Can anyone give some guidance with the following question:-
Prove that the Gauss-Newton Method applied to a linear system $Ax=b$ converges in one step to the solution of the normal equations
Can anyone give some guidance with the following question:-
Prove that the Gauss-Newton Method applied to a linear system $Ax=b$ converges in one step to the solution of the normal equations
Let $f$ be a function, in this case $f(x) = A x - b$. Then, with $p_n$ as the $n$-th iterand, this is Newton's method:
$y - f(p_n) = f'(p_n).(x - p_n)\;$ where $\;y = 0\;$ and $\;x = p_{n+1}\;$.
It's easy to see that $f'(x) = A$, so: $$A (p_{n+1} - p_n) = - (A p_n - b) $$ Start with $n=0$. Can you proceed from here?
The normal equations are:
$$ \mathbb{G}(\beta) = X^T(Y-X\beta)$$
The Gauss Newton method has that the root of $\mathbb{G}$ is iteratively found by:
$$\beta^{(1)} = \beta^{(0)} + \left( \frac{\partial \mathbb{G}}{\partial \beta} \right)^{-1} \mathbb{G}(\beta^{(0)})$$
Choose $\beta=0$. This gives you the usual OLS estimate and you can verify by plugging it into $\mathbb{G}$ that it is indeed the root.