Testing linearity for NARX models (Q1287103)

One considers the discrete-time NARX system \[ X_{n+1}= f(X_n,X_{n-1},\dots,X_{n-p+1}) -k(X_n,\dots,X_{n-p+1}) + \zeta_{n+1}, \] where \(f\) is the unknown function, \(k\) is the known stabilizing function and \(\{\zeta_{n}\}\) is the white noise. The test for linearity of the function \(f\) is proposed based on two estimates of the function \(f\) -- the recursive least squares estimator and the recursive kernel estimator.