Translation models revisited

Translation models have been defined as memoryless mappings of Gaussian elements which match exactly/approximately target marginal distributions/correlations. We extend this class of translation models to include memoryless mappings of non-Gaussian elements. It is shown that quantities of interest inferred from equivalent translation models, i.e., models which share the same marginal distributions and have similar second moments, can differ significantly. It is suggested to construct families of equivalent translation models and select members of these families which are optimal for given quantities of interest.