joshspeagle
commented
3 years ago Is setting dlogz_init=1e-10 an unrealistic value?
Oh goodness, that's really small. dynesty's default convergence criterion for the first initial run and the use of dlogz in general is to try and estimate the remaining fraction of the evidence that needs to be integrated over. The starting value of 0.01 means the remainder is <1% (and will in practice be less once the live points are "recycled" at the end of a batch/run), so 1e-10 is saying you want to estimate the integral until only <1e-8% remains before you terminate sampling. That leads to runtimes honing in on such a small area that you end up hitting numerical precision and sampling issues (notice that scale is 6.7e-13 when it should be of order ~0.1-ish). You shouldn't need to adjust dlogz to be any smaller than 1e-3 or 1e-4, and even that's if you want your evidence estimates to be really precise.
I'm not too sure how to get dynesty to always converge e.g. for mcmc I'd just increase the walkers and steps but I've found for dynesty that I need to alter the scale of the log-likelihood function - where logl = non-normalised chi2 - otherwise the code runs for a relatively minimal time and the convergence isn't great.
So a couple of things here. "Convergence" in nested sampling means different things from MCMC, and both aren't necessarily always well-defined by practitioners. Nested sampling is interested in estimating the integral (the posterior is a by-product), so "convergence" means you estimate this to some precision. Sometimes you can get excellent estimates of the evidence without very "thorough" sampling of the posterior, which is often what MCMC users think "convergence" means.
Alternately, if you mean it doesn't find the right solution or converges haphazardly or something else entirely, that might indicate other problems and is worth following up on in more detail (posting some trace plots, etc.). If you just want better-sampled posteriors, you can also play around with the stopping criteria in the Dynamic Nested Sampler (e.g., setting an effective sample size based on your MCMC runs after adjusting for the auto-correlation times).
The second is that if by altering the scale of your log-likelihood function you mean something like:
log-like = s (-0.5 chi2), where s is a scale factor
this is not the right thing to do. This is equivalent to raising your likelihood function to some power, which can substantially depart from the assumptions you've made when computing chi2 (e.g., the data is normally distributed relative to your model). If you mean something else, it would be good to get that assumption in more detail to better understand why/how it could impact the performance of the sampler. Thanks!
bjnorfolk
segasai
Hi,
I'm attempting to get dynesty to converge "nicely". My model has 19 parameters and I run it with:
Is setting
dlogz_init=1e-10an unrealistic value?I'm not too sure how to get dynesty to always converge e.g. for mcmc I'd just increase the walkers and steps but I've found for dynesty that I need to alter the scale of the log likelihood function - where logl = non-normalised chi2 - otherwise the code runs for a relatively minimal time and the convergence isn't great. If I leave the logl as just the non-normalised chi2 however, it prints the
Runtime error. Any tips would be greatly appreciated.Here's the full print out (since the code says it may be useful)