Quasi-random sequences

Low-discrepancy sequences for quasi-Monte Carlo integration. Convergence is O(1/N) instead of pseudo-random's O(1/√N), which is decisive for moderate-dimensional integrals.

sobol(n_points, n_dims, seed=None, scramble=True, dtype=torch.float32)

Sobol sequence backed by torch.quasirandom.SobolEngine. Owen scrambling is on by default for better equidistribution. Returns a (n_points, n_dims) tensor in [0, 1).

halton(n_points, n_dims, dtype=torch.float32)

Halton sequence using the first n_dims primes as bases. Simpler than Sobol but degrades above ~20 dimensions — use Sobol for d > 10. Returns a (n_points, n_dims) tensor in (0, 1).

latin_hypercube(n_points, n_dims, dtype=torch.float32, generator=None)

Latin Hypercube Sampling. Stratified — each row and column of the hypercube contains exactly one sample. Better space-filling than pure random for small sample sizes.