fix: guard sparse interferometer operator against unequal real/imag noise sigma

[Jammy2211]

Overview

InterferometerSparseOperator builds W~ = Re(FᴴWF) from a single precision operator computed with noise_map.real only (dataset.py:362). That is exact only when every visibility has equal real and imaginary sigma. With unequal sigmas the sparse InversionInterferometerSparse curvature matrix silently disagrees with the dense path (5e-16 → 5e-10..3e-2 relative, geometry dependent), on the single-mapper route too. Simulated and real datasets satisfy the equality, so it is latent — but nothing guards it. Found during #499/#500.

Plan

• Raise a clear DatasetException from Interferometer.apply_sparse_operator when any visibility has unequal real/imag sigma, naming the cause and the dense-path workaround.
• Unit test: unequal sigma raises; equal, non-uniform sigma passes.
• Docstring notes on apply_sparse_operator and InterferometerSparseOperator.
• Assess (in the PR, no implementation) whether a general two-operator extension is worth doing.

Detailed implementation plan

Affected Repositories

• PyAutoArray (primary)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoArray	main	clean (parallel worktree alongside numba-cpu-nnls-iteration-reduction — disjoint files)

Suggested branch: feature/sparse-interferometer-unequal-sigma-guard

Implementation Steps

1. autoarray/dataset/interferometer/dataset.py — at the top of apply_sparse_operator, if not np.allclose(self.noise_map.real, self.noise_map.imag): raise exc.DatasetException(...) with a message stating the W~ operator is built from the real-part sigma only and requires sigma_real == sigma_imag per visibility; suggest not calling apply_sparse_operator (dense mapping path) or equalising the sigmas.
2. Same file / inversion_interferometer_util.py — docstring notes on the precondition.
3. test_autoarray/dataset/interferometer/test_dataset.py — test__apply_sparse_operator__unequal_real_imag_noise_raises and an equal-but-non-uniform sigma control that succeeds.
4. PR body: short assessment of the two-operator generalisation (real- and imag-weighted precision operators combined in apply_operator).

Key Files

• autoarray/dataset/interferometer/dataset.py — guard + docstring
• test_autoarray/dataset/interferometer/test_dataset.py — tests

Original Prompt

Click to expand starting prompt

Sparse interferometer W~ path silently assumes equal real/imag noise sigma

Type: bug
Target: PyAutoArray
Repos:

• PyAutoArray
• workspaces
  Difficulty: medium
  Autonomy: supervised
  Priority: normal
  Status: formalised

Sparse interferometer W~ path silently assumes equal real/imag noise sigma

Type: bug
Target: PyAutoArray
Repos:

• PyAutoArray
  Difficulty: medium
  Autonomy: supervised
  Priority: normal

Problem

InterferometerSparseOperator builds the real-space operator W~ = Re(F^H W F) from a single
NUFFT precision operator. That reduction is only exact when every visibility has equal real and
imaginary noise sigma. With sigma_real != sigma_imag the sparse InversionInterferometerSparse
curvature matrix disagrees with the dense InversionInterferometerMapping path even on the
single-mapper route: measured 5e-16 relative with equal sigmas vs 5e-10 to 3e-2 with unequal
sigmas (geometry dependent). Pre-existing, not introduced by #500. SimulatorInterferometer and
real datasets satisfy the equality, so it is latent — but nothing guards it.

Fix

1. Raise a clear InversionException (or in Interferometer.apply_sparse_operator) when the noise
   map has unequal real/imag sigma for any visibility, with a unit test. Do this first.
2. Assess extending the operator to the general case (two precision operators, real- and
   imag-weighted, combined in apply_operator) — only if the cost is modest; otherwise the guard stands.

Context

Found 2026-08-28 while shipping #499 / #500. The misc/jax_assertions/fit_interferometer_sparse_operator.py parity
scripts in both test workspaces document the limitation in their
docstrings and deliberately use equal-sigma noise maps; they need no change.

[Jammy2211]

Shipped

#503 — Interferometer.apply_sparse_operator now raises DatasetException when any visibility has unequal real/imag noise sigma, naming the count, the max relative difference and both workarounds; docstring preconditions on apply_sparse_operator and InterferometerSparseOperator; two tests. Assessment in the PR: the general two-operator extension needs a second, pixel-sum-indexed kernel and is not worth it for any known dataset — not doing it.