Parameters (not transformed params) mysteriously taking values of NaN

[Daniel Rosenbloom]

Hi-

I'm running a model in Pystan (v 2.9), and every other iteration samples parameters (not transformed parameters!) with a value of NaN. How could this be?

My code is complicated and I'm still trying to put together a more readable minimal case that produces this error, but I hope that this description is informative enough for someone to point me in the right direction to understand this bug. I have some print statements at the start of the transformed parameters block that output the following parameter values every other update step (they match the initialization values):

SAMP1_purity: 0.914012

SAMP2_purity: 0.247057

SAMP1_frac_of_copiesAfterFirst_with_variant: [0.142037,0.281514,0.0910376,0.143587,0.319623,0.2782,0.280738,0.916913,0.560325,0.509864,0.283772,0.69511,0.65195,0.470809,0.289977,0.820358]

SAMP2_frac_of_copiesAfterFirst_with_variant: [0.715766,0.780833,0.261801,0.39491,0.466047,0.423018,0.0985913,0.433004,0.0576419,0.388036,0.467438,0.400268,0.210212,0.698076,0.320253,0.757758]

mutscale: 7

branching_ratio: 0.290607

subclonal_ccf_threshold: 0.290671

All parameters are bounded by (0,1), except for "mutscale" (any positive real) and "subclonal_ccf_threshold" (lower and upper bounds inputted as data).

Then, every other step my debug statements report that the parameter values have jumped to something like this:

SAMP1_purity: nan

SAMP2_purity: 0.0945089

SAMP1_frac_of_copiesAfterFirst_with_variant: [nan,0.676121,0.193679,0.071821,0.195353,0.451964,0.0692894,0.938367,0.113031,0.177153,0.283759,0.644974,0.78866,0.698519,0.569782,0.589965]

SAMP2_frac_of_copiesAfterFirst_with_variant: [0.918435,0.572717,0.285158,0.347458,0.525633,0.324575,0.183818,0.0963494,0.229748,0.261654,0.750091,0.825402,0.475699,0.923896,0.713328,0.737275]

mutscale: 2.05918

branching_ratio: 0.701423

subclonal_ccf_threshold: nan

These parameter values (appropriately) result in a log-likelihood of NaN and a Metropolis rejection -- after which the sampler returns to the original values, above. Each rejected sample has the same three parameters set to NaN (SAMP1_purity, SAMP1_frac_of_copiesAfterFirst_with_variant[1], and subclonal_ccf_threshold). With each successive failed sample, all other non-NaN parameters get closer to the original initialized value, until I receive this error: "RuntimeError: No acceptably small step size could be found. Perhaps the posterior is not continuous?" Of course, the issue is not that the non-NaN parameters require an impossibly small step size, but that no parameter should ever be set to NaN!

Has anyone ever encountered this sort of behavior before? I've been chopping up my code, but I can't yet isolate what might be causing the issue.

Thanks,

Daniel

[Bob Carpenter]

It'd be easy to create that behavior. You can get NaN
with underflow or overflow and then a divide-by-zero or adding
two infinities in our transforms from unconstrained to constrained.

It's impossible to say more without seeing
your model. A small example would be good.

- Bob

> --
> You received this message because you are subscribed to the Google Groups "Stan users mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.
> To post to this group, send email to stan-...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Daniel Rosenbloom]

Thanks for the advice. It was difficult to build a minimal counterexample that displayed this behavior, but one of the functions used to calculate my likelihood was regularly underflowing. After fixing that function, this particular bug went away.

-Daniel