
# The proximal map of total variation parallel

For given data $x\in\M^{\vect{n}}$ of dimension $\vect{n}=(n_1,\ldots,n_m)$, $m\in\mathbb N$, (usually $m=1,2$) this function computes the proximal map of the TV prior in parallel, returning all results in one array.

Let $\Grid = \{1,\ldots,n_1\}\times\{1,\ldots,n_2\}\times\cdots\times\{1,\ldots,n_m\}$ denote the pixel grid the data $x$ is defined on and

the forward neighbors. Then the TV prior is given by

Since each data item $x_{\vect{k}}$ appears in $m$ terms as forward neighbor and in $m$ terms in its “own” sum, computing the proximal map is challenging.

For the cyclic proximal point algorithm let

and analogously $N_{\mathrm{o},j}$ for the odd indices. Then the $2m$ sums of

contains each index at most once.

This function computes the proximal maps in parallel, i.e. returns an array $y\in\M^{\vect{n}\times 2m}$ containing all proximal maps

This function can be extended by carrying a different weight $\lambda_j$ for a difference in dimension $j\in\{1,\ldots,m\}$ and can also be extended to diagonal differences $\vect{k}+\vect{e}_{j_1}+\vect{e}_{j_2}$ when $\lambda_{ij}\neq 0$ is introduced, i.e. the weights $\lambda\in \mathbb R_{>0}^{m\times m}$ are given as a(n upper triangular) matrix.

For an implementation returning a cyclic evaluation of the $2m$ versions see proxTV.

### Matlab Documentation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
% proxParallelTV(M,x,lambda) compute all proxima the tv prior in parallel
% Items containing a NaN will be initalized with their first prox call.
% and the results saved in indepentend arrays concatenated at the first
% signular dimension.
%
% INPUT
%     M         : a manifold
%     x         : an dataset from M^{n_1,m_2,...,n_m} = M^n
%   lambda      : parameter of the proxes given as a vector, an entry for
%                 each dimension, or a matrix, the diagonal for the TV
%                 terms, the offdiagonals for diagonal differences.
%
% OPTIONAL
%   'FixedMask' : binary mask the size of data items in x to indicate fixed
%   data items
%   'DifferenceProx' : (@(x1,x2,lambda)
%   proxAbsoluteDifference(M,x1,x2,lambda))
%                      specify a prox for the even/odd TV term proxes, i.e.
%                      switch the classical TV by Huber.
%
% OUTPUT
%     y     : result of the (parallel) proximal maps, i,e, of size M^m,L,
%             where L is 2*sum(lambda>0), i.e. the number of parallel
%             proxes this term evaluates
% ---
% Manifold-valued Image Restoration Toolbox 1.2 | R. Bergmann | 2018-02-09