Moment approach for singular values distribution of a large auto-covariance matrix (Q503082)

Let \((\varepsilon_{t})_{t>0}\) be a sequence of \(p\)-dimensional independent random vectors with real entries and let \(X_{T}=\frac{1}{T} \sum_{t=s+1}^{s+T} \varepsilon_{t} \varepsilon_{t-s}^{*}\) be the lag-\(s\) auto-covariance matrix of \(\varepsilon_{t}\), where \(s\) is a fixed positive integer number. One of the main problems in the theory of random matrices is the analysis of the convergence of the sequence of the empirical spectral distribution (ESD) for a given sequence of symmetric or Hermitian random matrices \(\{A_{n}\}_{n\geq 0}\). Let us remind that the ESD of a Hermitian matrix \(A\) of size \(p\times p\) is defined by \(F^{A}(x)= \frac{1}{p}\sum_{j=1}^{p} \delta_{x_{j}},\) where \(x_{j}\), \(j=1,\dots,p\), are the eigenvalues of \(A\) which appear in the sum according to its multiplicity. The limit distribution of the sequence \(\{F^{A_{n}}(x)\}_{n\geq 0}\), which is usually nonrandom, is called the limiting spectral distribution (LSD) of the sequence \(\{A_{n}\}_{n\geq 0}\). The lag-\(0\) auto-covariance matrix of \(\varepsilon_{t}\) reduces to the standard sample covariance matrix and its property in large dimension is very well known in the literature. Very few results are known for the lag-\(s\) auto-covariance matrix. In such a direction, the case of the LSD for the symmetrized auto-covariance matrix has been studied in [\textit{B. Jin} et al., Ann. Appl. Probab. 24, No. 3, 1199--1225 (2014; Zbl 1296.60006)]. On the other hand, \textit{C. Wang} et al. [Ann. Appl. Probab. 25, No. 6, 3624--3683 (2015; Zbl 1328.60088)] analyzed the problem for its exact separation, which also provides the convergence of its largest eigenvalue. In the general case, since \(X_{T}\) is not symmetric, it is very well known that its singular eigenvalues are the square roots of the \(p\) nonnegative eigenvalues of \(A_{T}= X_{T}X_{T}^{*}\). Using the method of the Stieltjes transform, the LSD of \(A_{T}\) has been obtained in [\textit{Z. Li} et al., J. Multivariate Anal. 137, 119--140 (2015; Zbl 1338.60103)]. In the paper under review, assuming some conditions, the moments of the limiting distribution of the ESD of the matrix \(A_{T}\) are explicitly given (Theorem 3.1). On the other hand, using the moment method, the authors prove the convergence, almost surely, of the largest eigenvalue of the matrix \(A_{T}\) to the right edge of the LSD under an appropriate condition (Theorem 4.1). A new proof of the convergence of ESD of \(A_{T}\) to its LSD is presented. A distinctive feature here is that the matrix above can be considered as the product of four matrices involving \(\varepsilon_{t}\). A new methodology is required with respect to the existing literature on moment method in random matrix theory. Some complex recursion formulas related to enumeration of a particular family of ``walk paths'' on nonnegative integer numbers are provided.