A \(q\)-analogue of de Finetti's theorem (Q2380233)

Given \(q> 0\) a random binary sequence \(\varepsilon= (\varepsilon_1,\varepsilon_2,\dots)\) is \(q\)-exchangeable if \[ P(\varepsilon_1, \varepsilon_2,\dots, \varepsilon_{i-1}, \varepsilon_{i+1}, \varepsilon_i,\dots,\varepsilon_n)= f(\varepsilon_i- \varepsilon_{i+1})P(\varepsilon_1,\dots, \varepsilon) \] with \(f(t)= q^t\). The paper investigates what extension of De Finetti's theorem still holds. It turns out that there is no simple extension. For \(0< q< 1\) there is a bijection \(P\leftrightarrow\mu\) between probability laws \(P\) of \(\varepsilon\) and \(\mu\) a measure on the set \(\Delta_q= \{0,1,q,q^2,\dots\}\). The squence \(\varepsilon\) may be constructed \((*)\) by choosing \(x\in\Delta_q\) with probability distribution \(\mu\) and then taking a \(q\)-analogue of the Bernoulli process indexed by \(x\). The distribution of \(\varepsilon\) is determined by \(v_{nk}= P\) (\(\varepsilon_i=1\), \(i\leq k\), \(\varepsilon_i=0\), \(k< i\leq n\)). For the proof the authors define a weighted directed graph \(\Gamma(q)\) with vertex set \(\Gamma= \{(l,k): l,k= 0,1,\dots\}\) and outgoing edges \((l,k)\to(l+ 1,k)\) and \((l,k)\to(l,k+1)\) with weights \(1\) and \(q^l\), respectively. A path in \(\Gamma(q)\) has weight equal to the product of the weights of its edges. A Markov chain \(S= (S_n)\) on \(\Gamma(q)\) starting at \((0,0)\) and moving along its edges is central when for every fixed \((n-k, k)\) the conditional probability given \(S_n=(n-k,k)\) that \(S\) follows a path \(W\) from \((0,0)\) to \((n-k,k)\) is proportional to the weight of \(W\). There is a bijection between \(q\)-exchangeable binary sequences and central Markov chains on \(\Gamma(q)\). It is shown that \(P\,(S_n= (n-k,k))={n\brack k}v_{nk}\) with \(v_{nk}\) as above and \({n\brack k}\) the \(q\)-binomial generalisation of \({n\choose k}\). The \(v_{nk}\) satisfy a recurrence, viz. (9). The consruction \((*)\) may be seen as the barycentric decomposition of solutions of (9) into extreme solutions. These extreme solutions are found to be \[ \Phi_{nk}(x)= q^{k(k-n)}x^{n-k} \prod_{i<k} (1-xq^i) \] indexed by \(x\in\Delta_q\) and the decomposition is determined by \[ v_{nk}= \sum_{x\in\Delta_q} \Phi_{nk}(x)\,\mu(x) \] with \(\mu\) a certain measure on \(\Delta_q\). Examples studied: two \(q\)-analogues of the Bernoulli process and one of Pólya's urn process. Section 5 gives an interpretation of the theorem in terms of random subspaces in an infinite-dimensional vector space over \(\mathbb{F}_q\).