Robust Principal Component Analysis Based on Low-Rank and Block-Sparse Matrix Decomposition

Size: px
Start display at page:

Download "Robust Principal Component Analysis Based on Low-Rank and Block-Sparse Matrix Decomposition"

Transcription

1 1 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition Qiuwei Li Colorado School of Mines Gongguo Tang Colorado School of Mines rye Nehorai Washington University in St. Louis 1.1 Introduction Notations and Problem Setup The Methods of ugmented Lagrange Multipliers Implementation Numerical Simulations Convergence Behavior Comparison of Final Relative rror Comparison of The Matrices and xamine of effect of the block sparsity s 1.6 Conclusions References Introduction The classical principal component analysis (PC) [Jol02] is arguably one of the most important tools in high dimensional data analysis. However, the classical PC is not robust to gross corruptions on even only a few entries of the underlying low-rank matrix L containing the principal components. The robust principal component analysis (RPC) [C- SPW09, CLMW11, ZLW + 10] models the gross corruptions as a sparse matrix S superpositioned on the low-rank matrix L. The authors of [CSPW09,CLMW11,ZLW + 10] presented various incoherence conditions under which the low-rank matrix L and the sparse matrix B can be accurately decomposed by solving a convex program, the principal component pursuit: min + λ 1 subject to D = +, (1.1), where D is the observation matrix, λ is a tuning parameter, and and 1 denote the nuclear norm and the sum of the absolute values of the matrix entries, respectively. Many algorithms have been designed to efficiently solve the resulting convex program, e.g., the singular value thresholding [CCS08], the alternating direction method [YY09], the accelerated proximal gradient (PG) method [LGW + 09], and the augmented Lagrange multiplier method [LCWM09]. Some of these algorithms also apply stably when the observation matrix is corrupted by small, entry-wise noise [ZLW + 10]. The applications of RPC include video surveillance [CLMW11], face recognition [PGW + 12, LLY10], latent semantic indexing [MZWM10], image retrieval and management [ZYM10], and computer 1-1

2 1-2 Book title goes here vision [ZGLM12, WGS + 11], to name a few. We comment that, in certain applications, the sparse matrix S is actually of more interest than the low-rank matrix, which contains the principal components. However, RPC is not robust to column/row outliers in the case where is a column/row block sparse matrix. Many data sets arranged in matrix formats are of low-rank due to the correlations and connections within the data samples. The ubiquity of low-rank matrices is also suggested by the success and popularity of principal component analysis in many applications [Jol02]. xamples of special interest are network traffic matrices [JW + 10], and the data matrix formed in face recognition [WYG + 09], where the columns correspond to vectorized versions of face images. When a few columns of the data matrix are generated by mechanisms different from the rest of the columns, the existence of these outlying columns tends to destroy the low-rank structure of the data matrix. decomposition that enforces the low-rankness of one part and the block-sparsity of the other part would separate the principal components from the outliers.in this work, we design algorithms for low-rank and block-sparse matrix decomposition as a robust version of principal component analysis insensitive to column/row outliers. We use a convex program to separate the low-rank part and block-sparse part of the observed matrix. The decomposition involves the following model: D = +, (1.2) The observation matrix D is now the sum of the low-rank component and the blocksparse component. The block-sparse matrix contains mostly zero columns, with several non-zero ones corresponding to outliers. In order to eliminate ambiguity, the columns of the low-rank matrix corresponding to the outlier columns are assumed to be zeros. We demonstrate that the following convex program recovers the low-rank matrix and the block-sparsity matrix : min, + κ(1 λ) 2,1 + κλ 2,1 subject to D = +, (1.3) where, 2,1 denote respectively the Frobinius norm, and the l 1 norm of the vector formed by taking the l 2 norms of the columns of the underlying matrix. Note that the extra term κ(1 λ) 2,1 actually ensures the recovered has exact zero columns corresponding to the outliers. We design efficient algorithms to solve the convex program (1.3) based on the augmented Lagrange multiplier (LM) method. The LM method was employed to successfully perform low-rank and sparse matrix decomposition for large scale problems [LCWM09]. The challenges here are the special structure of the 2,1 norm, and the existence of the extra term κ(1 λ) 2,1. We demonstrate the validity of (1.3) in decomposition and compare the performance of the LM algorithm using numerical simulations. In future work, we will test the algorithms by applying them to face recognition and network traffic anomaly detection problems. Due to the ubiquity of low-rank matrices and the importance of outlier detection, we expect low-rank and block-sparse matrix decomposition to have wide applications, especially in computer vision and data mining. The paper is organized as follows. In Section 1.2, we introduce notations and the problem setup. Section 1.3 is devoted to the methods of augmented Lagrange multipliers for solving the convex program (1.3). We provide implementation details in Section 1.4. Numerical experiments are used to demonstrate the effectiveness of our method and the results are reported in Section 1.5. Section 1.6 summarizes our conclusions.

3 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition Notations and Problem Setup In this section, we introduce notations and the problem setup used throughout the paper. For any matrix R n p, we use ij to denote its ijth element, and j to denote its jth column The usual l p norm is denoted by p, p 1 when the argument is a vector. The nuclear norm, the Frobinius norm F, and the l 2,1 norm 2,1 of matrix are defined as follows: = i σ i (), (1.4) F = 2 ij, (1.5) i,j 2,1 = j j 2, (1.6) where σ i () is the ith largest singular value of. We consider the following model: D = +. (1.7) Here has at most s non-zero columns. These non-zero columns are considered as outliers in the entire data matrix D. The corresponding columns of are assumed to be zeros. In addition, has low rank, i.e., rank() r. We use the following optimization to recover and from the observation D: min, + κ(1 λ) 2,1 + κλ 2,1 subject to D = +. (1.8) This procedure separates the outliers in and the principal components in. s we will see in the numerical examples, simply enforcing the sparsity of and the low-rankness of would not separates and well. 1.3 The Methods of ugmented Lagrange Multipliers The ugmented Lagrange Multipliers (LM) method is a general procedure to solve the following equality constrained optimization problem: min f(x) subject to h(x) = 0, (1.9) where f : R n R and h : R n R m. The procedure defines the augmented Lagrangian function as L(x, λ; µ) = f(x) + λ, h(x) + µ 2 h(x) 2 2, (1.10) where λ R m is the Lagrange multiplier vector, and µ is a positive scalar. The LM iteratively solves x and the Lagrangian multiplier vector λ as shown in Table 1.1. To apply the general LM to problem (1.8), we define f(x) = + κ(1 λ) 2,1 + κλ 2,1, (1.11) h(x) = D (1.12)

4 1-4 Book title goes here with X = (, ). The corresponding augmented Lagrangian function is L(,, Y ; µ) = + κ(1 λ) 2,1 + κλ 2,1 + Y, D + µ 2 D 2 F, (1.13) where we use Y to denote the Lagrange multiplier. Given Y = Y k and µ = µ k, one key step in applying the general LM method is to solve min, L(,, Y k ; µ k ). We adopt an alternating procedure: for fixed = k, solve and for fixed = k+1, solve k+1 = argmin L(, k, Y k ; µ k ), (1.14) k+1 = argmin L( k+1,, Y k ; µ k ). (1.15) We first address (1.15), which is equivalent to solving min κλ 2,1 + µ k 2 D k Y k µ k 2 F. (1.16) Denote G = D k µ k Y k. The following lemma gives the closed-form solution of the minimization with respect to. Lemma 1 The solution to min η 2, G 2 F (1.17) is given by for j = 1,..., p. j = G j max ( ) 0, 1 η G j 2 (1.18) Proof: Note the objective function expands as = η 2, G 2 F n ( η j ) 2 j G j 2 2. (1.19) j=1 TBL 1.1 General augmented Lagrange multiplier method 1. initialize: Given ρ > 1, µ 0 > 0, starting point x s 0 and λ0; k 0 2. while not converged do 3. pproximately solve x k+1 = argmin x L(x, λ k ; µ k ) using starting point x s k Set starting point x s k+1 = x k+1. Update the Lagrange multiplier vector λ k+1 = λ k + µ k h(x k+1 ). 6. Update µ k+1 = ρµ k. 7. end while 8. output: x x k, λ λ k.

5 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition1-5 Now we can minimize with respect to j separately. Denote h(e) = η e e g 2 2 = η e First consider g 2 η. Cauchy-Schwarz inequality leads to h(e) η e [ e e, g + g 2 ] 2. (1.20) [ e e 2 g 2 + g 2 2] = 1 2 e (η g 2 ) e g 2 2 (1.21) 1 2 g 2 2, (1.22) where for the last inequality we used (1.21) is an increasing function on [0, ) if g 2 η. pparently, h(e) = g 2 /2 is achieved by e = 0 which is also unique. Therefore, if g 2 η, then the minimum of h(e) is achieved by the unique solution e = 0 In the second case when g 2 > η. Setting the derivative of h(e) with respect to e to zero yields ( ) η e + 1 = g. (1.23) e 2 Taking the l 2 norm of both sides yields e 2 = g 2 η > 0. (1.24) Plugging e 2 into (1.23) gives ( e = g 1 η ). (1.25) g 2 In conclusion, the minimum of (1.17) is achieved by with ( ( )) j = G j max 0, 1 η G j 2 (1.26) for j = 1, 2,..., p. We denote the operator solving (1.17) as T η ( ). So = T η (G) sets the columns of to be zero vectors if the l 2 norms of the corresponding columns of G are less than η, and scales down the columns otherwise by a factor 1 Now we turn to solve which can be rewritten as η G j 2. k+1 = argmin L(, k, Y k ; µ k ) (1.27) k+1 = { argmin + κ(1 λ) 2,1 + µ } k 2 G 2 F, (1.28)

6 1-6 Book title goes here where G = D k + 1 µ k Y k. We know that without the extra κ(1 λ) 2,1 term, a closed form solution is simply given by the soft-thresholding operator [CCS08]. Unfortunately, a closed form solution to (1.28) is not available. We use the Douglas/Peaceman-Rachford (DR) monotone operator splitting method [CP07,FSM07,FSM09] to iteratively solve (1.28). Define f 1 () = κ(1 λ) 2,1 + µ k 2 G 2 F and f 2 () =. For any β > 0 and a sequence, the DR itration for (1.28) is expressed as α t (0, 2) (j+1/2) = prox βf2 ( (j) ), (1.29) (j+1) = (j) + α j (prox βf1 (2 (j+1/2) (j) ) (j+1/2)). (1.30) Here for any proper, lower semi-continuous, convex function f, the proximity operators prox f ( ) gives the unique point prox f (x) achieving the infimum of the function z 1 2 x z f(z). (1.31) The following lemmas gives the explicit expression for the proximity operator involved in our DR iteration when f() = and f() = κ(1 λ) 2,1 + µ k 2 G 2 F. Lemma 2 For f( ) =, suppose the singular value decomposition of is = UΣV T, then the proximity operator is prox βf () = US β (Σ)V T, (1.32) where the nonnegative soft-thresholding operator is defined as S ν (x) = max(0, x ν), x 0, ν > 0, (1.33) for a scalar x and extended entry-wisely to vectors and matrices. For f( ) = η 2,1 + µ 2 G 2 F, the proximity operator is prox βf () = T βη 1+βµ ( ) + βµg. (1.34) 1 + βµ With these preparations, we present the algorithm for solving (1.8) using the LM, named as RPC-LBD, in Table 1.2. Note that instead of alternating between (1.14) and (1.15) many times until convergence, the RPC-LBD algorithm in Table 1.2 executes them only once. This inexact version greatly reduces the computational burden and yields sufficiently accurate results for appropriately tuned parameters. The strategy was also adopted in [LCWM09] to perform the low-rank and sparse matrix decomposition.

7 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition Implementation Parameter selection and initialization: The tuning parameters in (1.8) are chosen to be κ = 1.1 and λ = We do not have a systematic way to select these parameters. But these empirical values work well for all tested cases. For the DR iteration (1.29) and (1.30), we have β = 0.2 and α j 1. The parameters for the LM method are µ 0 = 30/ sign(d) 2 and ρ = 1.1. The low-rank matrix, the block-sparse matrix, and the Lagrange multiplier Y are initialized respectively to D, 0, and 0. The outer loop in Table 1.2 terminates when it reaches the maximal iteration number 500 or the error tolerance The error in the outer loop is computed by D k k F / D F. The inner loop for the DR iteration has maximal iteration number 20 and tolerance error The error in the inner loop is the Frobenius norm of the difference between successive (j) k s. Performing SVD: The major computational cost in the RPC-LBD algorithm is the singular value decomposition (SVD) in the DR iteration (Line 07 in Table 1.2). We usually do not need to compute the full SVD because only those that are greater than β are used. We use the PROPCK [Lar98] to compute the first (largest) few singular values. In order to both save computation and ensure accuracy, we need to predict the number of singular values of (j) k+1 that exceed β for each iteration. We adopt the following rule [LCWM09]: sv k+1 = { svnk + 1, if svn k < sv k ; min(svn k + 10, d), if svn k = sv k., where sv 0 = 10 and svn k = { svpk, if maxgap k 2; min(svp k, maxid k ), if maxgap k > 2. Here d = min(n, p), sv k is the predicted number of singular values that are greater than β, and svp k is the number of singular values in the sv k singular values that are larger than β. In addition, maxgap k and maxid k are respectively the largest ratio between successive singular values of (j) k+1 when arranged in a decreasing order and the corresponding index. TBL 1.2 Block-sparse and low-rank matrix decomposition via LM RPC-LBD(D, κ, λ) Input: Data matrix D R n p and the parameters κ, λ. Output: The low-rank part and the block-sparse part. 01. initialize: 0 D; 0 0; µ 0 = 30/ sign(d) 2 ; ρ > 0, β > 0, α (0, 1), k while not converged do 03. \\ Lines 4-11 solve k+1 = argmin L(, k, Y k ; µ k ). 04. G = D k + 1 µ Y k. k 05. j 0, (0) k+1 = G. 06. while not converged do 07. (j+1/2) k+1 = US β (Σ)V ( T where (j) ( k+1 = UΣV T is the SVD of (j) ) k+1. ) 08. (j+1) k+1 = (j) k+1 + α 09. j j end while T βκ(1 λ) 1+βµ k 11. k+1 = (j+1/2) k G = D k µ Y k. ( k 13. k+1 = T κλ G ). µ k 14. Y k+1 = Y k + µ k (D k+1 k+1 ). 15. µ k+1 = ρµ k. 16. k k end while 18. output: k, k. 2 (j+1/2) (j) k+1 k+1 +βµ k G 1+βµ k (j+1/2) k+1.

8 1-8 Book title goes here 1.5 Numerical Simulations Through this section, we consider only square matrices with n = p. The square matrix D is generated by follows: Construction of matrix D Construct matrix as the product of two Gaussian matrices with dimensions n r and r n. By this way, we get rank() = r. Normalize the columns of. Generate a support S of size s as a random subset of {1,..., p}. Let the columns of corresponding to the support S be sampled from the Gaussian distribution. Normalize the non-zero columns of. Set the columns of L corresponding to S to zeros. Construct D as the sum of and Convergence Behavior In this subsection, we mainly focus on the convergence behavior of the RPC-LBD and compare it with the classical RPC solved by the LM algorithm proposed in [LCWM09]. In order to illustrate the convergence behavior of these two algorithms, we apply these two algorithms to the same matrix D by running the same iterations K which is chosen as 20 here and then observe the evolution of the reconstruction error ρ (k), ρ (k), ρ (k) and ρ (k), respectively. The reconstruction error ρ (k), ρ (k), ρ (k) and ρ (k) are defined as follows. DFINITION 1.1 ssume (k) is the approximation to and (k) is the approximation to by running k iteration of the algorithm RPC- LBD. ssume (k) is the approximation to and (k) is the approximation to by running k iteration of the algorithm RPC. Then, ρ (k) (k) F F (1.35) ρ (k) (k) F F (1.36) ρ (k) (k) F F (1.37) ρ (k) (k) F F (1.38)

9 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition Reconstruction error of RPC RPC LBD FIGUR 1.1: Reconstruction error ρ n = 100, p = 100, r = 4, s = itration: k (k) and ρ (k) versus iteration number k for Case I: n = 100, p = 100, r = 4, s = 4 Figure 1.1 shows Reconstruction error ρ (k) and ρ (k) versus iteration number k for the case n = 100, p = 100, r = 4, s = 4 using RPC-LBD and RPC, respectively. Figure 1.2 shows Reconstruction error ρ (k) and ρ (k) versus iteration number k for the case n = 100, p = 100, r = 4, s = 4 using RPC-LBD and RPC, respectively. Case II: n = 200, p = 200, r = 8, s = 8 Figure 1.3 shows Reconstruction error ρ (k) and ρ (k) versus iteration number k for the case n = 200, p = 200, r = 8, s = 8 using RPC-LBD and RPC, respectively. Figure 1.4 shows Reconstruction error ρ (k) and ρ (k) versus iteration number k for the case n = 200, p = 200, r = 8, s = 8 using RPC-LBD and RPC, respectively. Case III: n = 300, p = 300, r = 12, s = 12 Figure 1.5 shows Reconstruction error ρ (k) and ρ (k) versus iteration number k for the case n = 300, p = 300, r = 12, s = 12 using RPC-LBD and RPC, respectively. Figure 1.6 shows Reconstruction error ρ (k) and ρ (k) versus iteration number k for the case n = 200, p = 300, r = 12, s = 12 using RPC-LBD and RPC, respectively.

10 1-10 Book title goes here 10 0 Reconstruction error of RPC RPC LBD FIGUR 1.2: Reconstruction error ρ n = 100, p = 100, r = 4, s = itration: k (k) and ρ (k) versus iteration number k for 10 0 Reconstruction error of RPC RPC LBD FIGUR 1.3: Reconstruction error ρ n = 200, p = 200, r = 8, s = itration: k (k) and ρ (k) versus iteration number k for

11 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition Reconstruction error of RPC RPC LBD FIGUR 1.4: Reconstruction error ρ n = 200, p = 200, r = 8, s = itration: k (k) and ρ (k) versus iteration number k for 10 0 Reconstruction error of RPC RPC LBD FIGUR 1.5: Reconstruction error ρ n = 300, p = 300, r = 12, s = itration: k (k) and ρ (k) versus iteration number k for

12 1-12 Book title goes here 10 0 Reconstruction error of RPC RPC LBD FIGUR 1.6: Reconstruction error ρ n = 300, p = 300, r = 12, s = itration: k (k) and ρ (k) versus iteration number k for Comment 5.1: s seen from the simulations, in all the cases, i.e., n = 100, n = 200, n = 300, the proposed algorithm RPC-LBD always outperforms the classical algorithm RP- C in term of the evolutions of the four indicators: ρl (k), ρ (k), ρ (k) and ρ (k). For example, from the comparison of the evolution of the indicators ρ (k) and ρ (k), we can see RPC would be quickly trapped into a local minimizer with about 9 iterations for all the three cases. In the contrary, the proposed RPC-LBD would not be that quickly trapped into a local minimizer hence it could approaches to a much more better approximations of the global minimizer, i.e., the true low rank matrix and this happens for all the three cases (n = 100, n = 200, n = 300 ). Furthermore, ρ (k) and ρ (k) are in fact the relative error between the approximation and the true low rank matrix L in k-th iteration. Thus, from the simulations, we can see the smallest relative error the classical algorithm RPC can achieved is at the order of 10 1 and the smallest relative error the proposed algorithm RPC-LBD can achieved is at the order of Since the classical algorithm RPC would be quickly be trapped into some local minimizer within only 9 iterations for all the three cases. It only need 20 iterations to have a good comparison between the evolution of the four indicators of the two algorithms. Here, r = round(0.04n) and s = round(0.3n) for all the three cases Comparison of Final Relative rror In the last subsection, we get a main idea of the convergence behaviour of the proposed algorithm RPC-LBD and the classical algorithm RPC. In this subsection, we will will compare the final results the two algorithms can achieve in term of the final relative recovery errors ρ, ρ, ρ and ρ, which is defined as follows.

13 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition1-13 DFINITION 1.2 ssume is the final approximation to and is the final approximation to by running the algorithm RPC-LBD until convergence. ssume is the final approximation to and is the final approximation to by running the algorithm RPC until convergence. Then, the final relative recovery errors are defined as follows: ρ F (1.39) F ρ F (1.40) F ρ F (1.41) F ρ F (1.42) F In Table 1.3, we compare the four final relative recovery errors of the RPC- LBD and the RPC: ρ, ρ, ρ and ρ. We consider n = 100, 200, 300, 400, 500, 600, 700, 800 and 900. Comment 5.2: The rank of the low-rank matrix is r = round(0.04n) and the number of the non-zero columns of the block sparse matrix is s = round(0.04n). We see that the RPC-LBD can accurately recover the low-rank matrix L and also the block sparse matrix. The relative recovery error of matrix by RPC- LBD ρ is at the order of In contrary, the relative recovery error of matrix L by RPC ρ ρ is at the order of Similarly, the relative recovery error of matrix by RPC-LBD is at the oder of However, the relative recovery error of matrix by RPC ρ is at the order of TBL 1.3 RPC. n Comparison of our algorithm (RPC-LBD) and the ρ ρ ρ ρ e e e e e e e e e e e e e e e e e e

14 1-14 Book title goes here Comparison of The Matrices and In this section, we will see the visible image of the different reconstructed matrices and by the proposed algorithm RPC-LBD and the classical algorithm RPC, respectively. We choose only the reconstructed matrices and for a visible sight since and are approximation to the block sparse matrix whose image version can give us a more obvious feeling. However, for the the reconstructed matrices and are approximation to the low rank matrix which, here, is the product of two gaussian matrix. Thus, in order to generate a more intuitively sight, in the following, we will see the visible image of the different reconstructed matrices and in the different cases.in the following case, we consider n = p, r = round(0.03n) and s = round(0.3n). Case I: n = 100, p = 100, r = 3, s = 30 In Figure 1.7, we show the original and those recovered by the RPC-LBD and the classical RPC. Here n = 100, p = 100, r = 3, s = 30. We can see that our algorithm RPC- LBD recovers the block sparse matrix near perfectly, while the RPC performs badly. ctually, the relative error for L recovered by the RPC-LBD ρ is while the relative error ρ is Case II: n = 200, p = 200, r = 6, s = 60 In Figure 1.8, we show the original and those recovered by the RPC-LBD and the classical RPC. Here n = 200, p = 200, r = 6, s = 60. We can see that our algorithm RPC- LBD recovers the block sparse matrix near perfectly, while the RPC performs badly. ctually, the relative error for recovered by the RPC-LBD ρ is while the relative error ρ is Case III: n = 300, p = 300, r = 9, s = 90 In Figure 1.9, we show the original and those recovered by the RPC-LBD and the classical RPC. Here n = 300, p = 300, r = 9, s = 90. We can see that our algorithm RPC- LBD recovers the block sparse matrix near perfectly, while the RPC performs badly. ctually, the relative error for recovered by the RPC-LBD ρ is while the relative error ρ is Comment 5.3: In the above cases, we set the block sparsity s of matrix as s = round(0.3n), which is already very high. Therefore, the classical algorithm RPC almost failed to recovery the block sparse matrix for all the cases. However, RPC-LBD can recovery exactly the block sparse matrix for all the cases and the relative error achieved by RPC-LBD is at the order of In contrast, the relative error achieved by RPC is about 30% which is very high. Hence, the proposed algorithm RPC-LBD is more robust to the factor of block sparsity s.

15 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition1-15 (a) (b) (c) FIGUR 1.7: For the case n = 100, p = 100, r = 3, s = 30. (a) The original ; (b) ; (c).

16 1-16 Book title goes here (a) (b) (c) FIGUR 1.8: For the case n = 200, p = 200, r = 6, s = 60. (a) The original ; (b) ; (c).

17 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition1-17 (a) (b) (c) FIGUR 1.9: For the case n = 300, p = 300, r = 9, s = 90. (a) The original ; (b) ; (c).

18 1-18 Book title goes here RPC RPC LBD Reconstruction error of sparsity: s FIGUR 1.10: Reconstruction error ρ and ρ versus block sparsity s varying from 10 to 50 with n = 100, p = 100, r = xamine of effect of the block sparsity s In Figure 1.10, we show the final relative error ρ and ρ corresponding to RPC and RPC-LBD, respectively, for block sparsity s varying from 10 to 50 with n = 100, p = 100, r = 4. In Figure 1.11, we show the final relative error ρ and ρ corresponding to RPC and RPC-LBD, respectively, for block sparsity s varying from 10 to 50 with n = 100, p = 100, r = 4. Comment 5.4: We set the block sparsity s of matrix varying in the range [ ]. Figure 1.10 and Figure 1.11 show the both RPC-LBD and RPC would behave worse when the block sparsity s becomes higher. One can see the classical algorithm RPC failed to recovery the block sparse matrix and the low rank matrix for all the different kinds of block sparsity. However, RPC-LBD can recovery exactly them when s = 10, 20. For s = 30, 40, 50, although RPC-LBD can not exactly recover the matrices,, its final relative error is still much smaller than those achieved by RPC, which, again, implies the proposed algorithm RPC-LBD is more robust to the block sparsity s.

19 Robust Principal Component nalysis Based on Low-Rank and Block-Sparse Matrix Decomposition RPC RPC LBD Reconstruction error of sparsity: s FIGUR 1.11: Reconstruction error ρ and ρ versus block sparsity s varying from 10 to 50 with n = 100, p = 100, r = Conclusions In this work we proposed a convex program for accurate low-rank and block-sparse matrix decomposition. We solved the convex program using the augmented Lagrange multiplier method. We used the Douglas/Peaceman-Rachford (DR) monotone operator splitting method to solve a subproblem involved in the augmented Lagrange multiplier method. We demonstrated the accuracy of our program and compared its results with those given by the robust principal component analysis. The proposed algorithm is potentially useful for outlier detection. References 1.. bdelkefi, Y. Jiang, W. Wang,. slebo, and O. Kvittem. Robust traffic anomaly detection with principal component pursuit. In Proceedings of the CM CoNXT Student Workshop, CoNXT 10 Student Workshop, pages 10:1 10:2, New York, NY, US, CM. 2. J-F Cai,. J. Candès, and Z. Shen. singular value thresholding algorithm for matrix completion. SIM J. on Optimization, 20(4): , J. Candès, X. Li, Y. Ma, and J. Wright. Robust principal component analysis? J. CM, 58(3):1 37, jun P. L. Combettes and J.. Pesquet. Douglas-rachford splittting approach to nonsmooth convex variational signal recovery. I Journal of Selected Topics in Signal Processing, 1(4): , Venkat Chandrasekaran, Sujay Sanghavi, Pablo Parrilo, and lan S Willsky. Sparse and low-rank matrix decompositions. In Communication, Control, and Computing, llerton th nnual llerton Conference on, pages I, M. J. Fadili, J. L. Starck, and F. Murtagh. Inpainting and zooming using sparse representations. The Computer Journal, 52:64 79, 2007.

20 1-20 Book title goes here 7. M. J. Fadili, J. L. Starck, and F. Murtagh. Inpainting and zooming using sparse representations. The Computer Journal, 52(1):64, I. T. Jolliffe. Principal component analysis. Springer series in statistics. Springer, R. M. Larsen. Lanczos bidiagonalization with partial reorthogonalization. Department of computer science, arhus University, Technical report, DIMI PB-357, code available at rmunk/propck/, Z. Lin, M. Chen, L. Wu, and Y. Ma. The augmented Lagrange multiplier method for exact recovery of corrupted low-rank matrices. UIUC Technical Report UILU- NG , nov Zhouchen Lin, rvind Ganesh, John Wright, Leqin Wu, Minming Chen, and Yi Ma. Fast convex optimization algorithms for exact recovery of a corrupted low-rank matrix. Computational dvances in Multi-Sensor daptive Processing (CMSP), 61, Guangcan Liu, Zhouchen Lin, and Yong Yu. Robust subspace segmentation by low-rank representation. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages , Kerui Min, Zhengdong Zhang, John Wright, and Yi Ma. Decomposing background topics from keywords by principal component pursuit. In Proceedings of the 19th CM international conference on Information and knowledge management, pages CM, Yigang Peng, rvind Ganesh, John Wright, Wenli Xu, and Yi Ma. Rasl: Robust alignment by sparse and low-rank decomposition for linearly correlated images. Pattern nalysis and Machine Intelligence, I Transactions on, 34(11): , Lun Wu, rvind Ganesh, Boxin Shi, Yasuyuki Matsushita, Yongtian Wang, and Yi Ma. Robust photometric stereo via low-rank matrix completion and recovery. In Computer Vision CCV 2010, pages Springer, J. Wright,. Yang,. Ganesh, S. Sastry, and Y. Ma. Robust face recognition via sparse representation. 31(2), feb Xiaoming Yuan and Junfeng Yang. Sparse and low-rank matrix decomposition via alternating direction methods. preprint, Zhengdong Zhang, rvind Ganesh, Xiao Liang, and Yi Ma. Tilt: transform invariant low-rank textures. International Journal of Computer Vision, 99(1):1 24, Zihan Zhou, Xiaodong Li, John Wright, mmanuel Candes, and Yi Ma. Stable principal component pursuit. In Information Theory Proceedings (ISIT), 2010 I International Symposium on, pages I, Guangyu Zhu, Shuicheng Yan, and Yi Ma. Image tag refinement towards low-rank, content-tag prior and error sparsity. In Proceedings of the international conference on Multimedia, pages CM, 2010.

Robust Principal Component Analysis Based on Low-Rank and Block-Sparse Matrix Decomposition

Robust Principal Component Analysis Based on Low-Rank and Block-Sparse Matrix Decomposition Robust Principal Component Analysis Based on Low-Rank and Block-Sparse Matrix Decomposition Gongguo Tang and Arye Nehorai Department of Electrical and Systems Engineering Washington University in St Louis

More information

Robust Principal Component Analysis

Robust Principal Component Analysis ELE 538B: Mathematics of High-Dimensional Data Robust Principal Component Analysis Yuxin Chen Princeton University, Fall 2018 Disentangling sparse and low-rank matrices Suppose we are given a matrix M

More information

Dense Error Correction for Low-Rank Matrices via Principal Component Pursuit

Dense Error Correction for Low-Rank Matrices via Principal Component Pursuit Dense Error Correction for Low-Rank Matrices via Principal Component Pursuit Arvind Ganesh, John Wright, Xiaodong Li, Emmanuel J. Candès, and Yi Ma, Microsoft Research Asia, Beijing, P.R.C Dept. of Electrical

More information

Analysis of Robust PCA via Local Incoherence

Analysis of Robust PCA via Local Incoherence Analysis of Robust PCA via Local Incoherence Huishuai Zhang Department of EECS Syracuse University Syracuse, NY 3244 hzhan23@syr.edu Yi Zhou Department of EECS Syracuse University Syracuse, NY 3244 yzhou35@syr.edu

More information

Robust PCA. CS5240 Theoretical Foundations in Multimedia. Leow Wee Kheng

Robust PCA. CS5240 Theoretical Foundations in Multimedia. Leow Wee Kheng Robust PCA CS5240 Theoretical Foundations in Multimedia Leow Wee Kheng Department of Computer Science School of Computing National University of Singapore Leow Wee Kheng (NUS) Robust PCA 1 / 52 Previously...

More information

Exact Recoverability of Robust PCA via Outlier Pursuit with Tight Recovery Bounds

Exact Recoverability of Robust PCA via Outlier Pursuit with Tight Recovery Bounds Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence Exact Recoverability of Robust PCA via Outlier Pursuit with Tight Recovery Bounds Hongyang Zhang, Zhouchen Lin, Chao Zhang, Edward

More information

ECE 8201: Low-dimensional Signal Models for High-dimensional Data Analysis

ECE 8201: Low-dimensional Signal Models for High-dimensional Data Analysis ECE 8201: Low-dimensional Signal Models for High-dimensional Data Analysis Lecture 7: Matrix completion Yuejie Chi The Ohio State University Page 1 Reference Guaranteed Minimum-Rank Solutions of Linear

More information

Linearized Alternating Direction Method with Adaptive Penalty for Low-Rank Representation

Linearized Alternating Direction Method with Adaptive Penalty for Low-Rank Representation Linearized Alternating Direction Method with Adaptive Penalty for Low-Rank Representation Zhouchen Lin Visual Computing Group Microsoft Research Asia Risheng Liu Zhixun Su School of Mathematical Sciences

More information

The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices

The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices Noname manuscript No. (will be inserted by the editor) The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices Zhouchen Lin Minming Chen Yi Ma Received: date / Accepted:

More information

The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices

The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices Noname manuscript No. (will be inserted by the editor) The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices Zhouchen Lin Minming Chen Yi Ma Received: date / Accepted:

More information

Robust Principal Component Pursuit via Alternating Minimization Scheme on Matrix Manifolds

Robust Principal Component Pursuit via Alternating Minimization Scheme on Matrix Manifolds Robust Principal Component Pursuit via Alternating Minimization Scheme on Matrix Manifolds Tao Wu Institute for Mathematics and Scientific Computing Karl-Franzens-University of Graz joint work with Prof.

More information

CSC 576: Variants of Sparse Learning

CSC 576: Variants of Sparse Learning CSC 576: Variants of Sparse Learning Ji Liu Department of Computer Science, University of Rochester October 27, 205 Introduction Our previous note basically suggests using l norm to enforce sparsity in

More information

A Counterexample for the Validity of Using Nuclear Norm as a Convex Surrogate of Rank

A Counterexample for the Validity of Using Nuclear Norm as a Convex Surrogate of Rank A Counterexample for the Validity of Using Nuclear Norm as a Convex Surrogate of Rank Hongyang Zhang, Zhouchen Lin, and Chao Zhang Key Lab. of Machine Perception (MOE), School of EECS Peking University,

More information

arxiv: v2 [stat.ml] 1 Jul 2013

arxiv: v2 [stat.ml] 1 Jul 2013 A Counterexample for the Validity of Using Nuclear Norm as a Convex Surrogate of Rank Hongyang Zhang, Zhouchen Lin, and Chao Zhang Key Lab. of Machine Perception (MOE), School of EECS Peking University,

More information

Shiqian Ma, MAT-258A: Numerical Optimization 1. Chapter 9. Alternating Direction Method of Multipliers

Shiqian Ma, MAT-258A: Numerical Optimization 1. Chapter 9. Alternating Direction Method of Multipliers Shiqian Ma, MAT-258A: Numerical Optimization 1 Chapter 9 Alternating Direction Method of Multipliers Shiqian Ma, MAT-258A: Numerical Optimization 2 Separable convex optimization a special case is min f(x)

More information

Contraction Methods for Convex Optimization and Monotone Variational Inequalities No.16

Contraction Methods for Convex Optimization and Monotone Variational Inequalities No.16 XVI - 1 Contraction Methods for Convex Optimization and Monotone Variational Inequalities No.16 A slightly changed ADMM for convex optimization with three separable operators Bingsheng He Department of

More information

ECE G: Special Topics in Signal Processing: Sparsity, Structure, and Inference

ECE G: Special Topics in Signal Processing: Sparsity, Structure, and Inference ECE 18-898G: Special Topics in Signal Processing: Sparsity, Structure, and Inference Low-rank matrix recovery via convex relaxations Yuejie Chi Department of Electrical and Computer Engineering Spring

More information

Solving Corrupted Quadratic Equations, Provably

Solving Corrupted Quadratic Equations, Provably Solving Corrupted Quadratic Equations, Provably Yuejie Chi London Workshop on Sparse Signal Processing September 206 Acknowledgement Joint work with Yuanxin Li (OSU), Huishuai Zhuang (Syracuse) and Yingbin

More information

Matrix Completion for Structured Observations

Matrix Completion for Structured Observations Matrix Completion for Structured Observations Denali Molitor Department of Mathematics University of California, Los ngeles Los ngeles, C 90095, US Email: dmolitor@math.ucla.edu Deanna Needell Department

More information

L 2,1 Norm and its Applications

L 2,1 Norm and its Applications L 2, Norm and its Applications Yale Chang Introduction According to the structure of the constraints, the sparsity can be obtained from three types of regularizers for different purposes.. Flat Sparsity.

More information

Conditions for Robust Principal Component Analysis

Conditions for Robust Principal Component Analysis Rose-Hulman Undergraduate Mathematics Journal Volume 12 Issue 2 Article 9 Conditions for Robust Principal Component Analysis Michael Hornstein Stanford University, mdhornstein@gmail.com Follow this and

More information

Non-convex Robust PCA: Provable Bounds

Non-convex Robust PCA: Provable Bounds Non-convex Robust PCA: Provable Bounds Anima Anandkumar U.C. Irvine Joint work with Praneeth Netrapalli, U.N. Niranjan, Prateek Jain and Sujay Sanghavi. Learning with Big Data High Dimensional Regime Missing

More information

arxiv: v1 [stat.ml] 1 Mar 2015

arxiv: v1 [stat.ml] 1 Mar 2015 Matrix Completion with Noisy Entries and Outliers Raymond K. W. Wong 1 and Thomas C. M. Lee 2 arxiv:1503.00214v1 [stat.ml] 1 Mar 2015 1 Department of Statistics, Iowa State University 2 Department of Statistics,

More information

Fast Algorithms for Structured Robust Principal Component Analysis

Fast Algorithms for Structured Robust Principal Component Analysis Fast Algorithms for Structured Robust Principal Component Analysis Mustafa Ayazoglu, Mario Sznaier and Octavia I. Camps Dept. of Electrical and Computer Engineering, Northeastern University, Boston, MA

More information

A Brief Overview of Practical Optimization Algorithms in the Context of Relaxation

A Brief Overview of Practical Optimization Algorithms in the Context of Relaxation A Brief Overview of Practical Optimization Algorithms in the Context of Relaxation Zhouchen Lin Peking University April 22, 2018 Too Many Opt. Problems! Too Many Opt. Algorithms! Zero-th order algorithms:

More information

Fantope Regularization in Metric Learning

Fantope Regularization in Metric Learning Fantope Regularization in Metric Learning CVPR 2014 Marc T. Law (LIP6, UPMC), Nicolas Thome (LIP6 - UPMC Sorbonne Universités), Matthieu Cord (LIP6 - UPMC Sorbonne Universités), Paris, France Introduction

More information

sparse and low-rank tensor recovery Cubic-Sketching

sparse and low-rank tensor recovery Cubic-Sketching Sparse and Low-Ran Tensor Recovery via Cubic-Setching Guang Cheng Department of Statistics Purdue University www.science.purdue.edu/bigdata CCAM@Purdue Math Oct. 27, 2017 Joint wor with Botao Hao and Anru

More information

Recovery of Low-Rank Plus Compressed Sparse Matrices with Application to Unveiling Traffic Anomalies

Recovery of Low-Rank Plus Compressed Sparse Matrices with Application to Unveiling Traffic Anomalies July 12, 212 Recovery of Low-Rank Plus Compressed Sparse Matrices with Application to Unveiling Traffic Anomalies Morteza Mardani Dept. of ECE, University of Minnesota, Minneapolis, MN 55455 Acknowledgments:

More information

FAST FIRST-ORDER METHODS FOR STABLE PRINCIPAL COMPONENT PURSUIT

FAST FIRST-ORDER METHODS FOR STABLE PRINCIPAL COMPONENT PURSUIT FAST FIRST-ORDER METHODS FOR STABLE PRINCIPAL COMPONENT PURSUIT N. S. AYBAT, D. GOLDFARB, AND G. IYENGAR Abstract. The stable principal component pursuit SPCP problem is a non-smooth convex optimization

More information

GoDec: Randomized Low-rank & Sparse Matrix Decomposition in Noisy Case

GoDec: Randomized Low-rank & Sparse Matrix Decomposition in Noisy Case ianyi Zhou IANYI.ZHOU@SUDEN.US.EDU.AU Dacheng ao DACHENG.AO@US.EDU.AU Centre for Quantum Computation & Intelligent Systems, FEI, University of echnology, Sydney, NSW 2007, Australia Abstract Low-rank and

More information

Fast Nonnegative Matrix Factorization with Rank-one ADMM

Fast Nonnegative Matrix Factorization with Rank-one ADMM Fast Nonnegative Matrix Factorization with Rank-one Dongjin Song, David A. Meyer, Martin Renqiang Min, Department of ECE, UCSD, La Jolla, CA, 9093-0409 dosong@ucsd.edu Department of Mathematics, UCSD,

More information

arxiv: v2 [cs.cv] 20 Apr 2014

arxiv: v2 [cs.cv] 20 Apr 2014 Robust Subspace Recovery via Bi-Sparsity Pursuit arxiv:1403.8067v2 [cs.cv] 20 Apr 2014 Xiao Bian, Hamid Krim North Carolina State University Abstract Successful applications of sparse models in computer

More information

A Local Non-Negative Pursuit Method for Intrinsic Manifold Structure Preservation

A Local Non-Negative Pursuit Method for Intrinsic Manifold Structure Preservation Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence A Local Non-Negative Pursuit Method for Intrinsic Manifold Structure Preservation Dongdong Chen and Jian Cheng Lv and Zhang Yi

More information

Coordinate Update Algorithm Short Course Operator Splitting

Coordinate Update Algorithm Short Course Operator Splitting Coordinate Update Algorithm Short Course Operator Splitting Instructor: Wotao Yin (UCLA Math) Summer 2016 1 / 25 Operator splitting pipeline 1. Formulate a problem as 0 A(x) + B(x) with monotone operators

More information

Lecture Notes 10: Matrix Factorization

Lecture Notes 10: Matrix Factorization Optimization-based data analysis Fall 207 Lecture Notes 0: Matrix Factorization Low-rank models. Rank- model Consider the problem of modeling a quantity y[i, j] that depends on two indices i and j. To

More information

2 Regularized Image Reconstruction for Compressive Imaging and Beyond

2 Regularized Image Reconstruction for Compressive Imaging and Beyond EE 367 / CS 448I Computational Imaging and Display Notes: Compressive Imaging and Regularized Image Reconstruction (lecture ) Gordon Wetzstein gordon.wetzstein@stanford.edu This document serves as a supplement

More information

arxiv: v1 [cs.it] 26 Oct 2018

arxiv: v1 [cs.it] 26 Oct 2018 Outlier Detection using Generative Models with Theoretical Performance Guarantees arxiv:1810.11335v1 [cs.it] 6 Oct 018 Jirong Yi Anh Duc Le Tianming Wang Xiaodong Wu Weiyu Xu October 9, 018 Abstract This

More information

Robust Component Analysis via HQ Minimization

Robust Component Analysis via HQ Minimization Robust Component Analysis via HQ Minimization Ran He, Wei-shi Zheng and Liang Wang 0-08-6 Outline Overview Half-quadratic minimization principal component analysis Robust principal component analysis Robust

More information

Low-rank Matrix Completion with Noisy Observations: a Quantitative Comparison

Low-rank Matrix Completion with Noisy Observations: a Quantitative Comparison Low-rank Matrix Completion with Noisy Observations: a Quantitative Comparison Raghunandan H. Keshavan, Andrea Montanari and Sewoong Oh Electrical Engineering and Statistics Department Stanford University,

More information

Proximal Gradient Descent and Acceleration. Ryan Tibshirani Convex Optimization /36-725

Proximal Gradient Descent and Acceleration. Ryan Tibshirani Convex Optimization /36-725 Proximal Gradient Descent and Acceleration Ryan Tibshirani Convex Optimization 10-725/36-725 Last time: subgradient method Consider the problem min f(x) with f convex, and dom(f) = R n. Subgradient method:

More information

Big Data Analytics: Optimization and Randomization

Big Data Analytics: Optimization and Randomization Big Data Analytics: Optimization and Randomization Tianbao Yang Tutorial@ACML 2015 Hong Kong Department of Computer Science, The University of Iowa, IA, USA Nov. 20, 2015 Yang Tutorial for ACML 15 Nov.

More information

Lecture 5 : Projections

Lecture 5 : Projections Lecture 5 : Projections EE227C. Lecturer: Professor Martin Wainwright. Scribe: Alvin Wan Up until now, we have seen convergence rates of unconstrained gradient descent. Now, we consider a constrained minimization

More information

About Split Proximal Algorithms for the Q-Lasso

About Split Proximal Algorithms for the Q-Lasso Thai Journal of Mathematics Volume 5 (207) Number : 7 http://thaijmath.in.cmu.ac.th ISSN 686-0209 About Split Proximal Algorithms for the Q-Lasso Abdellatif Moudafi Aix Marseille Université, CNRS-L.S.I.S

More information

Supplementary Materials: Bilinear Factor Matrix Norm Minimization for Robust PCA: Algorithms and Applications

Supplementary Materials: Bilinear Factor Matrix Norm Minimization for Robust PCA: Algorithms and Applications IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 16 Supplementary Materials: Bilinear Factor Matrix Norm Minimization for Robust PCA: Algorithms and Applications Fanhua Shang, Member, IEEE,

More information

Elastic-Net Regularization of Singular Values for Robust Subspace Learning

Elastic-Net Regularization of Singular Values for Robust Subspace Learning Elastic-Net Regularization of Singular Values for Robust Subspace Learning Eunwoo Kim Minsik Lee Songhwai Oh Department of ECE, ASRI, Seoul National University, Seoul, Korea Division of EE, Hanyang University,

More information

The convex algebraic geometry of rank minimization

The convex algebraic geometry of rank minimization The convex algebraic geometry of rank minimization Pablo A. Parrilo Laboratory for Information and Decision Systems Massachusetts Institute of Technology International Symposium on Mathematical Programming

More information

Compressed sensing. Or: the equation Ax = b, revisited. Terence Tao. Mahler Lecture Series. University of California, Los Angeles

Compressed sensing. Or: the equation Ax = b, revisited. Terence Tao. Mahler Lecture Series. University of California, Los Angeles Or: the equation Ax = b, revisited University of California, Los Angeles Mahler Lecture Series Acquiring signals Many types of real-world signals (e.g. sound, images, video) can be viewed as an n-dimensional

More information

Splitting Techniques in the Face of Huge Problem Sizes: Block-Coordinate and Block-Iterative Approaches

Splitting Techniques in the Face of Huge Problem Sizes: Block-Coordinate and Block-Iterative Approaches Splitting Techniques in the Face of Huge Problem Sizes: Block-Coordinate and Block-Iterative Approaches Patrick L. Combettes joint work with J.-C. Pesquet) Laboratoire Jacques-Louis Lions Faculté de Mathématiques

More information

arxiv: v1 [cs.cv] 1 Jun 2014

arxiv: v1 [cs.cv] 1 Jun 2014 l 1 -regularized Outlier Isolation and Regression arxiv:1406.0156v1 [cs.cv] 1 Jun 2014 Han Sheng Department of Electrical and Electronic Engineering, The University of Hong Kong, HKU Hong Kong, China sheng4151@gmail.com

More information

RECOVERING LOW-RANK AND SPARSE COMPONENTS OF MATRICES FROM INCOMPLETE AND NOISY OBSERVATIONS. December

RECOVERING LOW-RANK AND SPARSE COMPONENTS OF MATRICES FROM INCOMPLETE AND NOISY OBSERVATIONS. December RECOVERING LOW-RANK AND SPARSE COMPONENTS OF MATRICES FROM INCOMPLETE AND NOISY OBSERVATIONS MIN TAO AND XIAOMING YUAN December 31 2009 Abstract. Many applications arising in a variety of fields can be

More information

Exact Low-rank Matrix Recovery via Nonconvex M p -Minimization

Exact Low-rank Matrix Recovery via Nonconvex M p -Minimization Exact Low-rank Matrix Recovery via Nonconvex M p -Minimization Lingchen Kong and Naihua Xiu Department of Applied Mathematics, Beijing Jiaotong University, Beijing, 100044, People s Republic of China E-mail:

More information

Structured matrix factorizations. Example: Eigenfaces

Structured matrix factorizations. Example: Eigenfaces Structured matrix factorizations Example: Eigenfaces An extremely large variety of interesting and important problems in machine learning can be formulated as: Given a matrix, find a matrix and a matrix

More information

A Unified Approach to Proximal Algorithms using Bregman Distance

A Unified Approach to Proximal Algorithms using Bregman Distance A Unified Approach to Proximal Algorithms using Bregman Distance Yi Zhou a,, Yingbin Liang a, Lixin Shen b a Department of Electrical Engineering and Computer Science, Syracuse University b Department

More information

Tensor-Tensor Product Toolbox

Tensor-Tensor Product Toolbox Tensor-Tensor Product Toolbox 1 version 10 Canyi Lu canyilu@gmailcom Carnegie Mellon University https://githubcom/canyilu/tproduct June, 018 1 INTRODUCTION Tensors are higher-order extensions of matrices

More information

A Bregman alternating direction method of multipliers for sparse probabilistic Boolean network problem

A Bregman alternating direction method of multipliers for sparse probabilistic Boolean network problem A Bregman alternating direction method of multipliers for sparse probabilistic Boolean network problem Kangkang Deng, Zheng Peng Abstract: The main task of genetic regulatory networks is to construct a

More information

CS598 Machine Learning in Computational Biology (Lecture 5: Matrix - part 2) Professor Jian Peng Teaching Assistant: Rongda Zhu

CS598 Machine Learning in Computational Biology (Lecture 5: Matrix - part 2) Professor Jian Peng Teaching Assistant: Rongda Zhu CS598 Machine Learning in Computational Biology (Lecture 5: Matrix - part 2) Professor Jian Peng Teaching Assistant: Rongda Zhu Feature engineering is hard 1. Extract informative features from domain knowledge

More information

SPARSE signal representations have gained popularity in recent

SPARSE signal representations have gained popularity in recent 6958 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 57, NO. 10, OCTOBER 2011 Blind Compressed Sensing Sivan Gleichman and Yonina C. Eldar, Senior Member, IEEE Abstract The fundamental principle underlying

More information

Coordinate Update Algorithm Short Course Proximal Operators and Algorithms

Coordinate Update Algorithm Short Course Proximal Operators and Algorithms Coordinate Update Algorithm Short Course Proximal Operators and Algorithms Instructor: Wotao Yin (UCLA Math) Summer 2016 1 / 36 Why proximal? Newton s method: for C 2 -smooth, unconstrained problems allow

More information

LINEARIZED AUGMENTED LAGRANGIAN AND ALTERNATING DIRECTION METHODS FOR NUCLEAR NORM MINIMIZATION

LINEARIZED AUGMENTED LAGRANGIAN AND ALTERNATING DIRECTION METHODS FOR NUCLEAR NORM MINIMIZATION LINEARIZED AUGMENTED LAGRANGIAN AND ALTERNATING DIRECTION METHODS FOR NUCLEAR NORM MINIMIZATION JUNFENG YANG AND XIAOMING YUAN Abstract. The nuclear norm is widely used to induce low-rank solutions for

More information

Sparse and Low-Rank Matrix Decomposition Via Alternating Direction Method

Sparse and Low-Rank Matrix Decomposition Via Alternating Direction Method Sparse and Low-Rank Matrix Decomposition Via Alternating Direction Method Xiaoming Yuan a, 1 and Junfeng Yang b, 2 a Department of Mathematics, Hong Kong Baptist University, Hong Kong, China b Department

More information

Robust Sparse Recovery via Non-Convex Optimization

Robust Sparse Recovery via Non-Convex Optimization Robust Sparse Recovery via Non-Convex Optimization Laming Chen and Yuantao Gu Department of Electronic Engineering, Tsinghua University Homepage: http://gu.ee.tsinghua.edu.cn/ Email: gyt@tsinghua.edu.cn

More information

Machine learning for pervasive systems Classification in high-dimensional spaces

Machine learning for pervasive systems Classification in high-dimensional spaces Machine learning for pervasive systems Classification in high-dimensional spaces Department of Communications and Networking Aalto University, School of Electrical Engineering stephan.sigg@aalto.fi Version

More information

arxiv: v3 [math.oc] 18 Oct 2013

arxiv: v3 [math.oc] 18 Oct 2013 Noname manuscript No. (will be inserted by the editor) The Augmented Lagrange Multiplier Method for Exact Recovery of Corrupted Low-Rank Matrices Zhouchen Lin Minming Chen Yi Ma arxiv:1009.5055v3 [math.oc]

More information

Sparse representation classification and positive L1 minimization

Sparse representation classification and positive L1 minimization Sparse representation classification and positive L1 minimization Cencheng Shen Joint Work with Li Chen, Carey E. Priebe Applied Mathematics and Statistics Johns Hopkins University, August 5, 2014 Cencheng

More information

Lecture Notes 9: Constrained Optimization

Lecture Notes 9: Constrained Optimization Optimization-based data analysis Fall 017 Lecture Notes 9: Constrained Optimization 1 Compressed sensing 1.1 Underdetermined linear inverse problems Linear inverse problems model measurements of the form

More information

Least Squares Approximation

Least Squares Approximation Chapter 6 Least Squares Approximation As we saw in Chapter 5 we can interpret radial basis function interpolation as a constrained optimization problem. We now take this point of view again, but start

More information

ECE G: Special Topics in Signal Processing: Sparsity, Structure, and Inference

ECE G: Special Topics in Signal Processing: Sparsity, Structure, and Inference ECE 18-898G: Special Topics in Signal Processing: Sparsity, Structure, and Inference Low-rank matrix recovery via nonconvex optimization Yuejie Chi Department of Electrical and Computer Engineering Spring

More information

EE227 Project Report Robust Principal Component Analysis. Maximilian Balandat Walid Krichene Chi Pang Lam Ka Kit Lam

EE227 Project Report Robust Principal Component Analysis. Maximilian Balandat Walid Krichene Chi Pang Lam Ka Kit Lam EE227 Project Report Robust Principal Component Analysis Maximilian Balandat Walid Krichene Chi Pang Lam Ka Kit Lam May 10, 2012 May 10, 2012 1 Introduction Over the past decade there has been an explosion

More information

Contraction Methods for Convex Optimization and Monotone Variational Inequalities No.11

Contraction Methods for Convex Optimization and Monotone Variational Inequalities No.11 XI - 1 Contraction Methods for Convex Optimization and Monotone Variational Inequalities No.11 Alternating direction methods of multipliers for separable convex programming Bingsheng He Department of Mathematics

More information

Recovering any low-rank matrix, provably

Recovering any low-rank matrix, provably Recovering any low-rank matrix, provably Rachel Ward University of Texas at Austin October, 2014 Joint work with Yudong Chen (U.C. Berkeley), Srinadh Bhojanapalli and Sujay Sanghavi (U.T. Austin) Matrix

More information

Recent Developments of Alternating Direction Method of Multipliers with Multi-Block Variables

Recent Developments of Alternating Direction Method of Multipliers with Multi-Block Variables Recent Developments of Alternating Direction Method of Multipliers with Multi-Block Variables Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong 2014 Workshop

More information

On Optimal Frame Conditioners

On Optimal Frame Conditioners On Optimal Frame Conditioners Chae A. Clark Department of Mathematics University of Maryland, College Park Email: cclark18@math.umd.edu Kasso A. Okoudjou Department of Mathematics University of Maryland,

More information

An efficient ADMM algorithm for high dimensional precision matrix estimation via penalized quadratic loss

An efficient ADMM algorithm for high dimensional precision matrix estimation via penalized quadratic loss An efficient ADMM algorithm for high dimensional precision matrix estimation via penalized quadratic loss arxiv:1811.04545v1 [stat.co] 12 Nov 2018 Cheng Wang School of Mathematical Sciences, Shanghai Jiao

More information

Analysis of Greedy Algorithms

Analysis of Greedy Algorithms Analysis of Greedy Algorithms Jiahui Shen Florida State University Oct.26th Outline Introduction Regularity condition Analysis on orthogonal matching pursuit Analysis on forward-backward greedy algorithm

More information

Lecture 9: September 28

Lecture 9: September 28 0-725/36-725: Convex Optimization Fall 206 Lecturer: Ryan Tibshirani Lecture 9: September 28 Scribes: Yiming Wu, Ye Yuan, Zhihao Li Note: LaTeX template courtesy of UC Berkeley EECS dept. Disclaimer: These

More information

Augmented Lagrangian alternating direction method for matrix separation based on low-rank factorization

Augmented Lagrangian alternating direction method for matrix separation based on low-rank factorization Augmented Lagrangian alternating direction method for matrix separation based on low-rank factorization Yuan Shen Zaiwen Wen Yin Zhang January 11, 2011 Abstract The matrix separation problem aims to separate

More information

Robust Orthonormal Subspace Learning: Efficient Recovery of Corrupted Low-rank Matrices

Robust Orthonormal Subspace Learning: Efficient Recovery of Corrupted Low-rank Matrices Robust Orthonormal Subspace Learning: Efficient Recovery of Corrupted Low-rank Matrices Xianbiao Shu 1, Fatih Porikli 2 and Narendra Ahuja 1 1 University of Illinois at Urbana-Champaign, 2 Australian National

More information

ECE G: Special Topics in Signal Processing: Sparsity, Structure, and Inference

ECE G: Special Topics in Signal Processing: Sparsity, Structure, and Inference ECE 18-898G: Special Topics in Signal Processing: Sparsity, Structure, and Inference Sparse Recovery using L1 minimization - algorithms Yuejie Chi Department of Electrical and Computer Engineering Spring

More information

Bindel, Fall 2009 Matrix Computations (CS 6210) Week 8: Friday, Oct 17

Bindel, Fall 2009 Matrix Computations (CS 6210) Week 8: Friday, Oct 17 Logistics Week 8: Friday, Oct 17 1. HW 3 errata: in Problem 1, I meant to say p i < i, not that p i is strictly ascending my apologies. You would want p i > i if you were simply forming the matrices and

More information

Optimization for Compressed Sensing

Optimization for Compressed Sensing Optimization for Compressed Sensing Robert J. Vanderbei 2014 March 21 Dept. of Industrial & Systems Engineering University of Florida http://www.princeton.edu/ rvdb Lasso Regression The problem is to solve

More information

Dual and primal-dual methods

Dual and primal-dual methods ELE 538B: Large-Scale Optimization for Data Science Dual and primal-dual methods Yuxin Chen Princeton University, Spring 2018 Outline Dual proximal gradient method Primal-dual proximal gradient method

More information

Minimizing the Difference of L 1 and L 2 Norms with Applications

Minimizing the Difference of L 1 and L 2 Norms with Applications 1/36 Minimizing the Difference of L 1 and L 2 Norms with Department of Mathematical Sciences University of Texas Dallas May 31, 2017 Partially supported by NSF DMS 1522786 2/36 Outline 1 A nonconvex approach:

More information

EE 367 / CS 448I Computational Imaging and Display Notes: Image Deconvolution (lecture 6)

EE 367 / CS 448I Computational Imaging and Display Notes: Image Deconvolution (lecture 6) EE 367 / CS 448I Computational Imaging and Display Notes: Image Deconvolution (lecture 6) Gordon Wetzstein gordon.wetzstein@stanford.edu This document serves as a supplement to the material discussed in

More information

Robust Asymmetric Nonnegative Matrix Factorization

Robust Asymmetric Nonnegative Matrix Factorization Robust Asymmetric Nonnegative Matrix Factorization Hyenkyun Woo Korea Institue for Advanced Study Math @UCLA Feb. 11. 2015 Joint work with Haesun Park Georgia Institute of Technology, H. Woo (KIAS) RANMF

More information

Direct Robust Matrix Factorization for Anomaly Detection

Direct Robust Matrix Factorization for Anomaly Detection Direct Robust Matrix Factorization for Anomaly Detection Liang Xiong Machine Learning Department Carnegie Mellon University lxiong@cs.cmu.edu Xi Chen Machine Learning Department Carnegie Mellon University

More information

Sparse Solutions of an Undetermined Linear System

Sparse Solutions of an Undetermined Linear System 1 Sparse Solutions of an Undetermined Linear System Maddullah Almerdasy New York University Tandon School of Engineering arxiv:1702.07096v1 [math.oc] 23 Feb 2017 Abstract This work proposes a research

More information

arxiv: v1 [cs.cv] 17 Nov 2015

arxiv: v1 [cs.cv] 17 Nov 2015 Robust PCA via Nonconvex Rank Approximation Zhao Kang, Chong Peng, Qiang Cheng Department of Computer Science, Southern Illinois University, Carbondale, IL 6901, USA {zhao.kang, pchong, qcheng}@siu.edu

More information

GROUP-SPARSE SUBSPACE CLUSTERING WITH MISSING DATA

GROUP-SPARSE SUBSPACE CLUSTERING WITH MISSING DATA GROUP-SPARSE SUBSPACE CLUSTERING WITH MISSING DATA D Pimentel-Alarcón 1, L Balzano 2, R Marcia 3, R Nowak 1, R Willett 1 1 University of Wisconsin - Madison, 2 University of Michigan - Ann Arbor, 3 University

More information

arxiv: v3 [math.oc] 19 Oct 2017

arxiv: v3 [math.oc] 19 Oct 2017 Gradient descent with nonconvex constraints: local concavity determines convergence Rina Foygel Barber and Wooseok Ha arxiv:703.07755v3 [math.oc] 9 Oct 207 0.7.7 Abstract Many problems in high-dimensional

More information

Low-Rank Matrix Recovery

Low-Rank Matrix Recovery ELE 538B: Mathematics of High-Dimensional Data Low-Rank Matrix Recovery Yuxin Chen Princeton University, Fall 2018 Outline Motivation Problem setup Nuclear norm minimization RIP and low-rank matrix recovery

More information

Frank-Wolfe Method. Ryan Tibshirani Convex Optimization

Frank-Wolfe Method. Ryan Tibshirani Convex Optimization Frank-Wolfe Method Ryan Tibshirani Convex Optimization 10-725 Last time: ADMM For the problem min x,z f(x) + g(z) subject to Ax + Bz = c we form augmented Lagrangian (scaled form): L ρ (x, z, w) = f(x)

More information

Combining Sparsity with Physically-Meaningful Constraints in Sparse Parameter Estimation

Combining Sparsity with Physically-Meaningful Constraints in Sparse Parameter Estimation UIUC CSL Mar. 24 Combining Sparsity with Physically-Meaningful Constraints in Sparse Parameter Estimation Yuejie Chi Department of ECE and BMI Ohio State University Joint work with Yuxin Chen (Stanford).

More information

Low Rank Matrix Completion Formulation and Algorithm

Low Rank Matrix Completion Formulation and Algorithm 1 2 Low Rank Matrix Completion and Algorithm Jian Zhang Department of Computer Science, ETH Zurich zhangjianthu@gmail.com March 25, 2014 Movie Rating 1 2 Critic A 5 5 Critic B 6 5 Jian 9 8 Kind Guy B 9

More information

Optimization methods

Optimization methods Lecture notes 3 February 8, 016 1 Introduction Optimization methods In these notes we provide an overview of a selection of optimization methods. We focus on methods which rely on first-order information,

More information

Compressed Sensing and Robust Recovery of Low Rank Matrices

Compressed Sensing and Robust Recovery of Low Rank Matrices Compressed Sensing and Robust Recovery of Low Rank Matrices M. Fazel, E. Candès, B. Recht, P. Parrilo Electrical Engineering, University of Washington Applied and Computational Mathematics Dept., Caltech

More information

Lecture Notes 1: Vector spaces

Lecture Notes 1: Vector spaces Optimization-based data analysis Fall 2017 Lecture Notes 1: Vector spaces In this chapter we review certain basic concepts of linear algebra, highlighting their application to signal processing. 1 Vector

More information

Noisy Signal Recovery via Iterative Reweighted L1-Minimization

Noisy Signal Recovery via Iterative Reweighted L1-Minimization Noisy Signal Recovery via Iterative Reweighted L1-Minimization Deanna Needell UC Davis / Stanford University Asilomar SSC, November 2009 Problem Background Setup 1 Suppose x is an unknown signal in R d.

More information

Information Retrieval

Information Retrieval Introduction to Information CS276: Information and Web Search Christopher Manning and Pandu Nayak Lecture 13: Latent Semantic Indexing Ch. 18 Today s topic Latent Semantic Indexing Term-document matrices

More information

A new method on deterministic construction of the measurement matrix in compressed sensing

A new method on deterministic construction of the measurement matrix in compressed sensing A new method on deterministic construction of the measurement matrix in compressed sensing Qun Mo 1 arxiv:1503.01250v1 [cs.it] 4 Mar 2015 Abstract Construction on the measurement matrix A is a central

More information

Grassmann Averages for Scalable Robust PCA Supplementary Material

Grassmann Averages for Scalable Robust PCA Supplementary Material Grassmann Averages for Scalable Robust PCA Supplementary Material Søren Hauberg DTU Compute Lyngby, Denmark sohau@dtu.dk Aasa Feragen DIKU and MPIs Tübingen Denmark and Germany aasa@diku.dk Michael J.

More information