State-Space Subspace Identification methods obtain system estimates
in closed form, and this is in contrast to Maximum Likelihood
methods which, although provably consistent and statistically
efficient, require an iterative approach to solve an optimisation
problem (which is possibly non-convex) over the likelihood surface.
Particularly in signal processing and pattern recognition, the
so-called Expectation Maximisation (EM) method is a popular way of
performing these latter iterations. This paper establishes that a
subspace identification method can, in fact, be viewed as one
iteration of the EM algorithm. As such, a link between subspace and
Maximum Likelihood methods is established.