Select default number of CLAMP latent variables from an SVD

Chooses the default clamp_k used by the CLAMP solvers when the user does not provide one, and returns the scale used for downstream L1/L2 regularization. Multiple methods are available via method:

select_clamp_k(
  svdres,
  n_samples,
  svd_k,
  method = c("elbow", "permutation", "gavish_donoho", "scaleSVs"),
  data = NULL,
  B = 20
)

Arguments

svdres: An SVD result with a d component (output of compute_svd).
n_samples: Integer number of samples in the original matrix (i.e. ncol(Y)). Used by "scaleSVs" and "gavish_donoho".
svd_k: Integer upper bound on clamp_k (number of components actually computed in the SVD).
method: One of "elbow", "permutation", "gavish_donoho", "scaleSVs". Defaults to "elbow".
data: Raw data matrix. Required for "permutation" (row-normalized internally) and "gavish_donoho" (used for n_genes).
B: Number of permutations for "permutation".

Value

An integer: the selected number of latent variables.

Details

"elbow" (default): Elbow heuristic on the singular-value spectrum via num.pc(svdres, method = "elbow"). scale = svdres$d[clamp_k].
"permutation": Permutation test via num.pc(data, method = "permutation", B = B). Requires the raw row-normalized data matrix. scale = svdres$d[clamp_k].
"gavish_donoho": Gavish-Donoho optimal singular-value threshold via PCAtools::chooseGavishDonoho(). Requires the raw data matrix (used for n_genes). scale = svdres$d[clamp_k].
"scaleSVs": Previous behavior: getScaleFromSVs() linear-tail fit, clamp_k <- min(floor(k * 1.5), svd_k), scale from the fit.

Examples

set.seed(1)
Y <- matrix(rnorm(100 * 30), nrow = 100, ncol = 30)
svdres <- compute_svd(Y, k = 25)
select_clamp_k(svdres, n_samples = ncol(Y), svd_k = 25)
#> [1] 22