Вы находитесь на странице: 1из 32

Accepted Manuscript

Joint pilot placement and symbol design scheme for sparse Channel
Estimation in OFDM systems

Anthony Ngozichukwuka Uwaechia, Nor Muzlifah Mahyuddin

PII: S1874-4907(17)30284-7
DOI: https://doi.org/10.1016/j.phycom.2017.11.008
Reference: PHYCOM 465

To appear in: Physical Communication

Received date : 4 July 2017


Revised date : 30 September 2017
Accepted date : 27 November 2017

Please cite this article as: A.N. Uwaechia, N.M. Mahyuddin, Joint pilot placement and symbol
design scheme for sparse Channel Estimation in OFDM systems, Physical Communication (2017),
https://doi.org/10.1016/j.phycom.2017.11.008

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
Joint Pilot Placement and Symbol Design Scheme for
Sparse Channel Estimation in OFDM Systems
Anthony Ngozichukwuka Uwaechia, Nor Muzlifah Mahyuddin∗
School of Electrical and Electronic Engineering, Universiti Sains Malaysia, Seri
Ampangan, 14300, Nibong Tebal, Pulau Pinang, Malaysia.

Abstract
Existing investigations into the design of proper pilot placements for sparse
Channel Estimation (CE) in Orthogonal Frequency Division Multiplexing
(OFDM) system are based on the assumption that the pilot symbols are
equally-powered. However, this assumption may not necessarily exhibit low
coherence compressed CE. This paper investigates the joint pilot design prob-
lem for sparse CE in OFDM systems. In order to avoid the disjoint optimiza-
tion of the pilot symbol values and their placements, a joint pilot placement
and pilot symbol design scheme is proposed that optimizes over both the pilot
symbol values and their placements as a single design optimization problem.
The approach is based on minimizing the mutual coherence of the Fourier
submatrix used for the sparse signal recovery and therefore, can improve the
CE accuracy. Simulation results demonstrate that the proposed scheme is
effective and offer a better CE performance - in terms of Mean Square Error
(MSE) and Bit Error Rate (BER), when compared to former pilot place-
ment schemes that assume the equally powered pilot and other schemes that
jointly design the pilot symbols and their placement. It was also observed
that the proposed scheme can realize 18.75% improvement in bandwidth effi-
ciency with the same CE performance compared with the least squares (LS)
CE.
Keywords: Channel estimation, compressed sensing, pilot design, signal
reconstruction, sparse matrices, wireless communication.


Corresponding author.
Email address: eemnmuzlifah@usm.my (Nor Muzlifah Mahyuddin)

Preprint submitted to Nuclear Physics B September 30, 2017


1. Introduction
Orthogonal Frequency Division Multiplexing (OFDM) is a spectrally ef-
ficient multicarrier modulation technique, capable of establishing high-speed
digital transmission over frequency-selective fading channels [1, 2, 3]. OFDM
manages to decompose a dispersive frequency-selective fading channel into
multiple flat fading subcarriers with independent noises, which can typically
be compensated with a simple one-tap equalizer [1, 4, 5]. As in many wire-
less communication systems that use coherent detection, Channel Estimation
(CE) is critical for OFDM receiver design [6], as accurate knowledge about
the channel state can remarkably improve performance. In the conventional
CE techniques, such as Least-Squares (LS), the equally-spaced pilot pattern
is known to be optimal [6, 7]. However, it generally result in a lower spectral
efficiency, since smaller spacing in time between pilot symbols is required to
effectively track the channel time variations [2, 8].
Recently, under the assumption of channel sparsity, CE utilizing Com-
pressed Sensing (CS) can significantly reduce pilot overhead and offers a
desirably high spectral and energy efficiency for the same Mean Square Er-
ror (MSE) of CE [5, 9, 10, 11, 12]. This indeed has proven to be more efficient
over the conventional Least Squares (LS) [11, 13, 14]. Typically, CS focuses
on solving the following underdetermined system of linear equations

y = Φx + n, (1)
where y ∈ RM of size M consist of linear measurements of some k-sparse
signal x ∈ RN of size N (i.e., contains only k nonzero elements in x ), obtained
through the application of a measurement matrix Φ ∈ RM ×N with M < N
[15], while n ∈ RM denotes the additive noise vector, which can be modeled
either as a white noise vector, usually Gaussian [16] or as deterministic and
bounded [11, 16, 17].
Empirical studies have demonstrated that [9, 18, 19, 20, 21, 22], the appli-
cation of CS to sparse CE in OFDM systems can lead to both a significant re-
duction in pilot overhead -which improves spectral efficiency and the accuracy
of CE. Therefore, for an efficient CS-based sparse CE in OFDM systems, the
fewer number of pilots used for channel probing, should be properly designed
[23, 24]. However, while the optimal pilot placement, can be determined
by exhaustively searching all possible pilot placements, it is computationally
intensive and usually not feasible to achieve
 in practice [4, 21, 23]. For in-
stance, the exhaustive search over 256 16
, typically requires the generation of

2
all 1.008 × 1025 possible candidates which necessarily have to be examined.
In accordance with the Restricted Isometry Property (RIP) [25], it has been
shown that the randomly generated pilot pattern guarantees a high probabil-
ity of sparse recovery [14, 23]. Although the performance of such method is
usually statistically optimal [23], they are extremely slow for real-time imple-
mentation [14]. Therefore, most pilot pattern design schemes are proposed to
design deterministic pilot patterns [2, 14, 18, 20, 22, 26], which are typically
based on minimizing the coherence of the submatrix of the unitary Discrete
Fourier Transform (DFT) grid connected with the pilot subcarriers.
In [9], a technique that assumes equal power allocation assignment of
pilot symbols and uses the discrete stochastic approximation scheme to ob-
tain the optimal pilot design is proposed. However, the technique in [9],
only focuses on the optimization of the pilot placements of the equally pow-
ered pilot symbols. In [20], a technique that distributes equal power to pilot
symbols is studied using an iterative tree-based search scheme to optimize
the pilot placement for sparse CE in OFDM systems. Similarly, in [21], a
technique that allocates equal power to all pilot symbols and employs the
cross-entropy optimization technique to achieve the optimal pilot placement
is introduced. However, the optimization technique in [20] and [21] only
addresses the problem of the deterministic pilot placement for an equally
powered pilot symbols. Similarly, in [18], an equally powered pilot scheme
that optimizes pilot placements based on the summation of column corre-
lation of the measurement matrix is proposed. Therefore, for a successful
recovery of the target signal, the optimization method should consider the
design joint design of the pilot symbols and their placements.
In [27], a technique that jointly optimizes the pilot placement and power
distribution of the pilot symbols according to the Cyclic Different Set (CDS)
and the Almost Difference Set (ADS) is proposed. However, the proposed
scheme in [27] is only suitable for situations where either the CDS or ADS is
available. Moreover, in practical OFDM systems, it is not guaranteed that
a CDS or ADS will exist for every pair of (N, M ), where M out of N total
OFDM system subcarriers are utilized for pilot transmission. In [26] and [28]
a technique that jointly optimizes pilot symbols and their placements were in-
troduced for sparse CE in OFDM systems, but were addressed particularly as
a disjoint pilot placement and pilot symbol design sub-problems. Although,
the solutions derived from solving each sub-problem may be optimal, it may
not be possible to achieve the best optimal performance of the combined
solution. In [26], the pilot placement and pilot symbol optimization prob-

3
lem were formulated as a second-order cone programming (SOCP) problem,
and MOSEK (an interior-point optimizer) was applied in solving the conic
quadratic problem. However, the associated interior point technique in such
solver will arbitrarily slow down the convergence of loop optimization and
hence, necessarily exhibit a deterioration in performance since the the pri-
mal, dual, or gap residuals do not achieve the prescribed accuracy. In [28],
the (SOCP) optimization problem was approximated using a cost function,
and the Nesterov’s accelerated gradient technique was applied in converting
this problem to a first-order method for implementation. However, optimiz-
ing a non-differentiable function using the gradient descent method requires
a smoothing scheme, which necessarily leads to a non-equivalent set of con-
straints for the purpose of minimization. In [29] a scheme for designing joint
pilot placements and pilot symbols for OFDM systems operating in a high-
mobility scenario is proposed. This technique employs a discrete stochastic
approximation and uses an iterative algorithm to find the optimal pilot design
by minimizing the average coherence (a measure of the spread in columns)
rather than the mutual coherence (a measure of the worst-case correlation
of the similarity between columns) of the measurement matrix. Although
in the literature the average coherence measure usually results in guarantee-
ing good recovery properties, they are not better suited for recovering noisy
sparse signals as compared to the mutual coherence measure (usually referred
to as the worst-case coherence measure).
In this paper, in order to avoid the disjoint optimization of the pilot place-
ments and pilot symbol design sub-problems, a joint pilot placement and pilot
symbol design scheme is proposed. The proposed scheme optimizes over both
the pilot placements and pilot symbol values as a single design optimization
problem to minimize the mutual coherence (i.e., the worst-case correlation
between columns) of the measurement matrix. To address the resulting com-
binatorial search optimization problem, a local search optimization technique
is employed which starts with an initial population of candidate solution that
are randomly generated (a good choice, since randomized pilots are known
to be statistically optimal) and iteratively moves to the best neighbor among
all neighbors as a solution. Simulation results demonstrate that the pro-
posed pilot design scheme outperforms other different methods in terms of
constructing sensing matrices with lower mutual coherence and estimating
channels with lower MSE and system Bit Error Rate (BER).
The rest of this paper is organized as follows. Section 2 presents the
OFDM system model and the optimization objective of the pilot design

4
scheme. In Section 3, the proposed scheme is introduced. The simulation
results are presented in Section 4. Finally, Section 5 concludes the paper.
The notations used in this paper are defined as follows. Boldface lower
and upper case letters are used for vectors and matrices, respectively. k · k0
and k · k2 represent the l0 -norm and the l2 -norm, respectively. The super-
scripts (·)T and (·)H represent the transpose and the Hermitian transpose
of a matrix, respectively. CN represents the complex Gaussian distribution,
((·))N represents the modulo-N operation, #{·} represents the cardinality
of a set, \ represents the set exclusion, diag{·} represents the operator that
transforms a vector to a diagonal matrix. IN , CM ×N and R represent the
N th-order identity matrix, the set of M × N matrices in the complex field
and the real field, respectively.

2. System Model and Mutual Coherence Analysis


In this section, the system model is first described, followed by an overview
of the mutual coherence analysis.

2.1. System Model


The vast majority of the existing mathematical models of pilot-aided
OFDM CE essentially emerge from the assumption that the wireless com-
munication channel is a rich multipath [12, 19]. Contrarily, such peculiarly
rich multipath structure may be violated in most physical channels, and thus,
exhibits a particular structure in the channel impulse response (CIR), in the
sense that it is highly sparse [8, 19, 30]. Hence, the CIR as a time-variant
tap delay line is presented as
L−1
X
h(t, τ ) = hl (t)δ(τ − τl Ts ), (2)
l=0
where hl = h1 , h2 , . . . , hL is the complex channel tap gain and τl is the corre-
sponding time delay to each observed signal sample with a sampling period
of Ts and L representing the total number of propagation delay paths. In a
multipath channel with non-sample-spaced time delays, when τl is an integer,
the energy from all paths hl can be located and mapped to the CIR h by a
delay line with L taps. Therefore, Eq. (2) can be revised as
L
X
h(n) hl δ(n − τl ), (3)
l=1

5
where the complex channel tap (hl = h1 , h2 , . . . , hL ) gain are assumed to
be k-sparse. In this paper, an OFDM system with a comb-type pilot ar-
rangement is considered. The reason is that the comb-type pilot arrange-
ment system provides better resistance to fast fading channels, unlike the
block-type pilot arrangements [31]. In other words, they satisfy the need
for equalizing even as the channel changes from one OFDM block to an-
other. Let Z ∈ CM ×N denote a pilot selection matrix that selects M pilot
placement (using elements from an N -dimensional vector), in terms of the
pilot subcarriers, which are required for CE. Let N denote the total number
of subcarrier in each OFDM symbol (or, equivalently, the Discrete Fourier
Transform (DFT) size), with M subcarriers indexed with pilots in the form of
p = {p1 , p2 , . . . , pM } that belong to the set {1 ≤ p1 < p2 < . . . < pM ≤ N },
known a priori to the receiver. If the equivalent transmitted signal is denoted
as x(0), x(1), . . . , x(N − 1) where N − M are data symbols with M pilots,
the received signal vector can be expressed asfollows

y = XH + η = XFh + η, (4)

where y , [y(0), y(1), . . . , y(N − 1)]T denotes the received channel measure-
ments, X is an N × N diagonal matrix represented as X , diag{x(0), x(1),
. . . , x(N − 1)} which denotes the baseband-equivalent transmitted signal on
the main diagonal, H , [H(0), H(1), . . . , H(N − 1)]T denotes the sampled
Channel Frequency Response (CFR) vector, and η , [η(0), η(1), . . . , η(N −
1)]T ∼ CN (0, ση2 IN ) denotes the vector of additive white Gaussian noise
(AWGN) experienced by the channel, h , [h(0), h(1), . . . , h(L − 1)]T de-
notes the k-sparse received baseband CIR with length L, F denotes an
N × L partial DFT submatrix which contains only the first L columns of
a standard N × N DFT submatrix whose (m, n)th element of F is given by
[F ]m,n = √1N e−j2πmn/N , where 0 ≤ m ≤ N − 1 and 0 ≤ n ≤ L − 1. Therefore,
the received signals at pilot location can be expressed as

y (p) = X (p)F (p)h + η(p) = Ah + η(p), (5)

where y (p) , Zy , [y(p1 ), y(p2 ), . . . , y(pM )]T is an M × 1 vector which


denotes the received channel measurements at the pilot placement set p,
X (p) , ZXZ T , diag{x(p1 ), x(p2 ), . . . , x(pM )} denotes the transmitted
signal matrix at the pilot placement set p, η(p) , Z η , [η(p1 ), η(p2 ),
. . . , η(pM )]T ∼ CN (0, ση2 IM ) denotes the AWGN at pilot placement set p,
F (p) , ZF denotes an M × L DFT submatrix which contains only the rows

6
that corresponds to the pilot placement set p and the first L columns of a
standard N × N DFT submatrix and whose (m, n)th element of F is given
by [F ]m,n = √1N e−j2πmn/N , where 0 ≤ m ≤ M − 1 and 0 ≤ n ≤ L − 1 while
A , X (p)F (p) represents the designed M × L measurement matrix.
The mathematical model of Eq. (5) directly corresponds to a signal mea-
surement process that is nonadaptive, which senses a k-sparse signal h by
taking M linear measurements of the signal. Hence, h is reliably recov-
ered from the knowledge of the compressed measurement vector y and the
measurement matrix A. If the measurement matrix, A consist of a higher
number of rows than columns (M > L) then h can adequately be estimated
by LS CE method. However, if A consists of a lesser number of rows than
columns (M < L), then matrix A becomes ill-conditioned and the resulting
system of linear equations becomes underdetermined. Hence, the method
of LS which is capable of accurately estimating h under the condition of
M > L, becomes inaccurate. Since the number of rows (i.e., pilot signals) is
far lesser than the number of columns (i.e., channel coefficients), the system
significantly achieves a reduction in pilot overhead, and hence leads to the
conservation of spectral efficiency. However, through CS, h is accurately es-
timated and the originally transmitted signal is uniquely reconstructed on a
condition that the original signal is k-sparse.

2.2. Mutual Coherence Analysis


In CS-based sparse signal recovery, optimal pilot patterns are required to
successfully reconstruct the original sparse signal at the receiver. Although
the exhaustive search is the natural instinctive approach to select the optimal
pilot set, it is usually computationally infeasible [23]. Alternatively, to ob-
tain the optimal pilot pattern, the corresponding measurement matrix should
satisfy the Null Space Property (NSP) [11] or RIP [22]; but it is computa-
tionally intractable to verify if a given random measurement matrix satisfies
these properties [11, 32]. Fortunately, by minimizing the mutual coherence
of the measurement matrix in the CS framework, the corresponding pilot
pattern for CS-based CE can be optimized [23, 33]. This can be defined as
follows.

Definition 1 (Definition 1. of [34]): The mutual coherence of an M × L


measurement matrix A denoted as µ{A} is defined as the largest absolute
and normalized inner product correlation between any two separate columns

7
of A. This can formally be expressed as

|ha H
m , a n i|
µ{A} = max . (6)
0≤m,n≤L−1;m6=n ka m k2 · ka n k2

where a m is the mth column of A. Hence, µ{A} measures the maximum


possible linear dependency attained by any two columns of matrix A. The
following proposition show a connection between the RIP constant denoted
as δk and the mutual coherence denoted as µ [35].

Preposition 1: If a 1 , a 2 , . . . , a L−1 ∈ CM are the columns of A with mutual


coherence µ, and with column normalization such that a H m a m = 1. Therefore,
A will satisfy RIP of order k with constant δk = (k − 1)µ when δk < 1. In
other words, good measurement matrices can be constructed with low mutual
coherence.
Recent studies in [34, 36, 37], have established that the Basis Pursuit
(BP) and Orthogonal Matching Pursuit (OMP) algorithms can accurately
recover the k-sparse vector h, if the following theorem is satisfied.

Theorem 1 ([16, 34]): If F = [f1 , f2 , . . . , fN ], with fn ∈ RN is a rep-


resentation
PN basis for a real-valued, discrete-time signal x ∈ RN , and then
x = i=1 fi hi , where {hi }Ni=1 are sparse representation of the channel coeffi-
cients, which form an N × 1 vector h = [h1 , h2 , . . . , hN ]T . Then the signal x
is k-sparse with respect to F in some representation x = Fh if the following
condition is satisfied
 
1 1
k = khk0 < 1+ , (7)
2 µ{F }

where khk0 = #{hi 6= 0, i = 1, 2, . . . , N } is the l0 -norm measure that counts


the k-nonzero coefficients in h. Hence, h is the unique solution that is also
the sparsest representation of the signal x . Suppose further that the sensing
matrix X has been selected independent
P of the representation basis F , then
the vector h in y = Xx = X N f
i=1 i i h = XFh in some representation can
be solved by BP or OMP algorithm as described in the following corollary.

Corollary 1: If F is a representation basis for a real-valued, discrete-time


signal, and X is a sensing matrix independently selected, suppose that XF
satisfies the RIP if the representation y = Xx = XFh = Ah (where A

8
denotes the measurement matrix) satisfies the requirement
 
1 1
k = khk0 < 1+ , (8)
2 µ{XF }
such that if the measure of µ{XF } is as small as possible, then both BP
and OMP will manage to recover h perfectly and the deviation of the recon-
structed signal ĥ from the original signal h is bounded by
c2
kĥ − hk22 ≤ , (9)
1 − µ{XF }(2k − 1)
for some absolute constant c > 0 [29, 38], and thus can reconstruct x well.
This typically indicates that, for a k-sparse signal h, the smaller µ{A} is, the
better the approximation of h that can be obtained. According to Eq.(9),
it thus, appears that the reconstruction algorithm only manages to recover
h since µ{XF } > µ{F }. Hence, the performance deteriorates with higher
value of µ{XF } in terms of MSE. This, however, questions the efficacy of
mutual coherence towards the actual atomic behaviour of sparse represen-
tations and the performance of reconstruction algorithms [34, 39]. Various
optimization methods such as the MSE of the oracle estimator produces MSE
results that coincide with the unbiased Cramér-Rao bound (CRB). They are
very difficult to solve due to the nonconvex rank constraint [40]. Therefore,
if the grip of expectation is loosened with the anticipation that, a little por-
tion of signals with the corresponding representation cardinality will fail by
minimizing the mutual coherence, then values of khk0 considerably beyond
the above bound will still lead to successful CS whilst having the major ad-
vantage of lower computational complexity [34]. It is noteworthy to mention
that there are various bounds for random matrices, however a paraphrase of
an elegant result is thus presented.
If a measurement matrix, A ∈ RM ×N satisfies RIP of order 2k with Re-
stricted Isometry Constant (RIC) δ2k ∈ (0, 21 ] [11], then the required number
of linear measurements denoted as Mreq for successfully signal recovery is
given by
Mreq = O(klog(N/k)), (10)
where k and N denote the signal sparsity level and the size of the signal,
respectively. Consequently, Eq. (10) results in a decoding complexity of
O(Mreq log (N/Mreq )), which renders a substantial improvement over recent
solvers [41]. Moreover, from Eq. (10), it should be noted that, as the sparsity

9
level k (i.e., the number of nonzero coefficients) of the signal increases, the
number of required measurements Mreq also increases for the same values
of N . Hence, if sparsity is exploited (with far fewer nonzero coefficients),
the number of required measurements to successfully probe the channel is
reduced.

3. Optimal Pilot Design


In this section, the measurement matrix optimization metric is first pre-
sented, hereafter; the description of the proposed scheme then follows.

3.1. Pilot Design Optimization for CS


CS-based sparse CE requires optimal pilot set whose corresponding mea-
surement matrices have small mutual coherences in order to successfully ex-
ploit the inherently low-dimensionality of the sparse channel. Since the design
objective is to minimize the mutual coherence of the measurement matrix A,
a joint pilot placement and pilot symbol design scheme is proposed. The pro-
posed scheme optimizes over both the pilot placement and the pilot symbol
values as a single design optimization problem with the aim of minimizing
the mutual coherence of matrix A. Therefore, from Eq. (6), the mutual co-
herence of the measurement matrix A denoted as µ{A} is adopted as the
optimization design rule, where the objective function is to minimize µ{A},
is formulated in the expression that follows.
P
| M i=1 X(pi )e
−j2πpi (m)
X (pi )ej2πpi (n) |
µ{A} = max PM 2
i=1 |X (pi )|
0≤m<n≤L−1
P (11)
| M i=1 |x(p i )|2 j2πpi (n−m)
e |
= max PM .
2
i=1 |X (pi )|
0≤m<n≤L−1

According to [22, 23], studies have shown that under specific conditions on the
OFDM system sub-carriers N , and the number of subcarriers that transmit
the pilots signals M , the optimal pilot set that minimizes µ{A} can be
generated from a CDS based on the following definition.

Definition 2: A cyclic difference set (N, M, λ)-CDS is a subset {p1 , p2 ,


. . . , pM −1 } of integers modulo N such that each elements of 1, 2, . . . , N − 1
can be depicted as {pi − pj : i 6= j} modulo N in exactly λ number of ways
[27].

10
For example, a (23, 11, 5)-CDS is {1, 2, 3, 4, 6, 8, 9, 12, 13, 16, 18}, which
satisfies that any integer between 1 and 22 will occur and repeat exactly
λ = 5 times (i.e., λ = 11(11 − 1)/(23 − 1) = 5) in the set {Λ , ((βi −
βj ))23 |0 ≤ i ≤ j ≤ 10, i 6= j} with λ · N = 115 entries. Hence, the pilot set
generated from (N, M, λ)-CDS will meet the Welch’s bound on the maximum
cross-correlation amplitude and thus, yields the optimal pilot subset for the
equation in (11). Unfortunately, the availability of a CDS is not guaranteed
for every pilot size, since the presence of the difference sets is restricted to
some specific pair of (N, M ), and hence, unavailable in most practical OFDM
systems. Thus, proper pilot pattern design is desirable for sparsity-based CE
in OFDM systems.
In this case, since the periodic structure of the Fourier transform sub-
matrix only depends on q = n − m, Eq. (11) can, therefore, be re-expressed
as
P
| M 2 j2πpi q
i=1 |x(pi )| e |
µ{X(p)F(p)} = max P M
, (12)
2
i=1 |X (pi )|
0≤q≤L−1

where µ{A} = µ{X(p)F(p)}. Since lower µ{X(p)F(p)} lead to better


CS performance, then the objective of the pilot design scheme can be for-
mulated as one that minimizes µ{X(p)F(p)} over the pilot placement set
p = {pi }M
i=1 , and is presented as

Q = min µ{X (p)F (p)}. (13)


p

Consequently, the solution to the pilot design optimization problem of Eq. (13)
over the pilot placement can be expressed as

Λ(p)opt = arg min µ{X (p)F (p)}


p
P (14)
| M 2 j2πpi q
i=1 |x(pi )| e |
= arg min max P M
.
2
i=1 |X (pi )|
p 0≤q≤L−1

From the pilot design optimization problem of Eq. (14), it can be deduced
that the pilot symbol contribution in the mutual coherence measure is via
their magnitudes. Hence, optimal pilot symbol values and their placement
is therefore considered as a joint optimization problem, for proper pilot pat-
tern design. In this case, assume that there are R pilot symbol values that
corresponds to R pilot placement subsets {br }R r=1 . This will consequently

11
result in {b1 ∪ b2 ∪ · · · ∪ bR } = p (which signifies the placements of these pilot
symbol values in the set p), where p denotes the pilot placement set. Since
the impact of the pilot symbol values according to Eq. (14) is expressed as
Er , |X (par )|2 , where par ∈ br for ar = {1, 2, . . . M } and r = 1, 2, . . . R, the
objective function for the joint pilot design scheme over both pilot placement
and pilot symbols can therefore be reformulated as
PR P j 2π p q

| r=1 p a r ∈b r
E r e N ar |
Q = min max PR P . (15)
par ∈br Er
br ,Er 0≤q≤L−1
r=1

Hence, the solution to the pilot design optimization problem of Eq. (15) can
be re-expressed as
P P j 2π

| R
r=1 par ∈br Er e
p q
N ar |
Λ(p)opt = arg min max PR P . (16)
br ,Er 0≤q≤L−1
r=1 par ∈br E r

For simplicity, let Λ∗opt = {Λ(p)}∗opt represent the optimal pilot pattern over
both pilot placement and pilot symbol values. Since Eq. (16) is fundamentally
a combinatorial optimization problem, a local neighborhood search-based
heuristic is proposed to find a near optimal solution.

3.2. Proposed Pilot Design Scheme


To address the combinatorial search optimization problem of Eq. (16),
a local-neighborhood-search-based heuristic is employed to obtain the near-
optimal pilot pattern. Let b = {{p1 , . . . , pM } ∈ {1, . . . , N }} be the pilot
candidate placement set that selects neighborhood locations. Therefore, to
exhaustively examine candidates in the neighborhood of the current solu-
tion, only b tentative moves that belongs to part of the modified solution
needs to be searched. However, searching every vector in the neighborhood
extends the search space complexity since not every vector in the neighbor-
hood influences a reduction in MSE. Hence, through dimensionality reduc-
tion, a reduced neighborhood becomes crucial, to efficiently choose vectors
that will possibly influence a reduction in MSE while reducing complexity.
In this case, the search space of the pilot candidate set is reduced to the set
Ω , {{1, t, . . . , N − t}, t = #{p}}. Hence, the sets of pilot candidate is then
denoted as  


br = p par ∈ Ω , ar = 1, . . . , M (17)

12
Observe that, since the local search algorithm starts with an initial popula-
tion of candidate solution and iteratively moves to the best neighbor among
all neighbors as a solution, a neighborhood structure is required. Hence, the
neighbors of the pilot set p are defined as
 
0

N (p, br ) = |p̃ ∩ p| = M − 1 , p̃ 0 ∈ br (18)

It follows that to determine the effective neighbor set of the solution, is in a


way, synonymous with determining the effective move to make i.e., the move
that minimizes the mutual coherence in each iteration. It is noteworthy
to mention that the sets which have previously been crossed in the earlier
iterations necessarily needs to be eliminated, so as to bypass any feasible loop.
Additionally, in scenarios where no improving neighbor from the set of all
neighbor solutions (i.e., where all neighbor solutions are inferior to those in
the current set, in the form of minimum mutual coherence) is found, accepting
inferior solutions will allow the search to escape local minima. Therefore, the
algorithm selects the best-performing “non-improving neighbor” as a solution
during the current stage for the next iteration. This search process continues
iteratively until no improved solution is found after a d number of iterations.
In the proposed scheme, detailed descriptions for the joint pilot place-
ment and symbol design is presented as Algorithm 1. Some basic definitions
and descriptions that form the essential basis of the proposed scheme are
presented as follows. Define p and p̃ as different pilot placement sets. Al-
gorithm 1 starts by taking as input an initial feasible solution Λ∗0 with pilot
symbols Λ0 and pilot placement set p 0 = {{p1 , p2 , . . . , pM } ∈ {1 ≤ p1 < p2 <
· · · < pM ≤ N }}, which is randomly generated once as the initial solution
(since the randomly-generated pilot set are known to be theoretically opti-
mal [20]). The notation Λ∗best represents the optimal pilot design, and the the
notation Λ∗d represents the pilot design solution in the dth iteration which are
stored in the matrix Γd , while their corresponding mutual coherence value
are updated in µ∗min . To avoid becoming trapped in local minima (where
no improving neighbor solution is found), a criterion is employed that forces
the algorithm to select the best-performing “non-improving neighbor” after
Wmax consecutive failures as a solution during the current stage for the next
iteration. A local-variable counter Lcal is incremented in each stage of itera-
tion once there are no improving neighbors found within the neighborhood.
The entire search process continues iteratively until no improved solution is
found after a specific d number of iterations.

13
Algorithm 1 Joint pilot placement and symbol algorithm
Input: Initial pilot Λ∗0 = Λ0 (p 0 )∗opt with initial pilot symbols Λ0 and initial
pilot placement set p 0 ;
Output: The pilot signal Λ∗best with corresponding µmin
1: Initialization: Set Λ∗best = Λ∗0 , Set maximum value of threshold Wmax ,
set size of DFT subcarriers N , set size of p as M , set pilot symbol value
Er , |X (par )|2 , where par ∈ br for ar = {1, 2, . . . M } and r = 1, 2, . . . R,
(where R pilot symbol values corresponds to R pilot placement subsets
∗ ∗
{br }R
r=1 ). Set µmin = µ{Λ0 }, Set d = 1, Set Lcal = 1;
Iteration:
2: Determine the best neighbor of Λ∗d
PR P 2π
| r=1 par ∈br Er ej N par q |
Λ(p)∗d = arg min max PR P ,
par ∈br Er
br ,Er 0≤q≤L−1
r=1

3: Update: Γd
4: Γd = [Γd−1 Λ∗d ]
5: Compare Λ∗d with Λ∗best
6: if µ{Λ∗d } < µ∗min then
7: Λ∗best = Λ∗d
8: µ∗best = µ{Λ∗d }
9: Lcal = 1 {resets Lcal to “default” value ‘1’}
10: else
11: Lcal = Lcal + 1 {increment Lcal }
12: end if ;
13: Set d = d + 1, go to step 2 if Lcal ≤ Wmax and continue with a new
iteration.
14: return Λ∗best , and µ∗min

Nonetheless, if a new optimal solution is found that minimizes µ∗min for


Lcal ≤ Wmax , the algorithm resets the counter variable Lcal back to 1 and
continues the search process until Lcal > Wmax . Accordingly, the algorithm
begins at some initial probable random solution Λ∗0 ∈ br and iteratively
searches for a proper solution in the neighbourhood of Λ∗0 . If a proper solution
Λ∗best ∈ N (Λ∗0 ) is detected, then the search recommences in the neighbour-
hood, Λ∗best of the new solution. The algorithm stops and returns a feasible

14
solution Λ∗best , when Lcal > Wmax .

3.3. Complexity Analysis


The computational complexity of the proposed algorithm is deduced based
on the rate of growth of time taken with respect to the input, which is
equivalent to the number of complex multiplication required. The algorithm
consists of five main steps: Initialization, determination of the best solution
within the neighborhood, updating the search result matrix, incrementing
the iteration and the return step. However, the computational complexity
of the initialization process in step 1 can be omitted since the parameters
have initially been computed before run time and stored in memory to avoid
repeated computation. In step 2, the algorithm searches at most M (L − M )
pilot placements in X (p)F (p) = A (having M rows with L columns), and
therefore needs a computation in the order of O(LM ) to compute the mu-
tual coherence. Subsequently, it identifies the pilot signal with the least
value of mutual coherence with a computation complexity in the order of
O(M (L − M )). Step 4 requires no complex operation i.e., the worst case
run time is constant O(1). Step 5 requires O(M ) returning the M pilot
signal. Hence, the total complex multiplication required is in the order of
O(Wmax L2 M 2 ) where Wmax in this case is the number of times the algorithm
iterates before termination.

4. SIMULATION RESULTS
With the purpose of evaluating the proposed pilot design scheme, simu-
lation experiments were conducted and presented in this section. An OFDM
system with a data sequence modulated by 4QAM is considered with N =
256 subcarriers, where M = 16 subcarriers are assumed to be pilots unless
otherwise mentioned. The length L of the sparse Rayleigh multipath fading
channel h, is modeled with L = 50 taps, where k = 6 possitions are nonzero
randomly generated channel taps that are independently and identically dis-
tributed (i.i.d)CN(0, 1). For the recovery of the sparse channels, the OMP
[36] algorithm was applied and the MSE and BER performance obtained
were averaged over 1500 sparse channel realizations. In order to evaluate
the performance of the proposed scheme, the MSE and BER were adopted
to quantize the channel estimation errors. The entire simulations were per-
formed using MATLAB v8.5 (Release 2015a) on a PC Workstation equipped

15
Table 1: Pilot pattern design scheme by different methods for N = 256 and M = 16
OFDM system settings.
Method µ{A} Runtime (s) Pilot placement
PT1: Method of [18] 0.2305 48.219 7, 17, 20, 29, 33, 79, 85, 137,
156, 159, 165, 174, 178, 202,
206, 239
PT2: Method of [39] 0.2291 32.319 19, 74, 99, 104, 109, 129, 139,
144, 159, 174, 189, 199, 209,
234, 249, 254
PT3: Method of [23] 0.2317 20.025 8, 40, 48, 52, 72, 82, 99, 142,
145, 154, 158, 161, 183, 209,
212, 230
Proposed Scheme 0.2269 21.712 10, 40, 68, 91, 102, 112, 120,
132, 150, 168, 179, 189, 204,
236, 243, 252

with Intel Core i5-4460 CPU at 3.20GHz with 4GB installed Random Access
Memory (RAM) .
The first experiment was performed under (N, M, L) = (256, 16, 50) set-
tings of the OFDM system (i.e, utilizing M = 16 out of N = 256 subcarriers
for pilot signal transmission). This experiment compares the performance
of the proposed scheme with the methods of [23], [18] and [39] in terms of
the measurement matrix coherences µ{A}, complexity (which is provided
through the CPU runtime), and the designed pilot placement. The obtained
simulation results are presented in Table 1. It is observed that the com-
plexity of the proposed scheme is lower compared to the methods of [18]
and [39]. However, Method [23] exhibits a somewhat lower complexity than
the proposed scheme. This can be easily explained as the proposed scheme
devotes more time searching every vector within its neighborhood, and nec-
essarily advances towards increasing complexity in terms of runtime. It is
worth mentioning that, in order to obtain an improved result of optimization,
the corresponding measurement matrices associated with the pilot subcar-
riers should have small mutual coherence. Hence, in Table 1, the proposed
scheme is observed to be more effective in terms of the minimization of the
measurement matrix mutual coherence than the Method of [23], [18] and [39],
since it possesses a much smaller value of µ{A}. This implies that the pro-
posed scheme is able to amass more informative projection matrices, and will
drastically reduce the error for an equal number of measurements M , com-

16
1.2
Pilot symbol

0.8

0 20 40 60 80 100 120 140 160 180 200 220 240


Pilot placement

Fig. 1: The joint pilot placement and pilot symbol design.

pared to the different methods under investigation, or equivalently, achieve


comparable signal recovery performance with lower M . Fig. 1 illustrates the
power distribution of the pilot symbols of the proposed scheme. The height
of the stem indicates the power distribution among pilot symbols which is
not constrained to be of equal power. Nonetheless, in each OFDM frame,
the total power is normalized to one.
In the second experiment (conducted under (N, M, L) = (256, 13, 50)
OFDM system settings), the probability of exact signal recovery from noise-
less measurements was studied for all the different methods listed in Table 1
and with the proposed scheme. In this experiment, the Signal-to-Noise Ra-
tio(SNR) of the CIR, h is defined as
 
khk2
SNR(h) = 10log10 dB
kh − ĥk2

Hence, this experiment compares the sparse signal recovery probability with
respect to the sparsity of the channel h for SNR (h) ≥ 22dB. For ease of
notation in the figure, PT1, PT2 and PT3 are used to represent the three

17
1
PT2
0.9 PT1
PT3
0.8 Proposed Scheme

0.7
Recovery probability

0.6

0.5

0.4

0.3

0.2

0.1

0
2 4 6 8 10 12 14 16
Channel sparsity

Fig. 2: Recovery probability for different pilot design schemes (N, M, L) = (256, 16, 50).

different pilot patterns under consideration as listed in Table I, correspond-


ing to the Method of [18], Method of [39] and Method of [23], respectively.
Fig. 2 compares the sparse signal recovery probability with respect to the
sparsity level of the channel for SNR ≥ 22 dB. It can be observed that with
increasing number of sparsity k, the percentage of successful channel recovery
probability decreases for all the different pilot pattern design schemes. This
is due to the fact that as the level of sparsity k increases for the total num-
ber of allocable subcarriers N , the number of required measurements Mreq
also increases in accordance with Eq. (10). For higher sparsity levels, this
consequently renders the available pilots (i.e., M = 16) used in probing the
channel, to become insufficient, and hence necessarily exhibit an increased
deterioration in the signal recovery percentage. Nonetheless, it can be ob-
served that the proposed scheme is more robust and offered the best channel
recovery performance against channel sparsity compared to the methods of
[18] (PT1), [39] (PT2) and [23] (PT3).
Additionally, Fig. 3 and Fig. 4 show the MSE and BER performance
of the proposed scheme in comparison with the other different pilot place-

18
10 -1

10 -2
MSE

Random, M=16
10 -3 LS, M=52
LS, M=64
PT3, M=16
PT1, M=16
PT2, M=16
Proposed scheme, M=16

10 -4
0 5 10 15 20 25 30 35 40
SNR (dB)

Fig. 3: Comparison of channel estimation performance in terms of MSE performance for


different pilot design schemes (N, M, L) = (256, 16, 50).

ment methods under investigation as listed in Table 1. The equally-spaced


pilot pattern with pilot intervals of 5 and 4 (i.e., {1, 6, 11, . . . , 256} and
{1, 5, 9, . . . , 253}) are considered as it appears to be the best for the conven-
tional LS channel estimation. Furthermore, a randomly generated pilot pat-
tern {3, 51, 123, 134, 149, 152, 158, 160, 208, 212, 213, 215, 219, 231, 246, 254}
is also considered. The randomly generated pilot pattern is once gener-
ated by the pilot selection matrix Z ∈ CM ×N , which selects pilot place-
ment using elements from an N -dimensional vector in terms of the pilot
subcarriers (where N = 256 denotes the total number of OFDM subcar-
riers with M = 16 pilots subcarriers). It is worth mentioning that the
proposed scheme starts the local search by adopting the same placement
set {3, 51, 123, 134, 149, 152, 158, 160, 208, 212, 213, 215, 219, 231, 246, 254} as
the initial solution, with the aim of improving this pilot locations with re-
spect to pilot symbols. According to Table 1, it should be noted that all
the pilot design methods assign equal power to all pilot subcarriers except
for the proposed scheme. In Fig. 3 and Fig. 4, it is observed that, the pro-
posed scheme leads to a good CE performance in terms of MSE and BER as

19
10 -1

10 -2
BER

X: 25
10 -3 Random, M=16 Y: 0.002
LS, M=52
LS, M=64
PT3, M=16
PT1, M=16
PT2, M=16
Proposed scheme, M=16
10 -4
0 5 10 15 20 25 30 35 40
SNR (dB)

Fig. 4: Comparison of channel estimation performance in terms of BER performance for


different pilot design schemes (N, M, L) = (256, 16, 50).

compared to the other methods under consideration. The high-performance


enhancement in the estimation behavior of the proposed scheme could be
attributed to the joint design of pilot placement and its symbols as a sin-
gle design optimization problem. It is also observed that the proposed pilot
design scheme outperforms the uniformly random pilot pattern and other
different pilot pattern design schemes. Hence, this indicates the appropriate-
ness of the proposed scheme in designing proper pilot pattern. Moreover, at
higher SNR values (i.e good channel conditions) all schemes are observed to
perform very well in probing the channel, since the signal strength in com-
parison to the noise levels is known to be increasingly stronger. Moreover,
it is observed that the performance of the proposed scheme (with M = 16
pilots) reaches the performance of LS which utilizes 64 equally-spaced pilot
signals in the OFDM block. Hence, the proposed scheme by implication can
conserve 64 − 16 = 48 pilot subcarriers under the same CE performance,
leading to a ((64 − M )/N ) × 100 = ((64 − 16)/256) × 100 = 18.75% im-
provement in spectrum efficiency (where M = 16 pilot signals and N = 256
total number OFDM subscribers). In Fig. 4, it is observed that the proposed

20
Table 2: Pilot pattern design scheme by different methods for N = 256 and M = 13
OFDM system settings.
Method µ{A} Runtime (s) Pilot placement
PT4: Method of [27] 0.2687 22.153 1, 2, 4, 8, 13, 21, 31, 45, 66,
115, 140, 171, 218
PT5: Method of [28] 0.2682 20.371 1, 53, 97, 125, 139, 189, 198,
213, 217, 225, 237, 248, 252
PT6: Method of [22] 0.2722 32.617 1, 2, 4, 24, 41, 47, 56, 104,
109, 114, 128, 135, 139
PT7: Method of [9] 0.2713 41.219 6, 39, 78, 83, 84, 88, 104, 126,
139, 154, 157, 237, 243
PT8: Method of [29] 0.2679 24.035 14, 26, 34, 96, 99, 108, 120,
126, 161, 172, 202, 221, 223,
226, 233, 249
Proposed Scheme 0.2671 30.712 11, 40, 68, 102, 111, 120, 132,
168, 179, 189, 204, 243, 252

scheme achieves ≈ 15-dB SNR gain at BER = 0.002 compared to the LS CE


method that utilizes 52 pilots.
The third experiment was performed under (N, M, L) = (256, 13, 50) set-
tings of the OFDM system (i.e, utilizing M = 13 out of N = 256 subcarriers
for pilot signal transmission). The effectiveness of the proposed scheme is
compared with the joint pilot design schemes of the methods of [27] and [28]
that addresses the pilot design problem as a disjoint sub-problem. Another
Method, proposed in [29], which designs the pilot symbols and placement
jointly and uses the discrete stochastic approximation scheme to iteratively
obtain the optimal pilot design is also considered. Additionally, two other
different pilot design methods proposed in [9] and [22], which employs the
equally powered pilot symbols are also included in the investigation. The
mutual coherence of the measurement matrix µ{A}, the algorithm runtimes,
and the designed pilot placements obtained are listed in Table 2. For ease of
notation in the figure, PT4, PT5, PT6, PT7 and PT8 are used to represent
the five different pilot patterns, which corresponds to the Method of [27], [28],
[22], [9] and [29], respectively, as presented in Table 2. It is observed that the
measurement matrix coherence of the proposed scheme is far smaller com-
pared to the other schemes under investigation but with a somewhat increase
in runtime.
Fig. 5 compares the recovery probability for the different pilot design

21
1
PT7
0.9 PT6
PT4
0.8 PT5
PT8
0.7
Recovery probability

Proposed Scheme

0.6

0.5

0.4

0.3

0.2

0.1

0
2 4 6 8 10 12 14 16
Channel sparsity

Fig. 5: Recovery probability for different pilot design scheme (N, M, L) = (256, 13, 50).

schemes in Table 2, with reference to channel sparsity for SNR equal to 20


dB. It is observed that the proposed scheme is more robust and offered the
best channel recovery probability against channel sparsity as compared to
the methods of [9] (PT7), [22] (PT6), [28] (PT5), [27] (PT4) and [29] (PT8).
Fig. 6 and Fig. 7 depict the MSE and BER performance comparisons
of the proposed scheme with the other different pilot optimization schemes
listed in Table 2. A randomly generated pilot pattern {5, 45, 98, 134, 149,
152, 158, 160, 208, 209, 221, 237, 250} which was once generated is also con-
sidered. It is observed that the proposed scheme leads to a good CE perfor-
mance in terms of MSE and BER as compared to the other different pilot
design methods under consideration. The results strongly suggest that the
proper power distribution of the total available power between pilot symbols
and their placements can optimize the transmission over a frequency-selective
fading channel than equally distributing and placing the total power between
the pilot symbols.

22
10 -1
Random, M=13
PT7, M=13
PT6, M=13
PT4, M=13
PT5, M=13
PT8, M=13
Proposed scheme, M=13
MSE

10 -2

10 -3
0 5 10 15 20 25 30 35 40
SNR (dB)

Fig. 6: Comparison of channel estimation performance in terms of MSE performance for


different pilot design schemes (N, M, L) = (256, 13, 50).

5. CONCLUSION
In this paper, the optimal pilot pattern design for sparse CE in OFDM
systems was investigated using CS methods. Hence, a new deterministic
pilot design scheme was proposed for avoiding the disjoint design of pilot
symbols and their placements. The approach is based on minimizing the
mutual coherence of the Fourier submatrix used for the sparse signal re-
covery. The design formulation considers the joint optimization problem of
designing pilot symbol values and their placement as a single pilot design
optimization problem for sparse CE in OFDM systems. To address the com-
binatorial search optimization problem, a local search optimization technique
was employed which starts with an initial population of candidate solution
and iteratively moves to the best neighbor among all neighbors as a solution.
Simulation results show that the proposed scheme is effective and realizes
better performance than former pilot design methods. Additionally, the pro-
posed scheme can attain 18.75% improvement in bandwidth efficiency with
similar CE performance compared with the conventional Least Squares (LS)

23
10 -1
Random, M=13
PT7, M=13
PT6, M=13
PT4, M=13
PT5, M=13
PT8, M=13
Proposed scheme, M=13
BER

10 -2

10 -3
0 5 10 15 20 25 30 35 40
SNR (dB)

Fig. 7: Comparison of channel estimation performance in terms of BER performance for


different pilot design schemes (N, M, L) = (256, 13, 50).

CE that employs the equally-spaced pilot placement.

ACKNOWLEDGEMENT
This work was supported by the Universiti Sains Malaysia under RUI
grant [1001/PELECT/814206].

REFERENCES
[1] E. Kofidis, D. Katselis, A. Rontogiannis, S. Theodoridis, Preamble-
based channel estimation in OFDM/OQAM systems: A review, Sig.
Process. 93(7) (2013) 2038–2054. DOI: 10.1016/j.sigpro.2013.01.013

[2] X. He, R. Song and W. P. Zhu, Pilot Allocation for Distributed-


Compressed-Sensing-Based Sparse Channel Estimation in MIMO-
OFDM Systems, IEEE Trans. Veh. Tech. 65(5)(2016) 2990–3004. DOI:
10.1109/TVT.2015.2441743

24
[3] R. Chopra, D. Ghosh, and D. K. Mehra, Spectrum sensing for
OFDM signals using pilot induced cyclostationarity in the presence
of cyclic frequency offset, Phys. Commun. 24(2017) 182–194. DOI:
10.1016/j.phycom.2017.07.008

[4] R. Prasad, C. R. Murthy and B. D. Rao, Joint channel estimation and


data detection in MIMO-OFDM systems: A sparse Bayesian learn-
ing approach, IEEE Trans. Sig. Proc. 63(20)(2015) 5369–5382. DOI:
10.1109/TSP.2015.2451071

[5] E. Panayirci, H. Senol and M. Uysal, Sparse channel estimation and


equalization for OFDM-based underwater cooperative systems with
amplify-and-forward relaying, IEEE Trans. Sig. Proc. 64(1)(2016) 214–
228. DOI: 10.1109/TSP.2015.2477807

[6] Y. Liu, Z. Tan, H. Hu and L. J Cimini, Channel estimation for


OFDM, IEEE commun. Surv. Tut. 16(4) (2014) 1891–1908. DOI:
10.1109/COMST.2014.2320074

[7] B. C. Jung, J. K. Kwon and H. Jin, Sub-band spreading technique for


adaptive modulation in OFDM systems, Jour. Commun. Network. 10(1)
(2008) 71–78. DOI: 10.1109/JCN.2008.6388330

[8] S. Pejoski and V. Kafedziski, Estimation of sparse time dispersive chan-


nels in pilot aided OFDM using atomic norm., IEEE Wireless Commun.
Lett. 4(4)(2015) 397–400. DOI: 10.1109/LWC.2015.2425410

[9] C. Qi and L. Wu, A study of deterministic pilot allocation for sparse


channel estimation in OFDM systems, IEEE Commun. Lett. 16(5)(2012)
742-744. DOI: 10.1109/LCOMM.2012.032612.112553

[10] J. Lee and S. H. Lee, Low dimensional multiuser detection exploiting


low user activity, Jour. Commun. Network. 15(3) (2013) 283–291. DOI:
10.1109/JCN.2013.000051

[11] Y. C. Eldar and G. Kutyniok, Compressed sensing: theory and applica-


tions, Cambridge University Press, 2012.

[12] W. U. Bajwa and R. Calderbank, Why Gabor frames? Two fundamental


measures of coherence and their role in model selection, J. Commun.
Network. 12(4)(2010)289–307. DOI: 10.1109/JCN.2010.6388466

25
[13] D. Needell and J. A. Tropp, CoSaMP: Iterative signal recovery from
incomplete and inaccurate samples, Applied Comput. Harmonic Anal.
26(3)(2009) 301–321. DOI: 10.1016/j.acha.2008.07.002

[14] Y. Zhang, R. Venkatesan and O. A. Dobre, Novel Compressed Sensing-


Based Channel Estimation Algorithm and Near-Optimal Pilot Place-
ment Scheme, IEEE Trans. Wirel. Commun. 15(4)(2016)2590–2603.
DOI: 10.1109/TWC.2015.2505315

[15] F. Salahdine, N. Kaabouch, and H. El, A survey on compressive sensing


techniques for cognitive radio networks, Phys. Commun. 20 (2016) 61–
73. DOI: 0.1016/j.phycom.2016.05.002

[16] D. L. Donoho, I. Johnstone and A.Montanari, Accurate Prediction of


Phase Transitions in Compressed Sensing via a Connection to Mini-
max Denoising, IEEE Trans. Inf. Theory. 59(6) (2013) 3396–3433. DOI:
10.1109/TIT.2013.2239356

[17] E. J. Candès, J. Romberg, and T. Tao, Robust uncertainty princi-


ples: Exact signal reconstruction from highly incomplete frequency
information, IEEE Trans. Inf. Theory. 52(2)(2006) 489–509. DOI:
10.1109/TIT.2005.862083

[18] A. Kamali, M. R. A. Sahaf and A. M. D. Hosseini, A low complexity


DFT-matrix based pilot allocation algorithm for sparse channel esti-
mation in OFDM systems, AEU-Inter. J. Electronics. and Commun.
68(2)(2014) 85–89. DOI: 10.1016/j.aeue.2013.07.005.

[19] A. N . Uwaechia and N. M. Mahyuddin, A Review on Sparse Channel


Estimation in OFDM System Using Compressed Sensing, IETE Tech.
Rev. 34(5)(2016) 514–531. DOI: 10.1080/02564602.2016.1211966.

[20] C. Qi and L. Wu, Tree-based backward pilot generation for


sparse channel estimation, Electron. Lett. 48(9)(2012)501–503. DOI:
10.1049/el.2012.0010

[21] J. C. Chen, C. K. Wen and P. Ting, An efficient pilot design scheme


for sparse channel estimation in OFDM systems, IEEE Commun. Lett.
17(7)(2013)1352-1355. DOI: 10.1109/LCOMM.2013.051313.122933

26
[22] P. Pakrooh, A. Amini and F. Marvasti, OFDM pilot allocation for sparse
channel estimation, EURASIP J. Adv. Sig. Process. 2012(1)(2012)59.
DOI: 10.1186/1687-6180-2012-59

[23] C. Qi, G. Yue, L. Wu, Y. Huang and A. Nallanathan, Pilot Design


Schemes for Sparse Channel Estimation in OFDM Systems, IEEE Trans.
Veh. Tech. 64(4)(2015) 1493–1505. DOI: 10.1109/TVT.2014.2331085

[24] Ş. Şimşir and N. Tapnar, Pilot tones design using Grey Wolf Opti-
mizer for OFDM IDMA system, Phys. Commun. in press (2017). DOI:
10.1016/j.phycom.2017.08.010

[25] I. Singh, S. Kalyani and K. Giridhar, A Practical Compressed Sensing


Approach for Channel Estimation in OFDM Systems, IEEE Commun.
Lett. 19(12)(2015)2146–2149. DOI: 10.1109/LCOMM.2015.2487265

[26] C. Qi and L. Wu, Joint Design of Pilot Power and Pilot Pat-
tern for Sparse Cognitive Radio Systems, IEEE Trans. Veh. Technol.
64(11)(2015)5384–5390. DOI: 10.1109/TVT.2014.2374692

[27] M. Khosravi and S. Mashhadi, Joint pilot power and pattern de-
sign for compressive OFDM channel estimation, IEEE Commun. Lett.
19(1)(2015)50–53. DOI: 10.1109/LCOMM.2014.2371036

[28] R. Mohammadian, A. Amini and B. H. Khalaj, Compres-


sive Sensing-Based Pilot Design for Sparse Channel Estimation
in OFDM Systems, IEEE Comm. Lett. 21(1)(2017)4–7. DOI:
10.1109/LCOMM.2016.2613086

[29] X. Ren, W. Chen and M. Tao, Position-based compressed channel esti-


mation and pilot design for high-mobility OFDM systems, IEEE Trans.
Veh. Tech. 64(5)(2015)1918–1929. DOI: 10.1109/TVT.2014.2341712

[30] I. Angulo and J. Montalb, A Measurement-Based Multipath Chan-


nel Model for Signal Propagation in Presence of Wind Farms in
the UHF Band, IEEE Trans. Commun. 61(11)(2013)4788–4798. DOI:
10.1109/TCOMM.2013.101113.130144

[31] T. Suryani and G. Hendrantoro, Block-Type Pilot Arrangement with Al-


ternating Polarity for ICI Mitigation in Mobile OFDM Systems, J. Com-
mun. Softw. Syst. 9(3) (2013) 178–183. DOI: 10.24138/jcomss.v9i3.147

27
[32] A. S. Bandeira, E. Dobriban, D. G. Mixon and W. F. Sawin, Certifying
the Restricted Isometry Property is Hard, IEEE Trans. Inf. Theory.
59(6)(2013)3448–3450. DOI: 10.1109/TIT.2013.2248414

[33] S. H. S. Masoumian and B. M. Tazehkand, Greedy Deterministic Pilot


Pattern Algorithms for OFDM Sparse Channel Estimation, Wirel. Pers.
Commun. 84(2) (2015) 1119–1132. DOI: 10.1007/s11277-015-2679-5

[34] M. Elad, Optimized Projections for Compressed Sensing,


IEEE Trans. Signal Process. 55(12) (2007) 5695–5702. DOI:
10.1109/TSP.2007.900760

[35] J. D. Blanchard, M. Cermak and D. Hanle, Greedy Algorithms for Joint


Sparse Recovery, IEEE Trans. Signal Process. 62(7)(2014)1694–1704.
DOI: 10.1109/TSP.2014.2301980

[36] J. A. Tropp and A. C. Gilbert, Signal recovery from random mea-


surements via orthogonal matching pursuit, IEEE Trans. Inf. Theory.
53(12)(2007) 4655–4666. DOI: 10.1109/TIT.2007.909108

[37] S. K. Sahoo and A. Makur, Signal Recovery from Random Measure-


ments via Extended Orthogonal Matching Pursuit, IEEE Trans. Signal
Process. 63(10)(2015) 2572–2581. DOI: 10.1109/TSP.2015.2413384

[38] X. He, R. Song and W. P. Zhu, Optimal pilot pattern design for
compressed sensing-based sparse channel estimation in OFDM sys-
tems, Circuits, Syst. Signal Process. 31(4) (2012) 1379–1395. DOI:
10.1007/s00034-011-9378-6

[39] S. H. S. Masoumian and B. M. azehkand, Accurate OFDM Sparse Chan-


nel Estimation by Optimal Deterministic Pilot Pattern, IETE J. Res.
61(3) (2015) 230–235. DOI: 10.1080/03772063.2015.1009402

[40] Duarte-Carvajalino, J. Martin, and G. Sapiro, Learning to Sense Sparse


Signals: Simultaneous Sensing Matrix and Sparsifying Dictionary Op-
timization, IEEE Trans. Image Process. 18(7)(2009) 1395–1408. DOI:
10.1109/TIP.2009.2022459

[41] A. A. Moghadam and H. Radha, Complex sparse projections for


compressed sensing, 2010 44th Annual Conference on Information

28
Sciences and Systems (CISS), Princeton, NJ, 2010, pp. 1-6. DOI:
10.1109/CISS.2010.5464917

29

Вам также может понравиться