
# The proximal map of an absolute difference with Huber relaxation

Proximal map $\prox_{\lambda f}$ for the function $f\colon\M^2\to\R,\quad f(x,y) = h(d_{\M}(x,y))$, where $h$ relaxes the nondifferential point at $d=0$. The function $h$ is given as in [1] by

For small $s$ we obtain a parabola and for large s a linear increase with ascent $\omega\tau\sqrt{2}$. From both the proximal map of an absolute difference and its squared. The closed form of the proximal map also results from these two as

If a data item $x$ or $y$ contains a NaN it is set to the minimizer of the distance, i.e. initialized to the other argument of $f$.

### Matlab Documentation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
% proxAbsoluteDifferenceHuber(M,f1,f2,lambda,tau,omega) prox d(x,y) relaxed
% Given the Huber function
% h(t) = tau^2t^2 for t< omega/(sqrt(2)tau), omega*sqrt(2)*tau*t - omega^2/2
% i.e. for small t a suqared function (steered by tau) and for large t a
% linear part with ascent omega.
% lambda is the prox parameter, as for the absolute difference, this
% function inpaints NaN values and works on arbitrary manifolds
%
% INPUT
%   M         : A manifold
%  f1,f2   : data columns
%  lambda     : parameter within the proximal map
%  tau, omega : parameters of the Huber function
%
% OUTPUT
%  x1,x2   : resulting columns of the proximal map
% ---
% Manifold-valued Image Restoration Toolbox 1.0 ~ R. Bergmann, 2014-10-19