---
title: Convergence of Ergodic Averages
url: https://www.emergentmind.com/topics/convergence-of-ergodic-averages
type: topic
---

# Convergence of Ergodic Averages

The convergence of ergodic averages, i.e., the limit behavior of Cesàro-type sums over orbits of measure-preserving dynamical systems, stands at the heart of both classical and modern ergodic theory. This concept has been developed from the Birkhoff pointwise ergodic theorem to a rich tapestry of results regarding multiple averages, polynomial iterates, weighted configurations, commuting and non-commuting transformations, non-singular systems, as well as oscillation and fluctuation phenomena. The following entry provides a synthesis of key notions, principal convergence theorems, techniques, and open questions, referencing deep results and technical frameworks from recent research.

## 1. Foundational Definitions and Classical Results

Consider a standard Lebesgue probability space $(X, \mathcal{B}, \mu)$ and a (possibly vector-valued) collection of invertible, measure-preserving transformations $T_1, ..., T_k: X \to X$.

- **Ergodic Average:** The classical single-parameter average is
  $$
  A_N(f)(x) = \frac{1}{N} \sum_{n=0}^{N-1} f(T^n x)
  $$
  for $f \in L^1(\mu)$.

- **Multiple Average:** For bounded functions $f_1, ..., f_k$, the $k$-fold product average is
  $$
  A_N(f_1, ..., f_k)(x) = \frac{1}{N} \sum_{n=0}^{N-1} f_1(T_1^n x) \cdots f_k(T_k^n x)
  $$

- **Commuting/Non-commuting:** Transformations are either assumed to commute, $T_i T_j = T_j T_i$, or not, which leads to fundamentally different analysis [1006.5239].

- **Ergodicity/Weak Mixing:** $T$ is ergodic if every $T$-invariant set is trivial (measure $0$ or $1$); weak mixing means only constants are $L^2$–eigenfunctions.

The classical **Birkhoff Ergodic Theorem** guarantees for ergodic $T$ and $f \in L^1$:
$$
A_N(f)(x) \to \int_X f\,d\mu \quad \text{almost everywhere.}
$$

## 2. Pointwise Convergence for Multiple and Polynomial Averages

The extension of convergence from single to multiple ergodic averages, and to nonconventional configurations, reveals a complex landscape:

- **Commuting Transformations:** For $T_1, ..., T_d$ invertible, commuting, and measure-preserving, Tao's $L^2$–norm convergence [1705.01420] ensures
  $$
  \|A_N(f_1, ..., f_d) - F\|_{L^2} \to 0
  $$
  for some $F \in L^2(X)$. If the system is irreducible and all $T_i T_j^{-1}$ are ergodic, there exists a subsequence $N_k$ with
  $$
  A_{N_k}(f_1, ..., f_d)(x) \to \prod_{i=1}^d \int f_i\,d\mu\quad \text{a.e.}
  $$
  [1406.2608], [1705.01420].

- **Strictly Ergodic Models:** On topological models $(\hat{X}, \hat{T}_i)$, constructing such systems where all higher-order orbit closures remain strictly ergodic allows for uniform convergence of continuous functions. Pullback by factor/isomorphism reductions achieves almost-sure convergence for the measure-theoretic model [1406.5930].

- **Distal Systems:** For distal $(X,T)$ (arising via isometric extensions), pointwise convergence for averages such as
  $$
  \frac{1}{N}\sum_{n=0}^{N-1} \prod_{j=1}^d f_j(T^{jn}x)
  $$
  is established for $\mu$–a.e.\ $x$ [1406.5930], [1609.02529].

- **Non-commuting and Cubic Averages:** For systems not assuming commutativity, pointwise convergence for both cubic configurations and polynomial shifts is achieved using Host–Kra seminorm structural decompositions and nilsequence equidistribution methods [1006.5239].

- **Polynomial/Hardy Field Iterates:** Averages of the form
  $$
  \frac{1}{N} \sum_{n=1}^N \prod_{i=1}^k f_i(T_i^{\lfloor P_i(n) \rfloor} x)
  $$
  where $P_i$ are Hardy-field functions of distinct, non-integral polynomial growth, converge almost everywhere for $f \in L^p$ ($p > 1$), with full variational bounds available when $P_i \sim n^{c_i}, c_i \notin \mathbb{Z}$ [2411.07385].

## 3. Weighted, Bilateral, and Oscillatory Averages

Recent advances study convergence under various weightings and symmetry operations.

- **Möbius and Multiplicative Weights:** For Möbius or more general Siegel–Walfisz weights, averages of the form
  $$
  \frac{1}{N}\sum_{n\leq N} \mu(n) f_1(T^{P_1(n)} x) \cdots f_k(T^{P_k(n)} x)
  $$
  converge to zero almost everywhere [2401.03174].

- **Bilateral Averages:** Symmetric averages,
  $$
  B_n(f)(x) = \frac{1}{2n+1} \sum_{i=-n}^{n} f(T^i x)
  $$
  converge a.e.\ iff both forward and backward averages converge a.e., to a common (possibly infinite) limit. Oscillation around the limit is generic and infinite; almost everywhere, there are infinitely many upcrossings and downcrossings [2003.08280], [2404.11507].

- **Fluctuation and Non-monotonicity:** The convergence of ergodic averages (and related stochastic processes) is generically non-monotone, exhibiting infinite sign changes around the limit for "typical" $f$ in $L^1$. Coboundaries with barriers can exhibit non-fluctuation, but generic functions do not [2404.11507].

## 4. Maximal Inequalities and Limitations of Non-singular Approaches

- **Classical Maximal Inequality:** In measure-preserving systems, the Hardy–Littlewood/Birkhoff maximal function satisfies weak type $(1,1)$ control,
  $$
  \|\sup_{N\ge 1} |A_N f| \|_{1,\infty} \leq C \|f\|_1
  $$
  ensuring almost-everywhere convergence by density plus maximal estimate.

- **Failure in Non-Singular Systems:** There exist non-singular systems $(Y, \nu, S)$ where
  $$
  \|M(f)\|_{1,\infty} \leq C \|f\|_1
  $$
  fails; Birkhoff's maximal inequality breaks down and one cannot deduce pointwise convergence for the full sequence from density arguments [1406.2608]. Only subsequential convergence survives under weak-mixing + quotient ergodicity.

- **Ryll–Nardzewski–Tuleca Finitely Additive Charges:** The only available invariant measure equivalents may be finitely additive, insufficient for full measure-theory convergence [1406.2608].

## 5. Rates of Convergence, Slowdown, and Variational Bounds

No universal pointwise or $L^1$ rate of convergence exists for the ergodic theorem:

- **Krengel Effect (Arbitrarily Slow Convergence):** For any sequence $\psi(N)\searrow 0$, there exists $f\in L^1$ such that
  $$
  \limsup_{N\to\infty} |A_N f(x) - \int f| / \psi(N) = +\infty
  $$
  for a.e.\ $x$ [2211.06599], [2508.00463].

- **Sharp Quantitative Bounds for Special Cases:** For averages over $\mathbb{R}^d$-actions constructed over strictly convex sets, a universal rate $O(\|\mathbf{t}\|^{-(d+1)/2})$ is governed by spectral measure concentration and Hertz's Fourier decay [2506.16740].

- **Variational and Oscillation Inequalities:** For Hardy field sequences and polynomial iterates, variational norm bounds imply almost everywhere convergence and supply quantitative control over fluctuations [2411.07385]; for weighted averages, generic function oscillations are captured by jump counting and Baire category arguments [2404.11507].

## 6. Extensions: Topological Models, $C^*$-Dynamical Systems, and Movable Windows

- **Strictly Ergodic and CF-Nil Systems:** Model-theoretic constructions (Jewett–Krieger, Weiss) yield strictly ergodic topological models matching measurable pronilfactors. On CF-Nil($k$) systems, two-dimensional ergodic averages weighted by nilsequences converge pointwise for each $x$ [2510.17267].

- **Uniform Convergence in $C^*$-Algebras:** For unital $C^*$-dynamical systems $(\mathfrak{A}, \Phi)$ that are uniquely ergodic with respect to the fixed-point subalgebra, the Cesàro averages converge pointwise in norm, while twisted ("Wiener–Wintner") averages converge when the peripheral spectrum is appropriate [2003.04160].

- **Moving Averages and Cone Conditions:** Universal almost-everywhere convergence of moving averages
  $$
  M(v_n, L_n)^T f(x) = \frac{1}{L_n} \sum_{i=v_n}^{v_n+L_n-1} f(T^i x)
  $$
  holds if and only if standard ergodic averages converge completely, or if the cone condition is satisfied. Failure of the cone condition can produce strong sweep-out [2302.03400].

## 7. Open Problems and Future Directions

Important questions remain open:
- Can one prove full-sequence pointwise convergence for general multiple ergodic averages beyond strictly distal/nilpotent cases?
- What are the necessary and sufficient ergodic-theoretic conditions for pointwise convergence in non-singular and zero-entropy contexts?
- Are there spectral or topological obstructions to convergence for strictly ergodic but non-distal systems?
- What is the sharp quantitative behavior of fluctuations and rates for advanced configurations (e.g., polynomial weights, return-time sequences, or C*-algebra automorphisms)?

Convergence of ergodic averages, in all its variants—including rates, oscillations, maximum principles, multi-parameter and weighted settings—remains a central object of rigorous analysis and ongoing research in ergodic theory and related fields.

Source: https://www.emergentmind.com/topics/convergence-of-ergodic-averages