Consider a set $\Omega\subseteq\mathbb{R}^n$ being "sufficiently regular", for example being the image of a $C^1$ mapping from $\mathbb{R}^p$ for some $p\ge1$. We may then consider the mapping $$ g:\mathbb{R}^n \to \mathbb{R}^n \\ g(y) = \mathrm{argmin}_{\theta\in\Omega}\|y - \theta\|_2, $$ where $\mathrm{argmin}$ denotes the argument minimum. It may be that the argument minimum is not unique, hopefully, however, it will be in most cases.

I am interested in criteria on $\Omega$ to ensure that:

  1. The argument minimum defining $g$ is Lebesgue almost everwhere unique, such that $g$ is Lebesgue almost surely well-defined.

  2. $g$ is Lebesgue almost everywhere differentiable.

This problem is relevant in the context of for example shape restricted regression, see the article Meyer & Woodroofe (2000), where $\Omega$ is taken to be convex. In the case where $\Omega$ is convex, it is known that $g$ is well-defined, it is then ordinarily known as the "projection onto a convex set", also, $g$ is in this case a contraction and is in particular Lebesgue almost everywhere differentiable by Rademacher's theorem.

As regards the nonconvex case, it is easy to see that $g$ in general will not be everywhere well-defined, nor even everywhere continuous. If, for example, $\Omega$ is $\mathbb{R}^n$ minus a cone, moving $y$ along the axis of a cone, small perturbations of $y$ will make the projection onto $\Omega$ move wildly, and so $g$ will not have any chance of even being continuous on this axis.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.