We present an algorithm for learning a mixture of distributions.
The algorithm is based on spectral projection and
is efficient when the components of the mixture are logconcave
distributions.