Sei sulla pagina 1di 4

Least Squares Pursuit for Sparse Signal Recovery

Nicolae Cleju
Faculty of Electronics, Telecommunications and Information Technology
Gheorghe Asachi Technical University of Iasi, Romania
and
Institute of Computer Science, Romanian Academy, Iasi, Romania
Email: {nikcleju}@etti.tuiasi.ro

AbstractThis paper presents a novel algorithm for sparse algorithm in this paper is based on a similar greedy approach,
signal recovery, named Least Squares Pursuit, based on least- but with different prioritization of the internal constraints
squares minimization followed by choosing the atom with the compared to OMP.
largest resulting coefficient, in a greedy one-by-one fashion. It
follows a similar approach to that of Orthogonal Matching This paper is organized as follows. Section II presents
Pursuit, but with different prioritization of the constraints. We the Least Squares Pursuit algorithm and its relation with
propose an efficient implementation for Least Squares Pursuit, Orthogonal Matching Pursuit, and Section III presents an
derive theoretical guarantees for signal recovery, and finally efficient implementation procedure. Guarantees for successful
present promising simulation results. recovery are derived in Section IV, and simulation results
I. I NTRODUCTION are presented in Section V. Finally, conclusions are drawn in
Section VI.
Sparse signal recovery and its applications has been an
active research field in signal processing for more than a II. T HE L EAST S QUARES P URSUIT ALGORITHM
decade. It is now well known [1] that a signal x which has A. Background
a sparse decomposition ? in some basis or overcomplete
dictionary Directly solving (eq. 2) is not feasible because the `0 norm
x = D ? , with k? k0 = k, is making the problem NP-hard. The class of greedy pursuit
algorithms, featuring well-known algorithms like Matching
can be recovered from a reduced set of linear measurements Pursuit and Orthogonal Matching Pursuit, try to estimate
forming an acquisition matrix M , possibly contaminated with the support in a one-by-one fashion. Specifically, Orthogonal
noise z of energy  = kzk22 : Matching Pursuit (OMP) selects a new atom at step k as the
atom number i most correlated with the current residual:
y = M x + z.
i = arg max kDT r(k) | = kDT (y D (k) )k,
The recovery requires solving the NP-complete optimization
problem and then updates the current solution estimate (k) by pro-
jecting y onto the set of all selected atoms until this moment,
x = D arg min kk0 with ky M Dk22 < . (1)
T (k) . In other words, y is approximated as best as possible
This is known as compressed sensing [2], and it has been using only the atoms from T (k) , while the remaining atoms,
(k)
extensively researched in the last 10 - 15 years. which make up the complement set Tc , are not used at all.
In the noiseless case, which we consider in this paper,  = 0 This can be expressed as follows:
and the problem reduces to that of finding the sparsest soution (k) = arg min ky Dk22 subject to T (k) = 0, (3)
(in the `0 sense) of an undetermined equation system. We c

therefore consider our canonical problem defined as: where T (k) indicates the restriction of to the indices in
c
(k)
arg min kk0 subject to y = D. (2) Tc .

The idea of our algorithm is to swap the data-fidelity term
where, for brevity we renamed the product M D from (eq. 1) ky Dk22 with sparsity term T (k) = 0 in (eq. 3). Note
c
as simply D. This paper presents a novel greedy algorithm for that neither of the two terms is a priori more important than
solving this problem. the other, since the true solution ? with the true support
The optimization problem (eq. 2) is known to be NP-hard, T satisfies both terms perfectly. Their relative importance
therefore one must settle with suboptimal solutions. One of the is just a choice of algorithm design, and OMP chooses to
most common algorithms for (approximately) solving (eq. 2) strictly enforce the sparsity term T (k) = 0, while accepting
c
is Orthogonal Matching Pursuit [3][4] (OMP). OMP iteratively approximate data-fidelity and working to improve it with each
estimates the support of the solution by progressively adding, new atom. We propose to take the opposite route: always
at every iteration, the most likely atom, i.e. the one most enforce perfect data-fidelity, ky Dk22 = 0, while accepting
similar to the current approximation residual. Our proposed non-exact sparsity T (k) , working to improve it with each new
c
Algorithm 1 Least Squares Pursuit The name Least Squares Pursuit (LSP) is justified by the fact
1: (0)
Initialize: T ,
(0)
{1, 2, ...N }
Tc that the algorithm follows the same one-by-one atom selection
2: repeat as OMP (hence the name Pursuit), but the orthogonality
3: Update current solution: induced by OMPs projection step is replaced with a least
squares problem.
(k) = arg min kIT (k) k22 subject to y = D
c
III. E FFICIENT IMPLEMENTATION
4: Update current support: A naive implementation would solve (eq. 4) anew for (k) at
i = arg max kT (k) k every iteration. In the following, we describe a more efficient
c
implementation that starts from the solution at the previous
(k+1) (k)
T =T {i} step (k 1) and updates it. This is analogue to the update
step in OMP.
Tc(k+1) = Tc(k) \ {i}
Consider the initial least-squares solution:
5: until stop criterion
(0) = arg min kk22 subject to y = D.
Since T (0) is empty, this is the full least-squares solution of
selected atom. Thus our update step, given the estimate support the system, and thus it can be represented compactly using the
T (k) at step k, is: Moore-Penrose pseudoinverse of D, denoted here as D :
(k) = arg min T (k) subject to y = D. (4) (0) = D y
c

This approach was inspired by the Greedy Analysis Pursuit .


algorithm [5] designed for the cosparse model (also known Suppose that at step k we have the current estimate of the
as analysis sparsity), which follows the similar approach of support T (k) . The new solution estimate (k) from (eq. 4)
favoring the data-fidelity term over the sparsity one. differs from (0) by the fact that the coefficients of the selected
atoms T (k) are not anymore restricted.
B. Least Squares Pursuit The estimate (k) can be split in two parts, corresponding
(k)
Our proposed algorithm consists of iterating the two usual to the estimate support T (k) and the cosupport Tc , leading
steps of pursuit methods: (i) update the solution estimate given to:
(k) (k)
the current support estimate, and (ii) update the support by y = DT (k) T (k) + DT (k) (k) . (5)
c Tc
choosing one new atom to add.
At step k, considering the current support estimate T (k) and Let us further consider the orthogonal projection operator
(k)
(k)
the remaining atoms in set Tc , the algorithm estimates the P (k) on the span of DT :
solution via a weighted least-squares decomposition: (k) (k)
P (k) = DT (DT ) ,
k
= arg min kIT (k) k22 subject to y = D,
c as well as the complementary orthogonal projection operator
Here, IT (k) is a unit matrix which has 1s only on the locations on its complement subspace:
c
(k) (k) (k)
corresponding to the indices in Tc . Using IT (k) is just a
c
Pc(k) = I P (k) = I DT (DT ) .
notation trick equivalent to simply stating T (k) from (eq. 4),
c
since its effect is to constrain the norm only of the coefficients With every new atom added to T , the rank of operator P (k)
from the set T (k) (the coefficients corresponding to the current is incremented, whereas the complementary subspace spanned
c (k)
support T (k) are left unconstrained). by Pc decreases. As such, the corresponding subspaces
(k)
Thus, we seek the decomposition of y which has minimum spanned by Pc are nested as follows:
norm for the coefficients of the non-selected atoms, with no
Pc(0) Pc(1) ... Pc(k) ...
restriction for the selected ones.
Once the solution estimate is found, the support is updated (k)
Applying Pc to (eq. 5) results in
by adding the atom i with the largest coefficient in T (k) :
c (k)
Pc(k) y = Pc(k) DT (k) (k) = Pc(k) D (k) , (6)
i = arg max kT (k) k, c Tc
c
(k) (k)
i.e. the atom most likely to belong in the support (in other since Pc is orthogonal to DT .
(k)
words, the atom most likely not to belong in the co-support). Since we want (k) of minimum `2 norm, it will be the
Tc
The algorithm is summarized as Algorithm 1. The stop least squares solution of (eq. 6), i.e. orthogonal to the null
criterion of the iterations can be, just like for OMP, a fixed space of this system. The null space of (eq. 6) consists of
number of steps (if sparsity level is known beforehand) or a the original null space of D plus the subspace spanned by
(k)
sufficiently small value of the objective kIT (k) k22 . D DT , as can be seen from the following two observations:
c
(k)
1. This subspace is orthogonal to the system matrix Pc D: for some u, meaning that IT c (k) (k) lives in the space spanned
(k) (k) by D. Let N designate a basis for the null space of D (row-
Pc(k) DD DT = P c DT = 0. wise). Then, left multiplying with N yields 0:
(k)
2. The subspace spanned by D DT has dimension |T | = NT (k)
(k)
= 0.
(k)
k and thus precisely matches the null space dimension c Tc

of (eq. 6). Moreover, it is embedded in the row space of Considering that there were no wrong atom choices until the
the original dictionary D, since the columns of D and current iteration k, the original co-support Tc is completely
(k) (k)
the rows of D always span the same subspace. inside Tc . Therefore, we can split the set Tc into the
(k)
These considerations show that the subspace spanned by remaining support we still want to find, T? , and the whole
(k)
D DT is the additional null space brought by the decreasing true co-support Tc . Splitting the previous equation in this
(k) manner leads to:
range of Pc in addition to the original null space of D.
How does (k) differ from the previous estimate (k1) ? (k) (k)
NTc Tc = NT (k) (k) . (8)
They are both least-squares solutions of their respective ? T?
systems (eq. 6), and thus orthogonal to their null spaces The difference between the true sparse solution ? and
(k1) (k) (k) (k1)
D DT and D DT . But DT consists of DT together (k)
lives in the null space, since both solutions produce y.
with the last selected atom, henceforth denoted dk , and there- Therefore
fore we can obtain (k) by merely projecting and removing (k) ? = N T v,
from (k1) the part which lays in this additional null space
spanned by D dk : for some vector v. Restricting the above only to the rows in
the true co-support Tc , and considering that ? is zero for
(k) = (I D dk {D dk } ) (k1) . these atoms, leaves
Since dk is a vector, D dk is also a vector, and in this case (k)
Tc = NTTc v,
its pseudoinverse {D dk } is simply the transposed vector
(k)
divided by its squared norm. It follows that: therefore Tc lives in the span of NTTc . As such, it holds that:
1 T
(k) = (k1) T
D dk dTk D (k1) . (k)
Tc = NTc NTc Tc .
(k)
dTk D D dk
Replacing eq. 8 in the above leads to:
We can use the above equation for updating the estimate
(k) (k)
solution after every new atom is added. Tc = NTc NT (k) (k) .
? T?
Alternatively, we can consider the whole set DT (k) of atoms
selected until step k, and compute the solution at time k Thus, having
starting every time from the initial solution (0) = D y: kNTc NT (k) k, < 1
?

(k) = D y D DT (k) DT (k) T D T


D y, is a sufficient condition that
(k)
i.e. kTc k < kT (k) k ,
?

T thereby guaranteeing that atom selection via the largest coeffi-


(k) = D (In DT (k) DT (k) T D D )y. cient in (k) never goes wrong and the chosen atom is always
(k)
IV. G UARANTEE FOR SUCCESSFUL RECOVERY from the remaining true support T? .
(k)
Since the set T? at step k can be in general as large as
In this section, we derive a guarantee for successful signal
the original support T (e.g. before the first atom selection), the
recovery with Least Squares Pursuit. The approach is inspired
overall condition guaranteeing correct atom selection at every
by the similar guarantees of Greedy Analysis Pursuit and
iteration k is:
Orthogonal Matching Pursuit.
kNTc NT k, < 1.
We start from the unconstrained version of (eq. 4), i.e.:
V. S IMULATION RESULTS
k = arg min ky Dk22 + kIT c k2 . (7)

To investigate the performance of Least Squares Pursuit
where 0. The condition 0 ensures that the sparsity as a sparse recovery algorithm, we perform a recovery test
term kIT c k2 is infinitely less important than the the data on synthetic sparse data, and represent the phase-transition
fidelity term ky Dk22 , which means that the latter will be obtained.
essentially a strict constraint, thus leading to (eq. 4). Note that Specifically, we generate a random dictionary of size n
leads to the OMP problem, conversely. N = 200 240, and generate exact-sparse data by randomly
From the Lagrangian of this unconstrained formulation, we combining atoms from the dictionary. The data is then acquired
get: with an acquisition matrix of varying number of rows m (no
ITc (k) (k) = DT u, noise is added), and then we perform reconstruction. A signal
is considered successfully recovered if the relative error of the in the general unconstrained case. Also, the performance of
reconstruction is less than 106 . LSP with noisy measurements must be analyzed and compared
We plot the percentage of successfully recovered signals as with that of OMP. Finally, the OMP recovery guarantees are
a function of = m n (e.g. compression ratio of acquisition)
also valid for `1 -minimization based recovery [7], something
k
and = m (relative sparsity). The values of and go from which remains to be investigated in the case of LSP.
0.1 to 0.95 in steps of 0.05, and we average 100 recovered
ACKNOWLEDGEMENT
signals for every pair (, ).
The graphical representation of Fig. 1 constitutes a phase- This work was supported by a grant of the Romanian
transition plot [6] which indicates the area where Least National Authority for Scientific Research and Innovation,
Squares Pursuit is capable of successful reconstruction. Here, CNCS UEFISCDI, project number PN-II-RU-TE-2014-4-
white indicates 100% recovered signals, black 0%, and levels 0832 Medical signal processing methods based on com-
of gray indicate intermediate performance, correspondingly. pressed sensing; applications and their implementation.
R EFERENCES
0.95
[1] E. J. Candes, J. K. Romberg, and T. Tao, Stable signal
recovery from incomplete and inaccurate measurements,
Communications on Pure and Applied Mathematics, vol.
59, pp. 12071223, 2006, ISSN: 0010-3640.
[2] D. L. Donoho, Compressed sensing, IEEE Transactions
on Information Theory, vol. 52, no. 4, pp. 12891306,
2006.
0.5

[3] Y. C. Pati, R. Rezaiifar, and P. S. Krishnaprasad, Orthog-


onal matching pursuit: Recursive function approximation
with applications to wavelet decomposition, in Proc.
27th Annual Asilomar Conf. on Signals, Systems, and
Computers, 1993, pp. 4044.
[4] G. Davis, S. Mallat, and M. Avellaneda, Adaptive
0.1 greedy approximations, Constructive approximation,
0.1 0.5 0.95 pp. 5798, 1997.
[5] S. Nam, M. Davies, M. Elad, and R. Gribonval,
Fig. 1: Sparse recovery with LSP from noiseless measure- Cosparse analysis modeling - uniqueness and algo-
ments: percent of sparse signals successfully reconstructed. rithms, in Proc. ICASSP 2011, 2011, pp. 58045807.
DOI : 10.1109/ICASSP.2011.5947680.
White indicates 100% recovered signals, black 0%.
[6] D. L. Donoho and J. Tanner, Precise undersampling
The results show that Least Squares Pursuit is a viable theorems, Proceedings of the IEEE, vol. 98, no. 6,
algorithm for exact-sparse signal recovery, with good recovery pp. 913924, 2010. DOI: 10.1109/JPROC.2010.2045630.
capabilities in a wide area. The area of successful recovery [7] J. Tropp, Greed is good: Algorithmic results for sparse
(white zone) is comparable to the one obtained with similar approximation, Information Theory, IEEE Transactions
algorithms like Orthogonal Matching Pursuit or Greedy Anal- on, vol. 50, no. 10, pp. 22312242, Oct. 2004, ISSN:
ysis Pursuit (the latter for analysis-based recovery), which are 0018-9448. DOI: 10.1109/TIT.2004.834793.
however not included due to space reasons.
VI. C ONCLUSIONS AND FUTURE WORK
In this paper we proposed the Least Squares Pursuit algo-
rithm for sparse signal recovery, based on iteratively choosing
the largest atom from successive least-squares decomposition.
This approach is related to that of Orthogonal Matching
Pursuit, but with different prioritization of the data-fidelity
and sparsity constraints, and also with that of Greedy Analysis
Pursuit from the cosparse model. Although we only analyzed
the noiseless case, the results are promising and similar to
those of OMP.
Many avenues remain open for future work. Both OMP and
LSP can be considered as particular cases of an unconstrained
formulation (eq. 7) of the optimization problem, which opens
the way to having common theoretical analysis and guarantees

Potrebbero piacerti anche