A note on estimation by least squares for harmonic component models (Q4455675)

Let \(X_t=f_t(\vartheta)+\varepsilon_t\), \(t=1,\dots,n\), where \(\varepsilon_t\) is a stationary process with zero mean, \[ f_t(\vartheta)=\sum_{j=1}^k (A_j\cos\omega_j t+B_j\sin\omega_jt), \quad \omega_1>\dots>\omega_k,\quad \vartheta=(A_j,B_j,\omega_j)_{j=1}^k. \] The least squares estimator \((\widehat\omega_j)_{j=1}^k\) for \((\omega_j)_{j=1}^k\) is the minimizer of \(\min_{A_j,B_j}\sum_{t=1}^n (X_t-f_t(\vartheta))^2\). It is shown that if \(K_n/n\leq \omega_1,\dots,\omega_k\leq\pi-K_n/n\) and \(| \omega_j-\omega_i| \geq K_n/n\) \((j\not=i)\) for some \(K_n\) such that \(K_n/n\to 0\), \(K_n\to\infty\) for \(n\to\infty\), then \(n(\widehat\omega_i-\omega_i)\to 0\) in probability. For the one component model (\(k=1\)) no restrictions on \(\omega_1\) are needed. For the two-component model the author considers many particular cases and concludes that ``It seems certain that we obtain consistency of \((\widehat\omega_1,\widehat\omega_2)\) when \(\omega_i\) are not restricted to be different from 0 and \(\pi\) by a finite multiple of \(1/n\). We could undoubtedly establish this, but the number of special cases to be considered is so large to be quite unmanagable''.