Derivation of an EM algorithm for constrained and unconstrained multivariate autoregressive state-space (MARSS) models

This report presents an Expectation-Maximization (EM) algorithm for estimation of the maximumlikelihood parameter values of constrained multivariate autoregressive Gaussian state-space (MARSS) models. The MARSS model can be written: x(t)=Bx(t-1)+u+w(t), y(t)=Zx(t)+a+v(t), where w(t) and v(t) are multivariate normal error-terms with variance-covariance matrices Q and R respectively. MARSS models are a class of dynamic linear model and vector autoregressive model state-space model. Shumway and Stoer presented an unconstrained EM algorithm for this class of models in 1982, and a number of researchers have presented EM algorithms for specic types of constrained MARSS models since then. In this report, I present a general EM algorithm for constrained MARSS models, where the constraints are on the elements within the parameter matrices (B,u,Q,Z,a,R). The constraints take the form vec(M)=f+Dm, where M is the parameter matrix, f is a column vector of xed values, D is a matrix of multipliers, and m is the column vector of estimated values. This allows a wide variety of constrained parameter matrix forms. The presentation is for a time-varying MARSS model, where timevariation enters through the xed (meaning not estimated) f(t) and D(t) matrices for each parameter. The algorithm allows missing values in y and partially deterministic systems where 0s appear on the diagonals of Q or R.