Abstract
The recently proposed modified-compressive sensing (modified-CS), which utilizes the partially known support as prior knowledge, significantly improves the performance of recovering sparse signals. However, modified-CS depends heavily on the reliability of the known support. An important problem, which must be studied further, is the recoverability of modified-CS when the known support contains a number of errors. In this letter, we analyze the recoverability of modified-CS in a stochastic framework. A sufficient and necessary condition is established for exact recovery of a sparse signal. Utilizing this condition, the recovery probability that reflects the recoverability of modified-CS can be computed explicitly for a sparse signal with 
nonzero entries. Simulation experiments have been carried out to validate our theoretical results.
Citation: Zhang J, Li Y, Gu Z, Yu ZL (2014) Recoverability Analysis for Modified Compressive Sensing with Partially Known Support. PLoS ONE 9(2):
e87985.
https://doi.org/10.1371/journal.pone.0087985
Editor: Holger Fröhlich, University of Bonn, Bonn-Aachen International Center for IT, Germany
Received: August 7, 2013; Accepted: January 2, 2014; Published: February 10, 2014
Copyright: © 2014 Zhang et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was supported by the National High-tech R&D Program of China (863 Program) under grant 2012AA011601, the National Natural Science Foundation of China under grants 91120305, 61105121 and 61175114, the Natural Science Foundation of Guangdong under grant S2012020010945 and the Excellent Youth Development Project of Universities in Guangdong Province under grant 2012LYM 0057. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
A central problem in CS is the following: given an 
matrix 
(
), and a measurement vector 
, recover 
. To deal with this problem, the most extensively studied recovery method is the 
-minimization approach (Basis Pursuit) [1]–[5]
(1)This convex problem can be solved efficiently; moreover, 
probabilistic measurements are sufficient for it to recover a 
-sparse vector 
(i.e., all but at most 
entries are zero) exactly.
Recently, Vaswani and Lu [6]–[9], Miosso [10], [11], Wang and Yin [12], [13], Friedlander et.al [14], Jacques [15] have shown that exact recovery based on fewer measurements than those needed for the 
-minimization approach is possible when the support of 
is partially known. The recovery is implemented by solving the optimization problem.
(2)where T denotes the “known” part of support, 
, 
is a column vector composed of the entries of 
with their indices being in 
. This method is named modified-CS [6] or truncated 
minimization [12]. One application of the modified-CS is the recovery of (time) sequences of sparse signals, such as dynamic magnetic resonance imaging (MRI) [8], [9]. Since the support evolve slowly over time, the previously recovered support can be used as known part for later reconstruction.
As an important performance index of modified-CS, its recoverability, i.e., when is the solution of (2) equal to 
, has been discussed in several papers. In [6], a sufficient condition on the recoverability was obtained based on restricted isometry property. From the view of t-null space property, another sufficient condition to recover 
-sparse vectors was proposed in [12]. However, there always exist some signals that do not satisfy these conditions but still can be recovered. Specifically, in real-world applications, the known support often contains some errors. The existing sufficient conditions can not reflect accurately the recoverability of modified-CS in many cases. Therefore, it is necessary to develop alternative techniques for analyzing the recoverability of modified-CS.
In this paper, a sufficient and necessary condition (SNC) on the recoverability of modified-CS is derived. Then, we discuss the recoverability of modified-CS in a probabilistic way. The main advantage of our work is that, for a randomly given vector 
with 
nonzero entries, the exact recovery percentage of modified-CS can be computed explicitly under a given matrix 
and a randomly given 
that satisfied 
but includes 
errors, where 
denotes the size of the known support 
. Hence, this paper provides a quantitative index to measure the reliability of modified-CS in real-world applications. Simulation experiments validate our results.
Materials and Methods
1 A Sufficient and Necessary Condition for Exact Recovery
In this subsection, a SNC on the recoverability of modified-CS is derived. Firstly, we give some notations in the follows. The support of vector 
is denoted by 
, i.e. 
. Suppose 
can be split as 
, where 
is the unknown part of the support and 
is set of errors in the known part support 
. The set operations 
and 
stand for set union and set difference respectively. Let 
denote the solution of the model in (2) and 
denote the set of all subsets of 
. A SNC on the recoverability of modified-CS is given in the following theorem, which is an extension of a result in [16].
Theorem 1 For a given vector 
, 
, if and only if 
, the optimal value of the objective function of the following optimization problem is greater than zero, provided that this optimization problem is solvable:
(3)where 
.
The proof of this theorem is given in Appendix S1.
Remark 1: For a given measurement matrix 
, the recoverability of the sparse vector 
based on the model in (2) depends only on the index set of nonzeros of 
in 
and the signs of these nonzeros. In other words, the recoverability relies only on the sign pattern of 
in 
instead of the magnitudes of these nonzeros.
Remark 2: It follows from the proof of Theorem 1 that, even if 
contains several errors, Theorem 1 still holds.
Remark 3: Recently, the recoverability analysis of the modified-CS were reported in [6] and [12]. However, we establish a sufficient and necessary condition for the modified-CS to exactly reconstructs a sparse vector, which differs from the sufficient conditions proposed in these works.
2 Probability Estimation on Recoverability of Modified-CS
In this subsection, we utilize Theorem 1 to estimate the probability that the vector 
can be recovered by modified-CS, i.e., the conditional probability 
, where 
is defined as the number of nonzero entries of 
, 
and 
denote the size of 
and 
respectively. This probability reflects the recoverability of modified-CS, and is hereafter named as recovery probability.
Let 
denote the index set 
, it is easy to know that there are 
index subsets of 
with size 
. We denote these subsets as 
, 
. For each 
, there are 
subsets with size 
. We denote these subsets as 
. At the same time, for the set 
(the index set of the zero entries of 
), there are 
subsets with size 
. These subsets are denoted as 
. Firstly, we discuss the estimation of the recovery probability under the following assumption.
Assumption 1 The index set 
of the 
nonzero entries of 
can be one of the 
index sets 
, 
, with equal probability. The index set 
of 
errors in known support can be one of the 
index sets 
, 
, with equal probability. The index set 
of 
nonzero entries can be one of the 
index sets 
, 
, with equal probability. All the nonzero entries of the vector 
take either positive or negative sign with equal probability.
For a given vector 
and the known support 
, there is a sign column vector 
in 
. The recoverability of the vector 
only relates with the sign column vector 
(see Remark 1). Under the conditions that the index set of the nonzero entries of 
is 
and the known support 
is 
, it is easy to derive that 
contains 
indexes of the nonzeros of 
, where 
. Then there are 
sign column vectors. Among these sign column vectors, suppose that 
sign column vectors can be recovered, then 
is the probability of a vector 
being recovered by solving the modified-CS. Hence, following Assumption 1, the recovery probability is calculated by
(4)where 
, 
and 
. Because the measurement matrix 
is known, we can determine 
in (4) by checking whether the SNC (3) is satisfied for all the 
sign column vectors corresponding to the index set 
, 
and 
.
Because many practical situations such as Electroencephalogram (EEG) signals in wavelet domain do not completely satisfy those assumptions in “Assumption 1”. we further extend our analysis to more general case. Without loss of generality, we have the following assumption
Assumption 2 The index set 
of the 
nonzero entries of 
can be one of the 
index sets 
, 
, with probability 
. The index set 
of 
errors in known support can be one of the 
index sets 
, 
, with probability 
. The index set 
of 
nonzero entries can be one of the 
index sets 
, 
, with probability 
. All the nonzero entries of the vector 
take either positive or negative sign with probability 
or 
respectively.
Similarly, suppose the index set of the nonzero entries of 
is 
and the known support is 
, there are 
sign column vectors. Since all the nonzero entries of the vector 
take either positive or negative sign with probability 
or 
respectively, the probability of the sign pattern of vector 
equals one of 
sign column vectors is 
, where 
denotes the number of negative signs in this sign column vector and 
. Obviously, there are 
sign column vectors that has 
negative signs. Among these vectors, suppose that 
sign column vectors can be recovered, then 
is the probability of the vector 
being recovered by solving the modified-CS. Hence, under the Assumption 2, the recovery probability is calculated by
(5)where 
, 
and 
. Because the measurement matrix 
is known, we can determine 
in (5), 
, by checking whether the SNC (3) is satisfied for all the 
sign column vectors corresponding to the index set 
, 
and 
.
Remark 4: Equation (4) is a special case of equation (5) under the equal probability assumption.
However, the computational burden to calculate (5) increases exponentially as the problem dimensions increase. For each sign column vector 
and the corresponding index set 
, 
and 
, we denote the quads 
, where 
, 
, 
and 
. Suppose 
is a set composed by all the quads, there are 
elements in set 
. For each element of set 
, if the sign column vector 
can be recovered by modified-CS with a given measurement matrix 
and know support 
, we call the quad can be recovered. In (5), the estimation of recover probability need to check the total number of quads in set 
. When 
increases, the computational burden will increase exponentially. To avoid the computational burden problem, we state the following Theorem.
Theorem 2 Suppose that 
quads are randomly taken from set 
, where 
is a large positive integer (
), and 
of the 
quads can be recovered by solving modified-CS. Then
(6)The proof of this theorem is given in Appendix S2.
Remark 5: In real-world applications, by sampling randomly 
sign vectors with 
nonzero entries, we can check the number of the vectors that can be exact recovered by modified-CS with a random known support 
whose size is 
but contains 
errors. Suppose 
sign vectors can be recovered, the recovery probability 
can be computed approximately through calculating the ratio of 
.
Remark 6: It is well-known that certifying the restricted isometry property is hard, while based on the proposed method, the recoverability probability that reflects the recoverability of modified-CS can be computed explicitly.
From the proof of Theorem 2, the sampling numbers 
, which controls the precision in the approximation of (6), is related to the two-point distribution of 
other than the size of 
. Thus, there is no need for 
increasing exponentially as 
increases.
Results and Discussion
In this section, simulation examples on both synthesis data and real-world data have been conducted to demonstrate the validity of our theoretical results.
Example 1: In this example, the conclusion in Theorem 2 are demonstrated.
According to the uniform distribution in [−0.5, 0.5], we randomly generate three matrices 
(
) with (
, 
) = (7, 9), (48, 128) and (182, 1280) respectively. For matrices 
, 
and 
, we set (
, 
, 
) = (4, 2, 1), (20, 8, 2) and (60, 32, 4) respectively. As 
increases in their three cases, the number of sign vectors increases exponentially. For example, for 
, 
, the set 
contains approximately 
and 
elements respectively. Hence, for their three cases, we estimate the probabilities 
by the sampling method. For each case, we sample 
= 100, 500, 1000, 5000, 10000 respectively. The resultant probability estimates depicted in Fig. 1 indicate that 1) the estimation precision of the sampling method is stable in our experiments with different samplings. Therefore, we only need a very few samplings to obtain the satisfied estimation precision in real-world applications; 2) as 
increases in three cases, the sampling 
don't increase exponentially.
Example 2: Suppose 
was taken according to the uniform distribution in [−0.5, 0.5]. This example contains two parts in which the recovery probability estimates (4) and (5) are considered in simulation, respectively.
- All nonzero entries of the sparse vector
were drawn from a uniform distribution valued in the range [−1, +1]. Without loss of generality, we set 
. For a vector 
with 
nonzero entries, where 
= 2, 3,…, 7, we calculated the recovery probabilities by (4), where 
respectively. For every 
(
) nonzero entries, we also sampled 1000 vectors with random indices. For each vector, we solved the modified-CS with a randomly given 
, whose size equals to 
but contains 
errors, and checked whether the solution is equal to the true vector. Suppose that 
vectors can be recovered, we calculated the ratio 
as the recovery probability 
. The experimental results are presented in Fig. 2(a). Therein, solid curves denote the theoretic recovery probability estimated by (4). Dotted curves denote probabilities 
. Experimental results show that the theoretical estimates fit the simulated values very well.
- Now we consider the probability estimate (5). We suppose that all nonzero entries of the sparse vector
were drawn from a uniform distribution valued in the range [−0.5, +1]. Obviously, the nonzero entries of the vector 
take the positive sign with probability 
or the negative sign with probability 
. Similarly, we set 
. For a vector 
with 
nonzero entries, where 
= 2, 3,…, 7, we randomly generate the probabilities 
where 
. For an index set 
whose size equals to 
but contains 
errors, we randomly generate the probabilities 
where 
and 
where 
. The recovery probabilities are calculated by (5), where 
respectively. For every 
(
) nonzero entries, we also sampled 1000 vectors 
and 
that satisfy the assumption 2 with the above-generated probabilities. For each vector and 
, we solved the modified-CS and checked whether the solution is equal to the true vector. Finally, the 
can be calculated with the same way in the part I. We present the experimental results in Fig. 2(b). Therein, solid curves denote the theoretic recovery probability estimated by (5). Dotted curves denote probabilities 
. Experimental results show that in the general case, the theoretical estimates also fit the simulated values very well.
Example 3: In this example, we test on real-world ECG reconstruction to demonstrate the accuracy of the probability estimation by (5).
Firstly, eight ECG data have been chosen from the MIT-BIH arrhythmia database [17] as the test signals. Each data file includes two-channel ambulatory ECG recordings, and each channel contains 650000 binary data instances in a 16-bits data format, including the index and amplitude. In our simulation, ECG vector 
is extracted from the original data at the window size 
. A random sparse binary matrix [18] is used as our sensing matrix 
and we use D6 Daubechies wavelet dictionary 
to represent ECG segment, i.e.,
(7)It is well-known that vector 
is not strictly sparse, but can be approximated by 
-sparse vector. Therefore, to obtain the 
-sparse approximation 
of vector 
, we calculate the standard derivation 
of the high-frequency coefficients in vector 
and shrink the coefficients whose magnitudes are less than 3
to zero. We define the theoretic recovery of ECG segment 
as the SNC (3) can be satisfied for the sign pattern sign(
). On the other hand, for the recovery of a compressible vector 
, the best one can expect is that the solution of modified-CS and 
have their nonzero components at the same locations [19]. Considering the noise contamination, we think the ECG segment 
is recovered in practice if the solution of modified-CS and 
have the overwhelming majority (e.g. 95%) of their nonzero components at the same locations.
Hence, we randomly extracted 100 segments from each ECG data. According to the Theorem 2, we can estimate the recovery probability of modified-CS through calculating how many segments can be recovered, i.e., the recovery ratio. In our experiment, we suppose 
is the index set of low-frequency coefficients in vector 
. On the one hand, we check the SNC (3) for all the sign patterns sign(
) of these segments to obtain the theoretic recovery ratio; on the other hand, we obtain the practical recovery ratio by checking whether these ECG segments can be recovered in practice. For illustration, an original segment of record No. 100 in the MIT-BIH arrhythmia database and its wavelet coefficients are plotted in Fig. 3(a). At the same time, the reconstructed ones in time and wavelet domain are shown in Fig. 3(b). We present the experimental results of eight ECG data in Fig. 4. Therein, red curves denote the theoretic recovery probabilities. Blue curves denote practical recovery probabilities. Experimental results show that the proposed probability estimation is very accurate.
Conclusion
In this letter we study the recoverability of the modified-CS in a stochastic framework. A sufficient and necessary condition on the recoverability is presented. Based on this condition, the recovery probability of the modified-CS can be estimated explicitly. It is worth mentioning that Theorem 1 can be easy to extend to weighted-
minimization approach that was proposed in [20] for nonuniform sparse model. Moreover, the recovery probability estimation provides alternative way to find (numerically) the optimal set of weights in weighted-
minimization approach, which has the largest recovery probability to recover the signals.
Acknowledgments
The authors would like to thank anonymous reviewers and Academic Editor for the insightful and constructive suggestions.
Author Contributions
Conceived and designed the experiments: JZ YL. Performed the experiments: JZ ZG. Analyzed the data: JZ ZY. Contributed reagents/materials/analysis tools: JZ YL ZG ZY. Wrote the paper: JZ YL.
References
- 1.
Donoho DL (2006) Compressed sensing. IEEE Trans Inf Theory 52: 1289–1306.
- 2.
Candès EJ, Romberg J, Tao T (2006) Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information. IEEE Trans Inf Theory 52: 489–509.
- 3.
Candès EJ, Tao T (2006) Near-optimal signal recovery from random projections: universal encoding strategies? IEEE Trans Inf Theory 52: 5406–5425.
- 4.
Chen SS, Donoho DL, Saunders MA (2001) Atomic decomposition by basis pursuit. SIAM REV 43: 129–159.
- 5.
Candes E, Tao T (2005) Decoding by linear programming. IEEE Trans Inf Theory 51: 4203–4215.
- 6.
Vaswani N, Lu W (2010) Modified-CS: modifying compressive sensing for problems with partially known support. IEEE Trans Signal Process 58: 4595–4607.
- 7.
Vaswani N, Lu W (2009) Modified-CS: modifying compressive sensing for problems with partially known support. In: Proc IEEE Int Symp Inf Theory (ISIT). pp. 488–492.
- 8.
Lu W, Vaswani N (2009) Modified compressive sensing for real-time dynamic mr imaging. In: Proc IEEE Int Conf Image Proc (ICIP). pp. 3009–3012.
- 9.
Qiu C, Lu W, Vaswani N (2009) Real-time dynamic mri reconstruction using kalman filtered cs. In: Proc Int Conf Acoustics, Speech, Signal Processing(ICASSP). pp. 393–396.
- 10.
Miosso C, von Borries R, Argàez M, Velazquez L, Quintero C, et al. (2009) Compressive sensing reconstruction with prior information by iteratively reweighted least-squares. IEEE Trans Signal Process 57: 2424–2431.
- 11.
von Borries R, Miosso C, Potes C (2007) Compressed sensing using prior information. In: Computational Advances in Multi-Sensor Adaptive Processing, 2007. CAMPSAP 2007. 2nd IEEE International Workshop on. pp. 121–124.
- 12.
Wang Y, Yin W (2010) sparse signal reconstruction via iterative support detection. SIAM J Imag Sci 3: 462–491.
- 13.
Guo W, Yin W (2010) Edgecs: edge guided compressive sensing reconstruction. Technical Report TR10-02, Department of Computational and Applied Mathmatics, Rice University, Houston, TX.
- 14.
Friedlander M, Mansour H, Saab R, Yilmaz O (2010) Recovering compressively sampled signals using partial support information. to appear in the IEEE Trans Inf Theory.
- 15.
Jacques L (2010) A short note on compressed sensing with partially known signal support. Signal Processing 90: 3308–3312.
- 16.
Li Y, Amari S, Cichocki A, Guan C (2006) Probability estimation for recoverability analysis of blind source separation based on sparse representation. IEEE Trans Inf Theory 52: 3139–3152.
- 17.
Moody G, Mark R (2001) The impact of the mit-bih arrhythmia database. IEEE Eng Med Biol Mag 20: 45–50.
- 18.
Mamaghanian H, Khaled N, Atienza D, Vandergheynst P (2011) Compressed sensing for realtime energy-efficient ecg compression on wireless body sensor nodes. IEEE TransBiomed Eng 58: 2456–2466.
- 19.
Fuchs JJ (2005) Recovery of exact sparse representations in the presence of bounded noise. IEEE Trans Inf Theory 51: 3601–3608.
- 20.
Khajehnejad M, Xu W, Avestimehr S, Hassibi B (2011) Analyzing weighted ℓ1 minimization for sparse recovery with nonuniform sparse models. IEEE Trans Signal Process 59: 1985–2001.
