Ultranest hangs upon completion

[tomkimpson]

• UltraNest version: 3.5.7
• Python version: 3.9
• Operating System: OSX12.3.1

Description

I am testing out Ultranest, trying to do an inference on a simple 2-parameter likelihood.

Ultranest converges well; I can see from the live point display that the live points are converged very close to the true values, e.g.

Mono-modal Volume: ~exp(-61.48)   Expected Volume: exp(-62.36) Quality: ok

phi0 : +0.0000031416|                            +0.1999999999  *  +0.2000000001                                                                                                                                                                |+1.0000000000
omega: +0.000000010000046051808|                                                                                  +0.000000499999999999999  *  +0.000000500000000000001                                                                                    |+0.000000999995394840417

(true values are 0.20 for phi0 and 5e-7 for omega).

At this point, rather than finishing off, the code just hangs and does not complete.

What I Did

A psuedo example is as follows

def prior_transform(quantile_cube):

    transformed_parameters = np.empty_like(quantile_cube)
    # first parameter: a uniform distribution
    transformed_parameters[0] = 0.0 + np.pi * quantile_cube[0]

    # second parameter: a log-uniform distribution
    transformed_parameters[1] = 10**(-8 + 2 * quantile_cube[1])

    return transformed_parameters

def my_likelihood(params):

    p_copy = guessed_parameters.copy() # this is a dictionary defined globally 

    phi_var, omega_var = params 

    p_copy["phi0_gw"] = phi_var
    p_copy["omega_gw"] = omega_var

    ll = KF.likelihood(p_copy) # this is a function of the class `KF` which accepts a parameters dictionary and returns a likelihood

    return ll

param_names = ['phi0', "omega"]

import ultranest
print("Running sampler")
sampler = ultranest.ReactiveNestedSampler(param_names, my_likelihood, prior_transform)
result = sampler.run()
sampler.print_results()

I expect this could be due to the likelihood surface being very very narrow at the true values of the parameters. In this case, is there are way to force Ultranest to complete gracefully?

[JohannesBuchner]

Looks like you are probably hitting limitations of floating points. Between +0.1999999999 and +0.2000000001, there aren't a lot of values, so the proposal procedure may have difficulty finding new, unique points with higher likelihood. You could try reparametrizing the problem. Is it possible that the likelihood is a bit unrealistic (too informative)?