This is actually an exercise in Rudin's Real and Complex Analysis, $L^p$ spaces chapter. Could anyone help me out? Thanks in advance.

Motivation: It's well known that if we have a function $f$ which belongs to $L^p(0,1)$ for all $p\ge 1$. Then $\lim_{p\rightarrow \infty}\|f\|_p=\|f\|_{\infty}$ (moreover, $\|f\|_p$ is increasing in $p$). This is true even if $\|f\|_{\infty}=\infty$.

Question: How slow (fast) can $\|f\|_p$ grow when $\|f\|_{\infty}=\infty$? More precisely, given any positive increasing function $\Phi$ with $\lim_{p\rightarrow \infty}\Phi(p)=\infty$, can we always find a function $f$ which belongs to $L^p(0,1)$ for all $p\ge 1$, and $\|f\|_{\infty}=\infty$, such that $\|f\|_p\le (\ge)\Phi(p)$ for large $p$?

  • 1
    $\begingroup$ It's a very interesting question. I don't have the answer, but if we consider $f(x)=|\ln x|^{\alpha}$ where $\alpha>0$, we cover the case $\phi(p)=p^{\alpha}$. $\endgroup$ – Davide Giraudo Oct 26 '11 at 17:06
  • $\begingroup$ Nice observation. It took me a while to verify it. Is there a good way to see this? $\endgroup$ – Syang Chen Oct 27 '11 at 0:41
  • $\begingroup$ @Davide Giruado According to WolframAlpha, $\|f\|_p = \Gamma(p\alpha+1)^{1/p}$ for $f(x)=|\ln x|^\alpha$. How do you get $\phi(p)=p^\alpha$ from this? $\endgroup$ – Jeff Oct 27 '11 at 4:30
  • $\begingroup$ @Jeff, use Stirling's formula for Gamma function you will see this is true with "$=$" replaced by "$\approx$". $\endgroup$ – Syang Chen Oct 27 '11 at 6:39
  • $\begingroup$ You should add the condition that $\Phi$ is non-decreasing (or continuous), otherwise it's easy to construct $\Phi$ for which the condition trivially fails. $\endgroup$ – Yuval Filmus Oct 27 '11 at 14:11

I'll answer the question of whether $\|f\|_p$ can grow arbitrarily slowly; the answer is yes. I'm fairly certain that it can grow arbitrarily quickly as well, and, though I haven't given it much thought, I suspect a similar argument can be concocted. The problem doesn't seem to rely on the interval $(0,1)$, so what I write below doesn't either. If you really want to put everything in $(0,1)$, it is not hard to do so.

Let $\Phi$ be as you say. Here's the idea: We choose disjoint sets $E_n$ with positive (but as yet undetermined) measure, and we require that $f=\sum_{n=1}^{\infty} c_n\chi_{E_n}$, where $\{c_n\}$ is some sequence of positive numbers increasing to infinity, and $\chi_{E_n}$ is the characteristic function of $E_n$. This ensures that $f\notin L^{\infty}$. Assuming (as we may) that $\Phi(p)$ is bounded away from $0$, we see that the quotient $c_n^p/\Phi(p)^p$ is bounded in $p$ for each fixed $n$. So we are free to choose our sets $E_n$ so small in measure that $(c_n/\Phi(p))^pm(E_n)<2^{-n}$, independently of $p$. To conclude, we simply observe that $$ \begin{align*} \frac{\|f\|_p^p}{\Phi(p)^p} &= \frac{1}{\Phi(p)^p}\int\sum_{n=1}^{\infty}c_n^p\chi_{E_n}\\ &= \sum_{n=1}^{\infty}\frac{c_n^p}{\Phi(p)^p}m(E_n)\\ &< \sum_{n=1}^{\infty}2^{-n}=1. \end{align*} $$ This means $\|f\|_p<\Phi(p)$ for all $p$ (or, if you like, for all $p\in[a,\infty)$, where $\Phi(p)$ is bounded away from $0$ on $[a,\infty)$).

  • $\begingroup$ how can you choose $c_n$ such that $c_n^p/\Phi(p)^p$ is bounded in p? What are the $c_n$'s? $\endgroup$ – Syang Chen Oct 27 '11 at 14:07
  • 1
    $\begingroup$ @XianghongChen The sequence $\{c_n\}$ is an arbitrary sequence of positive numbers that increases to $\infty$ as $n\to\infty$. The fact that $c_n^p/\Phi(p)^p$ is bounded in $p$ for fixed $n$ has nothing to do with the choice of the sequence $\{c_n\}$ other than the fact that $c_n>0$. For any $r>0$, the ratio $r^p/\Phi(p)^p$ is bounded in $p$, simply because $\Phi(p)\to\infty$ with $p$. This is because, for sufficiently large $p$, you have $r/\Phi(p)<1$. For smaller $p$, the function $\Phi$ is bounded away from zero, so $r/\Phi(p)\leq r/\delta$ for some $\delta>0$. $\endgroup$ – Nick Strehlke Oct 27 '11 at 15:51
  • $\begingroup$ Here is how I understand it. Assume that $\Phi$ is increasing. Let $c_n=\Phi(n), m(E_n)=(2\Phi(n))^{−n}$. Is it correct? $\endgroup$ – Syang Chen Oct 28 '11 at 1:26
  • $\begingroup$ @Xianghong Those choices will certainly work. You can see why (as you probably did) by looking at $$ \left(\frac{c_n}{\Phi(p)}\right)^pm(E_n) = 2^{-n}\left(\frac{\Phi(n)^{p-n}}{\Phi(p)^{p}}\right). $$ Whenever $\Phi(p)>1$, the above quantity is $<2^{-n}$. That's because $\Phi(n)^{p-n}<1$ if $p<n$ and $\Phi(p)^p>\Phi(n)^{p-n}$ if $p>n$, because $\Phi$ is increasing. Of course, other choices will work too. $\endgroup$ – Nick Strehlke Oct 28 '11 at 4:04
  • $\begingroup$ The algorithm I had in mind was: (i) Fix your sequence $\{c_n\}$; (ii) maximize $g_n(p)=(cn/\Phi(p))^p$ in $p$ for each $n$, say $g_n(p)<M_n$ for all $p$; (iii) take $m(En)<1/(2^nM_n)$. By fixing a clever choice of $c_n$, you've essentially managed to do (ii) and (iii) in general. So you give up a little bit of choice and gain some efficiency. The reason I decided not to use a specific $\{c_n\}$ in writing my answer was that I wanted to make transparent the properties I was using, namely, that $c_n\to\infty$ with $n$ and $c_n>0$. (Had I thought of your choices I might have changed my mind. :)) $\endgroup$ – Nick Strehlke Oct 28 '11 at 4:26

Here's a similar solution for the case where we want the norm to grow arbitrarily fast. As mentioned in my comment, we must assume that $\Phi$ obtains a finite maximum on finite intervals (since the $p-$th norm is an increasing function in $p$).

Let $$C_n = \max(2^n, \max_{p \in [n,n+1]} \Phi(p)).$$ We construct a function $f$ such that $C_n \leq \|f\|_n < \infty$ for all natural $n$. The construction is of the general form $$ f = \sum_{n \geq 1} a_i \chi_{B_i}, \quad \mu(B_i) \triangleq \mu_i, \quad \sum_{i \geq 1} \mu_i < \infty. $$ We define $$ a_n = C_n^{n+1}, \mu_n = C_n^{-n^2}. $$ Since $C_n \geq 2^n$, the series $\sum_{i \geq 1} \mu_i$ clearly converges.

On the one hand, for every $n$ we have $$ \| f \|_n^n \geq a_n^n \mu_n = C_n^n. $$ On the other hand, for every $n$ we have $$ \| f \|_n^n = \sum_{i = 1}^n a_n^n \mu_n + \sum_{i \geq n+1} a_i^n \mu_i. $$ In order to show that the norm exists, it is enough to show that the latter series converges: $$ \sum_{i \geq n+1} a_i^n \mu_i = \sum_{i \geq n+1} C_i^{n(i+1) - i^2} \leq \sum_{i \geq n+1} C_i^{-1} < \infty. $$ We used $C_i \geq 1$ and $n(i+1) - i^2 \leq -1$, which follows from $n(n+2) - (n+1)^2 = -1$ and the fact that $n(i+1) - i^2$ is decreasing for $i \geq n/2$ (calculus).

  • $\begingroup$ Neat argument, Thanks! $\endgroup$ – Syang Chen Oct 28 '11 at 1:53

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.