Additive-bias subtraction in calibration zeroes the signal on constant-shear sims (m = −1)

[cailmdaley]

Found while running the image-sim m-bias chain end-to-end (#225): the m-bias came back at m ≈ −1 (no shear response at all) even after the estimator itself was fixed. The cause is a calibration step that is correct for data and silently fatal for constant-shear simulations.

Mechanism (exact, not just empirical)

scripts/calibration/calibrate_comprehensive_cat.py builds the science shear via calibration.get_calibrated_m_c():

# get_calibrated_quantities: correct ensemble metacal
g_corr = np.linalg.inv(R).dot(g_uncorr)          # R = ensemble 2x2 (shear + selection)

# get_calibrated_m_c: additive-bias removal — the problem for sims
c = np.mean(g_uncorr[comp])                       # "additive bias" := mean uncalibrated shear
g_corr_mc = g_corr - np.linalg.inv(R).dot(c)      # subtracted from every object

The multiplicative step is right. But the additive step subtracts each catalogue's own mean — and since ⟨g_corr⟩ = R⁻¹⟨g_uncorr⟩ = R⁻¹c, we get ⟨g_corr_mc⟩ = 0 exactly, for every sim. For real data this is the intended additive-systematics removal (mean cosmic shear ≈ 0, so the mean is PSF leakage / additive bias). For a constant-shear sim the mean is the injected signal, so the subtraction deletes exactly what the m-bias measures: each of 1p2z/1m2z/1z2p/1z2m ends up with ⟨e⟩ = 0 by construction, the response is a difference of zeros, and m = −1.

Evidence (grid_1, 1 tile, inverse-variance weighted, paired estimator)

shear column	response 1+m₁	response 1+m₂
e_uncal (raw)	+0.84	+0.92
e = g_corr_mc (what the m-bias currently reads)	+0.20	+0.02 ← signal gone
ensemble R⁻¹ on paired e_uncal, no additive step	+1.14 → m₁ = +0.14 ± 0.10	+1.27 → m₂ = +0.27 ± 0.11

The last row is a hand-diagnostic using only the per-object R_g** columns (it omits the selection-response part of the ensemble R, a few-% term), so it's the cross-check that the pipeline is otherwise healthy — not the final number.

Possible fixes — design call needed, since this touches the shared calibration path

1. Sim-mode flag (my leaning): an additive_correction: true/false knob on calibrate_comprehensive_cat.py, default true so the data path is provably unchanged; the image-sim mask config sets it false and the calibration writes g_corr (multiplicative-only) as e1/e2. Full ensemble R preserved; minimal diff.
2. Always expose g_corr as extra columns (e1_mult_only/e2_mult_only); the sim m-bias reads those. Non-invasive but adds columns for everyone.
3. The m-bias estimator applies R itself from e_uncal + R_g**. No shared-code change, but it can only apply the per-object shear response, not the global selection response — a good cross-check, not the final number.

Leaving the choice open — @-free ping via the PR thread. Note this finding is the image-sim gate working: it caught a real calibration/sim incompatibility before merge. Whether the data path's additive step is doing the right thing there is probably fine (mean shear ≈ 0 on data) but may deserve a look while we're here.

— Fable, on behalf of Cail

[martinkilbinger]

Not sure why this is a problem for image sims or constant shear. For real data we were able to compute and output the additive bias, independent of subsequent code that subtracts this bisa to get (m- and c-)calibrated shears. Why is this different here?

It is true that, for constant shear simulations, if we want to estimate this constant shear we need to use the non-c-calibrated (non-subtracted) shear estimates.

[cailmdaley]

There is no issue calculating the additive bias, but it's nulling the downstream constant-shear m-bias calculation because <e> = e_input in that case. As you said the solution is to not use the c-calibrated shear, but there are multiple ways to do this in practice, laid out as option 1/2/3 above. Which would you prefer?

[martinkilbinger]

I've implemented option 1:

• calibration.get_calibrated_m_c() gains an additive_correction keyword (default True, so
  the data path is unchanged). When False, the c subtraction is skipped; c and c_err are still
  computed and returned, so they remain available as diagnostics of what would have been
  subtracted.
• notebooks/calibrate_comprehensive_cat.py reads the flag from the metacal: section of the
  mask config and passes it through, printing a notice when the correction is disabled.
• The image-sim mask configs config/calibration/mask_v1.X.4_im_sim.yaml and
  mask_v1.X.9_im_sim.yaml set additive_correction: False with a comment pointing to this
  issue. Real-data configs are untouched and keep the current behaviour via the default.

Catalogues calibrated before this change have the signal subtracted, so any image-sim m-bias
derived from them is invalid; I'm rerunning calibrate_all _ mbias on the current grid run
to validate. Will report the recalibrated m here, and suggest we close the issue once it
comes out at the per-cent level.

Unrelated to the fix but noticed while reading the code: c_err is currently np.std(g_uncorr)
(population scatter, not the error on the mean) _ there's already a MKDEBUG TODO on it;
worth a separate small issue.

[martinkilbinger]

Implemented in #225

[martinkilbinger]

Preliminary results on 15 tiles without the subtracted additive-bias:

m1 = +0.1028 +-0.0605
c1 = -0.0003 +-0.0015
m2 = +0.1538 +-0.0633
c2 = -0.0027 +-0.0015

The estimates of the constant shear look good:
1p2z: = +0.02609 = -0.00188
1m2z: = -0.02506 = -0.00319
1z2p: = -0.00069 = +0.02507
1z2m: = +0.00027 = -0.02971

[cailmdaley]

Heads-up: the additive_correction fix (e38ace9) is on image_sims_val, not on the #225 branch — #225 moved calibration to scripts/calibration/, so the keyword wasn't reaching the workflow there. Mirrored into #225's layout with identical semantics (64ec1d7) and exercised end-to-end; with all weights also removed per #227, the first fully-repaired one-tile number is m1 = +0.274 ± 0.028, m2 = +0.231 ± 0.022 (details on #225 and shapepipe#803). The ±0.02 errors come from shapepipe#803's position-seeded fixnoise — percent-level σ_m now needs only ~4–7 tiles, so your percent-level close condition is within reach as soon as the mechanism behind the +0.25 central value is understood.

— Claude (Fable) on behalf of Cail

[cailmdaley]

Closing as resolved — the additive_correction flag (Martin's image_sims_val stack, mirrored into the image-sims branch layout) makes the additive-bias subtraction opt-out, and the sims configuration ships additive_correction: False for constant-shear runs.

Closing evidence from the corrected end-to-end chain (tile 233.293, grid_1, true amplitude g = 0.025): m1 = +0.019 ± 0.022, m2 = −0.015 ± 0.017 — the m ≈ −1 signal-zeroing this issue documents is gone, while the additive terms stay clean (c1 = −0.0026 ± 0.0084, c2 = −0.0033 ± 0.0075), which is exactly the split the fix was designed to preserve.

— Fable on behalf of Cail