% SIGNAL = sinusoids + noise % Setting up the signal parameters and time history N=100; a0=1.8; a1=1.5; theta1=1.3; a2=2; theta2=1.35; k=1:N; k=k(:); y=a0*randn(N,1)+a1*exp(1i*(theta1*k+2*pi*rand))+a2*exp(1i*(theta2*k+2*pi*rand));
IntroductionConsider a discretetime zeromean stationary process , , and let represents power density over frequency. Under suitable smoothness conditions, it is also Spectral estimation refers to esimating when only a finitesized observation record of the time series is available. Different schools of thoughts have evolved over the years based on varying assumptions and formalisms. Classical methods began first based on Fourier transform techniques and the periodogram, followed by the so called modern spectral estimation methods such as the Burg method, MUSIC, ESPRIT, etc.. The mathematics relates to the theory of the Szeg orthogonal polynomials and the stucture od Toeplitz covariances. In contrast, recently, the analysis of state covariance matrices, see e.g. [1, 2], suggested a general framework which alows even higher resolution. A series of generalized spectral estimation tools have been developed generalizing Burg, Capon, MUSIC, ESPRIT, etc.. In the following, we will overview some of these high resolution methods and the relevant computer software. Their use is shown via a representative academic examples. We want to resolve two sinusoids in based on noisy measurements of the time series where and . These are closer than the theoretical resolution limit of Fourier method. % SIGNAL = sinusoids + noise % Setting up the signal parameters and time history N=100; a0=1.8; a1=1.5; theta1=1.3; a2=2; theta2=1.35; k=1:N; k=k(:); y=a0*randn(N,1)+a1*exp(1i*(theta1*k+2*pi*rand))+a2*exp(1i*(theta2*k+2*pi*rand));
