Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Suppose I have a convex function $f(x)$ for which I can easily compute the proximal mapping

prox$_f(z) = \arg\min_{x} f(x) + \frac{1}{2}||x-z||^2_2$

is there a simple expression for the proximal mapping of $f(\mathcal{A}(x))$ where $\mathcal{A}(x)$ is linear? That is, can I express

$\arg\min_{x} f(\mathcal{A}(x)) + \frac{1}{2}||x-z||^2_2$

in terms of prox$_f(z)$?

share|cite|improve this question
up vote 5 down vote accepted

You might want to check the paper 'Proximal Splitting Methods in Signal Processing' by Combettes and Pesquet (*) which reviews the prox and the prox of typical operations on functions (translation, scaling, …).

In that paper, you will find that if $\mathbf A$ is a semi-orthogonal linear transform (i.e.: $\mathbf A\mathbf A^* = \nu \mathbf{Id}$ with $\nu>0$) then there is a closed form of the prox of $f\circ \mathbf A$. I don't think there is one for the general case (I'd be tempted to say that it would be in that paper if it existed!)

The expression is the following:

$$ \mathrm{prox}_{f\circ \mathbf A}(x) = x + \nu^{-1} \mathbf{A}^*\left(\mathrm{prox}_{\nu f}(\mathbf A x) - \mathbf A x\right) \tag{1}$$

Hope this helps a little!

(*) Paper by Combette & al

EDIT if you're interested in proving the above expression, here are the main steps of a possible proof:

  • Express the definition of the problem in a dual form (Fenchel-Rockafellar) and write the first order optimality conditions
  • Use Moreau's decomposition you should get something like $$ \nu^{-1}\mathbf A \mathbf d - y^* = \mathrm{prox}_{(\nu^{-1}f^\star)^\star}(\nu^{-1}\mathbf A\mathbf d)$$ where $y^*$ is the dual variable, and $^\star$ denotes the convex conjugate.
  • express the convex conjugate of $(\nu^{-1}f^\star)$
  • Shake the whole a little and using the optimality conditions, it should boil down to (1).

Another proof is using the fact that $\partial (f\circ \mathbf A) = \mathbf A^*\circ\partial f\circ \mathbf A$ (equality holds since $\mathbf A$ is surjective and hence the strong relint qualification condition holds, cf. Rockafellar's book). With Moreau decomposition $$x - \mathrm{prox}_{f\circ \mathbf A}(x) \in \mathbf A^*\partial f(\mathbf Ax)\in Im(\mathbf A^*)$$ multiply everything by $\mathbf A$ and it boils down to $$\mathbf A \mathrm{prox}_{f\circ\mathbf A}(x) = \mathrm{prox}_{\nu f}(\mathbf A x)$$ you can then use the decomposition of the prox in its orthogonal projection on $Im(\mathbf A^*)$ and on $ker(\mathbf A)$ and using the last equation you should end up with (1).

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.