Skip to content

GLASS v1.4.6.3_v2 mocks carry 17% more galaxies than the catalogue #319

Description

@cailmdaley

The GLASS v1.4.6.3_v2 mocks carry 17% more galaxies than the v1.4.6.3 catalogue they stand in for (60.4 M unit-weight per realisation vs 51.6 M weight-effective), with σ_e matching to 0.2%.
The mocks are therefore under-noised: their noise-dominated error bars are 16% smaller than the data's (0.73× in variance).

Two area conventions

The catalogue's n_eff = 4.957 arcmin⁻² is defined on the footprint mask, sp_validation/cosmo_inference/data/mask/mask_map_footprint_nside_4096.fits (binary nside-4096 map from the HealSparse pixel mask, 2894 deg²) — the area we use for covariances and quote in the papers.
The mock generator uses its own occupied-pixel mask, /n09data/guerrini/glass_mock_v1.4.6.3_v2/mask_nside4096.fits (create_mask.py: nside-4096 pixels containing at least one catalogue galaxy, 2757 deg²).

catalogue galaxies n_eff on the footprint mask σ_e per component
UNIONS v1.4.6.3 leak-corrected (/n17data/UNIONS/WL/v1.4.x/v1.4.6.3/unions_shapepipe_cut_struc_2024_v1.4.6.3.fits) 51.6 M weight-effective 4.957 0.2676
GLASS v1.4.6.3_v2 (/n09data/guerrini/glass_mock_v1.4.6.3_v2/results/unions_glass_sim_*_4096.fits) 60.4 M unit-weight 5.793 0.2682

Mock and data galaxies occupy the same pixels (footprints coincide to within 2%), so the count ratio is the whole noise difference: the C_ℓ^BB scatter of the 350 realisations is 0.72× the data covariance, against 0.74 predicted from the counts.

How it happened

make_unions_glass_sim.py (/n17data/sguerrini/sp_validation_copy/glass_mock/) takes the density as --neff, default 6.0905 arcmin⁻², and neither run_glass_mock.sh nor run_glass_mock_local.sh overrides it.
6.09 is the catalogue n_eff from the 2405 deg² era; when the higher-resolution footprint mask moved the area to 2894 deg² the catalogue n_eff became 4.957, but the default did not follow.
6.09 arcmin⁻² on the 2757 deg² occupied-pixel mask gives the 60.4 M galaxies per realisation.

Downstream consequences

  • Paper II compares mock scatter against the covariance machinery rerun at the mocks' own noise (CosmoCov with source_n_gal = 5.793 at area = 2894; NaMaster from the mock count map for C_ℓ^BB) and states the mismatch in the referee response. Not affected beyond that.
  • Papers III / IV state that the mocks reproduce the data's number density; the delivered catalogues do not.
    A covariance built at the data's n_eff exceeds the mock scatter by 16% in σ from the count mismatch alone.
    Paper IV (Fig. 4) reports iNKA error bars 15% above the mocks and OneCovariance within 5–10% of them, and attributes the iNKA gap to non-stationary noise; the count mismatch is a second contribution of the same size and sign. Nothing here quantifies the split.
  • Configuration- vs harmonic-space consistency checks calibrated on these mocks are too stringent: scatter measured on mocks with less noise than the data understates what the real catalogue would show.

— Claude on behalf of Cail

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions