Reaching the sparse-sampling limit for reconstructing a single peak in a 2D NMR spectrum using iterated maps

Abstract

Many of the ubiquitous experiments of biomolecular NMR, including R₂, R_1ρ, and CEST (Palmer et al, 2001; Loria et al, 1999; Neudecker et al, 2009; Palmer and Massi, 2006; van Zijl and Yadav, 2011; Walinda et al, 2018), involve acquiring repeated 2D spectra under slightly different conditions. Such experiments are amenable to acceleration using non-uniform sampling (NUS) spectral reconstruction methods that take advantage of prior information. We previously developed one such technique, an iterated maps method (DiffMap) that we successfully applied to 2D NMR spectra, including R₂ relaxation dispersion data (Frey et al, 2013). In that prior work, we took a top-down approach to reconstructing the 2D spectrum with a minimal number of sparse samples, reaching an undersampling fraction that appeared to leave some room for improvement. In this study, we develop an in-depth understanding of the action of the DiffMap algorithm, identifying the factors that cause reconstruction errors for different undersampling fractions. This improved understanding allows us to formulate a bottom-up approach to finding the lowest number of sparse samples required to accurately reconstruct individual spectral features with DiffMap. We also discuss the difficulty of extending this method to reconstructing many peaks at once, and suggest a way forward.

Introduction

NMR experiments aiming to study molecular structure and dynamics will often require the acquisition of large, multidimensional data sets. While these nD experiments (n>1) can reveal more information than a 1D experiment, they also take much longer, with durations proportional to the number of sampled points in the n-1 indirect dimension(s). In order to use the discrete Fourier transform to generate the spectrum, the points are “densely-sampled,” i.e., measuring all time-domain points lying on a Cartesian grid of uniform intervals consistent with the Nyquist-Shannon sampling theorem.

Sparse sampling, also known as non-uniform sampling (NUS), is a common approach to speeding up the acquisition of nD NMR data sets. The speed-up comes from skipping over particular indirect-dimension data points (saving entire 1D scans); the unmeasured points are typically set to zero to fill out the nD grid. Some alternative, non-Fourier processing algorithm must then be used to reconstruct the spectrum. Examples of such reconstruction approaches include maximum entropy (MaxEnt) (Hoch and Stern, 1996; Mobli and Hoch, 2008; Hyberts et al, 2010; Paramasivam et al, 2012), iterative soft thresholding (IST) (Stern et al, 2007), iteratively reweighted least squares (IRLS) (Schlossmacher, 1973; Kazimierczuk and Orekhov, 2012), multidimensional decomposition (MDD) (Orekhov and Jaravine, 2011), and compressed sensing (CS) (Kazimierczuk and Orekhov, 2011; Bostock and Nietlispach, 2018), among many others (Stanek and Koźmiński, 2010; Ying et al, 2017; Billeter, 2017; Rovnyak and Schuyler, 2018).

Motivated by the need to accelerate 3D MRI of solids experiments, we recently developed one such alternative reconstruction method, the so-called iterated maps approach, which is based upon Elser’s “Difference Map” algorithm (Fienup, 1982; Elser, 2003; Elser et al, 2007; Frey et al, 2013). In that paper, the iterated maps method (or “DiffMap” for short) was successfully applied to existing 2D NMR spectra (both liquid state and MAS of solids) and 3D MR images of solids. Our iterated maps approach is related to SIFT (Matsuki et al, 2009, 2010, 2011), as well as to POCS (Haacke et al, 1991). All of these methods take advantage of the fact that in many NMR/MRI experiments, the user knows where spectral features “should” and “should not” appear, even if the precise details of the spectrum (e.g., point-by-point amplitude) remain to be determined. This information can be exploited in a deterministic algorithm that nudges a sparsely-sampled data set towards its densely-sampled limit.

Many of the ubiquitous experiments of biomolecular NMR, including R₂,R_1ρ, and CEST experiments (Palmer et al, 2001; Loria et al, 1999; Neudecker et al, 2009; Palmer and Massi, 2006; van Zijl and Yadav, 2011; Walinda et al, 2018), are ideal targets for the iterated maps approach, since the series of ~ 20 2D spectra that are typically collected tend to be very similar to each other. Moreover, even if a few slices are densely sampled, sparsely sampling the rest will greatly compress the total duration of the experiment.

In our earlier work, one of the 2D NMR examples was the τ_cp = 0 slice of a larger, pseudo-3D data set, acquired as part of an R₂ dispersion experiment. For that τ_cp = 0 slice, we were able to lower the number of sampled indirect time points $N_{t_1}$ all the way down to an undersampling fraction of 58.6% of the dense data ($N_{t_1}/N_{t_1}^{\text{dense}}=75/128$), while still obtaining a high-quality spectrum. As we will explain below, our approach to making the spectrum entirely real causes M, the number of f₁ rows, to be given by $M=2N_{t_1}^{\text{dense}}=256$. So, defining L = 2 × 75 = 150, we had clearly pushed L well below M (Fig. 1). While this was a significant advance, we did not have a deep understanding of why the method broke down at L. In fact, there seemed to be additional room for improvement, since counting the number of significant points in each f₂ column n yielded even smaller values K(n)(i.e., M > L > K (n), see Fig. 1). Could we find a way to push $2N_{t_1}$ below L, and perhaps all the way down to ~ K(n), without hurting the quality of the reconstruction? In other words, what are the fundamental limits of this approach?

Fig. 1.

Background and motivation of this work. a Contour plot showing the central part of a M × N = 256 × 4096 2D spectrum from the pseudo-3D data set, with the mask(f) = 0 regions in gray. Axes show the frequency scaling (left and top) as well as the indices n = {0, 1, …, 4095} (bottom) and m = {0, 1, …, 255} (right). b Previously achieved and predicted limits to sparse sampling for the data set shown in a. Dense sampling corresponds to acquiring all M = 256 complex points (dashed line), or 128 complex t₁ ≥ 0 values. In prior work (Frey et al, 2013), we were able to push a global undersampling scheme down to L = 150 (dot-dashed line). However, for each f₂ column of the spectrum, the number of “significant” mask(f) ≠ 0 points K(n) (solid line) is less than L, indicating that even better results may be possible. In this paper, we attempt to push our sampling fraction down towards K(n), by focusing on reconstructing individual peaks in individual columns

To answer this question, we first needed to develop an improved understanding of how the iterated maps reconstruction method breaks down as $N_{t_1}$ is lowered, which included understanding the effects of multiple peaks, spectral tails, noise, and other artifacts. The behavior is very complicated across the whole 2D spectrum. To make progress, we decided to start with a simpler problem, by focusing on understanding the reconstruction of a single peak in a single f₂ column. We then set out to answer two main questions:

• (Q.I)For a high-quality reconstruction of a single peak p in a single f₂ column, what is the minimum number of sparsely-sampled data points, $N_{t_1}^p$?
• (Q.II)If we instead wish to reconstruct many (or all) peaks at once, what is the new minimum number of sparsely-sampled data points, $N_{t_1}^{*}$?

Answering question 1 will take us through most of the paper. As we will show, we can go very close to K(n)/2 for most peaks, because we can predict what values of $N_{t_1}$ are susceptible to large errors, as well as the likely range of the “residual” small errors. Question 2 will be addressed briefly at the end, where we show that it is hard to beat L/2 once the user tries to reconstruct many peaks simultaneously. A technique to get past this problem for certain types of pseudo-3D experiment is introduced in a companion paper (Rovny et al, 2019), which we will briefly discuss further at the very end of this paper.

Background

The NMR data set used to test our method

In this study of the DiffMap, we focus on reconstructing 2D slices drawn from an existing, pseudo-3D data set, which was previously studied in Frey et al (2013). This particular pseudo-3D data set is a series of twelve liquid-state 2D NMR spectra, from a ¹³CH₃ multiple-quantum CPMG relaxation dispersion experiment of an Isoleucine, Leucine, Valine ²H, ¹³C-methyl labeled sample of imidazole glycerol phosphate synthase (IGPS) from T. maritima (Lipchock and Loria, 2010; Lisi et al, 2018). The data was collected at 14.1 T and 30 °C with 120 t₁ increments and 1889 t₂ increments, using the sequence described by Korzhnev et al (2004). We zero-fill the data in both dimensions to the next largest power of 2, giving us 128 t₁ points and 2048 t₂ points, which we treat as the set of “measured” points to simplify the reconstruction problem. The ¹H carrier was centered at 0.75 ppm (not 4.7 ppm as incorrectly stated in Frey et al. (2013)) with a spectral width of 8500 Hz, while the ¹³C carrier was centered at 19.5 ppm with a spectral width of 3200 Hz. The twelve spectra correspond to τ_cp values of 0, 0.4167, 0.5, 0.625, 0.7682,1.0, 1.4286, 2.0, 2.5, 3.333, 5.0, and 10.0 ms. We will refer to these below by their respective indices $i_{{\tau }_{\text{cp}}}=1,2,\dots ,12$. The τ_cp = 0 ms slice ($i_{{\tau }_{\text{cp}}}=1$) of this data set was the subject of Figs. 2 & 3 of our earlier paper (Frey et al, 2013). We subsequently noticed that the spectral tails of that $i_{{\tau }_{\text{cp}}}=1$ slice are wider than all of the others, so we set that slice aside in the current analysis. Instead, we will assume the τ_cp = 0.4167 ms slice ($i_{{\tau }_{\text{cp}}}=2$) exists as a dense data set to make our ${\widehat{P}}_1$ mask, which will then be used to reconstruct sparsely-sampled versions of that slice and the remaining ten slices, $i_{{\tau }_{\text{cp}}}=2,3,4,\dots ,12$.

Iterated Maps Reconstruction (DiffMap)

The general approach to spectral reconstruction using iterated maps has been explained previously (Frey et al, 2013). We quickly review how the method is applied to the 2D NMR data studied here, while noting a few differences from the prior work.

We define the complex vector T(t) to be the densely-sampled data set in the 2D time-domain(t = t₁, t₂)), where t₁ and t₂ are the indirect and direct dimensions, respectively. We include the argument t to clarify that T is the time-domain data, and will use T(t) to refer to the value of T at coordinate t. As acquired, a densely sampled “States”-like data set is typically used to fill only the first quadrant of the 2D plane, with (t₁ ≥ 0, t₂ ≥ 0). As in Frey et al (2013), we instead use the same “States”-like data to fill all four quadrants of the time domain, such that the resulting M × N grid of complex data points has Hermitian symmetry about the origin (i.e., T (t)= T^*(−t)) (Frey et al, 2013; States et al, 1982; Mayzel et al, 2014). This ensures that the target dense spectrum, $\tilde{T}(f)= Ph(FT [T(t)])$, in the 2D frequency-domain (f = (f₁, f₂)) is purely real following a 2D complex discrete Fourier transform FT and a known phase correction Ph (see Supplementary Information). As explained below, forcing the spectrum’s phase to be real strengthens the ${\widehat{P}}_1$ mask used in our iterated maps approach, but the sharper spectral features that result from this initial step will also help with other approaches (Zhu and Bax, 1990; Mayzel et al, 2014; Ying et al, 2017). For all of the 2D spectra in this paper, M × N = 256 × 4098, where $M=2N_{t1}^{\text{dense}}$ and $N_{t_1}^{\text{dense}}=128$. In addition to specifying a point in the spectrum by the vector of frequencies f = (f₁, f₂), we can also use the corresponding indices m and n, which go from 0 to M − 1 and N − 1, respectively.

To simulate the effect of skipping particular experiments (due to NUS of the indirect dimension in a “States”-like experiment, we construct an initial “sparsely-sampled” data set $S^0(t)={\widehat{P}}_{0}T(t)$ where ${\widehat{P}}_0$ is a projection that sets to zero all skipped points, leaving the rest alone (see Supplementary Information). Thus, for the t₁ ≥ 0 rows of S⁰ (t), the number containing data is $N_{t_1}$_,, and the number filled with zeros is $N_{t_1}^{\text{dense}}-N_{t_1}$, where $N_{t_1}\le N_{t_1}^{\text{dense}}$. Moreover, the pattern of zero-filled rows is symmetric about t₁ = 0. For a given $N_{t_1}$, the choice of which particular t₁ rows are “measured” (i.e., left alone by ${\widehat{P}}_0$) is specified by the NUS sampling schedule. Many of the popular NUS schedules, such as Poisson Gap sampling or exponentially-biased sampling, sample t₁ rows more densely near t₁ = 0. However, since the iterated maps method appears to work better when the gaps between sampled points are kept to a minimum size, we prefer to use NUS schedules that spread the sparsely-sampled points out as uniformly as possible across the densely-sampled grid. In Frey et al (2013), we used a QUEST schedule (“QUasi-Even Sampling, plus jiTter”). In this paper, we turn off the small random jitter, and use deterministic quasi-even sampling, for reasons we explain below (see Reconstruction error source I: signal inside the mask).

As $N_{t_1}$ is lowered below $N_{t_1}^{\text{dense}}$, the resulting sparse spectrum ${\tilde{S}}^0(f)= Ph \left(\text{FT}\left[S^0(t)\right]\right)$ becomes a poor approximation of dense spectrum $\tilde{T}(f)$, with pronounced sparse-sampling artifacts that smear along the f₂ columns (see Supplementary Information). To do better than this, we use iterated maps to deterministically convert ${\tilde{S}}^0(f)\to {\tilde{F}}^i(f)\approx \tilde{T}(f)$.

The iterated method uses two projection operators, one in the frequency domain and one in the time domain, which enforce things that we know to be true about the dense data, similar to SIFT (Matsuki et al, 2009, 2010, 2011). In the frequency domain, the ${\widehat{P}}_1$ projection makes use of a predefined “mask function” mask(f). ${\widehat{P}}_1$ sets to zero any points in the mask (f) = 0 region, where we know that no signal exists. In addition, we exploit our knowledge of the signal’s phase and sign within the non-zero support mask regions to further strengthen ${\widehat{P}}_1$. Specifically, all imaginary parts are set to zero, and the negative (positive) real parts are set to zero where mask(f) = +1 (−1). Real signals of either sign are left alone in the “artifact regions,” where mask(f) = 2. The values of mask (f) can either be defined manually, or using an algorithm like this:

• $\text{mask}(f)=0\text{ if }\left|Re(\tilde{T}(f))\right|<\tau$

• $\text{mask}(f)=+1\text{ if }Re(\tilde{T}(f))\ge \tau$

• $\text{mask}(f)=-1\text{ if }Re(\tilde{T}(f))\le -\tau$

• mask(f) = +2 if the sign of $Re(\tilde{T}(f))$ is uncertain, and $\left|Re(\tilde{T}(f))\right|\ge \tau$

where τ is a positive threshold. In our case, the dense spectra from slices 2–12 of our pseudo-3D data set are similar enough that any one of them can be used to create masks for the others. Here, we use slice 2 to create our mask, and take τ to be ≈ 6 times the noise level of f₂ columns far from any signal or artifacts. We will revisit this choice of τ below (see Can K_p be lowered arbitrarily by increasing the ${\widehat{P}}_1$ mask threshold τ?). We also manually chose to use mask (f)= 2 for the at the f₁ edges of the spectrum, because that signal is an artifact whose sign is inconsistent across slices.

In the time domain, the ${\widehat{P}}_2$ projection resets measured points to their known values (i.e., the points in S⁰ (t) that were not zeroed by ${\widehat{P}}_0$). While these projections can be altered slightly to allow for noise-handling (Frey et al, 2013), we are not using noise-handling in this paper. This simplification will be helpful later (see Reconstruction error source II: signal outside the mask), since it enables us to develop a quantitative prediction of the size of reconstruction errors that arise from non-zero signal in the mask = 0 region.

Rather than simply alternating between these two projections, we combine them in a particular form of Elser’s Difference Map algorithm (Elser, 2003; Elser et al, 2007) which reproduces Fienup’s hybrid input-output map (Fienup, 1982). Our map is given by $\widehat{D}=\mathbb{1}+{\widehat{P}}_1\cdot \left(2{\widehat{P}}_2-\mathbb{1}\right)-{\widehat{P}}_2$, where $\mathbb{1}$ is the identity operator (Elser, 2003; Frey et al, 2013), and the iterative mapping becomes $S^i=\widehat{D}{S}^{i-1}$. Fourier transforms and phase corrections are applied as needed to reversibly convert between time and frequency domains. The final output of the iterated maps algorithm after i iterations is given by $F^i={\widehat{P}}_2{S}^i$. Most of the reconstructions we calculated required somewhere between 100 and 10,000 iterations to converge (see Supplemental Information for details). The DiffMap calculations were all performed in Igor Pro 8. We have also ported the code to Python 3, and verified that the Python and Igor results agree to within numerical precision. The Python code is available on GitHub at https://github.com/robbyblum/DiffMapcoDiffMap-Python, and will also be available on the NMRbox platform at http://NMRbox.org (Maciejewski et al, 2017).

Because the FT is a linear transformation, and because sparse-sampling is only used in the indirect dimension, the full 2D spectral reconstruction problem can be treated as a series of independent, 1D spectral reconstructions; i.e., each f₂ column can be reconstructed independently of the others. We took advantage of this fact to speed up our study of how the iterated maps method breaks down at low $N_{t_1}$. At a given $N_{t_1}$, we refer to the reconstructed 1D spectrum of a particular f₂ column with index n as ${\tilde{F}}_n^{i,N_{t_1}}\left(f_1\right)$. Furthermore, if we reconstruct a particular peak at f₁ index m well, then ${\tilde{F}}_{mn}^{i,N_{t_1}}\approx {\tilde{T}}_{mn}$ We define the reconstruction error of a point as a function of $N_{t_1}$ as $E_{mn}\left(N_{t_1}\right)={\tilde{F}}_{mn}^{i,N_{t_1}}-{\tilde{T}}_{mn}$. For a particular peak p at indices (m_p, n_p), we will refer to the reconstruction error as simply $E_p\left(N_{t_1}\right)$.

What is a reasonable target to shoot for when using NUS?

In many papers that use an NUS approach, the authors report reaching a low sampling fraction $N_{t_1}/N_{t_1}^{\text{dense}}$ (or its (n-1)-dimensional analog for nD experiments). However, as other authors have pointed out, a spectrum that has only one narrow feature should be easier to reconstruct than one with multiple peaks, even when both have the same $N_{t_1}^{\text{dense}}$ (Shchukina et al, 2017; Hyberts et al, 2014). Thus, a better metric to use when comparing the performance of NUS approaches is the ratio of the number of sparsely-sampled measurements to the number of significant points to be reconstructed. We can think of the number of significant points as the “information content” of the dense spectrum. We will adopt such a metric below.

Following the notation of Shchukina et al (2017), we define K(n), the number of significant points (relative to a threshold τ) of the nth f₂ column, as the information content of that 1D spectrum ${\tilde{T}}_n\left(f_1\right)$. We calculate K(n) by counting all of the mask =−1,+1, and +2 points in the nth column, and define K_p = K (n_p) for a particular peak p in column n_p. It is important to remember that K(n) is much larger than the total number of peaks in a column, since all real peaks have nonzero width. Because we fill all four quadrants of the time domain using the “States”-like data, the relevant number of sparsely-sampled points to compare to K(n) is $2N_{t_1}$, where $N_{t_1}$ is the number of complex points sampled with t₁ ≥ 0. Thus, we can characterize the compression achieved by NUS using the ratio $\mathcal{C}\equiv 2N_{t_1}/K(n)$.

Even though most NMR spectra have a higher density of non-zero points than the test images typically studied in the compressed sensing (CS) literature, we can compare our results to CS benchmarks. If we insert our parameters into expressions from the CS literature (see Eq. 9.40 of Foucart and Rauhut, 2013), we find that CS approaches are likely to yield high-quality reconstructions when:

$$2N_{t_1}>2K_p\text{ln}\left(M/K_p\right)(CSlimit\text{).}$$ (1)

For M = 256, and an average $\overline{K_p}=53.5$ (over all 114 peaks), the corresponding CS limit on the achievable compression ratio is ${\mathcal{C}}_{\text{CS}}>3.1$. This is already a significant improvement over the corresponding value for dense sampling, ${\mathcal{C}}_{\text{dense}}=4.8$. Note that the only assumption made by CS approaches is that the spectrum is “sparse.”

We may be able to beat this CS limit if we know where spectral features should not appear, since this constraint provides more information about the spectrum. In fact, our iterated maps approach has the potential to push down to the limit imposed by linear algebra (LA): that the number of “constraints” M − K_p must be greater than or equal to the number of “unknowns” $M-2N_{t_1}$:

$$2N_{t_1}\ge K_p(LAlimit).$$ (2)

To test this prediction, we tried to reconstruct an extremely simple, single-peak spectrum (Fig. 2a) that exactly satisfies the constraints of ${\widehat{P}}_1$ and ${\widehat{P}}_2$. In this ideal limit, as we lower $N_{t_1}$ below $N_{t_1}^{\text{dense}}$, there is essentially zero error in the reconstruction of the peak, all the way down to the ideal LA limit of $\mathcal{C}=1$ (Fig. 2b). Thus, $\mathcal{C}=1$ seems to be a reasonable compression target to shoot for with this method.

Fig. 2.

a “Idealized” spectrum, made by setting most of an actually-acquired spectrum, and its mask, to 0. b Error of the reconstructed peak amplitude for the “ideal” peak of interest shown in a. Vertical dashed line indicates K_p/2. The error is zero (within machine precision) until $N_{t_1}<K_p/2$. c The actually-acquired spectrum (for peak Leu50 δ², $i_{{\tau }_{\text{cp}}}=2$) that was simplified in a, indicating the important features. d Error of the reconstructed peak amplitude for peak Leu50 δ², $i_{{\tau }_{\text{cp}}}=2$. Many complicated features appear that did not arise in the “ideal” case. Notice that K_p has increased because there are more features in the real spectrum than in the simplified spectrum, which increases the size of the non-zero mask region. Note that in b and d we are graphing the absolute difference between the reconstructed and dense amplitudes of the peak of interest. We use this convention throughout the paper

Features of a Realistic Spectrum that Challenge the Iterated Maps Approach

Compared to the ideal case, a realistic 1D NMR spectrum has many additional features (e.g., Fig. 2c). The most obvious example is the presence of other peaks, which can be either positive or negative in sign, with their own corresponding mask regions (mask(f) = +1 or 1, respectively). There also can be artifact regions, which are usually unwanted spectral features that do not have a predictable sign, but which must be accounted for (mask(f) = 2). Finally, even in the regions “without signal” (mask(f) = 0), any real spectrum will have random noise, as well as spectral tails that fall below the mask cutoff τ All of these features make it harder for the iterated maps method to work as well as it does in the ideal case. This can be seen by comparing Fig. 2b to Fig. 2d: in the latter, even well above the $\mathcal{C}=1$ limit, the peak reconstruction error $E_p\left(N_{t_1}\right)$ is clearly non-zero, and it can swing wildly over narrow regions of $N_{t_1}$. In fact, these trends are even easier to see in Fig. 3, where we plot the reconstruction errors for the same peak used in Fig. 2c–d (Leu50 δ²) but for all eleven slices of the pseudo-3D data set. Minor variations in the dense spectrum across different slices lead to large spreads in $E_p\left(N_{t_1}\right)$ in certain regions of $N_{t_1}$, but not in others.

Fig. 3.

Error in the reconstructed amplitude for Leu50 δ², shown for all 11 slices at each $N_{t_1}$. The mask-collision metric (blue shading) indicates regions where significant errors are predicted to be possible (see description of “MCM” in main text). The linear algebra limit K_p/2 is indicated by the dashed line, below which errors grow very quickly (gray shading). The dense peak amplitude for Leu50 δ² (averaged over all 11 data slices) is 58.4 × 10⁵, and the mask cutoff is τ = 0.9 × 10⁵

Methods

In this section, we will discuss how the reconstruction error arises from two main categories of complications:

• ISignal inside the mask (mask (f) ≠ 0)
• IISignal outside the mask (mask(f) = 0)

In each case, we will present a metric that is able to describe the reconstruction error using available information. Category I causes the largest errors, so we discuss this first.

Reconstruction error source I: signal inside the mask

As pointed out by Maciejewski et al (2009), any NUS schedule leads to intraband aliases of all spectral peaks. These intraband alias features appear to move along the f₁ axis as $N_{t_1}$ is varied. As we studied the performance of our iterated maps method we noted a clear pattern: as $N_{t_1}$ is lowered, the biggest reconstruction errors for a given peak happen when there is a collision between its larger intraband alias features and the nonzero mask regions, or “mask collisions” for short. Qualitatively, at these $N_{t_1}$ values, ${\widehat{P}}_2$ mixes up the amplitudes at frequency points connected by intraband aliases, AND ${\widehat{P}}_1$ is unable to resolve the confusion. Mask collisions seem to matter for any nonzero mask(f) value (that is, +1, −1 or 2), so we will treat all these cases as being equivalent. The realization that mask collisions cause the worst errors motivated us to switch to a quasi-even sampling schedule.

To understand this further, recall that a non-uniform sampling schedule in the time domain has a corresponding point spread function (PSF) in the frequency domain (see Supplementary Information). The PSF determines how undersampling causes spectral information to be mixed among distinct f₁ points, which results in the intraband aliases. A periodic sampling pattern results in a coherent PSF with sharp side-lobe features, separated by nulls. A randomized sampling pattern results in an incoherent PSF, smearing the sidelobes, and filling the nulls (Maciejewski et al, 2009). Most NUS reconstruction approaches prefer the less coherent PSFs of sampling schedules with irregular spacings, because they tend to broaden and scramble the intraband alias features, which helps to reduce the extreme errors encountered as $N_{t_1}$ is lowered. Unfortunately, this broadening tends to fill in the nulls in the PSF, which increases the errors elsewhere. Nevertheless, this is a safe strategy that enables $N_{t_1}$ to be picked somewhat arbitrarily, as it smooths out the extreme errors in reconstruction.

However, because we are trying to push so aggressively to the lowest possible $N_{t_1}$ values, we will use the opposite strategy. Namely, we turn off the random jitter of QUEST sampling (Frey et al, 2013), in order to implement deterministic quasi-even sampling, since this sharpens the sidelobes of the corresponding PSF (see Supplementary Information).

As $N_{t_1}$ is lowered, this will increase the reconstruction errors due to mask collisions. However, these larger errors will now be restricted to smaller regions of $N_{t_1}$, which makes it easier to avoid them (see Fig. 3). At the same time, quasi-even sampling will also decrease the errors in the absence of mask collisions. As long as we can predict where the small errors will occur as a function of $N_{t_1}$, this aggressive strategy should result in the best possible reconstructions of the peak of interest.

Given a mask, we can identify potentially dangerous $N_{t_1}$ values for a particular f₁ peak in a particular f₂ column (e.g., the blue bands in Fig. 3 show the so-called danger zones). If we want to achieve a high-quality reconstruction of that peak, we should only use $N_{t_1}$ values in the complementary safe zones (e.g., the white bands in Fig. 3). We describe how we locate the danger zones next.

Finding the $N_{t_1}$ danger zones for a peak of interest

We expect the largest reconstruction errors for a given peak to occur in the danger zones of $N_{t_1}$, i.e., when significant intraband aliases of that peak collide with the mask. For a single peak of interest, we first need to know where its intraband aliases end up at a given $N_{t_1}$. We then use these intraband alias locations and the ${\widehat{P}}_1$ mask function to define our mask-collision metric (MCM).

For a given $N_{t_1}$, the t₁ ≥ 0 points in all sampling schedules discussed here can be described with a single expression:

$$t_1\in \left\{t_1^{\text{off}}+\delta t_1\text{Round}\left(\frac{N_{t_1}^{\text{dense}}}{N_{t_1}}k\right)\right\},$$ (3)

where $k\in \left\{0,1,\dots ,N_{t_1}-1\right\}$, δt₁ is the dwell time in the indirect dimension, and $t_1^{\text{off}}$ can either be 0 or δt₁/2. The case $N_{t1}^{\text{dense}}/N_{t_1}=1$ is what we mean by dense sampling. A given t₁ ≥ 0 sampling schedule provides the full sampling vector s(t₁) of length M to match our four-quadrant data, which in turn provides the PSF after a Fourier transform (see Supplementary Information for details) (Maciejewski et al, 2009).

We start by considering the simplest undersampled case: even undersampling (where $N_{t1}^{\text{dense}}/N_{t_1}$ is an integer greater than one) with $t_1^{\text{off}}=0$. For this case, sidelobes in the PSF appear at integer multiples of the sparse bandwidth $M{N}_{t_1}/N_{t_1}^{\text{dense}}$ away from the central peak at m₀ = M/2,

$$m_{\pm h}=m_0\pm 2N_{t_1}h\text{\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}mod\hspace{0.17em}}M,$$ (4)

where this defines the harmonic number h = 0, 1, 2, …. For instance, in Fig. 4a, $N_{t_1}=32$, so each harmonic is 64 points away from the previous one. If $m_0\pm 2N_{t_1}h$ falls outside the spectral width, the modulus operation (“mod M ”) brings it back in by “wrapping it around” to the far side of the window. Our earlier discussion of intraband aliasing (Frey et al, 2013, S.I.) included the $\pm 2N_{t_1}h$ term, but we failed to consider the wraparounds, which will turn out to be important in what follows.

Fig. 4.

a Point-spread function of a purely-even undersampling pattern for $N_{t_1}=32$ and $t_1^{\text{off}}=0$ (real part). b Point-spread function of a quasi-even sampling pattern for $N_{t_1}=34$ and $t_1^{\text{off}}=0$ (real part). More harmonics become differentiated. c Point-spread function of a sampling pattern for $N_{t_1}=34$ using quasi-even sampling and $t_1^{\text{off}}=\delta t_1/2$ (real part). In all three plots, harmonics ±h = 0, 1, 2, 3 are indicated

Truly even sampling is only possible for a few $N_{t_1}$, so we are usually forced to approximate it with quasi-even sampling ($N_{t1}^{\text{dense}}/N_{t_1}$ is not an integer, so the rounding in Eq. 3 matters). The PSF for one such case, $N_{t_1}=34$, is shown in Fig. 4b. While the number of nonzero points is much greater in this case, the locations of the sidelobes are still given by the same m _±h formula as before.

Finally, for our data set, we do not actually have $t_1^{\text{off}}=0$; instead, we have $t_1^{\text{off }}=\delta t_1/2$. This complication broadens the absorptive sidelobes of the PSF, making them appear somewhat more dispersive. Fig. 4c shows this effect at $N_{t_1}=34$. Because the harmonic sidelobes are no longer delta functions at exactly m_±h, we account for their nonzero width below.

Equation 4 gives us the location of the PSF sidelobes, but we are trying to find the location of the intraband aliases. The PSF, when convolved with a dense spectrum, generates a sparse spectrum. Thus, adapting Eq. 4, we know that the intraband aliases of a particular peak of interest at m_p will occur at ($m_p\pm 2N_{t_1}h\text{\hspace{0.17em}mod\hspace{0.17em}}M$) (i.e., replace m₀ with m_p in Eq. 4). From here, to see if an intraband alias is colliding with the mask, all we need to do is check whether the mask function (in the 1D column of interest) is nonzero at these locations. That is, for a given harmonic order h, we define a “hit function”

$$\text{Hit}\left(m_p,h,N_{t_1}\right)=\left(\text{mask}\left(m_p+2N_{t_1}h\text{\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}mod\hspace{0.17em}}M\right)\ne 0\right)+\left(\text{mask}\left(m_p-2N_{t_1}h\text{\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}\hspace{0.17em}mod\hspace{0.17em}}M\right)\ne 0\right).$$ (5)

There are technically an infinite number of intraband alias harmonics, but in practice, a given harmonic’s mask collisions do not seem to noticeably impact the quality of the DiffMap reconstruction until you get to a low enough $N_{t_1}$, which seems to be proportional to $N_{t_1}^{\text{dense}}/h$. Thus, for a given $N_{t_1}$, we establish a maximum harmonic number to check for collisions, h_max, given by

$$h_{\text{max}}=\text{Floor}\left(\beta \frac{N_{t_1}^{\text{ dense }}}{N_{t_1}}\right),$$ (6)

where we used β ≈ 0.82625 in this study (see the next section for further discussion).

Putting this all together, we are now able to define the Boolean mask collision metric (MCM). For a peak p at index m_p, $\text{MCM}\left(m_p,N_{t1}\right)$ predicts which $N_{t_1}$ are likely to result in a problematic collision between aliases and the mask:

$$\text{MCM }\left(m_p,N_{t_1}\right)=\left(\left[\sum _{h=1}^{h_{\text{max}}}\sum _{i=-1}^{+1}\text{Hit}\left(m_p+i,h,N_{t_1}\right)\right]>0\right).$$ (7)

Note that the sum over i = −1, 0, +1 is how we are accounting for the nonzero width of the harmonic sidelobes for the $t_1^{\text{off }}=\delta t_1/2$ case.

A schematic example of how this calculation works is shown in Fig. 5. The MCM divides the set of possible $N_{t1}$ into two categories: either a danger zone or a safe zone. Returning to Fig. 3, the blue bars show the danger zones, where $\text{MCM}\left(m_p,N_{t1}\right)=1$; all other $N_{t1}$ values are in safe zones, where $\text{MCM}\left(m_p,N_{t1}\right)=0$. Note that the MCM can be calculated as soon as the mask has been defined.

Fig. 5.

Schematic example of a mask collision metric calculation in a 1D spectrum, which requires only the peak-of-interest location (blue shading) and the ${\widehat{P}}_1$ support mask (pink shading). We wish to check for mask collisions for $N_{t_1}/N_{t_1}^{\text{dense}}=34/128$. In this case, h_max = Floor(3.11) = 3 (see Eq. 6), so we need to check three harmonics. We start at the location of the peak of interest (m_p =200) and go out (SW*34/128)*h (h= 1, 2, 3) in both directions, where SW is the sweep width. If this takes us off the edge of the spectrum, we wrap around to the far side of the spectrum. We then check each of the calculated locations for nonzero mask values. In this case, we find nonzero mask values at both the location of the second “up” harmonic (blue arrow) and that of the second “down” harmonic (black arrow), so Hit(200, 2, 34) = 2 (see Eq. 5), and the Boolean MCM(200, 34) = 1 (see Eq. 7)

Testing the predictive power of the mask collision metric

Now that we have defined the mask collision metric, we can easily apply it to all 114 peaks in our spectrum. For example, the danger zones for three additional peaks are shown in Fig. 6. The blue points show the absolute error in the peak reconstructions, calculated for all 11 slices of the pseudo-3D data set.

Fig. 6.

Examples of the MCM for three peaks, with reconstruction results. a Peak Val56 γ², with average dense amplitude 1.46 × 10⁷. b Peak Val18 γ¹, with average dense amplitude 5.72 × 10⁶. c Peak Ile52 δ¹, with average dense amplitude 5.66 × 10⁶. In each graph, the blue shading indicates which $N_{t_1}$ the metric flags as potentially bad, and the dots represent the error in the reconstructed peak amplitude $E_p\left(N_{t_1}\right)$, for all 11 slices of the pseudo-3D data set. The black dashed line indicates the linear algebra limit K_p/2 for the column

We can see some patterns start to emerge across Fig. 3 and Fig. 6. First, the reconstruction quality becomes terrible once $N_{t1}$ goes below the linear algebra limit, $K_p/2(\mathcal{C}=1)$. This makes sense, because in that region we do not expect to have enough information to solve the reconstruction problem uniquely.

Second, above $\mathcal{C}=1$, the metric does a very good job of locating dangerous $N_{t1}$ values which result in large reconstruction errors, which often appear as large variances in the error between the slices.

Third, above $\mathcal{C}=1$, we can see that the mask collision metric is somewhat conservative: it has more false positives than false negatives. For experimentalists, false negatives (predicted safe zones of $N_{t1}$ that nonetheless lead to bad reconstruction) would be the more significant problem, so we tried to construct the MCM so as to make them as rare as possible. As we explained in the last section, our MCM checks 3 adjacent points at each intraband alias location, which was the most systematic way we found to eliminate almost all false negatives without adding enormous numbers of false positives. We also wanted to limit false positives (predicted danger zones of $N_{t1}$ that turn out to have good reconstructions), since they exclude more $N_{t1}$ values than necessary. This motivated how we set β for Eq. 6. In particular, we were able to eliminate a large number of trivial false positives at very high $N_{t1}$. There are still remaining false positives which seem very hard to exclude, at least within our current understanding of the method.

In principle, β can be any positive number, but in our model, we have restricted it to the range 0 < β< 1. To choose a value for β, we used a procedure based on the peak-to-sidelobe ratio of the PSF for a given harmonic (see Supplementary Information for details). We were able to tune β to improve the accuracy of the MCM for our data, but we do not know how much it might need to be adjusted for other applications. In general, it is prudent to err on the side of high β as lowering it too far will introduce an increasing number of false negatives. It is actually perfectly acceptable to use a default value of β = 0.99 (see Supplementary Information for discussion of why we limit β< 1).

Before leaving this section, it is worth emphasizing two points. First, as we have defined it so far, $\text{MCM}\left(m_p,N_{t1}\right)$ is only predicting dangerous and safe $N_{t1}$ values for the reconstruction of a single peak of interest at f₁ index m_p; it says nothing about reconstruction error elsewhere in the spectrum. For example, as Fig. 7 shows, good (bad) reconstruction performance at m_p can be accompanied by bad (good) performance elsewhere. Second, while the magnitude of the peak of interest at f₁ index m_p has to be larger than our mask threshold τ, that is all that we have assumed to this point. It does not need to be an actual peak, such as a local maximum in the f₂ column, or a peak picked in the 2D spectrum.

Fig. 7.

Reconstruction performance examined in detail. a Reconstruction error in the amplitude of the Ile73 δ¹ peak (dots), with danger zones predicted by the MCM (blue shaded region) and the limit K_p/2 (dashed line). The results for $N_{t_1}=32$ and $N_{t_1}=34$ (larger dots) are examined in more detail in b and c, respectively. b “Good” reconstruction example, $N_{t_1}=32$. This is a good result despite the significant errors elsewhere since we are only interested in the peak Ile73 δ¹ (filled circle). c “Bad” reconstruction example, $N_{t_1}=34$. Most of the spectrum is reconstructed well, but this is a bad result since there is a noticeable error (more than 10%) at the location of the Ile73 δ¹ peak (filled circle)

Reconstruction error source II: signal outside the mask

As the last section demonstrates, the metric is good at identifying danger zones, or $N_{t1}$ values where the largest reconstruction errors can occur, due to significant intraband aliases colliding with the mask. The metric also picks out $N_{t1}$ values that fall in the complementary safe zones.

To understand DiffMap’s performance in the safe zones, we need to take into account a second, smaller source of reconstruction error. Our iterated maps method assumes that the mask we use in ${\widehat{P}}_1$ is correct; that is, there is NO signal outside of the support regions. This is just an approximation, since all real NMR spectra have noise. For the 2D data sets under study, the noise is very small compared to the peaks, but it is still nonzero. Furthermore, by making a mask using a threshold τ applied to a dense data set, you basically guarantee that there will be at least some tails of your peaks that fall outside the mask. Noise fluctuations and tails are actual spectral features of the dense spectrum, just like the significant peaks are. DiffMap treats them the same way as it does the large peaks, with one important distinction: the true location of these outside-the-mask features is disallowed by ${\widehat{P}}_1$. The ${\widehat{P}}_2$ operation will try to put them back in the right spot, and for large $N_{t_1}$, the final ${\widehat{P}}_2$ will mostly fix these features, but ${\widehat{P}}_1$ will always be trying to drive them out of the mask = 0 regions. So where do they end up in the spectrum? These features have intraband aliases just like the large peaks do, so for a given $N_{t1}$, they will go into the nonzero mask regions which collide with their intraband aliases. Thus, the signal outside the mask gets folded into the mask.

To describe the impact of signal from outside the mask on the reconstruction of points inside the mask, we will use a model in which the spectral features at different points outside the mask are drawn from a normal distribution with mean μ_v and standard deviation σ_v. Note that this is not really a correct description of this signal, because random electronic noise and spectral tails are somewhat different in character. Noise is incoherent, has a mean of zero, and is present everywhere in the spectrum. Spectral tails are coherent, usually positive-amplitude, and localized; in the context of a cutoff-generated mask, they will generally be found just outside the support regions. However, we have found that applying this model with μ_v and σ_v drawn from the calculated mean and standard deviation of the entire mask = 0 region in a given spectrum still gives us accurate bounding behavior for E_p. The main difference we see is that the fluctuations end up being more correlated in $N_{t1}$ than they would be in a purely random case.

Here we briefly describe the results of our model. The full derivations are in the Supplementary Information. We consider a spectrum with a single delta-function peak of amplitude A_dense at point m_p, with a corresponding single non-zero mask point. For simplicity, we assume we are in the $t_1^{\text{off}}=0$ case; simulations of the $t_1^{\text{off }}={\delta }_1/2$ case do not show a significant change in behavior. We call the reconstructed amplitude A_recon, and calculate the mean μ_E and the standard deviation σ_E of the peak reconstruction error E = A_recon − A_dense.

By considering the signal outside the mask in terms of its mean and standard deviation, we can consider the effect of each part separately. To calculate μ_E, we consider a delta-function spectrum with no noise and a DC offset μ_v. Then, the mean error is

$${\mu }_E\left(N_{t_1}\right)={\mu }_v\left(\frac{N_{t_1}^{\text{dense}}}{N_{t_1}}-1\right).$$ (8)

To calculate the standard deviation of the error σ_E, we again consider a delta-function spectrum, but now with added noise that has zero mean and standard deviation σ_v. From this, we find that the standard deviation of the error is

$${\sigma }_E\left(N_{t_1}\right)={\sigma }_v\sqrt{\frac{N_{t_1}^{\text{dense}}}{N_{t_1}}-1}.$$ (9)

Combining equations 8 and 9, we predict that 95% of the errors for $N_{t1}$ in the safe zones will lie within the range

$$E_{\text{SZ}}\left(N_{t_1}\right)={\mu }_E\left(N_{t_1}\right)\pm 2{\sigma }_E\left(N_{t_1}\right).$$ (10)

If we plot this prediction band against the actual reconstruction error for various peaks, we find that it does a very good job of explaining the observed error in the safe zones (see Fig. 8, as well as additional examples in the Supplementary Information). For the $N_{t1}$ values in the danger zones, the reconstruction error can push well outside this prediction band, which is exactly what we expect.

Fig. 8.

E_SZ prediction bands for peak Leu50 δ². ${\mu }_E\left(N_{t_1}\right)$ is plotted as the red line, the ±1σ_E band is the darker red shading, and the ±2σ_E band is the light red shading. As in Fig. 3, the danger zones of the MCM are plotted as blue bars, and $E_p\left(N_{t_1}\right)$ for all 11 data sets are plotted as blue dots

It is worth pointing out that we have accurately predicted the reconstruction error for all of the 11 sparsely-sampled slices of the pseudo-3D data set, using just the information from a single densely-sampled slice (slice 2). From this spectrum, we were able to calculate the mask, the MCM, and the E_SZ prediction bands.

Can K_p be lowered arbitrarily by increasing the ${\widehat{P}}_1$ threshold τ?

Yes, but the reconstruction error tends to grow as well. The amount of systematic shift of the reconstruction error in the safe zones depends on how you construct your mask. If you set a higher positive cutoff τ, more tails will fall outside the mask. This tends to increase μ_v and σ_v of the points outside the mask, which in turn increase μ_E and σ_E. If you set a lower cutoff, the opposite will happen. Thus, choosing a mask involves making trade-offs: a tighter mask (higher τ) will lower K_p and produce more safe zones (fewer intraband alias-mask collisions), but will have more error in those safe zones. A wider mask (lower τ) will have higher K_p and fewer safe zones (more intraband alias-mask collisions), but will have tighter prediction bands. Fig. 9 shows the effect of choosing different mask thresholds τ on K_p, the mask collision metric, and the E_SZ prediction bands for the peak of interest shown in Fig. 2c. Going from 9a to 9c lowers the minimum “safe” $N_{t1}$ for this peak by a factor of 12, but increases the mean and width of the prediction band at that lower $N_{t1}$ by a factor of 36.

Fig. 9.

Metrics for different choices of the mask cutoff τ. a τ= 90, 000. For a lower mask cutoff, the nonzero mask region is larger (covers more features), increasing the number of alias overlaps with the mask, and decreasing the amount of “noise” (including tails) outside the mask. The linear algebra limit K_p/2 is higher (dashed line) because the ${\widehat{P}}_1$ operator can impose fewer constraints. b τ = 500, 000. Raising the mask cutoff decreases the number of alias overlaps, increases the amount of “noise” outside the mask, and lowers the K_p/2 threshold. c τ = 2, 900, 000. Dramatically increasing the mask cutoff unreasonably removes almost all danger zones, sharply increasing the expected errors elsewhere. The dense peak amplitude is approximately 5.8 × 10⁶

In some applications, using a tight mask may be preferred, e.g., when the speed-up due to the drop in K_p is much more important than the associated increase in the size of the prediction band.

We actually began our current study with a larger τ inherited from our previous work, but as we came to understand this issue better, we decided that lowering τ to 6 times the noise level enabled us to reach the lowest possible $N_{t1}$ while maintaining acceptable reconstruction error.

Results

Recipe for finding $N_{t_1}^p$, and overall reconstruction performance

The success of the alias overlap metric for $N_{t_1}\ge K_p/2$ suggests a simple recipe for finding the lowest $N_{t_1}$ that will give an acceptable reconstruction of a given peak. You simply start at $N_{t_1}=K_p/2$, and step up in $N_{t1}$ until you find the lowest value in a safe zone of the MCM, which we call $N_{t_1}^p$ for peak p. Alternatively, the recipe can be expressed as

$$N_{t_1}^p=\text{min}\left(N_{t_1}\right)\text{for}\{\begin{array}{l}{N}_{t_1}>K_p/2\hfill \\ \text{MCM}\left(m_p,N_{t_1}\right)=0\hfill \end{array}.$$ (11)

Fig. 10a shows the values of $N_{t_1}^p$ found by applying this recipe for each assigned peak in the spectrum (blue points). Many of them are very close to K_p/2 (gray region), and only one of them is above L/2 (thick dashed line). Using these $N_{t_1}^p$ values, we build up a histogram of the absolute reconstruction error $E_p\left(N_{t_1}\right)$ across all 114 peaks, for all 11 slices, as shown in Fig. 10b. Most of the errors are smaller in magnitude than one mask cutoff (90,000), and 88% of the reconstructed peak amplitudes are within 5% of their dense values. We also see a slight positive offset of the center of the distribution, as we expect due to the presence of net-positive tails outside the mask.

Fig. 10.

Performance of the overall metric. a $N_{t_1}^p$ for each peak in the spectrum (blue dots), compared to the linear algebra limit K_p/2. The limit L/2= 75 achieved in the prior work (Frey et al, 2013) is also shown (dot-dashed line). b Histogram of the reconstructed peak errors for all 114 peaks in the spectrum, and all 11 2D slices from the pseudo-3D data set. The histogram shows a slight positive mean, which is anticipated by the fact that the “noise” outside the mask has positive mean μ_E, due to the tails of the actual features. Most of the errors lie within one mask cutoff of zero (± 90,000)

Fig. 10a plots the $N_{t_1}^p$ values in absolute terms, which naturally suggests a sampling fraction. However, this is not an especially useful metric for comparing different reconstruction methods, because of how much the sampling fraction depends on the original sparseness of the test spectra (Shchukina et al, 2017; Hyberts et al, 2014). Thus, we also plot the “bottom-up” ratio ${\mathcal{C}}_{\text{bu}}=2N_{t_1}^p/K_p$ in Fig. 11a, as well as plotting $2N_{t_1}^p$ vs. K_p in Fig. 11b. From these plots, we see that even for complicated spectra, we have been able to push down to an average $\overline{{\mathcal{C}}_{\text{bu}}}=1.25$, which is almost all the way down to our expected limit of $\mathcal{C}=1$ (Eq. 2).

Fig. 11.

Comparison between $2N_{t_1}^p$ and K_p using our metric. a The “bottom-up” compression ${\mathcal{C}}_{\text{bu}}=2N_{t_1}^P/K_p$ for each peak (red points). The average $\overline{{\mathcal{C}}_{\text{bu}}}$ is 1.25 (blue horizontal line), and the standard deviation is 0.3. The dashed lines show ${\mathcal{C}}_{\text{dense }}=4.8$, ${\mathcal{C}}_{\text{CS}}=3.1$, and ${\mathcal{C}}_{\text{td}}=2.8$. b Scatter plot of $2N_{t_1}^p$ vs. K_p for all peaks. A linear fit gives a line with slope 1.043 ± 0.046 and intercept 3.80 ± 1.34 (solid blue line)

To further benchmark these results, we compare ${\mathcal{C}}_{\text{bu}}$ to values for dense sampling ${\mathcal{C}}_{\text{dense }}=M/\overline{K_p}$, the CS prediction ${\mathcal{C}}_{\text{CS}}=2\text{ln}\left(M/\overline{K_p}\right)$, and the “top-down” limit reached in Frey et al (2013) ${\mathcal{C}}_{\text{td}}=L/\overline{K_p}$, where the average value $\overline{K_p}=53.5$ is used for all peaks (Fig. 11a). Every individual-peak ${\mathcal{C}}_{\text{bu}}$ achieved is lower than any of these benchmarks.

The prediction bands described in Eq. 10 can be combined with the recipe (Eq. 11) to predict the scale of the peak reconstruction error you should expect at $N_{t_1}^p$. Whether that error level is acceptable depends, of course, on your experimental needs. A small modification to the recipe allows it to account for this: if the error range is too large at the initial “$N_{t_1}^p$,” simply keep going up through the safe-zone $N_{t_1}$ values until the error range is sufficiently small. This expanded recipe can be written as

$$N_{ti}^p=\text{min}\left(N_{t_1}\right)\text{for}\{\begin{array}{l}{N}_{t_1}>K_p/2\hfill \\ \text{MCM}\left(m_p,N_{t_1}\right)=0\hfill \\ \text{max}\left(\left|E_{\text{SZ}}\right|\right)<E_{\text{max}}\hfill \end{array},$$ (12)

where E_max is the maximum acceptable noise level. See Supplemental Information for an example of a case when this might be necessary. That said, in Fig. 10 and Fig. 11, we simply used the lowest $N_{t_1}^p$ above K_p/2 (i.e., we applied Eq. 11 instead of Eq. 12).

If we build up from the reconstruction of a single peak at a time, to many peaks in parallel, what happens to $N_{t_1}^{*}$?

In order to answer (Q.I), we reduced the reconstruction problem to its smallest possible target: a single peak of interest in a single f₂ column. Using a deeper understanding of the iterated maps method, we pushed close to the theoretical minimum $N_{t_1}^p$ for each peak.

Turning to (Q.II), we want to see if we can build upon these results to attack the larger problem of reconstructing many, or all, peaks at one time (using a single global sampling number, $N_{t_1}^{*}$), as we did in our earlier study (Frey et al, 2013).

The mask collision metric MCM is different for each (f₁, f₂) point of interest, but it is easy to extend it to apply to a group of adjacent points in the same f₂ column: simply add together the metric for each point. Any $N_{t_1}$ values for which the combined metric is 0 should be safe for all of the points being considered. It would make sense to do this for a group of adjacent f₁ points, for instance, if one wants to do curve fits to interpolate peak locations. In that case, the combined metric tends to resemble the metric of the center point of the peak.

Extending the metric method beyond this case, however, starts to get trickier. For other non-adjacent points in the same column (a different peak, for instance), the metrics may not overlap very well, even though they use the same mask. The most challenging case is typically peaks in different f₂ columns, since there is no guarantee that the safe zones in the metrics will line up at all. This is simply because the masks in distant columns of a complicated spectrum are not very correlated with each other, so there is no reason to expect that you will be able to use the same $N_{t_1}$ for multiple peaks across the spectrum. Fig. 12 shows an example of this issue, using just three peaks of interest. For two of the peaks, their respective mask collision metrics have a safe zone at the same value of $N_{t_1}^p$. But for the third peak, that same $N_{t1}$ value is in its danger zone. Despite the fact that the third peak’s K_p and $N_{t_1}^p$ are lower than that of the other two, the “combined $N_{t_1}^p$” ($N_{t_1}^{*}$) for all three peaks is much higher.

Fig. 12.

MCM predictions for the three peaks in Fig. 6 (Val56γ²,Val18γ¹, and Ile52 δ¹). $N_{t_1}^p$ is highlighted for each peak (green dashed lines), calculated according to the recipe in Eq. 11. The peaks Val56γ² and Ile52 δ¹ (top and bottom) both have $N_{t_1}^p=30$, but in order to simultaneously satisfy the metric condition for those peaks and Val18γ¹ (middle), one must go all the way up to $N_{t_1}^{*}=49$. This problem compounds with each added peak

This problem means that, at least for the relatively complicated spectra analyzed here, the insights we have gleaned into the dependence of the local reconstruction performance on $N_{t1}$ cannot be applied to lower the global value of $N_{t_1}^{*}$ below the already-reached level of L/2. In fact, a strict application of the “union of sets” approach described above to the full set of 114 peaks would push $N_{t_1}^{*}$ well above the level of L/2. This is not reasonable, however, because we already know from the top-down approach in Frey et al (2013) that pushing down to L/2 is, in fact, globally acceptable. This apparent contradiction is likely due to the fact that the false positive rate of the MCM grows with increasing $N_{t1}$. This shows that while false positives are slightly inconvenient for a single peak, they are crippling when considering a large number of peaks.

Discussion

This paper was motivated by a simple question about the iterated maps approach: could we use a greatly improved understanding of the action of DiffMap on a particular 2D data set to reach an $N_{t1}$ below the previous minimum number of sparsely-sampled points ($2N_{t_1}^{*}=L=150$) achieved for that data set (Frey et al, 2013)? Focusing on this question narrowed the scope of this study. For example, we worked with the data as acquired, so any optimization of our NUS sampling schedules had to be consistent with the existing dense grid of points. Thus, even though the sharper PSF for the case of a t₁ = 0 point could lead to improved performance of our method, the current study is limited to the case of a t₁ = δ₁/2 point (see Fig. 4b,c). As another example, we did not directly compare the performance of DiffMap to any other NUS reconstruction methods. Nevertheless, by reporting the minimum compression ratios $\mathcal{C}$ achieved (as opposed to the undersampling fraction, $N_{t_1}^p/N_{t_1}^{\text{dense}}$), we hope make it easier to compare results across methods. The limited scope allowed us to do a deep dive into the performance of DiffMap.

We started by reducing the 2D problem to a much simpler one: understanding the reconstruction of a single peak of interest in a single f₂ column. Using quasi-even sampling, we developed a metric to predict the danger zones of $N_{t1}$ values, regions which are correlated with the largest reconstruction errors. We do not know how to predict the size of those errors, but we can avoid them altogether by using only $N_{t1}$ values in the complementary safe zones.

While reconstruction quality is typically better in the safe zones, it does degrade as $N_{t1}$ is lowered. We came up with a quantitative model for this smaller source of error, which can be traced back to the signal outside the mask folding back into the mask. Next, we made a recipe for defining and getting to the “minimum acceptable $N_{t_1}$ ” for a given peak, called $N_{t_1}^p$. We then applied this procedure to all 114 peaks, peak-by-peak, and it worked very well. In the end, we were able to get high quality reconstructions of each peak (for all 11 slices), while pushing down to $\overline{{\mathcal{C}}_{\text{bu}}}=1.25\pm 0.31$ across all 114 peaks, which is quite close to the fundamental limit on the compression ratio, $\mathcal{C}=1$.

To contextualize our results, we can compare to a nice recent study of NUS for pseudo-3D data done by Linnet and Teilum (2016). They did a very careful analysis of how many samples they needed to reconstruct a series of spectra, each with different amounts of information content. They define sparsity $\mathcal{S}$ as the fraction of data points in a contiguous (2D) region of the spectrum whose intensities (absolute value of amplitude) are less than six times the noise level, and coverage c as the minimum sampling fraction required to get an accurate reconstruction of a 2D spectrum using MDD. Conveniently, we also defined τ to be six times the noise level when creating our mask, so $\mathcal{S}$ is equivalent to $1-\overline{K}/M$, where $\overline{K}$ is the average K over the same 2D region. Meanwhile, c is a sampling fraction, so it is equivalent to what we call $N_{t_1}/N_{t_1}^{\text{dense}}$. Therefore, we can calculate an effective ${\mathcal{C}}_{\text{LT}}$ for their results which is very similar to $\mathcal{C}$ as we have defined it for our method: ${\mathcal{C}}_{\text{LT}}=c/(1-\mathcal{S})$. Across their four spectra(with sparsities ranging from $\mathcal{S}$= 92% to $\mathcal{S}$ = 71%), their average compression ratio was ${\overline{\mathcal{C}}}_{\text{LT}}\approx 2.8$, which is comparable to our previous ${\mathcal{C}}_{\text{td}}$. Our $\overline{{\mathcal{C}}_{\text{bu}}}$ is much lower than any of their ${\mathcal{C}}_{\text{LT}}$ values, but this direct comparison is not entirely fair, because $\overline{{\mathcal{C}}_{\text{bu}}}$ is the average of a set of separate 1D calculations, each focused on reconstructing a different peak, which do not share the same $N_{t_1}^p$ (or c) and K_p (or $\mathcal{S}$). However, this comparison does illustrate the benefit of focusing the reconstruction problem to a small portion of the spectrum.

Using the refined DiffMap approach we have described in this paper, a single peak of interest can be accurately reconstructed, using a number of measurements that is close to the fundamental minimum value. This result is immediately applicable to experiments requiring a particular spectral feature in a 2D spectrum to be repeatedly analyzed, as is done in the CEST experiment. In another example, after running an R₂ dispersion measurement, a user may want to get a higher density of τ_cp points for a particularly important residue. Examples such as these can take advantage of our understanding of the single peak reconstruction problem.

Our quantitative model for the smaller source of reconstruction error has an interesting implication: the signal-to-noise ratio (SNR) at the f₁ point¹ of a reconstructed peak of interest is always worse after using DiffMap, when compared to the SNR of the dense experiment. To see this, consider the ideal, best-case scenario of just “ideal noise” and no tails outside the mask, which in our error model implies μ_v= 0 and σ_v= σ_min, where σ_min is the standard deviation of the noise far from any peaks. In that case, if the SNR of the dense peak is A/σ_min, then the SNR of the reconstructed peak should drop to the lower value $\text{A}/\left({\sigma }_{\min }\sqrt{N_{t_1}^{\text{dense}}/N_{t_1}}\right)$. This expression is reminiscent of the SNR penalty suffered if fewer repetitions of an experiment are done. Of course, if the time saved by NUS is used to repeat the experiment $N_{t_1}^{\text{dense}}/N_{t_1}$ times, then this boosts the SNR by a factor of $\sqrt{N_{t_1}^{\text{dense}}/N_{t_1}}$. Unfortunately, even in this ideal case, the iterated maps SNR would just match the dense value of A/σ_min, in exactly the same time! In practice, a real spectrum has tails outside the mask, so μ_v ≠ 0 and σ_v > σ_min. Thus, the SNR of the iterated maps reconstruction is almost certainly going to be worse than the dense reconstruction. In our model, we assumed that ${\widehat{P}}_1$ and ${\widehat{P}}_2$ do noise-handling, so these conclusions be should be checked if is not the case. Still, we think it is safe thatto say that a peak of interest with the bare-minimum SNR in the dense spectrum is not a good candidate for iterated maps reconstruction.

The low values of ${\mathcal{C}}_{\text{bu}}$ presented here were achieved by analyzing individual features in particular 1D spectra. We tried to use our understanding of this case to find the minimum $2N_{t_1}$ required to reconstruct many, or all, peaks simultaneously. However, this bottom-up approach to finding a global $N_{t_1}^{*}$ did not work. Given the complicated spectrum under study, trying to reconstruct ∽ 3–4 peaks at once (out of 114) can be enough to push our recipe for $2N_{t_1}^{*}$ up to the previous value L = 150 found in the top-down approach (Frey et al, 2013). It is still possible that a broad search of the parameter space available to DiffMap (e.g., moving away from quasi-even sampling) could reveal new ways to improve global DiffMap performance for $2N_{t_1}<L$, but such a search is beyond the scope of this paper.

For the original problem of global 2D spectral reconstruction using iterated maps, the failure of our bottom-up approach to do any better than our top-down approach left us wondering if we could ever do better. In the end, we decided that there simply is not enough information in a single 2D slice to push below our previous limit of L (Frey et al, 2013). However, the fact that these 2D slices are part of a larger pseudo-3D data set suggests a way forward: use the correlations across 2D data sets as additional information that enables a global, high-quality reconstruction. In a companion paper, we develop a modified version of DiffMap, called coDiffMap (Rovny et al, 2019). Using coDiffMap, we show that we can improve results globally by exploiting the correlations that exist among 2D spectra in particular types of pseudo-3D experiment, adding information to our reconstruction method. This extra information enables us to push global $2N_{t_1}^{*}$ well below L, and even below $\mathcal{C}=1$ for certain cases (since the extra information across 2D slices adds to the information content of K_p). This alternate method also changes the scaling of the error that we described earlier in this paper. See Rovny et al (2019) for more details about coDiffMap.