38.04.01 · dynamics / ergodic-theorems

Measure-Preserving Systems, Poincaré Recurrence, and the Kac Formula

shipped3 tiersLean: none

Anchor (Master): Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 1-2; Cornfeld-Fomin-Sinai 1982 *Ergodic Theory* (Springer Grundlehren 245) Ch. 1, Ch. 10 (induced transformations, Kakutani-Rokhlin towers); Petersen 1983 *Ergodic Theory* (Cambridge) §2.4 (Kac's lemma); Aaronson 1997 *An Introduction to Infinite Ergodic Theory* (AMS) Ch. 1 (recurrence vs. conservativity)

Intuition Beginner

Imagine a closed box of gas, every molecule rushing about under fixed mechanical laws. Stir it once and let it run forever. A natural question is whether the gas could ever return, even approximately, to the exact arrangement it started in. Intuition says no — surely such an ordered state, once destroyed, is gone for good. The first great theorem of this subject says the opposite: for almost every starting arrangement, the system comes back arbitrarily close to where it began, and it does so not once but infinitely many times. This is Poincaré recurrence, and it is the surprising bedrock on which the whole theory of measure-preserving dynamics rests.

The hidden ingredient is that the dynamics preserves volume. As the system evolves, regions of states get stretched and folded, but never expanded or compressed in total size. A region of arrangements occupying one percent of all possibilities still occupies one percent after a million steps; only its shape has changed. When a fixed rule shuffles a finite pile of volume around forever without ever creating or destroying any, the pieces are forced to overlap their old positions again and again. There is simply nowhere for the volume to escape to.

Picture a single point hopping around a region under a fixed rule that preserves area. Mark a small patch and watch the point. If it started in the patch, it cannot wander away forever: if every visit landed somewhere new, the disjoint copies of the patch would eventually need more room than the whole region holds. So the point must come home. Once you know it returns at least once, the same argument applied again forces a second return, a third, and so on without end.

The takeaway: when a fixed rule reshuffles a fixed total volume of states, almost every state is recurrent — it returns close to itself infinitely often. The Kac formula then makes this quantitative, telling you that the average time between returns to a region of size $p$ is exactly $1/ p$ : small regions are visited rarely, in exact proportion to how small they are.

Visual Beginner

Picture a point moving around a square region by a fixed area-preserving rule, starting inside a small shaded patch $A$ . Each application of the rule moves it to a new spot; sometimes it lands back inside $A$ .

The wandering trajectory keeps coming back to the patch $A$ . The three side-by-side copies illustrate why: if returns never happened, the preimages of $A$ would stack up disjointly and overrun the whole square. The little table of return times shows the Kac formula in miniature — the average gap between visits to $A$ equals $1$ divided by the size of $A$ .

Worked example Beginner

We watch recurrence and return times for the simplest area-preserving rule on the circle: rotation by a fixed amount.

Step 1. The system. Take the circle as the numbers from $0$ up to $1$ , wrapping around so $1$ is the same point as $0$ . The rule is: add $a = 0.25$ and wrap. This rule preserves length on the circle, so it is measure-preserving. Starting from $0.05$ , the orbit is $0.05, 0.30, 0.55, 0.80, 0.05$ — after four steps it lands exactly back on $0.05$ .

Step 2. A region to watch. Let $A$ be the arc from $0$ to $0.25$ , which has length $0.25$ , that is one quarter of the circle. Our starting point $0.05$ is inside $A$ .

Step 3. First return. The orbit $0.05, 0.30, 0.55, 0.80, 0.05$ leaves $A$ immediately (the next point $0.30$ is outside the arc $[0, 0.25)$ ) and first comes back to $A$ at step $4$ , when it returns to $0.05$ . So for this point the first-return time is $4$ .

Step 4. The average return time. As the starting point ranges over the arc $A$ , some points return after $4$ steps and some after $3$ steps, depending on exactly where in $A$ they start. Averaging the return time over all starting points in $A$ , the Kac formula predicts the answer $1$ divided by the size of $A$ , namely $1/0.25 = 4$ .

What this tells us: the point keeps coming back to $A$ — that is Poincaré recurrence. And the average number of steps between visits is $4$ , exactly $1$ over the size $0.25$ of the region. A region taking up one quarter of the circle is visited, on average, once every four steps. That balance — rare visits to small regions, in precise proportion — is the content of the Kac formula.

Check your understanding Beginner

Exercise (easy, multiple choice).

Poincaré recurrence guarantees that, for a volume-preserving rule on a finite-volume space and a region $A$ of positive size:

A. Every single point of $A$ returns to $A$ after exactly one step B. Almost every point of $A$ returns to $A$ , in fact infinitely many times C. No point of $A$ ever leaves $A$ D. The point visits every other point of the space before returning

Hint

Recurrence is an "almost every" statement about returning, and once one return is forced, the argument repeats to force infinitely many.

Answer

B. Almost every point of $A$ returns to $A$ , infinitely many times.

Feedback-correct: recurrence is an almost-everywhere statement — a measure-zero set of points may fail to return — and the one-return argument iterates to give infinitely many returns. Feedback-wrong: A demands an exact one-step return, far too strong; C says points never leave, which is false in general; D describes a visiting-everywhere property that recurrence does not assert.

Formal definition Intermediate+

Throughout, $(X, B, μ)$ is a probability space — a measurable space 02.07.01 with $μ (X) = 1$ — and $T : X \to X$ is a measurable map.

Definition (measure-preserving system). The map $T$ is measure-preserving if $μ (T^{- 1} B) = μ (B)$ for every $B \in B$ , where $T^{- 1} B = {x \in X : T x \in B}$ . The quadruple $(X, B, μ, T)$ is then a measure-preserving system (m.p.s.). If in addition $T$ is a bijection with measurable inverse and $T^{- 1}$ is also measure-preserving, the system is invertible. The condition is stated on preimages, not images: for a non-invertible $T$ the forward image measure $μ (T B)$ may differ from $μ (B)$ .

Definition (Koopman operator). The Koopman operator $U_{T}$ acts on measurable functions by $(U_{T} f) (x) = f (T x)$ , i.e. $U_{T} f = f \circ T$ . Measure-preservation is equivalent to $U_{T}$ being an isometry of $L^{2} (μ)$ : for $f \in L^{2} (μ)$ , $∥ U_{T} f ∥_{2}^{2} = \int_{X} ∣ f \circ T ∣^{2} d μ = \int_{X} ∣ f ∣^{2} d μ = ∥ f ∥_{2}^{2},$ the middle equality being the change-of-variables identity $\int g \circ T d μ = \int g d μ$ that holds for measure-preserving $T$ . When $T$ is invertible, $U_{T}$ is unitary with $U_{T}^{- 1} = U_{T^{- 1}}$ ; the spectral theory of $U_{T}$ is the operator-theoretic shadow of the dynamics, in the sense made precise at 37.02.03.

Definition (ergodicity). A set $B \in B$ is invariant if $T^{- 1} B = B$ , and almost invariant if $μ (T^{- 1} B △ B) = 0$ . The system is ergodic if every almost-invariant set has $μ (B) \in {0, 1}$ , equivalently if every measurable $g$ with $g \circ T = g$ a.e. is a.e. constant.

Canonical examples. (i) Irrational rotation $T x = x + a mod 1$ on $([0, 1), Leb)$ with $a$ irrational: measure-preserving by translation invariance, ergodic. (ii) Doubling map $T x = 2 x mod 1$ on $([0, 1), Leb)$ : measure-preserving (each point has two preimages, each carrying half the measure) and non-invertible. (iii) Bernoulli shift on $X = {0, \dots, k - 1}^{N}$ with a product measure $μ = p^{\otimes N}$ and $T$ the left shift: the abstract model of an i.i.d. sequence 37.02.03. (iv) Gauss map $T x = 1/ x mod 1$ on $(0, 1)$ , preserving the Gauss measure $d μ = \frac{1}{l n 2} \frac{d x}{1 + x}$ ; its orbits encode continued-fraction expansions. (v) Toral automorphism $T = A mod 1$ on the torus $T^{2}$ for $A \in SL_{2} (Z)$ : preserves Lebesgue measure since $∣ det A ∣ = 1$ , and is ergodic precisely when $A$ has no eigenvalue that is a root of unity.

Definition (first-return time and induced transformation). Let $A \in B$ with $μ (A) > 0$ . The first-return time to $A$ is $r_{A} (x) = in f {n \geq 1 : T^{n} x \in A} \in {1, 2, \dots} \cup {\infty}, x \in A .$ By Poincaré recurrence (below) $r_{A} < \infty$ a.e. on $A$ . The induced (first-return) transformation $T_{A} : A \to A$ is $T_{A} (x) = T^{r_{A} (x)} (x)$ , defined a.e. on $A$ . We give $A$ the normalised measure $μ_{A} (\cdot) = μ (\cdot \cap A) / μ (A)$ .

Counterexamples to common slips Intermediate+

Measure-preservation is a statement about preimages. The doubling map $T x = 2 x mod 1$ satisfies $μ (T^{- 1} B) = μ (B)$ yet is two-to-one; the forward image of $[0, 1/2)$ is all of $[0, 1)$ , so $μ (T B) \neq = μ (B)$ in general. Checking measure-preservation on images is the standard error.
Recurrence needs finite total measure. Translation $T x = x + 1$ on $R$ with Lebesgue measure is measure-preserving but every point escapes to infinity and nothing recurs. The probability-space (or at least finite-measure) hypothesis is essential; infinite ergodic theory studies exactly the conservative/dissipative dichotomy that replaces it.
Recurrence does not assert ergodicity. A system can be wildly non-ergodic and still recurrent: the identity map $T = id$ returns every point in one step yet mixes nothing. Poincaré recurrence is a consequence of measure-preservation alone; ergodicity is an independent, stronger indecomposability condition.
The Kac formula needs ergodicity for the clean form. The identity $\int_{A} r_{A} d μ = 1$ (whence mean return time $1/ μ (A)$ ) requires the system to be ergodic so that the tower over $A$ exhausts $X$ . Without ergodicity, $\int_{A} r_{A} d μ = μ (⋃_{n \geq 0} T^{- n} A)$ equals the measure of the saturation of $A$ , which can be strictly less than $1$ .
The induced map's return time is at least $1$ , never $0$ . The infimum defining $r_{A}$ runs over $n \geq 1$ ; a point already in $A$ does not count as "returning" at time $0$ . Forgetting this off-by-one collapses the tower partition and breaks the Kac count.

Key theorem with proof Intermediate+

Theorem (Poincaré recurrence; Poincaré 1890). Let $(X, B, μ, T)$ be a measure-preserving system on a probability space and let $A \in B$ with $μ (A) > 0$ . Then for $μ$ -almost every $x \in A$ , the orbit returns to $A$ infinitely often: the set ${n \geq 1 : T^{n} x \in A}$ is infinite for a.e. $x \in A$ .

Proof. First show almost every point of $A$ returns at least once. Let $N = {x \in A : T^{n} x \in / A for all n \geq 1}$ be the set of points of $A$ that never return. We claim the sets $N, T^{- 1} N, T^{- 2} N, \dots$ are pairwise disjoint. Suppose $T^{- m} N \cap T^{- n} N \neq = \emptyset$ for some $0 \leq m < n$ , and pick $x$ in the intersection. Then $y = T^{m} x \in N \subseteq A$ and $T^{n - m} y = T^{n} x \in N \subseteq A$ with $n - m \geq 1$ . But $y \in N$ means $T^{k} y \in / A$ for every $k \geq 1$ , contradicting $T^{n - m} y \in A$ . Hence the preimages are pairwise disjoint.

By measure-preservation $μ (T^{- n} N) = μ (N)$ for every $n$ , so $n = 0 \sum \infty μ (T^{- n} N) = n = 0 \sum \infty μ (N) \leq μ (X) = 1.$ A sum of infinitely many copies of $μ (N)$ is finite only if $μ (N) = 0$ . Thus almost every point of $A$ returns to $A$ at least once.

To upgrade to infinitely many returns, apply the same result to the induced first-return map. Almost every $x \in A$ has $T^{r_{A} (x)} x \in A$ with $r_{A} (x) < \infty$ ; the point $T_{A} x = T^{r_{A} (x)} x$ again returns a.e., and iterating, $T_{A}^{k} x \in A$ for every $k$ , producing the strictly increasing sequence of return times $r_{A} (x) < r_{A} (x) + r_{A} (T_{A} x) < \dots$ . Hence ${n : T^{n} x \in A}$ is infinite for a.e. $x \in A$ . $□$

Bridge. Poincaré recurrence builds toward the entire quantitative theory of return times and appears again in the Kac formula below, where the bare "returns at least once" is sharpened into "returns after $1/ μ (A)$ steps on average". The foundational reason recurrence holds is the pigeonhole of finite measure against infinitely many equal-measure disjoint preimages — this is exactly the wandering-set obstruction, and a set on which it fails is by definition wandering. Putting these together with 37.02.03, recurrence is the qualitative skeleton whose flesh is Birkhoff's theorem: the time-average $\frac{1}{n} \sum_{k < n} 1_{A} \circ T^{k}$ converges to $μ (A)$ in the ergodic case, which is the central insight that visits to $A$ have asymptotic frequency $μ (A)$ , and the bridge is that a positive asymptotic frequency forces infinitely many visits, recovering recurrence as a corollary while adding the rate. The induced transformation $T_{A}$ introduced here is dual to the suspension/tower construction and generalises to the Kakutani-Rokhlin tower that organises the next theorem.

Exercises Intermediate+

Exercise 4 (medium, symbolic).

Prove that the induced transformation $T_{A}$ preserves the restricted measure $μ ∣_{A}$ , i.e. $μ (T_{A}^{- 1} E) = μ (E)$ for measurable $E \subseteq A$ , given that $r_{A} < \infty$ a.e. on $A$ .

Hint

Partition $A$ by the value of $r_{A}$ : write $A_{n} = {x \in A : r_{A} (x) = n}$ , so $T_{A} = T^{n}$ on $A_{n}$ . Sum the contributions and telescope using measure-preservation of $T$ .

Answer

Let $A_{n} = {x \in A : r_{A} (x) = n}$ , a measurable partition of $A$ (up to the null set where $r_{A} = \infty$ ). On $A_{n}$ , $T_{A} = T^{n}$ , so for $E \subseteq A$ , $T_{A}^{- 1} E = n \geq 1 ⨆ (A_{n} \cap T^{- n} E) .$ Set $C_{n} = T^{- n} E \cap {x : T x, \dots, T^{n - 1} x \in / A}$ , the points entering $E \subseteq A$ exactly at time $n$ for the first time. Then $T_{A}^{- 1} E = ⨆_{n} (C_{n} \cap A)$ and, because $T$ preserves $μ$ , the levels of the Kakutani skyscraper over $A$ have equal measure across each column. Concretely, defining the tower levels $B_{n, j} = T^{j} A_{n}$ for $0 \leq j < n$ , these are disjoint, $μ (B_{n, j}) = μ (A_{n})$ by measure-preservation, and $T$ maps level $j$ bijectively (mod $μ$ ) to level $j + 1$ . The return map $T_{A}$ sends the top of each column back into $A$ , and counting preimages of $E$ through the columns gives $μ (T_{A}^{- 1} E) = \sum_{n} μ (A_{n} \cap T^{- n} E)$ . The same column bijections give $\sum_{n} μ (A_{n} \cap T^{- n} E) = μ (E)$ , since each unit of measure of $E$ is hit once as the unique first entry. Hence $T_{A}$ preserves $μ ∣_{A}$ (equivalently, the normalised $μ_{A}$ ).

Exercise 5 (medium, symbolic).

Using the Kakutani skyscraper over $A$ (levels $T^{j} A_{n}$ for $0 \leq j < n$ , where $A_{n} = {r_{A} = n}$ ), prove Kac's formula $\int_{A} r_{A} d μ = μ (⋃_{n \geq 0} T^{- n} A)$ , and deduce $\int_{A} r_{A} d μ = 1$ when the system is ergodic.

Hint

The levels of the skyscraper are disjoint and partition the saturation of $A$ . Counting levels in the column of height $n$ gives exactly $n$ , which is $r_{A}$ on $A_{n}$ .

Answer

For each $n \geq 1$ the column over $A_{n}$ has levels $A_{n}, T A_{n}, \dots, T^{n - 1} A_{n}$ , which are pairwise disjoint (a point in two levels of the same column would return to $A$ before time $n$ ) and disjoint across columns. Their union is the Kakutani skyscraper $S = ⨆_{n \geq 1} ⨆_{j = 0}^{n - 1} T^{j} A_{n}$ , and one checks $S = ⋃_{m \geq 0} T^{- m} A$ up to measure zero: every point that ever visits $A$ in forward time sits at a definite height below the next entry into $A$ . Measure-preservation gives $μ (T^{j} A_{n}) = μ (A_{n})$ , so $μ (S) = n \geq 1 \sum j = 0 \sum n - 1 μ (T^{j} A_{n}) = n \geq 1 \sum n μ (A_{n}) = \int_{A} r_{A} d μ,$ the last equality because $r_{A} = n$ on $A_{n}$ . This is Kac's formula in the general form $\int_{A} r_{A} d μ = μ (S)$ . When $T$ is ergodic, $S = ⋃_{m} T^{- m} A$ is (mod $μ$ ) invariant and has positive measure, hence $μ (S) = 1$ , giving $\int_{A} r_{A} d μ = 1$ and mean return time $\int_{A} r_{A} d μ_{A} = 1/ μ (A)$ .

Exercise 7 (hard, symbolic).

Prove the Kakutani-Rokhlin tower lemma in the simplest case: if $(X, B, μ, T)$ is ergodic and aperiodic (no periodic orbit has positive measure), then for every $ε > 0$ and integer $n \geq 1$ there is a measurable base $B$ with $B, T B, \dots, T^{n - 1} B$ pairwise disjoint and $μ (⋃_{j = 0}^{n - 1} T^{j} B) > 1 - ε$ .

Hint

Take a small set $A$ of positive measure with $μ (A)$ small, form the induced first-return tower over $A$ , and use the columns of height $\geq n$ to extract a base; aperiodicity guarantees long columns carry most of the mass.

Answer

Choose a set $A$ with $0 < μ (A) < ε / n$ ; ergodicity plus aperiodicity allow $μ (A)$ arbitrarily small with the induced return times unbounded. Partition $A$ by return time, $A_{m} = {r_{A} = m}$ , and build the Kakutani skyscraper with columns of height $m$ over each $A_{m}$ . The whole space is covered (mod $μ$ ) by the columns since $T$ is ergodic, so $\sum_{m} m μ (A_{m}) = 1$ by Kac (Exercise 5). Within each column of height $m \geq n$ , slice the height into blocks of length $n$ from the bottom: the base $B$ is the union, over all columns of height $m \geq n$ and over each full block, of the bottom level of that block. By construction $B, T B, \dots, T^{n - 1} B$ are disjoint (they are $n$ consecutive distinct levels) and their union is everything in the tall columns except a remainder of fewer than $n$ levels per column. The discarded remainder has measure at most $n \cdot \sum_{m} μ (A_{m}) = n μ (A) < ε$ , and the short columns (height $< n$ ) contribute at most $\sum_{m < n} m μ (A_{m}) \leq n μ (A) < ε$ as well; absorbing constants, $μ (⋃_{j < n} T^{j} B) > 1 - ε$ . This is Rokhlin's lemma: aperiodic systems are approximable by towers of arbitrary height covering almost all of the space.

Exercise 8 (hard, symbolic).

Let $U_{T}$ be the Koopman operator of an invertible measure-preserving system on $L^{2} (μ)$ . Show $U_{T}$ is unitary, and that $T$ is ergodic if and only if $1$ is a simple eigenvalue of $U_{T}$ (the only $U_{T}$ -invariant functions are the constants).

Hint

Use measure-preservation for the isometry, invertibility for surjectivity. For the ergodicity equivalence, relate $U_{T} f = f$ to $f \circ T = f$ a.e. and recall the definition of ergodicity via invariant functions.

Answer

For $f \in L^{2} (μ)$ , $∥ U_{T} f ∥_{2}^{2} = \int ∣ f \circ T ∣^{2} d μ = \int ∣ f ∣^{2} d μ = ∥ f ∥_{2}^{2}$ by the change-of-variables identity for measure-preserving $T$ , so $U_{T}$ is an isometry. Invertibility of $T$ makes $U_{T}$ surjective with inverse $U_{T^{- 1}}$ (also an isometry), so $U_{T}$ is unitary; in particular its spectrum lies on the unit circle. Now $U_{T} f = f$ means $f \circ T = f$ a.e., i.e. $f$ is an invariant function. If $T$ is ergodic, every invariant function is a.e. constant, so the eigenspace for eigenvalue $1$ is one-dimensional — the constants — hence $1$ is a simple eigenvalue. Conversely, if the only invariant functions are constants, then for any almost-invariant set $B$ the indicator $1_{B}$ is (mod $μ$ ) invariant, hence a.e. constant, forcing $μ (B) \in {0, 1}$ ; so $T$ is ergodic. The eigenvalue $1$ is therefore simple exactly when the system is ergodic.

Advanced results Master

Theorem 1 (Poincaré recurrence, infinitely-often form; Poincaré 1890). For a measure-preserving system on a probability space and $A \in B$ with $μ (A) > 0$ , almost every $x \in A$ satisfies $T^{n} x \in A$ for infinitely many $n$ . Equivalently, the conservative part of the Hopf decomposition is all of $X$ : there are no wandering sets of positive measure. A set $W$ is wandering if $W, T^{- 1} W, T^{- 2} W, \dots$ are pairwise disjoint; the recurrence proof is precisely the statement that a probability-preserving system admits no positive-measure wandering set ^{[Poincaré 1890]}.

Theorem 2 (Kac's return-time formula; Kac 1947). Let $(X, B, μ, T)$ be ergodic and $A \in B$ with $μ (A) > 0$ . The first-return time $r_{A}$ is integrable on $A$ and $\int_{A} r_{A} d μ = 1, equivalently \int_{A} r_{A} d μ_{A} = \frac{1}{μ ( A )} .$ The mean first-return time to a set is the reciprocal of its measure. The integrand is built from the Kakutani skyscraper: $\int_{A} r_{A} d μ = μ (⋃_{n \geq 0} T^{- n} A)$ always, and ergodicity makes the saturation full. A refinement (the Kac distribution / Kakutani) records that the induced system $(A, μ_{A}, T_{A})$ is itself ergodic and that the return-time partition ${r_{A} = n}$ has $μ_{A} (r_{A} = n)$ decaying with the tower geometry ^{[Kac 1947]}.

Theorem 3 (induced and suspension transformations; Kakutani 1943). The induced map $T_{A}$ on $(A, μ_{A})$ is measure-preserving, and inversely every measure-preserving system arises as a suspension (Kakutani tower) over an induced system with a measurable roof function. Inducing and suspending are mutually inverse operations on the category of measure-preserving systems, and they preserve ergodicity. This is the structural mechanism by which return-time data is repackaged as a new dynamical system, and it underlies the Ambrose-Kakutani representation of measurable flows as flows under a function ^{[Kakutani 1943]}.

Theorem 4 (Kakutani-Rokhlin tower lemma; Rokhlin 1948). An aperiodic measure-preserving system on a Lebesgue probability space admits, for every $n \geq 1$ and $ε > 0$ , a base $B$ with $B, T B, \dots, T^{n - 1} B$ disjoint and $μ (⋃_{j < n} T^{j} B) > 1 - ε$ . The lemma reduces many qualitative questions to combinatorics on a finite tower and is the engine behind approximation theorems (weak and uniform approximation of $T$ by periodic transformations) and the genericity results of Halmos and Rokhlin. With an additional measurable partition one obtains the coloured Rokhlin lemma used in orbit-equivalence theory ^{[Rokhlin 1948]}.

Theorem 5 (Koopman linearisation; Koopman 1931). The assignment $T \mapsto U_{T}$ embeds measure-preserving dynamics into the unitary (invertible case) or isometric (general case) operators on $L^{2} (μ)$ , intertwining the dynamics with a linear operator: $U_{T} (f \circ T^{n}) = f \circ T^{n + 1}$ . Spectral invariants of $U_{T}$ — the maximal spectral type, multiplicity, point versus continuous spectrum — are isomorphism invariants of the system. Discrete spectrum (an orthonormal eigenbasis for $U_{T}$ ) characterises systems measurably isomorphic to rotations on compact abelian groups, by the Halmos-von Neumann theorem; continuous spectrum on the orthocomplement of the constants is equivalent to weak mixing ^{[Koopman 1931]}.

Synthesis. The five results are one architecture viewed from successive heights, and the foundational reason they cohere is that each reads the single fact "finite measure cannot host a positive-measure wandering set" at a finer resolution. Poincaré recurrence is that fact bare; Kac's formula is exactly the same fact counted, the Kakutani skyscraper turning the no-escape geometry into the arithmetic identity $\sum_{n} n μ (r_{A} = n) = μ (saturation)$ , and this is the central insight that return frequency and region size are reciprocal. The induced transformation is dual to the suspension, and putting these together with the Rokhlin tower lemma shows every aperiodic system is, up to $ε$ , a single finite column — the bridge from measure-preserving abstraction to concrete combinatorial models. Koopman's linearisation generalises all of this into operator theory, where recurrence becomes the spectral statement that $1$ lies in the spectrum and the Birkhoff/von Neumann averages of 37.02.03 become the projection onto the eigenspace at $1$ ; this is exactly the conditional-expectation-onto-invariants picture, now seen as the bottom of a spectral decomposition whose higher eigenvalues encode the finer mixing and entropy structure of the system.

Full proof set Master

Proposition 1 (no positive-measure wandering set). A measure-preserving system on a probability space has no wandering set of positive measure.

Proof. Let $W$ be wandering: $T^{- i} W \cap T^{- j} W = \emptyset$ for $i \neq = j$ . By measure-preservation $μ (T^{- n} W) = μ (W)$ for all $n$ , so $\sum_{n = 0}^{\infty} μ (T^{- n} W) = \sum_{n} μ (W)$ . Disjointness gives $\sum_{n = 0}^{\infty} μ (T^{- n} W) = μ (⨆_{n} T^{- n} W) \leq μ (X) = 1$ . A constant sum $\sum_{n} μ (W)$ is finite only if $μ (W) = 0$ . $□$

Proposition 2 (Kac integrand via the skyscraper). With $A_{n} = {x \in A : r_{A} (x) = n}$ and the columns $T^{j} A_{n}$ ( $0 \leq j < n$ ) disjoint, $\int_{A} r_{A} d μ = μ (⨆_{n} ⨆_{j < n} T^{j} A_{n})$ .

Proof. The levels $T^{j} A_{n}$ within a fixed column are disjoint because if $T^{j} x = T^{j^{'}} x^{'}$ for $x, x^{'} \in A_{n}$ and $0 \leq j < j^{'} < n$ then applying $T^{n - j^{'}}$ shows a return to $A$ before time $n$ , contradicting $x^{'} \in A_{n}$ . Distinct columns are disjoint by a similar return-time argument. Measure-preservation gives $μ (T^{j} A_{n}) = μ (A_{n})$ , so the skyscraper has measure $\sum_{n} \sum_{j = 0}^{n - 1} μ (A_{n}) = \sum_{n} n μ (A_{n})$ . Since $r_{A} \equiv n$ on $A_{n}$ and the $A_{n}$ partition $A$ (mod the null set ${r_{A} = \infty}$ , which is null by Proposition 1), $\int_{A} r_{A} d μ = \sum_{n} n μ (A_{n})$ , equal to the skyscraper's measure. $□$

Proposition 3 (ergodic Kac equals one). If $(X, B, μ, T)$ is ergodic and $μ (A) > 0$ , then $\int_{A} r_{A} d μ = 1$ .

Proof. The saturation $S = ⋃_{n \geq 0} T^{- n} A$ satisfies $T^{- 1} S \subseteq S$ , and since $μ (T^{- 1} S) = μ (S)$ with $T^{- 1} S \subseteq S$ forces $μ (S ∖ T^{- 1} S) = 0$ , $S$ is almost invariant. Ergodicity gives $μ (S) \in {0, 1}$ ; as $A \subseteq S$ has positive measure, $μ (S) = 1$ . By Proposition 2 the skyscraper equals $S$ (mod $μ$ ), so $\int_{A} r_{A} d μ = μ (S) = 1$ . $□$

Proposition 4 (induced map of an ergodic system is ergodic). If $T$ is ergodic and $μ (A) > 0$ , the induced transformation $T_{A}$ on $(A, μ_{A})$ is ergodic.

Proof. Let $E \subseteq A$ be $T_{A}$ -invariant, $T_{A}^{- 1} E = E$ (mod $μ_{A}$ ). Form the $T$ -saturation $E = ⋃_{n \geq 0} T^{- n} E$ within the skyscraper: a point flows up its column and re-enters $A$ inside $E$ iff its column base lies over $E$ , so $E$ consists of whole columns and satisfies $T^{- 1} E = E$ (mod $μ$ ). Ergodicity of $T$ gives $μ (E) \in {0, 1}$ . Intersecting back with $A$ , $E \cap A = E$ (mod $μ$ ), so $μ_{A} (E) = μ (E) / μ (A) \in {0, 1/ μ (A) \cdot μ (A)} = {0, 1}$ . Hence $T_{A}$ is ergodic. $□$

Connections Master

The ergodic theorems of Birkhoff, von Neumann, and Kingman 37.02.03 are the quantitative completion of recurrence: where Poincaré guarantees infinitely many returns, Birkhoff gives their asymptotic frequency $μ (A)$ and von Neumann gives the $L^{2}$ convergence of the Koopman averages to the projection onto invariants. This unit supplies the measure-preserving-system framework and the recurrence skeleton on which those convergence theorems are proved.
The product-measure and Fubini-Tonelli machinery 02.07.07 is what makes the Bernoulli-shift example a genuine measure-preserving system: the shift-invariance of the product measure $p^{\otimes N}$ , and the identification of the invariant $σ$ -algebra with a tail field, both rest on the product construction and the iterated-integral identity. The induced-transformation measure-preservation proof also uses Fubini-style summation over the return-time partition.
The $σ$ -algebra and measurable-space foundations 02.07.01 are the substrate: measure-preservation is a statement about $T^{- 1} B \subseteq B$ and preimage measures, the first-return time $r_{A}$ is measurable because ${r_{A} = n}$ is a finite Boolean combination of preimages, and the Kakutani skyscraper is a measurable partition. Every construction here lives or dies on the measurability bookkeeping established there.
Mixing and spectral theory of measure-preserving systems 38.05.01 take the Koopman operator introduced here as their primary object: weak mixing is continuous spectrum on the orthocomplement of the constants, and the Halmos-von Neumann theorem identifies discrete-spectrum systems with group rotations. The recurrence and induced-map technology of this unit is the measure-theoretic groundwork those spectral refinements build upon.
Entropy theory 38.06.01 uses Rokhlin towers as a central technical device: the Shannon-McMillan-Breiman theorem and the Kolmogorov-Sinai entropy of a partition are estimated by cutting the dynamics into finite columns, exactly the Kakutani-Rokhlin approximation proved here. The tower lemma is the bridge from the abstract measure-preserving system to the combinatorial entropy count.

Historical & philosophical context Master

Poincaré proved the recurrence theorem in his 1890 Acta Mathematica memoir on the three-body problem ^{[Poincaré 1890]}, the same work that uncovered homoclinic tangles and the origins of chaos. The theorem was immediately recognised as paradoxical: it appeared to contradict the thermodynamic arrow of time, since a gas in a box must, by Poincaré's result, eventually return arbitrarily close to any earlier configuration, including a highly ordered one. Zermelo pressed this Wiederkehreinwand (recurrence objection) against Boltzmann's H-theorem in 1896; Boltzmann's reply — that the recurrence times for a macroscopic gas are astronomically large, far exceeding the age of the universe — correctly located the resolution in the gap between qualitative possibility and quantitative timescale, a gap the Kac formula later made precise.

Mark Kac introduced the return-time formula in 1947 in the Bulletin of the American Mathematical Society ^{[Kac 1947]}, in the language of discrete stationary stochastic processes rather than abstract dynamics; the identity that the expected recurrence time equals the reciprocal of the state's probability is the probabilistic face of the measure-theoretic skyscraper. The induced-transformation construction that organises the proof is due to Shizuo Kakutani in 1943 ^{[Kakutani 1943]}, who also introduced the dual suspension (tower under a function) that, with Ambrose, represents measurable flows. The tower approximation lemma was established by Rokhlin in 1948 ^{[Rokhlin 1948]} as a step in showing that a generic measure-preserving transformation is not mixing, and it became one of the most-used technical tools of the subject. Koopman's 1931 observation ^{[Koopman 1931]} that composition with a measure-preserving map is a unitary operator on $L^{2}$ supplied the operator-theoretic language in which von Neumann proved the mean ergodic theorem within months, fusing the dynamical and the spectral viewpoints that the rest of ergodic theory would inhabit.

Bibliography Master

@article{Poincare1890,
  author  = {Poincar\'e, Henri},
  title   = {Sur le probl\`eme des trois corps et les \'equations de la dynamique},
  journal = {Acta Mathematica},
  volume  = {13},
  year    = {1890},
  pages   = {1--270}
}

@article{Kac1947,
  author  = {Kac, Mark},
  title   = {On the notion of recurrence in discrete stochastic processes},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {53},
  number  = {10},
  year    = {1947},
  pages   = {1002--1010}
}

@article{Koopman1931,
  author  = {Koopman, Bernard O.},
  title   = {Hamiltonian systems and transformations in Hilbert space},
  journal = {Proceedings of the National Academy of Sciences},
  volume  = {17},
  number  = {5},
  year    = {1931},
  pages   = {315--318}
}

@article{Kakutani1943,
  author  = {Kakutani, Shizuo},
  title   = {Induced measure preserving transformations},
  journal = {Proceedings of the Imperial Academy of Tokyo},
  volume  = {19},
  year    = {1943},
  pages   = {635--641}
}

@article{Rokhlin1948,
  author  = {Rokhlin, Vladimir A.},
  title   = {A general measure-preserving transformation is not mixing},
  journal = {Doklady Akademii Nauk SSSR},
  volume  = {60},
  year    = {1948},
  pages   = {349--351}
}

@book{Walters1982,
  author    = {Walters, Peter},
  title     = {An Introduction to Ergodic Theory},
  publisher = {Springer},
  series    = {Graduate Texts in Mathematics},
  volume    = {79},
  year      = {1982}
}

@book{CornfeldFominSinai1982,
  author    = {Cornfeld, Isaac P. and Fomin, Sergei V. and Sinai, Yakov G.},
  title     = {Ergodic Theory},
  publisher = {Springer},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {245},
  year      = {1982}
}

@book{Petersen1983,
  author    = {Petersen, Karl},
  title     = {Ergodic Theory},
  publisher = {Cambridge University Press},
  year      = {1983}
}

@book{Aaronson1997,
  author    = {Aaronson, Jon},
  title     = {An Introduction to Infinite Ergodic Theory},
  publisher = {American Mathematical Society},
  series    = {Mathematical Surveys and Monographs},
  volume    = {50},
  year      = {1997}
}

Prerequisites

02.07.01
02.07.07
37.02.03

Tier anchors

beginner: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 1 (informal: the orbit picture, recurrence); Coudène 2016 *Ergodic Theory and Dynamical Systems* (Springer) Ch. 1 (recurrence as the first theorem of the subject)
intermediate: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) §1.1-1.5 (measure-preserving transformations, recurrence, ergodicity, the Koopman operator); Petersen 1983 *Ergodic Theory* (Cambridge) §2.1-2.4
master: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 1-2; Cornfeld-Fomin-Sinai 1982 *Ergodic Theory* (Springer Grundlehren 245) Ch. 1, Ch. 10 (induced transformations, Kakutani-Rokhlin towers); Petersen 1983 *Ergodic Theory* (Cambridge) §2.4 (Kac's lemma); Aaronson 1997 *An Introduction to Infinite Ergodic Theory* (AMS) Ch. 1 (recurrence vs. conservativity)

References

Poincaré — Sur le problème des trois corps et les équations de la dynamique · Acta Mathematica 13 (1890), 1-270 (the recurrence theorem)
Kac — On the notion of recurrence in discrete stochastic processes · Bulletin of the American Mathematical Society 53 (1947), 1002-1010
Koopman — Hamiltonian systems and transformations in Hilbert space · Proceedings of the National Academy of Sciences 17 (1931), 315-318
Kakutani — Induced measure preserving transformations · Proceedings of the Imperial Academy of Tokyo 19 (1943), 635-641
Rokhlin — A general measure-preserving transformation is not mixing · Doklady Akademii Nauk SSSR 60 (1948), 349-351 (Rokhlin tower lemma)
Walters — An Introduction to Ergodic Theory · Springer GTM 79, 1982, Ch. 1 (measure-preserving systems, recurrence)
Cornfeld-Fomin-Sinai — Ergodic Theory · Springer Grundlehren 245, 1982, Ch. 1, Ch. 10 (induced transformations, towers)

Estimated time

beginner: 18m
intermediate: 55m
master: 95m