problem with random seed

[johannesbuchner]

problem with random seed

[jtlz2]

I wonder if https://software.intel.com/en-us/forums/topic/343147 is relevant.

[JohannesBuchner]

Hi,

MultiNest uses a distributed randomization algorithms as far as I know. Try decreasing your tolerance, e.g. to ln(3). Perhaps the parameter estimation is poor due to having few live points at the end.
Is your likelihood shape somehow odd?
If you can create a artificial likelihood function that reproduces this bug that would be very interesting as a test case.

Cheers,
Johannes

[jtlz2]

evtol=math.log(3.0) didn't help.

I doubled the number of live points to 2000 (with evtol as above), which didn't help either - I aborted the run since the number of likelihood calls was so high and convergence was not apparent.

The likelihood is not too weirdly-shaped - certainly the posteriors are unimodal gaussians.

It's true that the same problem happens even when the code is not wrapped with mpirun.

I suppose it's possible the sampler is stopping too early and that it would eventually have converged to the unbiased value.

Any other ideas?

[jtlz2]

This turned out to be an issue with the target sampling efficiency being too high - setting it to 0.1 dropped the actual efficiency to ~ 20 per cent, so that then the parameter estimates converged to the truth.