Eggbox problem intractable for Ultranest about ultranest HOT 3 CLOSED

You can try setting max_num_improvement_loops=0 to prevent too much refinement.

from ultranest.

Maybe first analytically calculate the number of posterior modes for the 5D problem to appreciate the difficulty of this problem. Other samplers giving answers quickly does not mean their answers are reliable, which is something you may care about.

from ultranest.

It may be 486 modes in 5 dimensions.

Since it is a product over the parameters, and each axis has a cos(x/2) function which has three +1 peaks,
I get 3^d positive interferences and then as many for negative interferences (-1 * -1 is also a peak).

This can be solved by convergence assuming the modes are similar, but if you set cluster_num_live_points (default 40), then it will want to add more live points for each newly found cluster.

Closing this for now, please reopen if there is something more to discuss / resolve.

from ultranest.