I am employing a sliding window technique to compute the sample mean for N (window) predicted residuals (r), instead of a single residual by using a formula defined in adaptive filtering method. When HtimeA(t,:) is the jacobian with size 1xt, RekfA is a scalar - the formula works.
if t > window
sumSekf = 0;
for l = 1:window,
Sekf = (1/window) * HtimeA(t,:)./ RekfA.^(1/2);
Sekf = (1/window) * sum( HtimeA (t-(l-1):t,:)*Ax_ekfA./ RekfA.^(1/2) );
sumSekf=sumSekf + Sekf*Sekf';
CovEnsembleekf=Sekf*Ax_ekfA*P_ekfA(:,:,t-1)*Ax_ekfA*Sekf' + (1/window);
CovDifferenceekf = CovTimeekf - CovEnsembleekf;
The problem arise when RekfA is not a scalar and HtimeA becomes C x C x t. There is an error of 'Transpose on ND array is not defined' for sumSekf=sumSekf + Sekf*Sekf';
Any advise on how to correct the formula for this case? Thank you.