NMR spectra are usually obtained from free-induction-decay signals by means of the computationally efficient fast Fourier transform. However, in spectral analysis by FFT, there are several inherent limitations such as low resolution for the truncated signal, or side lobes due to windowing. Recently, several least-squares alternative methods to the FFT have been proposed to reduce these problems ( J-IO). Barkhuijsen et al. applied a linear prediction (LP) method based on singular value decomposition (SVD) to FIDs (LPSVD) (1). The LPSVD method has many advantages as described in the following paragraph, although the SVD computation is very time consuming. First, no preprocessing such as windowing or zero filling is needed. Second, the procedure based on singular value decomposition enhances the numerical stability and enables one to distinguish signals from the noise. Third, the LPSVD method is a parametrical approach as it directly estimates the entire set, including frequency, damping factor, amplitude, and the phase of each component in the FID. Therefore, this approach is fundamentally different from the nonparametrical approach used in the maximum entropy method, in which the spectrum is directly estimated. This point is very important in quantitative analysis. As parametrical approaches need no complicated procedures to determine line intensities or line positions of spectral peaks, one can perform quantitative analysis more easily. A more efficient method than the LPSVD has been applied by Tirendi et al., by using the total least-squares (TLS) procedure (2). The difference between the LPSVD and TLS methods is that the TLS takes the effects of the noise in both the LP data matrix and the observation vector into consideration while the LPSVD considers only the LP data matrix (2, 11). This assumption is more realistic. Moreover, Cadzow et al. proposed using the autocorrelation-like matrix in the TLS procedure instead of the LP data matrix ( 7, 12). They insisted that the improvement in the performance made possible by using this approach in the noisy case warranted this modification. We compared these parametrical LP methods based on singular value decomposition and investigated the features and the efficiency of each method. First, we assume that the FID signal comprises K exponentially damped complex sinusoids plus a complex Gaussian white noise sequence W(n), sampled at regular times n At. The data sequence x(n) is given by