[Jammy2211]

Closes #642. Phase 1 of #638. Community proposal: PyAutoLabs/.github#39 (@HRSAstro), reference snippets attached there.

Summary

Lands the matrix-free interferometer data term and solver as numpy/scipy library utilities with parity tests against the current sparse (w-tilde) path. No fit wiring, no Settings flag, no JAX in this phase.

New module autoarray/inversion/inversion/interferometer/inversion_interferometer_matrix_free_util.py:

• sparse_mapping_matrix_from(inversion, mapper) — the mapper's mapping matrix as a scipy.sparse.csr_matrix from the existing slim triplets (equal to mapper.mapping_matrix exactly; a primary beam folds in via the triplets as on the sparse path).
• MatrixFreeDataOperator (from_inversion) — holds the w-tilde kernel, slim dirty image, mask and M/Mᵀ; wtilde_image_from(image_extent) applies W̃ by one FFT pair on the 2× padded extent grid (matches sparse_operator.apply_operator exactly); curvature_matvec(s) = Mᵀ W̃ M s; data_vector = Mᵀ d̃; kappa0 = W̃ zero-lag value.
• powell_preconditioner_from — P = κ₀ MᵀM + H (Powell et al. 2021), sparse LU factorised once.
• reconstruction_via_pcg_from — preconditioned CG on (F + H) s = D via scipy.sparse.linalg.cg.

Autoarray regularization matrices already carry their coefficient, so H is any existing regularization_matrix converted to sparse (the sparse Matérn precision is phase 2). Module docstring records the deferred JAX questions (fixed iteration count vs while_loop, static-nnz triplets, LU replacement under jit) and the single-mapper limitation.

API Changes

Additive only: one new utility module (not exported through aa). No existing behaviour changes.
See full details below.

Test Plan

test_autoarray/inversion/inversion/interferometer/test_matrix_free_util.py (10 tests, run on the nojax leg too), on a circular 12×12 mask, sub-size 2, 6×6 rectangular mesh, 40 random visibilities (DFT):

• M == mapper.mapping_matrix exactly (rectangular and Delaunay).
• curvature_matvec vs curvature_matrix @ v: 1.8e-15 (coefficient 1.0), 2.2e-16 (0.01); data_vector: 2.6e-16; wtilde_image_from == apply_operator exactly.
• PCG vs the unconstrained sparse-path reconstruction (use_positive_only_solver=False): 3.6e-11 and 7.2e-10 for two Constant coefficients (Witness target 1e-8).
• Preconditioner and LU solve checked against dense; a subprocess test blocks jax/jaxlib/nufftax and re-runs the module tests.
• Worktree runs: pytest test_autoarray/inversion -q → 1032 passed, 4 xfailed; nojax emulation 10 passed with no jax modules loaded.

Heart YELLOW acknowledged by the human on 2026-10-21 (manifest drift 1 mismatch; release validation stale: source moved since rehearsal (PyAutoArray)).

Full API Changes (for automation & release notes)

Added

• autoarray.inversion.inversion.interferometer.inversion_interferometer_matrix_free_util — sparse_mapping_matrix_from(inversion, mapper), MatrixFreeDataOperator (from_inversion, wtilde_image_from, curvature_matvec, data_vector, kappa0, params), powell_preconditioner_from(operator, regularization_matrix_sparse, coefficient), reconstruction_via_pcg_from(operator, regularization_matrix_sparse, coefficient, x0=None, rtol=1e-8, maxiter=5000).

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code