sparsax.lu_backend

sparsax.lu_backend(Ai, Aj, n, backend='auto', workload='solve')[source]

Choose the faster LU backend for a sparsity pattern.

KLU’s left-looking factorization wins on sparse graphs, such as contiguity and few-neighbour k-NN weights. UMFPACK’s multifrontal one, which hands dense fronts to BLAS, wins once those fronts grow. One number decides between them: KLU’s flops per factor entry, roughly the size of a typical front. Mean degree does not decide it, because the crossover degree falls as n grows.

The choice depends on the pattern alone, never on the values or on a timing, so a fixed seed reproduces the same draws: a sampler turns a 1e-15 difference in rounding into a different chain within a few dozen sweeps. Measuring the pattern costs one KLU factorization, once per pattern per process. The thresholds come from an Apple M3 Max; on other hardware they may shift, which moves near-ties.

Parameters:
Ai

COO pattern, concrete rather than traced. Duplicates and the order of entries do not matter.

Aj

COO pattern, concrete rather than traced. Duplicates and the order of entries do not matter.

n

matrix dimension.

backend='auto'

"auto" (default), or "klu" or "umfpack" to pin one.

workload='solve'

what the factors are mostly for: "solve" (default) or "logdet".

Returns:

LUBackend, whose solve, factor, solve_factor, logdet, logdet_factor, slogdet and rcond have the lu_* signatures. Tokens belong to the backend that issued them.