A quantum inspired approach to learning dynamical laws from data -- block-sparsity and gauge-mediated weight sharing

TL;DR

This paper introduces a scalable, robust method for learning dynamical laws from data using block-sparse tensor train representations inspired by quantum systems, extending to complex interactions and exploiting structural assumptions.

Contribution

It develops a novel block-sparsity constrained optimization algorithm with gauge-mediated weight sharing for efficient dynamical law recovery, extending tensor train methods to multi-mode interactions.

Findings

Achieves highly accurate recovery of dynamical laws in numerical experiments.

Demonstrates robustness to noise in data.

Improves performance over previous methods.

Abstract

Recent years have witnessed an increased interest in recovering dynamical laws of complex systems in a largely data-driven fashion under meaningful hypotheses. In this work, we propose a scalable and numerically robust method for this task, utilizing efficient block-sparse tensor train representations of dynamical laws, inspired by similar approaches in quantum many-body systems. Low-rank tensor train representations have been previously derived for dynamical laws of one-dimensional systems. We extend this result to efficient representations of systems with $K$-mode interactions and controlled approximations of systems with decaying interactions. We further argue that natural structure assumptions on dynamical laws, such as bounded polynomial degrees, can be exploited in the form of block-sparse support patterns of tensor-train cores. Additional structural similarities between interactions of certain modes can be accounted for by weight sharing within the ansatz. To make use of these structure assumptions, we propose a novel optimization algorithm, block-sparsity restricted alternating least squares with gauge-mediated weight sharing. The algorithm is inspired by similar notions in machine learning and achieves a significant improvement in performance over previous approaches. We demonstrate the performance of the method numerically on three one-dimensional systems -- the Fermi-Pasta-Ulam-Tsingou system, rotating magnetic dipoles and point particles interacting via modified Lennard-Jones potentials, observing a highly accurate and noise-robust recovery.