fast.svd: Fast Singular Value Decomposition

In corpcor: Efficient Estimation of Covariance and (Partial) Correlation

Description

fast.svd returns the singular value decomposition of a rectangular real matrix

M = U D V',

where U and V are orthogonal matrices with U' U = I and V' V = I, and D is a diagonal matrix containing the singular values (see svd).

The main difference to the native version svd is that fast.svd is substantially faster for "fat" (small n, large p) and "thin" (large n, small p) matrices. In this case the decomposition of M can be greatly sped up by first computing the SVD of either M M' (fat matrices) or M' M (thin matrices), rather than that of M.

A second difference to svd is that fast.svd only returns the positive singular values (thus the dimension of D always equals the rank of M). Note that the singular vectors computed by fast.svd may differ in sign from those computed by svd.

Usage

fast.svd(m, tol)

Arguments

m	matrix
tol	tolerance - singular values larger than tol are considered non-zero (default value: tol = max(dim(m))*max(D)*.Machine$double.eps)

Details

For "fat" M (small n, large p) the SVD decomposition of M M' yields

M M' = U D^2 U'

As the matrix M M' has dimension n x n only, this is faster to compute than SVD of M. The V matrix is subsequently obtained by

V = M' U D^(-1)

Similarly, for "thin" M (large n, small p), the decomposition of M' M yields

M' M = V D^2 V'

which is also quick to compute as M' M has only dimension p x p. The U matrix is then computed via

U = M V D^(-1)

Value

A list with the following components:

d	a vector containing the positive singular values
u	a matrix with the corresponding left singular vectors
v	a matrix with the corresponding right singular vectors

Author(s)

Korbinian Strimmer (https://strimmerlab.github.io).

See Also

svd, solve.

Examples

# load corpcor library
library("corpcor")


# generate a "fat" data matrix
n = 50
p = 5000
X = matrix(rnorm(n*p), n, p)

# compute SVD
system.time( (s1 = svd(X)) ) 
system.time( (s2 = fast.svd(X)) )


eps = 1e-10
sum(abs(s1$d-s2$d) > eps)
sum(abs(abs(s1$u)-abs(s2$u)) > eps)
sum(abs(abs(s1$v)-abs(s2$v)) > eps)

Example output

   user  system elapsed 
  0.037   0.000   0.064 
   user  system elapsed 
  0.022   0.000   0.023 
[1] 0
[1] 0
[1] 0

corpcor documentation built on Sept. 16, 2021, 5:12 p.m.