The performance of a long-baseline optical stellar interferometer is greatly enhanced if the instantaneous atmospheric delay ?(t ) can be tracked to within a fraction of a wavelength, permitting coherent integration of the optical correlation (fringe visibility). Real-time fringe tracking involves a control system that servos a rapidly responding pathlength compensator in real time. However, precise delay tracking can be achieved at somewhat lower signal levels by employing an off-line delaytracking system, in which the raw data measured by the interferometer are stored for subsequent analysis. Then the estimate of ? at time t is based on data collected both before and after time ? . An optimum delaytracking algorithm embraces the a priori statistics of the atmospheric delay process. Rather than simply estimating t at a point in time, a superior estimate of ? will be obtained by comparing all possible functions ?(t ) over a time period. Using Bayes’s theorem, the a posteriori probability density of any t(t ) function can be determined. An algorithm is developed that determines one or more functions that maximize that probability. Even the ambiguous estimates that result at lower signal levels can be employed for the coherent integration of optical correlation.