Date of Award
Doctor of Philosophy (PhD)
Low dimensional signal processing has drawn an increasingly broad amount of attention in the past decade, because prior information about a low-dimensional space can be exploited to aid in the recovery of the signal of interest. Among all the different forms of low di- mensionality, in this dissertation we focus on the synthesis and analysis models of sparse recovery. This dissertation comprises two major topics. For the first topic, we discuss the synthesis model of sparse recovery and consider the dictionary mismatches in the model. We further introduce a continuous sparse recovery to eliminate the existing off-grid mismatches for DOA estimation. In the second topic, we focus on the analysis model, with an emphasis on efficient algorithms and performance analysis.
In considering the sparse recovery method with structured dictionary mismatches for the synthesis model, we exploit the joint sparsity between the mismatch parameters and original sparse signal. We demonstrate that by exploiting this information, we can obtain a robust reconstruction under mild conditions on the sensing matrix. This model is very useful for
radar and passive array applications. We propose several efficient algorithms to solve the joint sparse recovery problem. Using numerical examples, we demonstrate that our proposed algorithms outperform several methods in the literature. We further extend the mismatch model to a continuous sparse model, using the mathematical theory of super resolution. Statistical analysis shows the robustness of the proposed algorithm. A number-detection algorithm is also proposed for the co-prime arrays. By using numerical examples, we show that continuous sparse recovery further improves the DOA estimation accuracy, over both the joint sparse method and also MUSIC with spatial smoothing.
In the second topic, we visit the corresponding analysis model of sparse recovery. Instead of assuming a sparse decomposition of the original signal, the analysis model focuses on the existence of a linear transformation which can make the original signal sparse. In this work we use a monotone version of the fast iterative shrinkage- thresholding algorithm (MFISTA) to yield efficient algorithms to solve the sparse recovery. We examine two widely used relaxation techniques, namely smoothing and decomposition, to relax the optimization. We show that although these two techniques are equivalent in their objective functions, the smoothing technique converges faster than the decomposition technique. We also compute the performance guarantee for the analysis model when a LASSO type of reconstruction is performed. By using numerical examples, we are able to show that the proposed algorithm is more efficient than other state of the art algorithms.
Martin Arthur, Viktor Gruev, Jr-Shin Li, Arye Nehorai, Heinz Schaettler
Permanent URL: https://doi.org/10.7936/K7HX19VH