Chevychase

[kmdalton]

This pull request will

• improve quadrature integration for French Wilson scaling
• add updated reference data from pymc3
• add improved tests for posterior parameter estimation

The major point is that we now have a much larger gold standard test data set in the form of Markov chain Monte Carlo integration performed in pymc3. I've ensured that the output of our algorithm stays very close to the MCMC output. I've improved numerical stability and decreased memory requirement by using Gauss-Chebyshev quadrature with degree 100. It is possible we can lower this degree setting if there are performance concerns. Currently, all posterior parameters are within 2.5% of the MCMC values as compared to 6% for the cctbx implementation of the classical algorithm.

I think this implementation constitutes a new state of the art. Two things we could potentially do to improve down the line:

• optimize the integration window size in order to maximize consistency with MCMC
• optimize the degree of the approximation
  Both of these things are set quite arbitrarily right now.

Here are some histograms of percent errors between rs and MCMC for the four posterior parameter estimates:

[codecov-io]

Codecov Report

Merging #48 (71f10b4) into master (e7b8c8f) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master      #48   +/-   ##
=======================================
  Coverage   99.17%   99.17%           
=======================================
  Files          34       34           
  Lines        1330     1341   +11     
=======================================
+ Hits         1319     1330   +11     
  Misses         11       11           

Flag	Coverage Δ
unittests	99.17% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
...alspaceship/algorithms/scale_merged_intensities.py	100.00% <100.00%> (ø)
reciprocalspaceship/dataset.py	99.02% <0.00%> (+0.02%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e7b8c8f...71f10b4. Read the comment docs.

[JBGreisman]

Looks good to me. I'm not crazy about including the standalone plotting script diagnostics.py, but if you find it useful I think its fine. It's at least in the testing > data subdirectory specific to french-wilson so its context is fairly clear.