Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

It is well-known that there are continuous curves $f:I \to \mathbb R^2$ (where $I \subset \mathbb R$ is an interval) whose image have positive measure (e.g Peano curve). I have read somewhere that if we require the curve to be differentiable evrywhere then this cannot happen; but if we require it to be almost everywhere differentiable, then it can happen! How could one proceed to prove the first statement and give a counterexample for the second?

share|improve this question
There is no discrepancy because there measure on $I$ is not related to the measure on $\mathbb{R}^2$ for the purposes of this problem. –  Trevor Wilson Sep 12 '12 at 2:58
My first thought is that the standard Peano and Hilbert space-filling curves appear to be differentiable outside a countable set, being the limits of curves with only finitely many points of non-differentiability. –  Kevin Carlson Sep 12 '12 at 3:00
Yes I did not say there was a disreptancy, I just found the second statement counterintuitive. –  Bernard Sep 12 '12 at 3:01
@KevinCarlson: The (graphs of the) standard nowhere differentiable functions like the blancmange or Weierstrass functions (restricted to a bounded interval) are also uniform limits of curves with only finitely many points of non-differentiability. –  t.b. Sep 12 '12 at 3:03
@t.b. indeed. So much for the easy route. –  Kevin Carlson Sep 12 '12 at 3:15

2 Answers 2

up vote 10 down vote accepted

Let $\phi : [0, 1] \to [0, 1]$ be the Cantor-Lebesgue function, and $\alpha : [0, 1] \to \Bbb{R}^n$ be a space-filling curve.

Since $\phi$ is stationary outside the Cantor set, it is locally constant almost everywhere. That is, $\beta := \alpha \circ \phi$ is also locally constant everywhere, allowing it to be differentiable a.e.. (In fact, $\beta' = 0$ a.e.!) On the other hand, $\phi$ is continuous and surjective. Thus $\beta$ is also a continuous path and the image of $\beta$ coincides with $\alpha$. That is, $\beta$ is also a space-filling curve. Therefore this serves as a counter-example.

Let $f : [0, 1] \to \Bbb{R}^n$ be a differentiable curve in $\Bbb{R}^n$ ($n \geq 2$). Let $\Gamma = f([0, 1])$ be the image of $f$ in $\Bbb{R}^n$.

Assuming that $|f'|$ is Lebesgue integrable, we can prove the first statement.

Theorem. [7.21, Rudin] If $f : [a, b] \to \Bbb{R}$ is everywhere differentiable and its derivative $f'$ is Lebesgue integrable, then $$ f(b) - f(a) = \int_{a}^{b} f'(t) \; dt. $$

This theorem immediately implies the following corollary:

Corollary. Let $f : [0, 1] \to \Bbb{R}^n$ be everywhere differentiable and its derivative $f'$ is Lebesgue integrable. Then for any $\epsilon > 0$, then there exists a partition $\Pi = \{ 0 = x_0 < \cdots < x_N = 1 \}$ such that for $$\epsilon_k = \sup \{|f(x) - f(y)| : x, y \in [x_{k-1}, x_k] \}, \quad (1 \leq k \leq N)$$ we have $$\epsilon_k \leq \epsilon \quad \text{and} \quad \sum_{k=1}^{N} \epsilon_k \leq \| f' \|_1.$$

Proof. Since $f'$ is Lebesgue integrable, it is absolutely continuous. Thus there exists $\delta > 0$ such that whenever a measurable subset $E \subset [0, 1]$ satisfies $|E| < \delta$, we have $\int_E |f'| < \epsilon$. Now let $\Pi = \{x_k\}$ be a partition of $[0, 1]$ into subintervals with length less than $\delta$. Then for each $x_{k-1} \leq x < y \leq x_k$, $$ |f(y) - f(x)| = \left| \int_{x}^{y} f' \right| \leq \int_{x}^{y} |f'| \leq \int_{x_{k-1}}^{x_k} |f'| < \epsilon. $$ Thus $$ \epsilon_k \leq \int_{x_{k-1}}^{x_k} |f'| < \epsilon $$ and the conclusion follows. ////

Remark. If $f'$ is bounded, then it is Lebesgue integrable. Also, in this case, the conclusion of the Corollary follows directly by mean value theorem.

Now let $\epsilon > 0$ and $\Pi$ be a corresponding partition of $[0, 1]$ by Corollary. Then we can cover $\Gamma$ by balls $B_{2\epsilon_k}(f(x_k))$ for $k = 1, \cdots, N$. Thus

$$ |\Gamma| \leq \sum_{k=1}^{N} \left|B_{2\epsilon_k}(f(x_k))\right| \leq \sum_{k=1}^{N} C_n \epsilon_k^n \leq C_n \epsilon^{n-1} \sum_{k=1}^{N} \epsilon_k \leq C_n \epsilon^{n-1} \| f' \|_1, $$

where $C_n$ is a constant depending only on the dimension $n$. (In fact, we can choose $C_n = |B_2|$.) Thus taking $\epsilon \to 0$, we have the desired result.

share|improve this answer
WOW! That was not as difficult as I thought, but still cool. Thanks! So how do we proceed to prove the first statement (which intuitively seems obvious)? –  Bernard Sep 12 '12 at 2:59

The other direction is called (mini)-Sard's Theorem. Page 205 in Appendix 1 of Guilleman and Pollack, Differential Topology. The mini version is just this: Let $U$ be an open set of $\mathbb R^n,$ and let $f:U \rightarrow \mathbb R^m$ be a smooth map. Then, if $m > n,$ we can conclude that $f(U)$ has measure zero in $\mathbb R^m.$

The full Sard's Theorem is about critical points, while we make no demands about the relative dimensions. Let $f:X \rightarrow Y$ be a smooth map of manifolds, and let $C$ be the set of critical points of $f$ in $X.$ Then $f(C)$ has measure zero in $Y.$

They say their proof is almost verbatim from John W. Milnor, Topology from the Differentiable Viewpoint. It appears Guillemin and Pollack, also Milnor, assume $C^\infty.$ However, Milnor refers back to Pontryagin(1955), English translation (1959). Evidently Pontryagin worked with weaker conditions, but we are not told what. http://en.wikipedia.org/wiki/Sard%27s_theorem

Alright, I've been looking stuff up elsewhere. The full Sard's theorem does depend on relative dimension. I currently think that mini-Sard make be true both for $C^1$ and for differentiable everywhere, which is a pretty weak hypothesis.

share|improve this answer
Provided $f$ is at least $C^1$ or some such regularity hypothesis. –  t.b. Sep 12 '12 at 3:17
@t.b., yes, thanks. It would appear the optimal result is Pontryagin (1955). I guess I can order a photocopy, but it takes a few days. –  Will Jagy Sep 12 '12 at 3:25
The regularity assumption $C^1$ for mini-Sard is enough as proved already in Sard's original paper, Theorem 4.1. I don't know off-hand how much this can be weakened. Whitney's example in A function not constant on a connected set of critical points is definitely worth having a look at. –  t.b. Sep 12 '12 at 3:33
@t.b. thanks. I've printed out Sard 1942, Thm. 4.1 says critical points. Critical point is when the rank of the matrix of first partials is smaller than both dimensions... –  Will Jagy Sep 12 '12 at 3:45
Oops... Guess it's bed time for me. Sorry about that. –  t.b. Sep 12 '12 at 3:50

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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