Daniel@0: function pot = linear_gaussian_to_cpot(mu, Sigma, W, domain, ns, cnodes, evidence)
Daniel@0: % LINEAR_GAUSSIAN_TO_CPOT Convert a linear Gaussian CPD  to a canonical potential.
Daniel@0: % pot = linear_gaussian_to_cpot(mu, Sigma, W, domain, ns, cnodes, evidence)
Daniel@0: %
Daniel@0: % We include any cts evidence, but ignore any discrete evidence.
Daniel@0: % (Use gaussian_CPD_params_given_dps to use discrete evidence to select mu, Sigma, W.)
Daniel@0: 
Daniel@0: odom = domain(~isemptycell(evidence(domain)));
Daniel@0: hdom = domain(isemptycell(evidence(domain)));
Daniel@0: cobs = myintersect(cnodes, odom);
Daniel@0: chid = myintersect(cnodes, hdom);
Daniel@0: cvals = cat(1, evidence{cobs});
Daniel@0: 
Daniel@0: %[g,h,K] = gaussian_to_canonical(mu, Sigma, W);
Daniel@0: Sinv = inv(Sigma);
Daniel@0: g = -0.5*mu'*Sinv*mu + log(normal_coef(Sigma));
Daniel@0: if isempty(W) | (size(W,2)==0)  % no cts parents
Daniel@0:   h = Sinv*mu;
Daniel@0:   K = Sinv;
Daniel@0: else
Daniel@0:   h = [-W'*Sinv*mu; Sinv*mu];
Daniel@0:   K = [W'*Sinv*W  -W'*Sinv';
Daniel@0:        -Sinv*W     Sinv]; 
Daniel@0: end
Daniel@0: 
Daniel@0: if ~isempty(cvals)
Daniel@0:   %[g, h, K] = enter_evidence_canonical(g, h, K, chid, cobs, cvals(:), ns);  
Daniel@0:   [hx, hy, KXX, KXY, KYX, KYY] = partition_matrix_vec(h, K, chid, cobs, ns);
Daniel@0:   y = cvals(:);
Daniel@0:   g = g + hy'*y - 0.5*y'*KYY*y;
Daniel@0:   if length(hx)==0 % isempty(X) % i.e., we have instantiated everything away
Daniel@0:     h = [];
Daniel@0:     K = [];
Daniel@0:   else
Daniel@0:     h = hx - KXY*y;
Daniel@0:     K = KXX;
Daniel@0:   end
Daniel@0: end
Daniel@0: 
Daniel@0: ns(odom) = 0;
Daniel@0: pot = cpot(domain, ns(domain), g, h, K);
Daniel@0: 
Daniel@0: