Rate of convergence in the Von Neumann ergodic theorem.

257 Views Asked by At

First let me state the Von Neumann ergodic theorem.

Suppose $U$ is a unitary operator on a Hilbert space $H$ and let $P$ be the orthogonal projection onto $\ker (U- I)$. Then we have for all $f \in H$ $$ \lim_{N \to \infty} \frac{1}{N} \sum_{n=0}^{N-1} U^n f = Pf.$$

This theorem I could prove with relative ease. But I found the following related exercise in my notes. Suppose we write $U = e^{iH}$ for some self-adjoint operator $H$ and we assume $\text{dist}( \sigma(H), 2\pi \mathbb{Z}) = \epsilon >0,$ where $\sigma(H)$ is the spectrum of $H$, what can we then say about the rate of convergence of the limit in the Von Neumann ergodic theorem?

A little note on $\text{dist}( \sigma(H), 2\pi \mathbb{Z})$: It isn't explained in the notes but I assume this is some kind of Hausdorff distance. You simply only consider $2 \pi \mathbb{Z} \cap [ - \lVert H \rVert, +\lVert H\rVert$], making the set compact and thus the Hausdorff distance well-defined. But I could be wrong, please tell me if you know the correct (or better) definition. I also suspect that due to the spectral mapping theorem, we would find that $\text{dist}(\sigma(U), 1) = \delta$ for some $\delta > 0$ but I'm not entirely sure.

Any help is much appreciated.

UPDATE:

I think I may have found a simple counterexample, in the sense that there's an operator $U_1$ without a spectral gap and $U_2$ with a spectral gap but in both cases the limits converge at the same rate. Consider $H = \mathbb{C}^2$ the following two unitary operators $$U_1 = \left( \begin{matrix} 1 & 0 \\ 0 & -1 \end{matrix} \right), U_2 = e^{i\alpha} I,$$ where $\alpha$ is some irrational number, ensuring that $e^{ni\alpha}$ never becomes 1. $U_1$ has an eigenvalue 1, so $\text{dist}(\sigma(U_1),1) = 0$ but $U_2$ only has eigenvalue $e^{i\alpha}$, so $\text{dist}(\sigma(U_1),1) >0$. Since the space is finite dimensional, we might as well look at convergence in operator norm. Note that $\ker (U_1- I) =\text{span} \{ e_1 \}$ where $e_1,e_2$ are the standard basis. Let $P$ be the projection onto this subspace. Then we see for odd $N$ $$ \lVert \frac{1}{N} \sum_{n=0}^{N-1} U_1^n - P \rVert = \lVert \left( \begin{matrix} 0 & 0 \\ 0 & \frac{1}{N}\sum_{n=0}^{N-1}(-1)^n \end{matrix} \right)\rVert = \frac{1}{N} $$ but $0$ for even $N$. Similarly for odd $N$ we have $$ \lVert \frac{1}{N} \sum_{n=0}^{N-1} U_2^n \rVert= \lVert \left( \begin{matrix} \frac{e^{iN\alpha} - 1}{N(e^{i\alpha} -1)} & 0 \\ 0 & \frac{e^{iN\alpha} - 1}{N(e^{i\alpha} -1)} \end{matrix} \right) \rVert = \frac{e^{iN\alpha} - 1}{N(e^{i\alpha} -1)} = O\left(\frac{1}{N}\right)$$ but $0$ for even $N$. So they both converge just as fast, despite one having a spectral gap and the other not. Please let me know if there's anything wrong with this argument.

1

There are 1 best solutions below

6
On

You are right about the distance, all is as you say.

On the rest, I suggest that you write $U$ in the form $$ U=\int_{S^1}\lambda\,d\mu(\lambda) $$ and do the computations using this form: for $f$ in the kernel of $U-I$ the average is $$ \frac{1}{N} \sum_{n=0}^{N-1} U^n f=f+\frac{1}{N} \sum_{n=1}^{N-1} \int_{S^1}\lambda^n\,d\mu(\lambda). $$ In view of the existence of $\delta$, the second term converges to zero with a speed related to the spectral gap with respect to $1$. That gives the speed that you want.

Notice that without the information about $\delta$ the second term still converges to zero (use dominated convergence), but you don't get anything about the speed.