asdex: Automatic Sparse Differentiation in JAX

DOI10.48550/arXiv.2610.12336MaRDI QIDQ7374655arXivFDO

Authors Guillaume Dalle, Adrian Hill

Publication date 8 October 2026

Published in arXiv

Tldr The asdex library accelerates automatic differentiation by leveraging input-agnostic sparsity patterns in Jacobians and Hessians, reducing the number of AD passes required to compute these matrices. asdex offers a standalone JAX implementation with efficient sparse derivatives for banded, block-banded, and other sparse functions.

Full work available at URL https://doi.org/10.48550/arXiv.2610.12336

arXiv classification

cs.LGcs.MSmath.NA

Cites work

• A versatile FEM framework with native GPU scalability via globally-applied AD
• Estimation of sparse hessian matrices and graph coloring problems
• Estimation of Sparse Jacobian Matrices
• Julia: a fresh approach to numerical computing
• New Acyclic and Star Coloring Algorithms with Application to Computing Hessians
• NonlinearSolve.jl: High-Performance and Robust Solvers for Systems of Nonlinear Equations in Julia
• NonlinearSolve.jl: high-performance and robust solvers for systems of nonlinear equations in Julia
• On the Estimation of Sparse Hessian Matrices
• Revisiting Sparse Matrix Coloring and Bicoloring
• What Color Is Your Jacobian? Graph Coloring for Computing Derivatives

This page was built for publication: asdex: Automatic Sparse Differentiation in JAX

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7374655)