Pith. sign in

REVIEW 4 cited by

A Natural Primal-Dual Hybrid Gradient Method for Adversarial Neural Network Training on Solving Partial Differential Equations

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2411.06278 v4 pith:S37HGGA6 submitted 2024-11-09 math.NA cs.LGcs.NAmath.OC

A Natural Primal-Dual Hybrid Gradient Method for Adversarial Neural Network Training on Solving Partial Differential Equations

classification math.NA cs.LGcs.NAmath.OC
keywords equationsalgorithmmethodgradientdifferentialhybridnaturalneural
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
abstract

We propose a scalable preconditioned primal-dual hybrid gradient algorithm for solving partial differential equations (PDEs). We multiply the PDE with a dual test function to obtain an inf-sup problem whose loss functional involves lower-order differential operators. The Primal-Dual Hybrid Gradient (PDHG) algorithm is then leveraged for this saddle point problem. By introducing suitable precondition operators to the proximal steps in the PDHG algorithm, we obtain an alternative natural gradient ascent-descent optimization scheme for updating the neural network parameters. We apply the Krylov subspace method (MINRES) to evaluate the natural gradients efficiently. Such treatment readily handles the inversion of precondition matrices via matrix-vector multiplication. An \textit{a posteriori} convergence analysis is established for the time-continuous version of the proposed algorithm for general linear PDEs. By incorporating appropriate boundary loss terms, we further obtain a refined \textit{a priori} convergence result for elliptic equations in divergence form. The algorithm is tested on various types of PDEs with dimensions ranging from $1$ to $50$, including linear and nonlinear elliptic equations, reaction-diffusion equations, and Monge-Amp\`ere equations stemming from the $L^2$ optimal transport problems. We compare the performance of the proposed method with several commonly used deep learning algorithms such as physics-informed neural networks (PINNs), the DeepRitz method and weak adversarial networks (WANs) using either the Adam or the L-BFGS optimizer. The numerical results suggest that the proposed method performs efficiently and robustly and converges more stably with higher accuracy.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Implicit Neural Optimal Transport via Fixed-Point Optimization

    math.OC 2026-05 unverdicted novelty 7.0

    A single-network fixed-point formulation for neural optimal transport eliminates adversarial min-max optimization and implicit differentiation while enforcing dual feasibility exactly.

  2. Implicit Neural Optimal Transport via Fixed-Point Optimization

    math.OC 2026-05 unverdicted novelty 7.0

    A single-network implicit neural optimal transport method that solves the c-transform via proximal fixed-point iteration for stable, non-adversarial training.

  3. Inf-Sup Neural Networks for High Dimensional PDEs

    math.NA 2026-07 conditional novelty 5.0

    Inf-sup neural networks with dual Lagrange multipliers solve high-dimensional PDEs with proven error decomposition for linear convection-diffusion and empirical success on nonlinear cases.

  4. Deep Tangent Bundle (DTB) method: a Deep Neural Network approach to compute solutions of PDES

    math.NA 2025-08 conditional novelty 5.0

    DTB approximates the spatial vector field of an evolution PDE by the span of the derivatives of a deep network, updates the solution directly via linear least squares, and adapts the network occasionally.