Compressive Sensing

Compressive Sensing
Compressive Sensing
Massimo Fornasier and Holger Rauhut
Austrian Academy of Sciences
Johann Radon Institute for
Computational and Applied
Mathematics (RICAM)
Altenbergerstrasse 69
A-4040, Linz, Austria
[email protected]
Hausdorff Center for Mathematics,
Institute for Numerical Simulation
University of Bonn
Endenicher Allee 60
D-53115 Bonn, Germany
[email protected]
April 18, 2010
Abstract
Compressive sensing is a new type of sampling theory, which predicts that sparse signals and images can be reconstructed from what
was previously believed to be incomplete information. As a main feature, efficient algorithms such as ℓ1 -minimization can be used for recovery. The theory has many potential applications in signal processing
and imaging. This chapter gives an introduction and overview on both
theoretical and numerical aspects of compressive sensing.
1
Introduction
The traditional approach of reconstructing signals or images from measured
data follows the well-known Shannon sampling theorem [94], which states
that the sampling rate must be twice the highest frequency. Similarly, the
fundamental theorem of linear algebra suggests that the number of collected
samples (measurements) of a discrete finite-dimensional signal should be at
least as large as its length (its dimension) in order to ensure reconstruction.
This principle underlies most devices of current technology, such as analog to
digital conversion, medical imaging or audio and video electronics. The novel
theory of compressive sensing (CS) — also known under the terminology of
compressed sensing, compressive sampling or sparse recovery — provides
a fundamentally new approach to data acquisition which overcomes this
1
common wisdom. It predicts that certain signals or images can be recovered
from what was previously believed to be highly incomplete measurements
(information). This chapter gives an introduction to this new field. Both
fundamental theoretical and algorithmic aspects are presented, with the
awareness that it is impossible to retrace in a few pages all the current
developments of this field, which was growing very rapidly in the past few
years and undergoes significant advances on an almost daily basis.
CS relies on the empirical observation that many types of signals or images can be well-approximated by a sparse expansion in terms of a suitable
basis, that is, by only a small number of non-zero coefficients. This is the
key to the efficiency of many lossy compression techniques such as JPEG,
MP3 etc. A compression is obtained by simply storing only the largest basis
coefficients. When reconstructing the signal the non-stored coefficients are
simply set to zero. This is certainly a reasonable strategy when full information of the signal is available. However, when the signal first has to be
acquired by a somewhat costly, lengthy or otherwise difficult measurement
(sensing) procedure, this seems to be a waste of resources: First, large efforts
are spent in order to obtain full information on the signal, and afterwards
most of the information is thrown away at the compression stage. One might
ask whether there is a clever way of obtaining the compressed version of the
signal more directly, by taking only a small number of measurements of the
signal. It is not obvious at all whether this is possible since measuring directly the large coefficients requires to know a priori their location. Quite
surprisingly, compressive sensing provides nevertheless a way of reconstructing a compressed version of the original signal by taking only a small amount
of linear and non-adaptive measurements. The precise number of required
measurements is comparable to the compressed size of the signal. Clearly,
the measurements have to be suitably designed. It is a remarkable fact that
all provably good measurement matrices designed so far are random matrices. It is for this reason that the theory of compressive sensing uses a lot of
tools from probability theory.
It is another important feature of compressive sensing that practical
reconstruction can be performed by using efficient algorithms. Since the
interest is in the vastly undersampled case, the linear system describing the
measurements is underdetermined and therefore has infinitely many solution. The key idea is that the sparsity helps in isolating the original vector.
The first naive approach to a reconstruction algorithm consists in searching for the sparsest vector that is consistent with the linear measurements.
This leads to the combinatorial ℓ0 -problem, see (3.4) below, which unfortunately is NP-hard in general. There are essentially two approaches for
2
tractable alternative algorithms. The first is convex relaxation leading to
ℓ1 -minimization — also known as basis pursuit, see (3.5) — while the second
constructs greedy algorithms. This overview focuses on ℓ1 -minimization. By
now basic properties of the measurement matrix which ensure sparse recovery by ℓ1 -minimization are known: the null space property (NSP) and the
restricted isometry property (RIP). The latter requires that all column submatrices of a certain size of the measurement matrix are well-conditioned.
This is where probabilistic methods come into play because it is quite hard to
analyze these properties for deterministic matrices with minimal amount of
measurements. Among the provably good measurement matrices are Gaussian, Bernoulli random matrices, and partial random Fourier matrices.
(a)
(b)
(c)
(d)
Figure 1: (a) 10-sparse Fourier spectrum, (b) time domain signal of length
300 with 30 samples, (c) reconstruction via ℓ2 -minimization, (d) exact reconstruction via ℓ1 -minimization
3
Figure 1 serves as a first illustration of the power of compressive sensing.
It shows an example for recovery of a 10-sparse signal x ∈ C300 from only
30 samples (indicated by the red dots in Figure 1(b)). From a first look at
the time-domain signal, one would rather believe that reconstruction should
be impossible from only 30 samples. Indeed, the spectrum reconstructed by
traditional ℓ2 -minimization is very different from the true spectrum. Quite
surprisingly, ℓ1 -minimization performs nevertheless an exact reconstruction,
that is, with no recovery error at all!
Sampling domain in the frequency plane
(a)
(b)
26 iterations
126 iterations
(c)
(d)
Figure 2: (a) Sampling data of the NMR image in the Fourier domain which
corresponds to only 0.11% of all samples. (b) Reconstruction by backprojection.
(c) Intermediate iteration of an efficient algorithm for large scale total variation
minimization. (d) The final reconstruction is exact.
An example from nuclear magnetic resonance imaging serves as a second
illustration. Here, the device scans a patient by taking 2D or 3D frequency
measurements within a radial geometry. Figure 2(a) describes such a sam4
pling set of a 2D Fourier transform. Since a lengthy scanning procedure
is very uncomfortable for the patient it is desired to take only a minimal
amount of measurements. Total variation minimization, which is closely
related to ℓ1 -minimization, is then considered as recovery method. For comparison, Figure 2(b) shows the recovery by a traditional backprojection algorithm. Figures 2(c), 2(d) display iterations of an algorithm, which was
proposed and analyzed in [40] to perform efficient large scale total variation
minimization. The reconstruction in Figure 2(d) is again exact!
2
Background
Although the term compressed sensing (compressive sensing) was coined
only recently with the paper by Donoho [26], followed by a huge research
activity, such a development did not start out of thin air. There were certain
roots and predecessors in application areas such as image processing, geophysics, medical imaging, computer science as well as in pure mathematics.
An attempt is made to put such roots and current developments into context
below, although only a partial overview can be given due to the numerous
and diverse connections and developments.
2.1
Early Developments in Applications
Presumably the first algorithm which can be connected to sparse recovery
is due to the French mathematician de Prony [71]. The so-called Prony
method, which has found numerous applications [62], estimates non-zero
amplitudes and corresponding frequencies of a sparse trigonometric polynomial from a small number of equispaced samples by solving an eigenvalue
problem. The use of ℓ1 -minimization appears already in the Ph.D. thesis of
B. Logan [59] in connection with sparse frequency estimation, where he observed that L1 -minimization may recover exactly a frequency-sparse signal
from undersampled data provided the sparsity is small enough. The paper
by Donoho and Logan [25] is perhaps the earliest theoretical work on sparse
recovery using L1 -minimization. Nevertheless, geophysicists observed in the
late 1970’s and 1980’s that ℓ1 -minimization can be successfully employed in
reflection seismology where a sparse reflection function indicating changes
between subsurface layers is sought [87, 80]. In NMR spectroscopy the idea
to recover sparse Fourier spectra from undersampled non-equispaced samples
was first introduced in the 90’s [96] and has seen a significant development
since then. In image processing the use of total-variation minimization,
which is closely connected to ℓ1 -minimization and compressive sensing, first
5
appears in the 1990’s in the work of Rudin, Osher and Fatemi [79], and was
widely applied later on. In statistics where the corresponding area is usually
called model selection the use of ℓ1 -minimization and related methods was
greatly popularized with the work of Tibshirani [88] on the so-called LASSO
(Least Absolute Shrinkage and Selection Operator).
2.2
Sparse Approximation
Many lossy compression techniques such as JPEG, JPEG-2000, MPEG or
MP3 rely on the empirical observation that audio signals and digital images
have a sparse representation in terms of a suitable basis. Roughly speaking
one compresses the signal by simply keeping only the largest coefficients. In
certain scenarios such as audio signal processing one considers the generalized situation where sparsity appears in terms of a redundant system — a
so called dictionary or frame [19] — rather than a basis. The problem of
finding the sparsest representation / approximation in terms of the given
dictionary turns out to be significantly harder than in the case of sparsity with respect to a basis where the expansion coefficients are unique.
Indeed, in [61, 64] it was shown that the general ℓ0 -problem of finding
the sparsest solution of an underdetermined system is NP-hard. Greedy
strategies such as Matching Pursuit algorithms [61], FOCUSS [52] and ℓ1 minimization [18] were subsequently introduced as tractable alternatives.
The theoretical understanding under which conditions greedy methods and
ℓ1 -minimization recover the sparsest solutions began to develop with the
work in [30, 37, 29, 53, 49, 46, 91, 92].
2.3
Information Based Complexity and Gelfand Widths
Information based complexity (IBC) considers the general question of how
well a function f belonging to a certain class F can be recovered from n
sample values, or more generally, the evaluation of n linear or non-linear
functionals applied to f [89]. The optimal recovery error which is defined as
the maximal reconstruction error for the “best” sampling method and “best”
recovery method (within a specified class of methods) over all functions in
the class F is closely related to the so-called Gelfand width of F [66, 21, 26].
Of particular interest for compressive sensing is F = B1N , the ℓ1 -ball in
RN since its elements can be well-approximated by sparse ones. A famous
result due to Kashin [56], and Gluskin and Garnaev [47, 51] sharply bounds
the Gelfand widths of B1N (as well as their duals, the Kolmogorov widths)
from above and below, see also [44]. While the original interest of Kashin
6
was in the estimate of n-widths of Sobolev classes, these results give precise
performance bounds in compressive sensing on how well any method may
recover (approximately) sparse vectors from linear measurements [26, 21].
The upper bounds on Gelfand widths were derived in [56] and [47] using
(Bernoulli and Gaussian) random matrices, see also [60], and in fact such
type of matrices have become very useful also in compressive sensing [26, 16].
2.4
Compressive Sensing
The numerous developments in compressive sensing began with the seminal work [15] and [26]. Although key ingredients were already in the air at
that time, as mentioned above, the major contribution of these papers was
to realize that one can combine the power of ℓ1 -minimization and random
matrices in order to show optimal results on the ability of ℓ1 -minimization
of recovering (approximately) sparse vectors. Moreover, the authors made
very clear that such ideas have strong potential for numerous application
areas. In their work [16, 15] Candès, Romberg and Tao introduced the
restricted isometry property (which they initially called the uniform uncertainty principle) which is a key property of compressive sensing matrices.
It was shown that Gaussian, Bernoulli, and partial random Fourier matrices [16, 78, 73] possess this important property. These results require many
tools from probability theory and finite dimensional Banach space geometry,
which have been developed for a rather long time now, see e.g. [58, 55].
Donoho [28] developed a different path and approached the problem of
characterizing sparse recovery by ℓ1 -minimization via polytope geometry,
more precisely, via the notion of k-neighborliness. In several papers sharp
phase transition curves were shown for Gaussian random matrices separating
regions where recovery fails or succeeds with high probability [31, 28, 32].
These results build on previous work in pure mathematics by Affentranger
and Schneider [2] on randomly projected polytopes.
2.5
Developments in Computer Science
In computer science the related area is usually addressed as the heavy hitters
detection or sketching. Here one is interested not only in recovering signals
(such as huge data streams on the internet) from vastly undersampled data,
but one requires sublinear runtime in the signal length N of the recovery
algorithm. This is no impossibility as one only has to report the locations
and values of the non-zero (most significant) coefficients of the sparse vector.
Quite remarkably sublinear algorithms are available for sparse Fourier re-
7
covery [48]. Such algorithms use ideas from group testing which date back to
World War II, when Dorfman [34] invented an efficient method for detecting
draftees with syphilis.
In sketching algorithms from computer science one actually designs the
matrix and the fast algorithm simultaneously [22, 50]. More recently, bipartite expander graphs have been successfully used in order to construct good
compressed sensing matrices together with associated fast reconstruction
algorithms [5].
3
Mathematical Modelling and Analysis
This section introduces the concept of sparsity and the recovery of sparse
vectors from incomplete linear and non-adaptive measurements. In particular, an analysis of ℓ1 -minimization as a recovery method is provided. The
null-space property and the restricted isometry property are introduced and
it is shown that they ensure robust sparse recovery. It is actually difficult
to show these properties for deterministic matrices and the optimal number
m of measurements, and the major breakthrough in compressive sensing results is obtained for random matrices. Examples of several types of random
matrices which ensure sparse recovery are given, such as Gaussian, Bernoulli
and partial random Fourier matrices.
3.1
Preliminaries and Notation
This exposition mostly treats complex vectors in CN although sometimes
the considerations will be restricted to the real case RN . The ℓp -norm of a
vector x ∈ CN is defined as
1/p

N
X
|xj |p  , 0 < p < ∞,
kxkp := 
j=1
kxk∞ := max |xj |.
j=1,...,N
(3.1)
For 1 ≤ p ≤ ∞, it is indeed a norm while for 0 < p < 1 it is only a quasiN or
norm. When emphasizing the norm the term ℓN
p is used instead of C
N
N
RN . The unit ball in ℓN
p is Bp = {x ∈ C , kxkp ≤ 1}. The operator norm
m
of a matrix A ∈ Cm×N from ℓN
p to ℓp is denoted
kAkp→p = max kAxkp .
kxkp =1
8
(3.2)
In the important special case p = 2, the operator norm is the maximal
singular value σmax (A) of A.
For a subset T ⊂ {1, . . . , N } we denote by xT ∈ CN the vector which
coincides with x ∈ CN on the entries in T and is zero outside T . Similarly, AT
denotes the column submatrix of A corresponding to the columns indexed
by T . Further, T c = {1, . . . , N } \ T denotes the complement of T and #T
or |T | indicate the cardinality of T . The kernel of a matrix A is denoted by
ker A = {x, Ax = 0}.
3.2
Sparsity and Compression
Compressive Sensing is based on the empirical observation that many types
of real-world signals and images have a sparse expansion in terms of a suitable basis or frame, for instance a wavelet expansion. This means that the
expansion has only a small number of significant terms, or in other words,
that the coefficient vector can be well-approximated with one having only a
small number of nonvanishing entries.
The support of a vector x is denoted supp(x) = {j : xj 6= 0}, and
kxk0 := | supp(x)|.
It has become common to call k · k0 the ℓ0 -norm, although it is not even a
quasi-norm. A vector x is called k-sparse if kxk0 ≤ k. For k ∈ {1, 2, . . . , N },
Σk := {x ∈ CN : kxk0 ≤ k}
denotes the set of k-sparse vectors. Furthermore, the best k-term approximation error of a vector x ∈ CN in ℓp is defined as
σk (x)p = inf kx − zkp .
z∈Σk
If σk (x) decays quickly in k then x is called compressible. Indeed, in order
to compress x one may simply store only the k largest entries. When reconstructing x from its compressed version the nonstored entries are simply set
to zero, and the reconstruction error is σk (x)p . It is emphasized at this point
that the procedure of obtaining the compressed version of x is adaptive and
nonlinear since it requires the search of the largest entries of x in absolute
value. In particular, the location of the non-zeros is a nonlinear type of
information.
The best k-term approximation of x can be obtained using the nonincreasing rearrangement r(x) = (|xi1 |, . . . , |xiN |)T , where ij denotes a permutation of the indexes such that |xij | ≥ |xij+1 | for j = 1, . . . , N − 1. Then
9
it is straightforward to check that

1/p
N
X
σk (x)p := 
rj (x)p  ,
0 < p < ∞.
j=k+1
and the vector x[k] derived from x by setting to zero all the N − k smallest
entries in absolute value is the best k-term approximation,
x[k] = arg min kx − zkp ,
z∈Σk
for any 0 < p ≤ ∞.
The next lemma states essentially that ℓq -balls with small q (ideally
q ≤ 1) are good models for compressible vectors.
Lemma 3.1. Let 0 < q < p ≤ ∞ and set r =
σk (x)p ≤ k−r ,
k = 1, 2, . . . , N
1
q
− p1 . Then
for all x ∈ BqN .
Proof. Let T be the set of indeces of the k-largest entries of x in absolute
value. The nonincreasing rearrangement satisfies |rk (x)| ≤ |xj | for all j ∈ T ,
and therefore
X
krk (x)q ≤
|xj |q ≤ kxkqq ≤ 1.
j∈T
1
Hence, rk (x) ≤ k− q . Therefore
X
X
p−q
p−q
σk (x)pp =
|xj |p ≤
rk (x)p−q |xj |q ≤ k− q kxkqq ≤ k− q ,
j ∈T
/
j ∈T
/
which implies σk (x)p ≤ k−r .
3.3
Compressive Sensing
The above outlined adaptive strategy of compressing a signal x by only
keeping its largest coefficients is certainly valid when full information on x
is available. If, however, the signal first has to be acquired or measured by
a somewhat costly or lengthy procedure then this seems to be a waste of
resources: At first, large efforts are made to acquire the full signal and then
most of the information is thrown away when compressing it. One may ask
whether it is possible to obtain more directly a compressed version of the
signal by taking only a small amount of linear and nonadaptive measurements. Since one does not know a priori the large coefficients, this seems a
10
daunting task at first sight. Quite surprisingly, compressive sensing nevertheless predicts that reconstruction from vastly undersampled nonadaptive
measurements is possible — even by using efficient recovery algorithms.
Taking m linear measurements of a signal x ∈ CN corresponds to applying a matrix A ∈ Cm×N — the measurement matrix —
y = Ax.
(3.3)
The vector y ∈ Cm is called the measurement vector. The main interest is in
the vastly undersampled case m ≪ N . Without further information, it is, of
course, impossible to recover x from y since the linear system (3.3) is highly
underdetermined, and has therefore infinitely many solutions. However, if
the additional assumption that the vector x is k-sparse is imposed, then the
situation dramatically changes as will be outlined.
The approach for a recovery procedure that probably comes first to mind
is to search for the sparsest vector x which is consistent with the measurement vector y = Ax. This leads to solving the ℓ0 -miminization problem
min kzk0
subject to Az = y.
(3.4)
Unfortunately, this combinatorial minimization problem is NP–hard in general [61, 64]. In other words, an algorithm that solves (3.4) for any matrix A
and any right hand side y is necessarily computationally intractable. Therefore, essentially two practical and tractable alternatives to (3.4) have been
proposed in the literature: convex relaxation leading to ℓ1 -minimization
— also called basis pursuit [18] — and greedy algorithms, such as various
matching pursuits [91, 90]. Quite surprisingly for both types of approaches
various recovery results are available, which provide conditions on the matrix A and on the sparsity kxk0 such that the recovered solution coincides
with the original x, and consequently also with the solution of (3.4). This is
no contradiction to the NP–hardness of (3.4) since these results apply only
to a subclass of matrices A and right-hand sides y.
The ℓ1 -minimization approach considers the solution of
min kzk1
subject to Az = y,
(3.5)
which is a convex optimization problem and can be seen as a convex relaxation of (3.4). Various efficient convex optimization techniques apply for
its solution [9]. In the real-valued case, (3.5) is equivalent to a linear program and in the complex-valued case it is equivalent to a second order cone
program. Therefore standard software applies for its solution — although
11
algorithms which are specialized to (3.5) outperform such standard software,
see Section 4.
The hope is, of course, that the solution of (3.5) coincides with the
solution of (3.4) and with the original sparse vector x. Figure 3 provides an
intuitive explanation why ℓ1 -minimization promotes sparse solutions. Here,
N = 2 and m = 1, so one deals with a line of solutions F(y) = {z : Az = y}
in R2 . Except for pathological situations where ker A is parallel to one of the
faces of the polytope B12 , there is a unique solution of the ℓ1 -minimization
problem, which has minimal sparsity, i.e., only one nonzero entry.
Az=y
l1
-ball
0
Figure 3: The ℓ1 -minimizer within the affine space of solutions of the linear
system Az = y coincides with a sparsest solution.
Recovery results in the next sections make rigorous the intuition that
ℓ1 -minimization indeed promotes sparsity.
For sparse recovery via greedy algorithms we refer the reader to the
literature [91, 90].
3.4
The Null Space Property
The null space property is fundamental in the analysis of ℓ1 -minimization.
Definition 3.1. A matrix A ∈ Cm×N is said to satisfy the null space property (NSP) of order k with constant γ ∈ (0, 1) if
kηT k1 ≤ γkηT c k1 ,
for all sets T ⊂ {1, . . . , N }, #T ≤ k and for all η ∈ ker A.
The following sparse recovery result is based on this notion.
12
Theorem 3.2. Let A ∈ Cm×N be a matrix that satisfies the NSP of order
k with constant γ ∈ (0, 1). Let x ∈ CN and y = Ax and let x∗ be a solution
of the ℓ1 -minimization problem (3.5). Then
kx − x∗ k1 ≤
2(1 + γ)
σk (x)1 .
1−γ
(3.6)
In particular, if x is k-sparse then x∗ = x.
Proof. Let η = x∗ − x. Then η ∈ ker A and
kx∗ k1 ≤ kxk1
because x∗ is a solution of the ℓ1 -minimization problem (3.5). Let T be the
set of the k-largest entries of x in absolute value. One has
kx∗T k1 + kx∗T c k1 ≤ kxT k1 + kxT c k1 .
It follows immediately from the triangle inequality that
kxT k1 − kηT k1 + kηT c k1 − kxT c k1 ≤ kxT k1 + kxT c k1 .
Hence,
kηT c k1 ≤ kηT k1 + 2kxT c k1 ≤ γkηT c k1 + 2σk (x)1 ,
or, equivalently,
kηT c k1 ≤
Finally,
2
σk (x)1 .
1−γ
kx − x∗ k1 = kηT k1 + kηT c k1 ≤ (γ + 1)kηT c k1 ≤
(3.7)
2(1 + γ)
σk (x)1
1−γ
and the proof is completed.
One can also show that if all k-sparse x can be recovered from y = Ax
using ℓ1 -minimization then necessarily A satisfies the NSP of order k with
some constant γ ∈ (0, 1) [53, 21]. Therefore, the NSP is actually equivalent
to sparse ℓ1 -recovery.
13
3.5
The Restricted Isometry Property
The NSP is somewhat difficult to show directly. The restricted isometry
property (RIP) is easier to handle and it also implies stability under noise
as stated below.
Definition 3.2. The restricted isometry constant δk of a matrix A ∈ Cm×N
is the smallest number such that
(1 − δk )kzk22 ≤ kAzk22 ≤ (1 + δk )kzk22 ,
(3.8)
for all z ∈ Σk .
A matrix A is said to satisfy the restricted isometry property of order k
with constant δk if δk ∈ (0, 1). It is easily seen that δk can be equivalently
defined as
δk =
max
kA∗T AT − Id k2→2 ,
T ⊂{1,...,N },#T ≤k
which means that all column submatrices of A with at most k columns are
required to be well-conditioned. The RIP implies the NSP as shown in the
following lemma.
Lemma 3.3. Assume that A ∈ Cm×N satisfies the RIP of order K = k + h
with q
constant δK ∈ (0, 1). Then A has the NSP of order k with constant
γ=
k 1+δK
h 1−δK .
Proof. Let η ∈ N = ker A and T ⊂ {1, . . . , N }, #T ≤ k. Define T0 = T and
T1 , T2 , . . . , Ts to be disjoint sets of indexes of size at most h, associated to a
nonincreasing rearrangement of the entries of η ∈ N , i.e.,
(3.9)
for all j ∈ Tℓ , i ∈ Tℓ′ , ℓ ≥ ℓ′ ≥ 1.
P
Note that Aη = 0 implies AηT0 ∪T1 = − sj=2 AηTj . Then, from the Cauchy–
Schwarz inequality, the RIP, and the triangle inequality, the following sequence of inequalities is deduced,
√
√
kkηT k2 ≤ kkηT0 ∪T1 k2
kηT k1 ≤
r
r
k
k
≤
kAηT0 ∪T1 k2 =
kAηT2 ∪T3 ∪···∪Ts k2
1 − δK
1 − δK
r
r
s
s
k X
1 + δK √ X
≤
kAηTj k2 ≤
kηTj k2 . (3.10)
k
1 − δK
1 − δK
|ηj | ≤ |ηi |
j=2
j=2
14
It follows from (3.9) that |ηi | ≤ |ηℓ | for all i ∈ Tj+1 and ℓ ∈ Tj . Taking the
sum over ℓ ∈ Tj first and then the ℓ2 -norm over i ∈ Tj+1 yields
|ηi | ≤ h−1 kηTj k1 ,
kηTj+1 k2 ≤ h−1/2 kηTj k1 .
and
Using the latter estimates in (3.10) gives
r
r
s−1
1 + δK k X
1 + δK k
kηT k1 ≤
kηTj k1 ≤
kηT c k1 ,
1 − δK h
1 − δK h
(3.11)
j=1
and the proof is finished.
Taking h = 2k above shows that δ3k < 1/3 implies γ < 1. By Theorem
3.2, recovery of all k-sparse vectors by ℓ1 -minimization is then guaranteed.
Additionally, stability in ℓ1 is also ensured. The next theorem shows that
RIP implies also a bound on the reconstruction error in ℓ2 .
Theorem 3.4. Assume A ∈ Cm×N satisfies the RIP of order 3k with δ3k <
1/3. For x ∈ CN , let y = Ax and x∗ be the solution of the ℓ1 -minimization
problem (3.5). Then
σk (x)1
kx − x∗ k2 ≤ C √
k
q
γ+1
1+δ3k
2
√ +γ , γ =
with C = 1−γ
2(1−δ3k ) .
2
Proof. Similarly as in the proof of Lemma 3.3, denote η = x∗ − x ∈ N =
ker A, T0 = T the set of the 2k-largest entries of η in absolute value, and
Tj ’s of size at most k corresponding to the nonincreasing rearrangement of
η. Then, using (3.10) and (3.11) with h = 2k of the previous proof,
s
1 + δ3k −1/2
k
kηT c k1 .
kηT k2 ≤
2(1 − δ3k )
q
1+δ3k
< 1. Lemma
From the assumption δ3k < 1/3 it follows that γ := 2(1−δ
3k )
3.1 and Lemma 3.3 yield
1
kηT c k2 = σ2k (η)2 ≤ (2k)− 2 kηk1 = (2k)−1/2 (kηT k1 + kηT c k1 )
γ+1
≤ (2k)−1/2 (γkηT c k1 + kηT c k1 ) ≤ √ k−1/2 kηT c k1 .
2
Since T is the set of 2k-largest entries of η in absolute value, it holds
kηT c k1 ≤ kη(supp x[2k] )c k1 ≤ kη(supp x[k] )c k1 ,
15
(3.12)
where x[k] is the best k-term approximation to x. The use of this latter
estimate, combined with inequality (3.7), finally gives
kx − x∗ k2 ≤ kηT k2 + kηT c k2
γ+1
√ + γ k−1/2 kηT c k1
≤
2
γ+1
2
√ + γ k−1/2 σk (x)1 .
≤
1−γ
2
This concludes the proof.
The restricted isometry property implies also robustness under noise on
the measurements. This fact was first noted in [16, 15]. We present the
so far best known result [43, 45] concerning recovery using a noise aware
variant of ℓ1 -minimization without proof.
Theorem 3.5. Assume that the restricted isometry constant δ2k of the matrix A ∈ Cm×N satisfies
δ2k <
3+
2
p
7/4
≈ 0.4627.
(3.13)
Then the following holds for all x ∈ CN . Let noisy measurements y = Ax+e
be given with kek2 ≤ η. Let x∗ be the solution of
min kzk1
subject to kAz − yk2 ≤ η.
(3.14)
Then
σk (x)1
kx − x∗ k2 ≤ C1 η + C2 √
k
for some constants C1 , C2 > 0 that depend only on δ2k .
3.6
Coherence
The coherence is a by now classical way of analyzing the recovery abilities of
a measurement matrix [29, 91]. For a matrix A = (a1 |a2 | · · · |aN ) ∈ Cm×N
with normalized columns, kaℓ k2 = 1, it is defined as
µ := max |haℓ , ak i|.
ℓ6=k
Applying Gershgorin’s disc theorem [54] to A∗T AT − I with #T = k shows
that
δk ≤ (k − 1)µ.
(3.15)
16
Several explicit examples of matrices are known which have small coherence
√
µ = O(1/ m). A simple one is the concatenation A = (I|F ) ∈ Cm×2m
of the identity matrix and the unitary Fourier matrix F ∈ Cm×m with
√
entries Fj,k = m−1/2 e2πijk/m . It is easily seen that µ = 1/ m in this
2
case. Furthermore, [82] gives several matrices A ∈ Cm×m with coherence
√
µ = 1/ m. In all these cases, δk ≤ C √km . Combining this estimate with the
recovery results for ℓ1 -minimization above shows that all k-sparse vectors x
can be (stably) recovered from y = Ax via ℓ1 -minimization provided
m ≥ C ′ k2 .
(3.16)
At first sight one might be satisfied with this condition since if k is very
small compared to N then still m might be chosen smaller than N and
all k-sparse vectors can be recovered from the undersampled measurements
y = Ax. Although this is great news for a start, one might nevertheless hope
that (3.16) can be improved. In particular, one may expect that actually a
linear scaling of m in k should be enough to guarantee sparse recovery by
ℓ1 -minimization. The existence of matrices, which indeed provide recovery
conditions of the form m ≥ Ck logα (N ) (or similar) with some α ≥ 1, is
shown in the next section. Unfortunately, such results cannot be shown
using simply the coherence because of the general lower bound [82]
s
N −m
1
(N sufficiently large).
∼√
µ≥
m
m(N − 1)
In particular, it is not possible to overcome the “quadratic bottleneck” in
(3.16) by using Gershgorin’s theorem or Riesz-Thorin interpolation between
k · k1→1 and k · k∞→∞ , see also [75, 81]. In order to improve on (3.16) one
has to take into account also cancellations in the Gramian A∗T AT − I, and
this task seems to be quite difficult using deterministic methods. Therefore,
it will not come as a surprise that the major breakthrough in compressive
sensing was obtained with random matrices. It is indeed easier to deal with
cancellations in the Gramian using probabilistic techniques.
3.7
RIP for Gaussian and Bernoulli Random Matrices
Optimal estimates for the RIP constants in terms of the number m of measurement matrices can be obtained for Gaussian, Bernoulli or more general
subgaussian random matrices.
Let X be a random variable. Then one defines a random matrix A =
A(ω), ω ∈ Ω, as the matrix whose entries are independent realizations of
17
X, where (Ω, Σ, P) is their common probability space. One assumes further
that for any x ∈ RN we have the identity EkAxk22 = kxk22 , E denoting
expectation.
The starting point for the simple approach in [4] is a concentration inequality of the form
2
P kAxk22 − kxk22 ≥ δkxk22 ≤ 2e−c0 δ m , 0 < δ < 1,
(3.17)
where c0 > 0 is some constant.
The two most relevant examples of random matrices which satisfy the
above concentration are the following.
1. Gaussian Matrices. Here the entries of A are chosen as i.i.d. Gaussian random variables with expectation 0 and variance 1/m. As shown
in [1] Gaussian matrices satisfy (3.17).
2. Bernoulli Matrices The entries of a Bernoulli matrices are inde√
pendent realizations of ±1/ m Bernoulli random variables, that is,
√
√
each entry takes the value +1/ m or −1/ m with equal probability.
Bernoulli matrices also satisfy the concentration inequality (3.17) [1].
Based on the concentration inequality (3.17) the following estimate on
RIP constants can be shown [4, 16, 63].
Theorem 3.6. Assume A ∈ Rm×N to be a random matrix satisfying the
concentration property (3.17). Then there exists a constant C depending
only on c0 such that the restricted isometry constant of A satisfies δk ≤ δ
with probability exceeding 1 − ε provided
m ≥ Cδ−2 (k log(N/m) + log(ε−1 )).
Combining this RIP estimate with the recovery results for ℓ1 -minimization
shows that all k-sparse vectors x ∈ CN can be stably recovered from a random draw of A satisfying (3.17) with high probability provided
m ≥ Ck log(N/m).
(3.18)
Up to the log-factor this provides the desired linear scaling of the number m
of measurements with respect to the sparsity k. Furthermore, as shown in
Section 3.9 below, condition (3.18) cannot be further improved; in particular,
the log-factor cannot be removed.
It is useful to observe that the concentration inequality is invariant under
unitary transforms. Indeed, suppose that z is not sparse with respect to the
18
canonical basis but with respect to a different orthonormal basis. Then
z = U x for a sparse x and a unitary matrix U ∈ CN ×N . Applying the
measurement matrix A yields
Az = AU x,
so that this situation is equivalent to working with the new measurement
matrix A′ = AU and again sparsity with respect to the canonical basis. The
crucial point is that A′ satisfies again the concentration inequality (3.17)
once A does. Indeed, choosing x = U −1 x′ and using unitarity gives
P kAU xk22 − kxk22 ≥ δkxk2ℓN = P kAx′ k22 − kU −1 x′ k22 ≥ δkU −1 x′ k2ℓN
2
2
−2 m
′
2
′
2
′
2
−c
δ
= P kAx k2 − kx k2 ≥ δkx kℓN ≤ 2e 0
.
2
Hence, Theorem 3.6 also applies to A′ = AU . This fact is sometimes referred
to as the universality of the Gaussian or Bernoulli random matrices. It does
not matter in which basis the signal x is actually sparse. At the coding
stage, where one takes random measurements y = Az, knowledge of this
basis is not even required. Only the decoding procedure needs to know U .
3.8
Random Partial Fourier Matrices
While Gaussian and Bernoulli matrices provide optimal conditions for the
minimal number of required samples for sparse recovery, they are of somewhat limited use for practical applications for several reasons. Often the
application imposes physical or other constraints on the measurement matrix, so that assuming A to be Gaussian may not be justifiable in practice.
One usually has only limited freedom to inject randomness in the measurements. Furthermore, Gaussian or Bernoulli matrices are not structured so
there is no fast matrix-vector multiplication available which may speed up
recovery algorithms, such as the ones described in Section 4. Thus, Gaussian
random matrices are not applicable in large scale problems.
A very important class of structured random matrices that overcomes
these drawbacks are random partial Fourier matrices, which were also the
object of study in the very first papers on compressive sensing [13, 16, 72, 73].
A random partial Fourier matrix A ∈ Cm×N is derived from the discrete
Fourier matrix F ∈ CN ×N with entries
1
Fj,k = √ e2πjk/N ,
N
19
by selecting m rows uniformly at random among all N rows. Taking measurements of a sparse x ∈ CN corresponds then to observing m of the entries
of its discrete Fourier transform x̂ = F x. It is important to note that the
fast Fourier transform may be used to compute matrix-vector multiplications with A and A∗ with complexity O(N log(N )). The following theorem
concerning the RIP constant was proven in [75], and improves slightly on
the results in [78, 16, 73].
Theorem 3.7. Let A ∈ Cm×N be the random partial Fourier matrix as
just
q described. Then the restricted isometry constant of the rescaled matrix
N
mA
satisfy δk ≤ δ with probability at least 1 − N −γ log
m ≥ Cδ−2 k log4 (N ).
3
(N )
provided
(3.19)
The constants C, γ > 1 are universal.
Combining this estimate with the ℓ1 -minimization results above shows
that recovery with high probability can be ensured for all k-sparse x provided
m ≥ Ck log4 (N ).
The plots in Figure 1 illustrate an example of successful recovery from partial
Fourier measurements.
The proof of the above theorem is not straightforward and involves Dudley’s inequality as a main tool [78, 75]. Compared to the recovery condition
(3.18) for Gaussian matrices, we suffer a higher exponent at the log-factor,
but the linear scaling of m in k is preserved. Also a nonuniform recovery
result for ℓ1 -minimization is available [13, 72, 75], which states that each
k-sparse x can be recovered using a random draw of the random partial
Fourier matrix A with probability at least 1 − ε provided m ≥ Ck log(N/ε).
The difference to the statement in Theorem 3.7 is that, for each sparse x,
recovery is ensured with high probability for a new random draw of A. It
does not imply the existence of a matrix which allows recovery of all k-sparse
x simultaneously. The proof of such recovery results do not make use of the
restricted isometry property or the null space property.
One may generalize the above results to a much broader class of structured random matrices which arise from random sampling in bounded orthonormal systems. The interested reader is referred to [72, 73, 75].
Another class of structured random matrices, for which recovery results
are known, consist of partial random circulant and Toeplitz matrices. These
correspond to subsampling the convolution of x with a random vector b at
20
m fixed (deterministic) entries. The reader is referred to [74, 75] for detailed
information. It is only noted that a good estimate for the RIP constants for
such types of random matrices is still an open problem. Further types of
random measurement matrices are discussed in [69, 93].
3.9
Compressive Sensing and Gelfand Widths
In this section a quite general viewpoint is taken. The question is investigated how well any measurement matrix and any reconstruction method
— in this context usually called the decoder — may perform. This leads to
the study of Gelfand widths, already mentioned in Section 2.3. The corresponding analysis will allow to draw the conclusion that Gaussian random
matrices in connection with ℓ1 -minimization provide optimal performance
guarantees.
Following the tradition of the literature in this context, only the realvalued case will be treated. The complex-valued case is easily deduced
from the real-case by identifying CN with R2N and by corresponding norm
equivalences of ℓp -norms.
The measurement matrix A ∈ Rm×N is here also referred to as the
encoder. The set Am,N denotes all possible encoder / decoder pairs (A, ∆)
where A ∈ Rm×N and ∆ : Rm → RN is any (nonlinear) function. Then, for
1 ≤ k ≤ N , the reconstruction errors over subsets K ⊂ RN , where RN is
endowed with a norm k · kX , are defined as
σk (K)X := sup σk (x)X ,
x∈K
Em (K, X) :=
sup kx − ∆(Ax)kX .
inf
(A,∆)∈Am,N x∈K
In words, En (K, X) is the worst reconstruction error for the best pair of
encoder / decoder. The goal is to find the largest k such that
Em (K, X) ≤ C0 σk (K)X .
Of particular interest for compressive sensing are the unit balls K = BpN for
N
0 < p ≤ 1 and X = ℓN
2 because the elements of Bp are well-approximated
by sparse vectors due to Lemma 3.1. The proper estimate of Em (K, X)
turns out to be linked to the geometrical concept of Gelfand width.
Definition 3.3. Let K be a compact set in a normed space X. Then the
Gelfand width of K of order m is
dm (K, X) :=
inf
Y ⊂X
codim(Y ) ≤ m
21
sup{kxkX : x ∈ K ∩ Y },
where the infimum is over all linear subspaces Y of X of codimension less
or equal to m.
The following fundamental relationship between Em (K, X) and the Gelfand
widths holds.
Proposition 3.8. Let K ⊂ RN be a closed compact set such that K = −K
and K + K ⊂ C0 K for some constant C0 . Let X = (RN , k · kX ) be a normed
space. Then
dm (K, X) ≤ Em (K, X) ≤ C0 dm (K, X).
Proof. For a matrix A ∈ Rm×N , the subspace Y = ker A has codimension
less or equal to m. Conversely, to any subspace Y ⊂ RN of codimension less
or equal to m, a matrix A ∈ Rm×N can be associated, the rows of which
form a basis for Y ⊥ . This identification yields
dm (K, X) =
inf
m×N
A∈R
sup{kηkX : η ∈ ker A ∩ K}.
Let (A, ∆) be an encoder / decoder pair in Am,N and z = ∆(0). Denote
Y = ker(A). Then with η ∈ Y also −η ∈ Y , and either kη − zkX ≥ kηkX or
k − η − zkX ≥ kηkX . Indeed, if both inequalities were false then
k2ηkX = kη − z + z + ηkX ≤ kη − zkX + k − η − zkX < 2kηkX ,
a contradiction. Since K = −K it follows that
dm (K, X) =
inf
m×N
A∈R
sup{kηkX : η ∈ Y ∩ K} ≤ sup kη − zkX
η∈Y ∩K
= sup kη − ∆(Aη)kX ≤ sup kx − ∆(Ax)kX .
η∈Y ∩K
x∈K
Taking the infimum over all (A, ∆) ∈ Am,N yields
dm (K, X) ≤ Em (K, X).
To prove the converse inequality, choose an optimal Y such that
dm (K, X) = sup{kxkX : x ∈ Y ∩ K}.
(An optimal subspace Y always exists [60].) Let A be a matrix whose rows
form a basis for Y ⊥ . Denote the affine solution space F(y) := {x : Ax = y}.
One defines then a decoder as follows. If F(y) ∩ K 6= ∅ then choose some
22
x̄(y) ∈ F(y) and set ∆(y) = x̄(y). If F(y) ∩ K = ∅ then ∆(y) ∈ F(y). The
following chain of inequalities is then deduced
Em (K, X) ≤ sup
y
≤
sup
x,x′ ∈F (y)∩K
sup
η∈C0 (Y ∩K)
kx − x′ kX
kηkX ≤ C0 dm (K, X),
which concludes the proof.
The assumption K + K ⊂ C0 K clearly holds for norm balls with C0 = 2
and for quasi-norm balls with some C0 ≥ 2. The next theorem provides a
two-sided estimate of the Gelfand widths dm (BpN , ℓN
2 ) [44, 27, 95]. Note that
the case p = 1 was considered much earlier in [56, 47, 44].
Theorem 3.9. Let 0 < p ≤ 1. There exist universal constants Cp , Dp > 0
such that the Gelfand widths dm (BpN , ℓN
2 ) satisfy
ln(2N/m) 1/p−1/2
≤ dm (BpN , ℓN
Cp min 1,
2 )
m
ln(2N/m) 1/p−1/2
≤ Dp min 1,
m
(3.20)
Combining Proposition 3.8 and Theorem 3.9 gives in particular, for large
m,
r
r
log(2N/m)
log(2N/m)
N N
C̃1
≤ Em (B1 , ℓ2 ) ≤ D̃1
.
(3.21)
m
m
This estimate implies a lower estimate for the minimal number of required
samples which allows for approximate sparse recovery using any measurement matrix and any recovery method whatsoever. The reader should compare the next statement with Theorem 3.4.
Corollary 3.10. Suppose that A ∈ Rm×N and ∆ : Rm → RN such that
σk (x)1
kx − ∆(Ax)k2 ≤ C √
k
for all x ∈ B1N and some constant C > 0. Then necessarily
m ≥ C ′ k log(2N/m).
23
(3.22)
Proof. Since σk (x)1 ≤ kxk1 ≤ 1, the assumption implies Em (B1N , ℓN
2 ) ≤
Ck−1/2 . The lower bound in (3.21) combined with Proposition 3.8 yields
r
log(2N/m)
−1/2
C̃1
.
≤ Em (B1N , ℓN
2 ) ≤ Ck
m
Consequently, m ≥ C ′ k log(eN/m) as claimed.
In particular, the above lemma applies to ℓ1 -minimization and consequently δk ≤ 0.4 (say) for a matrix A ∈ Rm×N implies m ≥ Ck log(N/m).
Therefore, the recovery results for Gaussian or Bernoulli random matrices
with ℓ1 -minimization stated above are optimal.
It can also be shown that a stability estimate in the ℓ1 -norm of the form
kx − ∆(Ax)k1 ≤ Cσk (x)1 for all x ∈ RN implies (3.22) as well [44, 24].
3.10
Applications
Compressive sensing can be potentially used in all applications where the
task is the reconstruction of a signal or an image from linear measurements,
while taking many of those measurements – in particular, a complete set
of measurements – is a costly, lengthy, difficult, dangerous, impossible or
otherwise undesired procedure. Additionally, there should be reasons to
believe that the signal is sparse in a suitable basis (or frame). Empirically,
the latter applies to most types of signals.
In computerized tomography, for instance, one would like to obtain an
image of the inside of a human body by taking X-ray images from different
angles. Taking an almost complete set of images would expose the patient
to a large and dangerous dose of radiation, so the amount of measurements
should be as small as possible, and nevertheless guarantee a good enough
image quality. Such images are usually nearly piecewise constant and therefore nearly sparse in the gradient, so there is a good reason to believe that
compressive sensing is well applicable. And indeed, it is precisely this application that started the investigations on compressive sensing in the seminal
paper [13].
Also radar imaging seems to be a very promising application of compressive sensing techniques [38, 83]. One is usually monitoring only a small
number of targets, so that sparsity is a very realistic assumption. Standard
methods for radar imaging actually also use the sparsity assumption, but
only at the very end of the signal processing procedure in order to clean up
the noise in the resulting image. Using sparsity systematically from the very
24
beginning by exploiting compressive sensing methods is therefore a natural
approach. First numerical experiments in [38, 83] are very promising.
Further potential applications include wireless communication [86], astronomical signal and image processing [8], analog to digital conversion [93],
camera design [35] and imaging [77].
4
Numerical Methods
The previous sections showed that ℓ1 -minimization performs very well in
recovering sparse or approximately sparse vectors from undersampled measurements. In applications it is important to have fast methods for actually solving ℓ1 -minimization problems. Two such methods – the homotopy (LARS) method introduced in [68, 36] and iteratively reweighted least
squares (IRLS) [23] – will be explained in more detail below.
As a first remark, the ℓ1 -minimization problem
min kxk1
subject to Ax = y
(4.1)
is in the real case equivalent to the linear program
min
2N
X
vj
subject to
j=1
v ≥ 0, (A| − A)v = y.
(4.2)
The solution x∗ to (4.1) is obtained from the solution v ∗ of (4.2) via x∗ =
(Id | − Id)v ∗ . Any linear programming method may therefore be used for
solving (4.1). The simplex method as well as interior point methods apply
in particular [65], and standard software may be used. (In the complex
case, (4.1) is equivalent to a second order cone program (SOCP) and can
also be solved with interior point methods.) However, such methods and
software are of general purpose and one may expect that methods specialized
to (4.1) outperform such existing standard methods. Moreover, standard
software often has the drawback that one has to provide the full matrix
rather than fast routines for matrix-vector multiplication which are available
for instance in the case of partial Fourier matrices. In order to obtain the full
performance of such methods one would therefore need to reimplement them,
which is a daunting task because interior point methods usually require much
fine tuning. On the contrary the two specialized methods described below
are rather simple to implement and very efficient. Many more methods are
available nowadays, including greedy methods, such as orthogonal matching
pursuit [91], CoSaMP [90], and iterative hard thresholding [7, 39], which
25
may offer better complexity than standard interior point methods. Due to
space limitations, however, only the two methods below are explained in
detail.
4.1
The Homotopy Method
The homotopy method – or modified LARS – [68, 67, 36, 33] solves (4.1)
in the real-valued case. One considers the ℓ1 -regularized least squares functionals
1
(4.3)
Fλ (x) = kAx − yk22 + λkxk1 , x ∈ RN , λ > 0,
2
and its minimizer xλ . When λ = λ̂ is large enough then xλ̂ = 0, and
furthermore, limλ→0 xλ = x∗ , where x∗ is the solution to (4.1). The idea of
the homotopy method is to trace the solution xλ from xλ̂ = 0 to x∗ . The
crucial observation is that the solution path λ 7→ xλ is piecewise linear, and
it is enough to trace the endpoints of the linear pieces.
The minimizer of (4.3) can be characterized using the subdifferential,
which is defined for a general convex function F : RN → R at a point
x ∈ RN by
∂F (x) = {v ∈ RN , F (y) − F (x) ≥ hv, y − xi for all y ∈ RN }.
Clearly, x is a minimizer of F if and only if 0 ∈ ∂F (x). The subdifferential
of Fλ is given by
∂Fλ (x) = A∗ (Ax − y) + λ∂kxk1
where the subdifferential of the ℓ1 -norm is given by
∂kxk1 = {v ∈ RN : vℓ ∈ ∂|xℓ |, ℓ = 1, . . . , N }
with the subdifferential of the absolute value being
{sgn(z)}, if z 6= 0,
∂|z| =
[−1, 1]
if z = 0.
The inclusion 0 ∈ ∂Fλ (x) is equivalent to
(A∗ (Ax − y))ℓ = λ sgn(xℓ )
∗
|(A (Ax − y)ℓ | ≤ λ
if xℓ 6= 0,
if xℓ = 0,
(4.4)
(4.5)
for all ℓ = 1, . . . , N .
As already mentioned above the homotopy method starts with x(0) =
xλ = 0. By conditions (4.4) and (4.5) the corresponding λ can be chosen
26
as λ = λ(0) = kA∗ yk∞ . In the further steps j = 1, 2, . . ., the algorithm
computes minimizers x(1) , x(2) , . . ., and maintains an active (support) set
Tj . Denote by
c(j) = A∗ (Ax(j−1) − y)
the current residual vector.
Step 1: Let
(1)
ℓ(1) := arg max |(A∗ y)ℓ | = arg max |cℓ |.
ℓ=1,...,N
ℓ=1,...,N
One assumes here and also in the further steps that the maximum is attained
at only one index ℓ. The case that the maximum is attained simultaneously
at two or more indexes ℓ (which almost never happens) requires more complications that will not be covered here. The reader is referred to [36] for
such details.
Now set T1 = {ℓ(1) }. The vector d ∈ RN describing the direction of the
solution (homotopy) path has components
(1)
(1)
dℓ(1) = kaℓ(1) k−2
2 sgn((Ay)ℓ(1) ) and dℓ = 0,
ℓ 6= ℓ(1) .
The first linear piece of the solution path then takes the form
x = x(γ) = x(0) + γd(1) = γd(1) ,
γ ∈ [0, γ (1) ].
One verifies with the definition of d(1) that (4.4) is always satisfied for x =
x(γ) and λ = λ(γ) = λ(0) − γ, γ ∈ [0, λ(0) ]. The next breakpoint is found
by determining the maximal γ = γ (1) > 0 for which (4.5) is still satisfied,
which is
)
(
(0) + c(1)
(0) − c(1)
λ
λ
ℓ
ℓ
,
.
(4.6)
γ (1) = min
1 − (A∗ A d(1) )ℓ 1 + (A∗ A d(1) )ℓ
ℓ6=ℓ(1)
Here, the minimum is taken only over positive arguments. Then x(1) =
x(γ (1) ) = γ (1) d(1) is the next minimizer of Fλ for λ = λ(1) := λ(0) − γ (1) .
This λ(1) satisfies λ(1) = kc(1) k∞ . Let ℓ(2) be the index where the minimum
in (4.6) is attained (where we again assume that the minimum is attained
only at one index) and put T2 = {ℓ(1) , ℓ(2) }.
Step j: Determine the new direction d(j) of the homotopy path by
solving
(j)
(j)
A∗Tj ATj dTj = sgn(cTj ),
(4.7)
27
which is a linear system of equations of size |Tj | × |Tj |, |Tj | ≤ j. Outside
(j)
the components in Tj one sets dℓ = 0, ℓ ∈
/ Tj . The next piece of the path
is then given by
x(γ) = x(j−1) + γd(j) ,
γ ∈ [0, γ (j) ].
The maximal γ such that x(γ) satisfies (4.5) is
(
)
(j−1) + c(j)
(j−1) − c(j)
λ
λ
(j)
ℓ
ℓ
,
γ+ = min
.
ℓ∈T
/ j
1 − (A∗ Ad(j) )ℓ 1 + (A∗ Ad(j) )ℓ
(4.8)
The maximal γ such that x(γ) satisfies (4.4) is determined as
(j−1)
(j)
γ− = min{−xℓ
ℓ∈Tj
(j)
/dℓ }.
(4.9)
Both in (4.8) and (4.9) the minimum is taken only over positive arguments.
(j) (j)
The next breakpoint is given by x(j+1) = x(γ (j) ) with γ (j) = min{γ+ , γ− }.
(j)
(j)
/ Tj providing the
If γ+ determines the minimum then the index ℓ+ ∈
(j)
(j)
minimum in (4.8) is added to the active set, Tj+1 = Tj ∪ {ℓ+ }. If γ (j) = γ−
(j)
(j)
then the index ℓ− ∈ Tj is removed from the active set, Tj+1 = Tj \ {ℓ− }.
Further, one updates λ(j) = λ(j−1) − γ (j) . By construction λ(j) = kc(j) k∞ .
The algorithm stops when λ(j) = kc(j) k∞ = 0, i.e., when the residual
vanishes, and outputs x∗ = x(j) . Indeed, this happens after a finite number
of steps. In [36] the following result was shown.
Theorem 4.1. If in each step the minimum in (4.8) and (4.9) is attained
in only one index ℓ, then the homotopy algorithm as described yields the
minimizer of the ℓ1 -minimization problem (4.1).
If the algorithm is stopped earlier at some iteration j then obviously it
yields the minimizer of Fλ = Fλ(j) . In particular, obvious stopping rules
may also be used to solve the problems
or
min kxk1
subject to kAx − yk2 ≤ ǫ
min kAx − yk2
subject to kxk1 ≤ δ.
(4.10)
(4.11)
The first of these appears in (3.14), and the second is called the lasso (least
absolute shrinkage and selection operator) [88].
The LARS (least angle regression) algorithm is a simple modification of
the homotopy method, which only adds elements to the active set in each
28
(j)
step. So γ− in (4.9) is not considered. (Sometimes the homotopy method
is therefore also called modified LARS.) Clearly, LARS is not guaranteed
any more to yield the solution of (4.1). However, it is observed empirically
— and can be proven rigorously in certain cases [33] — that often in sparse
recovery problems, the homotopy method does never remove elements from
the active set, so that in this case LARS and homotopy perform the same
steps. It is a crucial point that if the solution of (4.1) is k-sparse and the
homotopy method never removes elements then the solution is obtained after
precisely k-steps. Furthermore, the most demanding computational part at
step j is then the solution of the j × j linear system of equations (4.7). In
conclusion, the homotopy and LARS methods are very efficient for sparse
recovery problems.
4.2
Iteratively Reweighted Least Squares
This section is concerned with an iterative algorithm which, under the condition that A satisfies the NSP (see Definition 3.1), is guaranteed to reconstruct vectors with the same error estimate (3.6) as ℓ1 -minimization. Again
we restrict the following discussion to the real case. This algorithm has a
guaranteed linear rate of convergence which can even be improved to a superlinear rate with a small modification. First a brief introduction aims at
shedding light on the basic principles of this algorithm and their interplay
with sparse recovery and ℓ1 -minimization.
Denote F(y) = {x : Ax = y} and N = ker A. The starting point is the
t2
for t 6= 0. Hence, an ℓ1 -minimization can be
trivial observation that |t| = |t|
recasted into a weighted ℓ2 -minimization, with the hope that
arg min
x∈F (y)
x∗
N
X
j=1
|xj | ≈ arg min
x∈F (y)
N
X
j=1
x2j |x∗j |−1 ,
as soon as
is the desired ℓ1 -norm minimizer. The advantage of the reformulation consists in the fact that minimizing the smooth quadratic function
t2 is an easier task than the minimization of the nonsmooth function |t|.
However, the obvious drawbacks are that neither one disposes of x∗ a priori
(this is the vector one is interested to compute!) nor one can expect that
x∗j 6= 0 for all j = 1, . . . , N , since one hopes for k-sparse solutions.
Suppose one has a good approximation wjn of |(x∗j )2 + ǫ2n |−1/2 ≈ |x∗j |−1 ,
for some ǫn > 0. One computes
xn+1 = arg min
x∈F (y)
29
N
X
j=1
x2j wjn ,
(4.12)
and then updates ǫn+1 ≤ ǫn by some rule to be specified later. Further, one
sets
wjn+1 = |(xn+1
)2 + ǫ2n+1 |−1/2 ,
(4.13)
j
and iterates the process. The hope is that a proper choice of ǫn → 0 allows
the iterative computation of an ℓ1 -minimizer. The next sections investigate
convergence of this algorithm and properties of the limit.
4.2.1
Weighted ℓ2 -minimization
Suppose that the weight w is strictly positive which means that wj > 0 for
all j ∈ {1, . . . , N }. Then ℓ2 (w) is a Hilbert space with the inner product
hu, viw :=
N
X
wj uj vj .
(4.14)
j=1
Define
xw := arg min kzk2,w ,
z∈F (y)
(4.15)
1/2
where kzk2,w = hz, ziw . Because the k · k2,w -norm is strictly convex, the
minimizer xw is necessarily unique; it is characterized by the orthogonality
conditions
hxw , ηiw = 0,
for all η ∈ N .
(4.16)
4.2.2
An iteratively re-weighted least squares algorithm (IRLS)
An IRLS algorithm appears for the first time in the Ph.D. thesis of Lawson
in 1961 [57], in the form of an algorithm for solving uniform approximation
problems. This iterative algorithm is now well-known in classical approximation theory as Lawson’s algorithm. In [20] it is proved that it obeys a
linear convergence rate. In the 1970s, extensions of Lawson’s algorithm for
ℓp -minimization, and in particular ℓ1 -minimization, were introduced. In signal analysis, IRLS was proposed as a technique to build algorithms for sparse
signal reconstruction in [52]. The interplay of the NSP, ℓ1 -minimization, and
a reweighted least square algorithm has been clarified only recently in the
work [23].
The analysis of the algorithm (4.12) and (4.13) starts from the observation that
1
wt2 + w−1 ,
|t| = min
w>0 2
30
1
the minimum being attained for w = |t|
. Inspired by this simple relationship,
given a real number ǫ > 0 and a weight vector w ∈ RN , with wj > 0,
j = 1, . . . , N , one introduces the functional
N
J (z, w, ǫ) :=
1 X 2
zj wj + ǫ2 wj + wj−1 ,
2
j=1
z ∈ RN .
(4.17)
The algorithm roughly described in (4.12) and (4.13) can be recast as
an alternating method for choosing minimizers and weights based on the
functional J . To describe this more rigorously, recall that r(z) denotes the
nonincreasing rearrangement of a vector z ∈ RN .
Algorithm IRLS. Initialize by taking w0 := (1, . . . , 1). Set ǫ0 := 1. Then
recursively define, for n = 0, 1, . . . ,
xn+1 := arg min J (z, wn , ǫn ) = arg min kzk2,wn
z∈F (y)
and
ǫn+1
z∈F (y)
rK+1 (xn+1 )
,
:= min ǫn ,
N
(4.18)
(4.19)
where K is a fixed integer that will be specified later. Set
wn+1 := arg min J (xn+1 , w, ǫn+1 ).
w>0
(4.20)
The algorithm stops if ǫn = 0; in this case, define xj := xn for j > n. In
general, the algorithm generates an infinite sequence (xn )n∈N of vectors.
Each step of the algorithm requires the solution of a weighted least
squares problem. In matrix form
xn+1 = Dn−1 A∗ (ADn−1 A∗ )−1 y,
(4.21)
where Dn is the N × N diagonal matrix the j-th diagonal entry of which is
wjn . Once xn+1 is found, the weight wn+1 is given by
wjn+1 = [(xn+1
)2 + ǫ2n+1 ]−1/2 ,
j
4.2.3
j = 1, . . . , N.
Convergence properties
Lemma 4.2. Set L := J (x1 , w0 , ǫ0 ). Then
kxn − xn+1 k22 ≤ 2L J (xn , wn , ǫn ) − J (xn+1 , wn+1 , ǫn+1 ) .
31
(4.22)
Hence (J(xn , wn , ǫn ))n∈N is a monotonically decreasing sequence and
lim kxn − xn+1 k22 = 0.
n→∞
Proof. Note that J (xn , wn , ǫn ) ≥ J (xn+1 , wn+1 , ǫn+1 ) for each n = 1, 2, . . . ,
and
L = J (x1 , w0 , ǫ0 ) ≥ J (xn , wn , ǫn ) ≥ (wjn )−1 ,
j = 1, . . . , N.
Hence, for each n = 1, 2, . . . , the following estimates hold,
2[J (xn , wn , ǫn ) − J (xn+1 , wn+1 , ǫn+1 )]
≥ 2[J (xn , wn , ǫn ) − J (xn+1 , wn , ǫn )] = hxn , xn iwn − hxn+1 , xn+1 iwn
= hxn + xn+1 , xn − xn+1 iwn = hxn − xn+1 , xn − xn+1 iwn
=
N
X
j=1
wjn (xnj − xn+1
)2 ≥ L−1 kxn − xn+1 k22 ,
j
In the third line it is used that hxn+1 , xn − xn+1 iwn = 0 due to (4.16) since
xn − xn+1 is contained in N .
Moreover, if one assumes that xn → x̄ and ǫn → 0, then, formally,
J (xn , wn , ǫn ) → kx̄k1 .
Hence, one expects that this algorithm performs similar to ℓ1 -minimization.
Indeed, the following convergence result holds.
Theorem 4.3. Suppose A ∈ Rm×N satisfies the NSP of order K with constant γ < 1. Use K in the update rule (4.19). Then, for each y ∈ Rm ,
the sequence xn produced by the algorithm converges to a vector x̄, with
rK+1 (x̄) = N limn→∞ ǫn and the following holds:
(i) If ǫ = limn→∞ ǫn = 0, then x̄ is K-sparse; in this case there is therefore
a unique ℓ1 -minimizer x∗ , and x̄ = x∗ ; moreover, we have, for k ≤ K, and
any z ∈ F(y),
2(1 + γ)
σk (z)1 ;
(4.23)
kz − x̄k1 ≤
1−γ
1/2
P
2 + ǫ2
z
;
(ii) If ǫ = limn→∞ ǫn > 0, then x̄ = xǫ := arg minz∈F (y) N
j
j=1
(iii) In this last case, if γ satisfies the stricter bound γ < 1 −
32
2
K+2
(or,
equivalently, if
2γ
, that
K − 1−γ
2γ
1−γ
< K), then we have, for all z ∈ F(y) and any k <
kz − x̄k1 ≤ c̃σk (z)1 ,
#
"
2(1 + γ) K − k + 23
with c̃ :=
2γ
1−γ
K − k − 1−γ
(4.24)
As a consequence, this case is excluded if F(y) contains a vector of sparsity
2γ
.
k < K − 1−γ
Note that the approximation properties (4.23) and (4.24) are exactly
of the same order as the one (3.6) provided by ℓ1 -minimization. However,
in general, x̄ is not necessarily an ℓ1 -minimizer, unless it coincides with a
sparse solution.
The proof of this result is not included and the interested reader is referred
to [23, 39] for the details.
4.2.4
Local linear rate of convergence
It is instructive to show a further result concerning the local rate of convergence of this algorithm, which again uses the NSP as well as the optimality
conditions we introduced above. One assumes here that F(y) contains the
k-sparse vector x∗ . The algorithm produces a sequence xn , which converges
to x∗ , as established above. One denotes the (unknown) support of the
k-sparse vector x∗ by T .
For now, one introduces an auxiliary sequence of error vectors η n ∈ N
via η n := xn − x∗ and
En := kη n k1 = kx∗ − xn k1 .
Theorem 4.3 guarantees that En → 0 for n → ∞. A useful technical result
is reported next.
Lemma 4.4. For any z, z ′ ∈ RN , and for any j,
|σj (z)1 − σj (z ′ )1 | ≤ kz − z ′ k1 ,
(4.25)
(J − j)rJ (z) ≤ kz − z ′ k1 + σj (z ′ )1 .
(4.26)
while for any J > j,
33
′ ∈
Proof. To prove (4.25), approximate z by a best j-term approximation z[j]
′
Σj of z in ℓ1 . Then
′
σj (z)1 ≤ kz − z[j]
k1 ≤ kz − z ′ k1 + σj (z ′ )1 ,
and the result follows from symmetry. To prove (4.26), it suffices to note
that (J − j) rJ (z) ≤ σj (z)1 .
The following theorem gives a bound on the rate of convergence of En
to zero.
Theorem 4.5. Assume A satisfies the NSP of order K with constant γ.
2γ
2
Suppose that k < K − 1−γ
, 0 < ρ < 1, and 0 < γ < 1 − K+2
are such that
γ(1 + γ)
µ :=
1−ρ
1
1+
K +1−k
< 1.
Assume that F(y) contains a k-sparse vector x∗ and let T = supp(x∗ ). Let
n0 be such that
En0 ≤ R∗ := ρ min |x∗i |.
(4.27)
i∈T
Then, for all n ≥ n0 , we have
En+1 ≤ µ En .
(4.28)
Consequently, xn converges to x∗ exponentially.
Proof. The relation (4.16) with w = wn , xw = xn+1 = x∗ + η n+1 , and
η = xn+1 − x∗ = η n+1 , gives
N
X
(x∗i + ηin+1 )ηin+1 win = 0.
i=1
Rearranging the terms and using the fact that x∗ is supported on T , one
obtains
N
X
i=1
|ηin+1 |2 win = −
X
i∈T
x∗i ηin+1 win = −
X
i∈T
x∗i
η n+1 .
[(xni )2 + ǫ2n ]1/2 i
(4.29)
The proof of the theorem is by induction. Assume that En ≤ R∗ has already
been established. Then, for all i ∈ T ,
|ηin | ≤ kη n k1 = En ≤ ρ|x∗i |,
34
so that
|x∗i |
|x∗i |
|x∗i |
1
≤
,
=
≤
n
∗
n
n
2
2
1/2
|xi |
|xi + ηi |
1−ρ
[(xi ) + ǫn ]
(4.30)
and hence (4.29) combined with (4.30) and the NSP gives
N
X
i=1
|ηin+1 |2 win ≤
γ
1
kη n+1 k1 ≤
kη n+1
c k1
1−ρ T
1−ρ T
The Cauchy–Schwarz inequality combined with the above estimate yields
!
!
X
X
n 2
2 1/2
n+1 2
n+1 2 n
[(xi ) + ǫn ]
kηT c k1 ≤
|ηi | wi
i∈T c
i∈T c
=
N
X
i=1
≤
|ηin+1 |2 win
!
X
[(ηin )2 + ǫ2n ]1/2
i∈T c
γ
n
kη n+1
c k1 (kη k1 + N ǫn ) .
1−ρ T
!
(4.31)
n+1 is k-sparse and the algorithm
= 0, then xn+1
If ηTn+1
c
T c = 0. In this case x
n+1
has stopped by definition; since x
− x∗ is in the null space N , which
contains no k-sparse elements other than 0, one has already obtained the
6= 0, then cancelling the factor kηTn+1
solution xn+1 = x∗ . If ηTn+1
c k1 in (4.31)
c
yields
γ
(kη n k1 + N ǫn ) ,
kηTn+1
c k1 ≤
1−ρ
and thus
n+1
kη n+1 k1 = kηTn+1 k1 + kηTn+1
c k1 ≤ (1 + γ)kηT c k1 ≤
γ(1 + γ)
(kη n k1 + N ǫn ) .
1−ρ
(4.32)
Now, by (4.19) and (4.26) it follows
N ǫn ≤ rK+1 (xn ) ≤
kη n k1
1
(kxn − x∗ k1 + σk (x∗ )1 ) =
, (4.33)
K +1−k
K+1−k
since by assumption σk (x∗ )1 = 0. Together with (4.32) this yields the desired
bound,
γ(1 + γ)
1
n+1
En+1 = kη
k1 ≤
1+
kη n k1 = µEn .
1−ρ
K +1−k
In particular, since µ < 1, one has En+1 ≤ R∗ , which completes the induction step. It follows that En+1 ≤ µEn for all n ≥ n0 .
35
4.2.5
Superlinear convergence promoting ℓτ -minimization for τ <
1
The linear rate (4.28) can be improved significantly, by a very simple modification of the rule of updating the weight:
− 2−τ
2
2
2
,
)
+
ǫ
wjn+1 = (xn+1
n+1
j
j = 1, . . . , N, for any 0 < τ < 1.
This corresponds to the substitution of the function J with


N
X
τ
zj2 wj + ǫ2 wj + 2 − τ 1τ  ,
Jτ (z, w, ǫ) :=
2
τ w 2−τ
j=1
j
where z ∈ RN , w ∈ RN
+ , ǫ ∈ R+ . With this new up-date rule for the weight,
which depends on 0 < τ < 1, we have formally, for xn → x̄ and ǫn → 0,
Jτ (xn , wn , ǫn ) → kx̄kττ .
Hence such an iterative optimization tends to promote the ℓτ -quasi-norm
minimization.
Surprisingly the rate of local convergence of this modified algorithm is
superlinear; the rate is larger for smaller τ , and approaches a quadratic rate
as τ → 0. More precisely, the local error En := kxn − x∗ kττ satisfies
En+1 ≤ µ(γ, τ )En2−τ ,
(4.34)
where µ(γ, τ ) < 1 for γ > 0 sufficiently small. The validity of (4.34) is
restricted to xn in a (small) ball centered at x∗ . In particular, if x0 is close
enough to x∗ then (4.34) ensures the convergence of the algorithm to the
k-sparse solution x∗ , see Figure 4.
4.3
Numerical Experiments
Figure 5 shows a typical phase transition diagram related to the (experimentally determined) probability of successful recovery of sparse vectors by
means of the iteratively re-weighted least squares algorithm. For each point
of this diagram with coordinates (m/N, k/m) ∈ [0, 1]2 , we indicate the empirical success probability of recovery of a k-sparse vector x ∈ RN from m
measurements y = Ax. The brightness level corresponds to the probability.
As measurement matrix a real random Fourier type matrix A was used, with
entries given by
Ak,j = cos(2πjξk ), j = 1, . . . , N,
36
Comparison of the rate of convergence for different τ
5
10
τ=1
τ=0.8
τ=0.6
τ=0.56
τ=0.5, initial iter. with τ=1
0
Logarithmic error
10
−5
10
−10
10
−15
10
0
5
10
15
20
25
Number of iterations
30
35
40
45
Figure 4: The decay of logarithmic error is shown, as a function of the
number of iterations of IRLS for different values of τ (1, 0.8, 0.6, 0.56). We
show also the results of an experiment in which the initial 10 iterations are
performed with τ = 1 and the remaining iterations with τ = 0.5.
and the ξk , k = 1, ..., m, are sampled independently and uniformly at random
from [0, 1]. (Theorem 3.7 does not apply directly to real random Fourier
matrices, but an analogous result concerning the RIP for such matrices can
be found in [75].)
Figure 6 shows a section of a phase transition diagram related to the (experimentally determined) probability of successful recovery of sparse vectors
from linear measurements y = Ax, where the matrix A has i.i.d. Gaussian
entries. Here both m and N are fixed and only k is variable. This diagram establishes the transition from a situation of exact reconstruction for
sparse vectors with high probability to very unlikely recovery for vectors with
many nonzero entries. These numerical experiments used the iteratively reweighted least squares algorithm with different parameters 0 < τ ≤ 1. It
is of interest to emphasize the enhanced success rate when using the algorithm for τ < 1. Similarly, many other algorithms are tested by showing the
corresponding phase transition diagrams and comparing them, see [6] for a
detailed account of phase transitions for greedy algorithms and [28, 32] for
ℓ1 -minimization.
This section is concluded by showing applications of ℓ1 -minimization
methods to a real-life image recolorization problem [41, 42] in Figure 7. The
image is known completely only on very few colored portions, while on the
remaining areas only gray levels are provided. With this partial information,
the use of ℓ1 -minimization with respect to wavelet or curvelets coefficients
allows for high fidelity recolorization of the whole images.
37
Figure 5: Empirical success probability of recovery of k-sparse vectors
x ∈ RN from measurements y = Ax, where A ∈ Rm×N is a real random
Fourier matrix. The dimension N = 300 of the vectors is fixed. Each point
of this diagram with coordinates (m/N, k/m) ∈ [0, 1]2 indicates the empirical success probability of recovery, which is computed by running 100 experiments with randomly generated k-sparse vectors x and randomly generated
matrix. The algorithm used for the recovery is the iteratively re-weighted
least squares method tuned to promote ℓ1 -minimization.
5
Open Questions
The field of compressed sensing is rather young so there remain many directions to be explored and it is questionable whether one can assign certain
problems in the field already at this point the status of an “open problem”.
Anyhow, below we list two problems that remained unsolved until the time
of writing of this article.
5.1
Deterministic compressed sensing matrices
So far only several types of random matrices A ∈ Cm×N are known to satisfy
the RIP δs ≤ δ ≤ 0.4 (say) for
m = Cδ s logα (N )
38
(5.1)
Comparison of iterative re−weighted least squares for l1 and l1 → lτ minimization in Compressed Sensing
1
l1 minimization
Probability of exact reconstruction
0.9
l1 → lτ minimization
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
5
10
15
20
25
k.sparsity
Figure 6: Empirical success probability of recovery of a k-sparse vector
x ∈ R250 from measurements y = Ax, where A ∈ R50×250 is Gaussian. The
matrix is generated once; then, for each sparsity value k shown in the plot,
500 attempts were made, for randomly generated k-sparse vectors x. Two
different IRLS algorithms were compared: one with weights inspired by ℓ1 minimization, and the IRLS with weights that gradually moved during the
iterations from an ℓ1 - to an ℓτ -minimization goal, with final τ = 0.5.
for some constant Cδ and some exponent α (with high probability). This
is a strong form of existence statement. It is open, however, to provide
deterministic and explicit m × N matrices that satisfy the RIP δs ≤ δ ≤ 0.4
(say) in the desired range (5.1).
In order to show RIP estimates in the regime (5.1) one has to take into
account cancellations of positive and negative (or more generally complex)
entries in the matrix, see also Section 3.6. This is done “automatically”
with probabilistic methods but seems to be much more difficult to exploit
when the given matrix is deterministic. It may be conjectured that certain
equiangular tight frames or the “Alltop matrix” in [82, 70] do satisfy the
RIP under (5.1). This is supported by numerical experiments in [70]. It is
expected, however, that a proof is very hard and requires a good amount of
analytic number theory.
The best deterministic construction of CS matrices known so far uses
deterministic expander graphs [5]. Instead of the usual RIP, one shows
that the adjacency matrix of such an expander graph has the 1-RIP, where
the ℓ2 -norm is replaced by the ℓ1 -norm at each occurence in (3.8). The
1-RIP also implies recovery by ℓ1 -minimization. The best known deterministic expanders [17] yield sparse recovery under the condition m ≥
2
Cs(log N )c log (N ) . Although the scaling in s is linear as desired, the term
39
Figure 7: Iterations of the recolorization methods proposed in [41, 42] via ℓ1
and total variation minimization, for the virtual restoration of the frescoes
of A. Mantegna (1452), which were destroyed by a bombing during World
War II. Only a few colored fragments of the images were saved from the
disaster, together with good quality gray level pictures dated to 1920.
2
(log N )c log (N ) grows faster than any polynomial in log N . Another drawback is that the deterministic expander graph is the output of a polynomial
time algorithm, and it is questionable whether the resulting matrix can be
regarded as explicit.
5.2
Removing log-factors in the Fourier-RIP estimate
It is known [16, 73, 78, 75] that a random partial Fourier matrix A ∈ Cm×N
satisfies the RIP with high probability provided
m
≥ Cδ s log2 (s) log(N ).
log(m)
(The condition stated in (3.19) implies this one.) It is conjectured that one
can remove some of the log-factors. It must be hard, however, to improve this
to a better estimate than m ≥ Cδ,ǫ s log(N ) log(log N ). Indeed, this would
imply an open conjecture of Talagrand [85] concerning the equivalence of the
ℓ1 and ℓ2 norm of a linear combination of a subset of characters (complex
exponentials).
6
Conclusions
Compressive sensing established itself by now as a new sampling theory
which exhibits fundamental and intriguing connections with several mathematical fields, such as probability, geometry of Banach spaces, harmonic
40
analysis, theory of computability and information-based complexity. The
link to convex optimization and the development of very efficient and robust numerical methods make compressive sensing a concept useful for a
broad spectrum of natural science and engineering applications, in particular, in signal and image processing and acquisition. It can be expected that
compressive sensing will enter various branches of science and technology to
notable effect.
Recent developments, for instance the work [14, 76] on low rank matrix recovery via nuclear norm minimization, suggest new possible extensions of compressive sensing to more complex structures. Moreover, new
challenges are now emerging in numerical analysis and simulation where
high-dimensional problems (e.g., stochastic partial differential equations in
finance and electron structure calculations in chemistry and biochemistry)
became the frontier. In this context, besides other forms of efficient approximation, such as sparse grid and tensor product methods [10], compressive
sensing is a promising concept which is likely to cope with the “curse of
dimensionality”. In particular, further systematic developments of adaptivity in the presence of different scales, randomized algorithms, an increasing
role for combinatorial aspects of the underlying algorithms, are examples of
possible future developments, which are inspired by the successful history
of compressive sensing [84].
7
Cross-References
Compressive sensing has connections with the following chapters of the
book: Wavelets, Fourier Analysis, Compression, Astronomy, CT, Variational
Methods for Image Analysis, Numerical Methods for Variational Approach
in Image Analysis, Duality and Convex Minimization, Mumford Shah, Phase
Field Models, Iterative Solution Methods, Learning, Classification, Data
Mining, Radar, Sampling Methods, Linear Inverse Problems, Nonlinear Inverse Problems, Regularization Methods for Ill-Posed Problems, Seismic.
8
Recommended Reading
The initial papers on the subject are [13, 16, 26]. An introduction to
compressive sensing is contained in the monograph [45] by Rauhut and
Foucart under current preparation. Another introductory source are the
lecture notes [39, 75] of the summer school “Theoretical Foundations and
Numerical Methods for Sparse Recovery ”, held at RICAM in September
41
2009. The overview papers [12, 3, 11, 77] introduce to various theoretical and applied aspects of compressive sensing. A large collection of the
vastly growing research literature on the subject is available on the webpage
http://www.compressedsensing.com.
References
[1] D. Achlioptas. Database-friendly random projections. In Proc. 20th
Annual ACM SIGACT-SIGMOD-SIGART Symp. on Principles of
Database Systems, pages 274–281, 2001.
[2] F. Affentranger and R. Schneider. Random projections of regular simplices. Discrete Comput. Geom., 7(3):219–226, 1992.
[3] R. Baraniuk. Compressive sensing. IEEE Signal Process. Magazine,
24(4):118–121, 2007.
[4] R. G. Baraniuk, M. Davenport, R. A. DeVore, and M. Wakin. A simple
proof of the restricted isometry property for random matrices. Constr.
Approx., 28(3):253–263, 2008.
[5] R. Berinde, A. Gilbert, P. Indyk, H. Karloff, and M. Strauss. Combining geometry and combinatorics: A unified approach to sparse signal
recovery. preprint, 2008.
[6] J. D. Blanchard, C. Cartis, J. Tanner, and A. Thompson. Phase transitions for greedy sparse approximation algorithms. preprint, 2009.
[7] T. Blumensath and M. Davies. Iterative hard thresholding for compressed sensing. Appl. Comput. Harmon. Anal., 27(3):265–274, 2009.
[8] J. Bobin, J.-L. Starck, and R. Ottensamer. Compressed sensing in
astronomy. IEEE J. Sel. Topics Signal Process., 2(5):718–726, 2008.
[9] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge Univ.
Press, 2004.
[10] H.-J. Bungartz and M. Griebel. Sparse grids. Acta Numerica, 13:147–
269, 2004.
[11] E. Candès and M. Wakin. An introduction to compressive sampling.
IEEE Signal Process. Magazine, 25(2):21–30, 2008.
42
[12] E. J. Candès. Compressive sampling. In Proceedings of the International
Congress of Mathematicians, 2006.
[13] E. J. Candès, J., T. Tao, and J. Romberg. Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency
information. IEEE Trans. Inform. Theory, 52(2):489–509, 2006.
[14] E. J. Candès and B. Recht. Exact matrix completion via convex optimization. Found. Comput. Math.,, 9:717–772, 2009.
[15] E. J. Candès, J. Romberg, and T. Tao. Stable signal recovery from
incomplete and inaccurate measurements. Comm. Pure Appl. Math.,
59(8):1207–1223, 2006.
[16] E. J. Candès and T. Tao. Near optimal signal recovery from random
projections: universal encoding strategies? IEEE Trans. Inform. Theory, 52(12):5406–5425, 2006.
[17] M. Capalbo, O. Reingold, S. Vadhan, and A. Wigderson. Randomness
conductors and constant-degree lossless expanders. In Proceedings of the
Thirty-Fourth Annual ACM, pages 659–668 (electronic). ACM, 2002.
[18] S. S. Chen, D. L. Donoho, and M. A. Saunders. Atomic decomposition
by Basis Pursuit. SIAM J. Sci. Comput., 20(1):33–61, 1999.
[19] O. Christensen. An Introduction to Frames and Riesz Bases. Applied
and Numerical Harmonic Analysis. Birkhäuser, 2003.
[20] A. K. Cline. Rate of convergence of Lawson’s algorithm. Math. Comp.,
26:167–176, 1972.
[21] A. Cohen, W. Dahmen, and R. A. DeVore. Compressed sensing and
best k-term approximation. J. Amer. Math. Soc., 22(1):211–231, 2009.
[22] G. Cormode and S. Muthukrishnan. Combinatorial algorithms for compressed sensing. In CISS, 2006.
[23] I. Daubechies, R. DeVore, M. Fornasier, and C. Güntürk. Iteratively
re-weighted least squares minimization for sparse recovery. Comm. Pure
Appl. Math., 63(1):1–38, 2010.
[24] B. Do, P. Indyk, E. Price, and D. Woodruff. Lower bounds for sparse
recovery. In Proc. SODA, 2010.
43
[25] D. Donoho and B. Logan. Signal recovery and the large sieve. SIAM
J. Appl. Math., 52(2):577–591, 1992.
[26] D. L. Donoho. Compressed sensing. IEEE Trans. Inform. Theory,
52(4):1289–1306, 2006.
[27] D. L. Donoho. For most large underdetermined systems of linear equations the minimal l1 solution is also the sparsest solution. Commun.
Pure Appl. Anal., 59(6):797–829, 2006.
[28] D. L. Donoho. High-dimensional centrally symmetric polytopes with
neighborliness proportional to dimension. Discrete Comput. Geom.,
35(4):617–652, 2006.
[29] D. L. Donoho and M. Elad. Optimally sparse representation in general
(nonorthogonal) dictionaries via ell1 minimization. Proc. Natl. Acad.
Sci. USA, 100(5):2197–2202, 2003.
[30] D. L. Donoho and X. Huo. Uncertainty principles and ideal atomic
decompositions. IEEE Trans. Inform. Theory, 47(7):2845–2862, 2001.
[31] D. L. Donoho and J. Tanner. Neighborliness of randomly projected
simplices in high dimensions. Proc. Natl. Acad. Sci. USA, 102(27):9452–
9457, 2005.
[32] D. L. Donoho and J. Tanner. Counting faces of randomly-projected
polytopes when the projection radically lowers dimension. J. Amer.
Math. Soc., 22(1):1–53, 2009.
[33] D. L. Donoho and Y. Tsaig. Fast solution of l1-norm minimization
problems when the solution may be sparse. IEEE Trans. Inform. Theory, 54(11):4789–4812, 2008.
[34] R. Dorfman. The detection of defective members of large populations.
Ann. Statist., 14:436–440, 1943.
[35] M. Duarte, M. Davenport, D. Takhar, J. Laska, S. Ting, K. Kelly, and
R. Baraniuk. Single-Pixel Imaging via Compressive Sampling. Signal
Processing Magazine, IEEE,, 25(2):83–91, March , 2008.
[36] B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani. Least angle regression. Ann. Statist., 32(2):407–499, 2004.
44
[37] M. Elad and A. M. Bruckstein. A generalized uncertainty principle and
sparse representation in pairs of bases. IEEE Trans. Inform. Theory,
48(9):2558–2567, 2002.
[38] A. Fannjiang, P. Yan, and T. Strohmer. Compressed Remote Sensing
of Sparse Objects. preprint, 2009.
[39] M. Fornasier. Numerical methods for sparse recovery. In Theoretical
Foundations and Numerical Methods for Sparse Recovery, Radon Series
Comp. Appl. Math. deGruyter, in preparation.
[40] M. Fornasier, A. Langer, and C. B. Schönlieb. A convergent overlapping domain decomposition method for total variation minimization.
preprint, 2009.
[41] M. Fornasier and R. March. Restoration of color images by vector
valued BV functions and variational calculus. SIAM J. Appl. Math.,
68(2):437–460, 2007.
[42] M. Fornasier, R. Ramlau, and G. Teschke. The application of joint
sparsity and total variation minimization algorithms to a real-life art
restoration problem. Adv. Comput. Math., 31(1-3):157–184, 2009.
[43] S. Foucart. A note on guaranteed sparse recovery via ℓ1 -minimization.
Appl. Comput. Harmon. Anal., to appear.
[44] S. Foucart, A. Pajor, H. Rauhut, and T. Ullrich. The Gelfand widths
of ℓp -balls for 0 < p ≤ 1. preprint, 2010.
[45] S. Foucart and H. Rauhut. A Mathematical Introduction to Compressive
Sensing. Appl. Numer. Harmon. Anal. Birkhäuser, Boston, in preparation.
[46] J. J. Fuchs. On sparse representations in arbitrary redundant bases.
IEEE Trans. Inform. Theory, 50(6):1341–1344, 2004.
[47] A. Garnaev and E. Gluskin. On widths of the Euclidean ball. Sov.
Math., Dokl., 30:200–204, 1984.
[48] A. C. Gilbert, S. Muthukrishnan, S. Guha, P. Indyk, and M. Strauss.
Near-Optimal Sparse Fourier Representations via Sampling. In Proc.
STOC’02, pages 152 – 161. others, Association for Computing Machinery, 2002.
45
[49] A. C. Gilbert, S. Muthukrishnan, and M. J. Strauss. Approximation of
functions over redundant dictionaries using coherence. In Proceedings
of the Fourteenth Annual ACM-SIAM Symposium on Discrete Algorithms, Baltimore, MD, USA, January 12-14, 2003, pages 243–252.
SIAM and Association for Computing Machinery, 2003.
[50] A. C. Gilbert, M. Strauss, J. A. Tropp, and R. Vershynin. One sketch
for all: Fast algorithms for compressed sensing. preprint, 2006.
[51] E. Gluskin. Norms of random matrices and widths of finite-dimensional
sets. Math. USSR-Sb., 48:173–182, 1984.
[52] I. Gorodnitsky and B. Rao. Sparse signal reconstruction from limited
data using FOCUSS: a re-weighted minimum norm algorithm. IEEE
Trans. Signal Process., 45(3):600–616, 1997.
[53] R. Gribonval and M. Nielsen. Sparse representations in unions of bases.
IEEE Trans. Inform. Theory, 49(12):3320–3325, 2003.
[54] R. Horn and C. Johnson. Matrix Analysis. Cambridge University Press,
1990.
[55] W. B. Johnson and J. Lindenstrauss, editors. Handbook of the Geometry
of Banach Spaces Vol I. North-Holland Publishing Co., 2001.
[56] B. Kashin. Diameters of some finite-dimensional sets and classes of
smooth functions. Math. USSR, Izv., 11:317–333, 1977.
[57] C. Lawson. Contributions to the Theory of Linear Least Maximum
Approximation. PhD thesis, University of California, Los Angeles, 1961.
[58] M. Ledoux and M. Talagrand. Probability in Banach Spaces. Springer,
1991.
[59] B. Logan. Properties of High-Pass Signals. PhD thesis, Columbia University, 1965.
[60] G. G. Lorentz, M. von Golitschek, and Y. Makovoz. Constructive approximation: advanced problems. Springer, Berlin, 1996.
[61] S. G. Mallat and Z. Zhang. Matching pursuits with time-frequency
dictionaries. IEEE Trans. Signal Process., 41(12):3397–3415, 1993.
[62] S. Marple. Digital Spectral Analysis with Applications. Prentice - Hall,
1987.
46
[63] S. Mendelson, A. Pajor, and N. Tomczak Jaegermann. Uniform uncertainty principle for Bernoulli and subgaussian ensembles. Constr.
Approx., 28(3):277–289, 2009.
[64] B. K. Natarajan. Sparse approximate solutions to linear systems. SIAM
J. Comput., 24:227–234, 1995.
[65] Y. Nesterov and A. Nemirovskii. Interior-point polynomial algorithms
in convex programming, volume 13 of SIAM Studies in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, 1994.
[66] E. Novak. Optimal recovery and n-widths for convex classes of functions. J. Approx. Theory, 80(3):390–408, 1995.
[67] M. Osborne, B. Presnell, and B. Turlach. A new approach to variable
selection in least squares problems. IMA J. Numer. Anal., 20(3):389–
403, 2000.
[68] M. Osborne, B. Presnell, and B. Turlach. On the LASSO and its dual.
J. Comput. Graph. Statist., 9(2):319–337, 2000.
[69] G. E. Pfander and H. Rauhut. Sparsity in time-frequency representations. J. Fourier Anal. Appl., 16(2):233–260, 2010.
[70] G. E. Pfander, H. Rauhut, and J. Tanner. Identification of matrices having a sparse representation. IEEE Trans. Signal Process., 56(11):5376–
5388, 2008.
[71] R. Prony. Essai expérimental et analytique sur les lois de la Dilatabilité
des uides élastique et sur celles de la Force expansive de la vapeur de
leau et de la vapeur de lalkool, à différentes températures. J. École
Polytechnique, 1:24–76, 1795.
[72] H. Rauhut. Random sampling of sparse trigonometric polynomials.
Appl. Comput. Harmon. Anal., 22(1):16–42, 2007.
[73] H. Rauhut. Stability results for random sampling of sparse trigonometric polynomials. IEEE Trans. Information Theory, 54(12):5661–5670,
2008.
[74] H. Rauhut. Circulant and Toeplitz matrices in compressed sensing. In
Proc. SPARS’09, 2009.
47
[75] H. Rauhut. Compressive sensing and structured random matrices. In
Theoretical Foundations and Numerical Methods for Sparse Recovery,
Radon Series Comp. Appl. Math. deGruyter, to appear.
[76] B. Recht, M. Fazel, and P. Parillo. Guaranteed minimum rank solutions
to linear matrix equations via nuclear norm minimization. SIAM Rev.,
to appear.
[77] J. Romberg. Imaging via Compressive Sampling. IEEE Signal Process.
Magazine, 25(2):14–20, March, 2008.
[78] M. Rudelson and R. Vershynin. On sparse reconstruction from Fourier
and Gaussian measurements. Comm. Pure Appl. Math., 61:1025–1045,
2008.
[79] L. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based
noise removal algorithms. Physica D, 60(1-4):259–268, 1992.
[80] F. Santosa and W. Symes. Linear inversion of band-limited reflection
seismograms. SIAM J. Sci. Statist. Comput., 7(4):1307–1330, 1986.
[81] K. Schnass and P. Vandergheynst. Dictionary preconditioning for
greedy algorithms. IEEE Trans. Signal Process., 56(5):1994–2002,
2008.
[82] T. Strohmer and R. W. j. Heath. Grassmannian frames with applications to coding and communication. Appl. Comput. Harmon. Anal.,
14(3):257–275, 2003.
[83] T. Strohmer and M. Hermann. Compressed Sensing Radar. IEEE
Proc. Int. Conf. Acoustic, Speech, and Signal Processing, 2008, pages
1509–1512, 2008.
[84] E. Tadmor. Numerical methods for nonlinear partial differential equations. In Encyclopedia of Complexity and Systems Science. Springer,
2009.
[85] M. Talagrand. Selecting a proportion of characters. Israel J. Math.,
108:173–191, 1998.
[86] G. Tauböck, F. Hlawatsch, and H. Rauhut. Compressive Estimation
of Doubly Selective Channels: Exploiting Channel Sparsity to Improve
Spectral Efficiency in Multicarrier Transmissions. 2009.
48
[87] H. Taylor, S. Banks, and J. McCoy. Deconvolution with the ℓ1 -norm.
Geophys. J. Internat., 44(1):39–52, 1979.
[88] R. Tibshirani. Regression shrinkage and selection via the lasso. J. Roy.
Statist. Soc. Ser. B, 58(1):267–288, 1996.
[89] J. Traub, G. Wasilkowski, and H. Wo’zniakowski. Information-based
complexity. Computer Science and Scientific Computing. Academic
Press Inc., 1988.
[90] J. Tropp and D. Needell. CoSaMP: Iterative signal recovery from
incomplete and inaccurate samples. Appl. Comput. Harmon. Anal.,
page 30, 2008.
[91] J. A. Tropp. Greed is good: Algorithmic results for sparse approximation. IEEE Trans. Inform. Theory, 50(10):2231–2242, 2004.
[92] J. A. Tropp. Just relax: Convex programming methods for identifying
sparse signals in noise. IEEE Trans. Inform. Theory, 51(3):1030–1051,
2006.
[93] J. A. Tropp, J. N. Laska, M. F. Duarte, J. K. Romberg, and R. G.
Baraniuk. Beyond Nyquist: Efficient sampling of sparse bandlimited
signals. IEEE Trans. Inform. Theory, 56(1):520 –544, 2010.
[94] M. Unser. Sampling—50 Years after Shannon. Proceedings of the IEEE,
88(4):569–587, 2000.
[95] J. Vybiral. Widths of embeddings in function spaces. J. Complexity,
24(4):545–570, 2008.
[96] G. Wagner, P. Schmieder, A. Stern, and J. Hoch. Application of nonlinear sampling schemes to cosy-type spectra. J. Biomolecular NMR,
3(5):569, 1993.
49
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement