Precision and reproducibility

Data typesCopy Anchor Link

The random sketches default to the promoted floating dtype of the inputs, so single-precision inputs stay single precision and complex inputs get complex Gaussian sketches. The result has the same dtype as the inputs:

import numpy as np
import quimb.tensor as qtn

from src_method import compress

mps = qtn.MPS_rand_state(6, bond_dim=8, dtype=np.complex64, seed=0)
out = compress(mps.arrays, chi_out=4, seed=1)
assert all(a.dtype == np.complex64 for a in out)

An explicit dtype overrides the sketch dtype, and can promote the result: for instance, dtype=np.complex128 on single-precision real inputs returns a double-precision complex train.

Complex sketches are circularly symmetric, with real and imaginary parts sharing the unit variance, and draw from the same random stream as real ones: the same seed gives the same sketches in single and double precision up to rounding.

SeedingCopy Anchor Link

SRC is randomized, so two calls generally return slightly different trains. Pass an integer seed for reproducible results:

import numpy as np
import quimb.tensor as qtn

from src_method import compress

mpo = qtn.MPO_rand(6, bond_dim=8, seed=0)
first = compress(mpo.arrays, chi_out=4, seed=7)
second = compress(mpo.arrays, chi_out=4, seed=7)
assert all(np.array_equal(a, b) for a, b in zip(first, second))

ValidationCopy Anchor Link

chi_out must be a positive integer (NumPy integers are accepted, bool is not), and the trains must be open-boundary and agree on their length and the physical dimensions they join. Invalid input raises a TypeError or ValueError naming the culprit before any work is done.