Time warping is performed via a composition of warplets. The Bayesian model starts with one warplet and adds warplets one at a time until the warping action becomes negligible in the sense of having almost zero intensity or too narrow domains. The posterior distributions are used as prior distributions for the extended model in the next step. Warplets have an immediate interpretation as warping functions and the inverse warplet is trivial to obtain.
The Bayesian procedure starts with one warplet in the model and uses the posterior distributions as priors for a more extended model with one more warplet. The model is built with adding one warplet at a time and allows for amplitude variations.