Learn a standardizer from a matrix

A column with no spread cannot be divided by its standard deviation, so it keeps scale 1. Pass what to hear about it: the guard is silent otherwise, and a constant column is worth a word because nothing downstream will complain. Training converges, the posterior looks plausible, and the coordinate does nothing.

Usage

fit_standardizer(x, eps = 1e-08, what = NULL)

Arguments

x

The matrix to learn from.

eps

Standard deviations below this count as no spread.

what

Name of the argument x came from ("theta" or "x"), used in the warning. NULL, the default, warns about nothing. The standardize = FALSE path in prepare_simulations() builds a degenerate standardizer from a one-row zero matrix on purpose, and the diagnostics standardize draws they generated themselves.