Stochastic dynamical low-rank approximation method

Yu Cao, Jianfeng LuJournal of Computational Physics 2018 · DOI

Abstract

In computational physics and quantum mechanics, the "curse of dimensionality" often renders the simulation of high-dimensional stochastic systems intractable. This paper introduces the Stochastic Dynamical Low-Rank Approximation (SDLR) method, a novel framework that combines model reduction with Monte Carlo simulations. By projecting high-dimensional stochastic differential equations (SDEs) onto a moving low-rank manifold, the SDLR method provides a computationally efficient path to solving complex systems, such as the Lindblad equation, while maintaining high accuracy in capturing statistical moments.

Introduction

Modeling the evolution of complex physical systems frequently requires solving high-dimensional partial differential equations (PDEs). Whether analyzing the probability distribution of particles via the Fokker-Planck equation or the density matrix of an open quantum system via the Lindblad equation, the state space grows exponentially with the number of variables.

To mitigate this, researchers often employ Monte Carlo methods, which transform these PDEs into a collection of individual "particles" governed by Stochastic Differential Equations (SDEs). However, even these SDEs can remain high-dimensional and computationally expensive. The SDLR method addresses this challenge by applying low-rank model reduction directly to the stochastic processes, identifying a simplified subspace where the most significant dynamics occur.

Methodology

Tangent Space Projection

The SDLR method extends the deterministic Dynamical Low-Rank Approximation (DLRA) into the stochastic domain. Instead of attempting to solve the full high-dimensional system, the dynamics are restricted to a manifold of low-rank measures. The authors utilize a tangent space projection to ensure that the simplified evolution remains as close as possible to the true high-dimensional trajectory in the $L^2$ sense.

The SDLR Algorithm

The core of the SDLR approach is a coupled system of equations that governs the evolution of the reduced model:

  1. Basis Evolution (ODE): An ordinary differential equation that evolves the coordinate system (the basis) of the low-rank subspace.
  2. Particle Evolution (SDE): A stochastic differential equation that evolves the individual particles within that simplified subspace.

The Itô Correction

A significant contribution of this work is the inclusion of a specific Itô correction term. Unlike previous "Dynamical Orthogonal" (DO) methods, the SDLR framework accounts for the unique geometric interactions between noise and the low-rank manifold. This correction is vital for maintaining accuracy when the system is subjected to stochastic fluctuations.

Theoretical Insights

The Commuting Diagram in Quantum Mechanics

One of the most theoretically robust findings is the "commuting diagram" for quantum systems. The authors demonstrate that for the Lindblad equation, the order of operations is interchangeable: one can either simplify the master equation first and then "unravel" it into SDEs, or unravel it first and then apply the low-rank approximation. Both paths yield the same result, confirming the method's theoretical consistency.

SDLR vs. Dynamical Orthogonal (DO) Methods

While the DO method is a popular choice for tracking the mean (average) of a system, the authors show that SDLR is significantly more accurate at capturing the second moment (variance and correlations). This makes SDLR particularly valuable for applications where quadratic statistics—such as energy or density matrices—are the primary variables of interest.

Numerical Benchmarks

The authors evaluated the SDLR method across three distinct physical models:

  1. Geometric Brownian Motion (20D): A high-dimensional stochastic process used to verify the method's ability to capture decaying energy states.
  2. Stochastic Burgers’ Equation: A classic non-linear SPDE used in fluid dynamics. In this test, SDLR achieved a second-moment error ratio of approximately 0.57 to 0.81 compared to the DO method, representing a substantial improvement in statistical accuracy.
  3. Quantum Damped Harmonic Oscillator: A standard model for open quantum systems. The results showed that a rank-5 approximation was nearly indistinguishable from the full rank-21 simulation, allowing for a reduction of over 75% in system complexity without significant data loss.

Conclusion and Future Work

The SDLR method proves highly effective for "dissipative" systems—those that naturally lose energy or settle into simpler states over time—as their dynamics naturally collapse onto low-rank subspaces. Theorem 2 of the paper further proves that if a system is inherently low-rank, SDLR will recover the exact dynamics perfectly.

However, limitations remain. The method's performance can degrade under conditions of extremely high noise (diffusivity), where particles may be forced out of the low-rank manifold. Additionally, the current implementation requires a fixed rank ($r$) to be chosen a priori. Future research will focus on developing adaptive schemes to adjust the rank dynamically and extending the framework to infinite-dimensional Hilbert spaces to handle SPDEs more naturally.

Summary updated

BibTeX
@article{cao2018stochastic,
  author = {Yu Cao and Jianfeng Lu},
  title = {{Stochastic dynamical low-rank approximation method}},
  journal = {Journal of Computational Physics},
  year = {2018},
  doi = {10.1016/j.jcp.2018.06.058},
}
Figure