In this work, the functional approximation of stored energy functions characterizing the homogenized behavior of hyperelastic composites has been addressed by resorting on a probabilistic interpretation. Upon considering the macroscopic Cauchy–Green tensor as a tensor-valued random variable characterized by an information-theoretic model, a surrogate model for the homogenized stored energy function has been obtained through an uncertainty propagation step, making use of a polynomial chaos expansion. The efficiency of the approach was next demonstrated on two benchmark problems, and compared with an interpolation-based scheme proposed elsewhere. The functional framework further allowed for the precise definition of closest approximations (and that of the associated residual) of arbitrary stored energy functions in well-defined classes, such as Ogden-type potentials. The relevance of the methodology, as well as its sensitivity with respect to some critical parameters related to the microstructure (such as the elastic contrast and anisotropy), were finally investigated by considering two multiscale problems and a structural problem relying on a non-concurrent coupling. The definition of such closest approximations is particularly appealing from an engineering point of view, since it allows multiscale information to be readily transferred into commercial finite element codes.