1: \begin{abstract}

2: This work considers the problem of learning the structure of multivariate

3: linear tree models, which include a variety of directed tree graphical

4: models with continuous, discrete, and mixed latent variables such as

5: linear-Gaussian models, hidden Markov models, Gaussian mixture models, and

6: Markov evolutionary trees.

7: The setting is one where we only have samples from certain observed

8: variables in the tree, and our goal is to estimate the tree structure

9: (\emph{i.e.}, the graph of how the underlying hidden variables are

10: connected to each other and to the observed variables).

11: We propose the Spectral Recursive Grouping algorithm, an efficient and

12: simple bottom-up procedure for recovering the tree structure from

13: independent samples of the observed variables.

14: Our finite sample size bounds for exact recovery of the tree structure

15: reveal certain natural dependencies on underlying statistical and

16: structural properties of the underlying joint distribution.

17: Furthermore, our sample complexity guarantees have no explicit dependence

18: on the dimensionality of the observed variables, making the algorithm

19: applicable to many high-dimensional settings.

20: At the heart of our algorithm is a spectral quartet test for determining

21: the relative topology of a quartet of variables from second-order

22: statistics.

23: \end{abstract}

24: