The power spectrum of truncated variables is developed as the transform of the correlation function of data smoothed by an arbitrary mean, as a function of both lag and averaging intervals. It is shown that a large class of techniques for computing spectra can be readily derived as special cases of this more general approach. These include the conventional approach of Taylor as well as various methods in current use. The need for further study of the numerical effects of the approximations used in applying these techniques to finite records of discrete data is indicated.