Closed saraedum closed 2 years ago
@fchapoton I hope that this fixes the failing CI runs for your pull requests.
this reveals some failing doctests
@slel @videlec
File "/usr/share/miniconda/envs/test/lib/python3.9/site-packages/surface_dynamics/interval_exchanges/cover.py", line 1091, in surface_dynamics.interval_exchanges.cover.PermutationCover.?
Failed example:
c.lyapunov_exponents_H_plus(isotypic_decomposition=True) # abs tol 0.05
Expected:
[[],
[0.2857, 0.2857],
[0.5714, 0.5714],
[0.2857, 0.2857]]
Got:
[[],
[0.35299143514907344, 0.3543090645855721],
[0.6224970232357228, 0.5916136365203184],
[0.353330129706186, 0.35358205413175214]]
Tolerance exceeded in 5 of 6:
0.2857 vs 0.35299143514907344, tolerance 7e-2 > 5e-2
0.2857 vs 0.3543090645855721, tolerance 7e-2 > 5e-2
0.5714 vs 0.6224970232357228, tolerance 6e-2 > 5e-2
0.2857 vs 0.353330129706186, tolerance 7e-2 > 5e-2
0.2857 vs 0.35358205413175214, tolerance 7e-2 > 5e-2
Does the expected tolerance have any real meaning here or should we just increase it to make these tests pass?
The tolerance sort of have some meaning. This is a Monte Carlo algorithm, so the only thing we can bound is in expectation. Here, it should be rather small.
Anyway, the culprit doctest was a duplicate of the one just above which has an extra parameter nb_iterations=2**19
. The latter does increase the number of iterations (and improve the convergence). I just removed the duplicated doctest in 48c3807.
Checklist
doc/news/
. No, this is not really newsworthy.