In this paper we develop a new method to identify noncausal AR models
that are driven by non-Gaussian i.i.d. input. Under a few moderate ass
umptions, the necessary and sufficient conditions for rebuilding the A
R models from second and third order statistics are derived. It is sho
wn that the AR model parameters are directly related to the solution o
f an eigenproblem. Based on this approach we present a method of AR mo
del identification, applying eigendecomposition. Unique identification
of AR models is guaranteed up to sign and linear phase ambiguities. M
odel order determination is not crucial in the method. If the order is
overestimated, several equivalent AR models with different linear pha
ses will be obtained.