$\newcommand{\al}{\alpha}$ $\newcommand{\ga}{\gamma}$ $\newcommand{\e}{\epsilon}$

Let $X,Y$ be Riemannian manifolds, such that $\dim(X) > \dim(Y)$.

I am trying to prove the following statement (mentioned by Gromov in his book on metric geometry):

There is no arcwise isometry (i.e length preserving map) from $X$ to $Y$.

However, the naive attempt to prove this hits an obstacle which I do not see how to pass:

Suppose by contradiction $f:X \to Y$ is an arcwise isometry. Then $f$ is $1$-Lipschitz, hence differeniable almost everywhere (by Rademacher's theorem).

Question: Let $p \in X$ be a point where $f$ is differentiable. Is $df_p:T_pX \to T_{f(p)}Y$ an isometry?

(This will imply the claim of course).

Here is what happens when trying to show this naively:

Let $v \in T_pX$, and let $\al:[0,1] \to X$ be a smooth path s.t $\al(0)=p,\dot \al(0)=v$.

Then $|\dot \alpha(s)| = |\dot \alpha(0)|+\Delta(s)=|v|+ \Delta(s)$ where $\lim_{s \to 0} \Delta(s) =\Delta(0)= 0$, thus

$$ (1) \, \, L(\alpha|_{[0,t]})=\int_0^t |\dot \alpha(s)| ds=t|v|+\int_0^t \Delta(s) ds.$$

$\al$ is Lipschitz and $f$ is $1$-Lipschitz, so $\ga:= f \circ \al$ is Lipschitz. By theorem 2.7.6 in the book ``A course in metric geometry'' ( Burago,Burago,Ivanov) it follows that:

$$ (2) \, \, L(\ga|_{[0,t]})=\int_0^t \nu_{\ga}(s) ds, $$

where $\nu_{\ga}(s):=\lim_{\e \to 0} \frac{d\left( \ga(s),\ga(s+\e) \right)}{|\e|}$ is the speed of $\ga$.

Note that $\nu_{\ga}(0)= |\dot \ga(0)|=|df_p(v)|$.

Using the assumption $f$ preserves lengths, we would now like to compare $(1),(2)$ and take derivatives at $t=0$, to get $$|v|=\frac{d}{dt}\left. \right|_{t=0} L(\al|_{[0,t]})=\frac{d}{dt}\left. \right|_{t=0} L(\ga|_{[0,t]})=|df_p(v)|.$$

However, it seems that the last equality is false in general; Even when the speed of a Lipschitz curve and the derivative of its length exist at a point, they do not need to be equal.

It seems that the only thing we can say is that $ \frac{d}{dt}\left. \right|_{t=0} L(\ga|_{[0,t]}) \ge \nu_{\ga}(0) =|df_p(v)|$, so we are left with $|v| \ge |df_p(v)|$ which doesn't help.

Is there a way to "fix" the proof above?

  • 5
    $\begingroup$ Any metric-defined notion of dimension would work. For example Lebesgue covering dimension is defined for metric spaces. It behaves as expected for manifolds and metric isometries of this sort give you a lower bound on the dimension of the target. So it's an immediate contradiction. All the results you need are in Munkres's point-set topology textbook, for example. $\endgroup$ Dec 13, 2016 at 22:35
  • 5
    $\begingroup$ Doesn't this follow from the fact that an isometry is a homeomorphism? $\endgroup$
    – Igor Rivin
    Dec 14, 2016 at 1:21
  • 2
    $\begingroup$ @IgorRivin: Asaf's definition of isometry allows for maps to be not bijections -- he allows covering spaces, for example. It also allows for isometric embeddings. He's trying to make a calculus-ish argument that the domain's dimension needs to be less than or equal the target space's. $\endgroup$ Dec 14, 2016 at 7:19
  • $\begingroup$ @RyanBudney I might be dense, but since $d(f(x), f(y)) = d(x, y),$ how can an isometry fail to be bijective? $\endgroup$
    – Igor Rivin
    Dec 14, 2016 at 15:19
  • $\begingroup$ @IgorRivin These length-preserving maps need not be metric isometries, take for instance $\alpha(t)=(\cos t,\sin t)$. Also, even when they are injective isometries, they are not always surjective ($(x \to (x,0)$). $\endgroup$ Dec 14, 2016 at 16:12

2 Answers 2


Your proof is correct, but you need to add words "amost everywhere" at ane more place.

We use Rademacher's theorem and lemma about length of curve, which says that if a curve parametrized by length then its velocity is 1 almost everywhere, see 2.7.4 in Metric Geometry by Burgo, Burago and Ivanov.

Fix a chart $U_{\subset\mathbb{R}^n}\to X$ and a vector $v\in \mathbb{R}^n$. Let $g$ be the induced metric on $U$.

Note that from the lemma we get that $|d_pv|=|v|_g$ almost everywhere. Repeat the same for $N=\tfrac12\cdot n\cdot(n+1)$ vectors $v_1,\dots,v_N$ in general position. We get $|d_pv_i|=|v_i|_g$ for almost all $p$ and all $i$. It follows that for all $w$,the identity $|d_pw|=|w|_g$ holds for almost all $p$. Hence $\dim X \le \dim Y$.

You may want to check

  • $\begingroup$ Thanks, however, I am still not sure this finishes the problem. As far as I can see, the problem is that we need to know the differential at a single fixed point preserevs lengths of vectors, and I do not see a trivial way to ensure this based on your argument. (I have written another "answer" below where I expressed my concerns explicitly). Also, I think it is interesting to know whether it's true that the differential is an isometry at every point of differentiability. Do you think this is true? $\endgroup$ Dec 14, 2016 at 8:10
  • 1
    $\begingroup$ @AsafShachar for your second question. It is easy to construct a length pereserving map $\mathbb{R}\to\mathbb{R}$ which differential vanish at some point. Passing to product maps, you can do the same for $\mathbb{R}^n\to\mathbb{R}^n$. $\endgroup$ Dec 14, 2016 at 15:16
  • $\begingroup$ @AsafShachar I updated the answer, hope it is clear now. $\endgroup$ Dec 14, 2016 at 15:34
  • $\begingroup$ @Thanks. I thought I understood, though now I see that something is still not clear to me. I agree you can restrict yourself to a chart, and I also see why it would suffice to prove this for a fixed vector $v \in \mathbb{R}^n$. I am still not convinced how exactly do you deduce that (after fixing $v$) , $|df_p(v)|=|v|$ for almost every $p$ in $U$. Can you please elaborate on this? $\endgroup$ Dec 22, 2016 at 11:11
  • $\begingroup$ @AsafShachar At almost all points the differential is defined and LINEAR. Does it help? $\endgroup$ Dec 22, 2016 at 21:15

I am completenig some details based on Anton's answer:

We prove the following theorem:

Let $X,Y$ be Riemannian manifolds, and let $f:X \to Y$ be a length preserving map. Then $df$ is an isometry almost everywhere.

The proof is composed of 3 steps:

Step I:

It is enough to prove the claim where $X=(\mathbb{R}^n,g)$ where $g$ is an arbitrary metric.

Step II:

Assume step I, i.e we consider $f:(\mathbb{R}^n,g) \to Y$. Let $v \in \mathbb{R}^n$ fixed . Then $|df_x(v)|_{f(x)}=|v|_x$ for almost every $x \in \mathbb{R}^n$.

Step III:

Conclude that $df$ is an isometry almost everywhere.

We begin with proofs of steps I,III, since these are easy.

Proof of step I:

$f$ is $1$-Lipschitz, hence by Rademacher's Thm it is differentiable a.e. Since any manifold admits a countable cover of charts, we finished.

Proof of step III:

$f:(\mathbb{R}^n,g) \to Y$.

Let $v_1,...,v_n$ be a basis for $\mathbb{R}^n$. By applying step II for every $v \in Q=\{v_1,...,v_n,v_1+v_2,v_1+v_3,...,v_{n-1}+v_n\}$, we get that $|df_x(v)|_{f(x)}=|v|_x$ for all $v \in Q$ and for almost every $x \in \mathbb{R}^n$. Now it is a simple linear algebra fact that at all these points $x$, $df_x$ is an isometry. Indeed,

$$ \langle df_x(v_i),df_x(v_j) \rangle = \frac{1}{2}(|df_x(v_i)+df_x(v_j)|^2 - |df_x(v_i)|^2 - |df_x(v_j)|^2) = \frac{1}{2}(|df_x(v_i+v_j)|^2 - |v_i|^2 - |v_j|^2) $$ $$ = \frac{1}{2}(|v_i+v_j|^2 - |v_i|^2 - |v_j|^2) = \langle v_i,v_j \rangle.$$

Comment: Here we see why restricting to a chart was so helpful! We can use the "same" basis vectors $v_i$ for different points $x$ (on a general non-parallelizable manifold, this is meaningless).

Proof of step II:

Fix $v \in \mathbb{R}^n$. Then $|df_x(v)|_{f(x)}=|v|_x$ for almost every $x \in \mathbb{R}^n$.


Fix $x \in \mathbb{R}^n$, and define $\alpha_x(t)=x+tv$. Putting $\gamma_x=f \circ \alpha_x$ (this is a Lipschitz path), we get by theorem 2.7.6 that $$ (1): \, \, \nu_{\alpha_x}(t) =\frac{d}{dt}\left. \right|_{t=0} L(\alpha_x|_{[0,t]})=\frac{d}{dt}\left. \right|_{t=0} L(\ga_x|_{[0,t]})=\nu_{\ga_x}(t) \, \, \text{ for almost every $t$}$$

For any path $\beta$ in a Riemannian manifold, at every point of differentiability, it holds that $\nu_{\beta}(t)=|\dot \beta(t)|_{\beta(t)}$.

Specializing this for $\beta = \alpha_x$ (and combining $(1)$) we obtain

$$ (1)': \, \, |v|_{\alpha_x(t)}=\nu_{\ga_x}(t) \, \, \text{ for almost every $t$}$$

Specializing this for $\beta = \gamma_x$ leads to

observation I:

$\nu_{\gamma_x}(t)=|df_{\alpha_x(t)}(v)|_{f(\alpha_x(t))}$ for those $t$ where $f$ is differentiable at $\alpha_x(t)$.

We define the sets

$ B:=\{(x,t) \in \mathbb{R}^n \times \mathbb{R} | \, \, \nu_{\gamma_x}(t)=|v|_{\alpha_x(t)} \, \, \text{ and } \, f \, \text{ is differentiable at } \, \alpha_x(t) \} \subseteq \mathbb{R}^{n+1}, $

$B_x=\{ t \in \mathbb{R} | \, \, (x,t) \in B \} \subseteq \mathbb{R}$ and

$B^t= \{x\in\Bbb R^n: (x,t)\in B\} \subseteq \mathbb{R}^n$.

We also define $h:\mathbb{R}^{n+1} \to \mathbb{R}^n$ by $h(x,t)=\alpha_x(t)$.

$h(B)=\{\alpha_x(t) | \, \, \nu_{\gamma_x}(t)=|v|_{\alpha_x(t)} \, \, \text{ and } \, f \, \text{ is differentiable at } \, \alpha_x(t) \}=\{\alpha_x(t) | \, \, |df_{\alpha_x(t)}(v)|_{f(\alpha_x(t))}=|v|_{\alpha_x(t)} \}$

where the last equality follows from observation I.

Hence, it is enough to show $h(B)^c$ has measure zero in $\mathbb{R}^n$.

This essentially follows from Tonelli's theorem:

Since $h$ is surjective, $h(B)^c \subseteq h(B^c)$. Thus, it is enough to show $\mu(B^c)=0$ in $\mathbb{R}^{n+1}$.

$$ B^c=\{ (x,t) | \, \, \neg ( \nu_{\gamma_x}(t)=|v|_{\alpha_x(t)} ) \} \cup \{ (x,t) | \, \, f \, \text{ is not differentiable at } \, \alpha_x(t) \}:=A \cup D$$

$(1)'$ implies that $\lambda_1(A_x)=0$ for every $x \in \mathbb{R}^n$, hence (by Tonelli) $\mu(A)=0$.

$D^t = \{ x \in \mathbb{R}^n | f \, \text{ is not differentiable at } \, x+tv \} = W -tv$ (where $W \subseteq \mathbb{R}^n$ is the set of points of non-differentiability of $f$), so $\lambda_n(D^t)=\lambda_n(W)=0$

Again, Tonelli implies $\mu(D)=0$.


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.