Different seeds can give identical samples (noise seed has only 30 bits) #261

Hi! A quick one that's probably worth settling before 1.0.0, since the fix changes seeded samples.

The sampler turns the user's seed into the noise seed with only 30 bits of entropy (src/tsim/sampler.py):

channel_seed = int(np.random.default_rng(seed).integers(0, 2**30))

seed=None also draws just 30 bits (integers(0, 2**30)). By the birthday bound, a sweep over seeds 0…N has about N²/2³¹ colliding pairs. Seeds 0–300,000, for example, contain 46 pairs that give the same noise. For a noise-only circuit those pairs return bit-identical samples:

import numpy as np, stim, tsim

c = tsim.Circuit(str(stim.Circuit.generated(
    "surface_code:rotated_memory_z", distance=3, rounds=3,
    after_clifford_depolarization=0.01)))
a = c.compile_detector_sampler(seed=37430).sample(100_000)
b = c.compile_detector_sampler(seed=43909).sample(100_000)
np.array_equal(a, b)   # True: both seeds map to channel seed 644760479

With statevector components, the noise columns still match and only the component outcomes differ. Runs that combine several seeds, or treat them as independent, end up double-counting shots.

Possible fix: pass the full seed through, e.g. seed the channel sampler from np.random.default_rng([seed, 0]) or np.random.SeedSequence(seed).spawn(...), and give seed=None a 64-bit draw. That changes the samples you get for a given seed, so it's nicer to do before 1.0.0 than after.

Reproduces on main (e3a465b) and on #259 (a9af140).

🤖 Generated with Claude Code