Automatic Sparse Differentiation in JAX.

asdex exploits sparsity structure to efficiently materialize Jacobians and Hessians. It implements a custom Jaxpr interpreter that uses abstract interpretation to detect sparsity patterns from the computation graph, then uses graph coloring to minimize the number of AD passes needed.

pip install asdex

Or with uv:

uv add asdex

import asdex
import jax
import jax.numpy as jnp

def f(x):
    return (x[1:] - x[:-1]) ** 2

x_sample = jnp.zeros(50)  # sample input for sparsity pattern detection
jac_fn = jax.jit(asdex.jacobian(f, x_sample))
# ColoredPattern(49×50, nnz=98, sparsity=96.0%, JVP, 2 colors)
#   2 JVPs (instead of 49 VJPs or 50 JVPs)
# ⎡⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎤   ⎡⣿⎤
# ⎢⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⎥ → ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⠀⠀⎥   ⎢⣿⎥
# ⎢⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⢦⡀⎥   ⎢⣿⎥
# ⎣⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠉⎦   ⎣⠉⎦

for x in inputs:
    J = jac_fn(x)

Instead of 49 VJPs or 50 JVPs, asdex computes the full sparse Jacobian with just 2 JVPs.

Since sparsity detection and coloring can be expensive on large problems, we recommend saving and reusing colored patterns:

import jax.numpy as jnp
from asdex import jacobian_coloring
from asdex import ColoredPattern, jacobian_from_coloring

# Compute coloring once...
x = jnp.zeros(1000)
coloring = jacobian_coloring(f, x)
coloring.save("colored.npz")

# ...load and reuse later
coloring = ColoredPattern.load("colored.npz")
jac_fn = jax.jit(jacobian_from_coloring(f, coloring))

The full ASD pipeline:

• Sparse Jacobians and Hessians: one VJP/JVP/HVP per color, with automatic (or user-defined) mode selection.
• Sparsity detection: finds global sparsity patterns valid for all inputs.
• Graph coloring: row, column, and symmetric coloring minimize AD passes.
• Correctness verification against vanilla JAX.

You already know your sparsity pattern?

• Manually provide sparsity patterns: supply a known pattern from dense, COO, or BCOO formats.
• Precompute, save & load: reuse a colored pattern across inputs, or persist it by saving and loading.

An interface mirroring JAX:

• Multiple inputs and outputs: supports multi-argument functions via argnums, as well as multiple return values.
• PyTree inputs and outputs: sparse differentiation through arbitrary nested PyTrees.
• Auxiliary outputs: supports has_aux=True for functions returning (output, aux).
• Value and derivative: value_and_jacobian / value_and_hessian return the primal value f(x) without a redundant forward pass.

And more:

• Multiple output formats: decompression to BCOO, dense JAX arrays, NumPy, and SciPy (COO/CSR/CSC) arrays.
• Bounded memory: chunk_size caps parallel AD passes for large color counts.
• Visualizations: spy plots and braille pattern previews.

• Getting Started: step-by-step tutorial
• How-To Guides: task-oriented recipes
• Explanation: how and why it works
• API Reference: full API documentation
• Contributing: guidelines for collaborating on asdex
• AI Policy: guidelines for LLM contributions

Prior work on ASD by asdex's authors Adrian Hill (@adrhill) and Guillaume Dalle (@gdalle), as well as Alexis Montoison (@amontoison):

• An Illustrated Guide to Automatic Sparse Differentiation, Hill, Dalle, Montoison (2025)
• Sparser, Better, Faster, Stronger: Efficient Automatic Differentiation for Sparse Jacobians and Hessians, Hill & Dalle (2025)
• Revisiting Sparse Matrix Coloring and Bicoloring, Montoison, Dalle, Gebremedhin (2025)
• SparseConnectivityTracer.jl, Hill & Dalle
• SparseMatrixColorings.jl, Dalle & Montoison
• DifferentiationInterface.jl, Dalle & Hill

Prior and concurrent (partial) attempts at ASD in JAX:

• sparsejac: coloring and decompression
• sparsediffax: coloring and decompression (by asdex's @gdalle)
• jax-nansparse: sparsity detection using NaN propagation
• JAX-AMG: specialized ASD module for algebraic multigrid methods
• tatva: specialized ASD module for FEM
• See discussion in JAX issue #1032

Adrian Hill gratefully acknowledges funding from the German Federal Ministry of Education and Research under the grant BIFOLD26B.

This package is built with Claude Code, based on previous, hand-written work by the same authors in the Julia programming language, as noted above. These works in turn stand on the shoulders of giants, notably Andreas Griewank, Andrea Walther, and Assefaw Gebremedhin.

The asdex logo was designed by @overripemango.

If you use asdex in your research, please cite:

@software{asdex2026,
  author = {Hill, Adrian and Dalle, Guillaume},
  title = {asdex: Automatic Sparse Differentiation in JAX},
  url = {https://github.com/adrhill/asdex},
  doi = {10.5281/zenodo.18788242}
}