Condensed Matter

Many-body localization generalizes the concept of Anderson localization (i.e. single particle localization) to isolated interacting systems, where many-body eigenstates in the presence of sufficiently strong disorder can be localized in a region of Hilbert space even at nonzero temperature. This is an example of ergodicity breaking, which manifests failure of thermalization or more specifically the break down of eigenstate-thermalization hypothesis.

In this talk, I enquire into the quasi many-body localization in topologically ordered states of matter, revolving around the case of Kitaev toric code on the ladder geometry, where different types of anyonic defects carry different masses induced by environmental errors.  Our study verifies that the presence of anyons generates a complex energy landscape solely through braiding statistics, which suffices to suppress the diffusion of defects in such clean, multicomponent anyonic liquid. This nonergodic dynamics suggests a promising scenario for investigation of quasi many-body localization.  Our results unveil how self-generated disorder ameliorates the vulnerability of topological order away from equilibrium. This setting provides a new platform which paves the way toward impeding logical errors by self-localization of anyons in a generic, high energy state, originated exclusively in their exotic statistics.

Three fundamental factors determine the quality of a statistical learning algorithm: expressiveness, generalization and optimization.  The classic strategy for handling these factors is relatively well understood.  In contrast, the radically different approach of deep learning, which in the last few years has revolutionized the world of artificial intelligence, is shrouded by mystery.  This talk will describe a series of works aimed at unraveling some of the mysteries revolving expressiveness, arguably the most prominent factor behind the success of deep learning.  I will begin by showing that state of the art deep learning architectures, such as convolutional networks, can be represented as tensor networks -- a computational model commonly employed in quantum physics.  This connection will inspire the use of quantum entanglement for defining measures of data correlations modeled by deep networks.  Next, I will turn to a quantum max-flow / min-cut theorem characterizing the entanglement captured by tensor networks.  This theorem will give rise to new results that shed light on expressiveness in deep learning, and in addition, provide new tools for deep network design.

Works covered in the talk were in collaboration with Yoav Levine, Or Sharir, David Yakira and Amnon Shashua.

Remarkable recent experiments have observed Mott insulating behavior and superconductivity in moire superlattices of twisted bilayer graphene near a magic twist angle. However, the nature of the Mott insulator, origin of superconductivity and an effective model remain to be determined. I will present our understanding of these phenomena. We propose a Mott insulator with intervalley coherence that spontaneously breaks U(1) valley symmetry, and describe a mechanism that selects this order over the competing magnetically ordered states favored by the Hunds coupling. We also identify symmetry related features of the nearly flat bands that are key to understanding the strong correlation physics and constrain any lattice model description. First, although the charge density is concentrated on the triangular lattice sites of the moire pattern, the Wannier states of a tight-binding model must be centered on different sites which form a honeycomb lattice. Next, spatially localizing electrons derived from the nearly flat bands necessarily breaks valley and other symmetries within any mean-field treatment, which is suggestive of a valley-ordered Mott state, and also dictates that additional symmetry breaking is present to remove symmetry-enforced band contacts. Tight-binding models describing the nearly flat mini-band are derived, which highlight the importance of further neighbor hopping and interactions. We discuss the consequences of this picture.

Recently, a web of quantum field theory dualities was proposed linking several problems in the study of strongly correlated quantum critical points and phases in two spatial dimensions. These dualities follow from a relativistic flux attachment duality, which relates a Wilson-Fisher boson with a unit of attached flux to a free Dirac fermion. While several derivations of members of the web of dualities have been presented thus far, none explicitly involve the physics of flux attachment, which in relativistic systems affects both statistics and spin. We discuss how this can be achieved in models of relativistic current loops, where the concept of relativistic flux attachment can be made precise. In this context, we provide simple, explicit “derivations” of members of the web of dualities. We describe some implications of this work for relativistic composite fermion theories arising in condensed matter physics, as well as new possibilities for deriving additional dualities using these techniques.

Recent quench experiments in a quantum simulator of interacting Rydberg atoms demonstrated surprising long-lived, periodic revivals from certain low entanglement states, while apparently quick thermalization from others. Motivated by these findings, I will in this talk analyze the dynamics of a family of kinetically constrained spin models related to the experiments. By introducing a manifold of locally entangled spins, representable by a low-bond dimension matrix product state (MPS), I will derive "semiclassical" equations of motion for them. I find that they host isolated, unstable periodic orbits, the presence of which captures the long-lived oscillations and gives rise to slow relaxation of local observables from certain initial configurations. This thus represents a form of weak breaking of ergodicity in dynamics. Our results are reminiscent of the phenomenon of quantum scarring in single-particle chaotic systems which is rooted in classical unstable periodic orbits, and complement the explanation of the recurrences given by [Nature Physics (2018), doi:10.1038/s41567-018-0137-5], in terms of motion over special nonergodic many-body eigenstates, suggestively dubbed `quantum many-body scars'. 

We demonstrate that extremely long range correlations may develop in systems that start from equilibrium and are then rapidly cooled (or driven in other ways). Amongst other things, these correlations suggest a collapse of the viscosity data of glass formers. This collapse is found to be obeyed over 16 decades of relaxation times in experimental data on all known types of supercooled fluids. 

Over the last decade, there have been enormous gains in machine learning technology primarily driven by neural networks. A major reason neural networks have outperformed older techniques is that the cost of optimizing them scales well with the size of the training dataset. But neural networks have the drawback that they are not very well understood theoretically.

Recent work by several groups has explored an alternative approach to creating machine learning model functions based on tensor networks, which are a technique for parameterizing many-body quantum wavefunctions. The cost of training tensor network models scales similarly to the cost of training neural networks. In addition, their relatively simple, linear structure has provided good theoretical understanding of their properties, and underpins many powerful techniques to optimize and manipulate them. 

After introducing tensor network machine learning models, I will discuss some of the techniques to optimize them and results for supervised and generative machine learning tasks. These techniques can automatically adapt the number of parameters and suggest interesting interpretations and extensions for 'deep' tensor network variants. I will conclude by discussing a recent tensor network based proposal to formulate hybrid quantum-classical algorithms for machine learning with quantum computers.

Time series data contains useful information on the phase of a system. Here we propose the use of recurrent neural networks (LSTM) to learn and extract such information in order to classify phases and locate phase boundaries. We demonstrate this on a many-body localized model, and attempt to interpret the learned behavior by looking at individual LSTM cells. We also discuss the validity of the learned model and investigate its limits.

Recently, we have shown that deep neural networks can be used to solve the Schrödinger Equation (https://journals.aps.org/pra/abstract/10.1103/PhysRevA.96.042113), classical spin models (https://doi.org/10.1103/PhysRevE.97.032119), and 2d-materials such as graphene and boron-nitride (https://doi.org/10.1016/j.commatsci.2018.03.005).

I will argue that these "deep models" can be used to simulate and understand matter in a way which was not previously possible. Specifically, I will show how our recently reported extensive deep neural networks (https://arxiv.org/abs/1708.06686) can be used to infer the properties of meso-scale materials based on training data generated from much smaller structural motifs (evaluated using electronic structure methods such as density functional theory). Extensive deep neural networks scale as O(N) and can be efficiently evaluated in parallel using petascale computational platforms.