Estimates of derivatives of random functions I

If one looks for an optimal (by criterion of minimal variance) linear estimate of s′(t) from the observation of u(t) = s(t) + n(t), where n(t) is noise and s(t) is useful signal, then one can derive an integral equation for the weight function of optimal estimate. This integral equation is often difficult to solve and, even if one can solve it, it is difficult to construct the corresponding filter. In this paper an optimal estimate of s′ on a subset of all linear estimates in sought and it is shown that this quasioptimal estimate is easy to calculate, the corresponding filter is easy to construct, and the error of this estimate differs little from the error of optimal estimates. It is also shown that among all estimates (linear and nonlinear) of s′ for ∥n∥ ? δ and ∥s″∥ ? M the best estimate is given by Δ_hu = (2h)^?1 u(t + h) ? u(t ? h)] with $\text{h = (}\text{2δ}\text{M}\text{)}{}^{\mathrm{<mtext>1</mtext><mtext>2</mtext>}}$.