Hey everyone,

I am currently running some chains for Power Spectrum, Bispectrum and Joint analysis including cosmology and local PNG.

Within these runs, I experience counter-intuitively different performances (even order of magnitude for mostly Gaussian posteriors) of pocoMC. While the Bispectrum with 11 parameters gives me great performance and converges after 220k likelihood evaluations, corresponding to 46 iterations, Power Spectrum with 16 parameters took 3.5 million steps with 80 iterations, and Joint with 18 parameters still has not converged with more than 2 million evaluations and 76 iterations so far.

The parameter space is for all these runs mostly Gaussian and exhibits just within 3 parameters (related to PNG) a star formation within themselves, which does not seem to leak towards other parameters. The hope was to utilize the preconditioner of pocoMC to let the Normalizing Flow learn that degeneracy and sample the parameter space more efficiently. This seemed to have worked out within the Bispectrum, but unfortunately not in Joint and Power Spectrum.

In all these runs, I am using the same Normalizing Flow and following settings within the CosmoSIS framework:

[pocomc]
n_effective = 512
n_active = 256
flow = nsf6
precondition = True
dynamic = True
n_total = 4096
n_evidence = 0

While we do know that the Power Spectrum exhibits larger degeneracies in some nuissance parameters, typically worsening the convergence within most other samplers without preconditioner, the Joint typically exhibits a similar convergence (at least without PNG) as the Bispectrum. Therefore, I was wondering if in that case the increase of the dimensionality is increasing the necessary steps within the Joint and if there might be a way of accelerating Power Spectrum and Joint to make it comparable with the Bispectrum.

Even after reading through the docs and the background theory, I am still not able to resolve the issue.

Therefore, I wanted to ask for possibly more optimal settings for this specific problem to hopefully increase the performance by a lot.

Thank you very much for your help in advance!

Greetings,
Dennis