Since the Borel -algebra of is generated by the projections , , we only have to show that for all and all : . Now by definition and since we conclude that
Remark: In probability the projections are independent and identically distributed -valued random variables, a so called i.i.d. sequence in with distribution .
The following result is a topological version of the previous proposition. Recall the following definitions from topology:
A subset of a metric space is called a -set if it's the intersection of a sequence of open sets.
A basis for the topology of is a collection , , of open subsets, such that for every open subset of there is a subset such that .
For example, the irrational numbers form a subset of the reals. Any closed or open subset of a metric space is a subset, but any countable subset of an uncountable complete metric space is not ! If is a dense subset of a metric space , then the collection
is a basis for the topology of .
If is separable and locally compact and is continuous, then , is a linear operator on satisfying . By the Riesz Representation Theorem the space of all finite signed Borel measures on is isometrically isomorphic to the dual of : for any there is exactly one signed Borel measure such that for all .
Using this identification the dual (or adjoint) mapping is given by : indeed for all we have by definition and the transformation theorem of measuer theory
The Banach space is the space of all finite signed Borel measures on equipped with the total variation norm: . If in addition is discrete, then . A formally more general type of operators are so called Markov operators, which we are going to discuss in the subsequent section.
Let us put for and :
Then is a probability measure, is measurable and by the Markov property is the conditional probability that given , i.e.
These conditions simply say that is a regular conditional probability for given . Therefore we also have
Finally the Markov operator is given by
The mapping , , is called a Markovian transition function.
Since iff we infer from e.g. the extended Markov property:
Finally let be a probability measure on and put for all :
Then is a probability measure on .
Formally the Markov operator is only defined on the vector space of bounded, measurable functions on . Now the existence of an invariant measures allows us to define as a positive, linear contraction on the Hilbert space :
1. By Jensen's inequality we have for all and all :
and by invariance for :
Since is dense in , there is a unique bounded, linear extension of to and this extension is obviously a positive contraction.
Let us remark that there is another commonly used notion of ergodicity for stochastic matrices (irreducible and aperiodic states), which is a bit stronger than the notion we employ! However, we will never refer to that stronger notion.
If is supported on , then for all bounded is harmonic iff for all and all satisfying : . Hence for harmonic functions we have .
Remark: Starting at the sequence converges 'weakly' to a random variable , whose distribution is the harmonic measure on with respect to . Hence for all continuous the function is the harmonic extension of into the interior of .
Invariant measures on finite sets
For every finite stochastic matrix we have , because is an eigenvalue for and thus is an eigenvalue for - the spectrum of is the complex conjugate of the spectrum of ! Thus . Yet we don't know if contains a measure.
1. Suppose , then the associated Markov operator and its adjoint are given by
Let be the set of probability measures on , is compact and convex. Since we infer from Brouwer's fixed point theorem that there is some such that .
2. Alternatively and more elementary we may take any and define
Then and thus any accumulation point of the sequence is a fixed point of . This also shows that converges to if is the unique invariant probability measure. Hence for all
If , then for all : and therefore must be constant
Noteworthy the alternative proof also works if the sequence has some accumulation point with respect to some metric , say, which is weaker than the metric defined by the norm, i.e. for some constant .
Computationally the presented proofs don't work well because averaging gives a pretty slow algorithm. Usually a certain convex combination of the powers of yields a stochastic matrix with strictly positive entries (cf. exam) and the sequence for any converges (exponentially fast) to the invariant measure of , which is also the invariant measure of , provided it's unique.
This C code generates text by means of an -Markov chain. Short texts such as Time will almost be reproduced (for ), wheras long texts such as Mann get jumbled (for ).
Putting in we see that the group , , also determines :
The notion of a complete vector field allows us to write as . By the group property we get:
and thus is equivalent to:
That's the case for a diffusion on a compact Riemannian manifold. In general the diffusion semigroup generated by sends any onto a smooth function. If the Riemannian manifold is complete and satisfies a certain geometric condition then is Feller. The associated Markov process , , is reversible with respect to the speed measure - in case the Markov process is called Brownian motion on with drift.
, , for some . Compute the density . The Jacobi polynomials are eigen functions of for the eigenvalues and they form a complete orthogonal set. For these polynomials are called ultraspherical or Gegenbauer polynomials.
, , . Compute the density . The Hermite polynomials are eigen functions of for the eigenvalues and they form a complete orthogonal set.
, , for some . Compute the density . The Laguerre polynomials are eigen functions of for the eigenvalues and they form a complete orthogonal set.
We just talk about convolution semigroups on : suppose , , is a family of probability measures on such that for all : , where
is the convolution of and . The -stable distribution defined in exam is a prominent example. Here are a few more: