Skip to contents

Pairwise Wasserstein Distances Along Projection Directions Between Several Empirical Distributions

Source: R/sliced_wasserstein.R
compute_all_distances.Rd

This function computes the squared 2-Wasserstein distances between the projections of several empirical directions along specified projection directions. The projection directions can optionally be transformed by means of a linear map. The functions can output the squared distances along each projection direction, or the squared distances averaged over projection directions.

Usage

compute_all_distances(
  distributions,
  thetas,
  A = NULL,
  verbose = TRUE,
  keep_projections = TRUE,
  test_idx = NULL
)

Arguments

distributions

A list of matrices representing empirical distributions.

thetas

A matrix, each row of which represents a projection direction.

A

Optionally, a matrix used to transform each projection direction.

verbose

If TRUE, show progress.

keep_projections

If TRUE, the distance matrix for each projection direction is output. If FALSE, the distance matrices for the different projection directions are averaged.

test_idx

Optionally, a vector of indices. If supplied, skip the distance computations between distribution pairs whose indices both occur in this vector.

Value

If keep_projections = TRUE, a list of squared-distance matrices, one for each projection direction; otherwise, a matrix with the averaged squared distances.

Examples

M1 <- matrix(rnorm(50), ncol = 5)
M2 <- matrix(rnorm(50), ncol = 5)
M3 <- matrix(rnorm(250), ncol = 5)
# Sliced Wasserstein:
my_directions <- generate_directions(20, 5)
compute_all_distances(list(M1, M2, M3), my_directions,
  keep_projections = FALSE, verbose = FALSE)
#>           [,1]      [,2]      [,3]
#> [1,] 0.0000000 0.3726307 0.2292539
#> [2,] 0.3726307 0.0000000 0.2697058
#> [3,] 0.2292539 0.2697058 0.0000000
# Marginal Wasserstein distances:
marginal_wass <- compute_all_distances(list(M1, M2, M3), diag(1, 5),
  keep_projections = TRUE, verbose = FALSE)
marginal_wass[[3]] # along third dimension
#>            [,1]      [,2]       [,3]
#> [1,] 0.00000000 0.4029302 0.09973663
#> [2,] 0.40293015 0.0000000 0.35042965
#> [3,] 0.09973663 0.3504296 0.00000000
# Reweight projection directions
A <- diag(c(4, 0.5, 3, 2, 1))
shear_wass <- compute_all_distances(list(M1, M2, M3), diag(1, 5), A = A,
  keep_projections = TRUE, verbose = FALSE)
shear_wass[[3]]
#>           [,1]     [,2]      [,3]
#> [1,] 0.0000000 3.626371 0.8976297
#> [2,] 3.6263714 0.000000 3.1538668
#> [3,] 0.8976297 3.153867 0.0000000