TanH Bijector

distreqx.bijectors.tanh.Tanh (AbstractFowardInverseBijector, AbstractInvLogDetJacBijector)

A bijector that computes the hyperbolic tangent.

The log-determinant implementation in this bijector is more numerically stable than relying on the automatic differentiation approach used by Lambda, so this bijector should be preferred over Lambda(jnp.tanh) where possible.

When the absolute value of the input is large, Tanh becomes close to a constant, so that it is not possible to recover the input x from the output y within machine precision. In cases where it is needed to compute both the forward mapping and the backward mapping one after the other to recover the original input x, it is the user's responsibility to simplify the operation to avoid numerical issues. One example of such case is to use the bijector within a Transformed distribution and to obtain the log-probability of samples obtained from the distribution's sample method. For values of the samples for which it is not possible to apply the inverse bijector accurately, log_prob returns NaN. This can be avoided by using sample_and_log_prob instead of sample followed by log_prob.

---