David Wierichs

Results 63 comments of David Wierichs

I think this is a problem with the `h` keyword argument of `finite_diff`, which defaults to `1e-7`, which is not suitable for `shots=1000` :grimacing: ```python import pennylane as qml import...

The default is indeed `param_shift`, but for `op=qml.prod(qml.I(0), qml.RX(theta, 1))` we have `op.grad_method="F"` (because we added `op.grad_recipe=[none]*op.num_params` in this PR). This makes `param_shift` fall back to `finite_diff` :grimacing:

> Definitely those big values are unexpected I agree that they are not nice. We should fix the handling of `h` in `finite_diff`... >On the other hand, the gradient breaks...