To explore the Perturb and Combine idea for estimating probability densities, we study mixtures of tree structured Markov networks derived by bagging combined with the Chow and Liu maximum weight spanning tree algorithm, or by pure random sampling. We empirically assess the performances of these methods in terms of accuracy, with respect to mixture models derived by EM-based learning of Naive Bayes models, and EM-based learning of mixtures of trees. We find that the bagged ensembles outperform all other methods while the random ones perform also very well. Since the computational complexity of the former is quadratic and that of the latter is linear in the number of variables of interest, this paves the way towards the design of efficient density estimation methods that may be applied to problems with very large numbers of variables and comparatively very small sample sizes.