Incoherence of Partial-Component Sampling in multidimensional NMR

Hatef Monajemi, David L. Donoho, Jeffrey C. Hoch, Adam D. SchuylerarXiv 2017 · DOI · arXiv

Abstract

In NMR spectroscopy, undersampling in the indirect dimensions causes reconstruction artifacts whose size can be bounded using the so-called coherence. In experiments with multiple indirect dimensions, new undersampling approaches were recently proposed: random phase detection (RPD) and its generalization, partial component sampling (PCS). The new approaches are fully aware of the fact that high-dimensional experiments generate hypercomplex-valued free induction decays; they randomly acquire only certain low-dimensional components of each high-dimensional hypercomplex entry. We provide a classification of various hypercomplex-aware undersampling schemes, and define a hypercomplex-aware coherence appropriate for such undersampling schemes; we then use it to quantify undersampling artifacts of RPD and various PCS schemes.

Introduction

Multidimensional NMR spectroscopy is a cornerstone of structural biology, yet its utility is often limited by the time required to sample indirect dimensions parametrically. To mitigate this, Non-Uniform Sampling (NUS) has become standard, allowing researchers to skip data points and reduce acquisition time from weeks to days. However, this efficiency comes at a cost: undersampling introduces reconstruction artifacts that can obscure real signals.

Traditionally, the severity of these artifacts is quantified by "coherence." While effective for standard NUS, traditional coherence metrics fail to account for the hypercomplex nature of modern sampling schemes like Random Phase Detection (RPD) and Partial Component Sampling (PCS). These methods do not merely skip time points (indels); they sample specific components of the hypercomplex data at those points. This paper introduces a hypercomplex-aware coherence metric to bridge this gap.

Classification of Hypercomplex Sampling Schemes

The authors categorize sampling strategies based on how they handle two distinct dimensions: the "indel" (the time-domain grid point) and the "component" (the hypercomplex part of the signal).

Hierarchy of hypercomplex-aware sampling schedules.

As illustrated in the hierarchy above, schemes range from traditional Uniform Sampling (US) to complex Partial Component Sampling (PCS). In PCS, the sampling schedule is randomized across both indels and components, offering a more granular approach to data reduction than standard NUS, which typically samples full components at a subset of indels.

Methodology: Hypercomplex Algebra and Coherence

To analyze these schemes, the authors utilize a $2^d$-dimensional commutative hypercomplex algebra ($H_d$) to model phase-sensitive detection in $d$ dimensions. By developing a matrix isomorphism that maps hypercomplex multiplications to real-valued matrix-vector products, the researchers can apply compressed sensing theory to hypercomplex data.

The core contribution is the Hypercomplex Coherence ($\mu_H$). This metric accounts for the singular values of the inner products between hypercomplex sampling columns. The authors prove a recovery theorem: a $k$-sparse hypercomplex spectrum can be uniquely recovered via $\ell_1$-minimization if the number of spikes $k$ satisfies the bound:

$$k < \frac{1 + \mu_H^{-1}}{2}$$

Results and Discussion

Numerical experiments on simulated 3D NMR grids (ranging from $20^3$ to $200^2$) validated the theoretical framework. Monte Carlo simulations revealed that randomization is the primary defense against artifacts. Specifically, spreading randomness across both indels and components (as in PCS) significantly lowers coherence compared to deterministic schemes.

Visualization of a Partial Component Sampling (PCS) scheme.

The study also identified scaling laws for coherence. For a finite number of samples $N$, coherence $\mu(N)$ decays following $N^{-(d-k)/2}$, where $d$ is the dimension and $k$ represents non-randomized dimensions. Interestingly, while PCS offers more flexibility, it did not show a massive synergistic advantage over NUS in all regimes, suggesting that the benefit of PCS lies in its ability to provide a more tunable "knob" for optimization rather than a fundamental shift in artifact suppression.

Conclusion and Future Work

This work provides a rigorous mathematical foundation for evaluating modern NMR sampling strategies. By accounting for hypercomplex structures, the $\mu_H$ metric allows for a more accurate prediction of reconstruction quality.

Future research must address the limitations of this idealized model, particularly the effect of Free Induction Decay (FID) envelopes, which introduce exponential signal decay not captured by simple spike models. Additionally, the computational overhead of the matrix isomorphism approach for 4D and 5D datasets remains a challenge for practical implementation in high-throughput environments.

Summary updated

Abstract

In NMR spectroscopy, undersampling in the indirect dimensions causes reconstruction artifacts whose size can be bounded using the so-called {\it coherence}. In experiments with multiple indirect dimensions, new undersampling approaches were recently proposed: random phase detection (RPD) \cite{Maciejewski11} and its generalization, partial component sampling (PCS) \cite{Schuyler13}. The new approaches are fully aware of the fact that high-dimensional experiments generate hypercomplex-valued free induction decays; they randomly acquire only certain low-dimensional components of each high-dimensional hypercomplex entry. We provide a classification of various hypercomplex-aware undersampling schemes, and define a hypercomplex-aware coherence appropriate for such undersampling schemes; we then use it to quantify undersampling artifacts of RPD and various PCS schemes.
BibTeX
@misc{monajemi2017incoherence,
  author = {Hatef Monajemi and David L. Donoho and Jeffrey C. Hoch and Adam D. Schuyler},
  title = {{Incoherence of Partial-Component Sampling in multidimensional NMR}},
  year = {2017},
  eprint = {1702.01830},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1702.01830},
}
Figure