In this thesis, we study approximation and learning methods which provide sparse representations. These methods allow to analyze very redundant data-bases thanks to learned atoms dictionaries. Being adapted to studied data, they are more efficient in representation quality than classical dictionaries with atoms defined analytically. We consider more particularly multivariate signals coming from the simultaneous acquisition of several quantities, as EEG signals or 2D and 3D motion signals. We extend sparse representation methods to the multivariate model, to take into account interactions between the different components acquired simultaneously. This model is more flexible that the common multichannel one which imposes a hypothesis of rank 1. We study models of invariant representations: invariance to temporal shift, invariance to rotation, etc. Adding supplementary degrees of freedom, each kernel is potentially replicated in an atoms family, translated at all samples, rotated at all orientations, etc. So, a dictionary of invariant kernels generates a very redundant atoms dictionary, thus ideal to represent the redundant studied data. All these invariances require methods adapted to these models. Temporal shift-invariance is an essential property for the study of temporal signals having a natural temporal variability. In the 2D and 3D rotation invariant case, we observe the efficiency of the non-oriented approach over the oriented one, even when data are not revolved. Indeed, the non-oriented model allows to detect data invariants and assures the robustness to rotation when data are revolved. We also observe the reproducibility of the sparse decompositions on a learned dictionary. This generative property is due to the fact that dictionary learning is a generalization of K-means. Moreover, our representations have many invariances that is ideal to make classification. We thus study how to perform a classification adapted to the shift-invariant model, using shift-consistent pooling functions.
