[Jammy2211]

Summary

Correction to #727 (part of #726). _pixelized_source_flux returned the plain surface integral Σᵢ sᵢ Aᵢ over the source mesh, but every other flux latent in autolens/analysis/latent.py is a sum over data pixels (∫ I d²θ / A_pix), and the reconstruction sᵢ is in those per-data-pixel units (the mapping matrix maps s onto image-pixel values). The pixelized magnification therefore scaled with the data's pixel scale: on the Euclid pipeline's 0.1"/pixel simulated dataset it came out 1095, exactly 100× the plausible 10.95 (the Sersic control on the same scene is 15.15). PyAutoLens' fixtures use 1" pixels (A_pix = 1), which is why #727's tests could not see it.

The helper now divides by the data pixel area read from fit.dataset.grids.lp — the same grid the light-profile flux term is evaluated on, so both terms of total_source_flux share one convention. Grid2D.pixel_area is a Python float, so the per-sample jax.jit latent path is unaffected. If no pixel area resolves (no pixel_area, no pixel_scales on the grid or dataset) the helper returns NaN with a once-per-process warning rather than a silently unscaled number. Docstrings on _pixelized_source_flux, total_source_flux, total_source_flux_mujy and magnification state the convention and why the ratio is then the physical ∫ I d²θ / ∫ S d²β.

Measured after the fix on the Euclid euclid_dr1_like scene (Delaunay, 150 Hilbert + 30 edge points, PyPI autoarray's Voronoi areas): magnification 10.95, total_source_flux 1.257 (Sersic control 15.15 / 0.868). The accuracy check against truth stays the human acceptance step on #726.

API Changes

Changed behaviour, no signature changes:

• total_source_flux, total_source_flux_mujy and magnification for a pixelized source are now Σ sᵢAᵢ / A_pix based rather than Σ sᵢAᵢ — a factor 1/A_pix (100× at 0.1"/pixel) relative to fix: magnification latent finite for pixelized sources #727, which was never released. Light-profile-only fits and 1"/pixel data are numerically unchanged.
  See full details below.

Test Plan

• test_autolens/analysis/test_latent.py: 34 passed (3 new — a rectangular pixelized fit at 0.5"/pixel pins Σ sA / 0.25 (4.561 → 1.140, exactly the pixel-area factor); an unresolvable pixel area gives NaN and exactly one warning; a fixed-field-of-view Sersic-behind-isothermal scene observed at 0.4" and 0.2" gives pixelized magnifications within a factor 2 of each other (48.97 / 61.22), which the undivided value violated by 4×). With the division reverted exactly the two value-bearing tests fail.
• Full suite python -m pytest test_autolens/ -n auto: 620 passed, 1 xfailed
• CI green on the Python 3.12 / 3.13 matrix

Full API Changes (for automation & release notes)

Added

• autolens.analysis.latent._data_pixel_area(fit) -> Optional[float] — private helper; fit.dataset.grids.lp.pixel_area, falling back to the product of pixel_scales on the grid then the dataset; None if none resolves.

Changed

• autolens.analysis.latent._pixelized_source_flux(fit, xp=np) — returns Σ sᵢAᵢ / A_pix (was Σ sᵢAᵢ in fix: magnification latent finite for pixelized sources #727); NaN + one warning when A_pix cannot be resolved.
• total_source_flux, total_source_flux_mujy, magnification — pixelized-source values rescaled by 1/A_pix accordingly; unchanged for light-profile-only sources.

Removed

• None.

Migration

• None. fix: magnification latent finite for pixelized sources #727 was merged today and is unreleased; no released version carries the unscaled value.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_011kyfKgDB1rMkcQsn19ow4T

---

Generated by Claude Code