Let $f$ be Lebesgue integrable on $[0,1]$, and define $$\large P_nf=n\sum_{k=1}^n\int_{(k-1)/n}^{k/n}f\,d\mu\cdot\chi_{[\frac{k-1}{n},\frac{k}{n}]}.$$

How do we show that $$\lim_{n\to\infty}\int_0^1 |f-P_nf|\,d\mu=0$$?

Thank you very much.

My attempts:

I would think in the direction of either Dominated Convergence Theorem, Monotone Convergence Theorem, or Fatou's Lemma (these are the only 3 tricks I know).

Firstly, I would like to prove $P_nf\to f$ a.e. I can see it intuitively, but not sure how to prove it. I can see that $P_nf$ is trying to partition $[0,1]$ into $n$ subintervals, then for any $x$, it is in one of the intervals $[\frac{k-1}{n},\frac{k}{n}]$, so that $P_nf(x)=n\int_{(k-1)/n}^{k/n}f(x)\,d\mu\to n\cdot\frac{1}{n}f(x)=f(x)$. (How to show this rigorously?)

I can't seem to use DCT due to the "$n$" in $P_nf$ which grows to infinity, so it is hard to find an explicit dominating function. Similarly, MCT seems hard since it is hard to show that $|P_nf|$ is increasing ($n$ is increasing but the "narrowing limits integral" may be decreasing) . That leaves Fatou's Lemma, and I am stuck.

Thanks for any help once again!

  • 1
    Prove it first for $f \in C([0,1])$ (it essentially follows from uniform continuity) and then estimate an $L^1$ function by a sequence of continuous functions. It is actually more difficult to prove convergence a.e. I think. $L^1$ convergence is much more natural here. – User8128 Oct 18 '16 at 4:29
  • Thanks. I haven't learnt how to estimate an $L^1$ function by continuous functions. Can it be done using Lusin theorem? – yoyostein Oct 18 '16 at 5:52
  • 1
    You can use Lebesgue's differentiation theorem to make your argument for pointwise a.e. convergence rigorous. – Dominique R.F. Oct 18 '16 at 6:09
  • 1
    I'm not sure Lusin's theorem will help here [though, of course, it might]. I'm simply talking about that fact that $C([0,1])$ is dense in $L^1([0,1])$. Thus for any $f \in L^1([0,1])$, we can find a sequence $f_n \in C^1([0,1])$ so that $\| f - f_n \|_{L^1([0,1])} \to 0.$ – User8128 Oct 19 '16 at 13:48
up vote 2 down vote accepted

A straight forward computation yields \begin{align} \int^1_0|f-P_n f|\ d\mu =&\ \int^1_0\left|\sum^n_{k=1} f(x)\chi_{[\frac{k-1}{n},\frac{k}{n}]}-n\sum^n_{k=1} \int^{k/n}_{(k-1)/n} f\ d\mu\cdot \chi_{[\frac{k-1}{n},\frac{k}{n}]}\right|\ d\mu\\ \leq&\ \sum^n_{k=1}\int^{k/n}_{(k-1)/n}\left|f(x)-n\int^{k/n}_{(k-1)/n} f\ d\mu \right|\ d\mu\\ =&\ n\sum^n_{k=1}\int^{k/n}_{(k-1)/n}\left|\int^{k/n}_{(k-1)/n} f(x)-f(y)\ d\mu(y) \right|\ d\mu(x)\\ \leq&\ n\sum^n_{k=1} \int^{k/n}_{(k-1)/n}\int^{k/n}_{(k-1)/n}|f(x)-f(y)|\ d\mu(x)d\mu(y). \end{align} If $f$ is continuous on $[0, 1]$, then it's uniformly continuous on $[0, 1]$, which mean for each $\epsilon>0$ there exists $\delta>0$ such that $|f(x)-f(y)|<\epsilon$ whenever $|x-y|<\delta$.

In our case since $x, y \in [(k-1)/n, k/n]$, then for big enough $N$ we have that $|x-y|<\delta$. Hence it follows \begin{align} n\sum^n_{k=1} \int^{k/n}_{(k-1)/n}\int^{k/n}_{(k-1)/n}|f(x)-f(y)|\ d\mu(x)d\mu(y) < \epsilon n\sum^n_{k=1} \int^{k/n}_{(k-1)/n}\int^{k/n}_{(k-1)/n} d\mu(y)d\mu(x) = \epsilon. \end{align}

Thus, if we could approximate $f$ by a continuous function then we are essentially done.

Fix $\epsilon>0$. By Lusin's Theorem, there exists $g_\epsilon$ continuous such that \begin{align} \int^1_0 |f-g_\epsilon|<\epsilon \end{align} and likewise we have \begin{align} \int^1_0 |P_n f-P_n g_\epsilon|<\epsilon. \end{align} Important Note: the $g_\epsilon$ that you get from Lusin's theorem is actually equal to $f$ expect on a set of small measure.

  • Thanks a lot. When we are using Lusin Theorem, we also need this fact right: If $f$ is integrable, then for any measurable $|A|<\delta$, then $\int_A |f|<\epsilon$? How would we show that $P_nf$ is integrable? Essentially I can understand except the last line $\int_0^1 |P_nf-P_ng_\epsilon|<\epsilon$. Thanks a lot. – yoyostein Oct 25 '16 at 3:37
  • Just one more question: Why do we need $\int_0^1 |P_nf-P_ng_\epsilon|<\epsilon$? I don't see it used explicitly anywhere. – yoyostein Oct 25 '16 at 3:41
  • 1
    Think of $P_n f$ as $\sum a_k \chi_{k}$ and see what you when you integrate. – Jacky Chong Oct 25 '16 at 3:41
  • 1
    You need $\int^1_0 |P_n f-P_n g_\epsilon|<\epsilon$ because $\int |f-P_n f| \leq \int |f-g|+|g-P_ng|+|P_ng-P_n f|$. – Jacky Chong Oct 25 '16 at 3:42
  • Ok, very nice. Thanks and have a great week ahead! – yoyostein Oct 25 '16 at 3:49

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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