Johannes Buchner

If the proposal is bad (too few steps, no jumps are accepted), your live points can thin out to be linearly dependent. Then you get a numerical problem. I think...

Actually, looking back at your error stack trace, I think the issue is elsewhere. update_region tries to compute a new region. Sometimes this fails for numerical reasons (e.g., cannot invert...

Could you try this patch: https://github.com/JohannesBuchner/UltraNest/compare/feature-regionlinalgexceptions Perhaps you can try to retrigger the issue.

I think you are running the code from the repo (perhaps it is in your PYTHONPATH). In that case, you need `python3 setup.py build_ext --inplace`

Please try again, I updated the branch. Good to see you are reproducing it well :-)

What does sampler.live_points_healthy say? Can you dump out what the live points look like (sampler.region.u) and investigate if they are weird (e.g., linearly dependent)? The efficiency printed to stdout should...

Yes, the first region building doesn't have the same try-catch protection at the moment.

Perhaps you can put a few prints in the step sampler code to see what is being proposed, what its likelihood was, and whether it was accepted. Then you should...

This implies the stepsampler made n steps and no proposal was accepted. Very strange if that happens with the slice samplers, because the shrink the interval until a new point...

Perhaps put a print in the two branches of this if: https://github.com/JohannesBuchner/UltraNest/blob/master/ultranest/stepsampler.py#L595 It should tell you the proposed point (unew, pnew) and its likelihood Lnew, and whether it was accepted...