Runtime errors during `param_unconstrain` because of precision issues about bridgestan

[WardBrian]

The tolerance is a hard coded value here: https://github.com/stan-dev/math/blob/46f3c59e63851451ed1bf9b64c8f57c89fa49e8e/stan/math/prim/err/constraint_tolerance.hpp#L12

It would be nice if this was at least a #define so we could override it easily at compile time.

[mjhajharia]

okay makes sense, yeah that's unfortunate, for high dimensional simplexes this would definitely keep creating problems.

[WardBrian]

I think using more sig_figs in CmdStan (or CmdStan offering a binary format which has full double precision) are probably the best “solutions” for now, but it might be worth opening an issue in Stan-Dev/math about making that constant something configurable

[sethaxen]

Is there a compile flag for disabling this check entirely?

[roualdes]

From what I see, there is not a flag for disabling this. @WardBrian would you mind filing an issue in Stan Math?

[WardBrian]

I think the issue @mjhajharia opened for a configurable tolerance is better than a flag, since one could just set it to a very high tolerance and effectively disable it that way

[WardBrian]

The next version of Stan will let this be configured at compile time