tfp.experimental.substrates.jax.math.log1psquare

View source on GitHub

Numerically stable calculation of log(1 + x**2) for small or large |x|.

Aliases:

tfp.experimental.substrates.jax.math.log1psquare(
    x,
    name=None
)

For sufficiently large x we use the following observation:

log(1 + x**2) =   2 log(|x|) + log(1 + 1 / x**2)
              --> 2 log(|x|)  as x --> inf

Numerically, log(1 + 1 / x**2) is 0 when 1 / x**2 is small relative to machine epsilon.

Args:

  • x: Float Tensor input.
  • name: Python string indicating the name of the TensorFlow operation. Default value: 'log1psquare'.

Returns:

  • log1psq: Float Tensor representing log(1. + x**2.).