How to calculate the subgradient of $|Ax+b|$.

241 Views Asked by At

I know We have the chain rule if $f(x)=g(Ax)$, then $∂f(x)=A^T∂g(Ax)$. I cannot really understand that. Since if the $f(x) = g(Ax)$, then we have $∂f(x)=A^T∂g(Ax) = A^T∂f(x)$.

For example, we have $f(x) = |2x_1 + 3x_2|$, then by this rule we have $∂f(x)=A^T∂g(Ax) = [2,3]^T∂|2x_1 + 3x_2|$, it seems we get back to origin again. Could you calculate the subgradient of the example by this rule in detail?

1

There are 1 best solutions below

0
On BEST ANSWER

The chain rule still applies with appropriate modifications and assumptions, however since the 'inner' function is affine one can compute the subgradient directly.

Let $n(x) = \|x\|$, whatever the norm in question is.

Note that $\nu \in \partial n(x)$ iff $n(x+h)-n(x) \ge \langle \nu, h \rangle $ for all $h$.

Then $f(x+h)-f(x) \ge \langle \xi, Ah \rangle = \langle A^T \xi, h \rangle$ for all $h$ and so we see that $\partial f(x) = A^T \partial n(Ax+b)$.

In the example above, this becomes $\partial f(x) = \begin{cases} \{ - \begin{bmatrix} 2 \\ 3 \end{bmatrix} \}, & 2x_1+3x_2 <0 \\ \{ t \begin{bmatrix} 2 \\ 3 \end{bmatrix} \}_{t \in [-1,1]}, & 2x_1+3x_2 =0 \\ \{ + \begin{bmatrix} 2 \\ 3 \end{bmatrix} \}, & 2x_1+3x_2 >0 \\ \end{cases}$.