Trouble solving for the Jacobian update formula in Broyden's method

253 Views Asked by At

I'm having trouble understanding how to update this formula (it's Broyden's method in multiple dimensions) by solving the following equations:

$$A^{(m)}(x^{(m)}-x^{(m-1)}) = f(x^{(m)})-f(x^{(m-1)})$$ and $$A^{(m)}v=A^{(m-1)}v,$$ where $A^{(m)}$ is a matrix, $x^{(m)}$ is the $m^{th}$ iteration, and for all vectors v orthogonal to $(x^{(m)}-x^{(m-1)})$.

These two equations should lead to $$A^{(m)}=$$ $$A^{(m-1)} + \frac{f(x^{(m)})-f(x^{(m-1)})-A^{(m-1)}(x^{(m)}-x^{(m-1)})}{(x^{(m)}-x^{(m-1)})^T(x^{(m)}-x^{(m-1)})}(x^{(m)}-x^{(m-1)})^T$$.

How does one find the last equation?

1

There are 1 best solutions below

0
On BEST ANSWER

You want $A_{+1}s=d$ while $A_{+1}v=Av$ for $v\perp s$. The second condition means that $A_{+1}$ is a rank-one update of $A$, $$ A_{+1}=A+ws^\top. $$ Now insert this into the first condition $$ d=A_{+1}s=As+ws^\top s\implies w = \frac{d-As}{s^\top s}. $$


The orthogonality comes from the minimization in the Frobenius norm, $A_{+1}$ is there the closest matrix to $A$ such that the first condition holds. As Lagrange function this gives (using the notation $A:B={\rm trace}(AB^T)$) \begin{align} L(X,w)&=\frac12\|X-A\|_F^2-w^\top(Xs-d) \\ &=\frac12(X-A):(X-A)-(ws^\top):X+w^\top d \\ &=\frac12\|X-A-ws^\top\|^2-(ws^\top):A-\frac12(w^\top s)_F^2+w^\top d \end{align} with the stationarity condition $$ 0=(X-A)-ws^\top $$ which implies the invariance on the space orthogonal to $s$.