Asymptotics of regressions with stationary and nonstationary residuals. (Q2574563)

The paper deals with the discrete-time autoregressive process \(y_t=\beta ^{\text{T}}_0x_t+e_t\), \(t=1,2,\dots \), with nonstochastic \(d\)-dimensional regressors \(x_t\) and residuals \(e_t\) fulfilling \(e_1=\eta _1\), \(e_{t+1}=\rho _0e_t+\eta _{t+1}\), where \(\eta _t\), \(t=1,2,\dots \), is a sequence of i.i.d. random variables, \(E\,\eta _t=0\), \(\text{Var}\,\eta _t=\sigma _0^2>0\). The pseudo maximum likelihood estimators \(\widehat \theta \) and \(\widehat \sigma \) of the parameter vector \(\theta _0=(\beta _0,\rho _0)\) and the variance \(\sigma _0^2\), respectively, are studied in all possible cases to which different values of the autocorrelation coefficient \(\rho _0\) can lead, i.e., stationary, unit root, and explosive. Three main results are obtained. First, the consistency of simultaneously estimating \((\theta _0,\sigma _0^2)\) with \((\widehat \theta ,\widehat \sigma )\) is established. Second, for \(\widehat \theta \) normalized by means of the sample information matrix and \(\widehat \sigma \), the limiting distribution is found, implying in particular the asymptotic independence of all \(d+1\) components of the normalized estimator, the asymptotic normality of the components corresponding to \(\beta _0\), and the fact that also the component corresponding to \(\rho _0\) has a proper, nondegenerate limiting distribution, which is normal in the stationary case. Finally, the limiting distribution of the likelihood ratio statistics is derived for the tests of four important hypotheses: \(\rho _0=1\), \(\rho _0=0\), \(\beta _0=0\), and \(\theta _0=(\beta _0,\rho _0)=(0,1)\). The results are obtained assuming the parameter-dependent uniform asymptotic negligence of the sequence of regressors, although a bit stronger, parameter-independent assumptions are initially stated, inspired by \textit{L. Fahrmeir} and \textit{H. Kaufmann} [Ann. Stat. 13, 342--368 (1985; Zbl 0594.62058)]. Their proofs rely on the classical machinery of martingale triangular arrays, random walk and Donsker's principle. They are clearly structured and with the exception of several typos, which occur especially in the proof of the second result, easily understandable.