Skip to contents

Repair a numeric vector onto the unit simplex

Source: R/statistical_metrics.R
repair_simplex.Rd

Clips numerical under/overflow and renormalises so that \(\mathbf{1}^{\mathsf{T}}\boldsymbol{p}=1\) and \(\boldsymbol{p}\ge\mathbf{0}\). This is a repair / renormalisation step for estimated proportions, not a Euclidean projection onto the simplex and not a statistical-identifiability constraint.

When open = TRUE, entries on the closed-simplex boundary are nudged into the relative interior so that isometric_log_ratio() is defined (ILR / Marquardt / Newton starts). That open-simplex path is the former role of starting_simplex().

Usage

repair_simplex(
  p,
  tolerance = 100 * .Machine$double.eps,
  open = FALSE,
  nms = NULL
)

Arguments

p

Numeric vector \(\boldsymbol{p}\in\mathbb{R}^{J}\).

tolerance

Non-negative tolerance for treating entries as zero (default 100 * .Machine$double.eps).

open

Logical; if TRUE, push the repaired vector off the boundary into the open simplex.

nms

Optional names attached to the returned vector (e.g. cell-type colnames).

Value

Numeric vector on the simplex \(\Delta^{J-1}\) (open when open = TRUE).

See also

compositions::clo() for compositional closure.

Examples

repair_simplex(c(0.2, 0.3, 0.5 + 1e-12))
#> [1] 0.2 0.3 0.5
repair_simplex(c(1, 0, 0), open = TRUE)
#> [1] 1.000000e+00 2.220446e-14 2.220446e-14