Suppose $M \subseteq \mathbb R^n$ is an $m$-dimensional compact embedded manifold, possibly with boundary, with typically $n$ much larger than $m$. (Realistic values I am interested in: $n$ somewhere between $1000$ and $100000$ and $m$ up to, say, 30.) The manifold $M$ has an induced measure from $\mathbb R^n$.

We call a subspace $V \subseteq \mathbb R^n$ with orthogonal projection $\pi_V : \mathbb R^n \to V$ representative if the set $$ \{m \in M \mid \#\pi_V^{-1}(\pi_V(m)) > 1\} $$ has measure zero. That is, the projection restricted to $M$ is "injective except for a set of measure zero". My question is:

What is the least dimension of such a $V$ in general?

Clearly it must be greater than $m$ (just consider the sphere in $\mathbb R^{m+1}$); does $m+1$ suffice? If not, are $\mathcal O(m)$ dimensions sufficient?

  • $\begingroup$ It's a foundational result in knot theory that two-dimensional $V$ works for $m=1$,$n=3$ - see e.g. math.stackexchange.com/questions/446091/…. My hunch (based mostly on the proof of the Whitney embedding theorem) is that something around $2m$ will work in general. $\endgroup$ – Anthony Carapetis Apr 26 '17 at 13:17
  • 1
    $\begingroup$ In fact I think the proof of the easy Whitney embedding theorem given by Hirsch (Differential Topology, Theorem 3.5) can be very slightly modified (just making the projection orthogonal) to show that $\pi_V|_M$ is an embedding for almost every $(2m+1)$-dimensional subspace $V$. $\endgroup$ – Anthony Carapetis Apr 26 '17 at 13:42
  • $\begingroup$ @AnthonyCarapetis Thank you very much! If you want to put that into an answer, I'd be happy to accept. (Although I still wonder if $n+1$ suffices for reasons of curiosity.) $\endgroup$ – Mees de Vries Apr 26 '17 at 13:55

The proof given of Whitney's easy embedding theorem by Hirsch in Differential Topology works by iterating the following result:

If $M^m$ is an embedded submanifold of $\mathbb R^q$ and $q>2m+1$ then there is a vector $v \in S^{q-1} \setminus \mathbb R^{q-1}$ such that the projection $p_v$ along $v$ on to $\mathbb R^{q-1}$ is an embedding $M^m \to \mathbb R^{q-1}$.

Since $v$ is transverse to $\mathbb R^{q-1}$, the projection along $v$ from $\mathbb R^{q-1}$ on to $v^\perp$ is an isomorphism; and composing this with $p_v$ we get the orthogonal projection $\pi_v : M^m \to v^\perp$, which is thus an embedding. Thus iterating this we get a $2m+1$-dimensional subspace $V$ such that the orthogonal projection on to $V$ is an embedding on $M$.

Now, since we don't need an embedding, we can see how much further similar reasoning can get us. Let's assume we have an embedding $M \to \mathbb R^n$ and try to find a representative projection on to a $n-1$-dimensional subspace.

Let $\Delta = \{(p,p) : p \in M\}$ be the diagonal of $M \times M$ and define the secant map $\sigma: M \times M \setminus \Delta \to S^{n-1}$ by $$\sigma(p,q)=\frac{p-q}{|p-q|}.$$

Then for any direction $v \in S^{n-1}$, $\pi_{v^\perp}(p)=\pi_{v^\perp}(q)$ if and only if $\sigma(p,q)= \pm v$; so the singular set you want to be measure-zero can be described as $\pi_1 X \cup \pi_2 X$, where $X = \sigma^{-1}(v)$ and $\pi_i$ are the projections on to the factors of $M \times M$. Thus it suffices to find a $v$ such that $\dim X < m$.

By Sard's theorem we can choose $v$ to be a regular value of $\sigma$, so that $X$ is a submanifold of $M\times M \setminus \Delta$. Either $X$ is empty and we're done, or the dimension of $X$ is just the dimension of the kernel of $D \sigma$ at a point in $X$. Since $v$ is a regular value we know $D\sigma$ has rank $n-1$, and thus (by rank-nullity) $\dim X = 2m - n + 1$. This is less than $m$ exactly when $n > m + 1$ as you conjectured.

Unfortunately, we can't just iterate this argument like we do with the Whitney embedding theorem, since we need the $M$ we start with to be a submanifold, but the projection might not be a submanifold. Thus if $n >> m$ the best I can do for you is $2m$, by doing this once after applying the embedding theorem.

However, the fact that the single-dimension reduction works all the way down to $m+1$ makes me think it's quite likely that there's some way to get around this, and that $m+1$ dimensions always work. Perhaps some more geometrically intricate argument with a Grassmannian instead of $S^{n-1}$, or perhaps making the iteration work by relaxing the regularity hypotheses. Anyone have any ideas?


Your Answer

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

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