Skip to contents

Doubly-stochastic scaling via Sinkhorn-Knopp iterations

Source: R/sinkhorn.R
sinkhorn.Rd

Alternates row and column normalization until the matrix is approximately doubly stochastic. When the chosen backend supports resident broadcast and reduction kernels, the hot loop stays on device via resident_handle and returns a deferred adgeMatrix bound to the resident result.

Usage

sinkhorn(
  A,
  max_iter = 200L,
  tol = 1e-08,
  check_every = 5L,
  eps = 1e-15,
  mode = "fast",
  backend = NULL,
  return_info = FALSE
)

Arguments

A

A dense numeric matrix or adgeMatrix. Sparse inputs are not yet supported in this surface.

max_iter

Maximum number of Sinkhorn iterations.

tol

Convergence tolerance on the maximum row/column sum error.

check_every

Check convergence every check_every iterations.

eps

Floor applied to row/column sums before division.

mode

Execution mode used when coercing a plain matrix. Default "fast" allows accelerated backends to use lower precision.

backend

Backend name used when coercing a plain matrix. Ignored when A is already an adgeMatrix.

return_info

When TRUE, return convergence metadata alongside the scaled matrix.

Value

By default, an adgeMatrix. With return_info = TRUE, a list containing result, iterations, converged, row_error, col_error, backend, and method.

See also

dist_matrix

Examples

A <- abs(matrix(rnorm(16), nrow = 4)) + 0.1
S <- sinkhorn(A, max_iter = 50L)
# Row sums should be close to 1
rowSums(as.matrix(S))
#> [1] 1 1 1 1