I want to find the P matrix that minimizes the function $f(P) = r(P)^Tr(P)$ with $r(P) = PX - x$ using the Gauss-Newton method.
with $X$ being a constant $m \times k$ matrix, $x$ being a constant $n \times k$ matrix, and $P$ being an $n \times m$ matrix.
I got the Jacobian matrix using the method described in here, $J = I_n \otimes X$.
Using the formula on this page, the Gauss-Newton equation should look like:
$$ P_{n+1} = P_n - (J^TJ)^{-1}J^Tr(P_n) $$
using properties of the Kronecker product from here
$$ (J^TJ)^{-1} = ((I_n \otimes X)^T(I_n \otimes X))^{-1} = (I_n \otimes X^TX)^{-1} = I_n \otimes (X^TX)^{-1} $$
and working it out on paper it seemed that: $$ J^Tr(P_n) = (I_n \otimes X)^T r(P_n) = r(P_n) \otimes X^T $$
Combining the two parts we get $$ (I_n \otimes (X^TX)^{-1})(r(P_n) \otimes X^T) = r(P_n) \otimes (X^TX)^{-1}X^T$ $$
$r(P_n)$ is an $n\times k$ matrix and $(X^TX)^{-1}X^T$ is an $ m \times k $ matrix so doing the kron product to them produces a $nm \times kk$ matrix.
Which is so close.
If it was $r(P_n)((X^TX)^{-1}X^T)^T$ instead of $r(P_n) \otimes (X^TX)^{-1}X^T$ it would produce a $n \times m $ matrix which I need for Gauss-Newton method equation from the top to work.
Can someone let me know where I went wrong?
gtransposed in the final equation? SinceHis an mn x mn matrix, andp_n,gare both mnx1 matrices, the update term in the final equation is: (mn x mn)(mn x 1) (mn x 1)(mn x 1) which doesn't work, but if I useg^TI get (mn x mn)(mn x 1) (1 x mn)(mn x 1) which gives me an (mn x 1) vector which is what I want. – Kyle Nov 19 '21 at 02:45ris evaluated atp_n. sor(p_n)is an nk x 1 matrix. andgis also evaluated atp_nsog(p_n)is an nm x 1 matrix. How do I evaluatedH^-1atp_n? – Kyle Nov 19 '21 at 03:24H^-1doesn't rely onp_nso it doesn't do anything,H^-1(p_n) = H^-1. – Kyle Nov 19 '21 at 03:29