MAINT: signal: force LTI state-space matrices to have inexact type

[roryyorke]

Reference issue

Closes gh-13125.

What does this implement/fix?

Forces LTI state space arrays A, B, C, and D to have inexact type.

Previously, when matrices were all of integer-type, step, lsim, etc. gave unexpected results.

This will restrict StateSpace to float32, float64, complex64, and complex128 types. So, a possible objection to the change is that it would previously have allowed more general types for ABCD, but I don't know if this is intended or required.

[e-q]

An alternate approach could be to use np.result_type to calculate the correct output dtype, instead of just inheriting A's, in places like this:

scipy/scipy/signal/ltisys.py

Line 2060 in 19acfed

xout = zeros((n_steps, n_states), sys.A.dtype)

[roryyorke]

An alternate approach could be to use np.result_type to calculate the correct output dtype, instead of just inheriting A's, in places like this:

scipy/scipy/signal/ltisys.py

Line 2060 in 19acfed

xout = zeros((n_steps, n_states), sys.A.dtype)

That needs at least, and possibly only, changes in lsim (where you've indicated) and dlsim. lsim2 doesn't need this change. This would be less invasive than what I've proposed, in the sense that it only changes "simulation" (step, lsim, etc.) behaviour. On the other hand, I think the ABCD matrices are naturally real or complex, and making this change in only one place (ignoring the docstring changes) is reasonable.

I don't mind making the other change.

[tylerjereddy]

The PR has been inactive for almost 4 years, I'll bump the milestone, just ping me if it is actually close to ready to merge for branching ~Sunday.