Adaptive estimation for bifurcating Markov chains (Q2405167)

For each \(m=0,1,\dots\), \(\mathbb{G}_{m}\) is defined by \[ \mathbb{G}_m:=\begin{cases} \{1\} &\text{if }m=0\\ \{0,1\}^{m} &\text{if }m\geq1.\end{cases} \] Let \(\mathbb{T}:=\bigcup_{m=0}^{\infty}\mathbb{G}_{m}\). For \(m=0,1,\dots\), and \(u\in \mathbb{G}_{m}\) set \(|u|=m\) and define \[ \begin{aligned} u0 &= (u,0)\in \mathbb{G}_{m+1},\\ u1 &= (u,1)\in \mathbb{G}_{m+1}.\end{aligned} \] Consider the following model. Let \((E,\mathcal{E})\) be a measurable space; \(\mathcal{P}:(\mathcal{E}\otimes\mathcal{E})\times E\to[0,1]\) a Markov kernel from \((E,\mathcal{E})\) to \((E\times E,\mathcal{E}\otimes \mathcal{E})\), that is {\parindent=6mm \begin{itemize}\item[{\(\cdot\)}] for each \(x\in E\), \(B\mapsto \mathcal{P}(B,x)\) is a probability on \((E\times E,\mathcal{E}\otimes \mathcal{E})\), \item[{\(\cdot\)}] for each \(B\in (\mathcal{E}\otimes\mathcal{E})\), \(x\mapsto \mathcal{P}(B,x)\) is a measurable function. \end{itemize}} Let \(\left(\Omega,\mathcal{F},(\mathcal{F}_{m})_{m\geq0},\mathbb{P}\right)\) be a filtered probability space, that is {\parindent=6mm \begin{itemize}\item[{\(\cdot\)}] \(\Omega\) is a nonempty set, \item[{\(\cdot\)}] \(\mathcal{F},\mathcal{F}_0,\mathcal{F}_1,\dots\) are \(\sigma\)-fields of \(\Omega\) such that \(\mathcal{F}_0\subset \mathcal{F}_1\subset \mathcal{F}_2\subset \dots \mathcal{F}\), \item[{\(\cdot\)}] \(\mathbb{P}\) is a probability on \((\Omega,\mathcal{F})\). \end{itemize}} For each \(u\in \mathbb{T}\), \(X_{u}:\Omega\to \mathbb{R}\) is \(\mathcal{F}_{|u|}\)-measurable. For each \(g:(E\times E)\times E\to \mathbb{R} (\mathcal{E}\otimes\mathcal{E})\otimes\mathcal{E}\)-measurable and bounded, \(\mathcal{P}_{g}:E\to \mathbb{R}\) is defined by \[ \mathcal{P}_{g}(x)=\int_{E\times E}g(y,z,x)\mathcal{P}(dy,dz,x),\;x\in E. \] \((X_{u})\) is a bifurcating Markov chain if \[ E\left(\prod_{u\in \mathbb{G}_{m}}g_{u}(X_{u},X_{u0},X_{u1})|\mathcal{F}_{m}\right)=\prod_{u\in \mathbb{G}_{m}}\mathcal{P}_{g_{u}}(X_{u}) \] for every \(m\geq0\) and any family of (bounded) measurable functions \((g_{u})_{u\in \mathbb{G}_{m}}\). For each \(n\geq0\), \(\mathbb{T}_{n}\) is defined by \(\mathbb{T}_{n}:=\bigcup_{m=0}^n\mathbb{G}_{m}\). Assume we observe \(X^n:=(X_{u})_{u\in \mathbb{T}_{n}}\). For each \(m\geq0\), \(Y_{m}\) is a random variable defined by \[ \begin{aligned} Y_0 &= X_0, \\ Y_{m} &= X_{0,\varepsilon_1,\dots,\varepsilon_{m}},\; m\geq1,\end{aligned} \] where \((\varepsilon_{m})_{m\geq1}\) is a sequence of independent Bernoulli variables with parameter \(1/2\), independent of \((X_{u})_{u\in \mathbb{T}}\). \((Y_{m})_{m\geq1}\) has transition probability \(\mathcal{Q}=\frac12 (\mathcal{P}_0+\mathcal{P}_1)\) where for each \(A\in E\) \[ \begin{aligned} \mathcal{P}_0(x,A) &= \int_{y\in A}\int_{z\in E}\mathcal{P}(dy,dz,x),\; x\in E,\\ \mathcal{P}_1(x,A) &= \int_{z\in A}\int_{y\in E}\mathcal{P}(dy,dz,x),\; x\in E.\end{aligned} \] By previous results it is known that if \((Y_{m})_{m\geq1}\) is ergodic with distribution \(\nu\) then \[ \frac{1}{|\mathbb{G}_{n}|}\sum_{u\in \mathbb{G}_{n}}g(X_{u})\to \int_{E}g\cdot d\nu \] for appropriate functions \(g\). In this paper, estimators of \(\nu\), \(\mathcal{Q}\), and \(\mathcal{P}\) are constructed under certain conditions of regularity. It also proves that the proposed estimators have good properties in the minimax sense. The proofs of the results are rigorous and detailed but are hard to follow for those who do not have good knowledge about the theory of stochastic processes. A section of this paper presents in detail examples of successful application of both the model and the proposed estimation method. The list of references is extensive and updated.