Singular value decomposition for bigmemory::big.matrix inputs

Compute the singular value decomposition (SVD) of a bigmemory::big.matrix without materialising it as a base R matrix. Blocks of rows are streamed through BLAS before LAPACK is invoked so that even moderately large matrices can be decomposed efficiently.

Usage

svd_bigmatrix(
  xpMat,
  nu = -1L,
  nv = -1L,
  block_size = 1024L,
  method = c("dgesdd", "dgesvd")
)

Arguments

xpMat

Either a bigmemory::big.matrix or an external pointer such as mat@address that references the source big.matrix.

nu

Number of left singular vectors to return. Use a negative value to request the default of min(nrow, ncol) vectors and zero to skip returning u entirely.

nv

Number of right singular vectors to return. Use a negative value to request the default of min(nrow, ncol) vectors and zero to skip returning v entirely.

block_size

Number of rows to process per block when streaming data into BLAS kernels. Larger values can improve throughput at the cost of additional temporary memory.

method

LAPACK backend used to compute the decomposition. The default uses the divide-and-conquer routine dgesdd and falls back to dgesvd when required.

Value

A list with components u, d, and v analogous to base R's svd() output. When nu or nv are zero the corresponding matrix has zero columns.

Examples

set.seed(42)
mat <- bigmemory::as.big.matrix(matrix(rnorm(20), nrow = 5))
svd_res <- svd_bigmatrix(mat, nu = 2, nv = 2)
svd_res$d
#> [1] 4.319302 3.283675 1.685888 1.091984