TreeCorr's tree approximation changes the means of ξ± by up to ~0.9σ on SP_v1.4.6.3's reporting grid (bin_slop at its default, b ≈ 0.1), depending on how the jackknife patches split the tree. #399 and #400 make the published means independent of the patch layout, and they compute them at b ≤ 0.01. The patched pass, which only supplies covariances (ξ± jackknife, ρ jackknife, M_ap²), stays at TreeCorr's default tolerance, because tightening it costs about 6× per job.
Open question: is the jackknife covariance at default tolerance accurate enough? Each patch's statistic carries its own approximation error, and the spread of those errors enters the jackknife variance. To settle it, compare the jackknife covariance of the SP_v1.4.6.3 reporting ξ± at default bin_slop against bin_slop = 0.036 (b ≈ 0.01) for the same layout (diagonal ratio, correlation matrix, χ² of the data under each), and check whether the difference is negligible against the covariance's own noise.
Evidence for the means: scratch-wf/d2t-parity/xi-patch/ on candide; figures in #399 and #400.
— Claude Opus 5.5 on behalf of Cail
TreeCorr's tree approximation changes the means of ξ± by up to ~0.9σ on SP_v1.4.6.3's reporting grid (bin_slop at its default, b ≈ 0.1), depending on how the jackknife patches split the tree. #399 and #400 make the published means independent of the patch layout, and they compute them at b ≤ 0.01. The patched pass, which only supplies covariances (ξ± jackknife, ρ jackknife, M_ap²), stays at TreeCorr's default tolerance, because tightening it costs about 6× per job.
Open question: is the jackknife covariance at default tolerance accurate enough? Each patch's statistic carries its own approximation error, and the spread of those errors enters the jackknife variance. To settle it, compare the jackknife covariance of the SP_v1.4.6.3 reporting ξ± at default bin_slop against bin_slop = 0.036 (b ≈ 0.01) for the same layout (diagonal ratio, correlation matrix, χ² of the data under each), and check whether the difference is negligible against the covariance's own noise.
Evidence for the means:
scratch-wf/d2t-parity/xi-patch/on candide; figures in #399 and #400.— Claude Opus 5.5 on behalf of Cail