Variational Monte Carlo Renormalization Group
Yantao Wu
A Dissertation Presented to the Faculty of Princeton University in Candidacy for the Degree of Doctor of Philosophy
Recommended for Acceptance by the Department of Physics Adviser: Roberto Car
May 2021 © Copyright by Yantao Wu, 2021. All rights reserved. Abstract
The renormalization group is an important method to understand the critical behaviors of a statistical system. In this thesis, we develop a stochastic method to perform the renormalization group calculations non-perturbatively with Monte Carlo simulations. The method is variational in nature, and involves minimizing a convex functional of the renormalized Hamiltonians. The variational scheme overcomes critical slowing down, by means of a bias potential that renders the coarse-grained variables uncorrelated. When quenched disorder is present in the statistical system, the method gives access to the flow of the renormalized Hamiltonian distribution, from which one can compute the critical exponents if the correlations of the renormalized couplings retain finite range. The bias potential again reduces dramatically the Monte Carlo relaxation time in large disordered systems. With this method, we also demonstrate how to extract the higher-order geometrical information of the critical manifold of a system, such as its tangent space and curvature. The success of such computations attests to the existence and robustness of the renormalization group fixed-point Hamiltonians. In the end, we extend the method to continuous-time quantum Monte Carlo simulations, which allows an accurate determination of the sound velocity of the quantum system at criticality. In addition, a lattice energy-stress tensor emerges naturally, where the continuous imaginary-time direction serves as a ruler of the length scale of the system.
iii Acknowledgements
The life in Princeton has been a long, rewarding, and fulfilling time. I would like to first thank my advisor, Roberto Car, for his guidance and mentorship. To me, he is a passionate, open-minded, and thoughtful advisor. He is interested and extremely knowledgeable in many areas of physics, and has shaped my interests as a young researcher. I am often moved by his passion, energy, extreme focus, and child-like curiosity of the unknowns. It has been a privilege and absolute pleasure to have worked with him. I would also like to thank members of the Car group and the Selloni group with whom I have spent countless hours in Frick: Hsin-Yu Ko, Linfeng Zhang, Fausto Martelli, Biswajit Santra, Marcos Calegari Andrade, Bo Wen, Xunhua Zhao, Yixiao Chen, Clarissa Ding, and Bingjia Yang. I want to give my special thanks to Hsin-Yu, Linfeng, Marcos, and Yixiao for stimulating scientific discussions. I would like to thank my friends and classmates in the physics department for friendship and the bits of physics that I have learned from them here and there: Jiaqi Jiang, Xinran Li, Junyi Zhang, Sihang Liang, Xiaowen Chen, Xue Song, Jingyu Luo, Huan He, Jie Wang, Zheng Ma, Zhenbin Yang, Yonglong Xie, Zhaoqi Leng, Yunqin Zheng, Jingjing Lin, Jun Xiong, Bin Xu, Erfu Su, and Wentao Fan. I would like to thank professor Sal Torquato for stimulating discussions and reading this thesis. I would like to thank professor David Huse for advices in research and serving in my committee. I thank professor Robert Austin for guiding me through the experimental project in biophysics. I thank Kate Brosowsky for being a patient helper as I navigate through the graduate student life in the physics department. I also would like to thank professor Lin Lin for hosting me in Berkeley, and professor Lin Wang for hosting me in Beijing. Life in Princeton would not have been this enjoyable without the many friends that I have made during graduate school. In particular, I would like to thank the following iv people for banters, encouragement, and lots of fun: Tianhan Zhang, Kaichen Gu, Lunyang Huang, Wenjie Su, Youcun Song, Wenxuan Zhang, Yezhezi Zhang, Yi Zhang, Yinuo Zhang, Pengning Chao, Lintong Li, Fan Chen, and the Princeton Chinese soccer team. I also thank the Terascale Infrastructure for Groundbreaking Research in Science and Engineering (TIGRESS) High-Performance Computing Center and Visualization Laboratory at Princeton University. In the end, I would like to thank my family and, in particular, my mom for their unconditional support, without whom none of this would be possible. I have always come to my mom for encouragement and advice in the lows of life. From her, I have learned sharing, humor, and never giving up. I dedicate this thesis to her.
v To my mom.
vi Contents
Abstract...... iii Acknowledgements...... iv List of Tables...... x List of Figures...... xiv
1 Introduction1
2 Variational approach to Monte Carlo renormalization group7 2.1 An introduction to real-space renormalization group...... 7 2.1.1 Scaling operators and critical exponents...... 10 2.2 Variational principle of the renormalized Hamiltonian...... 12 2.3 Monte Carlo sampling of the biased ensemble...... 15 2.4 Results of the renormalized couplings constants...... 16 2.5 Critical exponents in variational Monte Carlo renormalization group. 18 2.6 Appendix...... 21
2.6.1 The coupling terms Sα ...... 21
3 Tangent space and curvature to the critical manifold of statistical system 23 3.1 The critical manifold of a statistical system...... 23 3.2 The critical manifold tangent space...... 25
vii 3.2.1 Critical manifold tangent space in the absence of marginal operators...... 25 3.2.2 Critical manifold tangent space in the presence of marginal operators...... 26 3.2.3 The Normal Vectors to Critical Manifold Tangent Space... 27 3.3 Numerical results for CMTS...... 28 3.3.1 2D Isotropic Ising model...... 28 3.3.2 3D Istropic Ising Model...... 29 3.3.3 2D Anistropic Ising Model...... 31 3.3.4 2D Tricritical Ising Model...... 32 3.4 Curvature of the critical manifold...... 34 3.5 Tangent space to the manifold of critical classical Hamiltonians repre- sentable by tensor networks...... 36 3.5.1 Monte Carlo renormalization group with tensor networks... 37 3.5.2 Numerical Results...... 40
4 Variational Monte Carlo renormalization group for systems with quenched disorder 47 4.1 The renormalization group of statistical systems with quenched disorder 48 4.2 Numerical results...... 53 4.2.1 2D dilute Ising model...... 53 4.2.2 1D transverse field Ising model...... 55 4.2.3 2D spin glass...... 56 4.2.4 2D random field Ising model...... 58 4.2.5 3D random field Ising model...... 59 4.3 Time correlation functions in the biased ensemble...... 61 4.4 Appendix...... 61 4.4.1 Optimization details for the 2D DIM...... 61 viii 4.4.2 Couplings in the computation of critical exponents of the dilute Ising model...... 64 4.4.3 Optimization details for the 3D RFIM...... 65
5 Variational Monte Carlo renormalization group for quantum sys- tems 68 5.1 Continuous-time Monte Carlo simulation of a quantum system.... 68 5.2 The MCRG for continuous-time quantum Monte Carlo...... 71 5.3 The sound velocity of a critical quantum system...... 72 5.3.1 Q = 2: The Ising model...... 74 5.3.2 Q = 3 and 4 ...... 76 5.4 The energy-stress tensor...... 78 5.5 Appendix...... 81 5.5.1 Spacetime correlator at large distance and low temperature.. 81
6 Outlook 83
Bibliography 85
ix List of Tables
0 2.1 Leading even (e) and odd (o) eigenvalues of ∂Kα at the approximate ∂Kβ fixed point found with VRG, in both the unbiased and biased ensembles. The number in parentheses is the statistical uncertainty on the last digit, obtained from the standard error of 16 independent runs. 13 (5) coupling terms are used for even (odd) interactions. See Sec. 2.6.1 for a detailed description of the coupling terms. The calculations used 106 MC sweeps for the 45 × 45 and 90 × 90 lattices, and 5 × 105 sweeps for the 300 × 300 lattice...... 20
3.1 Pαβ for the isotropic Ising model. α indexes rows corresponding to
the three renormalized constants: nn, nnn, and . The fourth row of the table at the Onsager point shows the exact values. β = 2, 3, and 4 respectively indexes the component of the normal vector to
CMTS corresponding to coupling terms nnn, , and nnnn. β = 1
corresponds to the nn coupling term and Pα1 is always 1 by definition. The simulations were performed on 16 cores independently, each of which ran 3 × 106 Metropolis MC sweeps. The standard errors are cited as the statistical uncertainty...... 30
x 3.2 Pαβ for the odd coupling space of the isotropic Ising model. α indexes rows corresponding to the four renormalized odd spin products: (0, 0), (0, 0)-(0,1)-(1,0), (0, 0)-(1, 0)-(-1,0) and (0, 0)-(1,1)-(-1,-1), where the pair (i, j) is the coordinate of an Ising spin. The simulations were performed on 16 cores independently, each of which ran 3 × 106 Metropolis MC sweeps. The standard errors are cited as the statistical uncertainty...... 30
3.3 Pαβ for the 3D isotropic Ising model. The two rows in the table correspond to the two different α which respectively index the nn and the nnn renormalized constants. β runs from 1 to 8, corresponding to
(0) the following spin products, Sβ (σ): (0, 0, 0)-(1, 0, 0), (0, 0, 0)-(1, 1, 0), (0, 0, 0)-(2, 0, 0), (0, 0, 0)-(2, 1, 0), (0, 0, 0)-(1, 0, 0)-(0, 1, 0)-(0, 0, 1), (0, 0, 0)-(1, 0, 0)-(0, 1, 0)-(1, 1, 0), (0, 0, 0)-(2, 1, 1), and (0, 0, 0)-(1, 1, 1), where the triplet (i, j, k) is the coordinate of an Ising spin. 16 independent simulations were run, each of which took 3 × 105 Metropolis MC sweeps. The simulations were performed at the
nearest-neighbor critical point with Knn = 0.22165...... 31
3.4 Pαβ for the 2D anisotropic Ising model. α indexes rows corresponding
to the four renormalized constants: nnx, nny, nnn, and . β = 2 − 6 respectively indexes the component of the normal vector to CMTS
corresponding to coupling terms nny, nnn, , nnnnx, and nnnny. β = 1
corresponds to the nnx coupling term and Pα1 is always 1 by definition. 32 3.5 The couplings used in the computation of CMTS for the 2D tricritical Ising model...... 33
3.6 Pαβ for the 2D tricritical Ising model. α indexes rows corresponding to the first five renormalized couplings listed in Table 3.5, which also gives the couplings for β = 2 − 6...... 33
xi 3.7 au1 + bu2 computed from Table 3.6 for α = 3 − 5 and β = 2 − 6... 34
3.8 κβη at the same three critical points as in Table 3.1, calculated from
2 (n) (0) (0) ∂ Knn /∂Kβ ∂Kη . The exact curvature for β = nn and η = nnn at the Onsager point is also shown [28]...... 35 3.9 The tensor elements which are related to one another by symmetry.. 41 (n,0) Aαβ 2 3.10 The matrix Pαβ = (n,0) for the isotropic 2D square Ising model. A 256 Aα1 lattice was used with the renormalization level n = 5. The simulations were performed on 16 cores independently, each of which ran 3 × 106 Metropolis MC sweeps. The mean is cited as the result and twice the standard error as the statistical uncertainty...... 42 3.11 The tensor elements which belong to distinct symmetry classes. Only one representative of each class is listed...... 44 (n,0) Aαβ 3.12 The matrix Pαβ = (n,0) for the isotropic 2D square three-state Potts Aα1 model. A 2562 lattice was used with the renormalization level n = 5. The simulations were performed on 16 cores independently, each of which ran 9 × 105 Metropolis MC sweeps. The mean is cited as the result and the standard error as the statistical uncertainty...... 45 3.13 The tensor elements which belong to distinct symmetry classes. Only one representative of each class is listed...... 46
5.1 The couplings used for d = 1 TFIM. Note that when α = 2 and 3, the couplings are themselves isotropic between space and time...... 75 5.2 The renormalized constants for the d = 1 TFIM. For each n, L = P = 8 × 2n. VMCRG is done with 4000 variational steps. During each variaional step, the MC sampling is done on 16 cores in parallel, where each core does MC sampling of 20000 Wolff steps. The optimization step is µ = 0.001. The number in the paranthesis is the uncertainty on the last digit...... 75 xii (4) 5.3 The renormalized constants for the d = 2 TFIM. L = P = 128. K1,x and (4) K1,yz are respectively the renormalized nearest neighbor spin constants along the time and the space direction at n = 4...... 76 5.4 The couplings used for d = 1, Q = 3 and 4 Potts model. Note that when α = 2, the coupling is itself isotropic between space and time.. 76 5.5 The renormalized constants for the d = 1, Q = 3 Potts model. For each n, L = P = 8 × 2n. When n = 0 to 3, the simulations are done with Metropolis local updates with 1000 variational steps. When n = 0, 1, 2, each variational step uses 100 sweeps of MC averaging in parallel on 8 cores. When n = 3, each variational step uses 500 sweeps. For n = 4 and 5, the simulation details are the same as in Table 5.2...... 77 5.6 The renormalized constants for the d = 1, Q = 4 Potts model. For each n, L = P = 8 × 2n. The simulation details are the same as in Table 5.5. 78
xiii List of Figures
2.1 Variation of the renormalized coupling constants over five VMCRG iterations on a 300 × 300 lattice. Each iteration has 1240 variational steps, each consisting of 20 MC sweeps. 16 multiple walkers are used for the ensemble averages in Eqs. 2.13 and 2.14. For clarity, we only show the four largest renormalized couplings after the first iteration. Top:
Simulation starting with Knn = 0.4365. Bottom: Simulation starting
with Knn = 0.4355...... 18
2.2 (color online). Time correlation of the estimator A = S0(σ)S0(µ) on
45 × 45 and 90 × 90 lattices (Eq. 5.11). S0 is the nearest neighbor term in the simulations of Table 2.1...... 21
3.1 Left: part of a tensor network representing a 2D classical system. ijkl... represent tensor indices. Right: a single tensor in the network. Its four
tensor indices are labeled as i0i1i2i3. A grey circle represents a lattice site, or equivalently a tensor index. A green box represents a tensor.. 38 3.2 Optimization trajectory of the tensor network renormalized constants for the three-state Potts model on a 162 lattice at K = 1.005053. All 81 renormalized constants are independently optimized and shown. Each curve represents one coupling term...... 44 3.3 The tensor in cubic-lattice tensor network. It is associated with 8 spins. 45
xiv 0 4.1 Distribution of Knn for a DIM with KDI = 0.60 (left), 0.609377 (middle), and 0.62 (right). n denotes RG iteration. All figures have the same scale...... 55
4.2 Distribution of Knny for the Trotter approximation of the TFIM with
KTFIM = 0.935 (left), 1.035 (middle), and 1.135 (right)...... 56 4.3 Distribution of nearest neighbor couplings for a spin glass model with
KSG = 1.2...... 58 4.4 Distribution of the renormalized nearest neighbor (left) and local magne-
tization (right) coupling constants for the 2D RFIM with KRFIM = 0.8
and h0 = 1...... 59 4.5 Distribution of the renormalized coupling constants for the 3D RFIM
with h0 = 0.35, and KRFIM = 0.25 (left), 0.264 (middle), and 0.28 (right). The top row is for the nearnest neighbor couplings. The middle row is for the magnetization couplings. The bottom row is for the next nearest neighbor couplings...... 60
4.6 Distribution of Knn for the 3D random field Ising model with h0 = 0.35
and KRFIM = 0.255, 0.257, 0.262, 0.266, 0.27 from left to right. The data are collected from one sample with L = 64...... 61
hM(t)M(0)i 4.7 Time correlation functions CM (t) = hM 2i of the system magneti- 1 P zation M = N i σi, in the biased and unbiased ensemble for various lattice sizes. For the spin glass model, we take the magnetization to
be in the ρ-system. In the biased ensemble, the bias potential VJmin is obtained by VMCRG for b = 2 with 1 RG iteration. The data are collected, shown from top to bottom, for the dilute Ising model with
KDI = KDI,c, the Trotter approximation of the transverse field Ising
model with KTFIM = 1.0, the 2D spin glass with KSG = 1.2, and the
2D random field Ising model with KRFIM = 0.8, h0 = 1.0,...... 62
xv 4.8 Time correlation functions CM (t) in the biased and unbiased ensemble for various lattice sizes for the 3D random field Ising model with
KRFIM = 0.27, h0 = 0.35. The relaxation for the 3D RFIM in the unbiased ensemble is very slow, and an inset is placed in the center of the figure to show the unbiased time correlation functions more clearly. Also note that in the biased ensemble of the 3D RFIM, the time correlation function eventually converges to the average magnetization squared, hMi2, which in the random field Ising model is not necessarily zero due to the random fields...... 63 4.9 Renormalized nearest neighbor constants at criticality during the opti- mization of Ω[V ] for one bond realization of a 128 × 128 dilute Ising
model, for the first 4 RG iterations. KDI = 0.609377...... 63 4.10 Renormalized nearest neighbor constants at criticality during the opti- mization of Ω[V ] for one bond realization of a 256 × 256 dilute Ising
model for the 5th RG iteration. KDI = 0.609377...... 64 4.11 Renormalized constants for the local magnetization coupling during the optimization of Ω[V ] for one bond realization of a 643 random field
Ising model, for the first 3 RG iterations. KRFIM = 0.264, h0 = 0.35 . 67 4.12 Renormalized constants for the nearest neighbor product coupling during the optimization of Ω[V ] for one bond realization of a 643 random
field Ising model, for the first 3 RG iterations. KRFIM = 0.264, h0 = 0.35 67
xvi Chapter 1
Introduction
This dissertation describes a variational method based on Monte Carlo (MC) sampling to carry out the renormalization group (RG) calculations non-perturbatively on lattice spin systems. We name this method the variational Monte Carlo renormalization group (VMCRG). RG theory originated in the context of quantum field theory to deal with the ultraviolet divergences of a continuum theory [1]. Its connection to statistical physics and critical phenomena was first realized and articulated by Wilson [2,3]. In a field theory where an ultraviolet momentum cutoff is required to make the theory well- defined, the precise value of the ultraviolet cutoff often does not affect the field theory’s prediction of physical results, such as a scattering amplitude. The independence of the physical observables on the cutoff values originates from the fact that the physical observables are often long-wave-length and low-energy quantities, while the cutoff value corresponds to the short-wave-length and high-energy details of a theory. The freedom in choosing any cutoff values gives a field theory predictive power. The momentum cutoff corresponds to the lattice spacing of a statistical system, and the freedom in choosing any lattice spacing is also present there. In both a field theory and a statistical system, the way by which to implement this freedom is the renormalization group.
1 First realized by Wilson, the implementation of this freedom itself gives non-trivial information on the macroscopic behavior of a statistical system. The experimental motivation for RG in statistical physics is critical phenomena. In the 1960s, it was discovered, both in experiments and in exactly solvable models, that at phase transitions, the thermodynamic variables often exhibit power-law singularities. The exponents of these power-law singularities were found to be very close in seemingly very different systems. This gave the notion of universality. In addition, it was found that many critical exponents satisfy very simple scaling relations among each other, leaving only a small number, most commonly two or three, of them independent. This pushed people’s attention to scaling relations. Motivated by these phenomenologies, Wilson was able to come up with RG to understand them at a deeper level. The Wilsonian RG, in a sense, is a theory of theories. The many microscopic theories are defined by different lattice spacings or microscopic Hamiltonians. An RG calculation computes how one theory gets transformed into another theory under a scale transformation. When a statistical system is posed at criticality, signaled by thermodynamic quantities becoming singular, invariance emerges when theories are transformed by scale dilations, i.e. the RG calculation exhibits a fixed-point. To explicitly carry out an RG calculation, however, is difficult. While Wilson spelled out the general procedure to carry out an RG calculation, his early calculations [2] were based on the diagrammatic expansions of partition functions, and were thus perturbative. Since Wilson’s seminal work, there has been strong interest in methods to compute the renormalized coupling constants and the critical exponents in a non-perturbative fashion. This goal has been achieved with the Monte Carlo renormalization group (MCRG) approach of Swendsen. In 1979, he introduced a method to compute the critical exponents, which did not require explicit knowledge of the renormalized Hamiltonian [4]. A few years later, he solved the problem of calculating the renormalized coupling constants for the Ising model, using
2 an equality due to Callen [5] to write the correlation functions in a form explicitly depending on the couplings. Swendsen was able to convert doing an RG calculation to measuring certain correlation functions in the system, which the Monte Carlo simulation was ready to do. A serious limitation of Swendsen’s MCRG is that the MC simulations can often be very difficult. In pure systems, i.e. systems with no quenched disorder, the MC simulations suffers from critical slowing down at criticality. In systems with quenched disorder, the MC simulations face challenges such as frustration in spin glasses and large field pinning down in random field models, even not at criticality. Inspired by the enhanced sampling techniques [6] in free energy sampling, we developed a variational approach to MCRG that alleviates these sampling difficulties greatly. In addition, for disordered systems, one needs the evolution of the distribution of the renormalized coupling constants, which requires their explicit computation. In our approach, the coupling constants are explicitly obtained by minimizing a certain variational principle. The critical exponents also derive from the same principle and Swendsen’s formulae emerge as a special case. Here we provide summaries of the individual chapters of this dissertation. In Chapter2, we first review the variational functional whose utility was initially demon- strated in free energy sampling. We then formulate it in a form that is useful to MCRG. We will identify the coarse-grained spins in RG as the order parameters whose Landau free energy profiles are often computed through statistical sampling [6]. The renormalized Hamiltonian will be numerically obtained through an MC simulation. A truncation error will occur and will be discussed in detail. The critical couplings of a system can be obtained through looking for the couplings that render the renormalized Hamiltonian of the system to flow into a non-trivial fixed-point Hamiltonian. The Jacobian matrix of the RG transformation will also be obtained in the same MC simulation, which, if computed at the fixed-point Hamiltonian, allows one to obtain
3 the critical exponents of a statistical system. We will demonstrate how the critical slowing down is greatly reduced in the variational MC simulation. The work in this chapter has been published previously as the following refereed journal article:
• Yantao Wu and Roberto Car. “Variational Approach to Monte Carlo Renor- malization Group”. In: Phys. Rev. Lett. 119 (22 2017), p. 220602. doi: 10.1103/PhysRevLett.119.220602.
In Chapter3, we show how to use VMCRG to extract higher order information of the critical manifold of a statistical system. Here, the zeroth order information of the critical manifold is the location of the critical couplings, and the first order information is the tangent space of the critical manifold, and so forth. We pay special attention to the Jacobian matrix of the RG transformation. We will show that due to the reduction in critical slowing down, one is able to sample the entire Jacobian matrix well, for the models considered. The kernel of this matrix gives the tangent space of the critical manifold a statistical system. This determination will be free of truncation error. The curvature of the critical manifold can also be accessed with higher-order correlation functions in the MC and thus more statistical noise. The work in this chapter has been published previously as the following refereed journal article:
• Yantao Wu and Roberto Car. “Determination of the critical manifold tangent space and curvature with Monte Carlo renormalization group”. In: Phys. Rev. E 100 (2 2019), p. 022138. doi: 10.1103/PhysRevE.100.022138.
• Yantao Wu. “Tangent space to the manifold of critical classical Hamiltonians representable by tensor networks”. In: Phys. Rev. E 100 (2 2019), p. 023306. doi: 10.1103/PhysRevE.100.023306.
In Chapter4, we show how to use VMCRG to study statistical systems with quenched disorder. In disordered systems, one focuses on the renormalization of 4 the distribution of quenched Hamiltonians, and the scaling variables parametrize the renormalized distributions. We show how to extract these scaling information from VMCRG. We will also demonstrate that sampling difficulties associated with the quenched disorder can be greatly alleviated by VMCRG. The work in this chapter has been published previously as the following refereed journal article:
• Yantao Wu and Roberto Car. “Monte Carlo Renormalization Group for Classical Lattice Models with Quenched Disorder”. In: Phys. Rev. Lett. 125 (19 2020), p. 190601. doi: 10.1103/PhysRevLett.125.190601.
In Chapter5, we show how to use VMCRG to study quantum mechanical system simulatable by a continuous-time MC. One maps the partition function of a quantum system into its continuous-time path-integral representation. The quantity that can be extracted very accurately is the sound velocity of a critical quantum system with conformal symmetry. In addition, one can use the continuous nature of the time dimension to determine the energy stress tensor of a statistical system on a discrete lattice. The work in this chapter has been published previously as the following refereed journal article:
• Yantao Wu and Roberto Car. “Continuous-time Monte Carlo renormalization group”. In: Phys. Rev. B 102 (1 2020), p. 014456. doi: 10.1103/Phys- RevB.102.014456.
In addition to the work mentioned above, VMCRG has also indirectly inspired the following research work during this Ph.D:
• Yantao Wu and Roberto Car. “Quantum momentum distribution and quantum entanglement in the deep tunneling regime”. In: The Journal of Chemical Physics 152.2 (2020), p. 024106. doi: 10.1063/1.5133053.
5 • Yantao Wu. “Nonequilibrium renormalization group fixed points of the quantum clock chain and the quantum Potts chain”. In: Phys. Rev. B 101 (1 2020), p. 014305. doi: 10.1103/PhysRevB.101.014305.
• Yantao Wu. “Dynamical quantum phase transitions of quantum spin chains with a Loschmidt-rate critical exponent equal to 12 ”. In: Phys. Rev. B 101 (6 2020), p. 064427. doi: 10.1103/PhysRevB.101.064427.
• Yantao Wu. “Time-dependent variational principle for mixed matrix product states in the thermodynamic limit”. In: Phys. Rev. B 102 (13 2020), p. 134306. doi: 10.1103/PhysRevB.102.134306.
• Yantao Wu. “Dissipative dynamics in isolated quantum spin chains after a local quench”. In: (2020). In Review. arXiv: 2010.00700 [cond-mat.stat-mech].
6 Chapter 2
Variational approach to Monte Carlo renormalization group
2.1 An introduction to real-space renormalization
group
As mentioned in Chapter1, the theory of RG is not only significant in statistical mechanics, but also in field theory. In field theory, the major goal of RG is to demonstrate the insensitivity of long-wave length quantities on the momentum cutoff. In statistical mechanics, the goal is very different. RG aims to reveal the underlying reason of the singular behavior of thermodynamic variables when a statistical system is fine-tuned to be at the special critical parameters. In doing so, it makes the notion of scale invariance explicit. It is remarkable that while the problems in field theory and in statistical mechanics look so different on the surface, they are only different facets of the same physical structure, i.e. the insensitivity of the low-energy quantities on the high-energy details of the theory in the ultraviolet limit for field theory, and in the thermodynamic limit for statistical mechanics.
7 Here we focus on statistical mechanics and briefly review the theory of real-space RG in its context. The central quantity of a statistical system is its partition function. When at criticality, the partition function depends non-analytically on the system parameters, such as the temperature, pressure, volume, etc. To isolate the non-analytic nature of the system from its analytic background, the RG procedure extracts an analytic part from the partition function in each iteration while keeping its non- analytic part intact. After many iterations of peeling off the analytic parts, the non-analytic part of the partition function will be isolated out and its structure will become apparent. To be concrete, let us consider a statistical mechanical system in d spatial dimen- sions with spins σ and Hamiltonian H(0)(σ),
(0) X (0) H (σ) = Kβ Sβ(σ) (2.1) β
where Sβ(σ) are the coupling terms of the system, such as nearest neighbor spin
(0) (0) products, next nearest neighbor spin products, etc., and K = {Kβ } are the corresponding coupling constants. Here we call the original Hamiltonian before any RG transformation the zeroth level renormalized Hamiltonian, hence the notation (0) in the superscript. RG considers a flux in the space of Hamiltonians (4.1) under scale transformations that reduce the linear size of the original lattice by a factor b, with b > 1. The motivation to consider a scale transformation is in part due to the early solutions of exactly solvable models [7], where it was discovered that the correlation functions at criticality exhibit power-law, instead of exponential, decay with distance. A power-law dependence is essentially scale-less. For example, consider a dimensionless quantity which decays as a power-law as a function of distance r: q(r) = 1/(ar)b = a−b1/rb, the quantity a with dimension inverse distance can always be factored out as a
8 trivial multiplicative factor. This is very different from an exponential function q(r) = exp(−ar) where a cannot be factored out and is intrinsic. Thus a critical statistical system is thought to be scale-invariant 1. In a real-space RG calculation, one defines coarse-grained spins σ0 in the renormalized system with a conditional probability T (σ0|σ) that effects a scale transformation with scale factor b. T (σ0|σ) is the probability of σ0 given spin configuration σ in the original system. The majority rule block spin in the Ising model proposed by Kadanoff [8] is one example of the coarse-grained variables. T (σ0|σ) can be iterated n times to define the nth level coarse-graining T (n)(µ|σ) realizing a scale transformation with scale factor bn:
X X T (n)(µ|σ) = .. T (µ|σ(n−1)) ··· T (σ(1)|σ) (2.2) σ(n−1) σ(1)
T (n) defines the nth level renormalized Hamiltonian H(n)(µ) up to a constant g(K(0)) independent of µ [9]:
X (0) H(n)(µ) ≡ − ln T (n)(µ|σ)e−H (σ) + g(K(0)) σ (2.3) X (n) (0) = Kα Sα(µ) + g(K ) α
(n) where {Kα } are the nth level renormalized coupling constants associated with the
(0) coupling terms Sα(µ) defined for the nth level coarse-grained spins. Here g(K ) is made unique by requiring that the identity coupling term in H(n)(µ) have a coupling constant of zero. g(K(0)) is the analytic-part of the free energy, or equivalently the partition function, mentioned in the second paragraph of this section. Repeated ad infinitum, the RG transformations generate a flux in the space of Hamiltonians, in which all possible coupling terms appear, unless forbidden by symmetry. For example, in an Ising model with no magnetic field, only even spin products appear. The
1A simplified understanding could be simply noting that at criticality, the correlation length diverges, and a scale is missing in the theory 9 space of the coupling terms is, in general, infinite. However, perturbative [3] and non-perturbative [4] calculations suggest that only a finite number of couplings should be sufficient for a given desired degree of accuracy.
2.1.1 Scaling operators and critical exponents
Because the RG procedure generates a flow in the coupling space, there are generally three possibilities for the asymptotic behaviors of the RG flow: i) flows into a fixed- point, ii) forms a periodic orbit, or iii) follows a chaotic trajectory. The last two possibilities do occur [10], but the first possibility is by far the most common scenario in equilibrium phase transitions, and we will focus exclusively on it in this thesis. A statistical system is critical when its RG flows goes into a fixed-point that requires the fine tuning of the system coupling constants for the RG flow to go into. In the simplest case, one only needs to fine-tune one parameter. Let us consider this case first, and denote the critical fixed-point by K∗ 2. Let the controlling coupling constant, for example the temperature, be T and its critical value be Tc. We also define t ≡ T − Tc. If t is different from but sufficiently close to 0, the RG flow will start from K(0), approach K∗, stay around K∗ for a while, and then eventually stray away to one of the non-critical fixed-points. When the RG flow is in the vicinity of K∗, its behavior can be linearized around K∗:
(n+1) (n+1) (n+1) (n+1) ∗ ∂K (n) ∗ ∂K (n) δK ≡ K − K = (n) (K − K ) = (n) δK (2.4) ∂K K∗ ∂K K∗
∂K(n+1) Let the eigenvalues and the eigenvectors of ∂K(n) be λi and φi, ordered by the K∗ (n) descending magnitude of λi. We also define the scaling variables ui as the coordinates 2The critical fixed-point is typically parametrized by an infinite number of coupling constants. But in practice one uses a finite approximations. K∗ is a vector in both cases.
10 (n) of δK in the eigenbasis {φi}:
(n) X (n) (n+1) (n) δK ≡ ui φi, ui = λiui (2.5) i
∗ When |λi| > 1, the RG flow is repulsive against K and φi is called a relevant scaling
∗ operator. When |λi| < 1, the RG flow is attractive to K and φi is called an irrelevant
scaling operator. When |λi| = 1, φi is called a marginal scaling operator.
When t = 0, all the relevant ui must be zero, because otherwise the RG flow will not end up on the critical fixed-point. This means that if one only needs to fine-tune
one parameter for criticality, there must only be one relevant ui. That is, when t = 0,
u1 = 0. If it takes a finite number, n, of RG transformations to bring the system
∗ (n) into the linear regime around Kc , one expects that ui depends analytically on t. (n) In particular, to the leading order of t, u1 = a δt, where a depends on n. Near
criticality, the free energy per site can be separated into a singular part, fs, which
depends non-analytically on t, and a regular part, fg, which depends analytically on t. When a finite number of RG iterations are done, the free energy per site has the following scaling relation [11]
−nd (n) fs(t) = b fs(u1 ) (2.6) −nd (n) −nd (n+1) −nd d (n) fs(λ1t) = b fs(λ1u1 ) = b fs(u1 ) = b b fs(u1 )
Thus,
−d −nd n fs(t) = b fs(λ1t) = b fs(λ1 t) (2.7) which implies that
d/y fs(t) = f±|t| , where y = logb λ1 (2.8)
11 where f+ is for t > 0, and f− for t < 0. Thus, d/y determines all the critical exponents controlled by the critical fixed-point. This explains the connection between the critical exponents and the Jacobian matrix of the RG transformation. Thus, in a numerical calculation of RG, one needs to determine first the critical value of the control parameter, and the RG Jacobian matrix at the critical fixed-point.
2.2 Variational principle of the renormalized Hamil-
tonian
To calculate accurately the renormalized Hamiltonian, H(n), is generally very difficult. Perturbation theory has been very successful [3], but at the same time very difficult and unable to take into non-perturbative effects. See [12] for a heroic effort in applying the perturbation theory to φ4 theory. A numerical but non-perturbative approach is to study the RG flow with MC sampling, where the correlation functions sampled should provide sufficient information to determine the RG flow. In the proximity of a critical point, the coarse-grained spins µ displays a divergent correlation length, originating critical slowing down of local MC updates. This can be avoided by modifying the distribution of the µ by adding to the Hamiltonian H(n)(µ) a biasing potential V (µ) to force the biased distribution of the coarse-grained spins,
pV (µ), to be equal to a chosen target distribution, pt(µ). For instance, pt can be the constant probability distribution. Then the µ have the same probability at each lattice site and act as uncorrelated spins, even in the vicinity of a critical point. It turns out that V (µ) obeys a powerful variational principle that facilitates the sampling of the Landau free energy [6], which follows from the variational principle of the generalized Legendre transformation. In the present context, we define the
12 functional Ω[V ] of the biasing potential V (µ) by:
P −[H(n)(µ)+V (µ)] µ e X Ω[V ] = log + p (µ)V (µ), (2.9) P e−H(n)(µ) t µ µ
where pt(µ) is a normalized known target probability distribution. As demonstrated in [6], the following properties hold:
1. Ω[V ] is a convex functional with a lower bound.
2. The minimizer, Vmin(µ), of Ω is unique up to a constant and is such that:
(n) H (µ) = −Vmin(µ) − log pt(µ) + constant (2.10)
3. The probability distribution of the µ under the action of Vmin is:
(n) e−(H (µ)+Vmin(µ)) pVmin (µ) = (n) = pt(µ) (2.11) P −(H (µ)+Vmin(µ)) µ e
The above three properties lead to the following MCRG scheme.
First, we approximate V (µ) with VJ(µ), a linear combination of a finite number
of terms Sα(µ) with unknown coefficients Jα, forming a vector J = {J1, ..., Jα, ..., Jn}.
X VJ(µ) = JαSα(µ) (2.12) α
Then the functional Ω[V ] becomes a convex function of J, due to the linearity of the
expansion, and the minimizing vector, Jmin, and the corresponding Vmin(µ) can be found with a local minimization algorithm using the gradient and the Hessian of Ω:
∂Ω(J) = −hSα(µ)iVJ + hSα(µ)ipt (2.13) ∂Jα
13 ∂2Ω(J) = hSα(µ)Sβ(µ)iVJ − hSα(µ)iVJ hSβ(µ)iVJ (2.14) ∂Jα∂Jβ
Here h·iVJ is the biased ensemble average under VJ and h·ipt is the ensemble average
under the target probability distribution pt. The first average is associated to the Boltzmann factor exp{−(H(n)(µ) + V (µ))} and can be computed with MC sampling
(see Sec. 2.3). The second average can be computed analytically if pt is simple enough.
h·iVJ always has inherent random noise, or even inaccuracy, and some sophistication is required in the optimization problem. Following [6], we adopt the stochastic optimization procedure of [13], and improve the statistics by running independent MC simulations, called multiple walkers, in parallel. For further details, consult [6].
(n) The renormalized Hamiltonian H (µ) is given by Eq. 2.10 in terms of Vmin(µ).
Taking a constant pt, we have modulo a constant:
(n) X H (µ) = −Vmin(µ) = (−Jmin,α)Sα(µ) (2.15) α
In this finite approximation the renormalized Hamiltonian has exactly the same terms
of Vmin(µ) with renormalized coupling constants
(n) Kα = −Jmin,α. (2.16)
The relative importance of an operator Sα in the renormalized Hamiltonian can be
estimated variationally in terms of the relative magnitude of the coefficient Jmin,α.
When Jmin,α is much smaller than the other components of Jmin, the corresponding
Sα(µ) is comparably unimportant and can be ignored. The accuracy of this approx-
imation could be quantified by measuring the deviation of pVmin (µ) from pt(µ). In
the case of the Ising model, for example, if pt(µ) is the uniform distribution, any
spin correlators should vanish under pt and to determine how close pVmin , one simply
14 measures spin correlators in the fully optimized biased ensemble, and any deviation from zero would indicate an approximation.
2.3 Monte Carlo sampling of the biased ensemble
It is not entirely clear how to sample the observable O(µ) in the biased ensemble, so we explain here the details of the sampling.
P O(µ)e−(H(n)(µ)+V (µ)) hO(µ)i = µ V P e−(H(n)(µ)+V (µ)) µ (2.17) P O(µ)T (n)(µ|σ)e−(H(0)(σ)+V (µ)) = µ,σ P (n) −(H(0)(σ)+V (µ)) µ,σ T (µ|σ)e
Thus, the state space of the MC sampling should the product space of µ and σ: {(µ, σ)}. Let the proposal probability be:
g(µ0, σ0|µ, σ) = g(σ0|σ)T (n)(µ0|σ0) (2.18) with the acceptance probability
g(σ|σ0) A = min 1, e−(∆H+∆V ) (2.19) g(σ0|σ) where ∆H = H(0)(σ0) − H(0)(σ), ∆V = V (µ0) − V (µ). It is easy to prove that this Metropolis MC scheme satisfies the detailed balance. Then if one bases the sampling on the local Metropolis move [14], g(σ|σ0)/g(σ0|σ) = 1, and one uses an acceptance probability of e−(∆H+∆V ). If the MC sampling is based on the Wolff algorithm [15], then the acceptance probability is simply min(1, e−V ), because e−∆H g(σ|σ0)/g(σ0|σ) = 1 in the Wolff algorithm. Note that one can also define the expected value of hO(µ, σ)i according to the ensemble in the second line of Eq. 2.17. 15 2.4 Results of the renormalized couplings constants
There are two ways to carry out the VMCRG to compute the renormalized coupling constants for the nth RG iteration. The first way is to perform the VMCRG with T (1) to obtain H(1) and then use this H(1) as the starting Hamiltonian for another VMCRG calcualtion with T (1) to obtain H(2). This process is repeated n times to obtain H(n). The second way is to perform the VMCRG with T (n) once to obtain H(n). The drawback of the first way is that H(i) is truncated for all i < n, which gives a relatively large truncation error. But the comparative advantage is that in this scheme each VMCRG is done with a small block size b, and the critical slowing down is essentially eliminated. While in the second way the block size is bn and statistical correlation builds up within a block, so critical slowing down is only reduced partially. In this chapter, we focus exclusively on the first way, allowing for some truncation error. In Chapter3, we use the second way. To illustrate the method, we present a study of the Ising model on a 2D square lattice in the absence of a magnetic field. We adopt 3 × 3 block spins with the majority rule. 26 coupling terms were chosen initially, including 13 two-spin and 13 four-spin products. One preliminary iteration of VMCRG was performed on a 45 × 45 lattice starting from the nearest-neighbor Hamiltonian. The coupling terms with renormalized coupling constants smaller than 0.001 in absolute value were deemed unimportant and dropped from further calculations. 13 coupling terms, including 7 two-spin and 6 four-spin products, survived this criterion and were kept in all subsequent calculations. Each calculation consisted of 5 VMCRG iterations starting with nearest-neighbor
coupling, Knn, only. All the subsequent iterations used the same lattice of the initial iteration. Standard Metropolis MC sampling [14] was adopted, and the calculations
16 were done at least twice to ensure that statistical noise did not alter the results significantly.
In Fig. 2.1, results are shown for a 300 × 300 lattice with two initial Knn, equal to
0.4355 and to 0.4365, respectively. When Knn = 0.4365, the renormalized coupling constants increase over the five iterations shown, and would increase more dramatically with further iterations. Similarly, they decrease when Knn = 0.4355. Thus, the critical
coupling Kc should belong to the window 0.4355 − 0.4365. The same critical window is found for the 45 × 45, 90 × 90, 150 × 150, and 210 × 210 lattices. Because each iteration is affected by truncation and finite size errors, less iterations for the same rescaling factor would reduce the error. For example, 4 VMCRG iterations with a 2 × 2 block have the rescaling factor of a 16 × 16 block. The latter is computationally more costly than a calculation with 2 × 2 blocks, but can still be performed with modest computational resources. Indeed, with a 16 × 16 block, RG iterations on a 128 × 128 lattice gave a critical window 0.4394 − 0.4398, very close to the exact value,
Kc ∼ 0.4407, due to Onsager [16]. The statistical uncertainty of the calculated renormalized coupling constants is smaller with the variational method than with the standard method [17]. For example,
using VMCRG and starting with Knn = 0.4365 on a 300 × 300 lattice, we found a renormalized nearest-neighbor coupling equal to 0.38031 ± 0.00002 after one RG iteration with 3.968 × 105 MC sweeps. Under exactly the same conditions (lattice
size, initial Knn, coupling terms and number of MC sweeps) we found instead a renormalized nearest-neighbor coupling equal to 0.3740 ± 0.0003 with the standard method. In the VMCRG calculation we estimated the statistical uncertainty with the block averaging method [18], while we used the standard deviation from 14 independent calculations in the case of the standard method. A small difference in the values of the coupling constants calculated with VMCRG and the standard method is to be
17
✁ ✂✄
✁ ✂✞✄
✁ ✂✞
✁ ✂✝✄
✖ ✗✘ ✙ ✗✚✛ ✖ ✗✜✢✣ ✤✥ ✙
✁ ✂✝
✡
✦✖ ✧ ✖ ✗✘ ✙ ✗✚✛ ✖ ✗✜✢✣ ✤✥ ✙
✆
✁ ✂ ✄
✠
★
✙✧ ✖ ✗✘ ✙✗ ✚ ✛ ✖ ✗✜✢✣ ✤✥ ✙
✆
✁ ✂
☎
✁ ✂ ✄
✩✪
✖ ✗✘ ✙ ✗✚✛ ✚ ✘ ✙✗
☎
✁ ✂
✁ ✂✁✄
✁