ImageSimMBias computes m/c on the full pooled sample with a paired object bootstrap. Object resampling assumes galaxies are independent, but tile-level systematics (PSF model, background, selection) correlate objects within a tile, so the bootstrap can underestimate the error.
- Carry a tile id per object through the extracted cut catalogue (patch name from
final_cat_*.hdf5).
- Add a leave-one-tile-out jackknife for m/c errors, reported alongside the bootstrap; disagreement between the two is itself a tile-systematics diagnostic.
- Replace
mbias_cumulative.yaml / plot_convergence (stateful across runs, mixes code versions) with an in-run convergence curve: m on the first k tiles, k=1..N, jackknife errors.
Raised by Fabian (Slack, 2026-09-14); he'd read the cumulative plot as per-tile averaging — the estimator is already full-sample.
ImageSimMBiascomputes m/c on the full pooled sample with a paired object bootstrap. Object resampling assumes galaxies are independent, but tile-level systematics (PSF model, background, selection) correlate objects within a tile, so the bootstrap can underestimate the error.final_cat_*.hdf5).mbias_cumulative.yaml/plot_convergence(stateful across runs, mixes code versions) with an in-run convergence curve: m on the first k tiles, k=1..N, jackknife errors.Raised by Fabian (Slack, 2026-09-14); he'd read the cumulative plot as per-tile averaging — the estimator is already full-sample.