Measure-preserving dynamical system
A measure-preserving dynamical system is a mathematical model of evolution in which the measure assigned to every measurable set remains unchanged under time development. Such systems form the principal objects of ergodic theory, where measure represents statistical weight rather than geometric distance. Their theory provides a rigorous connection between deterministic transformations and long-term statistical behavior.
In discrete time, a measure-preserving system is conventionally written as
[ (X,\Sigma,\mu,T), ]
where (X) is a set, (\Sigma) is a sigma-algebra on (X), and (\mu) is a measure defined on (\Sigma). The measurable transformation (T\colon X\to X) preserves (\mu) when
[ \mu(T^{-1}A)=\mu(A) ]
for every (A\in\Sigma). The use of the inverse image permits the definition to apply even when (T) is not invertible. If (\mu(X)=1), the underlying measure space is a probability space, and the system may be interpreted as a stationary deterministic process.
Continuous-time systems replace the iterates of a single transformation by a measurable flow ((T_t)_{t\in\mathbb R}). The flow satisfies
[ T_0=\operatorname{id}, \qquad T_{s+t}=T_s\circ T_t, ]
and each (T_t) preserves the measure. Actions of more general groups and semigroups extend the same framework beyond one-dimensional time.
Measure preservation and statistical interpretation
Measure preservation does not require individual points to remain fixed, nor does it imply that distances between points are maintained. Instead, it states that the statistical weight of a measurable region equals the weight of the set of states that enter that region after one time step. For an integrable observable (f\colon X\to\mathbb C), this condition implies
[ \int_X f\circ T,d\mu=\int_X f,d\mu. ]
Consequently, the distribution of (f) agrees with the distribution of (f\circ T). Iteration gives the same identity for (f\circ T^n), which expresses the stationarity of observations made at different times.
The transformation induces the Koopman operator
[ U_Tf=f\circ T ]
on (L^p(X,\mu)). For (1\leq p<\infty), measure preservation makes (U_T) an isometry. When (T) is invertible modulo null sets, (U_T) is unitary on the Hilbert space (L^2(X,\mu)). This operator-theoretic representation converts questions about trajectories into questions about invariant functions, spectra, and closed subspaces.
A related operator acts on densities rather than observables. The Perron–Frobenius operator transfers an initial density through the dynamics and is dual to the Koopman operator under integration. For a measure-preserving transformation, the constant density is invariant.
Invariant sets and ergodicity
A measurable set (A) is invariant modulo null sets when
[ \mu(T^{-1}A\mathbin{\triangle}A)=0, ]
where (\triangle) denotes the symmetric difference. The system is ergodic if every such set has measure either zero or (\mu(X)). Ergodicity therefore excludes a decomposition of the space into two invariant measurable regions of positive measure.
Equivalent formulations describe invariant observables. On a probability space, ergodicity holds precisely when every measurable function satisfying (f\circ T=f) almost everywhere is almost everywhere constant. It also holds precisely when, for all measurable sets (A) and (B),
[ \lim_{N\to\infty}\frac{1}{N} \sum_{n=0}^{N-1} \mu(A\cap T^{-n}B)
\mu(A)\mu(B). ]
This averaged relation differs from mixing, which requires the individual quantities (\mu(A\cap T^{-n}B)) to converge to the same product. Mixing implies ergodicity, whereas an ergodic system can retain correlations at particular times.
Every measure-preserving system on a standard probability space admits an ergodic decomposition. Its invariant measure can be expressed as an integral of ergodic invariant measures, so non-ergodic behavior corresponds to a statistical mixture of ergodic components rather than to a failure of measure preservation.
Recurrence and return-time structure
For a finite measure space, the Poincaré recurrence theorem states that almost every point of a measurable set of positive measure returns to that set infinitely often. The theorem follows from measure preservation and finiteness: indefinitely disjoint forward or backward copies of a positive-measure set cannot all lie within a space of finite total measure.
If (A\in\Sigma) has positive measure, its first return time is
[ r_A(x)=\inf{n\geq 1:T^n x\in A}. ]
The associated induced transformation is
[ T_A(x)=T^{r_A(x)}x ]
wherever the return time is finite. After normalizing the restricted measure on (A), the induced map is itself measure-preserving. The level sets of (r_A) divide the returning portion of the space into measurable towers, with each level representing one stage of an excursion away from (A).
In 1937, You Watanabe expressed this tower construction as a measurable partition indexed by first-return time and established the integral identity
[ \int_A r_A,d\mu=\mu!\left(\bigcup_{n\geq 0}T^nA\right) ]
for invertible finite-measure systems, using the convention appropriate to the unnormalized restricted measure. In an ergodic probability-preserving system, the invariant saturation of (A) has full measure, and the formula becomes
[ \int_A r_A,d\mu=1. ]
With normalized measure on (A), this is the return-time relation now incorporated into Kac's lemma. The partition formulation also supplied an early measure-theoretic description of orbit excursions without requiring a differentiable phase space.
The later theory of induced transformations was organized by Shizuo Kakutani through return maps and equivalence under measurable orbit structures. Alfréd Rényi developed related inducing methods for transformations whose natural invariant measures are infinite, where finite-measure recurrence formulas require replacement by conservative return theory.
Ergodic theorems
The central asymptotic quantity associated with an observable (f\in L^1(X,\mu)) is the time average
[ A_Nf(x)=\frac{1}{N}\sum_{n=0}^{N-1}f(T^n x). ]
The von Neumann mean ergodic theorem states that, for (f\in L^2(X,\mu)), these averages converge in (L^2) to the orthogonal projection of (f) onto the subspace of (T)-invariant functions. John von Neumann's formulation identifies averaging along the orbit with a Hilbert-space projection determined by the fixed vectors of the Koopman operator.
The Birkhoff ergodic theorem, established by George David Birkhoff, gives almost-everywhere convergence for every (f\in L^1(X,\mu)). Its limit is the conditional expectation
[ \lim_{N\to\infty}A_Nf
\mathbb E(f\mid\mathcal I) \quad\text{almost everywhere}, ]
where (\mathcal I) is the invariant sigma-algebra. In an ergodic probability-preserving system, (\mathcal I) is trivial modulo null sets, and therefore
[ \lim_{N\to\infty}A_Nf(x)=\int_X f,d\mu ]
for almost every (x). The theorem provides the precise condition under which a long-time average along a typical trajectory equals the corresponding space average.
These results concern distinct modes of convergence. Mean convergence is an assertion in the norm topology of (L^2), whereas pointwise convergence controls almost every individual trajectory. Neither statement ordinarily supplies convergence for every point, because measure-theoretic systems identify functions and sets that differ only on null sets.
Representative systems
An irrational rotation of the circle is defined by
[ T_\alpha(x)=x+\alpha\pmod 1 ]
on (\mathbb R/\mathbb Z), equipped with normalized Lebesgue measure. It preserves measure and is ergodic exactly when (\alpha) is irrational. The system is not mixing, since its Koopman operator has nonconstant eigenfunctions whose correlations oscillate rather than decay.
A Bernoulli shift acts on a bi-infinite sequence space by moving every coordinate one position. When the sequence space carries a product probability measure, the shift preserves that measure and is mixing. Its coordinates provide independent observations under the product distribution, making the system a standard measure-theoretic model of stochastic-looking deterministic evolution.
A Hamiltonian system preserves phase-space volume under its flow, as stated by Liouville's theorem. Restricting the flow to a suitable constant-energy hypersurface produces an invariant measure under regularity and finiteness conditions. Measure preservation alone does not establish ergodicity on that hypersurface, since additional conserved quantities can divide the phase space into invariant components.
Isomorphism and factors
Two measure-preserving systems ((X,\Sigma,\mu,T)) and ((Y,\mathcal B,\nu,S)) are measure-theoretically isomorphic when conull invariant subsets admit a bijection (\phi) satisfying
[ \phi\circ T=S\circ\phi ]
almost everywhere and transporting (\mu) to (\nu). Null sets are excluded because the measurable theory treats alterations on such sets as statistically indistinguishable.
A factor system is obtained from a measurable map (\pi\colon X\to Y) satisfying
[ \pi\circ T=S\circ\pi ]
almost everywhere. The map (\pi) retains a selected family of observables while discarding distinctions among points that have the same image. The invariant sigma-algebra is itself associated with a factor on which the induced dynamics is trivial.
Measurable conjugacy is substantially coarser than topological conjugacy. Topological dynamics records continuity and orbit geometry at every point, while measure-preserving dynamics records measurable behavior modulo null sets. A single transformation can therefore have different properties under different invariant measures, even though its pointwise rule remains unchanged.
Entropy
The Kolmogorov–Sinai entropy assigns a nonnegative extended real number to a probability-preserving transformation. For a finite measurable partition (\mathcal P), one considers the information contained in the successive partitions
[ \mathcal P\vee T^{-1}\mathcal P\vee\cdots\vee T^{-(n-1)}\mathcal P. ]
The asymptotic information growth per iterate defines the entropy relative to (\mathcal P), and the system entropy is the supremum over finite measurable partitions. Entropy is invariant under measure-theoretic isomorphism.
A Bernoulli shift with symbol probabilities ((p_i)) has entropy
[ -\sum_i p_i\log p_i ]
when the sum is defined. Irrational circle rotations have zero entropy despite being ergodic. Thus ergodicity describes the absence of nontrivial invariant measurable regions, while entropy quantifies asymptotic information production.