Affected Repositories

• PyAutoArray (primary)
• PyAutoGalaxy (stacked follow-on: parity test, guards only if needed)

Branch Survey

Suggested branch: feature/matrix-free-interferometer-wiring (both repos)
Worktree: ~/Code/PyAutoLabs-wt/matrix-free-interferometer-wiring/

Implementation Steps (PyAutoArray)

1. autoarray/inversion/inversion/interferometer/matrix_free.py — InversionInterferometerMatrixFree(InversionInterferometerSparse). Subclassing the sparse FFT class gives it the data vector, mask, transformer and mapped-data dicts, and is what MatrixFreeDataOperator.from_inversion already expects. A cached operator property builds M and the W~ kernel action. Overrides:
   • reconstruction: reconstruction_via_pcg_from, or reconstruction_positive_via_interior_point_from when use_positive_only_solver.
   • log_det_curvature_reg_matrix_term: Powell log-det + SLQ via log_det_curvature_reg_via_slq_from (module defaults: 10 probes, 40 steps, fixed seed).
   • log_det_regularization_matrix_term: the regularization's exact sparse term.
   • regularization_term, fast_chi_squared: via the sparse H and operator.curvature_matvec (inherited versions read the dense F).
   • curvature_matrix, curvature_matrix_diag, curvature_reg_matrix, curvature_reg_matrix_reduced, reconstruction_covariance_matrix, reconstruction_noise_map, errors: raise InversionException naming the matrix-free route and pointing at interferometer_route="fft".
   • Constructor precondition check (mirrors interferometer_numba/sparse.py::_check_preconditions), raising for: non-NumPy xp; any AbstractLinearObjFuncList; mapper count != 1; over-sampling (sub_fraction != 1); non-trivial primary beam; mesh not RectangularUniform; regularization without regularization_matrix_sparse_from / regularization_factor_sparse_from (decision: only MaternSparse supplies the sparse factor the SLQ probes need, so other schemes raise rather than densify — flagged in the PR body).
2. autoarray/settings.py — accept "matrix_free" in interferometer_route (property + constructor docstring, documented like "fft"/"numba"). autoarray/config/general.yaml — extend the comment; default stays null.
3. autoarray/inversion/inversion/factory.py (inversion_interferometer_from, sparse-operator branch ~L228) — if settings.interferometer_route == "matrix_free" return the new class before the numba gate; the class's own checks raise, so the factory never falls back.
4. autoarray/inversion/plot/inversion_plots.py — add exc.InversionException to the two except clauses around reconstruction_noise_map_dict (panel 6 and save_reconstruction_csv) so the visualizer skips the noise map on this route.
5. inversion_interferometer_matrix_free_util.py docstring — move the wiring line out of "Remaining work (phase 5)" and name the class + Settings value.
6. Tests test_autoarray/inversion/inversion/interferometer/test_matrix_free.py (reuse fixtures from test_matrix_free_util.py): factory dispatch; reconstruction vs FFT route to 1e-6 of max; log evidence within the documented 0.5%; positive-only parity vs FFT positive-only; every unsupported configuration raises; every dense attribute raises; Settings typo raises ValueError. NumPy/scipy only — a JAX-backend raise test uses pytest.importorskip("jax").

Implementation Steps (PyAutoGalaxy, after the PyAutoArray PR merges)

7. Expected no source change: FitInterferometer builds through the factory with its Settings, and log_evidence reads only the two log-det terms, regularization_term and fast_chi_squared. Add to test_autogalaxy/interferometer/test_fit_interferometer.py: rectangular MaternSparse fit with interferometer_route="matrix_free" vs "fft" at the witness tolerances, plus one positive-only case. Guard any visualizer/aggregator dense-attribute touch found while testing.

Key Files

• autoarray/inversion/inversion/interferometer/inversion_interferometer_matrix_free_util.py — phase 1–4 utilities the class composes
• autoarray/inversion/inversion/interferometer/sparse.py — parent class
• autoarray/inversion/inversion/interferometer_numba/sparse.py — precondition pattern to mirror
• autoarray/inversion/inversion/factory.py — dispatch
• autoarray/settings.py, autoarray/config/general.yaml — the option
• autoarray/inversion/regularization/matern_sparse.py — sparse matrix/factor/log-det providers
• test_autoarray/inversion/inversion/interferometer/test_matrix_free_util.py — fixtures and tolerances

Witness

In CI: FitInterferometer with Settings(interferometer_route="matrix_free") on a rectangular MaternSparse case agrees with the FFT sparse route (reconstruction 1e-6 relative, log evidence within the 0.5% SLQ tolerance stated in the module docstring), and every unsupported configuration raises InversionException.

Matrix-free interferometer inversion, phase 5a: make it selectable from a fit

Type: feature
Target: PyAutoArray
Repos:

• PyAutoArray
• PyAutoGalaxy
  Autonomy: supervised
  Priority: high
  Parent: active/matrix_free_interferometer_inversion.md (PyAutoArray#638)
  Community: Matrix-free solution to make very large meshes more efficient for Interferometer .github#39 (@HRSAstro)
  Consequence: judge
  Heart-override: human approved "Human override for Heart RED (development only)" on 2026-10-26 (RED set = release integrate, workspace smoke, worktree drift; none touch PyAutoArray/PyAutoGalaxy)

Context

Phases 1–4 of #638 are merged (PyAutoArray #645, #650, #651, #652). They are released in
2026.10.23.1 and later. They exist only as utility functions in
autoarray/inversion/inversion/interferometer/inversion_interferometer_matrix_free_util.py:

• the operator data term and Powell-preconditioned PCG;
• aa.reg.MaternSparse;
• the evidence log-dets via exact sparse LU + stochastic Lanczos quadrature;
• positivity via primal-dual interior point with active-set polish.

The module docstring says: "None of it is wired into fits and there is no Settings flag."
This task wires it in. It is the "wiring" item of
draft/research/autolens_profiling/matrix_free_interferometer_phase5.md. Split that draft:
this task is 5a; the profiling note and the workspace example stay in phase 5 as 5b/5c.

What

1. A new inversion class InversionInterferometerMatrixFree (next to sparse.py in
   autoarray/inversion/inversion/interferometer/). It holds M (scipy.sparse mapping matrix),
   the W~ kernel FFT action and the sparse regularization. It provides:
   • reconstruction (PCG, or interior point when use_positive_only_solver);
   • log_det_curvature_reg_matrix_term / log_det_regularization_matrix_term
     (phase 3 functions);
   • mapped_reconstructed_data / mapped_reconstructed_image.
     Dense-only attributes (curvature_matrix, curvature_reg_matrix, its inverse, errors)
     raise a clear InversionException naming the matrix-free route. Do not silently build them.
2. A Settings option in autoarray/settings.py, extending the existing
   interferometer_route (None / "fft" / "numba") with "matrix_free". Document it like
   the existing values and add a packaged config default (off). Dispatch it in
   autoarray/inversion/inversion/factory.py around line 228 (the
   dataset.sparse_operator is not None branch). Where the matrix-free route can't represent the
   inversion, raise rather than fall back silently. Those cases are: multiple mappers, linear light
   profiles / func lists, over-sampling, a non-trivial primary beam, a non-rectangular mesh, the
   JAX backend.
3. PyAutoGalaxy: FitInterferometer reaches it through Settings with no other API change. Check
   that log_evidence, figure_of_merit and the visualizer don't touch the dense attributes on
   this route. Where they do, skip or guard them.
4. Tests (PyAutoArray + PyAutoGalaxy):
   • on a small rectangular MaternSparse interferometer case, FitInterferometer via
     interferometer_route="matrix_free" matches the FFT sparse route: reconstruction to 1e-6
     of its max, log evidence within the documented SLQ tolerance;
   • positive-only parity, as in phase 4;
   • the unsupported cases raise;
   • a nojax leg needs importorskip where relevant.
5. Update the module docstring's "Remaining work (phase 5)" list to say wiring is done.

Out of scope

• The profiling campaign. Matrix-free rows are being added as experiments/ cells in the
  likelihood-lifecycle 2b interferometer campaign in autolens_profiling, owned by another session.
  Don't touch autolens_profiling.
• The workspace example.
• Uncertainty maps (stochastic diagonal).
• Multiple mappers and linear light profiles.
• A JAX path.

Witness

In CI: FitInterferometer with Settings(interferometer_route="matrix_free") on a rectangular
MaternSparse case agrees with the FFT sparse route (reconstruction 1e-6 relative, log evidence
within the SLQ tolerance stated in the module docstring), and every unsupported configuration
raises InversionException.