Proximal Mapping for maximum of linear and quadratic function

968 Views Asked by At

I was wondering if there is an efficient way of calculating the proximal mapping of the following function $f : \mathbb{R}^3 \rightarrow \mathbb{R}$, $b_i \in \mathbb{R}^3$, $c_i \in \mathbb{R}$ : $$ f(y) = \max \{ \frac{1}{2} \| y \|^2 + b_1^T y + c_1, b_2^T y + c_2 \}, $$ i.e. $$\arg \underset{y \in \mathbb{R}^3} \min \, f(y) + \frac{1}{2 \sigma} \| y - y^0 \|^2.$$

Right now I'm considering to use subgradient descent, but I was wondering if there is a more efficient variant.

1

There are 1 best solutions below

0
On

Perhaps someone can find an analytical formula for this prox operator, but if we're going to evaluate it numerically then one option is to reformulate our optimization problem as \begin{align} \operatorname{minimize}_{y,t} & \quad t + \frac{1}{2 \sigma} \|y - y^0 \|^2 \\ \text{subject to} & \quad t \geq b_2^T y + c_2 \\ & \quad t \geq \frac12 \|y \|^2 + b_1^T y + c_1. \end{align} This problem can then be solved using interior point methods. (For example we could solve this using CVX in Matlab.)

This trick of introducing $t$ is used a lot in convex optimization.