Block‐sparse Signal Recovery Via Minimisation Method
Wendong Wang,Jianjun Wang,Zili Zhang
DOI: https://doi.org/10.1049/iet-spr.2016.0381
IF: 1.819
2018-01-01
IET Signal Processing
Abstract:IET Signal ProcessingVolume 12, Issue 4 p. 422-430 Research ArticleFree Access Block-sparse signal recovery via minimisation method Wendong Wang, Wendong Wang School of Computer and Information Science, Southwest University, Chongqing, People's Republic of ChinaSearch for more papers by this authorJianjun Wang, Jianjun Wang School of Mathematics and Statistics, Southwest University, Chongqing, People's Republic of ChinaSearch for more papers by this authorZili Zhang, Corresponding Author Zili Zhang zhangzl@swu.edu.cn School of Computer and Information Science, Southwest University, Chongqing, People's Republic of China School of Information Technology, Deakin University, Locked Bag, 20000 Geelong, VIC, AustraliaSearch for more papers by this author Wendong Wang, Wendong Wang School of Computer and Information Science, Southwest University, Chongqing, People's Republic of ChinaSearch for more papers by this authorJianjun Wang, Jianjun Wang School of Mathematics and Statistics, Southwest University, Chongqing, People's Republic of ChinaSearch for more papers by this authorZili Zhang, Corresponding Author Zili Zhang zhangzl@swu.edu.cn School of Computer and Information Science, Southwest University, Chongqing, People's Republic of China School of Information Technology, Deakin University, Locked Bag, 20000 Geelong, VIC, AustraliaSearch for more papers by this author First published: 01 June 2018 https://doi.org/10.1049/iet-spr.2016.0381Citations: 10AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Motivated by the recently emerged method for sparse signal recovery, in this study, the authors make an ongoing effect to extend this methodology to the setting of block sparsity, which directly leads to the proposed method for block-sparse signal recovery. Some theoretical results are induced to guarantee the validity of proposed method. In particular, the obtained recovery condition rigorously includes the one induced by Yin et al., and the obtained error estimate can be used to model both the (block-) sparse and non-sparse signals, which is more comprehensive than that induced by Yin et al. which applies only to the sparse signals. The authors also derive an alternating direction method of multipliers (ADMM)-based algorithm to tackle the induced optimisation problem. Some experimental results that are based on the synthetic block-sparse signals and the real-world foetal electrocardiogram signals further demonstrate the better performance of the method when it is compared with the state-of-the-art group-lasso method and method for 0 < q < 1. 1 Introduction Over last 10 years, abundant literature both in the theoretical and applied fields have witnessed the overwhelming growth and development of compressed sensing (CS) pioneered by Donoho, Candès and their collaborators [1-3]. So far CS has not only been regarded as an alternative to Nyquist-Shannon sampling theorem for signal processing, but has also gradually become a powerful tool to solve many other practical problems such as image processing [4], face recolonisation [5], subspace clustering [6] and even the Sudoku [7], a world-wide popular number game which is usually solved by hand for fun. In short, CS mainly aims to recover a sparse signal from a small set of compressed measurements , where with and are the measurement matrix and additive noise, respectively. Mathematically, it can be realised by solving an -norm minimisation where bounds the -norm of noise and is determined in advance. As is known to all, the success of CS heavily relies on the sparsity assumption on signals. However, some signals in real world may exhibit additional sparse structures. A typical example is the block-sparse (also named group-sparse) structure, i.e. the non-zero elements in such a signal are assembled in few fixed blocks, which is also our main concern in this paper. Such structured sparse signals naturally arise in various applications. Prominent examples include foetal electrocardiogram (FECG) [8], colour image [4], motion segmentation [9] and machine learning [10]. Without loss of generality, we assume there are m blocks with block size in x. Then, we can write the signal x as where denotes the i th block of x. If x has at most s non-zero blocks, we refer to such vector x as block s -sparse signal. Accordingly, the measurement matrix can also be written as where and denote the i th column vector and the j th sub-block matrix of , respectively. To tackle such block-sparse signals, a block version of the -norm minimisation method, i.e. the mixed -norm [Note that the mixed -norm defined in multiple measurement vectors (MMVs) community also has notion . However, these two kinds of norms have totally different input requirements and definitions. See, e.g. [11] for more details on MMVs case.] minimisation method, was studied in [12-14], which solves (1)The method was intended for the recovery of block-sparse signals and its mechanism is using the and norms simultaneously. Specifically, the -norm characterises the intra-block cooperation in each , and the -norm characterises the inter-block sparsity in newly formed vector . Some typical algorithms to solve (1) include group-lasso [12], block-BP [13] and so on. Inspired by the success of non-convex method [15, 16] in conventional CS, some researchers later investigated a mixed method [4, 17-20] by solving The method well inherits the strong sparsity-promoting capacity of the method, and hence leads to a better performance than previous method in dealing with the block-sparse signals. In order to investigate the theoretical performance of both the and methods, some theoretical tools were developed. Among these, block restricted isometry property, i.e. block-RIP (see Definition 1) is the most popular and widely used one. Block-RIP in fact can be regarded as an extended version of the RIP. It is known in conventional CS that the RIP has always played a critical role in establishing the theoretical analysis of sparse signal recovery and this point can be supported by large amount of literature (see [21-24] and more). Although random Gaussian matrices and random Bernoulli matrices have been proved to possess RIP with high possibility, however, in terms of the coherence of matrix [1, 25], which is defined as they often have small coherence [To the best of the authors' knowledge, no literature has established the strict mathematical characterisation between coherence and RIP, and a recent discussion on their relations can be found in [25].]. Considering that many problems in practice, such as electroencephalography source localisation [26], direction-of-arrival estimation [27] and radar detection [28], may suffer a highly coherent measurement matrix, a natural question is: can conventional and methods still efficiently tackle the sparse signal recovery problems when highly coherent measurement matrices are confronted? The recent work initiated by Yin et al. [1] provided a negative answer to the question above, and instead they proposed an method to deal with such cases. They have proved theoretically that when measurement matrix obeys some conditions related to RIP, their method can exactly recover any sparse signal. Besides, their simulation results also further show that such novel method has the superior performance than previous and methods [29, 30] when highly coherent measurement matrices are confronted. Unfortunately, this method is not applicable for block-sparse signals due to the lack of consideration of their block structures. In this paper, we present an ongoing effort to extend the method to the setting of block-sparsity, which directly leads to our method (2)From the perspective of block coherence of matrix (see Definition 2), our method is totally different from previous and methods which heavily rely on the measurement matrices with small block coherence to function well, and hence offers an addition to the existing block-sparse frames. The main contributions of this paper can be summarised as follows: We extend the method for sparse signal recovery to the setting of block-sparsity, and hence propose an method for block-sparse signal recovery, which expands the potential applications of method. We theoretically investigate the proposed method, and the obtained recovery condition and the error estimate are proved to better than those induced for previous method in [1]. We derive an ADMM-based algorithm to solve our induced optimisation problem, and also compare it with some state-of-the-art methods to recover the synthetic block-sparse signals and the real-world FECG signals. The obtained results further confirm the good performance of our proposed method, specifically when it is used to deal with the highly block-coherent measurement matrices. The rest of this paper is organised as follows. In Section 2, we first introduce some preliminaries and then present two main theorems, one for block-structured signals recovery and another for construction of the highly block-coherent matrices. An efficient ADMM-based algorithm for solving the induced optimal problem is presented in Section 3. In Section 4, we conduct some experiments to support the validity of our method. The conclusions and future work are addressed in Section 5. 2 Preliminaries and main results We begin with introducing some preliminaries, which will be used later. Definition 1.(block-RIP): A measurement matrix is said to satisfy the block-RIP with constant if is the smallest value such that holds for all vectors that are block s-sparse. Definition 2.(block coherence): Given matrix , we define its block coherence as As an extreme case of block partition, i.e. d = 1 or m = N, block-RIP and block coherence will naturally degenerate to the conventional RIP and coherence, respectively. In such context, will be denoted by for convenience. In addition, given any index set and vector , we define as the complement of E in , the support of x as , and we also define the best s -block approximation of x as where counts the number of elements in . Furthermore, will represent the restriction of x to the index set E, which is defined as when and otherwise. Besides, we also need following lemma. Lemma 1.For any vector that has s non-zero blocks, we have Proof.By constructing a map with , the proof of Lemma 1 can be attributed to that of following conclusion: holds for all . Since above conclusion has been proved to be true in Lemma 2.1(b) in [1], so we complete the proof of our Lemma 1. □ With preparations above, we now present our main results. Theorem 1.Let . If the measurement matrix satisfies block-RIP with (3)then for any x satisfying and the solution to (2), we have (4)where and is defined in (16). Remark 1.Theorem 1 provides a sufficient condition for recovery of the block-structured signals via method. The obtained results not only extend the sparse recovery to the setting of block-sparsity, but also provide a more complete error estimation (4) since a best s -block approximation error is well embedded in the upper bound on error, which makes our error estimate become available to both the (block-)sparse and non-sparse signals. In contrast, the error estimate for method induced in [1] applies only to the sparse signals. What is more, if one sets , our condition (3) will directly degenerate to (5)Note that Yin et al. [1] also provided a similar condition, which takes the form (6)where and . Actually, the condition (6) is rigorously included in our condition (5), or in other words, our condition (5) is much easier to be satisfied than condition (6), and one can find the detailed proof in [31]. For completeness, we prove it as follows. Due to (6) and its deduced inequality , we directly have holds for all . Remark 2.Theorem 1 indicates that method can exactly recover any block s -sparse signals with in the absence of noise. i.e. when (3) holds. Compared with some existing recovery conditions based on block-RIP such as [14, 32] for method and [17, 18] for method, condition (3) for our method might be much too stringent. However, one will see later in experimental part that our method is comparable to the method when general random Gaussian measurement matrices are considered, and is superior to both the group-lasso and methods for block-sparse signal recovery in face of a highly block-coherent measurement matrix. In addition, since some inequalities in the proof of Theorem 1 such as (10) are not sharp, the resulting conclusions such as the sufficient condition (3) and the coefficients and in (4) become relatively stringent. However, it can be expected that these problems can be well addressed in future when sharper inequalities are induced. Proof.We start with denoting and . Naturally, we have . Since is the solution to (2), a direct result is (7)As to the left hand of (7), we have (8)As to the right hand of (7), we have (9)Combining (8) and (9), we can further simplify (7) into (10)Next, we decompose into a series of vectors such that where corresponds to the indices of the s largest -norm blocks of ; to the indices of the next s largest -norm blocks of and so on. Then, for every , from Lemma 1, it follows that Since every has at most s blocks, then we have Therefore (11)where we have used the inequalities Denoting and applying the fact together with (10) and (11), yield (12)Applying the triangle inequality, we have This, together the block-RIP, gives Using the block-RIP again, we directly have (13)and Now, we have (14)Taking (12) and (13) into consideration, we can further simplify (14) into (15)where (16)Combining (15) and (12), together with the assumption (3), gives where and are defined in Theorem 1. □ By means of the well-known block-RIP, Theorem 1 offers a theoretical frame for our method to recover the block-structured signals. In next section, we will do some experiments to conform the performance of our method when both the general random Gaussian measurement matrices and the highly block-coherent measurement matrices are confronted. Following theorem provides a way to generate the highly block-coherent matrices easily. Theorem 2.Given any non-zero matrix , we generate matrix where is an orthogonal matrix. Then, as to the block coherence of , we have (a) (b) Proof.The left inequality in (a) is from the Theorem 2.3 in [33], and the right one in (a) is a direct result of Cauchy–Schwarz inequality. Since (b) is easy to prove, we leave it to the interested readers. □ Remark 3.Theorem 2 states that the construction of any highly block-coherent matrices can be attributed to the conventional highly coherent matrix problems. In addition, considering that the coherence of matrix is ≤1, naturally the matrix whose block coherence tends to be 1 will be viewed as a highly block-coherent matrix. 3 Algorithm implementation of method In this section, we will discuss some implementation aspects of the optimisation problem given in (2). Due to the non-convexity of the objective function and the difficulty in estimating properly, problem (2) cannot be solved easily and conveniently. Using the popular penalty function method [34], we first present an unconstrained version of (2), i.e. (17)where is a regularisation parameter. Inspired by the ideas in [1], we will approximate iteratively the solution to (17) by solving following problem in order: (18)where when and otherwise. In fact, such approximation strategy is feasible, since using a similar argument as that in [1] one can prove that any non-zero limit point of sequence generated iteratively from (18) is a stationary point of (17). To solve (18), we resort to an ADMM algorithm [30]. We start with reformulating it as (19)and then we can form the augmented Lagrangian of (19) as where is a penalty parameter and z is the Lagrange multiplier. Now, the ADMM procedure for (19) can be presented as (20) (21) Specifically, we have Algorithm 1 (see Fig. 1). Fig. 1Open in figure viewerPowerPoint Algorithm 1: solving minimisation (2) via ADMM In practice, the factors affecting the convergence of Algorithm 1 (Fig. 1) mainly come from the stopping criterions for both the inner and outer loops and the parameters in these two loops. Following the suggestions in [1, 30], we stop the inner loop, i.e. the ADMM procedure, when or where , . As to the stopping criterion for the outer loop, we will discuss it in the next section. 4 Numerical experiments In this section, we compare our method with method [1], group-lasso method [30] and the method [17] to recover the (nearly) block-sparse signals. Two sorts of signals including the synthetic block-sparse signals and the real-world FECG signals (which can be viewed as nearly block-sparse signals) are used as the test signals. In addition, we set the regularisation parameter and for group-lasso method, and use their default settings for other methods without specification. For our Algorithm 1 (Fig. 1), we always set . Throughout the experiments, we generate the block s -sparse signals as follows: (i) s block indices are chosen from uniformly at random; (ii) the entries restricted to these block indices are generated from a standard Gaussian distribution and the entries out of the selected s block indices are set to zero. Two kinds of matrices are used as the measurement matrices for synthetic block-sparse signal recovery. They are random Gaussian matrices and synthetic highly block-coherent matrices. Let H be a Hadamard matrix of order d, according to Theorem 2, we construct the highly block-coherent matrices by using where and is a randomly oversampled partial discrete cosine transform (DCT) matrix with here is the i th column vector of matrix , and whose components are uniformly and independently sampled from [0, 1]. The number F is a parameter controlling the coherence performance of P. Fig. 2 plots the obtained block-coherence result on such with and . Fig. 2Open in figure viewerPowerPoint Parameter F versus block coherence of matrix It is easy to see that tends to get larger as F increases. Since a bigger F will lead to a bigger block coherence of , for convenience we will use F to represent the block coherence when confusion is not caused. All the experiments are implemented in Matlab and the codes will be released on webpage: https://github.com/DongSylan/Block-L12-ADMM. 4.1 Results on random Gaussian matrices We start with the comparison of synthetic block-sparse signal recovery by using the random Gaussian matrices as the measurement matrices. During this sort of experiments, we set , , and chose . Before moving on, we have to determine a proper regularisation parameter since it plays a vital role in our Algorithm 1 (Fig. 1). Fig. 3 a plots the and relative error (RelError) results obtained by our Algorithm 1 (Fig. 1) for recovering the block 12-sparse signals with block size . Here RelError is defined as where and are the input and output signals. It is easy to see that is relatively suitable. By fixing , we further study the convergence of our Algorithm 1 (Fig. 1) when block 12-sparse signals with block size are considered. The obtained results can be found in Fig. 3 b, where . It is easy to see that, our method converges fast and its accuracy is higher than that of method. Without loss of generality, we uniformly set their outer-loop stopping criterion as or . With above preparations, we now turn to compare our method with method, group-lasso method and the method with , 0.5, 0.8 in recovering the block-sparse signals. By fixing the block size , Fig. 4 a plots the block-sparsity and success rate results obtained by these methods. If the relative error satisfies , the recovery is considered a success. One can directly see that, (i) method with performs best, and our method ranks the second, followed by the method with , 0.2 and group-lasso method; (ii) all the block-structured methods perform much better than non-block-structured method. In fact, deep exploration on block structure of sparse signals helps to largely improve the performance of existing methods. To confirm this conclusion, by fixing the number of non-zero elements of signal to be recovered as 64, we investigate the success rates influenced by block s -sparse signals with different block size d. Obviously, in such case, we have . It is easy to see from Fig. 4 b that, an increasing block size d, or in other words a decreasing block-sparsity , often leads to a better performance of all the block-structured methods. However, such change in block-sparse signals has nothing to do with method. In summary, our method is not the best one, but it is still comparable to the method. In fact, when highly block-coherent measurement matrices are confronted, our method will give full play to its advantages. Fig. 3Open in figure viewerPowerPoint selection and convergence analysis for Algorithm 1 (Fig. 1) (a) selection, (b) Convergence analysis Fig. 4Open in figure viewerPowerPoint Block-sparse signal recovery with random Gaussian matrices (a) Block-sparsity versus success rate, (b) Block size versus success rate 4.2 Results on block-coherent matrices In this part, some experiments based on the block-coherent matrices are conducted to verify the performance of our method. During these experiments, we always set , , and chose . First, we investigate the performance of methods influenced by measurement matrices with different block coherences. Here block 8-sparse signals with block size are used as the test signals. One can directly see in Fig. 5 a that, with the increasing of F, or say block coherence, the recovery performance of all the methods tends to be poor except the method, however, our method outperforms the rest of the methods when F is fixed. In addition, it is notable that method performs better when F turns bigger, but far worse than the rest block-structured methods when F tends to be smaller. Actually, the performance of our method can be further improved provided the spikes of a block-sparse signal are sufficient separated. Specifically, a block s -sparse signal x is called as a spiked block s -sparse signal if and only if the elements in satisfy for certain positive integer L. Here we also parallel call L as the minimum separation [1] which characters a signal's block-sparsity distribution, and we also set . Fig. 5 b plots the recovery performance for such spiked block 8-sparse signals with different block-coherent matrices. It is easy to see that our method is not sensitive to the measurement matrices with different block coherences, and therefore nearly successfully recovers all the signals. However, the rest methods show quite different performance: the performance of method is easy to be affected by both F and q, i.e. a bigger F, a worse performance, however a bigger q, a better performance; as to group-lasso method, its performance declines when F increases; the performance of method here is consistent with the one concluded from Fig. 5 a. Without loss of generality, we set and for the rest of the experiments, which is sufficient to guarantee a highly block-coherent matrix and a spiked block-sparse signal. Fig. 5Open in figure viewerPowerPoint Block coherence versus success rate with block 8-sparse signals (a) General block-sparse case, (b) Spiked block-sparse case In what follows, we consider the recovery of block-sparse signals by using the highly block-coherent measurement matrices. One can see the corresponding results with different block sizes from Fig. 6 directly. It shows that, although block sizes are chosen from set {2,4,8} by order, the obtained results share some common characteristics. First, with the increasing of the block-sparsity, a fixed measurement number M becomes more and more unable to meet the minimal requirements of CS theory to guarantee the exact recovery, and hence all the methods tend to be poor. However, our method still outperforms the rest of the methods in the case that block size d is fixed. Second, the performance of method heavily relies on q, and a bigger q is recommended to promote the success rate. Third, the block-structured methods including group-lasso method, method and method can perform much better than conventional method when block size tends to be larger. Meanwhile there still exist some significant differences. First, when block size , method outperforms the method with , 0.5, and group-lasso method, however is worse than method with and our method. Second, with the increasing of block size, method tends to be the worst one. Third, proposed selection of the q will make method perform better than group-lasso method and an example can be found in Fig. 6 b. The claim that the performance of our method can be further improved in the case that block-sparse signals have spike structure is proved once again in Fig. 7. Compared with the corresponding results in Fig. 6, this conclusion is obvious. It is notable that under such situation, our method performs much better than rest methods when the block size of signals increases. Fig. 6Open in figure viewerPowerPoint Success rates with different block sizes for block-sparse signal recovery (a) , (b) , (c) Fig. 7Open in figure viewerPowerPoint Success rates with different block sizes for spiked block-sparse signal recovery (a) , (b) , (c) Since above experiments are conducted in the noiseless cases, in order to verify the robustness of above methods, we test them to recover the noise signals. To this end, we first generate the test signal x by creating a spiked block 4-sparse signal x with block size , and then synthesise noisy with . We present the signal-to-noise ratio (SNR) results in Table 1. It is easy to see that, with the increasing of (noise level), the performance of all the methods tends to be weak, however our method almost has the best performance (except ). Our SNR result in fact is not the best one when , but it is far better than the method with and the method. Table 1. SNR results with different noise levels by different methods 3.60 −0.62 −3.43 −4.98 −5.91 8.23 1.89 −1.50 −3.01 −4.16 13.34 5.00 1.15 −0.32 −2.12 group-lasso 15.16 7.78 4.54 1.85 −0.02 3.43 0.84 −0.87 −2.35 −3.25 18.28 10.08 5.00 2.20 −0.37 The bold values indicate the biggest values which corresponds to the best SNR performance. 4.3 Results on FECG signals To further testify the performance of our method in some real-world applications, we apply our method, together with the methods used above to recover the FECG signals [35]. In fact, this kind of signals have been investigated by CS and application communities for many purposes, see for example [8] and the references within. Fig. 8 plots a segment of such FECG signals. In this segment, the parts from 20 to 60 and from 200 to 250 time points can be viewed as two significant non-zero blocks, and other parts can be viewed as concatenations of zero blocks. Roughly, this segment can be viewed as a block 2-sparse signal (corrupted by signal noise). Generally, the location of non-zero elements in FECG is not known in advance, so it becomes difficult to use the block-structured methods directly. To tackle this problem, similar with [8], we assume this segment is evenly divided into 10 blocks with block size . Besides, as [8] did in each experiment, we also use the same matrix as the measurement matrix. This measurement matrix is generated by randomly creating a sparse binary matrix of the size . Its each column contains 15 entries of 1 s, while other entries are zero. To boost the best performance of above methods, we let their regularisation parameters be chosen from , and output their best recovered signals (judged by RelError) as the final results. Fig. 9 plots the results obtained by method, method with , group-lasso method and our method. It is easy to see that, both our method and the method with perform much better than other methods, and are very close to the original segment. However, from the viewpoint of relative error, our method shows slightly better than method with . Next, we use the same measurement matrix to compress the whole FECG signals shown in Fig. 10 to verify the ability of our method. Since each FECG signal shown in Fig. 10 is 2500-dimensional, we first evenly divide each of them into ten segments and then recover those segments in order. The obtained results are presented in Table 2. One can see that our method ranks first, followed by the method for , group-lasso method and the method in order, which conforms again the performance of our method. Fig. 8Open in figure viewerPowerPoint Segment from FECG signals Fig. 9Open in figure viewerPowerPoint Recovered results by different methods (a) method with RelError = 0.62, (b) Group-lasso method with RelError = 0.40, (c) method (with ) with RelError = 0.31, (d) method with RelError = 0.27 Fig. 10Open in figure viewerPowerPoint Whole FECG signals Table 2. RelError results by different methods (a) (b) (c) (d) (e) (f) (g) (h) 0.54 0.45 0.50 0.72 0.45 0.36 0.39 0.38 group-lasso 0.45 0.34 0.38 0.49 0.34 0.29 0.30 0.32 0.43 0.30 0.35 0.49 0.30 0.25 0.26 0.28 0.41 0.28 0.34 0.48 0.29 0.23 0.25 0.27 The bold values indicate the smallest values which corresponds to the best relative error performance. 5 Conclusions and future work In this paper, we propose an method for block-sparse signal recovery, and also establish some theoretical analysis for proposed method including recovery condition and error estimate, which are proved to be better than those obtained in [1] for previous method. Our numerical experiments that are based on the synthetic block-sparse signals and the real-world FECG signals further confirm the superior performance of proposed methods in dealing with block-sparse signals. Potential future work includes characterising the method and our method with some non-RIP tools such as coherence [23, 36], and applying our method to solve some practical problems related to highly block-coherent measurement matrices and/or block-sparse signals. 6 Acknowledgments This work was supported by Natural Science Foundation of China under grant number 61273020, Fundamental Research Funds for the Central Universities under grant number XDJK2015A007 and National High Technology Research and Development Program 863 under grant number 2013AA013801. 7 References 1Yin, P.H., Lou, Y.F., He, Q. et al: 'Minimization of ℓ1-2 for compressed sensing', SIAM J. Sci. Comput., 2015, 37, (1), pp. A536– A563 2Donoho, D.L.: 'Compressed sensing', IEEE Trans. Inf. Theory, 2006, 52, (4), pp. 1289– 1306 3Candès, E.J., Romberg, J., Tao, T.: 'Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information', IEEE Trans. Inf. Theory, 2006, 52, (2), pp. 489– 509 4Majumdar, A., Ward, R.K.: 'Compressed sensing of color images', Signal Process., 2010, 90, (12), pp. 3122– 3127 5Wright, J., Yang, A.Y., Ganesh, A. et al: 'Robust face recognition via sparse representation', IEEE Trans. Pattern Anal. Mach. Intell., 2009, 31, (2), pp. 210– 227 6Patel, V.M., Nguyen, H.V., Vidal, R.: 'Latent space sparse and low-rank subspace clustering', IEEE J. Sel. Top. Signal Process., 2015, 9, (4), pp. 691– 701 7Babu, P., Pelckmans, K., Stoica, P. et al: 'Linear systems, sparse solutions, and Sudoku', IEEE Signal Process. Lett., 2010, 17, (1), pp. 40– 42 8Zhang, Z., Jung, T.P., Makeig, S. et al: 'Compressed sensing for energy-efficient wireless telemonitoring of noninvasive fetal ECG via block sparse Bayesian learning', IEEE Trans. Biomed. Eng., 2013, 60, (2), pp. 300– 309 9Vidal, R., Ma, Y.: 'A unified algebraic approach to 2-D and 3-D motion segmentation and estimation', J. Math. Imag. Vis., 2006, 25, (3), pp. 403– 421 10Huang, S., Yang, Y., Yang, D. et al: 'Class specific sparse representation for classification', Signal Process., 2015, 116, pp. 38– 42 11Steffens, C., Pesavento, M., Pfetsch, M.E.: ' A compact formulation for the ℓ2,1 mixed-norm minimization problem', arXiv preprint arXiv: 1606.07231, 2016 12Yuan, M., Lin, Y.: 'Model selection and estimation in regression with grouped variables', J. R. Statist. Soc., 2006, 68, (1), pp. 49– 67 13Van Den Berg, E., Friedlander, M.P.: 'Probing the pareto frontier for basis pursuit solutions', SIAM J. Sci. Comput., 2008, 31, (2), pp. 890– 912 14Eldar, Y.C., Mishali, M.: 'Robust recovery of signals from a structured union of subspaces', IEEE Trans. Inf. Theory, 2009, 55, (11), pp. 5302– 5316 15Chartrand, R.: 'Exact reconstruction of sparse signals via nonconvex minimization', IEEE Signal Process. Lett., 2007, 14, (10), pp. 707– 710 16Xu, Z.B., Guo, H.L., Wang, Y. et al: 'Representative of L1/2 regularization among Lq (0 < q ≤ 1) regularizations: an experimental study based on phase diagram', Acta Autom. Sin., 2012, 38, (7), pp. 1225– 1228 17Wang, Y., Wang, J.J., Xu, Z.B.: 'On recovery of block-sparse signals via mixed ℓ2 /ℓq (0 < q ≤ 1) norm minimization', EURASIP J. Adv. Signal Process., 2013, 2013, (1), pp. 1– 17 18Yin, H.T., Li, S.T., Fang, L.Y.: 'Block-sparse compressed sensing: non-convex model and iterative re-weighted algorithm', Inverse Probl. Sci. Eng., 2013, 21, (1), pp. 141– 154 19Wang, Y., Wang, J.J., Xu, Z.B.: 'Restricted p -isometry properties of nonconvex block-sparse compressed sensing', Signal Process., 2014, 104, pp. 188– 196 20Wang, J.J., Zhang, J., Wang, W.W. et al: 'A perturbation analysis of nonconvex block-sparse compressed sensing', Commun. Nonlinear Sci. Numer. Simul., 2015, 29, (1), pp. 416– 426 21Candès, E.J.: 'The restricted isometry property and its implications for compressed sensing', C. R. Math. Acad. Sci. Paris, Ser. I, 2008, 346, (9), pp. 589– 592 22Mo, Q., Li, S.: 'New bounds on the restricted isometry constant δ2k', Appl. Comput. Harmon. Anal., 2011, 31, (3), pp. 460– 468 23Foucart, S., Rauhut, H.: ' A mathematical introduction to compressive sensing' ( Springer, New York, USA, 2013) 24Cai, T.T., Zhang, A.R.: 'Sparse representation of a polytope and recovery of sparse signals and low-rank matrices', IEEE Trans. Inf. Theory, 2014, 60, (1), pp. 122– 132 25Cai, T.T., Wang, L., Xu, G.W.: 'Stable recovery of sparse signals and an oracle inequality', IEEE Trans. Inf. Theory, 2011, 56, (7), pp. 3516– 3522 26Gorodnitsky, I.F., George, J.S., Rao, B.D.: 'Neuromagnetic source imaging with FOCUSS: a recursive weighted minimum norm algorithm', Electroencephalogr. Clin. Neurophysiol., 1995, 95, (4), pp. 231– 251 27Malioutov, D., Cetin, M., Willsky, A.S.: 'A sparse signal reconstruction perspective for source localization with sensor arrays', IEEE Trans. Signal Process., 2005, 53, (8), pp. 3010– 3022 28Ender, J.H.G.: 'On compressive sensing applied to radar', Signal Process., 2010, 90, (5), pp. 1402– 1414 29Lai, M.J., Xu, Y.Y., Yin, W.T.: 'Improved iteratively reweighted least squares for unconstrained smoothed ℓq minimization', SIAM J. Numer. Anal., 2013, 51, (2), pp. 927– 957 30Boyd, S., Parikh, N., Chu, E. et al: 'Distributed optimization and statistical learning via the alternating direction method of multipliers', Found. Trends Mach. Learn., 2011, 3, (1), pp. 1– 122 31Wang, W.D., Wang, J.J., Zhang, Z.L.: 'Robust signal recovery with highly coherent measurement matrices', IEEE Signal Process. Lett., 2017, 24, (3), pp. 304– 308 32Lin, J.H., Li, S.: 'Block sparse recovery via mixed ℓ2 /ℓ1 minimization', Acta Math. Sin., 2013, 29, (7), pp. 1401– 1412 33Calderbank, R., Thompson, A., Xie, Y.: 'On block coherence of frames', Appl. Comput. Harmon. Anal., 2015, 38, (1), pp. 50– 71 34Bazaraa, M.S., Shetty, C.M.: ' Nonlinear programming theory and algorithms' ( Wiley, New York, 1979) 35Moor, B.D.: ' DaISy: database for the identification of systems'. Available at http://www.esat.kuleuven.ac.be/sista/daisy, November 2011 36Birnbaum, T., Eldar, Y.C., Needell, D.: ' Tolerant compressed sensing with partially coherent sensing matrices', arXiv preprint, arXiv:1608.05094, 2016 Citing Literature Volume12, Issue4June 2018Pages 422-430 FiguresReferencesRelatedInformation