Handle callback dt at t=0 separately

[ChrisRackauckas]

This is a simple fix to the zero callback in https://discourse.julialang.org/t/diffeq-tolerances-on-continuouscallbacks/53181

[devmotion]

Why test for zero here specifically and not the initial time point? And it seems maybe the problem could also be solved without this branch by specifying a smaller dt?

[ChrisRackauckas]

The problem is due to eps(integrator.t) at t==0 being absurdly small.

[devmotion]

Oh I got the impression the problem was that abst was too large before from the discussion on discourse. Since you removed one part of the max call the new value should be less than or equal to the old value - but if it the eps part was too small it should actually be the same, shouldn't it?

[ChrisRackauckas]

The only reason that other part is there is because of you use eps(integrator.t) at zero you get:

julia> eps(0.0)
5.0e-324

which isn't helpful. At all other spots it's adapted to the relative size of the values in the problem, so it's a good estimate. So that max was there to cover that other case. But it seems like integrator.dt/10000 ended up dominating the max even when it shouldn't, so I've relegated it to exactly the one spot where it matters.

[devmotion]

Ahh now I understand, thanks!

[klaff]

Pardon my ignorance, but is this the right way to pick up the changes in master branch of DiffEqBase?

begin # define independent package environment
	import Pkg
	Pkg.activate(mktempdir())
	Pkg.add("Plots"); using Plots
	Pkg.add(name="DiffEqBase",rev="master");
	Pkg.add(name="DifferentialEquations"); using DifferentialEquations
	Pkg.add("PlutoUI"); using PlutoUI
end

I'm not seeing a change in the behavior.

[ChrisRackauckas]

I just tagged. Get the new version.

[klaff]

Much better on the bouncing ball. I still have a similar failure in my "real" simulation which uses VectorContinuousCallback. I haven't found an error in my code yet but it's much more complicated so not a good MWE.

EDIT: The bouncing ball MWE rewritten with VectorContinuousCallback also works correctly (by which I mean it doesn't miss the cross until the time between bounces approaches eps(t)).

[klaff]

Hi Chris, I have another example which still fails with DiffEqBase = v6.53.5. Could still be my error but I've stared at it too long and need some other eyes on it. I would also like to discuss whether I'm approaching this in a reasonable fashion or am I trying to use a screwdriver as a hammer.

Where would be best to post this example? (Here, new issue, new discourse thread, ?) Thanks.

[ChrisRackauckas]

New issue.