MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Let $E$ be a closed subspace of $L^2[0,1]$. Suppose that $E\subset{}L^\infty[0,1]$. Is it true that $E$ is finite dimensional?

PS. This is actually a question from the real analysis qualifier. I came across it as I was teaching qualifier preparation course, and was solving problems from old qualifiers. So, though it might follow from some advanced theory of Banach spaces, I am most interested in the 'elementary' solution, using only methods from standard real analysis course. Note: if $E\subset{}C[0,1]$, then it is a problem from Folland, and there is a solution there. However, it does not work for $L^\infty$, not without some trick.

share|cite|improve this question
Sorry, I left a reference (to Rudin's book) as an answer without properly reading your question, and in particular the fact that you wanted a proof avoiding functional-analytic tricks – Yemon Choi Jan 19 '11 at 19:12
Since it seems that using the open mapping theorem is allowed, perhaps the reference to Rudin might still be of interest? It's Theorem 5.2 in Functional Analysis (2nd ed) – Yemon Choi Jan 20 '11 at 1:46
up vote 27 down vote accepted

Another solution: as Mikael wrote, $||f||_{\infty} \leq C ||f||_2$ for every $f \in E$. Let $f_1,\ldots,f_n$ be an orthonormal family in your subspace. Then for every $x \in [0,1]$, $f_1(x)^2+\ldots+f_n(x)^2 \leq ||f_1(x)f_1+\ldots+f_n(x)f_n||_{\infty} \leq C \|f_1(x)f_1+\ldots+f_n(x)f_n\|_2$ $$=C \sqrt{f_1(x)^2+\ldots+f_n(x)^2},$$

and by squaring we get $f_1(x)^2+\ldots+f_n(x)^2 \leq C^2$, and integrating gives $n \leq C^2$.

share|cite|improve this answer
I like this!!!! – Nate Eldredge Jan 19 '11 at 17:06
Many thanks! This is actually the trick to extend the solution from Folland from $C[0,1]$ to $L^\infty$. – Rostyslav Kravchenko Jan 19 '11 at 18:44
This is the "right" elementary solution and is what the makers of the qualifying exam had in mind. At a deeper conceptual level, the problem is obvious (the 2-summing norm of the identity operator on any $n$-dimensional space is $\sqrt {n}$). – Bill Johnson Jan 19 '11 at 22:53

First note that by the closed graph theorem, there is a $C$ such that $\|f\|_\infty \leq C\|f\|_2$ for any $f$ iin $E$. (this can also be checked directly if you consider that the closed graph theorem is advanced Banach space theory).

Assume $E$ is infinite dimensional, and take $(f_n)_{n\ge 1}$ an orthonormal basis of $E$. Let $A_n$ be the subset of $[0,1]$ on which the real part of $f_n$ is greater than $1/10$. Replacing if necessary $f_n$ by $-f_n$ or $i f_n$ or $-if_n$, we can assume that $\|Re(f_n)^+\|_2 \ge 1/2$, so that the measure of $A_n$ is greater than some constant $\delta$ depending on $C$ only ($\delta = (1/2^2-1/10^2)/C^2$ works for example).

Since the integral of $1_{A_1}+1_{A_2}+...+1_{A_n}$ is greater than $n\delta$, there exist $i_1,...i_k$ with $k$ being the integer part of $n\delta$, such that $A_{i_1},...,A_{i_k}$ have non-trivial intersection. Then the sum $1/\sqrt k (f_{A_{i_1}}+...f_{A_{i_k}})$ has $L^2$ norm $1$, but $L^\infty$ norm greater than $\sqrt k/10$ (because its real part is greater than $\sqrt k /10$ on a non-trivial subset). A contradiction.

share|cite|improve this answer
Could you give a hint on how you can check $\|f\|_{\infty} \leq C \|f\|_{2}$ without appealing to the open mapping theorem (or closed graph theorem)? – Theo Buehler Jan 19 '11 at 16:45
In fact I was not very honest: the proof I had in mind was more or less a proof of the closed graph theorem (see Anyway, here it goes. A first attempt would be to take a sequence $g_n$ in the unit ball of $E$ such that $\|g_n\|_\infty > 4^n$, consider $f=\sum_n \omega_n 2^{-n} g_n$ with a good choice of $\omega_n\in \mathbb C$ with $|\omega_n|=1$, and hope that $f$ is not bounded. I do not believe that this works in general, but if the $g_n$ are chosen carefully this will work... – Mikael de la Salle Jan 19 '11 at 22:05
... In fact $\|g_n\|_\infty>4^n$ means that there is $A_n \subset [0,1]$ of positive measure such that the (modulus of the) average $E[g|A_n]$ of $g_n$ on $A_n$ is greater than $4^n$. Changing $g_n$ if necessary, we can assume that $g_n$ is "the worst" such function, ie that $E[g_n|A]= c_n > 4^n$, and that $|E[g|A]| \leq (c_n+1) \|g\|_2$ for any $g \in E$. So define $f_0=g_0$ and $f_{n+1} = f_n+ 2^{-n} e^{-i\theta_n} g_{n+1}$, where $\theta_{n}$ is an argument of $E[f_n|A_{n+1}]$, so that $|E[f_{n+1}|A_{n+1}]|\geq c_n/2^n$... – Mikael de la Salle Jan 19 '11 at 22:15
Hum, I noticed that I already got it wrong: I should have replaced the $2^{-n}$ by $3^{-n}$ in the definition of $f_{n+1}$, and I would conclude that $|E[f_n|A_n]|\geq c_n 3^{-n+1}$. Then $f=\lim f_n$ belongs to $E$, and $\|f-f_n\|_2 \leq 3^{-n+1}/2$, so that $|E[f|A_n]| \ge |E[f_n|A_n]|-(c_n+1)\|f_f_n\|_2 \ge 3^{-n+1}(c_n -1)/2$. This goes to $\infty$, and proves that $f$ is not in $L^\infty$. – Mikael de la Salle Jan 19 '11 at 22:38
Thank you very much, I think I see what you're saying and I'll have to have a closer look at Sokal's paper which looks really neat! While trying to answer the original question I came up with something similar as you're sketching now and I was afraid that a substantially simpler argument escaped me - I interpreted using the open mapping theorem as "cheating" from the PS in the question... – Theo Buehler Jan 19 '11 at 23:11

Here's one solution. There may be cleaner ones.

Let $E$ be as supposed. The natural inclusion $T : L^\infty([0,1]) \hookrightarrow L^2([0,1])$ is bounded, so $E = T^{-1}(E)$ is therefore also closed in $L^\infty$. By the open mapping theorem, it follows that $T^{-1}$ is bounded on $E$, so there exists $C$ such that for all $f \in E$, $||f||_\infty \le C ||f||\_2$. Now if $||f||\_2 = 1$, we have $||f||\_\infty \le C$, and so by noting $$1 = \int |f|^2 \le C^2 m(|f| > \epsilon) + \epsilon^2$$ and taking, say, $\epsilon = 1/2$, we have $m(|f| > 1/2) \ge 1/4C^2$.

Now suppose $E$ is infinite dimensional; then it contains an $L^2$-orthonormal sequence $\{f_n\}$. By replacing $f_n$ by $-f_n$ as necessary we may assume that for each $f_n$, $m(f_n > 1/2) \ge 1/8C^2$. By a pigeonhole argument there is a set $A$ of positive measure where $f_{n_k} > 1/2$ for infinitely many $n_k$ (edit: actually I am not sure about this step). Now $f = \sum_k k^{-1} f_{n_k}$ converges in $L^2$ and so $f \in E$; since the $L^2$ and $L^\infty$ norms are equivalent on $E$, the sum also converges uniformly to $f$ on a set of full measure. But this implies that $f = +\infty$ a.e. on $A$, which is absurd.

share|cite|improve this answer
Oops, Mikael beat me. – Nate Eldredge Jan 19 '11 at 15:44
It is funny that we had exactly the same construction. I just do not follow your use of the pigeohole principle. – Mikael de la Salle Jan 19 '11 at 16:29
Hm, in retrospect, neither do I. – Nate Eldredge Jan 19 '11 at 17:07
@Nate, I think that you are taking $A = \limsup \{f_n > 1/2\}$ and observing that $m(A) \ge \limsup m(f_n > 1/2) \ge 1/8C^2$. – L Spice Jan 19 '11 at 22:35
@L Spice: That's what I was thinking, but unfortunately, $A = \limsup \\{ f_n > 1/2\\}$ doesn't quite do the trick. It gives us that for each $x \in A$, there is some sequence $f_{n_k}$ with $f_{n_k}(x) > 1/2$ for all $n_k$; unfortunately, the particular sequence that works may vary with $x$, and it is not clear to me that there is a single sequence that works for all $x \in A$ (or even a positive-measure subset). – Nate Eldredge Jan 19 '11 at 23:06

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.