CircuitMPS.sample with seed returns the same bitstring for every shot

[kevinsung]

What happened?

• Create a CircuitMPS that flips a coin.
• Draw 50 samples from it.
  • Do this with seed=None. Behavior is as expected.
  • Do this with seed set to some random integers. In each instance, the same bit is measured every single time.

What did you expect to happen?

The same bit should not be returned every time. It's a coin flip.

Minimal Complete Verifiable Example

import quimb.tensor

circ = quimb.tensor.CircuitMPS(1)
circ.apply_gate("H", 0)

shots = 50
for seed in [None, 1613, 3024, 7943]:
    print(f"seed = {seed}")
    samples = "".join(circ.sample(shots, seed=seed))
    print(f"\t{samples}")

Relevant log output

seed = None
	11110111110001001110000100011101110001101110101001
seed = 1613
	11111111111111111111111111111111111111111111111111
seed = 3024
	00000000000000000000000000000000000000000000000000
seed = 7943
	00000000000000000000000000000000000000000000000000

Anything else we need to know?

It looks like the same Generator is being recreated before each shot:

quimb/quimb/tensor/tensor_1d.py

Line 3209 in 1960e05

rng = np.random.default_rng(seed)

Instead, the Generator should be created only once, before sampling, and reused for every shot.

This is probably unrelated but I noticed that here quimb sets the random number generator seed in a way that is discouraged:

quimb/quimb/tensor/circuit.py

Line 3260 in 1960e05

np.random.seed(seed)

Instead, a Generator should be created with np.random.default_rng and used.

Environment

• Arch Linux
• Python 3.11.9
• quimb 1960e05

[jcmgray]

Thanks for catching this quickly and the PR! And yes moving to the generator style everywhere would definitely be preferable.

[kevinsung]

Thanks for catching this quickly and the PR! And yes moving to the generator style everywhere would definitely be preferable.

Opened #249 to track moving to generator style.