Method predict

bayes/src/infRtFlt.cpp:151–203

Bayes_base::Float
 Information_root_scheme::predict (Linrz_predict_model& f, const FM::ColMatrix& invFx, bool linear_r)
/* Linrz Prediction: using precomputed inverse of f.Fx
 * Precondition:
 *   r(k|k),R(k|k)
 * Postcondition:
 *   r(k+1|k) computed using QR decomposition see Ref[1]
 *   R(k+1|k)
 *
 * r can be computed in two was:
 * Either directly in the linear form or  using extended form via R*f.f(x)
 *
 * Requires LAPACK geqrf for QR decomposition (without PIVOTING)
 */
{
	if (!linear_r)
		update ();		// x is required for f(x);

						// Require Root of correlated predict noise (may be semidefinite)
	Matrix Gqr (f.G);
	for (Vec::const_iterator qi = f.q.begin(); qi != f.q.end(); ++qi)
	{
		if (*qi < 0)
			error (Numeric_exception("Predict q Not PSD"));
		column(Gqr, qi.index()) *= std::sqrt(*qi);
	}
						// Form Augmented matrix for factorisation
	const std::size_t x_size = x.size();
	const std::size_t q_size = f.q.size();
						// Column major required for LAPACK, also this property is using in indexing
	DenseColMatrix A(q_size+x_size, q_size+x_size+unsigned(linear_r));
	FM::identity (A);	// Prefill with identity for top left and zero's in off diagonals

	Matrix RFxI (prod(R, invFx));
	A.sub_matrix(q_size,q_size+x_size, 0,q_size) .assign (prod(RFxI, Gqr));
	A.sub_matrix(q_size,q_size+x_size, q_size,q_size+x_size) .assign (RFxI);
	if (linear_r)
		A.sub_column(q_size,q_size+x_size, q_size+x_size) .assign (r);

						// Calculate factorisation so we have and upper triangular R
	DenseVec tau(q_size+x_size);
	int info = LAPACK::geqrf (A, tau);
	if (info != 0)
			error (Numeric_exception("Predict no QR factor"));
						// Extract the roots, junk in strict lower triangle
	R = UpperTri( A.sub_matrix(q_size,q_size+x_size, q_size,q_size+x_size) );
    if (linear_r)
		noalias(r) = A.sub_column(q_size,q_size+x_size, q_size+x_size);
	else
		noalias(r) = prod(R,f.f(x));	// compute r using f(x)

	return UCrcond(R);	// compute rcond of result
}


Bayes_base::Float