sky_sampling test converges incorrectly with large amount of samples.

[nonl4331]

The sky sampling test checks if the numerically integrated solution from the generated PDF matches the sampled solution. With large amounts of samples (> 10^6) it becomes evident that while they are very close there are slight differences that cannot be attributed to noise.

The test fails in this case as expected. While the test does fail the image produced should still be unbiased since the sampled pdf does appear to integrate to 1. Furthermore, since the sampled result is similar to the numerically integrated solution this should have a negligible impact on performance.

It is likely that this issue is due to floating-point errors. It is unclear from which bit of the code the error is coming from.