Markov operators and OS-positive processes (Q1820410): Difference between revisions

Let \((X_ 1,\mu_ 1)\) and \((X_ 2,\mu_ 2)\) be probability spaces. A Markov operator is a bounded linear operator \(A: L^ 2(X_ 1,\mu_ 1)\to L^ 2(X_ 2,\mu_ 2)\) such that Af\(\geq 0\) a.e. whenever \(f\geq 0\) a.e. and such that \(A1=1\) and \(A^*1=1\). Markov operators are continuous infinite dimensional analogues of rectangular doubly stochastic matrices. The author shows that every Markov operator A is factorable in the sense that there is a probability space (Y,\(\nu)\) and a minimal set \((V_ 1,V_ 2)\) of Markov isometries \(V_ i: L^ 2(X_ i,\mu_ i)\to L^ 2(Y,\nu)\) such that \(V_ 2*V_ 1=A\) and that this decomposition is essentially unique up to multiplication by a unitary Markov operator. Here \((V_ 1,V_ 2)\) is said to be a minimal set if \(\cup V_ i(L^{\infty}(X_ i,\mu_ i))\) generates \(L^{\infty}(Y,\nu)\) as \(W^*\)-algebras. A Markov isometry V is also shown to be precisely those Markov operators such that \(V\upharpoonright L^{\infty}\) is a multiplication operator. In general, a bounded linear operator A on \(L^ 2(X,\mu)\) is said to be factorable if there is a (perhaps infinite) measure space (Y,\(\nu)\) and an order preserving operator \(B: L^ 2(X,\mu)\to L^ 2(Y,\nu)\) such that \(B^*B=A\). If A is a factorable Markov operator, the author shows that B may be taken to be a Markov operator. In general, if A is a self- adjoint order preserving operator, then A is factorable if and only if A is in the \(\sigma\)-weak closure of the cone generated by \(f\otimes f\) \((f\in L^ 2(X,\mu)\), \(f\geq 0)\). This implies that the set of factorable Markov operators forms a weakly compact convex set in the algebra of bounded linear operators on \(L^ 2(X,\mu)\). Also if \(A\geq 0\) is a Markov operator, then a factorization of A (when it exists) is in fact a dilation of a Markov projection, i.e., there is a Markov isometry \(V: L^ 2(X,\mu)\to L^ 2(\tilde X,{\tilde \mu})\) and a Markov projection E on \(L^ 2(\tilde X,{\tilde \mu})\) with \(A=V^*EV.\) Let (X,\({\mathfrak a},\phi,P)\) be a full Markov system consisting of a compact Hausdorff space X, a \(C^*\)-algebra \({\mathfrak a}\) with identity, a state \(\phi\) on \({\mathfrak a}\), and a continuous linear map P of C(X) onto the space of continuous linear maps on \({\mathfrak a}\) such that (i) \(P_ f\geq 0\) whenever \(f\geq 0\), (ii) \(P_ 1\) fixes the unit of \({\mathfrak a}\) and (iii) \(\phi (P_ f(a)*a)\geq 0\) for all \(a\in {\mathfrak a}\) and \(f\geq 0\) such that \(clos\{P_ f(1)|\) \(f\geq 0\}={\mathfrak a}^+\). For example, let \({\mathfrak a}\) be a separable \(C^*\)-algebra with tracial state \(\phi\), let \(\{e_ n\}\) be a subset of \({\mathfrak a}^+\) which generates \({\mathfrak a}^+\) such that \(\sum \| e_ n\| \leq 1\), let \(X={\mathbb{N}}\cup \{\infty \}\) (the one-point compactification of \({\mathbb{N}})\) and let \(P_ f(a)=\sum f(n)(e_ n)^{1/2}a(e_ n)^{1/2}\). Back to the general case again, let \(\mu\) be the Baire probability measure satisfying \(\int fd\mu =\phi (P_ f(1))\). The positive bounded linear operator on \(L^ 2(X,\mu)\) defined by \(<Af,g>=\phi (P_ f(1)P_{\bar g})\) is a positive Markov operator. Then the author proves that the operator A is factorable if and only if the \(C^*\)-algebra \(\pi_{\phi}({\mathfrak a})\) is abelian. Here \(\pi_{\phi}\) is the GNS representation. This result allows the author to construct a nonfactorable Markov operator A using \(n\times n\) complex matrices \(M_ n\). Here A is the convolution operator \(Af=f*\psi\) on \(L^ 2(U,m)\) where U is the unitary group of the \(M_ n\), m is the Haar measure on U, and \(\psi (u)=| <u\zeta,\zeta >|^ 2\) for \(u\in U.\) The author now discusses the relation of Markov operators to OS-positive processes, i.e., Osterwalder-Schrader positive processes. Incidentally, the author provides the reader with a careful introduction to the ideas from stochastic processes needed for the remainder of the paper. Here let (\(\Omega\),p) be a probability space and let \(\Sigma\) be a standard Borel space. A stochastic process with state space \(\Sigma\) is a sequence \(\{X_ n|\) \(n\in {\mathbb{Z}}\}\) of Borel functions \(X_ n: \Omega \to \Sigma\) such that the \(W^*\)-algebra generated by the functions \(\omega \to f(X_ n(\omega))\) (f: \(\Sigma\to {\mathbb{C}}\) is a bounded Borel function) is \(L^{\infty}(\Omega,p)\). A Markov process is a process \(\{X_ n\}\) with \(E(f(X_ n)| X_{n_ 1},...,X_{n_ m}\}=E(f(X_ n)| X_{n_ m}\}\) for all bounded Borel functions f and for all \(n_ 1<...<n_ m<n\). Here E denotes the usual conditional expectation from probability theory. Now let \(\{X_ n\}\) be a stochastic process. Then \(\{X_ n\}\) is said to be stationary (resp., symmetric) if \(p\{X_{-n+1}\in E_{- n},...,X_{n+1}\in E_ n\}=p\{X_{-n},...,X_ n\in E_ n\}\) (resp., if \(p\{X_{-n}\in E_{-n},...,X_ n\in E_ n\}=p\{X_ n\in E_{- n},...,X_{-n}\in E_ n\})\) for all Borel sets \(E_ i\). A stationary processes \(\{X_ n\}\) is called an OS-positive process if, for every \(n\geq 1\) and every bounded Borel function \(f: \Sigma\) \({}^{n+1}\to {\mathbb{C}}\), the \(L^ 2(\Omega,p)\) inner product \(<f(X_ 0,X_ 1,...,X_ n),f(X_ 0,X_{-1},...,X_{-n})>\) is nonnegative. \textit{A. Klein} [Z. Wahrscheinlichkeitstheorie Verw. Geb. 40, 115-124 (1977; Zbl 0343.60028)] showed that every Gaussian OS-positive process \(\{X_ t|\) \(t\in {\mathbb{R}}\}\) (N.B. the use of the time variable t) has an extension to a Markov process. An extension of a process \(\{X_ t\}\) is another process \(\{\bar X{}_ t\}\) with state space \({\tilde \Sigma}\) together with a Borel function \(\psi\) : \({\tilde \Sigma}\to \Sigma\) such that \(X_ t=\psi \circ \tilde X_ t\) for all t. In the second part of this paper, the analogous problem for the process \(\{X_ n|\) \(n\in {\mathbb{Z}}\}\) is considered. To each process (\(\Omega\),p) one can associate a canonical Markov operator K as follows: let \(E_+\) be the projection of \(L^ 2(\Omega,p)\) on the subspace generated by \(\{f(X_ n)|\) \(N\geq 0\); f bounded, Borel function on \(\Sigma\}\) and let \(K=E_+R\) restricted to \(E_+(L^ 2(\Omega,p))\). Here R is the unique unitary operator on \(L^ 2(\Omega,p)\) with \(Rf(X_{-n},...,X_ n)=f(X_ n,...,X_{-n})\) which exists by a symmetric hypothesis. Then the author proves that if an OS- process can be extended to a symmetric Markov process, then its canonical Markov operator K is factorable. Using the material on factorable Markov operators from the earlier sections, he constructs examples of OS- processes that are not extendable to Markov processes.