Identifying the multifractional function of a Gaussian process (Q1273015)

A real Gaussian process \(X_t\), \(t\in [0,1]\), is called multifractional if its covariance function has the form \(\Sigma(t,s)= \int_R f(t,\lambda)\overline{f(s,\lambda)} d\lambda\), where \(f(t,\lambda)= (e^{it\lambda}- 1)a(t,\lambda)|\lambda|^{-\alpha(t)- 1/2}\) with \(\alpha(t)\) and \(a(t,\lambda)\) sufficiently smooth. \(\alpha(t)\) is called multifractional function. It is proved that a multifractional Gaussian process behaves locally around a point \(t\) as a fractional Brownian motion with index \(\alpha(t)\). Next, an estimator \(\widehat\alpha(t)\) of \(\alpha(t)\) based on the discrete observations of one sample path is proposed. \(\widehat\alpha(t)\) is expressed in terms of generalized quadratic variations of \(X\). Strong consistency of \(\widehat\alpha(t)\) is proved and asymptotic mean-square error is estimated.