Bug with infinite timespan in SDEProblem

[eviatarbach]

With ODEProblem, the timespan can be set to be infinite, and this is useful for the iterator interface. For example, this feature is used in DynamicalSystemsBase.jl for creating a tangent integrator. However, doing this with SDEProblem results in the simulation time being set to Inf after one time step.

Example:

f(u, p, t) = [1.0]
p = SDEProblem(f, f, [1.0], (0.0, Inf))
i = init(p, EM(); dt=.02)

Then, running step!(i) results in i.t becoming Inf.

[ChrisRackauckas]

On Julia v1.6 with StochasticDiffEq v6.35? I cannot reproduce.

[eviatarbach]

Thank you for looking into this!

Yes, this is on Julia 1.6.2 with StochasticDiffEq v6.35.0. Weird that it doesn't do the same on your end. I'm pretty sure it's due to this line

@fastmath abs(ttmp - tstop) < 100eps(max(integrator.t,tstop)) ? (integrator.t = tstop) : (integrator.t = ttmp)

in integrator_utils.jl. When tstop is Inf, 100eps(max(integrator.t,tstop)) evaluates to NaN, and since a comparison with NaN is false, it should take the integrator.t = ttmp branch. But since @fastmath isn't supposed to deal with NaNs it gives an unexpected result on my end:

julia> t = 0.0; ttmp = 0.02; tstop = Inf;

julia> abs(ttmp - tstop) < 100eps(max(t,tstop)) ? tstop : ttmp
0.02

julia> @fastmath abs(ttmp - tstop) < 100eps(max(t,tstop)) ? tstop : ttmp
Inf

Maybe it would make sense to add an explicit check for isinf(tstop)?

[ChrisRackauckas]

odd, your test case gives Inf for me. But yes, try adding an isinf for safety.