Skip to main content

Changelog of null-tests performed in POLARBEAR large-patch analysis

Keywords

POLARBEAR, Null tests, Reproducible research

Changelog of null-tests performed in POLARBEAR large-patch analysis

Below is the changelog of all null-test runs performed in the POLARBEAR analysis in part V. They are reproduced here for reference.

Note:

  • Semantic versioning is used: major.minor.bugfix. Major versions all share the same map-making, for example v0.6 through v0.6.6. Minor versions differ in post-mapmaking, for example how they are grouped to coadd, how the spectra are generated, etc.
  • Since v0.13.2.1: revert to plotting \(C_l\) instead.
  • Since v0.6.2, \(C_l B_l^2\) is plotted instead.
  • Since v0.2.1, the spectra are plotted without the dof factor of \(\frac{l (l+1)}{2 \pi}\). That is, the previous spectra are the so-called \(D_l\).
  • If you see a + bj, interpret a as the value and b as the error-bar. It is just a shortcut to represent elements of \(\mathbb{R}^2\) as elements of \(\mathbb{C}\).
  • Changelog:
  • v1.0.2: The only bug fix is the patch that fix ground template mentioned below:
    • Dominic found that tweaking how AnalysisBackend construct the ground template can remove the stripes. I applied a patch at AnalysisBackend commit 07b6610. Merge to cleanup?
    • What the patch does:
      • before the patch, a point-source masked TOD is used to construct the ground template
      • with the patch, TOD without point-source mask is used to construct the ground template
  • v1.1.1: (superseded by fix in v1.0.2) Only real-mapmaking for debugging stripes in I-maps: change group template from 0.5 to 1.
  • v1.1: (superseded by fix in v1.0.2) Only real-mapmaking for debugging stripes in I-maps: change group template from 0.24 to 0.5.
  • v1.0.1: add --deconvolve to real mapmaking, everything else identical. That is, sign-flip-noise from real map is not regenerated.
  • v1.0.0: identical to v1.0.0-beta except no. of TTEEBB simulation increased from 358 to 512.
  • v1.0.0-beta:
    • updated theory spectra input using Planck 2018 values and CLASS.
    • point source mask:
      • boolean mask in mapmaking changed to 3.5 arcmin (from default 5 arcmin)
      • map to pseudo-spectra mask changed to Gaussian (from \(C^\infty\) taper with inner and outer radius) for mode-coupling niceness, with FWHM at 10.5 arcmin.
    • removed null tests that involve random splits, reducing from 24 to 21 null tests.
    • 128 TTEE, TTBB simulations; 512 TTEEBB simulations with 358 only included in this report.
  • v0.13.4:
    • zero-padding on maps changed from 1.3x to 2x
    • “naïve” Wiener filter on temperature weights removed
  • v0.13.3: use sign-flip noise in TT/EE/BB/noise simulations
  • v0.13.2.1: null-spectra plotting changing from \(C_l B_l^2\) to \(C_l\)
  • v0.13.2: Fixed a bug that made v0.13.1 wrong. That is, this one has the correct output from --correct_poleff
  • v0.13.1: --correct_poleff is applied to real map to correct for the polarization efficiency which effectively rescale the TOD.
  • v0.13:
    • add point sources: fixed_sources_pccsinlp from pointing/sources.py
    • OLEA projection
    • completely new pipeline for post-mapmaking
      • no intermediate binning, only 1 single binning step is used in the final step
      • mode-coupling:
        • bug fixed
        • with and without point-source mask are compared
      • pseudo-spectra:
        • fully implement no_diag variant for every case (also the cross cases has been unified to avoid bugs)
        • should be numerically more stable (due to how the old one binning things)
      • new filter transfer function definition: \(\widetilde{C}_l = \sum_{\ell'} F_\ell M_{\ell \ell'} B_{\ell'}^2 C_{\ell'}\)
      • there’s only 1 single mask (per temperature/polarization, that is, the final mask) used across all observations (“10”-day cycles). Previously, each observation has their own mask, where as the mode-coupling matrices are calculated once using the final mask, which is what’s done in the low-\(\ell\) pipeline. But since the low-\(\ell\) pipeline calculated the mode-coupling by simulation rather than analytical like the high-\(\ell\) pipeline, this shouldn’t be a problem for low-\(\ell\).
      • post mapmaking was more expensive than mapmaking, and it is now much cheaper. Debug QoS can almost accommodate the whole post-mapmaking calculation.
    • binning is done slightly differently in null-spectra
  • v0.12:
    • use 1 arcmin.
    • pseudo-spectra pipeline requires more memory than Cori I has available per node. So only maps are produced but no spectra.
  • v0.11:
    • largepatch4 datasets
    • merged AnalysisBackend/cleanup/4254c87 to the high-l post-making pipelines
    • applied the high-l data selection on top of it: No. of sub-scans: before: 50902628; after: 47995975. ~94% efficiency (requiring pairing would result in only 40405768 sub-scans. ~84% efficiency.)
  • v0.10: Same as v0.9, except:
    • regenerate null-flags after data-selection to ensure its correctness (on splitting the available data in a half)
    • no data-selection that requires top-bottom pairing.
    • 24 simulations for TTEE/TTBB; 48 simulations for TTEEBB. This is primarily for better p-values, and also as a scaling test.
    • No THS_BHS null-test because of some unknown bug.
  • v0.9.1: same as v0.9, except not using nodiag variant mentioned in v0.6.5.
  • v0.9: similar to v0.8, but apply an intelligent data-selection on correlations, and also add the pairing data-selection from v0.6.
  • v0.8.1: same as v0.8, except not using nodiag variant mentioned in v0.6.5.
  • v0.8: apply data selection to cut half amount of data by relations between detectors. Caveats:
    • this aggressive data-cut is for showing if it does move the null-spectra or not. Different data-selection schemes will be tested to preserve more data.
    • null flags aren’t regenerated (that is, this additional data selection is done after the original data-selection and null-flag generations)
    • LEAK_BY_BOLO, LHS_RHS, THS_BHS cannot be generated (probably just a data volume problem between null-splits)
    • the last l-bin can be very large (again perhaps just data volume problem)
    • QU_PIXEL with fine bin is very bad (perhaps it’s the same problem with the mismatch between fine-bin and large-bin result that we can’t believe the fine-bin spectra)
  • v0.7.1: same as v0.7, except not using nodiag variant mentioned in v0.6.5.
  • v0.7: turn off common-mode filter, otherwise same as v0.6.5.
  • v0.6.6: same as v0.6.4, except not using nodiag variant mentioned in v0.6.5.
  • v0.6.5:
    • back to v0.6.2 config, except
    • in real-maps, replace form_2d_cross_spectra with nodiag variant. If the noise between 2 null-splits are correlated, this would improve it. See Neil’s “Expanding out the QUIET cross null spectrum”.
  • v0.6.4: use roughly quarter volume of data to generate null-spectra, to test if some of the null-spectra is really noise-like (THS_BHS is missing due to an error)
  • v0.6.3: Coadd by season (instead of by 10-day cycle), that is, from 37 observations down to 3. Note that the vertical axis is still \(C_l B_l^2\).
  • v0.6.2: plot \(C_l B_l^2\) instead.
  • v0.6.1: add noise to TTEEBB sim.
  • v0.6:
    • back to v0.3’s configuration, except
    • with most null-tests in v0.2 plus new ones
    • new gain from Anh
    • add data-selection that requires t/b-pairing
    • add a null-test RANDOM_BUT_PAIR_BY_BOLO1 that is random but both t/b-pair always come together.
    • --width changed to 45 (between 45 to 70.3 there’s only tiny bit of signals that’s not worth saving, considering the extra computation requirements that would impose.)
    • changed definition of get_cross_weight in my pseudo-spectra pipeline from \(\frac{1}{w} = \frac{1}{2} \left( \frac{1}{w_1} + \frac{1}{w_2} \right)\) to \(\frac{1}{w} = \frac{1}{w_1} + \frac{1}{w_2}\)
    • (accidentally) no white-noise in the TTEEBB simulations
    • backward incompatible HDF5 maps: new-ish mode-coupling and pseudo-spectra pipelines with a different HDF5 maps structure (for example maps are saved in 2d array.)
  • v0.5:
    • applies heavy data-selection on T, in the hopes that it suppress the I to Q/U leakage in high-\(l\).
    • add TOP_BOTTOM split.
    • (--mask-cross hasn’t been applied due to a misunderstanding of what the code does. But the diff. is at ~4% level in the apodization masks, also known as tpweight/tpmask, which should not make too much difference in the perceived severe failing of the null-tests.)
  • v0.4: identical with v0.3 except for --notch_line_width: 0.04, to see if it helps the null spectra at all.
  • v0.3:
    • config:
      • null-tests covers only the 4 worst failing null-test from v0.2: QU_PIXEL, LEAK_BY_BOLO, AMP_2F_BY_BOLO, AMP_4F_BY_BOLO
        • add RANDOM_BY_BOLO1
      • --groundlr to separately subtract the ground template on L/R scan
      • --notch_lines --notch_line_width: 0.01 instead of --nofft: perform FFT and then applies notch filtering
    • data:
      • Satoru’s new HDF5
      • Anh’s new patch center (should be affecting the output maps only but not the spectra)
      • Anh’s new gain
      • Neil’s notch filter
      • Kolen’s new coadd (then doesn’t save the I/Q/U and have a bunch of computational improvements)
  • v0.2.3.: “gain matched” between the two splits for each spectra
  • v0.2.2:
    • large bin at 300, fine bin at 30 (just above correlation length)
    • add EE \(\chi\)-statistics.