A comparison of the least squares method and the Burg method for autoregressive spectral analysis

The performance of autoregressive (AR) spectral estimates based on two different methods for computing the AR coefficients are compared. They are the recursive method as stated by Burg, which minimizes the residual power with respect to only one coefficient, and the straightforward but computationally less efficient least squares method (LSM) which minimizes the residual power with respect to all the AR coefficients simultaneously. It is shown that when the input signal consists of two equal-leveled sinusoids in white noise, the LSM estimate is highly superior with respect to resolution, positional bias, and spurious peaks in the spectrum.