In this paper, we study sparse spike deconvolution over the space of complex-valued measures when the input measure is a finite sum of Dirac masses. We introduce a modified version of the Beurling Lasso (BLasso), a semi-definite program that we refer to as the Concomitant Beurling Lasso (CBLasso). This new procedure estimates the target measure and the unknown noise level simultaneously. Contrary to previous estimators in the literature, theory holds for a tuning parameter that depends only on the sample size, so that it can be used for unknown noise level problems. Consistent noise level estimation is standardly proved. As for Radon measure estimation, theoretical guarantees match the previous state-of-the-art results in Super-Resolution regarding minimax prediction and localization. The proofs are based on a bound on the noise level given by a new tail estimate of the supremum of a stationary non-Gaussian process through the Rice method.More precisely, pioneering works were proposed in [12] treating inverse problems on the space of Borel measures and in [13], where the Super-Resolution problem was investigated via Semi-Definite Programming and a groundbreaking construction of a "dual certificate". Exact recovery (in the noiseless case), minimax prediction and localization (in the noisy case) have been performed using the Beurling Lasso (BLasso) estimator [2,38,25,37] which minimizes the total variation norm over complex-valued Borel measures. Noise robustness (as the noise level tends to zero) has been thoroughly investigated in [22]; the reader may also consult [23,20,24] for more details. Change point detection and grid-less spline decomposition are studied in [7,19]. Several interesting extensions, such as deconvolution over spheres, have been also recently provided in [6,9,8]. For more general settings, we refer to the work of [30].
Concomitant Beurling Lasso: adapting to the noiseOur proposed estimator is an adaptation to the Super-Resolution framework of a methodology first developed for sparse high dimensional regression. In the latter case, the joint estimation of the parameter and of the noise level has first been considered in [33,1], though without any theory. It was based on concomitant estimation ideas that could be traced back to the work of Huber [29]. The formulation we consider in this work appeared first in [33, 1] with a statistical point of view, as well as in [40] with a game theory flavor. Note that interestingly, both approaches rely on the notion of robustness. An equivalent definition of this estimator was proposed and extensively studied independently in [5] under the name Square-root Lasso. The formulation we investigate is also closer to the one analyzed in [36] under the name Scaled-Lasso. Yet, we adopt the terminology of "Concomitant Beurling Lasso" in reference to the seminal paper [33]. Last but not least, our contribution borrows some ideas from the stimulating lecture notes [39].Remark that an alternative formulation was investigated in [35] with a particular aim at Gaussian m...