probabilistic-numerics/probnum

Zero linop matmul bug

timweiland opened this issue · 1 comments

timweiland commented

Describe the bug
pn.linops.Zero does not implement the _matmul semantic of a LinearOperator correctly. Currently, it is implemented via
matmul = lambda x: np.zeros(x.shape, np.result_type(x, self.dtype))

To Reproduce

z = pn.linops.Zero((2, 1))
x = np.array([[42.]])
z @ x

will throw an AssertionError inside of __matmul__ of LinearOperator, because the result has shape (1, 1).

Expected behavior
The result should have shape (2, 1). I think the correct implementation should take the shape of x and replace the value at index -2 with self.shape[0].

System:

  • OS: Arch Linux
  • Python 3.10.6, Probnum: latest commit

Additional context
If this is actually a bug and I'm not misunderstanding something, I can make a quick PR to get this resolved.

JonathanWenger commented

Yes this looks very much like a bug to me. I agree that the expected behavior should return a zero vector of self.shape[0]. A PR would be great, thanks!

  • 👍1