“Ersatz” and “Hybrid” NMR Spectral Estimates Using the Filter Diagonalization Method

Phase-sensitive spectral estimation by the hybrid filter diagonalization method

A more robust way to obtain a high-resolution multidimensional NMR spectrum from limited data sets is described. The Filter Diagonalization Method (FDM) is used to analyze phase-modulated data and cast the spectrum in terms of phase-sensitive Lorentzian "phase-twist" peaks. These spectra are then used to obtain absorption-mode phase-sensitive spectra. In contrast to earlier implementations of multidimensional FDM, the absolute phase of the data need not be known beforehand, and linear phase corrections in each frequency dimension are possible, if they are required. Regularization is employed to improve the conditioning of the linear algebra problems that must be solved to obtain the spectral estimate. While regularization smoothes away noise and small peaks, a hybrid method allows the true noise floor to be correctly represented in the final result. Line shape transformation to a Gaussian-like shape improves the clarity of the spectra, and is achieved by a conventional Lorentzian-to-Gaussian transformation in the time-domain, after inverse Fourier transformation of the FDM spectra. The results obtained highlight the danger of not using proper phase-sensitive line shapes in the spectral estimate. The advantages of the new method for the spectral estimate are the following: (i) the spectrum can be phased by conventional means after it is obtained; (ii) there is a true and accurate noise floor; and (iii) there is some indication of the quality of fit in each local region of the spectrum. The method is illustrated with 2D NMR data for the first time, but is applicable to n-dimensional data without any restriction on the number of time/frequency dimensions.

Filter diagonalization using a "sensitivity-enhanced basis": improved performance for noisy NMR spectra

The Filter Diagonalization Method (FDM) has been used to process NMR data in liquids and can be advantageous when the spectrum is sparse enough, the lines are sharp and Lorentzian, raw sensitivity is adequate, and the measured time-domain data is short, so that the Fourier Transform spectrum exhibits distorted line shapes. Noise can adversely impact resolution and/or frequency accuracy in FDM spectral estimates. Paradoxically, more complete data can lead to worse FDM spectra if there is appreciable noise. However, by modifying the numerical method, the FDM noise performance improves significantly, without apparently losing any of the existing advantages. The two key modifications are to adjust the FDM basis functions so that matrix elements computed from them have less noise contribution on average, and to regularize each dimension of a multidimensional spectrum independently. The modifications can be recommended for general-purpose use in the case of somewhat noisy, incomplete data.