Core Concepts
This paper establishes the mathematical foundations of the Dynamical Low Rank Approximation (DLRA) method for high-dimensional stochastic differential equations, including a rigorous formulation and analysis of the Dynamically Orthogonal (DO) approximation as a specific parametrization of DLRA.
Abstract
The paper focuses on the theoretical foundations of dynamical low-rank methods for stochastic differential equations (SDEs). It starts by revisiting the Dynamically Orthogonal (DO) equations for SDEs, which were previously derived using the assumption of time differentiability of the solution, a property that SDE solutions do not possess. The authors derive the DO equations without using time derivatives, by pushing the differentiability to the spatial basis while keeping the stochastic basis non-differentiable.
The paper then shows the equivalence between the DO formulation and a parameter-independent DLRA formulation for SDEs. The main result is the proof of local existence and uniqueness of both the DO and DLRA equations under standard assumptions. As part of the existence result, the authors characterize the interval of existence in terms of the linear independence of the stochastic DO basis. They also show that while the DO solution ceases to exist at the explosion time, the DLRA can be continuously extended beyond that point. Finally, the authors provide a sufficient condition for global existence of the DO solution.