Matrix factorization for multivariate time series analysis (Q2008611)

The authors deal with the observations of multivariate series \(X = (x_{i,t})_{(i,t)\in [1,d]\times [1,T]}\), \[ X = M +\varepsilon = UV\Lambda + \varepsilon, \] where \(\varepsilon\) is a noise process, \(M\) is a matrix of rank \(k\in[1, T]\), the matrices \(U\in {\mathcal M}_{d,k}(\mathbb R)\), and \(V\in {\mathcal M}_{k,\tau}(\mathbb R)\) are unknown, while \(\tau\leq T\) and \(\Lambda\in {\mathcal M}_{\tau,r}(\mathbb R)\), such that \(\text{rank}(\Lambda) =\tau\), are known. By multiplying both sides of the equation by the pseudo-inverse \(\Lambda^+ = \Lambda^*(\Lambda \Lambda^+)^{-1}\), we obtain the ``simplified model'' \[ \widetilde{X }= \widetilde{M} +\tilde{\varepsilon}, \] with \(\widetilde{X }= X\Lambda^+\), \(\widetilde{M } = UV\) and \(\tilde{\varepsilon} = {\varepsilon}\Lambda^+\). In this model, the estimation \[ \widehat{\widetilde{M}}_S\in \arg \min_{A\in S}\tilde{r}(A), \] where \[ \tilde{r} (A) = \|A - \widetilde{X}\|^2_F, \quad A \in {\mathcal M}_{d,\tau}(\mathbb R). \] The estimator of \(M\) is defined as \(\widehat{{M}}_S=\widehat{\widetilde{M}}_S\Lambda\). The authors study the statistical performances of this estimator. There is proved that when the series exhibit some additional structure like periodicity or smoothness, it is possible to improve the classical rates of convergence. All the results are first stated under a known structure, that is, we assume that we know the rank \(k\), the period \(\tau\) or the smoothness of the series. A model selection procedure is provided at the end of the paper that allows to obtain the same rates of convergence without assuming this prior knowledge.