A learning scheme by sparse grids and Picard approximations for semilinear parabolic PDEs

TL;DR

This paper introduces a probabilistic neural network-based scheme using sparse grids and Picard iterations to efficiently solve high-dimensional semi-linear parabolic PDEs, effectively mitigating the curse of dimensionality.

Contribution

It develops a new learning algorithm combining sparse grid approximation, Picard iterations, and neural networks, with proven convergence and polynomial complexity growth in high dimensions.

Findings

Convergence of the proposed scheme under smallness conditions.

Polynomial complexity growth in the PDE dimension.

Numerical validation showing competitive performance.

Abstract

Relying on the classical connection between Backward Stochastic Differential Equations (BSDEs) and non-linear parabolic partial differential equations (PDEs), we propose a new probabilistic learning scheme for solving high-dimensional semi-linear parabolic PDEs. This scheme is inspired by the approach coming from machine learning and developed using deep neural networks in Han and al. [32]. Our algorithm is based on a Picard iteration scheme in which a sequence of linear-quadratic optimisation problem is solved by means of stochastic gradient descent (SGD) algorithm. In the framework of a linear specification of the approximation space, we manage to prove a convergence result for our scheme, under some smallness condition. In practice, in order to be able to treat high-dimensional examples, we employ sparse grid approximation spaces. In the case of periodic coefficients and using pre-wavelet basis functions, we obtain an upper bound on the global complexity of our method. It shows in particular that the curse of dimensionality is tamed in the sense that in order to achieve a root mean squared error of order ${\epsilon}$, for a prescribed precision ${\epsilon}$, the complexity of the Picard algorithm grows polynomially in ${\epsilon}^{-1}$ up to some logarithmic factor $ |log({\epsilon})| $ which grows linearly with respect to the PDE dimension. Various numerical results are presented to validate the performance of our method and to compare them with some recent machine learning schemes proposed in Han and al. [20] and Hur\'e and al. [37].