Inconsistent test for splines derivative

[GiovanniCanali]

Describe the bug
The derivative tests in test_spline.py and test_spline_surface.py occasionally fail. This is not due to an issue in the spline implementation but to a logical flaw in the tests themselves.

The spline order ranges from 1 to 8, which means the underlying polynomial degree ranges from 0 to 7. When the tests compare derivatives against values computed via automatic differentiation, they sometimes attempt to compute a second derivative of a degree-0 polynomial (a constant). PyTorch’s autograd (called by our grad) returns None for the derivative of a constant, and taking a second derivative of that value raises a TypeError (“NoneType is not subscriptable”).

In short: the tests incorrectly assume that higher-order derivatives always exist, but autograd cannot compute second derivatives of constants, leading to sporadic failures.

To reproduce
Modify, for instance, test_spline_surface.py to force the order to be always 1 (degree 0):

orders = [random.randint(1, 1) for _ in range(2)]

[dario-coscia]

Do we want torch to return None or do we want it to return zero if deriving a constant?

[GiovanniCanali]

Do we want torch to return None or do we want it to return zero if deriving a constant?

This is actually a torch.autograd behaviour. To guarantee fully compatibility, I would align to their choice.