Variational Inference Using a Differentiable Multigrid Linear Solver

DOI10.48550/arXiv.2608.00760MaRDI QIDQ7284569arXivFDO

Authors Philipp Frank, Dominik Jüstel, Philipp Mertsch, Vo Hong Minh Phan, Laurin Söding, Andrés Ramírez, Philipp Haim, David Föger, Torsten A. Enßlin, Philipp Gschwandtner, Hanieh Zandinejad, Ralf Kissmann

Publication date 1 August 2026

Published in arXiv

Tldr We develop DMGS, a differentiable multigrid solver for steady-state diffusion-absorption problems that integrates with JAX and NIFTy, allowing efficient Jacobian-vector products for variational inference; it generalizes to 32 independent datasets, reduces peak memory usage, and outperforms native JAX runtime on an inverse problem involving diffuse radiative transfer.

Full work available at URL https://doi.org/10.48550/arXiv.2608.00760

arXiv classification

stat.COmath.NAphysics.med-ph

Cites work

• An introduction to the adjoint approach to design
• Automated derivation of the adjoint of high-level transient finite element programs
• Bringing PDEs to JAX with forward and reverse modes automatic differentiation
• Gradient-Based Markov Chain Monte Carlo for Bayesian Inference With Non-differentiable Priors
• NIFTY - Numerical Information Field Theory - a versatile Python library for signal inference
• NIFTY - Numerical Information Field Theory - a versatile Python library for signal inference
• Stan: A Probabilistic Programming Language

This page was built for publication: Variational Inference Using a Differentiable Multigrid Linear Solver

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284569)