In this thesis, the method of "classification with empirically observed statistics"--also known as empirical classification, empirical detection, universal classification, and type-based detection--is configured and applied to the despreading/detection receiver operation of a spread-spectrum (SS), code division multiple access (CDMA) communications system. In static and Rayleigh-fading environments, the empirical detector is capable of adapting to unknown noise environments in a superior manner than the linear matched-filter despreader/detector, and done with reasonable amounts of training. Compared to the optimum detector, when known, the empirical detector always approaches optimal performance, again, with reasonable amounts of training. In an interference-limited channel, we show that the single-user likelihood-ratio detector, which is the optimum single-user detector, can greatly outperform the matched filter in certain imperfect power-control situations. The near optimality of the empirical detector implies that it, too, will outperform the matched filter in these situations. Although the empirical detector has the added cost of requiring chip-based phase synchronization, its consistent and superior performance in all environments strongly suggests its application in lieu of the linear detector for SS/CDMA systems employing long, pseudo-random spreading. In order to apply empirical classification to digital communications, we derive the empirical forced-decision detector and show that it is asymptotically optimal over a large class of empirical classifiers.