R/normalized.exponential.R

In nickpoison/nltsa: Nonlinear Time Series Analysis

#' Normalized exponential of a vector
#' 
#' Compute exp(x)/sum(exp(x)) in a numerically stable way 
#'
#' Computing exp(x)/sum(exp(x)) can lead to numerical instability if some of the components
#' of x are very large or very small. This procedure recenters by the median to ensure 
#' a bit more stability. It is not bullet-proof but is sufficient for practical cases.
#' 
#' The computation of the median of a vector takes O(N log N), whereas taking the mean takes 
#' O(N). The mean can thus be used instead of the median if this function turns out to be 
#' a bottleneck, but the median is more robust to the few extreme components that typically 
#' cause numerical instability.
#'
#' @param x Vector 
#' @return exp(x)/sum(exp(x))
#' 
#' @export
normalized.exponential <- function(x) {
  x <- x-max(x)
  ex <- exp(x)
  ex/sum(ex)
}