Maximum Mean Discrepancy (MMD) with Gaussian RBF Kernel

Evaluates the distributional discrepancy between empirical reference and generated/simulated single-cell high-dimensional profiles or latent embeddings.

Usage

calc_mmd(
  ref_mat,
  sim_mat,
  cells_as_cols = TRUE,
  sigma = NULL,
  max_cells = 500,
  max_features = 1000
)

Arguments

ref_mat

Matrix of reference cells (cells x features or features x cells).

sim_mat

Matrix of simulated cells (cells x features or features x cells).

cells_as_cols

Logical, whether cells are columns. If TRUE (default for scRNA-seq), the matrix is transposed so rows represent cells.

sigma

Optional kernel bandwidth. If NULL, uses the median pairwise distance heuristic.

max_cells

Maximum number of cells to subsample for speed (default 500).

max_features

Maximum number of features to retain for speed (default 1000).

Value

A list with MMD, squared MMD, and the kernel bandwidth sigma used.

Examples

ref <- matrix(stats::rnorm(100), 50, 2)
sim <- matrix(stats::rnorm(100), 50, 2)
calc_mmd(ref, sim)
#> $mmd
#> [1] 0.05175825
#> 
#> $mmd_squared
#> [1] 0.002678917
#> 
#> $bandwidth_sigma
#> [1] 6.336369
#>