Gatti/DES weights pathological on image sims: N_eff ~1–5% of N, inflates σ(m) ~12×

[cailmdaley]

Second finding from the end-to-end image-sim m-bias run (#225), distinct from the additive-subtraction bug: even with a paired (shape-noise-cancelling) estimator, the m-bias error stayed at ±1.5 instead of the expected ~±0.1. The w_des weights from compute_weights_gatti are degenerate on the sims — a handful of objects carry essentially all the weight.

Measurement (grid_1, 1 tile, per sim)

        N     N_eff   (% of N)   w: min      med     max      max/med   top-10 %wt
1z2z   1684   51.3    3.0%       4.2e-2   1.05e1   4.5e3    426        29%
1p2z   1668   21.2    1.3%       1.3e-3   9.7e0    8.3e3    858        37%
1m2z   1683   29.4    1.7%       2.6e-3   1.1e1    5.6e3    510        40%
1z2p   1685   12.7    0.8%       1.8e-2   9.9e0    1.4e4    1401       55%
1z2m   1675   90.9    5.4%       2.2e-2   9.0e0    3.0e3    332        23%

(N_eff = (Σw)²/Σw²; a healthy shear weight sits at 50–90% of N.) σ(m) scales as 1/√N_eff, so this alone inflates the error ~8–12×, and the dominant few objects also skew the central value:

        weight   N_eff    m1                  m2
        unit     1563     −0.880 ± 0.124      −1.002 ± 0.113
        w_iv     1500     −0.805 ± 0.108      −0.979 ± 0.099
        w_des      ~28    −2.139 ± 1.533      −0.431 ± 0.332

(All rows still show the m ≈ −1 central value — that's #226, a separate bug. The weight is the error story.) w_iv behaves normally, which is what isolates the problem to compute_weights_gatti — most likely the SNR/size binning going degenerate, or a variance term blowing up, on the sims' narrow magnitude/size range.

Why it can't be side-stepped

For the m-bias to be science-applicable it must be measured under the weight the science analysis actually uses — so w_des needs to be fixed at source, not swapped for w_iv (which serves as the diagnostic proving the estimator and pairing are sound). Possibly related to the broader ngmix-v2.0 weight-map work (inverse-variance weights, shapepipe#604 stream).

— Fable, on behalf of Cail

[martinkilbinger]

DES weights are very noisy for a small number of objects. In the (some) 20x20 grid of SNR and resolution, some pixels might have few or no galaxy for a small number of galaxies/tiles. Might have to run more tiles to see this behaviour go away?

[cailmdaley]

@fabianhervaspeters says that we shouldn't be using weights at all for m-bias

[cailmdaley]

Closing as resolved — no weights anywhere in the sim m-bias is now the standing configuration (w_col: null, global_R_weight: null in the image-sims branch config), with w_des retained only as a diagnostic knob.

Confirmation from the corrected single-tile baseline (tile 233.293, grid_1, paired unweighted estimator, and the true injected amplitude g = 0.025 — see the amplitude discussion around shapepipe#796): m1 = +0.019 ± 0.022, m2 = −0.015 ± 0.017, c1 = −0.0026 ± 0.0084, c2 = −0.0033 ± 0.0075. The σ(m) ≈ ±1.5 inflation this issue documents is gone; σ(m) ≈ 0.02 on a single tile with noise pairing (shapepipe#803).

The related estimator-flavor question (w_des appearing in the #796 m-bias snippet) is with Fabian on that thread; it doesn't affect this issue's verdict since the chain no longer uses weights anywhere.

— Fable on behalf of Cail