The linear model, in which a set of observations is assumed to be given by a linear combination of columns of a matrix, has long been the mainstay of the statistics and signal processing literature. One particular challenge for inference under linear models is understanding the conditions on the dictionary under which reliable inference is possible. This challenge has attracted renewed attention in recent years since many modern inference problems deal with the "underdetermined" setting, in which the number of observations is much smaller than the number of columns in the dictionary. This paper makes several contributions for this setting when the set of observations is given by a linear combination of a small number of groups of columns of the dictionary, termed the "block-sparse" case. First, it specifies conditions on the dictionary under which most block subdictionaries are well conditioned. This result is fundamentally different from prior work on block-sparse inference because (i) it provides conditions that can be explicitly computed in polynomial time, (ii) the given conditions translate into near-optimal scaling of the number of columns of the block subdictionaries as a function of the number of observations for a large class of dictionaries, and (iii) it suggests that the spectral norm and the quadratic-mean block coherence of the dictionary (rather than the worst-case coherences) fundamentally limit the scaling of dimensions of the well-conditioned block subdictionaries. Second, this paper investigates the problems of block-sparse recovery and block-sparse regression in underdetermined settings. Near-optimal block-sparse recovery and regression are possible for certain dictionaries as long as the dictionary satisfies easily computable conditions and the coefficients describing the linear combination of groups of columns can be modeled through a mild statistical prior.