Differential equations and dynamical systems

Ergodic theory

Studies the long-term average behavior of dynamical systems, connecting them to measure-preserving transformations.

IntuitionIf you spin forever by the same odd angle, do you eventually visit everywhere?

Mark a point on a circular dial and rotate it, again and again, by a fixed angle that is an irrational fraction of a full turn — say, the golden angle of about $137.5^\circ$ , as sunflower seeds and pinecone scales grow. Because the angle is irrational, the point never returns exactly to where it started, and remarkably it eventually comes arbitrarily close to every point on the circle, visiting each small arc with a frequency proportional to the arc's length. No randomness is involved anywhere — the rule is a single rigid rotation, repeated forever — yet the long-run statistics look exactly as if the point were chosen uniformly at random. This equidistribution phenomenon, and the question of which average properties a deterministic rule produces over infinite time, is the subject of ergodic theory.

Drag the theta slider repeatedly through the golden angle

137.5^\circ

: the marked point never repeats exactly, yet after enough turns it has swept out an evenly-spread, dense set of points on the circle — the geometric picture behind equidistribution and ergodicity of the irrational rotation

T(x)=x+\alpha \pmod 1

UndergraduateMeasure-preserving transformations

Definition: Measure-preserving transformation

Let $(X,\mathcal F,\mu)$ be a probability space (a set $X$ , a $\sigma$ -algebra $\mathcal F$ of measurable subsets, and a measure $\mu$ with $\mu(X)=1$ ). A map $T:X\to X$ is measure-preserving if $\mu(T^{-1}A) = \mu(A)$ for every $A\in\mathcal F$ — the measure of the set of points that will land in $A$ equals the measure of $A$ itself. Intuitively: applying $T$ never creates or destroys probability mass, only rearranges where it sits.

\mu(T^{-1}A) = \mu(A) \quad \forall\, A \in \mathcal{F}

Three examples anchor the theory. The irrational rotation $T(x)=x+\alpha \pmod 1$ on the circle $[0,1)$ with $\alpha$ irrational preserves ordinary (Lebesgue) length, since rotating an arc does not change its length. The doubling map $T(x)=2x\bmod 1$ also preserves Lebesgue measure — it is exactly 2-to-1, and each of the two preimage branches of an interval is compressed by a factor of $2$ , so their total length exactly reconstructs the original. And a Bernoulli shift (independent coin flips, shifted one step at a time) preserves the natural product probability measure on the space of infinite coin-flip sequences, since shifting a sequence of independent flips one step still leaves independent, identically distributed flips.

T^{-1}A=A \;\Longrightarrow\; \mu(A)\in\{0,1\}

A measure-preserving $T$ is ergodic if every $T$ -invariant set is essentially trivial: $T^{-1}A=A \;\Longrightarrow\; \mu(A)\in\{0,1\}$ for every measurable $A$ . Equivalently, the system cannot be split into two pieces of positive measure that $T$ never mixes together — there is no nontrivial way to decompose the dynamics. Both the irrational rotation and the doubling map are ergodic with respect to Lebesgue measure, but for entirely different reasons, as the table below makes precise.

Ergodicity and mixing for three canonical examples
Transformation	Ergodic?	Mixing?
Irrational rotation $T(x)=x+\alpha \pmod 1$	Yes — every orbit is equidistributed	No — rigid rotation never mixes two arcs together
Doubling map $T(x)=2x\bmod 1$	Yes	Yes — correlations between distant times decay to zero
Identity map $T(x)=x$	No (unless $X$ is a single point) — every set is invariant	No

UndergraduateBirkhoff's ergodic theorem and Poincaré recurrence

Theorem: Birkhoff's pointwise ergodic theorem

Let $T$ be a measure-preserving transformation of a probability space $(X,\mathcal F,\mu)$ and $f\in L^1(\mu)$ . Then the time averages $\frac{S_nf(x)}{n}$ converge for $\mu$ -almost every $x$ to a $T$ -invariant limit $\bar f(x)$ with $\int_X \bar f\,d\mu = \int_X f\,d\mu$ . If moreover $T$ is ergodic, then $\bar f(x)=\int_X f\,d\mu\ \text{a.e.}$ — the time average along a single orbit equals the space average.

Why is it true?

This is the rigorous version of an intuition physicists had used for decades without proof (Boltzmann's ergodic hypothesis in statistical mechanics): to compute the long-run average of some quantity, you can either watch one system evolve for a very long time, or average over a whole ensemble of systems at one instant — and for ergodic systems these two very different-sounding computations give exactly the same answer.

Proof

Step 1 (invariant limsup and liminf). Define $f^*(x)=\limsup_{n\to\infty}\frac{S_nf(x)}{n}$ and $f_*(x)=\liminf_{n\to\infty}\frac{S_nf(x)}{n}$ . Since $S_{n+1}f(x) = f(x) + S_nf(Tx)$ , dividing by $n+1$ and letting $n\to\infty$ shows $f^*(Tx)=f^*(x)$ and $f_*(Tx)=f_*(x)$ : both are $T$ -invariant functions.

Step 2 (the maximal ergodic theorem). For $\lambda\in\mathbb R$ , let $A_\lambda=\{x: \sup_n S_nf(x)/n>\lambda\}$ be the set where the running time average ever exceeds $\lambda$ . The key technical lemma (proved by considering $g=f-\lambda$ and the maximum of the partial sums $M_n(x)=\max(0,S_1g(x),\dots,S_ng(x))$ , then using $M_n(Tx)\ge S_kg(Tx)$ termwise and integrating over the set where $M_n>0$ ) gives $\int_{A_\lambda} f\,d\mu \ge \lambda\,\mu(A_\lambda)$ .

Step 3 (squeezing f and f_ together). Suppose for contradiction that $f^*>f_*$ on a set of positive measure; then there exist rationals $\alpha<\beta$ with $E=\{x: f_*(x)<\alpha<\beta<f^*(x)\}$ of positive measure. $E$ is $T$ -invariant (since $f^*,f_*$ are), so we may restrict attention to $E$ . Applying the maximal inequality to $f-\beta$ on $E$ forces $\int_E f\,d\mu\ge\beta\mu(E)$ , and applying it to $\alpha-f$ similarly forces $\int_E f\,d\mu\le\alpha\mu(E)$ ; since $\alpha<\beta$ and $\mu(E)>0$ these contradict each other. Hence $f^*=f_*$ almost everywhere, so the limit $\bar f(x)=\lim_n S_nf(x)/n$ exists a.e. and is $T$ -invariant.

Step 4 (matching the integrals, and the ergodic case). A dominated-convergence argument (truncating $f$ and controlling the tails using the maximal inequality again) shows $\int_X \bar f\,d\mu = \int_X f\,d\mu$ . Finally, if $T$ is ergodic, the $T$ -invariant function $\bar f$ must be constant almost everywhere (by the very definition of ergodicity applied to its level sets $\{\bar f\le c\}$ , each of which is $T$ -invariant and hence has measure $0$ or $1$ ); combined with the equality of integrals, that constant must be $\int_X f\,d\mu$ , giving $\bar f(x)=\int_X f\,d\mu\ \text{a.e.}$ .

Theorem: Poincaré's recurrence theorem

Let $T$ be a measure-preserving transformation of a probability (or more generally finite-measure) space $(X,\mathcal F,\mu)$ and let $A\in\mathcal F$ with $\mu(A)>0$ . Then almost every point of $A$ returns to $A$ infinitely often: for almost every $x\in A$ , $T^nx\in A$ for infinitely many $n\ge1$ .

Why is it true?

If space is finite and nothing is ever destroyed (measure-preservation), a region cannot keep sending points off to entirely new, never-before-visited territory forever — eventually the system has to start revisiting where it has already been, simply because there is nowhere new left to put the returning measure.

Proof

Step 1 (points that never return). Let $A_0=\{x\in A: T^nx\notin A\ \forall n\ge1\}$ be the points of $A$ that never come back to $A$ . The sets $A_0, T^{-1}A_0, T^{-2}A_0,\dots$ are pairwise disjoint: if $x\in T^{-i}A_0\cap T^{-j}A_0$ with $i<j$ , then $T^ix\in A_0$ but also $T^j x = T^{j-i}(T^ix) \in A$ with $j-i\ge1$ , contradicting that $T^ix\in A_0$ never returns to $A$ . So $T^{-i}A_0 \cap T^{-j}A_0=\varnothing\ (i\ne j)$ .

Step 2 (the never-return set has measure zero). Since $T$ is measure-preserving, $\mu(A_0)=\mu(T^{-i}A_0)$ for every $i$ . If $\mu(A_0)>0$ , the countably many pairwise disjoint sets $T^{-i}A_0$ ( $i=0,1,2,\dots$ ) would all have this same positive measure, so their union would have infinite total measure — impossible since $\mu(X)<\infty$ (indeed $\mu(X)=1$ ). Hence $\mu(A_0)=0$ .

Step 3 (finitely-often visitors also have measure zero). Let $B=\{x\in A: x\ \text{returns to}\ A\ \text{only finitely often}\}$ . Writing $B$ as a countable union over $k$ of (essentially) the never-return set of $T^kA$ under the shifted dynamics, $B=\bigcup_{k\ge0} T^{-k}\{x\in T^kA: x\ \text{never returns to}\ T^kA\}$ , each term has measure zero by exactly the Step 1–2 argument applied to $T^kA$ in place of $A$ (using $\mu(T^kA)=\mu(A)$ ). By countable subadditivity, $\mu(B)=0$ .

Step 4 (conclusion). Every $x\in A\setminus B$ (which has full measure in $A$ , since $\mu(B)=0$ ) returns to $A$ infinitely often by definition of $B$ . This is exactly the statement of the theorem.

AdvancedMixing: a stronger form of ergodicity

Definition: Strong mixing

A measure-preserving $T$ is (strongly) mixing if $\lim_{n\to\infty}\mu(T^{-n}A\cap B)=\mu(A)\mu(B)$ for all measurable $A,B$ : the fraction of $B$ that lands back in $A$ after $n$ steps converges to what it would be if $A$ and $B$ were statistically independent. Mixing implies ergodicity (take $B=A$ with $T^{-1}A=A$ : then $\mu(A)=\mu(A\cap A)\to\mu(A)^2$ , forcing $\mu(A)\in\{0,1\}$ ), but not conversely.

Why is the irrational rotation ergodic but not mixing? Take $A=B$ a small arc: rotating it by $n\alpha$ just moves the same small arc rigidly around the circle, so $T^{-n}A\cap A$ is either empty or (for infinitely many $n$ , since the rotation is equidistributed) very close to all of $A$ again — the overlap fraction never settles down to the independent value $\mu(A)^2$ , it keeps oscillating back up near $\mu(A)$ . The doubling map, by contrast, stretches and folds every small interval across the whole space exponentially fast, genuinely scrambling it — the mechanism responsible for its mixing, and ultimately for treating chaotic maps like $r=4$ logistic dynamics as "as good as random" for statistical purposes.

UndergraduateReal-World Applications and Worked Examples

Ergodic theory turns "average behavior over infinite time" into a computable, provable statement, which is exactly what statistical mechanics needs to justify replacing a single physical system's long-run average with an ensemble average (Boltzmann's ergodic hypothesis), what data compression needs to define the entropy rate that bounds how much a source can be compressed (via the Kolmogorov–Sinai entropy of the associated shift), what search engines like Google's PageRank need to guarantee a unique long-run visiting frequency for a random web surfer (the invariant measure of a Markov chain), and what cryptographic pseudo-random generators built from chaotic maps need to justify treating their output as statistically random. In every case, ergodic and mixing properties are what license treating a single deterministic trajectory as if it carried statistical information about the whole system.

Example: How often does an irrational rotation visit a given arc?

Let $T(x)=x+\alpha\pmod 1$ with $\alpha$ irrational act on the circle $[0,1)$ with Lebesgue measure, and let $A=[0,0.3)$ . For a generic starting point $x_0$ , what fraction of the first $n$ iterates $x_0,x_1,\dots,x_{n-1}$ land in $A$ , as $n\to\infty$ ?

Solution

This is exactly a Birkhoff ergodic theorem computation with $f=\mathbf 1_A$ , the indicator function of $A$ : the time average $\frac1n\sum_{k=0}^{n-1}\mathbf 1_A(x_k)$ is precisely the fraction of the first $n$ iterates landing in $A$ .

Irrational rotations are a classical example of an ergodic transformation with respect to Lebesgue measure (this is Weyl's equidistribution theorem in disguise: any $T$ -invariant set, when expanded in a Fourier series, forces all nonzero Fourier coefficients to vanish because rotation multiplies them by $e^{2\pi i k\alpha}\ne1$ , leaving only a constant function).

Since $T$ is ergodic, Birkhoff's theorem applies in its strongest form: the time average equals the space average for every generic starting point, not just on average over starting points. The space average is $\int_X \mathbf 1_A\,d\mu = \mu(A) = 0.3-0 = 0.3$ .

So the long-run fraction of time the orbit spends in $A$ is exactly $0.3$ , regardless of which $x_0$ you start from (outside a measure-zero exceptional set) — a rigorous version of "the deterministic rotation behaves, statistically, just like picking a uniformly random point in $A$ each time."

Example: How much can a biased data source be compressed? The Kolmogorov–Sinai entropy of a Bernoulli shift

A data source emits independent bits, each $1$ with probability $p=0.3$ and $0$ with probability $0.7$ — modeled ergodically as the Bernoulli shift on sequences with the product measure $(0.3,0.7)$ . Compute its entropy rate (the Kolmogorov–Sinai entropy of the shift, equal to the Shannon entropy of a single symbol here), which by Shannon's source coding theorem is the minimum number of bits per symbol needed, on average, to losslessly compress this source.

Solution

For an i.i.d. (Bernoulli) source, the Kolmogorov–Sinai entropy of the shift map reduces to the ordinary Shannon entropy of a single symbol: $h=-\sum_i p_i\log_2 p_i$ .

Substituting $p_1=0.3$ , $p_0=0.7$ : $h=-0.3\log_2 0.3-0.7\log_2 0.7$ .

Compute each term: $-0.3\log_2 0.3 = 0.3\times1.737=0.521$ bits, and $-0.7\log_2 0.7=0.7\times0.515=0.361$ bits (using $\log_2 0.3\approx-1.737$ and $\log_2 0.7\approx-0.515$ ).

Adding these, $h\approx0.881\ \text{bits}$ . So on average, no lossless code can compress this source below about $0.881$ bits per symbol (noticeably less than the $1$ bit/symbol needed for a fair coin, precisely because the bias makes the source more predictable, hence more compressible) — and Shannon's theorem guarantees this bound is achievable.

Common mistake. Ergodicity and mixing are often conflated, but they are genuinely different: ergodicity only says a system cannot be split into two invariant pieces (equivalently, time averages equal the space average), while mixing additionally says correlations between events decay to zero over time. Every mixing system is ergodic, but the irrational rotation above is a standard counterexample showing ergodic does not imply mixing. A second common error is over-reading the Poincaré recurrence theorem: it guarantees that almost every point of

A

returns to

A

infinitely often, but says nothing about when — the expected return time can be astronomically long (by Kac's lemma, the average return time to

A

1/\mu(A)

, which is enormous for a tiny set

A

), and the theorem gives no information about returning to the exact same point, only to the set

A

Historical note

The subject's motivating question came from 19th-century statistical mechanics: Ludwig Boltzmann's ergodic hypothesis (1870s) proposed that a gas's trajectory through phase space eventually passes arbitrarily close to every state consistent with its energy, justifying replacing time averages with the much easier-to-compute ensemble averages. Henri Poincaré proved his recurrence theorem in 1890, as a technical step within his prize memoir on the three-body problem, showing that bounded mechanical systems must eventually return arbitrarily close to earlier configurations — a result initially seen as almost paradoxical (how can entropy increase if the system returns?), resolved only once the astronomically long recurrence times were properly appreciated. Rigorous ergodic theory as a mathematical field began in 1931-1932, when George Birkhoff and John von Neumann, working independently within months of each other, each proved a version of the ergodic theorem — von Neumann's $L^2$ mean version appeared first but was published slightly later, while Birkhoff's stronger pointwise (almost-everywhere) version appeared in the Proceedings of the National Academy of Sciences in December 1931, finally placing Boltzmann's hypothesis on rigorous mathematical footing over half a century after it was proposed.

Henri Poincaré John von Neumann

Research frontier as of 2026

Ergodic theory has become an essential tool far outside its origins in statistical mechanics. Peter Sarnak's Möbius disjointness conjecture (proposed around 2010) predicts that the Möbius function from number theory is asymptotically uncorrelated with the orbit of any deterministic system of zero topological entropy — a precise ergodic-theoretic formulation of "the primes look random to any sufficiently non-chaotic clock," verified for many specific classes of systems but open in general as of 2026. Hillel Furstenberg's ergodic-theoretic reproof of Szemerédi's theorem on arithmetic progressions (1977) launched the field of ergodic Ramsey theory, whose multiple recurrence theorems continue to be extended to ever more general group actions and configurations. On the more applied side, rigorously computing or bounding entropy and mixing rates for realistic high-dimensional systems — from molecular dynamics simulations to training dynamics of neural networks viewed as noisy dynamical systems — remains largely open, and understanding how quickly quantum many-body systems "thermalize" (approach ergodic-like behavior, formalized via the Eigenstate Thermalization Hypothesis) is an active area connecting ergodic theory to condensed-matter physics and quantum information.

A transformation $T:X\to X$ preserves a probability measure $\mu$ , meaning $\mu(T^{-1}A) = \mu(A)$ . For which sets $A$ must this equation hold?

By Birkhoff's ergodic theorem, for the ergodic irrational rotation $T(x)=x+\alpha \pmod 1$ and $A=[0,0.3)$ , what is the long-run fraction of time a generic orbit spends in $A$ ?

According to the Poincaré recurrence theorem, for a measure-preserving system on a finite-measure space with $\mu(A)>0$ , almost every point of $A$ ...

A source emits i.i.d. bits with $P(1)=0.3$ . Its entropy rate (the Kolmogorov–Sinai entropy of the associated Bernoulli shift), rounded to two decimals in bits, is closest to:

References

Peter Walters (1982). An Introduction to Ergodic Theory
George D. Birkhoff (1931). Proof of the Ergodic Theorem
John von Neumann (1932). Proof of the Quasi-Ergodic Hypothesis
Hillel Furstenberg (1981). Recurrence in Ergodic Theory and Combinatorial Number Theory