GPU-native randomized SVD (Halko et al. 2011)

Computes a truncated SVD via randomized projection entirely on the GPU. All QR, matmul, and SVD steps stay on device; a single mlx_eval materializes the results. Falls back to irlba::svdr on CPU if no GPU backend with rsvd support is active.

Usage

rsvd(x, k, n_oversamples = 10L, n_iter = 2L, ...)

Arguments

x: An adgeMatrix or plain numeric matrix.
k: Number of singular values/vectors to compute.
n_oversamples: Extra columns for the random projection (default 10). Increasing this improves accuracy at modest cost.
n_iter: Number of power-iteration passes (default 2). More passes give better accuracy for matrices with slowly decaying spectra.
...: Ignored (for forward compatibility).

Value

A list with components u (m x k), d (length-k singular values, decreasing), and v (n x k).

References

Halko, N., Martinsson, P. G., & Tropp, J. A. (2011). Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions. SIAM Review, 53(2), 217-288.

Examples

A <- matrix(rnorm(200), nrow = 20)
res <- rsvd(A, k = 3L)
length(res$d)
#> [1] 3