Add calibrated global min-PTE threshold for many null tests

[cailmdaley]

Problem. A null test that runs many correlated statistics (E/B estimators, bin pairs, scale cuts) cannot use a fixed per-test PTE cut. The global false-positive rate is unknown because it depends on how the statistics correlate and how well each PTE is calibrated.

Change. Adds calibrate_min_pte to sp_validation.statistics. It takes an (n_mocks, n_stats) PTE matrix from noise-only realisations and returns:

• the α-quantile of the per-mock minimum PTE, which is the global threshold, with an order-statistic interval that makes no distributional assumption;
• the effective number of independent tests k_eff, from 1 − (1 − t)^k = α;
• .global_pte(data_ptes): the fraction of mocks whose minimum PTE is ≤ the data's, with a Wilson interval.

It works with any statistic and any number of tomographic bins. An optional two-sided mode uses 2 min(p, 1−p). It needs only numpy and scipy.

Verification. Five new tests in src/sp_validation/tests/test_statistics.py:

• independent uniform PTEs recover the Šidák threshold and k_eff = k;
• the same holds for two-sided PTEs;
• perfectly correlated statistics give k_eff = 1;
• the global p-value matches a hand-worked example;
• the global p-value is uniform under a correlated null.

🤖 Generated with Claude Code