6
$\begingroup$

In a uniformly convex Banach space $$x_n\stackrel{w}\to x \ \ \text{and} \ \ ||x_n||\to ||x|| \ \ \text{implies} \ ||x_n-x||\to 0.$$ Can you help me to solve it?

Thanks in advance.

$\endgroup$
  • $\begingroup$ The case $x = 0$ is immediate. So suppose $x\neq 0$. Consider a linear functional $\lambda$ with $\lVert\lambda\rVert = 1$ and $\lambda(x) = \lVert x\rVert$. $\endgroup$ – Daniel Fischer Jun 22 '14 at 22:51
  • $\begingroup$ @Daniel Fischer: do you mean $\Vert \lambda(x) \Vert = \Vert x \Vert$? $\lambda(x) = \Vert x \Vert$ is not linear, or do I miss something? $\endgroup$ – Robert Lewis Jun 22 '14 at 22:57
  • 1
    $\begingroup$ @RobertLewis $x$ is a specific value, the weak limit. It might have been less confusing if that had been called $x_\ast$ for example, then I would have written $\lambda(x_\ast) = \lVert x_\ast\rVert$. It's just the value at one point that I prescribe. $\endgroup$ – Daniel Fischer Jun 22 '14 at 23:00
  • 1
    $\begingroup$ It may help if you assume $\lVert x\rVert = 1$, since that may make things closer to the definition of a uniformly convex space. What can you say about $x+x_n$, and how does that, together with the uniform convexity imply $\lVert x_n-x\rVert\to 0$? $\endgroup$ – Daniel Fischer Jun 22 '14 at 23:25
  • 1
    $\begingroup$ There is no reason why we should have $\lambda(x_n-x)= \lVert x_n-x\rVert$. $\endgroup$ – Daniel Fischer Jun 22 '14 at 23:26
2
$\begingroup$

The case $x = 0$ is immediate, since $\lVert x_n\rVert\to 0$ is just the norm convergence of $(x_n)$ to $0$.

So let's suppose $x\neq 0$, and, without loss of generality, $\lVert x\rVert = 1$. By the Hahn-Banach theorem, there is a continuous linear functional $\lambda$ with $\lVert\lambda\rVert = 1$ and $\lambda(x) = 1$.

Since $x_n \xrightarrow{w} x$, we have $\lambda(x_n) \to \lambda(x) = 1$, and therefore

$$\lim_{n\to\infty} \lambda\left(\tfrac{1}{2}(x_n+x)\right) = 1,$$

which implies

$$\liminf_{n\to\infty} \left\lVert \tfrac{1}{2}(x_n+x)\right\rVert \geqslant 1.$$

On the other hand, since $\left\lVert\frac{1}{2}(x_n+x)\right\rVert \leqslant \frac{1}{2}(\lVert x_n\rVert + \lVert x\rVert)$, we have

$$\limsup_{n\to\infty} \left\lVert \tfrac{1}{2}(x_n+x)\right\rVert \leqslant 1,$$

so

$$\lim_{n\to\infty} \left\lVert\tfrac{1}{2}(x_n+x)\right\rVert = 1.$$

But a space is uniformly convex if and only if for all sequences $(a_n)$ and $(b_n)$ with $\lim\limits_{n\to\infty} \lVert a_n\rVert = 1 = \lim\limits_{n\to\infty} \lVert b_n\rVert$ and $\lim\limits_{n\to\infty} \left\lVert \frac{1}{2}(a_n+b_n)\right\rVert = 1$ it follows that $\lim\limits_{n\to\infty} \lVert a_n - b_n\rVert = 0$.

Pick $a_n = x_n$ and $b_n = x$.

This characterisation of uniform convexity may be part of the definition, or be a lemma to be proved. In that case, we only need one direction here:

Suppose $\lVert a_n - b_n\rVert \not\to 0$. Then there is an $\varepsilon > 0$ such that $\lVert a_{n_k} - b_{n_k}\rVert > 2\varepsilon$ for subsequences $(a_{n_k})$ and $(b_{n_k})$. We may assume that this holds for the full sequences. For that $\varepsilon$, by uniform convexity, there is a $\delta > 0$ such that

$$\left\lVert \tfrac{1}{2}(u+v)\right\rVert \leqslant 1 - \delta$$

for all $u,v$ with $\lVert u\rVert = 1 = \lVert v\rVert$ and $\lVert u-v\rVert \geqslant \varepsilon$.

Then, setting $\alpha_n = \frac{1}{\lVert a_n\rVert}a_n$ and $\beta_n = \frac{1}{\lVert b_n\rVert}b_n$ ($a_n = 0$ or $b_n = 0$ can happen only finitely often, we drop these terms), we have

$$\begin{align} \left\lVert\alpha_n-\beta_n\right\rVert &\geqslant \lVert a_n-b_n\rVert - \lVert a_n-\alpha_n\rVert - \lVert b_n-\beta_n\rVert\\ &= \lVert a_n-b_n\rVert - \lvert 1 - \lVert a_n\rVert\rvert - \lvert 1-\lVert b_n\rVert\rvert\\ &> 2\varepsilon - \lvert 1 - \lVert a_n\rVert\rvert - \lvert 1-\lVert b_n\rVert\rvert\\ &> \varepsilon \end{align}$$

for $n$ so large that $\lvert 1-\lVert a_n\rVert\rvert < \varepsilon/2$ and $\lvert 1-\lVert b_n\rVert\rvert < \varepsilon/2$. But then

$$\begin{align} \left\lVert \tfrac{1}{2}(a_n+b_n)\right\rVert &\leqslant \left\lVert \tfrac{1}{2}(\alpha_n+\beta_n)\right\rVert + \tfrac{1}{2}(\lVert \alpha_n - a_n\rVert + \lVert \beta_n-b_n\rVert)\\ &\leqslant 1-\delta + \tfrac{1}{2}(\lVert \alpha_n - a_n\rVert + \lVert \beta_n-b_n\rVert)\\ &\leqslant 1 - \frac{\delta}{2} \end{align}$$

for $n$ so large that also $\lvert 1-\lVert a_n\rVert\rvert < \delta/2$ and $\lvert 1-\lVert b_n\rVert\rvert < \delta/2$, contradicting the assumption $\left\lVert \frac{1}{2}(a_n+b_n)\right\rVert \to 1$.


Summary: In a uniformly convex space, the curvature of spheres is "sufficiently large" in all directions that the intersection of a spherical shell $a-\varepsilon < \lVert y\rVert < a+\varepsilon$ with a "thickened supporting hyperplane" $a-\varepsilon < \operatorname{Re} \lambda(y) < a+\varepsilon$ becomes arbitrarily small as $\varepsilon \to 0$.

$\endgroup$

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.