[On dev branch] New BrownianInterval hangs when `dt` is very small about torchsde HOT 7 CLOSED

How small is small? I've tested this down to dt=1e-7 (so 100 times smaller than the default dt_min) and things seem fine. (It terminates in 80 seconds on my machine.)

from torchsde.

Good point, I'm observing values of tb - ta to be as small as 1.7763568394002505e-15. This is really a result of the while loop in base_solver.py still stepping forward despite curr_t being set to ts[-1].

This is clearly due to limited floating point precision, though the precise reason I have not figured out yet.

from torchsde.

Are you saying there's still a bug in base_solver.py somewhere?

This has made me realise, though that we can get an arbitrarily small timestep on the final step. I'll update the heuristics for creating the dependency tree to be a bit smarter.

from torchsde.

That is clearly the case, i.e. curr_t = ts[-1] < out_t=ts[-1].

from torchsde.

Can you provide a repro? Just eyeballing the integration code nothing really looks wrong to me.

from torchsde.

Can you provide a repro? Just eyeballing the integration code nothing really looks wrong to me.

Try running python -m diagnostics.ito_scalar on branch dev-remove-tuple. The problem occurs regardless of which random seed I use on my machine.

I think ultimately, we should step pass the boundary and pull back later on.

from torchsde.

Right, figured it out - this was a bug in BInterval, not in the solver. I'll have a PR in a few moments.

from torchsde.