The time series considered have jumps in their spectral distribution function; that is, the series is the sum of a 'signal' component, comprising a finite linear sum of pure sine-waves, and a 'noise' component, having continuous spectral density function. Given a set of observations from such a time series the primary problem is to estimate the 'signal' frequencies, the power in each component of the signal, and the 'noise' spectral density at these frequencies. The essence of the method used is as follows. For a given set of observations from such a series, and for each frequency that might yield a signal component, several estimates of the spectral density are made, using spectral windows of different bandwidths. To a first approximation, the noise component of the estimate is the same for every window, while the part of the estimate due to the signal is inversely proportional to the bandwidth of the window. Thus using a regression technique, one can separate the signal power from the noise spectral density at the given frequency and estimate these two quantities. These ideas are developed as follows. After a historical introduction, the early part of the thesis is devoted to the 'probability' aspects of the problem. First some results are proved that apply to the 'noise' series or any stationary time series. They give extensions and refinements of early approximations for the expected value of the spectral estimate, and for the covariance between two spectral estimates; these include the rates at which the limiting values are attained.