Sensitivity analysis of solutions of the harmonic inversion problem: are all data points created equal?

We consider the harmonic inversion problem, and the associated spectral estimation problem, both of which are key numerical problems in NMR data analysis. Under certain conditions (in particular, in exact arithmetic) these problems have unique solutions. Therefore, these solutions must not depend on the inversion algorithm, as long as it is exact in principle. In this paper, we are not concerned with the algorithmic aspects of harmonic inversion, but rather with the sensitivity of the solutions of the problem to perturbations of the time-domain data. A sensitivity analysis was performed and the counterintuitive results call into question the common assumption made in "super-resolution" methods using non-uniform data sampling, namely, that the data should be sampled more often where the time signal has the largest signal-to-noise ratio. The numerical analysis herein demonstrates that the spectral parameters (such as the peak positions and amplitudes) resulting from the solution of the harmonic inversion problem are least susceptible to the perturbations in the values of data points at the edges of the time interval and most susceptible to the perturbations in the values at intermediate times.

The extended Fourier transform for 2D spectral estimation.

We present a linear algebraic method, named the eXtended Fourier Transform (XFT), for spectral estimation from truncated time signals. The method is a hybrid of the discrete Fourier transform (DFT) and the regularized resolvent transform (RRT) (J. Chen et al., J. Magn. Reson. 147, 129-137 (2000)). Namely, it estimates the remainder of a finite DFT by RRT. The RRT estimation corresponds to solution of an ill-conditioned problem, which requires regularization. The regularization depends on a parameter, q, that essentially controls the resolution. By varying q from 0 to infinity one can "tune" the spectrum between a high-resolution spectral estimate and the finite DFT. The optimal value of q is chosen according to how well the data fits the form of a sum of complex sinusoids and, in particular, the signal-to-noise ratio. Both 1D and 2D XFT are presented with applications to experimental NMR signals.

Pseudotime Schrödinger equation with absorbing potential for quantum scattering calculations.

The Schrödinger equation (Hpsi) (r) = [E+u(E)W(r)]psi(r) with an energy-dependent complex absorbing potential -u(E)W(r), associated with a scattering system, can be reduced for a special choice of u(E) to a harmonic inversion problem of a discrete pseudotime correlation function y(t) = phi(T)U(t)phi. An efficient formula for Green's function matrix elements is also derived. Since the exact propagation up to time 2t can be done with only approximately t real matrix-vector products, this gives an unprecedently efficient scheme for accurate calculations of quantum spectra for possibly very large systems.

The multidimensional filter diagonalization method.

The theory and numerical aspects of the recently developed multidimensional version of the filter diagonalization method (FDM) are described in detail. FDM can construct various "ersatz" or "hybrid" spectra from multidimensional time signals. Spectral resolution is not limited by the time-frequency uncertainty principle in each separate frequency dimension, but rather by the total joint information content of the signal, i.e., N(total) = N(1) x N(2) x vertical ellipsis x N(D), where some of the interferometric dimensions do not have to be represented by more than a few (e.g., two) time increments. It is shown that FDM can be used to compute various reduced-dimensionality projections of a high-dimensional spectrum directly, i.e., avoiding construction of the latter. A subsequent paper (J. Magn. Reson. 144, 357-366 (2000)) is concerned with applications of the method to 2D, 3D, and 4D NMR experiments.

The multidimensional filter diagonalization method.

The theory of the multidimensional filter diagonalization method (FDM) described in the previous paper (V. A. Mandelshtam, 2000, J. Magn. Reson. 144, 343-356 (2000)) is applied to NMR time signals with up to four independent time variables. Direct projections of the multidimensional time signals produce new kinds of 2D spectra. The resolution obtained by FDM can be far superior to that obtained by conventional phase-sensitive FT processing, and correlation peaks in heteronuclear and homonuclear experiments can be condensed to sharp singlets, removing all spin-spin couplings. Examples of singlet-HSQC and singlet-TOCSY spectra show big gains in resolution. It is not necessary to have a finely digitized spectrum, in which the individual multiplet components are resolved, for the methods to work. Examples of FDM spectra, ranging from simple organic molecules and steroids to metalloproteins, are shown.

Pure absorption electron spin echo envelope modulation spectra by using the filter-diagonalization method for harmonic inversion.

Harmonic inversion of electron spin echo envelope (ESEEM) time-domain signals by filter diagonalization is investigated as an alternative to Fourier transformation. It is demonstrated that this method features enhanced resolution compared to Fourier-transform magnitude spectra, since it can eliminate dispersive contributions to the line shape, even if no linear phase correction is possible. Furthermore, instrumental artifacts can be easily removed from the spectra if they are narrow either in time or frequency domain. This applies to echo crossings that are only incompletely eliminated by phase cycling and to spurious spectrometer frequencies, respectively. The method is computationally efficient and numerically stable and does not require extensive parameter adjustments or advance knowledge of the number of spectral lines. Experiments on gamma-irradiated methyl-alpha-d-glucopyranoside show that more information can be obtained from typical ESEEM time-domain signals by filter-diagonalization than by Fourier transformation.

Semiclassical spectra and diagonal matrix elements by harmonic inversion of cross-correlated periodic orbit sums.

Semiclassical spectra weighted with products of diagonal matrix elements of operators A(alpha), i.e., g(alphaalpha')(E)= summation operator(n)/(E-E(n)), are obtained by harmonic inversion of a cross-correlation signal constructed of classical periodic orbits. The method provides highly resolved semiclassical spectra even in situations of nearly degenerate states, and opens the way to reducing the required signal lengths to shorter than the Heisenberg time. This implies a significant reduction of the number of orbits required for periodic orbit quantization by harmonic inversion.

Reference deconvolution, phase correction, and line listing of NMR spectra by the 1D filter diagonalization method.

We describe a new way to attack the problem of identifying and quantifying the number of NMR transitions in a given NMR spectrum. The goal is to reduce the spectrum to a tabular line list of peak positions, widths, amplitudes, and phases, and to have this line list be of high fidelity. In this context "high fidelity" means that each true NMR transition is represented by a single entry, with no spurious entries and no missed peaks. A high fidelity line list allows the measurement of chemical shifts and coupling constants with good accuracy and precision and is the ultimate in data compression. There are two parts to the problem. The first is to overcome common imperfections: the non-Lorentzian lineshapes that can arise whenever the magnetic field inhomogeneity is less than perfect, and nonzero time delays that cause frequency-dependent phase errors. The second is to fit the spectral features to a model of Lorentzian lines. We use the recently developed filter diagonalization method (FDM) to accomplish the reference deconvolution, the phase correction, and the fitting, and show good progress toward the goal of obtaining a high fidelity line list.