I have been trying to learn about Ergodic Theorem for a while and now I have a problem I can't solve.

Assume $T$ is a measure preserving transform and $X_n\rightarrow X$ everywhere. Also, assume that $E(\sup_n |X_n|)<\infty$. Then I need to show $\lim_{n\rightarrow\infty}\frac{1}{n}\sum_{k=0}^{n-1}{X_k(T^k(\omega))}$ converges with probability 1.

It's easy to show $X$ is integrable, so I was trying to use ergodic theorem (Birkhoff's theorem) to prove this but I couldn't get better results.

Any suggestion is really appreciated.

  • $\begingroup$ "converges with probability 1 to some random variable with expectation equal to E(X)" is a bit odd. $\endgroup$ – mookid Mar 6 '14 at 23:49
  • $\begingroup$ Sorry, do you mean it's wrong? or the way I have expressed it is wired? $\endgroup$ – Cohlan Mar 6 '14 at 23:51
  • $\begingroup$ Isn't it "converges with probability 1 to E(X)" instead? $\endgroup$ – mookid Mar 6 '14 at 23:54
  • $\begingroup$ I think once we can prove that thing converges with probability 1, the rest would be pretty easy. So maybe I should edit the question. $\endgroup$ – Cohlan Mar 6 '14 at 23:54
  • $\begingroup$ put the "\" before Sup and lim in this case ;) $\endgroup$ – mookid Mar 6 '14 at 23:55

I realize what I did in the other answer was maybe overcomplicated. Define $D_l:=\sup_{k\geqslant l}|X_k-X|$ (which is integrable for each $l$), and notice that by Birkhoff's ergodic theorem, $$\frac 1n\sum_{j=0}^{n-1}D_l\circ T^j\to\mathbb E[D_l\mid\mathcal I]\quad \mbox{a.e.}$$ with the same notations as in the other answer. Since $$\frac 1n\sum_{j=0}^{n-1}|X_k-X|\circ T^j=\frac 1n\sum_{j=0}^{l-1}|X_j-X|\circ T^j+\frac 1n\sum_{j=l}^{n-1}\sup_{k\geqslant l}|X_k-X|\circ T^j\\ \leqslant \frac 1n\sum_{j=0}^{l-1}|X_j-X|\circ T^j+\frac 1n\sum_{j=0}^{n-1}\sup_{k\geqslant l}|X_k-X|\circ T^j+ \frac 1n\sum_{j=0}^{l-1}\sup_{k\geqslant l}|X_k-X|\circ T^j,$$ we obtain that for almost every $\omega$ and any $l$, $$\limsup_{n\to +\infty}\frac 1n\sum_{j=0}^{n-1}|X_k-X|\circ T^j(\omega)\leqslant \mathbb E[D_l\mid\mathcal I](\omega).$$ By monotone convergence, $\lim_{l\to +\infty}\mathbb E[D_l\mid\mathcal I](\omega)=0$ almost everywhere and we are done.

| cite | improve this answer | |
  • $\begingroup$ Thanks for this one as well. $\endgroup$ – Cohlan Mar 8 '14 at 20:52
  • $\begingroup$ You are welcome. $\endgroup$ – Davide Giraudo Mar 8 '14 at 20:55

This is a kind of "uniform ergodic theorem" and extends naturally the case $X_k=X$ for each $k$.

Notice that $X_k=X_k-X+X$ and by Birkhoff's ergodic theorem, $$\frac 1n\sum_{k=0}^{n-1}X\circ T^k\to \mathbb E[X\mid\mathcal I]\quad\mbox{ a.s.},$$ where $\mathcal I$ denotes the $\sigma$-algebra of invariant sets, that is, $\mathcal I=\{A, T^{-1}(A)=A\}$.

If we manage to show that $$\frac 1n\sum_{k=0}^{n-1}(X_k-X)\circ T^k\to 0\quad\mbox{ a.s.},$$ we will be done. Replacing the $X_k$'s by $|X_k-X|=:Y_k$ we are reduced to show that $$\frac 1n\sum_{k=0}^{n-1}Y_k\circ T^k\to 0\quad\mbox{ a.s.}$$ under the assumption $Y_k\to 0$ a.s. and $\sup_kY_k$ is integrable.

Consider a sequence $(\delta_l)_{l\geqslant 0}$ of positive real numbers such that for each $l$, $\mathbb E[\chi_A\sup_nY_n]\lt 2^{-l}$ if $\mu(A)\lt \delta_l$. We then use Egoroff's theorem: for each $l$, there is a measurable set $A_l$ such that $$\lim_{k\to+\infty}\sup_{\omega\in A_l}Y_k=0\quad\mbox{ and }\mu(\Omega\setminus A_l)\lt\delta_l.$$ Define $M_n:=\frac 1n\sum_{k=0}^{n-1}Y_k\circ T^k=M'_{n,l}+M''_{n,l}$ with $M'_{n,l}:=\frac 1n\sum_{k=0}^{n-1}(\chi_{A_l}Y_k)\circ T^k$ and $M''_{n,l}:=\frac 1n\sum_{k=0}^{n-1}(\chi_{\Omega\setminus A_l}Y_k)\circ T^k$.

Since for each $k$, $\chi_{A_l}Y_k\leqslant \sup_{\omega\in A_l}Y_k(\omega)$, we have $$\tag{1} 0\leqslant M'_{n,l}\leqslant \frac 1n\sum_{k=0}^{n-1}\sup_{\omega\in A_l}Y_k(\omega).$$

Notice that $$0\leqslant \mathbb E[\sup_nM''_{n,l}]\leqslant \mathbb E[\sup_n\max_{0\leqslant k\leqslant n-1}\chi_{\Omega\setminus A_l}Y_k]\leqslant \delta_l\leqslant 2^{-l},$$ hence by the Borel-Cantelli lemma, $$\tag{2}\lim_{l\to\infty}\sup_{n\geqslant 1}M''_{n,l}=0\mbox{ a.s.}$$ Combining (1) and (2), it follows $$0\leqslant M_n\leqslant \frac 1n\sum_{k=0}^{n-1}\sup_{\omega\in A_l}Y_k(\omega)+ \sup_mM''_{m,l},$$ hence taking the $\limsup_{n\to+\infty}$, we obtain for each $l$, $$0\leqslant \limsup_{n\to \infty}M_n\leqslant \sup_mM''_{m,l}.$$ Now let $l$ going to infinity to conclude.

| cite | improve this answer | |
  • $\begingroup$ That's great, Thank you very much. $\endgroup$ – Cohlan Mar 8 '14 at 20:51

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.