Let $M$ be a compact, connected, oriented, smooth Riemannian manifold with non-empty boundary. Let $f:M \to M$ be a smooth orientation preserving isometric immersion.

Is it true that $f(\partial M) \subseteq \partial M$?

Edit: In this paper, it is proved that any "local isometry" from a compact, connected metric space into itself is a homeomoprhism, where by a "local isometry", they mean "locally preserving distance".

Thus, if every isometric immersion is locally preserving distance, then our $f$ will be a homeomorphism (hence will map boundary to boundary).

Is it true that isometric immersion $\Rightarrow$ locally preserving distance?


A positive answer to this question would imply that every smooth orientation preserving isometric imersion $M \to M$ is a Riemannian isometry (see details below**), Moreover, if two manifolds can be isometrically immersed in each other, then they are isometric (see remark 1 in the "updates and remarks" of this question).

In particular, to refute the conjecture "$f(\partial M) \subseteq \partial M$" it is enough to find such an immersion which is not surjective, or at least not an isometry.


(1) Compactness is essential. Look at $M=[0,\infty) \,,\,f(x)=x+1$.

(2) The claim clearly holds in dimension $1$ ($M$ must be a closed interval).

** Indeed, since $f(\partial M) \subseteq \partial M$ and it's easy to see that $f(M^o) \subseteq M^o$ we get that $f(M^o)$ is clopen in $M^o$, thus $f(M^o)=M^o$. Since $f(M)$ is closed in $M$, and contains $M^o$, we conclude $f(M)=M$, and so $f(\partial M)= \partial M,f(M^o)=M^o$.

So, $f$ is a surjective $1$-Lipschitz map from a compact space to itself. (The $1$-Lipshictzity follows since isometric immersions preserve lengths of paths). Thus, it is a metric isometry. Hence, by the positive answer to this question, $f$ is smooth, and in fact a Riemannian isometry.

  • $\begingroup$ How about this: If not, then some $x\in\partial M$ would map to an interior point of $M$. As $f$ is an immersion, there would be a small open set of $M$ not contained in the image of $f$. Therefore the volume of $f(M)$ would be strictly less than the volume of $M$, which contradicts the condition that $f$ is an isometric immersion. $\endgroup$
    – Neal
    Nov 4, 2016 at 18:26
  • 2
    $\begingroup$ Unfortunately, I am not sure this is so simple (but perhaps I am wrong); 1) I am not sure if the image will be an (immersed) submanifold (I only know this is the case for images of injective immersions). 2) I am not sure how you conclude there are points outside the image of $f$, locally around the "bad" boundary point (which is imapped to the interior) this is true- the image will only contain a "half-open" set, but $f$ does not have to be injective, so in theory we can still reach everything. 3) I am not sure $f$ being an immersion implies the volume of $f(M)$ must equal the volume of $M$. $\endgroup$ Nov 4, 2016 at 18:46
  • $\begingroup$ I don't understand your argument if you don't assume injectivity. Why does this not imply that the standard covering maps of the circle are isometries? $\endgroup$
    – mme
    Nov 6, 2016 at 10:46
  • 2
    $\begingroup$ @AsafShachar I found this paper ams.org/journals/proc/1982-085-04/S0002-9939-1982-0660621-7 $\endgroup$
    – erz
    Feb 26, 2017 at 6:43
  • 2
    $\begingroup$ @erz That's an interesting paper. Thanks. I am not sure if every isometric immersion is "locally preserving distance", though. I have edited the question to mention this direction of "attack". $\endgroup$ Feb 26, 2017 at 10:27

3 Answers 3


Let me mimic the argument of Całka:

Note that the map $f$ is short. Set $$X=\bigcap_n\, f^n(M)$$ Note that $f|_X$ is an isometry.

Note that for any positive integer $n$, any small convex ball in the interior of $M$ is mapped by $f^n$ isometrically. In particular $X$ has nonempty interior.

Indeed, a $2{\cdot}r$-ball in the interior is mapped to a $2{\cdot} r$-ball and the map is length-preserving. It follows that the corresponding $r$-ball maps isometrically to an $r$-ball.

Assume $X\ne M$. Fix $z\in M\backslash X$ which lies in a small ball centered in $X$. Note that the distance from $f^n(z)$ to $X$ is constant. On the other hand the set of partial limits of $f^n(z)$ have to lie in $X$, a contradiction.

So, $f$ is an isometry as short map from a compact space to it self. In particular $f(\partial M)=\partial M$

  • $\begingroup$ Anton, I do not understand how you got the uniform lower bound. For instance, why is $X$ not contained in the boundary of $M$? $\endgroup$
    – Misha
    Mar 19, 2017 at 2:52
  • $\begingroup$ @Misha Note that small convex ball in the interior of $M$ is mapped isometrically. In particular $X$ has nonempty interior. $\endgroup$ Mar 19, 2017 at 3:13
  • $\begingroup$ If $f$ is distance preserving on a small convex ball $B$ (in the interior of $M$), is $f$ also distance preserving on $f(B)$? $\endgroup$
    – user_1789
    Mar 19, 2017 at 13:18
  • 1
    $\begingroup$ Thank you, I understand this argument, if $f$ is injective on the $2\cdot r$-ball $B_{2r}(x)$. This injectivity of $f$ is certainly valid, if $r=r(x)$ is sufficiently small. So far it is unclear to me, how this behaves under iteration: If $f$ is injective on $B_{2r}(x)$, is $f$ injective on $f(B_{2r}(x))$? $\endgroup$
    – user_1789
    Mar 19, 2017 at 15:15
  • 1
    $\begingroup$ @AntonPetrunin: Yes, of course, small open balls in the interior map isometrically, I do not see how you conclude that $int(X)\ne\emptyset$. $\endgroup$
    – Misha
    Mar 19, 2017 at 15:20

The following relies on the answer of Anton Petrunin and his explanations.

Set $X:=\cap_{n\geq 1} f^n(M)$.

Claim 0: Each $x\in X$ is a limit point of a sequence $f^n(y)$ for some $y\in M$. Moreover, if $x\in X\cap M^\circ$, then $y$ can be chosen in $M^\circ$.

Proof of Claim 0: Assume $x\in X$. Then $x=f^n(y_n)$ for some $y_n\in M$ and each $n\geq 1$. After passing to a subsequence, we may assume that $y_{n_l}\rightarrow y$.
Fix $\varepsilon>0$ and $K\geq 1$; we show that for some $k\geq K$, $f^k(y)\in B_\varepsilon(x)$.
Choose $l\geq 1$, such that $k:=n_l\geq K$ and $y_{k}\in B_\varepsilon(y)$. Then, since $f$ is distance non-increasing, $dist(f^k(y),f^k(y_k))\leq dist(y,y_k)<\varepsilon$. The 'moreover' statement follows, since we can replace $y$ by $f^{k_0}(y)$, and choose $k_0$ such that $f^{k_0}(y)\in M^\circ$; the latter is possible, since a subsequence of $f^n(y)$ converges to $x\in M^\circ$.

Claim 1: For any $x\in M^\circ$ and $r>0$ such that $B_r(x)\subset M^\circ$ and $n\geq 1$, we have $f^n(B_r(x))=B_r(f^n(x))$.

We defer a proof of Claim 1 to the end.

Claim 2: $X^\circ$ is nonempty, indeed it coincides with the set of all limit points of sequences $f^n(y)$ for $y\in M^\circ$.

Proof Claim 2: By Claim 0, each $x\in X\cap M^\circ \supset X^\circ$ is such a limit point. Conversely, if $x$ is such a limit point, then $x\in B_r(f^k(y))=f^k(B_r(y))\subset f^k(M)$ for (certain) arbitrary large $k$; here we choose $r$ such that $B_r(y)\in M^\circ$ and use Claim 1. Therefore $x\in X=\cap_{n\geq 1} f^n(M)$, since the intersection is decreasing. Actually the same argument shows, that $B_s(x)\subset X$ for each $s<r$ (or also $s\leq r$). As a consequence $x\in X^\circ$.

Claim 3: $X\cap M^\circ$ is open in $M^\circ$. (Clearly it is also closed, as $M$ is compact and nonempty by Claim 2).

Proof Claim 3: Combining Claim 1 and 2, we see that $X\cap M^\circ=X^\circ=X^\circ \cap M^\circ$ is open in $M^\circ$.

As $M^\circ$ is connected, Claim 3 implies:
Claim 4: $X\cap M^\circ=M^\circ$. Since $X\subset M$ is closed, it follows that $X=M$, and thus $f(M)=M$.

Conclusion:Claim 4 implies that $f$ is a distance preserving homeomorphism by Całka, Corollary 3.5 and Corollary 4.4. In particular, $f(\partial M)=\partial M$.

A (standard) proof of Claim 1: First assume n=1: Since $f$ is $1$-Lipschitz, we have $f(B_r(x))\subset B_r(f(x))$ for any $x\in M^\circ$. To show equality, let $y\in B_r(f(x))$ arbitrary. There exists a path (of constant speed) $\gamma:[0,1]\rightarrow M^\circ$ of length $<r$ with $\gamma(0)=f(x)$ and $\gamma(1)=y$.

There exists a unique lifting of $\gamma$, i.e. a path $\eta:[0,1]\rightarrow M^\circ$ with $f\circ \eta=\gamma$ and $\eta(0)=x$.

(The set $T\subset [0,1]$ of $t$'s for which there is a unique lifting of $\gamma$ on the intervall $[0,t]$ contains $0$. $T$ is open since $f$ is a local diffeomorphism on $M^\circ$. In addition, $T$ is closed: If $t_k\in T$ converges to $t\in [0,1]$, then there is a unique lift $\eta_0$ of $\gamma$ on $[0,t)$. Since $M$ is compact and $\eta_0$ is Lipschitz, we can extend it to a path $\eta_1:[0,t]\rightarrow M$, and we have automatically $f(\eta_1(s))=\gamma(s)$ for $s\in [0,t]$. Since $length(\eta_1)=length(\gamma\vert_{[0,t]})<r$ as $f$ is an infinitesimal isometry, we have $\eta_1(t)\in B_r(x)\subset M^\circ$ and therefore $t\in T$, as desired.)

Since $f$ is an (infinitesimal) isometry, length($\eta$)=length$(\gamma)<r$ and thus $\eta(1)\in B_r(x)$.

Therefore $y=f(\eta(1))\in f(B_r(x))$, as claimed for $n=1$.

For arbitrary $n\geq 2$, note that $f^{n-1}(B_r(x))=B_r(f^{n-1}(x))$ by induction.
The domain invariance theorem (applied locally to $f^{n-1}$) implies that $f^{n-1}(M^\circ)\subset M^\circ$; hence if $B_r(x)\in M^\circ$, then so is $f^{n-1}(B_r(x))$. Applying Claim 1 for $n=1$ to the ball $B_r(f^{n-1}(y))$ yields Claim 1 for $n$.


If by an isometric immersion you mean infinitesimally isometric (each tangent mapping is a linear isometry) then the obvious mapping $M_2\to M_1$ in example 4 of the following paper is an isometric immersion which is not surjective and not injective and does not map boundary points to boundary points. There, $M_1$ and $M_2$ are diffeomorphic but not isometric.

  • Franz W. Kamber, Peter W. Michor: Completing Lie algebra actions to Lie group actions. Electron. Res. Announc. Amer. Math. Soc. 10 (2004) 1-10. pdf

So I assume that by an isometric immersion you mean isometric in the sense of distance preserving. But then, $f$ is injective: If $f(x)=f(y)$ for $x\ne y$ we cannot have $0=dist(f(x),f(y)) = dist(x,y)>0$. Thus $f(M)$ is a submanifold with boundary of $M$ and the arguments given in the remarks above show that the measure of $f(M)$ (which equals the measure of $M$) must be smaller that the measure of $M$, if $f$ is not surjective.

  • 1
    $\begingroup$ Yes, I meant an isometric immersion in the infinitesimal sense. Your example is interesting. However, I specifically asked about maps from a Riemannian manifold $M$ to itself. That is, I assume $M_1=M_2$ as Riemannian manifolds. (I already know some "counterexamples" for isometric immersions which map a boundary point to the interior if you allow the manifolds to be different). $\endgroup$ Dec 10, 2016 at 11:27

Your Answer

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

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