[IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA...
Transcript of [IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA...
![Page 1: [IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA (2010.06.13-2010.06.18)] 2010 IEEE International Symposium on Information Theory - Recovery thresholds](https://reader037.fdocuments.us/reader037/viewer/2022100123/5750ab821a28abcf0ce00479/html5/thumbnails/1.jpg)
Recovery thresholds for �1 optimization in binarycompressed sensing
Mihailo StojnicSchool of Industrial Engineering
Purdue University
West Lafayette IN 47907, USA
Email: [email protected]
Abstract—Recently, [3], [7] theoretically analyzed the successof a polynomial �1 optimization algorithm in solving an under-determined system of linear equations. In a large dimensionaland statistical context [3], [7] proved that if the number ofequations (measurements in the compressed sensing terminology)in the system is proportional to the length of the unknownvector then there is a sparsity (number of non-zero elementsof the unknown vector) also proportional to the length of theunknown vector such that �1 optimization succeeds in solvingthe system. In this paper, we consider the same problem whileadditionally assuming that all non-zero elements elements areequal to each other. We provide a performance analysis ofa slightly modified �1 optimization. As expected, the obtainedrecoverable sparsity proportionality constants improve on theequivalent ones that can be obtained if no information aboutthe non-zero elements is available. In addition, we conducteda sequence of numerical experiments and observed that theobtained theoretical proportionality constants are in a solidagreement with the ones obtained experimentally.Index Terms: compressed sensing, �1 optimization
I. INTRODUCTION
In last several years the area of compressed sensing has
been the subject of extensive research. One of the most
crucial compressed sensing problems is finding the sparsest
solution of an under-determined system of equations. While
this problem had been known for a long time it is the work
of [3] and [7] that rigorously proved for the first time that a
sparse enough solution can be recovered by solving a linear
program in polynomial time. These results generated enormous
amount of research with possible applications ranging from
high-dimensional geometry, image reconstruction, single-pixel
camera design, decoding of linear codes, channel estimation in
wireless communications, to machine learning, data-streaming
algorithms, DNA micro-arrays, magneto-encephalography etc.
(more on the compressed sensing problems, their importance,
and wide spectrum of different applications can be found in
excellent references [1], [5], [13], [21]–[23], [31]).
As is well known, solving an under-determined system of
linear equations assumes finding x such that
Ax = y (1)
where (in the compressed sensing terminology) A is an m×n(m < n) measurement matrix and y is an m×1 measurement
vector. Standard compressed sensing context assumes that x is
an n × 1 unknown k-sparse vector (see Figure 1; here and in
the rest of the paper, under k-sparse vector we assume a vector
that has at most k nonzero components). The main topic of
this paper will be compressed sensing of the so-called ideally
sparse signals (more on the so-called approximately sparse
signals can be found in e.g. [4], [33]). Also, in the rest of the
paper we will assume the so-called linear regime, i.e. we will
assume that k = βn and that the number of the measurements
is m = αn where α and β are absolute constants independent
of n. The following �1 optimization approach to solving (1)
k
n
m =
A xy
Fig. 1. Model of a linear system; vector x is k-sparse
attracted a great deal of attention in recent years
min ‖x‖1
subject to Ax = y. (2)
Quite remarkably, for certain statistical matrices A in [3], [7]
the authors were able to show that if α and n are given then
any unknown vector x with no more than k = βn (where β is
an absolute constant dependent on α and explicitly calculated
in [3], [7]) non-zero elements can be recovered, with over-
whelming probability, by solving (2). (Under overwhelming
probability we in this paper assume a probability that is no
more than a number exponentially decaying in n away from
1.) As expected, this assumes that y was in fact generated by
that x and given to us. (More on practically very important
scenario when the available measurements are noisy versions
of y can be found in [3], [32].)
The above described scenario is the standard compressed
sensing setup. Such a setup does not assume availability of any
a priori knowledge about the structure of the unknown k-sparse
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
1593978-1-4244-7892-7/10/$26.00 ©2010 IEEE ISIT 2010
![Page 2: [IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA (2010.06.13-2010.06.18)] 2010 IEEE International Symposium on Information Theory - Recovery thresholds](https://reader037.fdocuments.us/reader037/viewer/2022100123/5750ab821a28abcf0ce00479/html5/thumbnails/2.jpg)
signal x. In certain application one can however encounter
signals that have an a priori known structure (see, e.g. [1],
[14], [15], [28]–[30], [34]) and then try to design the recovery
algorithms that would exploit the knowledge of the signal
structure. In this paper we will also consider signals that have
certain structure. Namely, we will be interested in the recovery
of the so-called binary sparse signals x. For the simplicity
of the exposition we will under binary sparse signals assume
signals whose components are either zero or one (more on this
or similar discrete type of signals as well as on their potential
applications can be found in e.g. [6], [8], [11], [19], [20],
[31]). Consequently, under binary k-sparse signals x we will
consider signals that have k components equal to one and the
remaining ones equal to zero. It is also relatively easy to note
that the binary sparse signals are a subclass of the so-called
non-negative sparse signals studied in e.g. [11], [18], [27],
[35]. To recover the non-negative sparse signals one usually
considers a modification of (3). That modification assumes
adding positivity constraints on each of the components of x(see, e.g. [11], [18], [27], [35]). In a similar fashion to recover
binary k-sparse x we will consider the following modification
of (2) (see e.g. [11])
min ‖x‖1
subject to Ax = y
0 ≤ xi ≤ 1, 1 ≤ i ≤ n. (3)
Let βw be the maximum allowable value of the constant βsuch that (3) finds given binary k-sparse solution of (1) with
overwhelming probability. The value of βw is often referred
to as the weak threshold (for more on the definitions of the
weak threshold see, e.g. [7], [10], [27]). For a specific group
of randomly generated matrices A, we will in this paper
determine the values of βw for the entire range of α, i.e. for
0 ≤ α ≤ 1.
II. BASIC THEOREMS
In this section we introduce two useful theorems that will
be of key importance in our subsequent analysis. First we
introduce a null-space characterization of the matrix A that
guarantees that the solutions of (1) and (3) coincide. Since the
analysis will clearly be irrelevant with respect to what partic-
ular location of the nonzero elements are chosen, we can for
the simplicity of the exposition and without loss of generality
assume that the components x1,x2, . . . ,xn−k of x are equal
to zero and the components xn−k+1,xn−k+2, . . . ,xn of x are
equal to one. (One should note that while for our analysis it
is assumed that the components x1,x2, . . . ,xn−k of x are
equal to zero this fact is not known to the algorithm). Under
these assumptions we have the following theorem (similar
characterizations adopted in different contexts can be found
in [12], [17], [26], [28], [33], [35]).
Theorem 1: (Nonzero part of x has fixed location; all non-
zero elements of x are equal to one) Assume that an m × nmeasurement matrix A is given. Let x be a binary k-sparse
vector whose nonzero components are equal to one and let
x1 = x2 = · · · = xn−k = 0. Further, assume that y = Axand that w is an n × 1 vector. Let I = {1, 2, . . . , n − k} and
Ic = {n − k + 1, n − k + 2, . . . , n}. If (∀w ∈ Rn|Aw =0,wi ≥ 0, i ∈ I,wi ≤ 0, i ∈ Ic) it holds that
−n∑
i=n−k+1
wi <n−k∑i=1
wi (4)
then the solution of (3) is the binary k-sparse solution of (1)
Proof: Follows directly from the corresponding results in
e.g. [26], [35] by adding the obvious restriction wi ≤ 0, i ∈Ic.
Having matrix A such that (4) holds would be enough for the
solution of (3) to coincide with the binary k-sparse solution
of (1). If one assumes that m and k are proportional to n(the case of interest in this paper) then the construction of
the deterministic matrices A that would satisfy (4) is not
an easy task (in fact, one may say that it is one of the
most fundamental open problems in the area of theoretical
compressed sensing). However, turning to random matrices
significantly simplifies things. As we will see later in the paper,
the random matrices A that have the null-space uniformly
distributed in the Grassmanian will turn out to be a very
convenient choice. The following phenomenal result from [16]
that relates to such matrices will be the key ingredient in the
analysis that will follow.
Theorem 2: ( [16] Escape through a mesh) Let S be a
subset of the unit Euclidean sphere Sn−1 in Rn. Let Y be
a random (n − m)-dimensional subspace of Rn, distributed
uniformly in the Grassmanian with respect to the Haar mea-
sure. Let
w(S) = E supw∈S
(hT w) (5)
where h is a random column vector in Rn with i.i.d. N (0, 1)components. Assume that w(S) <
(√m − 1
4√
m
). Then
P (Y ∩ S = 0) > 1 − 3.5e−(√
m− 14√
m−w(S)
)2
18 . (6)
Remark: Gordon’s original constant 3.5 was substituted by
2.5 in [24]. Both constants are fine for our subsequent analysis.
III. PROBABILISTIC ANALYSIS OF THE NULL-SPACE
CHARACTERIZATIONS
In this section we probabilistically analyze validity of the
null-space characterization given in Theorem 1. We will show
how one can obtain the values of the weak threshold βw for
the entire range 0 ≤ α ≤ 1 based on such an analysis.
A. Weak threshold
As masterly noted in [24] Theorem 2 can be used to
probabilistically analyze (4). Following the procedure of [27]
we set Sw
Sw = {w ∈ Sn−1|−∑i∈Ic
wi ≥∑i∈I
wi,wi ≥ 0, i ∈ I,wi ≤ 0, i ∈ Ic}(7)
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
1594
![Page 3: [IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA (2010.06.13-2010.06.18)] 2010 IEEE International Symposium on Information Theory - Recovery thresholds](https://reader037.fdocuments.us/reader037/viewer/2022100123/5750ab821a28abcf0ce00479/html5/thumbnails/3.jpg)
and
w(Sw) = E supw∈Sw
(hT w) (8)
where h is a random column vector in Rn with i.i.d. N (0, 1)components and Sn−1 is the unit n-dimensional sphere. Let
Y be an (n − m) dimensional subspace of Rn uniformly
distributed in Grassmanian. Furthermore, let Y be the null-
space of A. Then as long as w(Sw) <(√
m − 14√
m
), Y
will miss Sw (i.e. (4) will be satisfied) with probability no
smaller than the one given in (6). More precisely, if α = mn
is a constant (the case of interest in this paper), n,m are
large, and w(Sw) is smaller than but proportional to√
m then
P (Y ∩ Sw = 0) −→ 1. This in turn is equivalent to having
P ((∀w ∈ Rn|Aw = 0,wi ≥ 0, i ∈ I,wi ≤ 0, i ∈ Ic),
−∑i∈Ic
wi <∑i∈I
wi) −→ 1
which according to Theorem 1 means that the solution of (3)
coincides with probability 1 with the binary k-sparse solution
of (1). For any given value of α ∈ (0, 1] a threshold value
of β can be then determined as a maximum β such that
w(Sw) <(√
m − 14√
m
). That maximum β will be exactly
the value of the weak threshold βw. If one is only concerned
with finding a possible value for βw it is easy to note that
instead of computing w(Sw) it is sufficient to find its an upper
bound. However, as we will soon see, to determine as good
values of βw as possible, the upper bound on w(Sw) should
be as tight as possible. The main contribution of this work as
well as of [27] is a fairly precise estimate of w(Sw). As in
[27] we can write
w(h, Sw) = maxw∈Sw
(hT w)
= maxw∈Sw
(∑i∈I
hiwi +∑i∈Ic
(−hi)(−wi)). (9)
Let
h(I) = (h1,h2, . . . ,hn−k)T
and
h(Ic) = (−hn−k+1,−hn−k+2, . . . ,−hn)T .
Further, let h(I)(i) be the i-th smallest of the elements of h(I). In
a similar fashion, let h(Ic)(i) be the i-th smallest of the elements
of h(Ic). Set
h̄ = (h(I)(1),h
(I)(2), . . . ,h
(I)(n−k),h
(Ic)(1) ,h(Ic)
(2) , . . . ,h(Ic)(k) )T .
Then one can simplify (9) in the following way
w(h, Sw) = maxy∈Rn
h̄T a
subject to ai ≥ 0, 0 ≤ i ≤ nn∑
i=n−k+1
ai ≥n−k∑i=1
ai
n∑i=1
a2i ≤ 1. (10)
Let z ∈ Rn be a column vector such that zi = 1, 1 ≤ i ≤ (n−k), zi = −1, n−k+1 ≤ i ≤ n, and zi = 0, n−k+1 ≤ i ≤ n.
Repeating the derivation given in [27] one has, based on the
Lagrange duality theory, that there are c(1)w = (1 − θ
(1)w )n ≤
(n − k) and c(2)w = (1 − θ
(2)w )n ≤ k such that
w(Sw) = E(w(h, Sw)) ≤ O(√
ne−nconst.)
+ (En−k∑
i=c(1)w +1
h̄2i + E
n∑i=n−k+c
(2)w +1
h̄2i
− (E(h̄T z) − E∑c(1)
wi=1 h̄i + E
∑n−k+c(2)w
i=n−k+1 h̄i)2
n − c(1)w − c
(2)w
)12 , (11)
where h̄i is the i-th element of vector h̄ and const is an
absolute constant independent of n. One should note the
appearance of c(2)w which corresponds to the fact that now
(differently from [27]) all components of a in (10) are non-
negative. Furthermore, similarly to what was established in
[27], one can establish the following two relations that c(1)w
and c(2)w satisfy
(E(h̄T z) − E∑c(1)
wi=1 h̄i + E
∑n−k+c(2)w
i=n−k+1 h̄i)
(n − c(1)w − c
(2)w )/(1 − ε)
≥ F−1c
((1 + ε)c(1)
w
n(1 − β)
)
(E∑c(1)
wi=1 h̄i − E(h̄T z) − E
∑n−k+c(2)w
i=n−k+1 h̄i)
(n − c(1)w − c
(2)w )/(1 + ε)
≥ F−1c
((1 − ε)c(2)
w
nβ
).
(12)
As in [27], F−1c (·) in the inequalities in (12) stands for the
inverse cdf of the zero-mean, unit variance Gaussian random
variable and ε > 0 is an arbitrarily small constant independent
of n. When n is large the first term on the right hand side of
the inequality in (11) goes to zero. Finding c(1)w and c
(2)w that
satisfy (12) and minimize the second term on the right hand
side of the inequality in (11) is then enough to determine
a valid upper bound on w(Sw) and therefore to compute
the values of the weak threshold. The theorem that we will
present below summarizes how an attainable value of the weak
threshold can be computed. It follows by literally repeating
each of the remaining steps of the procedure established in
[27] while relying on results from [2], [25].
To facilitate the presentation of the theorem, we before
proceeding further first simplify notation . Let erfinv be the
inverse of the standard error function associated with zero-
mean unit variance Gaussian random variable. Further, let
f1(θ, β) be the following function of arguments θ and β
f1(θ, β) = 2(erfinv(21 − θ
1 − β− 1))2. (13)
Also, let f2(θ, β) be the following function of arguments θand β
f2(θ, β) =
⎧⎨⎩(1 − β) − (1 − θ) + (1−β)
√f1(θ,β)√
2πef1(θ,β)2(1−θ)1−β ≥ 1
(1 − β) − (1 − θ) − (1−β)√
f1(θ,β)√2πef1(θ,β) otherwise.
(14)
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
1595
![Page 4: [IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA (2010.06.13-2010.06.18)] 2010 IEEE International Symposium on Information Theory - Recovery thresholds](https://reader037.fdocuments.us/reader037/viewer/2022100123/5750ab821a28abcf0ce00479/html5/thumbnails/4.jpg)
Finally, let f3(θ(1), θ(2), β) be the following function of argu-
ments θ(1), θ(2), and β
f3(θ(1), θ(2), β) =(1 − β)e−(erfinv(
2(1−θ(1))1−β −1))2
√2π
− βe−(erfinv(2(1−θ(2))
β −1))2
√2π
. (15)
Theorem 3: (Weak threshold; non-zero elements of x are
equal to one) Let A be an m × n measurement matrix in (1)
with the null-space uniformly distributed in the Grassmanian.
Let the unknown x in (1) be binary k-sparse. Further, let the
location of the nonzero elements of x be arbitrarily chosen but
fixed. Assume that the non-zero elements of x are equal to one.
Let k, m, n be large and let α = mn and βw = k
n be constants
independent of m, n, and k. Let ε > 0 be an arbitrarily small
constant and let functions f1(·, ·), f2(·, ·), and f3(·, ·, ·) be as
defined in (13), (14), and (15), respectively. Further let θ(1)w ,
(βw ≤ θ(1)w ≤ 1), θ
(2)w , (1 − βw/2 ≤ θ
(2)w ≤ 1), be such that
(1 − ε)f3(θ(1)w , θ
(2)w , βw)
θ(1)w + θ
(2)w − 1
≥√
2erfinv(2(1 + ε)(1 − θ
(1)w )
1 − βw− 1)
(1 + ε)f3(θ(1)w , θ
(2)w , βw)
θ(1)w + θ
(2)w − 1
≤ −√
2erfinv(2(1 − ε)(1 − θ
(2)w )
βw−1).
(16)
If α and βw further satisfy
α > f2(θ(1)w , β)+f2(θ(2)
w , 1−β)−(f3(θ
(1)w , θ
(2)w , βw)
)2
θ(1)w + θ
(2)w − 1
(17)
then the solution of (3) coincides with overwhelming proba-
bility with the binary k-sparse solution of (1).
Proof: Follows from the previous discussion and analysis
presented in [27].
The results for the weak threshold obtained from the above
theorem are presented in Figure 2. For the comparison we also
show the values of the weak threshold for the so-called non-
negative signals. The threshold values obtained in that case
correspond to the ones computed in [27] and (as demonstrated
in [27]) match the best currently known ones from [9], [10].
Since, binary signals are a subclass of the non-negative signals
(and as such are clearly more structured) one would then
expect that they have higher recoverable sparsity thresholds.
Figure 2 hints that the �1 optimization algorithm from (3)
indeed recovers higher sparsity. Also, it is interesting to note
that if α → 12 then β
α → 1. This means that in binary
compressed sensing one needs at most m = n2 measurements
no matter how sparse the signal is (more on this phenomenon
the interested reader can find in an excellent recent work [20]).
Another interesting point is that in binary compressed sensing
the understanding of sparsity can be twofold. Traditionally
one views sparsity as the number of non-zero components
of x. However, given the symmetry of the binary signals, it
is relatively easy to see that the sparsity can be viewed as
the number of zeros as well (a simple signal transformation
x′i = 1 − xi, 1 ≤ i ≤ n, translates one sparsity to another).
This, on the other hand, implies that if one can recover
all signals of sparsity (understood in the traditional way as
the number of non-zeros) less than or equal to k with mmeasurements, then with the same m measurements one can
recover all signals of sparsity larger than or equal to (n − k)as well.
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
α
β/α
Weak threshold, l1
optimization, xi∈ {0,1}
Signed x [Donoho/Tanner; Stojnic] x
i∈ {0,1} Present paper
Fig. 2. Weak threshold; �1 optimization, xi ∈ {0, 1}IV. NUMERICAL EXPERIMENTS
In this section we briefly discuss the results that we obtained
from numerical experiments. In all our numerical experiments
we fixed n = 200. We then generated matrices A of size
m × n with m2 ∈ {10, 20, 30, 40, 45, 49, 59, . . . , 99}. The
components of the measurement matrices A were generated as
i.i.d. zero-mean unit variance Gaussian random variables. For
each m we generated k-sparse signals x for several different
values of k from the transition zone (the locations of non-
zero elements of x were chosen randomly and fixed). For
each combination (k, m) we generated 200 different problem
instances and recorded the number of times the �1 optimization
algorithm from (3) failed to recover the correct k-sparse x. The
obtained data are then interpolated and graphically presented
in Figure 3. The color of any point in Figure 3 shows the
probability of having �1 optimization from (3) succeed for a
combination (α, β) that corresponds to that point. The colors
are mapped to probabilities according to the scale on the right
hand side of the figure. The simulated results can naturally be
compared to the weak threshold theoretical prediction. Hence,
we also show in Figure 3 the theoretical value for the weak
threshold calculated according to Theorem 3 (and shown in
Figure 2). We observe that the simulation results are in a good
agreement with the theoretical calculation.
V. CONCLUSION
In this paper we considered recovery of sparse signals from
a reduced number of linear measurements. We additionally
assumed that the sparse signal is binary which essentially
means that the non-zero components are equal to one. We
provided a theoretical performance analysis of a modified
polynomial �1 optimization algorithm when used for recovery
of such signals. Under the assumption that the measurement
matrix A has a basis of the null-space distributed uniformly in
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
1596
![Page 5: [IEEE 2010 IEEE International Symposium on Information Theory - ISIT - Austin, TX, USA (2010.06.13-2010.06.18)] 2010 IEEE International Symposium on Information Theory - Recovery thresholds](https://reader037.fdocuments.us/reader037/viewer/2022100123/5750ab821a28abcf0ce00479/html5/thumbnails/5.jpg)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
α
β/α
Recoverable sparsity, n=200, xi∈ {0,1}
l1 optimization fails
l1 optimization succeeds
theoretical threshold, xi∈{0,1}
Fig. 3. Experimentally recoverable sparsity; �1 optimization xi = {0, 1}the Grassmanian, we derived lower bounds on the values of the
recoverable weak thresholds in the so-called linear regime, i.e.
in the regime when the recoverable sparsity are proportional to
the length of the unknown vector. Obtained threshold results
suggest that unknown sparse signals with additional binary
structure can have higher recoverable sparsity.
The results presented in this paper seem encouraging and
one may wonder if further extensions are possible. One of
the most important observations is that the sparse signals that
we considered in this paper are discrete (in fact very simple
discrete signals). For such simple discrete signals we were
able to show that one can improve the recoverable thresholds.
Naturally, one can ask if results of similar flavor can be
obtained for more general discrete signals. For example, it
would be very interesting (especially given the importance of
digital signals in practical considerations) to see if one can
design an adaptation of the standard �1 optimization so that
the unknown signals with components, say, from an integer set
larger than the binary one considered here can be recovered
as well. This will be the subject of a future work.REFERENCES
[1] R. Baraniuk, V. Cevher, M. Duarte, and C. Hegde. Model-basedcompressive sensing. available online at http://www.dsp.ece.rice.edu/cs/.
[2] A. Barvinok and A. Samorodnitsky. Random weighting, asymptoticcounting, and inverse isoperimetry. Israel Journal of Mathematics,158:159–191, 2007.
[3] E. Candes, J. Romberg, and T. Tao. Robust uncertainty principles: exactsignal reconstruction from highly incomplete frequency information.IEEE Trans. on Information Theory, 52:489–509, December 2006.
[4] A. Cohen, W. Dahmen, and R. DeVore. Compressed sensing and bestk-term approximation. Journal of the American Mathematical Society,22(1), January 2009.
[5] S. F. Cotter and B. D. Rao. Sparse channel estimation via matching pur-suit with application to equalization. IEEE Trans. on Communications,50(3), 2002.
[6] W. Dai and O. Milenkovic. Weighted superimposed codes and con-strained integer compressed sensing. IEEE Trans. on InformationTheory, 55(9):2215–2219, September 2009.
[7] D. Donoho. High-dimensional centrally symmetric polytopes withneighborlines proportional to dimension. Disc. Comput. Geom.,35(4):617–652, 2006.
[8] D. Donoho, A. Maleki, and A. Montanari. Message-passing algo-rithms for compressed sensing. Proc. National Academy of Sciences,106(45):18914–18919, Nov. 2009.
[9] D. Donoho and J. Tanner. Sparse nonnegative solutions of underdeter-mined linear equations by linear programming. Proc. National Academyof Sciences, 102(27):9446–9451, 2005.
[10] D. Donoho and J. Tanner. Thresholds for the recovery of sparse solutionsvia l1 minimization. Proc. Conf. on Information Sciences and Systems,March 2006.
[11] D. Donoho and J. Tanner. Counting the face of randomly projectedhypercubes and orthants with application. 2008. available online athttp://www.dsp.ece.rice.edu/cs/.
[12] D. L. Donoho and X. Huo. Uncertainty principles and ideal atomic de-compositions. IEEE Trans. Inform. Theory, 47(7):2845–2862, November2001.
[13] M. Duarte, M. Davenport, D. Takhar, J. Laska, T. Sun, K. Kelly, andR. Baraniuk. Single-pixel imaging via compressive sampling. IEEESignal Processing Magazine, 25(2), 2008.
[14] Y. C. Eldar and M. Mishali. Robust recovery of signals from a structuredunion of subspaces. 2008. available at arXiv:0807.4581.
[15] A. Ganesh, Z. Zhou, and Y. Ma. Separation of a subspace-sparse signal:Algorithms and conditions. IEEE International Conference on Acoustics,Speech and Signal Processing, pages 3141–3144, April 2009.
[16] Y. Gordon. On Milman’s inequality and random subspaces which escapethrough a mesh in Rn. Geometric Aspect of of functional analysis, Isr.Semin. 1986-87, Lect. Notes Math, 1317, 1988.
[17] R. Gribonval and M. Nielsen. Sparse representations in unions of bases.IEEE Trans. Inform. Theory, 49(12):3320–3325, December 2003.
[18] M. A. Khajehnejad, A. G. Dimakis, W. Xu, and B. Hassibi. Sparserecovery of positive signals with minimal expansion. Preprint, 2009.available at arXiv:0902.4045.
[19] O. L. Mangasarian and M. C. Ferris. Uniqueness of integer solutionof linear equations. Data Mining Institute Technical Report 09-01, July2009. available online at http://www.cs.wisc.edu/math-prog/tech-reports.
[20] O. L. Mangasarian and B. Recht. Probability of unique inte-ger solution to a syatem of linear equations. Data Mining Insti-tute Technical Report 09-02, September 2009. available online athttp://www.cs.wisc.edu/math-prog/tech-reports.
[21] F. Parvaresh, H. Vikalo, S. Misra, and B. Hassibi. Recovering sparse sig-nals using sparse measurement matrices in compressed dna microarrays.IEEE Journal of Selected Topics in Signal Processing, 2(3):275–285,June 2008.
[22] B. Recht, M. Fazel, and P. A. Parrilo. Guaranteed minimum-ranksolution of linear matrix equations via nuclear norm minimization. 2007.available online at http://www.dsp.ece.rice.edu/cs/.
[23] J. Romberg. Imaging via compressive sampling. IEEE Signal ProcessingMagazine, 25(2):14–20, 2008.
[24] M. Rudelson and R. Vershynin. On sparse reconstruction from Fourierand Gaussian measurements. Comm. on Pure and Applied Math., 61(8),2007.
[25] S. M. Stigler. The asymptotic distribution of the trimmed mean. Analysisof Statistics, 1:472–477, 1973.
[26] M. Stojnic. A simple performance analysis of �1-optimization incompressed sensing. ICASSP, International Conference on Acoustics,Signal and Speech Processing, April 2009.
[27] M. Stojnic. Various thresholds for �1-optimization in compressedsensing. Preprint, 2009. available at arXiv:0907.3666.
[28] M. Stojnic, F. Parvaresh, and B. Hassibi. On the reconstruction of block-sparse signals with an optimal number of measurements. IEEE Trans.on Signal Processing, August 2009.
[29] J. Tropp, A. C. Gilbert, and M. Strauss. Algorithms for simultaneoussparse approximation. part i: Greedy pursuit. Signal Processing, Aug2005.
[30] E. van den Berg and M. P. Friedlander. Joint-sparse recovery frommultiple measurements. Preprint, 2009. available at arXiv:0904.2051.
[31] H. Vikalo, F. Parvaresh, and B. Hassibi. On sparse recovery ofcompressed dna microarrays. Asilomor conference, November 2007.
[32] M. J. Wainwright. Sharp thresholds for high-dimensional and noisyrecovery of sparsity. Proc. Allerton Conference on Communication,Control, and Computing, September 2006.
[33] W. Xu and B. Hassibi. Compressed sensing over the grassmannmanifold: A unified analytical framework. 2008. available online athttp://www.dsp.ece.rice.edu/cs/.
[34] A. C. Zelinski, L. L. Wald, K. Setsompop, V. K. Goyal, and E. Adal-steinsson. Sparsity-enforced slice-selective mri rf excitation pulsedesign. IEEE Trans. on Medical Imaging, 27(9):1213–1229, Sep. 2008.
[35] Y. Zhang. A simple proof for recoverability of ell-1-minimization (ii):the nonnegative case. Rice CAAM Department Technical Report TR05-10, 2005. available at http://www.dsp.ece.rice.edu/cs/.
ISIT 2010, Austin, Texas, U.S.A., June 13 - 18, 2010
1597