Modeling with mixtures is a powerful method in the statistical toolkit that can be used for representing the presence of sub-populations within an overall population. In many applications ranging from financial models to genetics, a mixture model is used to fit the data. The primary difficulty in learning mixture models is that the observed data set does not identify the sub-population to which an individual observation belongs. Despite being studied for more than a century, the theoretical guarantees of mixture models remain unknown for several important settings. In this thesis, we look at three groups of problems. The first part is aimed at estimating the parameters of a mixture of simple distributions. We ask the following question: How many samples are necessary and sufficient to learn the latent parameters? We propose several approaches for this problem that include complex analytic tools to connect statistical distances between pairs of mixtures with the characteristic function. We show sufficient sample complexity guarantees for mixtures of popular distributions (including Gaussian, Poisson and Geometric). For many distributions, our results provide the first sample complexity guarantees for parameter estimation in the corresponding mixture. Using these techniques, we also provide improved lower bounds on the Total Variation distance between Gaussian mixtures with two components and demonstrate new results in some sequence reconstruction problems. In the second part, we study Mixtures of Sparse Linear Regressions where the goal is to learn the best set of linear relationships between the scalar responses (i.e., labels) and the explanatory variables (i.e., features). We focus on a scenario where a learner is able to choose the features to get the labels. To tackle the high dimensionality of data, we further assume that the linear maps are also "sparse", i.e., have only few prominent features among many. For this setting, we devise algorithms with sub-linear (as a function of the dimension) sample complexity guarantees that are also robust to noise. In the final part, we study Mixtures of Sparse Linear Classifiers in the same setting as above. Given a set of features and the binary labels, the objective of this task is to find a set of hyperplanes in the space of features such that for any (feature, label) pair, there exists a hyperplane in the set that justifies the mapping. We devise efficient algorithms with sub-linear sample complexity guarantees for learning the unknown hyperplanes under similar sparsity assumptions as above. To that end, we propose several novel techniques that include tensor decomposition methods and combinatorial designs.