unable to understand ultranest corner plot

[shiningsurya]

• UltraNest version: 3.4.4
• Python version: 3.8
• Operating System: centos 7

Description

Given some parameters (3 angles + 1 angular velocity) as initial parameters, model integrates in time and produces some (time, measurement). Objective is to given a set of (time, measurement), estimate the initial parameters which fit the data.

i have simulated a mock dataset (time, measurement) using some set of parameters. Evaluating the likelihood on this dataset using that same set of parameters returns zero.

Ultranest explores until L=-2e4 and produces corner plot like this. corner.pdf The angle parameters have wrap=True. frac_remain is set to 1e-9.

What i expect

i expect ultranest to go until L=1e-9. Is there any way to sample more points during high likelihoods?

In a way, i am using ultranest as an optimization tool to find all the parameters which can fit the data. This is because the model has a lot of degeneracies. Is there a way to get the list of all the high likelihood samples?

Can you please clarify the difference between various types of chains? (equal_weighted_post.txt, weighted_post.txt) Also, is there any documentation on how to interpret the points and tree hdf5? i could not find what various columns of points.hdf5 are.

thanks in advance and sorry for the trouble

[hposborn]

Also, is there any documentation on how to interpret the points and tree hdf5? i could not find what various columns of points.hdf5 are.

I would also like to echo this request - I cannot figure out what the columns in points.hdf5 refer to. I seem to have 2N+1 columns, where N is the number of parameters, but exactly what each column is and how to use them is not documented anywhere.

[JohannesBuchner]

Yes, the output files should be documented better.

• equal_weighted_post.txt: equally weighted posterior samples (similar to a Markov chain)
• weighted_post.txt: posterior samples with a weight attached. This is made by NS directly, and the above is produced from this. However, carrying the weights around is cumbersome.
• points.hdf5: this is sort of an internal file, but the columns are (used here for example https://github.com/JohannesBuchner/UltraNest/blob/master/ultranest/integrator.py#L795): likelihood threshold under which the point was sampled, likelihood of the point, a quality indicator (0 for MLFriends, otherwise the number of steps in the step sampler), u-space (unit cube) coordinates, p-space (transformed parameters) coordinates.

[JohannesBuchner]

Regarding the original question: the posterior looks flat, and sampling at higher likelihood (which have a tiny region) will not change this. Look at your loglikelihood function and check why the values are always the same.

Maybe you have a bug in your loglikelihood function which returns similar values. For example, are you sure you are returning the log of a likelihood?

[JohannesBuchner]

I am updating the docs, and saw that the output files are briefly documented here: https://johannesbuchner.github.io/UltraNest/performance.html

It's a bit hidden, so I intend to make it more discoverable.

[JohannesBuchner]

I think this is fixed with the new docs and latest release. Please try and let me know.