Draft: Add distribution tests to look for specific error modes #47
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Directly testing that the probability of an event is as expected can be considerably more powerful and can require less memory than performing a Kolmogorov-Smirnov test; however, this approach is much less flexible.
mstoeckl
force-pushed
the
binary-distr-test
branch
from
447b350 to
c85b82a
Compare
1 task
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
CHANGELOG.mdentrySummary
This introduces a file in distr_test to add tests focused on detecting cases where a distribution sampler is biased on a particular event.
Motivation
Some of the distribution sampling issues that I have found are hard to discover with a Kolmogorov-Smirnov or similar general distribution test, because the issues only occur with large input parameters or low probabilities, and would require a lot of time and memory to detect with a general test.
This is marked as draft for now, as the changes are still a bit rough.
Details
The last bit tests systematically confirm that the floating point precision issue for Binomial indeed only occurs at n > 2^53; and that the Geometric sampler has no equivalent issue.
test_binomial_endpointscan be used to confirm that the change in endpoint rounding for the BTPE method in #43 is reasonable -- currently the probability of 0 or 20 being sampled is ~50% higher than expected on Binomial(20, 0.5), while with #43 the event does not occur at a rate clearly distinguishable from the ideal probability.Since rand_distr reexports Bernouilli, it may as well have a distribution test for it.