On a probability space, a measure-preserving transformation is ergodic when each measurable set satisfying modulo null sets has probability zero or one. Such sets form the invariant sigma-algebra. Equivalently, every measurable invariant observable is constant almost everywhere: its real and imaginary level sets are invariant, and the zero-or-one condition forces its distribution to be concentrated at one value. Conversely, the indicator function of an invariant set is an invariant observable.[ref]
254A: Topics in Ergodic Theory by Terence Tao. Lecture9, Theorem3, characterizations of ergodicity
Ergodicity says that the measure cannot split into two invariant pieces of positive probability. It does not say every orbit visits every point, or that all correlations decay. A cyclic permutation of a finite set with uniform probability is ergodic because its invariant sets are unions of whole cycles and there is just one cycle. The identity on two positive-mass points is not ergodic, since either singleton is invariant.
In this probability setting, the Birkhoff ergodic theorem gives almost everywhere for every integrable . On an infinite-measure space, invariant-set ergodicity still has a definition, but the stated equivalence using only integrable invariant functions and the probability-average formula cannot be transferred without additional qualifications.
Ergodicity is a concept from statistical mechanics and dynamical systems theory that describes the behavior of systems over time. In general terms, a system is considered ergodic if its time averages are equivalent to its ensemble averages. This means that a sufficiently long observation of a single trajectory (or the time evolution of a single state of the system) will provide the same statistical properties as observing a large number of different states of the system at a single point in time (the ensemble).
New to topics? Read the docs here!