1: \begin{abstract}

2:   This work considers the problem of estimating the parameters of

3:   \emph{negative mixture models}, i.e. mixture models that possibly

4:   involve negative weights. The contributions of this paper are as

5:   follows. (i) We show that every rational probability distributions

6:   on strings, a representation which occurs naturally in spectral

7:   learning, can be computed by a negative mixture of at most two

8:   probabilistic automata (or HMMs).  (ii) We propose a method to

9:   estimate the parameters of negative mixture models having a specific

10:   tensor structure in their low order observable moments. Building

11:   upon a recent paper on tensor decompositions for learning latent variable models, 

12:   we extend this work to the broader setting

13:   of tensors having a symmetric decomposition with positive \emph{and

14:     negative} weights. We introduce a generalization of the

15:   \emph{tensor power method} for complex valued tensors, and establish

16:   theoretical convergence guarantees. (iii) We show how our approach

17:   applies to \emph{negative Gaussian mixture models}, for which we

18:   provide some experiments.

19: \end{abstract}

20: