Conditioning of partial nonuniform Fourier matrices with clustered nodes Thanks: The research of DB and LD is supported in part by AFOSR grant FA9550-17-1-0316, NSF grant DMS-1255203, and a grant from the MIT-Skolkovo initiative. The research of GG and YY is supported in part by the Minerva Foundation.

In many applications of interest, the columns of $\mathbf{V}_{N}$ as above arise from sampling the exponential functions $\left\{\exp\left(\imath\omega t_{j}\right)\right\}_{j=1}^{s}$ at equispaced points $\omega_{k}={k\over N}\Omega,\;|k|\leq N$, where $\Omega>0$ is a quantity which is frequently called the bandlimit or bandwidth, and the nodes $\{t_{j}:={N\xi_{j}\over\Omega}\}$ represent some relevant physical parameters, such as angles of arrival, locations of point sources etc. Therefore, in these cases it is more natural to regard $\{t_{j}\}$ and $\Omega$ as the primary variables instead of $\{\xi_{j}\}$ and $N$, while in fact thinking about the scenario where $N$ can be very large. According with this philosophy, we shall be primarily interested in the continuous limit $N\to\infty$.

With the above definition, the Gramian matrix $\mathbf{V}_{N}\left(\boldsymbol{x},\Omega\right)^{H}\mathbf{V}_{N}\left(\boldsymbol{x},\Omega\right)$ becomes in the limit $N\to\infty$ the kernel matrix with respect to the well-known $\sinc$ kernel.

For any matrix $\mathbf{G}\in\mathbb{C}^{s\times s}$, and a matrix $\mathbf{V}\in\mathbb{C}^{N\times s}$ with $N\geq s$, we denote as usual

The main subject of the paper is the scaling of the smallest eigenvalue of $\mathbf{G}$ and the smallest singular value of $\mathbf{V}_{N}$, when some of the nodes of $\boldsymbol{x}$ nearly collide (become very close to each other).

Let $\mathbf{V}_{N}$ be as defined in (2.2), i.e. a rectangular Vandermonde matrix with nodes $z_{j,N}=\exp\left(\imath\xi_{j,N}\right)$ on the unit circle with $\xi_{j,N}=t_{j}{\Omega\over N}$, $j=1,\ldots,s$. Denote $\Delta_{N}:=\min_{i\neq j}|\xi_{i,N}-\xi_{j,N}|$.

Several more or less equivalent bounds on $\sigma_{\min}\left(\mathbf{V}_{N}\right)$ are available in the “well-separated” case $N\Delta_{N}>const$, using various results from analysis and number theory such as Ingham and Hilbert inequalities, large sieve inequalities and Selberg’s majorants [23, 30, 34, 3, 31, 32, 19, 9].

The tightest bound was obtained in [3] (slightly improving Moitra’s bound from [30]), where it was shown that (in our notations we substitute $N\rightarrow 2N+1$) if $2N+1>{2\pi\over\Delta_{N}}$ then

In our setting, we have $\Delta_{N}={\Delta\Omega\over N}$ and so as $N\to\infty$ we obtain, assuming $\Delta\Omega\geq\pi$, that

The case $\Delta\Omega\ll 1$, or, equivalently, $\min_{i\neq j}|\xi_{i,N}-\xi_{j,N}|\ll{1\over N}$, turns out to be much more difficult to analyze. All known results provide sharp bounds only in the particular case when all the nodes are clustered together, or approximately equispaced.

If all the nodes $t_{j}$ are equispaced, say $t_{j}=t_{0}+j\Delta,\;j=1,\dots,s$, then the matrix $\mathbf{G}$ is the so-called prolate matrix, whose spectral properties are known exactly [43, 40]. Indeed, we have in this case

and therefore $\mathbf{G}=\frac{\pi}{\Omega\Delta}\boldsymbol{Q}(s,W)$ where $\boldsymbol{Q}(s,W)$ is the matrix defined in [40, eq. (21)]. The smallest eigenvalue of $\boldsymbol{Q}(s,W)$, denoted by $\lambda_{s-1}(s,W)$ in the same paper, has the exact asymptotics for $W$ small, given in [40, eqs. (64,65)]:

which gives

The same scaling was shown using Szego’s theory of Toeplitz forms in [15] – see also Subsection 3.2. The authors showed that there exist $C>0$ and $y^{*}>0$ such that for $\Omega\Delta<y^{*}$

To conclude the above discussion, defining the “super-resolution factor” as

we have that

It turns out that the bound (2.7) is too pessimistic if only some of the nodes are known to be clustered. Consider for instance the configuration $\boldsymbol{x}=\left(t_{1}=\Delta,\;t_{2}=2\Delta,\;t_{3}=-\frac{\pi}{4}\right)$, then, as can be seen in Figure 3.1, we have in fact $\lambda_{\min}\left(\mathbf{G}\left(\boldsymbol{x},\Omega\right)\right)\approx\left(\Delta\Omega\right)^{2}$, decaying much slower than $(\Delta\Omega)^{4}$ – which would be the bound given by (2.7).

In this paper we bridge this theoretical gap. We consider the partially clustered regime where at most $2\leq\ell\leq s$ neighboring nodes can form a cluster (there can be several such clusters), with two additional parameters $\rho,\tau,$ controlling the distance between the clusters and the uniformity of the distribution of nodes within the clusters.

The different parameters are illustrated in Figure 3.2.

Our main result is the following generalization of (2.7) for clustered configurations.

The proof of Theorem 3.2 is presented in Subsection 4.3 below. It is based on the “decimation-and-blowup” technique, previously used in the context of super-resolution in [1, 2, 6, 7, 8] and references therein. In a nutshell, the main idea is to choose an appropriate “decimation” parameter $\lambda\approx\Omega$ such that the “inflated” nodes in the vector $\lambda\boldsymbol{x}$ (considered in the wrap-around sense) are separated by $\lambda\Delta\approx\Omega\Delta$ from its cluster neighbors, and by a constant from the other nodes. Then we fix sufficiently large $N$ and divide the $2N+1$ rows of $\mathbf{V}_{N}$ into groups of $s$ rows, separated by $\lambda N\over\Omega$. Each of the resulting square Vandermonde matrices can be explicitly estimated (the inverses have well-known behaviour), and has smallest singular value of the order ${1\over\sqrt{N}}\left(\Delta\Omega\right)^{\ell-1}$. The main technical part is to show that such $\lambda$ exists, and it is proved in Lemma 4.1 by a union bound argument, showing that the measure of all “bad” values of $\lambda$ (causing a collision of at least two nodes) is small. The condition on $N$ in (3.2) is obtained by accurate counting of how many such “bad” intervals exist.

Our next result is the analogue of (3.2) for the Vandermonde matrix $\mathbf{V}_{N}$ as in (2.1), albeit under an extra assumption that the nodes are restricted to the interval $\frac{1}{s^{2}}\left(-\frac{\pi}{2},\frac{\pi}{2}\right]$.

Returning back to Theorem 3.2, it turns out that the bound (3.3) is asymptotically optimal.

The proof of Theorem 3.9 is presented in Subsection 4.4.

Finally we conclude with the optimal scaling for the condition number of $\mathbf{V}_{N}=\mathbf{V}_{N}\left(\boldsymbol{x},\Omega\right)$, which is of interest to some applications.

The problem of (sparse) super-resolution is to recover discrete, point-like objects from their noisy and bandlimited spectral measurements. It arises in many fields such as frequency estimation, sampling theory, array processing, astronomical imaging, seismic imaging, nonuniform FFT, statistics, radar signal detection, error correction codes, and others [4, 12, 13, 16, 10, 20, 28, 25, 35]. Our main results have direct implications for the problem of super-resolution under sparsity constraints, in the so-called ‘‘on-grid’’ model²² 2 Note that the results in the previous section are valid for “off-grid” setting, as the nodes $\{t_{j}\}$ can have arbitrary real values in $\left(-{\pi\over 2},{\pi\over 2}\right]$..

Consider the problem of reconstructing $\mu\in{\cal R}$ from approximate spectral data $\widehat{\mu}\left(\omega\right)$ restricted to some interval $\omega\in\left[-\Omega,\Omega\right]$. Here the Fourier transform $\widehat{\mu}$ is defined as

The measurement space $L_{2}\left(\left[-\Omega,\Omega\right]\right)$ contains complex-valued square-integrable functions supported on $\left[-\Omega,\Omega\right]$, with the norm

Proceeding as in [16, 15], we define the minimax error for this problem as follows.

Using arguments very similar to [33, 16, 15, 26] and the novel bounds of Theorem 3.2 and Theorem 3.9, we obtain the optimal rate for the minimax error for clustered on-grid super-resolution.

For the proof, see Subsection 4.5 below. This result generalizes [15, 26] (where the scaling ${\cal E}\asymp\mathrm{SRF}^{2\ell-1}\varepsilon$ was derived for $\ell=s$), as well as [33] (where it was shown that for positive $a_{j}$ it holds that ${\cal E}\lessapprox\mathrm{SRF}^{2\ell}\varepsilon$, with a comparable definition of the Rayleigh regularity $\ell$).

A different but closely related setting was considered in the seminal paper [16], where the measure $\mu$ was assumed to have infinite number of spikes on a grid of size $\Delta$, with one spike per unit of time on average, but whose local complexity was constrained to have not more than $R$ spikes per any interval of length $R$ (such $R$ is called the “Rayleigh index”). It was shown in [16] that the minimax recovery rate for such measures scales like $\mathrm{SRF}^{\alpha}$ where $2R-1\leq\alpha\leq 2R+1$. Our partial cluster model can therefore be regarded as the finite-dimensional version of these “sparsely clumped” measures with finite Rayleigh index, showing the same scaling of the error – polynomial in $\mathrm{SRF}$ and exponential in the “local complexity” of the signal.

If the grid assumption is relaxed, then one might wish to measure the accuracy of recovery $\|\widetilde{\mu}-\mu\|$ by comparing the locations of the recovered signal $\widetilde{\mu}$ with the true ones $\{t_{j}\}$. In this case, there are additional considerations which are required to derive the minimax rate, and it is possible to do so under the partial clustering assumptions. See [1, 8] for details, where we prove that ${\cal E}\asymp\mathrm{SRF}^{2\ell-1}\Delta\varepsilon$ in this scenario, for uniform bound on the noise $\|e\|_{\infty}:=\sup_{|\omega|\leq\Omega}\left|e\left(\omega\right)\right|\leq\varepsilon\lessapprox\mathrm{SRF}^{1-2\ell}$. The extreme case $\ell=s$ has been treated recently in [6, 7].

In the case of well-separated spikes (i.e. clusters of size $\ell=1$), a recent line of work using $\ell_{1}$ minimization ([12, 11, 17, 14] and the great number of follow-up papers) has shown that the problem is stable and tractable.

Therefore, the partial clustering case is somewhat mid-way between the extremes $\ell=1$ and $\ell=s$, and while our results in this paper (and also in [8]) show that it is much more stable than in the unconstrained sparse case, it is an intriguing open question whether provably tractable solution algorithms exist.

Several candidate algorithms for sparse super-resolution are well-known – MUSIC, ESPRIT/matrix pencil, and variants; these have roots in parametric spectral estimation [42]. In recent years, the super-resolution properties of these algorithms are a subject of ongoing interest, see e.g. [18, 29, 38, 26, 27] and references therein. Smallest singular values of the partial Fourier matrices $\mathbf{V}_{N}$, for finite $N$, play a major role in these works, and therefore we hope that our results and techniques may be extended to analyze these algorithms as well.

Here we introduce the uniform blowup of a node vector $\boldsymbol{x}=\left(t_{1},\dots,t_{s}\right)$ by a positive parameter $\lambda$, and study the effect of such a blowup mapping on the minimal wrap-around distance between the mapped nodes.