Multi-kernel unmixing and super-resolution using the modified matrix pencil method

, we derive an algorithm for this problem for estimating the source parameters of each group, along with precise non-asymptotic guarantees. Our approach involves estimating the group parameters sequentially in the order of increasing scale parameters, i.e., from group

1

to

L

. In particular, the estimation process at stage

1 l e q l l e q L

involves (i) carefully sampling the tail of the Fourier transform of

y

, (ii) a emph{deflation} step wherein we subtract the contribution of the groups processed thus far from the obtained Fourier samples, and (iii) applying Moitra's modified Matrix Pencil method on a deconvolved version of the samples in (ii).

Review text

In this excellent paper, the authors consider a generalization of the standard super-resolution problem, the so-called \textit{multi-kernel unmixing super-resolution problem}. They consider signals \[ x(t) = \sum_{\ell = 1}^L x_{\ell}(t)\,, \quad t\in \mathbb R\,, \] where \[ x_{\ell}(t) = \sum_{j=1}^K u_{\ell,j}\, \delta(t - t_{\ell,j})\,, \quad \ell=1,\ldots ,L\,, \] are streams of Dirac's with locations \(t_{\ell,j}\in [0,\,1)\) and coefficients \(u_{\ell,j}\in \mathbb C\). Let \(g_{\ell}(t) = g(t/\mu_{\ell})\), \(\ell=1,\ldots, L\) be Gaussian kernels with \(g(t) = \exp(- t^2/2)\) and positive variance parameters \(\mu_1 < \ldots < \mu_L\). Let \(f(s) = (Fy)(s)\) be the Fourier transform of \[ y(t) = \sum_{\ell = 1}^L (g_{\ell} \ast x_{\ell})(t) \] such that \[ f(s) = \sum_{\ell = 1}^L (F g_{\ell})(s)\, (F x_{\ell})(s)\,. \] Then the aim of the multi-kernel unmixing super-resolution problem is to recover estimates of all locations \(t_{\ell,j}\) and coefficients \(u_{\ell,j}\) from few noisy samples of \(f(s)\). This problem has important practical applications such as in single molecule spectroscopy, spike sorting in neuron activity analysis, DNA sequencing, or spike hunting in galaxy spectra. The main idea of the algorithm is based on stepwise computation of the locations and coefficients for \(\ell = 1,\ldots,L\). In the first step \(\ell = 1\) for noiseless data, one uses the fact that \(f(s)/(Fg_1)(s)\) is approximately equal to the exponential sum \[ (F x_1)(s) = \sum_{j=1}^K u_{1,j}\, \exp(2\pi {\mathrm i}\,t_{1,j}\,s)\,, \quad s \ge s_1 \gg 1\,. \] Using a modified matrix pencil method, one can determine \(t_{1,j}\) and \(u_{1,j}\) for \(j=1,\ldots, K\). In the second step \(\ell = 2\), one considers \(f(s) - (F g_1)(s)\, (F x_1)(s)\) for \(s\ge s_2> s_1\) and computes \(t_{2,j}\) and \(u_{2,j}\) for \(j=1,\ldots, K\) by a similar technique and so on. A highlight of these results is the fact that in all estimates the authors present explicit bounds. In many numerical experiments, signals with \(L=4\) and \(K\in \{2,\,3,\,4,\,5\}\) are considered.

Recommendations

Cites work

Cited in

(4)

This page was built for publication: Multi-kernel unmixing and super-resolution using the modified matrix pencil method

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2291571)