BUG: Avoid solve_ivp failure when ts is empty

[Patol75]

Reference issue

Closes gh-11700
Closes gh-12100

What does this implement/fix?

While using solve_ivp, if the user provides t_eval, it is possible that execution will fail if no solution evaluations are completed before the solver exits. In particular, gh-11700 and gh-12100 respectively illustrate that the Radau solver fails and a terminal event occurs before any solution evaluation completes. I am proposing here a simple fix, taking advantage of the fact that, in Python, bool([]) is False. As a result, hstack is only executed if elements have been added to the initially empty ts list; it is initially empty only if t_eval is provided.

Additional information

None

[tylerjereddy]

Is it worth adding a test based on one of the issue reports? I suppose it might just be a small loop that would have produced an error before but no longer does?

I did notice that in one of the issues there is contention about whether or not this is a bug (I'm not an expert).

[Patol75]

Well, it is mainly about handling the specific case where the solver exits without having written any solution, a case that both linked issues document. Writing elif ts: instead of else: avoids calling numpy.hstack on an empty list, which otherwise yields an error. Note that this if block occurs at the very end of solve_ivp, right before the return statement. It turns out that OdeResult, which is called by the latter return, is able to handle empty solutions without any changes.

It is definitely a bug. The contention comes from a disagreement about how solve_ivp is used, but it is actually irrelevant to the issue as the error is raised if a terminal event occurs earlier than any t_eval value.

I am pushing a commit with a new test.

[rgommers]

This is indeed a clear bug, and the test looks good to me. The new output seems good:

  message: 'A termination event occurred.'
     nfev: 11
     njev: 0
      nlu: 0
      sol: None
   status: 1
  success: True
        t: []
 t_events: [array([7.])]
        y: []
 y_events: [array([[0.41155802, 0.2420958 ]])]

[Patol75]

It looks like the Linux Tests / Nightly CPython (3.1) (pull_request) test is taking a large amount of time - can someone with the right privileges stop it?

[andyfaff]

I stopped it

[rgommers]

LGTM, thanks @Patol75