On correlation calculus for multivariate martingales (Q2368169)

Let \(M:\Omega\times[0,\infty)\to R^ n\) and \(m:\Omega\times[0,\infty)\to\mathbb{R}^ m\) be locally square integrable martingales, \(\langle m,M\rangle:\Omega\times[0,\infty)\to\mathbb{R}^{m\times n}\) be their predictable covariation process, \(c(m,M)_ t=\langle m\rangle_ t- \langle m,M\rangle_ t\langle M\rangle^{-1}_ t\langle M,m\rangle_ t\), where \(\langle M\rangle_ t\) is supposed to be invertible. The process \(c(m,M)\) characterizes the correlation between \(m\) and \(M\). For example, when \(m\) and \(M\) are Gaussian martingales, \(c(m,M)\) has an interpretation as an \(({\mathcal L}_ 2\)-)projection error; in a statistical context \(c(m,M)\) is a measure of deficiency when somebody compares an arbitrary estimator with an optimal one. The authors omit the restriction that \(\langle M\rangle_ t\) is invertible and replace \(\langle M\rangle_ t^{-1}\) by Moore-Penrose inverse \(\langle M\rangle^ +_ t\). The properties of \(\langle M\rangle^ +\) and of \(P=\langle M\rangle\langle M\rangle^ +\) are investigated. Right continuity of \(c(m,M)\) is established and principal integral representation for \(c(m,M)\) is presented. Linear dependence between \(m\) and \(M\) is defined by \(c(m,M)=0\) and is characterized by the property that there exists such constant random matrix \(C\) that \(m=CM\).