fix: imaging/delaunay log_evidence -inf via rank-stripped regularisation (#69)

[Jammy2211]

Summary

Fixes _developer#69 — likelihood/imaging/delaunay.py was returning log_evidence = -inf from both step-by-step and inv-matrix rebuilds, while FitImaging.log_evidence returned a finite 26288.321397. The bug was in the script's manual recomputation, not in the library.

Validated end-to-end: script now exits 0, both regression assertions PASS at 26288.321397.

Root cause

compute_log_evidence computed slogdet on the full regularization_matrix, which is rank-deficient: the script's lens model includes linear MGE light profiles whose linear parameters share the inversion's linear system but are not regularised. Their rows / columns in the regularisation matrix are zero, so the full matrix is singular and slogdet returns -inf for log_det(H). The same applies to F + H if you slice through those zero rows.

PyAutoArray's reference Inversion.log_evidence chain handles this correctly by computing the log-det terms on the reduced block — regularization_matrix_reduced / curvature_reg_matrix_reduced slice the matrices down to just the mapper-pixel rows and columns (positive-definite, Cholesky-safe). See autoarray/inversion/inversion/abstract.py::log_det_regularization_matrix_term for the reference implementation.

Fix

• Precompute inversion.regularization_matrix_reduced, inversion.curvature_reg_matrix_reduced, and inversion.mapper_indices eagerly outside the timed region.
• Change compute_log_evidence to:
  • Slice the reconstruction by mapper_indices for the s^T H s regularisation term.
  • Use Cholesky-based log-det (2 * sum(log(diag(L)))) on the reduced PD matrices — matching the reference implementation.
• chi² and the noise-normalisation term still use the full reconstruction and full noise map, since those are per-pixel data terms over the masked image, not regularised-matrix terms.

Validation

Locally on PyAutoLens 2026.5.14.2 / Intel i9-10885H CPU:

  Eager regression assertion PASSED: log_evidence matches 26288.321397
  Regression assertion PASSED: log_evidence matches 26288.321397

Whole script exits 0 (was exit 1 with -inf vs 26288 mismatch).

Refs

• The upstream source-of-truth _developer/jax_profiling/jit/imaging/delaunay.py has the same bug. A follow-up PR there should mirror this fix — added as a TODO note in the F1 follow-up tracking comment.
• Surfaced by today's full-script run against the canonical autolens_profiling repo (Phase 5 ship validation).

🤖 Generated with Claude Code