Linear regression with delta-modulated integrated signals (Q2266550)

Let \(f(t)=\alpha +\beta t+e(t)\) where e is a stationary random process with autocorrelation \(\rho\) (t). Suppose \(g(t)=\Sigma_ i\eta_ i\delta (t-t_ i)\) is a pulse readout of an instrument that performs differential pulse code modulation on the time integral of f; i.e., on any interval [\(\lambda\),\(\mu\) ] \(| \int^{\mu}_{\lambda}[f- g]dt| \leq \eta\), and \(| \eta_ i| =\eta.\) It is demonstrated that using g(t) to estimate \(\alpha\), \(\beta\) is very treacherous if \(\beta\) is small. Approximations to unbiased estimators for these parameters are given with errors of order O(\(\eta\) /T) and \(O(\eta /T^ 2)\), respectively, where T is the time interval used for the estimates. Variances are also derived. For the determinate case the theory gives accuracy estimates for expansion in Legendre polynomials.