The distance of from a point $x$ to a set $A$ is defined by $$ d(x,S) = \inf\{d(x,a)\mid a\in A\}, $$ where you may choose the setting to be $\mathbb R^n$, a Banach space or a complete metric space.

This is of interest in constructive analysis and my own theory called abstract Stone duality, as explained below, but I believe that my question is meaningful to classical real analysts, whose intuition I am seeking.

Here $A$ is fixed and I am considering the real valued function $d(x)=d(x,A)$.

My picture is that $A$ is the sea and we are on an island that is foggy (so you can't see the sea) but you can measure the altitude, which is equal to the (nearest) distance to the sea. So the geography of the island consists of conical hills.

Although the altitude is a scalar (it doesn't tell you which way to get to the sea), so long as you're not at the summit of a hill, there is a unique path that goes straight down to the sea on a $45^\circ$ slope.

My conjecture is that $D\cup A$ is dense, where $D$ is the subspace on which $d(x)$ is differentiable.

For any point $x$, even a summit, there is a straight line $[x,a]$ of length $d(x)$, with $a\in A$. Let's call any open segment of such a line a downward path and let $D_0$ be the union of all such segments. Then $D_0\cup A$ is dense.

By simple Euclidean geometry (if we're working in $\mathbb R^n$), any two downward paths through a point of $D_0$ are co-linear.

If $D_0$ is open then it coincides with $D$ and my conjecture follows.

Is this necessarily true?

Here is the context of my question:

In Bishop-style constructive analysis, a subset $S$ of a metric space is said to be located if $f(x,S)$ is a real number and not just a lower real. Many classical theorems become constructively valid once locatedness is added as a hypothesis.

In various forms of constructive topology (including mine, called abstract Stone duality), there is a concept called overtness that is lattice-dual to compactness. For example, familiar results using with words closed, compact and Hausdorff have duals with the words open, overt and discrete.

However, overtness is invisible (vacuously true) in the classical setting, so I am trying to show how it is related to more familiar ideas, by concentrating on the metric case.

What is important is not the overt subspace $A$ qua subspace, but the predicate $\lozenge U$ on open subspaces of whether $U\cap A$ is inhabited. This satisfies $$ \lozenge \bigcup\{U_i\mid i\in I\} \Longrightarrow \exists i\in I. \lozenge U_i. $$

Since $\lozenge$ preserves unions in this sense, we need only consider its values at \emph{basic} open subspaces, which are the balls $B_r(x)$ in the case of a metric space. So we write $$ d(x) \lt r \iff \lozenge B_r(x). $$

This function has the following properties: $$ d(x) < r \iff \exists r'. d(x) < r' < r $$ $$ d(x) < r \land d(x,y)\lt s \Longrightarrow d(y)< r+s $$ $$ d(x) < r \land \epsilon\gt 0 \Longrightarrow \exists y. d(x,y)< r\land d(y)<\varepsilon $$

In fact $d(x)$ is in general only an upper real number, so the function $d$ is upper semi-continuous. They are real-valued and continuous iff the subspace is also closed; for simplicity we assume this here.

The subspace in question is $A = \{a\mid d(a)= 0\}$, where $d(a)=0$ means $\forall\varepsilon\gt 0.d(a)<\varepsilon$.

It is not difficult to show that this function $d$ is exactly the point--set distance $d(x,A)$ above.

However, where this may be of interest to classical real and numerical analysts is this analogy with the Newton--Raphson algorithm:

Newton's idea is that, given an approximant $x_0$ to a solution to $f(x)=0$ where $f$ is differentiable, we get a better approximant $$ x_1=x_0-\delta(x_0) \quad\mbox{where}\quad \delta(x) = f(x)/f'(x), $$ at least in certain good circumstances.

My observation is that $\delta(x)$ is like $d(x)$ above, except that

  • $\delta$ is a vector, but $d$ is a scalar; and

  • $d$ satisfies a triangle law, but $\delta$ doesn't.

Adding the triangle law is easy in logic but not of course in numerical analysis. The reason for asking about differentiability is to turn a scalar into a vector.

I would like then to find some weaker notion than the join-preserving operator that would be an abstraction of the Newton--Raphson algorithm.

The connection between locatedness and overtness above was first considered by Bas Spitters. My draft paper about these topics is Overt Subspaces of ${\mathbb R}^n$

NB This question is not my mathematical territory, so I do not know what are the appopriate tags, so others are welcome to change these.

  • 1
    $\begingroup$ I think this may be related mathoverflow.net/questions/247694/… $\endgroup$ May 17, 2019 at 14:19
  • 3
    $\begingroup$ Not sure how much you care about this case but in an infinite dimensional Banach space the norm (i.e. the distance to the set $\{0\}$) can fail to be differentiable anywhere, depending on what you mean by differentiable. The norm in $\ell^1$ is nowhere Fréchet differentiable, for instance. $\endgroup$ May 17, 2019 at 14:21

1 Answer 1


For finite dimensional spaces, as distance functions are Lipschitz, they are differentiable almost everywhere. So a stronger form of your conjecture is true, namely the complement of $D\cup A$ has measure zero. In infinite dimensions this fails in general as remarked in the comments. I'm not sure what happens for infinite dimensional Hilbert spaces though. Also, the set of points where the distance function to $A$ is non differentiable is known as the medial axis of the complement of $A$, and a great deal is known about its properties.

Note: http://annals.math.princeton.edu/wp-content/uploads/annals-v157-n1-p05.pdf seems to be a relevant reference for positive results in the infinite dimensional case

  • $\begingroup$ Thanks, that's very useful, but I wonder whether you might fill in some details of what you say. $\endgroup$ May 17, 2019 at 15:21
  • 1
    $\begingroup$ about which point? The fact used in the first sentence is called Rademacher's theorem, should I add. $\endgroup$
    – alesia
    May 17, 2019 at 15:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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