Sequential Gaussian approximation for nonstationary time series in high dimensions (Q6635728)

Let \(X_1,\ldots,X_n\) be random \(d\)-dimensional vectors with finite \(q\)th moments. The authors study Gaussian approximation of the partial sums \(S_k=\frac{1}{\sqrt{n}}\sum_{t=1}^kX_t\), \(k=1,\ldots,n\), in the case where the dimension \(d\) grows with \(n\). In particular, they construct a sequential coupling of these partial sums with a Gaussian process \(\mathcal{Y}_k\) such that \(\max_{k\leq n}\lVert S_k-\mathcal{Y}_k\rVert=\mathcal{O}_P(\tau_n)\) with \(\tau_n=\sqrt{\log(n)}d^{\frac{3}{4}-\frac{1}{2q}}n^{\frac{1}{2q}-\frac{1}{4}}\) in the case where the \(X_t\) are independent. This allows the dimension to grow at a faster rate than previous results: for large \(q\) we may take \(d=o(n^{\frac{1}{3}-\alpha})\) for some \(\alpha>0\). The authors further extend their high-dimensional approximation result to non-stationary time series, and give statistical applications to a sequential test for the mean and changepoint detection. Finally, a bootstrap procedure for statistical inference is also proposed.