If $f : E \to F$ is a continuous map of Banach spaces, with bounded Fréchet derivative. Then

$x_0,x_1 \in E\Rightarrow \|f(x_1) − f(x_0)\| ≤ M\|x_1 − x_0\|$ where $M = \sup \|f'(x)\|.$

The most efficient way to prove this, as far as I know, is to apply Hahn-Banach. Alternatively, one starts with $g(t)= f(x_0 + t(x_1 − x_0));\ 0\le t\le 1$ and reduces the problem to the real case. But, of course, the Hahn-Banach theorem is a (relatively) big gun, and to do the integral if one uses $g$, then integration in Banach spaces must be dealt with (via regulated functions, for example.)

The only elementary proof I have seen uses the dot product (Apostol, Rudin), but then of course, one must assume an inner product.

Below I sketch a very basic proof, but it required more effort than I expected, so my question is: can one get it cheaper, using only elementary means (very basic facts about normed spaces and the definition of derivative)?

Assume for convenience that $x_0=f(x_0)=0$ and consider the segment $\{tx_1:\ 0\le t\le 1\}.$

We have $\|f(x)\|\le M\|x\|+\|r(x)\|\cdot\|x\|$ where $\frac{\|r(x)\|}{\|x\|}\to 0$ as $x\to 0.$

So, if $\epsilon>0,$ we can choose $\delta>0$ such that $t\le \min \left(1,\frac{\delta}{\|x_1\|}\right)\Rightarrow \|f(tx_1)\| ≤ (M + \epsilon)t\|x_1\|$. To prove the claim, it will suffice to prove that $t=1$ satisfies this last inequality, so to that end,

let $\tau=\sup\{r>0:\|f(tx_1)\| ≤ (M + \epsilon)t\|x_1\|\text{for all}\ t\in [0,r]\}$, so that $\tau\le 1$ and $\tau\in \{r>0:\|f(tx_1)\| ≤ (M + \epsilon)t\|x_1\|\text{for all}\ t\in [0,r]\}$ (because $f$ is continuous.)

Now, toward a contradiction, suppose that $\tau<1$ and choose $c>0$ and small enough so that $(\tau+c)x_1$ is on the segment.

Now, there is a $\delta' > 0$ such that

$\|f(x) − f(\tau x_1)\| \le (M + \epsilon)||x − \tau x_1\|<\epsilon$ whenever $\|x − \tau x_1\| <\delta'.$

So, since $\|(\tau + c)x_1 − \tau x_1\| = c\|x_1\|<\delta' $ if $c$ is small enough, we have, with $x=(\tau + c)x_1$,

$\|f((\tau+c) x_1)-f(\tau x_1)\| \le (M+\epsilon)\|(\tau+c)x_1-\tau x_1\|=(M+\epsilon)\|cx_1\|$

and so finally,

$f((\tau+c)x_1)\le (M+\epsilon)\|cx_1\|+\|f(\tau x_1)\|\le $

$(M+\epsilon)\|cx_1\|+ (M+\epsilon)\|\tau x_1\|=(M+\epsilon)(c+\tau)\|x_1\|$,

which contradicts the fact that $\tau$ is a supremum.

  • $\begingroup$ This is Theorem 1.1.1 in Hörmander's The Analysis of Partial Differential Operators. He needs 8 lines for the proof. $\endgroup$ – Jochen Mar 29 at 11:13

Here is another proof. At the core, the arguments are similar to yours (but with fewer $\delta$s).

Assume that there are $x_1,y_1\in E$ and $\epsilon>0$ such that $$ \|f(y_1)-f(x_1) \| > (M+\epsilon) \|y_1-x_1\|. $$ By triangle inequality, one of the following inequalities are satisfied: $$ \|f(y_1)-f(\frac{x_1+y_2}2) \| > (M+\epsilon) \|y_1-\frac{x_1+y_2}2\|, $$ $$ \|f(\frac{x_1+y_2}2) - f(x_1)\| > (M+\epsilon) \|\frac{x_1+y_2}2-x_1\|. $$ Inductively, we can construct elements $(x_n,y_n)$ on the line segment $[x_1,y_1]$ with $\|x_n-y_n\| = 2^{1-n} \|x_1-y_1\|$, $[x_{n+1},y_{n+1}]\subseteq[x_n,y_n]$ and $$ \|f(y_n)-f(x_n) \| > (M+\epsilon) \|y_n-x_n\|. $$ Then $x_n,y_n\to x$ for some $x\in [x_n,y_n]$. And $$ \|f(y_n)-f(x_n) \| \le \|f(y_n)-f(x)\| + \|f(x_n)-f(x)\|. $$ By Frechet differentiability of $f$ at $x$ there we get for $n$ large enough $$\begin{split} \|f(y_n)-f(x_n) \| &\le \|f(y_n)-f(x)\| + \|f(x_n)-f(x)\| \\ &\le (M+\epsilon/2) (\|y_n-x\| + \|x-x_n\|) \\& = (M+\epsilon/2)\|y_n-x_n\|, \end{split}$$ which is a contradiction.

Let me try a direct proof in the flavor of your post. The proof works just with Gateaux differentiability of $f$ (and does not need continuity). Let $x,h\in E$ be given. Let $\epsilon>0$. Define $$ I = \{ t\in [0,1]: \|f(x+th)-f(x)\| \le (M+\epsilon) t\|h\|\}. $$ Clearly, $0\in I$. Assume $[0,t]\subset I$ for some $t\ge0$. Then by Gateaux differentiability of $f$ at $x+th$, there is $s_0>0$ such that $$ \|f(x+(s+t)h)-f(x+th)\| \le (M+\epsilon) s\cdot \|h\| $$ for $s \in (0, s_0)$. By triangle inequality, $$ \|f(x+(s+t)h)-f(x)\| \le (M+\epsilon) (s+t)\cdot \|h\| \quad \forall s\in (0,s_0). $$ Hence $[0,t+s_0)\cap [0,1]\subset I$.

Now suppose $[0,t)\subset I$ for some $t>0$. By Gateaux differentiability of $f$ at $x+th$, there is $s_0>0$ such that $$ \|f(x+(t-s)h)-f(x+th)\| \le (M+\epsilon) s\cdot \|h\| $$ for $s \in (0, s_0)$. Taking $s\in (0,s_0$ such that $t-s\in I$, and using triangle inequality as above, it follows $t\in I$ and $[0,t]\subset I$.

By induction on the reals, $I=[0,1]$: Assume $I\ne[0,1]$. Let $t = \inf [0,1]\setminus I$. Then $t\ne0$ because there is $s_0>0$ such that $[0,s_0)\subset I$. By definition of inf, $[0,t)\subset I$. Using the steps above: $[0,t]\subset I$, $[0,t+s_0)\subset I$ for some $s_0>0$.

Since $\epsilon$ in the definition of $I$ was arbitrary, the claim follows.

Using these arguments, one can prove for Gateaux differentiable $f$ the standard mean value inequality $$ \|f(x+h)-f(x) \| \le \sup_{t\in [0,1)}\|f'(x+th)\| \|h\|, $$ with supremum taken over the half-open interval.

Can this be reduced to a supremum over the open interval $(0,1)$?

  • $\begingroup$ You only need completeness of the (at most two-dimensional) space spanned by $x_1$ and $y_1$. $\endgroup$ – gerw Mar 29 at 7:58
  • $\begingroup$ @gerw so true....... $\endgroup$ – daw Mar 29 at 8:00

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.