simplify the normalization in gpdfit

[avehtari]

Seth Axen pointed out that

w_theta <- 1 / vapply(jj, FUN.VALUE = numeric(1), FUN = function(j) {
  sum(exp(l_theta - l_theta[j]))
})

Is doing normalization in O(n^2), while it can be simplified to O(n) and also made more stable as

w_theta <- exp(l_theta - matrixStats::logSumExp(l_theta))

[ParadaCarleton]

@sethaxen Is this taken from some of the changes I made while writing ParetoSmooth.jl, or is it a separate performance improvement I should add to ParetoSmooth myself? (I keep forgetting what all the weird little apply functions in R do.)

[sethaxen]

@ParadaCarleton no, this isn't from ParetoSmooth.jl. If the improvement isn't in there, it should be.

[jgabry]

This looks good, thanks! I need to investigate why the Windows checks are failing but haven't had a chance yet. I'm guessing it's unrelated to this PR, but want to double check before merging.

[jgabry]

The windows test failure is happening here:

loo/tests/testthat/test_loo_and_waic.R

Lines 123 to 127 in bffab8c

	test_that("loo.array runs with multiple cores", {
	loo_with_arr1 <- loo(LLarr, cores = 1, r_eff = NA)
	loo_with_arr2 <- loo(LLarr, cores = 2, r_eff = NA)
	expect_identical(loo_with_arr1$estimates, loo_with_arr2$estimates)
	})

the error message in the github actions log says that the error occurs because the objects are "equal but not identical". This usually means that the value is the same but some attribute differs, but I can't think of any reason how this PR could cause that.

Does anyone have access to a windows machine? If so could you try running this one test locally?

[codecov-commenter]

Codecov Report

Merging #187 (769f34f) into master (c91f8be) will decrease coverage by 0.00%.
The diff coverage is 100.00%.

❗ Current head 769f34f differs from pull request most recent head 6b87ccb. Consider uploading reports for the commit 6b87ccb to get more accurate results

@@            Coverage Diff             @@
##           master     #187      +/-   ##
==========================================
- Coverage   95.15%   95.15%   -0.01%     
==========================================
  Files          28       28              
  Lines        2663     2662       -1     
==========================================
- Hits         2534     2533       -1     
  Misses        129      129              

Impacted Files	Coverage Δ
R/gpdfit.R	100.00% <100.00%> (ø)

---

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 c91f8be...6b87ccb. Read the comment docs.

[jgabry]

reminder to self: @mike-lawrence nicely offered remote access to a windows machine, so remember to take him up on that offer

[jgabry]

Ok so it's now passing on Windows despite us not changing anything. The only thing I did was wait a few weeks and rerun the github actions workflow. Anyway, I think we can go ahead and merge this.