 2 , siao ftecreainconfig- correlation the of estimator d nu ot al iuainusing simulation Carlo Monte antum tal. et ‡ muigDie 147 Singapore 117417, Drive, omputing eso httecasfiainof classification the that show We . n weKa Lee Kuan Hwee and ij,Tko1209,Japan 192-0397, Tokyo hioji, n eerh(A*STAR), Research and atm 3-50 Japan 330-8570, Saitama s n h aaantcphase paramagnetic the and ase ao h lsia Ymodel. XY classical the of ta 1,1]peetdaohrtp fclus- of type another presented 16] [15, v,185,Singapore 168751, Ave, sWay, is 3 , 4 , ration 5 , 6 § et 2
calculated from the spin configuration. The upper figures (a) spin (b) correlation (c) improved are snapshots at the low temperature, T = 0.9, and the
0 0 0 lower figures are those at the high temperature, T =1.06.
5 5 5 Temperatures are measured in the unit of the coupling J.
10 10 10 We note that the exact second-order transition tempera-
15 15 15 ture T for this model is known as 1/ ln(1 + √3) 0.995. c ≈ 20 20 20 Spins are displayed in one of three colors, red, green, or 25 25 25 blue. The ordinary correlation takes a value of 1, 1/2, 30 30 30 or +1/4, whereas the improved correlation takes a− value 0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30
-1/2 1/4 1 0 1/2 1 of 1, 0, or +1/2. The both of correlations from +1 to 1 are mapped in gray scale from 255 (white) to 0 (black).− (d) spin (e) correlation (f) improved The permutation of three-state spins yields an essen- tially identical configuration, and the correlation configu- 0 0 0
5 5 5 rations are invariant under the permutation. The borders
10 10 10 of FK clusters for spin configuration are drawn by lines
15 15 15 in Figs. 1(a) and (d). They are copied in improved cor-
20 20 20 relation configuration. The border of the largest cluster
25 25 25 is drawn by yellow thick line for convenience.
30 30 30 At high temperatures, the spin configurations and the 0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30 correlation configurations are randomly distributed, and -1/2 1/4 1 0 1/2 1 the fluctuation of these quantities gives the susceptibil- FIG. 1. Examples of the spin configuration si ((a), (d)), ity. In the improved correlation, the cancellation among { } correlation configuration gi(L/2) ((b), (e)), and improved different FK clusters are automatically satisfied. Fig- { } correlation configuration g˜i(L/2) ((c), (f)) of the 2D 3-state { } ure 1(e) and 1(f) show the difference between the two Potts model. The upper figures ((a)-(c)) are snapshots at the correlation configurations. While the ordinary correla- low temperature, T = 0.9, and the lower figures ((d)-(f)) are tion configuration in Fig. 1(e) fluctuates in space, a cou- those at the high temperature, T = 1.06. The borders of FK ple of brighter areas in the largest cluster show the im- clusters for spin configuration are drawn by lines. They are copied in improved correlation configuration. proved correlation in Fig. 1(f). As a supplemental material, we provide animations of the spin configuration, the correlation configuration, and We consider the configuration of the spin correlation the improved correlation configuration for the 2D Ising with the distance of the half of the system size, L/2. We model (mp4 files) for convenience [22]. The animations note that this type of correlation function was used along at various temperatures are compared at the low temper- with the generalized scheme for the probability-changing ature (T =2.1), at Tc = 2.269, and at the high temper- cluster algorithm [21]. For the q-state Potts model (in- ature (T =2.4). The system sizes are L=32 and 64. cluding the Ising model), the correlation between two We use the same technique of supervised learning as spins becomes 1 for the same spin pair, whereas it be- Shiina et al. [3] for the classification of the phases of spin comes 1/(q 1) for the pair of different states. In systems. We consider a fully connected neural network the improved− estimator− for the cluster representation, the implemented with a standard TensorFlow library [23] us- correlation becomes 1 for the spin pair belonging to the ing the 100-hidden unit model to classify the ordered, the same FK cluster, whereas it becomes 0 for the spins of BKT, and the disordered phases. For the input layer, we different clusters. When the embedded algorithm for the use the improved correlation configurations g˜i(L/2) . { } continuous spins is used, the projection of spins onto a We have used a cross-entropy cost function supplemented randomly chosen reflection axis is made. We denote the with an L2 regularization term. The neural networks site-dependent correlation configuration as gi(L/2). For were trained using the Adam method [24]. actual calculation, we treat the average value of the x- We first analyzed the 2D 3-state Potts model. The out- direction and the y-direction, that is, put layer averaged over a test set as a function of T for the 2D 3-state Potts model is shown in Fig. 2(a). The prob-
gi(L/2)=(g[sxi,yi ,sxi+L/2,yi ]+ g[sxi,yi ,sxi,yi+L/2])/2, abilities of predicting the phases, the disordered or the (1) ordered, are plotted for each temperature. The system where g[s,s′] denotes a spin-spin correlation between a sizes are L = 32, 48, and 64. The samples of T within the spin pair s and s′. ranges 0.85 T 0.94 and 1.06 T 1.15 were used In Fig. 1, we show the examples of the spin configu- for the training≤ data.≤ We have not≤ used≤ the samples close ration s , correlation configuration g (L/2) , and im- to T for the training data because we assumed the situa- { i} { i } c proved correlation configuration g˜i(L/2) of the 2D 3- tion that the exact Tc is not known. For a whole temper- state Potts model. The spin configuration{ is} generated by ature range, around 35,000 training data sets are used, the Monte Carlo simulation, and the correlation config- and we use 500 test data sets for each temperature. Ten uration and the improved correlation configurations are independent calculations were performed to provide error 3
(a) lation, but the difference is small. The advantage of the improved estimator appears at high enough temperatures 1.0 (compare Fig. 1(f) with 1(e)). We next consider the 2D 6-state clock model. Be- ferro phase para phase 0.8 cause of the discreteness, there are two transitions. One
1.0 L=32(para) is a higher BKT transition, T2, between the disordered 0.6 0.8 L=32(ferro) and BKT phases, and the other is a lower transition, L=48(para) layer 0.6 L=48 L=48(ferro) T1, between the BKT and ordered phases. The output 0.4 0.4 L=64(para) layer averaged over a test set as a function of T for the 0.2 L=64(ferro) Output 0.0 2D 6-state clock model is shown in Fig. 2(b). The sys-
0.2 0.98 1.00 1.02 tem sizes are L = 32, 48, and 64. The samples of T improved ordinary within the ranges 0.4 T 0.64, 0.77 T 0.83, and ≤ ≤ ≤ ≤ 0.0 0.96 T 1.2 were used for the low-temperature, mid- range≤ temperature,≤ and high-temperature training data, 0.85 0.90 0.95 1.00 1.05 1.10 1.15 T/J respectively. The recent numerical estimates of T1 and T2 for the 6-state clock model are 0.701(5) and 0.898(5), (b) respectively [25]. This figure corresponds to Fig. 4(a) of Ref. [3], and the present figure again shows the successful classification into the three phases.
GFSSP QIBTF #,5 QBSB QIBTF We have classified the phases of transitions by means of the machine-learning approach by Shiina et al. [3] using - QBSB - #,5 improved correlation configuration. There is no apprecia- - GFSSP ble difference of accuracy between the use of the correla- - QBSB - #,5 tion configuration and that of the improved correlation - GFSSP - QBSB configuration. The result indicates that the machine- 0VUQVU MBZFS - #,5 learning based phase classification is robust; that is, the - GFSSP phase classification does not discriminate the improved correlation configuration from the ordinary one. Many applications of the loop updating method have been done for quantum systems. Here, we consider 5+ the quantum spin 1/2 XY model in two dimensions, which clearly demonstrated the utility of the loop algo- FIG. 2. (a) The output layer averaged over a test set as a rithm [20]. The Hamiltonian is written as function of T for the 2D 3-state Potts model. The system sizes are L = 32, 48, and 64. The samples of T within the ranges H = J (SˆxSˆx + SˆySˆy). (2) − i j i j 0.85 T 0.94 and 1.06 T 1.15 are used for the training hXi,ji data.≤ In the≤ inset, the comparison≤ ≤ is made between the results of improved correlation and those of ordinary correlation in Here, the spin operators Sˆx,y are one-half of the Pauli c √ the case of L = 48. The exact T , 1/ ln(1 + 3) 0.995, matrices σx,y. The summation is taken over the nearest- is shown by dashed line for convenience. (b) The same≈ plot for the 2D 6-state clock model. The system sizes are the neighbor pairs. This model exhibits the BKT transition same. The samples of T within the ranges 0.4 T 0.64, at around T =0.342 [20]. 0.77 T 0.83, and 0.96 T 1.2 are used for≤ the training≤ We performed the quantum Monte Carlo simulation data.≤ ≤ ≤ ≤ using the loop algorithm, and calculated the spatial cor- relation with the distance of L/2. A D-dimensional quan- tum system can be treated as a (D +1)-dimensional clas- sical system with an extra dimension of imaginary time. analysis. This figure corresponds to Fig. 2(a) of Ref. [3], In calculating the spatial correlation, the summation over and we again observe that the neural network could suc- the imaginary-time axis is taken. The Sˆx-component of cessfully classify the disordered and ordered phases using the correlation function is calculated as [26] the improved correlation configuration. In the inset of Fig. 2(a), we show the comparison of the results of im- 4 β β proved correlation (the present study) and those of the x ˆx r ˆx r g˜ri,rj = 2 S ( i, τ1)S ( j , τ2) dτ1dτ2 previous study [3] of ordinary correlation in the case of β Z0 Z0 β β L = 48. We used the same conditions for both training 1 r r = 2 δℓ( i, τ1; j, τ2) dτ1dτ2, (3) data and test data of improved and ordinary correlations β Z0 Z0 produced from the same spin configurations. The point that the probabilities of predicting two phases are 50% where δ ( ) is the function that returns 1 if the loop of ℓ · is slightly more close to the exact critical temperature, the position ri and the time τ1 and that of the position shown in dashed line in the inset, for the improved corre- rj and the time τ2 belong to the same loop, whereas 4
(a) quantum T = 0.24 (b)quantum T = 0.44 (a)
0 0 TQJO 9: 5 5
10 10 QBSB QIBTF 15 15 #,5 QIBTF - QBSB 20 20 - #,5 - QBSB
25 25 - #,5 - QBSB 30 30 - #,5 0VUQVU MBZFS
0 5 10 15 20 25 30 0 5 10 15 20 25 30
0 1/2 1 0 1/2 1
(c) classical T = 0.6 (d) classical T = 1.2
0 0 5+ 5 5
10 10 (b)
15 15 SPUBUPS UP TQJO 9: 20 20
25 25 QBSB QIBTF 30 30 #,5 QIBTF - QBSB 0 5 10 15 20 25 30 0 5 10 15 20 25 30 - #,5
0 1/2 1 0 1/2 1 - QBSB - #,5 - QBSB i - #,5 FIG. 3. Comparison of snapshots of g˜ (L/2) of the 2D 0VUQVU MBZFS quantum ((a), (b)) and classical ((c), (d)){ XY models.} Ex- amples of snapshots below TBKT ((a), (c)) and those above TBKT ((b), (d)) are displayed.
returns 0 otherwise. Due to the O(2) symmetry of the 5+ model, the Sˆy-component of the correlation function is exactly the same as the Sˆx-component [26]. The factor FIG. 4. (a) The output layer averaged over a test set as a 4 is introduced for the comparison of the spin one-half function of T for the 2D spin 1/2 XY model. The system sizes are L = 32, 48, and 64. The samples of T within the system with the classical model. We checked our calcu- ranges 0.16 T 0.32 and 0.38 T 0.54 are used for lation by the consistency with the precise calculations at the training≤ data.≤ (b) The classification≤ ≤ of the quantum XY T = 0 [27, 28]. model using the training data of the classical XY model (plane It is instructive to compare the correlation configura- rotator). The samples of T within the ranges 0.50 T 0.84 tions of the quantum XY model and the classical XY and 0.96 T 1.30 for the classical XY model are≤ used≤ for ≤ ≤ model. In Fig. 3, examples of the snapshots of g˜i(L/2) the training data. of two models are displayed. At high temperatures{ above} TBKT ((b), (d)), both improved configurations represent the behavior of finite correlation length. At low temper- mate for the infinite system, TBKT =0.342 [20], although atures below TBKT ((a), (c)), they are different from the this temperature becomes lower as the system size in- high-temperature configurations and at the same time creases. We also tested the classification of the quan- they are different from the behavior of the ordered state, tum XY model using the training data of the classical which was shown in Fig. 1(c). (Note that the precise esti- model. For the classical model, not only the classical XY mate of the BKT temperature of the classical XY model model (plane rotator) but also the anisotropic Heisenberg is TBKT = 0.8929 [29].) model with XY interaction was treated. This anisotropic The classification of the BKT and paramagnetic phases Heisenberg model has out-of-plane fluctuation and the of the spin 1/2 XY model using the machine-learning BKT transition temperature is slightly lowered at around technique is shown in Fig. 4(a). The samples of T within TBKT = 0.70 [30, 31]. In Fig. 4(b), we show the result the ranges 0.16 T 0.32 and 0.38 T 0.54 were of the classification of the quantum XY model using the used for the BKT-temperature≤ ≤ and≤ high-temperature≤ training data of the classical XY model (plane rotator). training data, respectively. If we estimate the value of We reproduced the BKT transition of the quantum XY TBKT from the point that the probabilities of predicting model. The same conclusion was obtained when using two phases are 50%, this temperature becomes around the anisotropic Heisenberg model as the training data. T = 0.40. It is slightly higher than the precise esti- The classification into two phases is slightly sharper for 5 the anisotropic Heisenberg model than the classical XY It is not trivial whether loop clusters in quantum spin model (plane rotator). The opposite direction, using the systems can be identified with FK clusters in classical training data of the quantum model in the classification spin systems [32]. In this study, we clarified that the of the classical models, was also successful. phase classification using machine learning does not dis- criminate between loop clusters and FK clusters. The To summarize, we have proposed a method to use the BKT transition of the present study is a thermal phase improved estimator of the correlation configuration in the transition. The investigation of a quantum phase transi- machine-learning study of the phase classification of spin tion at T = 0 will be interesting. For future studies, we models. For the classical spin systems, we have demon- may list up several models for spin and charge degrees strated the machine-learning studies of the 2D 3-state of freedom with loop algorithms. Examples are several Potts model (the second-order transition) and the 2D quantum spin models, strongly-correlated electron mod- 6-state clock model (the BKT transition). The results els, hard-core boson models, optical lattices, etc. were compared with those of the previous study [3] using Another direction of the future study is related to the the ordinary correlation instead of the improved correla- inverse renormalization group approach [33]. Efthymiou tion. The method was also applied to the quantum Monte et al. [34] have proposed a method to increase the size Carlo simulation using the loop algorithm. We treated of lattice spin configuration using super-resolution, deep the spin 1/2 quantum XY model, and analyzed the BKT convolutional neural networks. At high temperatures, transition of the model. We emphasize that the classifi- however, there is a problem that the noise is largely ran- cation scheme based on the training data of the classical dom and difficult to learn. The present improved corre- XY model can be used for the phase classification of the lation configuration could reduce this difficulty at high quantum model. It indicates the universality of the phase temperatures. transition, and at the same time, the generalized feature of the phase classification based on the machine learn- The authors thank Hiroyuki Mori for valuable dis- ing. We also point out the effectiveness of the improved cussions. This work was supported by a Grant-in- estimators in the loop algorithm to bridge classical and Aid for Scientific Research from the Japan Society for quantum Monte Carlo simulations. the Promotion of Science Grant Number JP16K05480, JP16K05482. KS is grateful to the A*STAR (Agency for We have opened a door to using the improved estima- Science, Technology and Research) Research Attachment tors for the machine-learning study of quantum systems. Programme (ARAP) of Singapore for financial support.
[1] G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, Society of Japan 73, 1379 (2004). N. Tishby, L. Vogt-Maranto, and L. Zdeborov´a, Rev. [19] J. Gubernatis, N. Kawashima, and P. Werner, Quan- Mod. Phys. 91, 045002 (2019). tum Monte Carlo Methods: Algorithms for Lattice Mod- [2] J. Carrasquilla and R. G. Melko, Nature Physics 13, 431 els (Cambridge University Press, 2016). (2017). [20] K. Harada and N. Kawashima, Phys. Rev. B 55, R11949 [3] K. Shiina, H. Mori, Y. Okabe, and H. K. Lee, Scientific (1997). Reports 10, 2177 (2020). [21] Y. Tomita and Y. Okabe, Phys. Rev. B 66, 180401(R) [4] V. L. Berezinskii, Sov. Phys. JEPT 32, 493 (1971). (2002). [5] V. L. Berezinskii, Sov. Phys. JEPT 34, 610 (1972). [22] See Supplemental Material at [URL will be inserted by [6] J. M. Kosterlitz and D. J. Thouless, Journal of Physics publisher] for animations of the spin configuration, the C: Solid State Physics 6, 1181 (1973). correlation configuration, and the improved correlation [7] J. M. Kosterlitz, Journal of Physics C: Solid State Physics configuration for the 2D Ising model (mp4 files). 7, 1046 (1974). [23] M. Abadi et al., https://tensorflow.org, [8] R. H. Swendsen and J.-S. Wang, Phys. Rev. Lett. 58, 86 arXiv:1603.04467 [cs.DC]. (1987). [24] D. P. Kingma and J. Ba, arXiv:1412.6980 [cs.LG]. [9] U. Wolff, Phys. Rev. Lett. 62, 361 (1989). [25] T. Surungan, S. Masuda, Y. Komura, and Y. Okabe, [10] P. Kasteleyn and C. Fortuin, J. Phys. Soc. Jpn. Suppl. Journal of Physics A: Mathematical and Theoretical 52, 26, 11 (1969). 275002 (2019). [11] C. Fortuin and P. Kasteleyn, Physica 57, 536 (1972). [26] R. Brower, S. Chandrasekharan, and U.-J. Wiese, Phys- [12] U. Wolff, Nuclear Physics B 322, 759 (1989). ica A: Statistical Mechanics and its Applications 261, [13] U. Wolff, Nuclear Physics B 334, 581 (1990). 520 (1998). [14] Y. Tomita and Y. Okabe, Phys. Rev. Lett. 86, 572 (2001). [27] A. W. Sandvik and C. J. Hamer, Phys. Rev. B 60, 6588 [15] H. G. Evertz, G. Lana, and M. Marcu, Phys. Rev. Lett. (1999). 70, 875 (1993). [28] H.-Q. Lin, J. S. Flynn, and D. D. Betts, Phys. Rev. B [16] H. G. Evertz, Advances in Physics 52, 1 (2003). 64, 214411 (2001). [17] B. B. Beard and U.-J. Wiese, Phys. Rev. Lett. 77, 5130 [29] M. Hasenbusch, Journal of Physics A: Mathematical and (1996). General 38, 5869 (2005). [18] N. Kawashima and K. Harada, Journal of the Physical [30] H. G. Evertz and D. P. Landau, Phys. Rev. B 54, 12302 6
(1996). (2020). [31] T. Figueiredo, J. Rocha, and B. Costa, Physica A: Sta- [33] D. Ron, R. H. Swendsen, and A. Brandt, Phys. Rev. tistical Mechanics and its Applications 488, 121 (2017). Lett. 89, 275701 (2002). [32] Y. Nonomura and Y. Tomita, Phys. Rev. E 101, 032105 [34] S. Efthymiou, M. J. S. Beach, and R. G. Melko, Phys. Rev. B 99, 075113 (2019).