High-dimensional dynamic factor models: a selective survey and lines of future research

For all N, f^N_{\chi} is a rational spectral density.

We suppose that there exists N_0\geq q such that, from N_0 onwards, the rank of f^N_{\chi} is independent of N. Such rank is denoted by q.

The state dimension of a minimal, stable and miniphase state space realization corresponding to f_{\chi}^N is independent of N from a certain N_1 onwards. Such state dimension is denoted by n.

Weak cross-sectional dependence of (\xi^N_t). The eigenvalues \omega^N_{\xi,1}(\lambda) are uniformly bounded, i.e. there exists B>0 such that \omega^N_{\xi,1}(\lambda)<B for all N and \lambda.

Strong cross-sectional dependence of (\chi^N_t). The eigenvalues \omega^N_{\chi,s}(\lambda), s=1,\ldots,q, diverge as N\to \infty, \lambda almost everywhere in [-\pi,\ \pi].

We suppose that P_0=I_N and that P(z) and Q(z) are left coprime [deistler12, see, p.41]), i.e. the matrix (P(z)\ Q(z)) has rank N for all z\in \mathbb C.

For q=r, we assume that k(0)=I_N, i.e. Q_0=I_N holds. For q<r, we assume that the top q\times q submatrix of k(0) is (1) non-singular (which is the case generically) and (2) lower triangular (with non-zero diagonal elements).

The r largest eigenvalues of the covariance matrix \gamma_\chi(0) diverge as N\to \infty.

holds.