# Expectation maximization of moving average with binary source input

I am trying to do blind system identification of a univariate linear FIR model:

I am unsure if the approach is correct or not and any help to further proceed with the maximization will be great. Thank you.

• Your equations (1) and (2) do not make sense to me. How is $u(t)$ used? What is $w(t)$ in the paragraph below (2). It looks like (1) should read something like: $z(t) = h_1 z(t-1) + h_2 z(t-2) + w(t)$ for this to start to make sense... Or is it something else? – Peter K. Jul 1 '15 at 11:56
• I had made a terrible blunder with the notations, thank you for telling me. I have corrected them. – Ria George Jul 1 '15 at 16:08

I think your E-step is correct(only one term missing in the last expression: $-N\ln\sigma_w$ ). To obtain the M-step you have to differentiate with respect to all your parameters. You don't have to include $u(n)$ in $\theta$, since it's defined by $p$. So you compute $\frac{\partial Q}{\partial\theta}=0$. Solving for $\theta$, and this is your new $\theta$ i.e. $\theta_{m+1}$. E.g. for $h$ we gets $\frac{\partial Q}{\partial h}=-(-2\sum y_nz_n^s+2h\sum P_n^s)/(2\sigma_w^2)=0$.
Solving: $h=h_{m+1}=\sum y_nz_n^s(\sum P_n^s)^{-1}$.
-For $p$: $Q'(p)=\sum z_n^s/p -(N-\sum z_n^s)/(1-p)=0,p=\sum z_n^s/N$.
-For $\sigma_w$: $Q'(\sigma_w)=-(\sum y_n^2-2h^T\sum y_nz_n^s+h^Th\sum P_n^s)\cdot (-2/(2\sigma_w^3))-N/\sigma_w=0, \sigma_w^2=(\sum y_n^2-2h^T\sum y_nz_n^s+h^Th\sum P_n^s)/N$