See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/264595377

Compressive Sensing Based Channel Estimation for OFDM Systems Under Long
Delay Channels

Article  in  IEEE Transactions on Broadcasting · June 2014

DOI: 10.1109/TBC.2014.2315913

CITATIONS READS
84 547

5 authors, including:

Wenbo Ding Fang Yang

Georgia Institute of Technology Tsinghua University
84 PUBLICATIONS   806 CITATIONS    137 PUBLICATIONS   1,159 CITATIONS   

SEE PROFILE SEE PROFILE

Linglong Dai Jian Song

Tsinghua University Tsinghua University
198 PUBLICATIONS   5,484 CITATIONS    266 PUBLICATIONS   2,548 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

NOMA - Non-Orthogonal Multiple Access View project

The combination of millimeter-wave and non-orthogonal multiple access (mmWave-NOMA) View project

All content following this page was uploaded by Wenbo Ding on 11 May 2015.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON BROADCASTING, VOL. 60, NO. 2, JUNE 2014 313

Compressive Sensing Based Channel Estimation for

OFDM Systems Under Long Delay Channels
Wenbo Ding, Fang Yang, Senior Member, IEEE, Changyong Pan, Senior Member, IEEE,
Linglong Dai, Member, IEEE, and Jian Song, Senior Member, IEEE

Abstract—Time-domain synchronous orthogonal frequency Union (ITU) in Dec. 2011 [1], DTMB has attracted lots
division multiplexing (TDS-OFDM) has advantages in spectral of interests from both academia and industry. As the core
efficiency and synchronization. However, its iterative interfer- technique of DTMB, time-domain synchronous orthogonal
ence cancellation algorithm will suffer from performance loss
especially under severely fading channels with long delays and frequency division multiplexing (TDS-OFDM) distinguishes
has difficulty supporting high-order modulations like 256 QAM, the standard cyclic prefix OFDM (CP-OFDM) by replacing
which may not accommodate the emerging ultra-high definition CP with the prior known pseudo noise (PN) sequence as
television service. To solve this problem, a channel estimation the guard interval (GI) [2]. The PN sequence can also work
method for OFDM under the framework of compressive sensing as the training sequence (TS) for both synchronization and
(CS) is proposed in this paper. Firstly, by exploiting the signal
structure of recently proposed time-frequency training OFDM channel estimation (CE) at the receiver side, which saves
scheme, the auxiliary channel information is obtained. Secondly, a large amount of frequency-domain pilots commonly used
we propose the auxiliary information based subspace pursuit in CP-OFDM. Consequently, TDS-OFDM usually has higher
(A-SP) algorithm to utilize a very small amount of frequency- spectral efficiency under the same condition [3]. Additionally,
domain pilots embedded in the OFDM block for the exact channel faster and reliable synchronization could be also achieved by
estimation. Moreover, the obtained auxiliary channel information
is adopted to reduce the complexity of the classical SP algorithm. TDS-OFDM [4], [5].
Simulation results demonstrate that the CS-based OFDM out- However, the main drawback of TDS-OFDM is that the
performs the conventional dual pseudo noise padded OFDM and TS and the OFDM block will cause mutual inter-block
CS-based TDS-OFDM schemes in both static and mobile envi- interferences (IBI) to each other. Thus, iterative interfer-
ronments, especially when the channel length is close to or even ence cancellation algorithm with high complexity has to be
larger than the guard interval length, where the conventional
schemes fail to work completely. adopted for CE and channel equalization in TDS-OFDM sys-
tems [6]. Unfortunately, this will result in an open problem
Index Terms—Channel estimation (CE), compressive sensing of TDS-OFDM: Under the severely fading channels, it is
(CS), long delays, orthogonal frequency division multiplexing
(OFDM), ultra-high definition television (UHDTV). difficult for the iterative algorithm to perfectly remove the
IBI when the maximum channel delay spread is large, which
is common in the single frequency network (SFN) environ-
I. I NTRODUCTION ment [7]. This will cause the degradation of the whole system
performance and the difficulty to support the high-order mod-
ITH the official approval of the digital televi-
W sion/terrestrial multimedia broadcasting (DTMB) as
the fourth international digital television terrestrial broadcast-
ulations like 256 QAM [2] to accommodate the emerging
ultra-high definition television (UHDTV) service requirement
(DVB-T2 has claimed to support the 256QAM for UHDTV
ing (DTTB) standard by International Telecommunications
services [8]).
Manuscript received February 3, 2013; revised March 25, 2014; accepted Some alternative solutions have been proposed to solve this
April 3, 2014. Date of publication May 16, 2014; date of current version problem [9]–[12]. One exciting solution is the dual-PN padded
June 4, 2014. This work was supported in part by National Key Basic Research
Program of China under Grant 2013CB329203, in part by China Electric OFDM (DPN-OFDM) scheme, whereby the PN sequence is
Power Research Institute (CEPRI) under Grant 20142000191, and in part duplicated twice to make the second PN sequence immune
by National High Technology Research and Development Program of China from the IBI caused by the preceding OFDM block [9]. Thus,
under Grant 2012AA011704.
W. Ding, F. Yang, C. Pan, and L. Dai are with the Department the second received PN sequence can be directly used for CE,
of Electronic Engineering, and also with the Research Institute of which avoids the iterative interference cancellation algorithm
Information Technology, Tsinghua National Laboratory for Information with high complexity, and improves the performance over
Science and Technology (TNList), Tsinghua University, Beijing 100084,
China (e-mail: [email protected]; [email protected]; severely fading channels as well. However, the spectral effi-
[email protected]; [email protected]). ciency of the DPN-OFDM scheme is significantly decreased
J. Song is with the Department of Electronic Engineering, Tsinghua by the doubled length of the PN sequence. For example, when
National Laboratory for Information Science and Technology (TNList),
Tsinghua University, Beijing 100084, China, and also with the National the length of the single PN sequence is 1/8 that of the OFDM
Engineering Laboratory for DTV, Beijing 100191, China (e-mail: block, the spectral efficiency of TDS-OFDM is 88.89%, while
[email protected]). it is reduced to 80% in DPN-OFDM. To prevent the spec-
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. tral efficiency loss, the compressive sensing (CS) theory is
Digital Object Identifier 10.1109/TBC.2014.2315913 exploited to solve the problem of TDS-OFDM, whereby the
0018-9316 c 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
314 IEEE TRANSACTIONS ON BROADCASTING, VOL. 60, NO. 2, JUNE 2014

IBI-free region of small size within the received PN sequence

is utilized to reconstruct the propagation channel with high
dimension [10]. However, such CS-based TDS-OFDM scheme
cannot work once the maximum channel delay spread is close
to the GI length due to the reduced size of the required obser-
vations. Furthermore, to the best of the authors’ knowledge,
there is no effective solution to improve the CE accuracy
for channels with long delays larger than the GI length in
TDS-OFDM up to now [11], [12].
To enhance the CE performance under severely fading chan-
nels with long delays, we propose the CE method under the
framework of CS. The work is inspired by the recently pro- Fig. 1. System models of the two conventional schemes. (a) DPN-OFDM.
(b) CS-based TDS-OFDM.
posed time-frequency training OFDM (TFT-OFDM) scheme
which is modified from TDS-OFDM [13], [14]. The spe-
cific contributions of this paper can be summarized as
follows: 1) Compared to the conventional TDS-OFDM or CP-
OFDM which merely relies on either the time-domain TS or
frequency-domain pilots for channel recovery [5], the proposed
CE method uses the PN sequence to acquire the coarse chan-
nel path delay estimation, while the exact channel impulse
response (CIR) estimation depends on a very small amount of
frequency-domain pilots embedded in the OFDM block based
on CS. 2) Unlike the conventional scheme of TDS-OFDM
where the IBI caused by the preceding OFDM block to the
current TS has to be cancelled completely to achieve good
performance, it is not necessary to remove such interference
in the proposed scheme since the received PN sequence is
only used for the coarse channel path delay estimation. The
main CE task is transferred to the pilots in the OFDM block,
which could combat the maximum channel delay spread close
to or even larger than the GI length. 3) With the use of CS
and sparse channel nature, the number of pilots embedded in
the OFDM block could be significantly reduced (about 1%
of the total sub-carrier number), and hence high spectral effi-
ciency can still be maintained. 4) Moreover, the coarse channel
Fig. 2. Proposed frame structure and the corresponding CS-based CE for
path delay estimation is used as the auxiliary information to the TFT-OFDM scheme.
reduce the complexity of the classical CS algorithm, making
it applicable for practical systems.
The remainder of this paper is organized as follows. The II. S YSTEM M ODEL
system model is introduced in Section II. The proposed CS-
based CE method is addressed in Section III. Section IV Fig. 1 presents the system model of two conventional
presents the performance analysis of the proposed method. schemes. As shown in Fig. 1(a), the DPN-OFDM scheme uses
The simulation results are demonstrated in Section V. Finally, the second received PN sequence immune from IBI for CE.
the conclusions are provided in Section VI. Thus, the iterative IBI removal can be avoided, but significant
Notation: We use boldface letters to denote matrices and loss in spectral efficiency will be introduced [9]. In Fig. 1(b),
column vectors; FN×N denotes the N-point fast Fourier trans- the CS-based TDS-OFDM scheme utilizes the IBI-free region
form (FFT) matrix √ with the (n + 1, k + 1)th entry being of size G within the received PN sequence for CE, but the max-
exp(−j2π nk N)/ N; 0 denotes the zero vector; ⊗ represents imum estimated channel length is limited to L = M − G + 1,
the circular correlation; (·)T , (·)H , (·)−1 , (·)† , diag(·), and ·p where M is the length of the PN sequence [10]. When the
denote the transpose, conjugate transpose, matrix inversion, channel length L becomes larger, the size of the required
Moore-Penrose matrix inversion, diagonal matrix, and lp norm IBI-free region becomes smaller, which leads to severe perfor-
operations, respectively; xS is generated by restricting the vec- mance deterioration due to the reduced number of observations
tor x to its S largest components, and sup (xS ) denotes the in CS.
location set of the S largest components of x; x| denotes the Unlike the conventional TDS-OFDM or CP-OFDM where
entries of the vector x in the set of ;  represents the sub- the training information only exists in either the time or fre-
matrix comprising the  columns of , and  represents the quency domain [15], [16], Fig. 2 shows that TFT-OFDM has
sub-matrix comprising the  rows of ; Finally, E(·) stands training information in both the time and frequency domains
for the expectation operator. for every signal symbol, i.e., the time-domain TS and the
DING et al.: COMPRESSIVE SENSING BASED CHANNEL ESTIMATION FOR OFDM SYSTEMS UNDER LONG DELAY CHANNELS 315

frequency-domain pilots scattered over the signal bandwidth where  is the Toeplitz matrix of size N × N determined by
are jointly used in  TFT-OFDM. TheT ith TFT-OFDM sig- the transmitted OFDM block xi , hiN is the N-length CIR vec-
nal symbol si = si0 , si1 , · · · , siM+N−1 is composed of the tor extended from the original L-length CIR vector hi with
 T N − L zeros, and wi = [ω0i , ω1i , . . . , ωN−1
i ]T denotes the addi-
known PN sequence c = c0 , c1 , · · · , cM−1 of length M
 T tive white Gaussian noise (AWGN) with zero mean and the
and the OFDM block xi = x0i , x1i , · · · , xN−1 i of length N,
variance σ 2 .
i.e.,
    The frequency-domain OFDM block ỹi =
c c  i i T
s = i
i
= i , (1) ỹ0 , ỹ1 , · · · , ỹN−1
i after cyclicity reconstruction can be
x (M+N)×1 FH
N×N x̃ (M+N)×1 represented as
 T
where x̃i = x̃0i , x̃1i , · · · , x̃N−1
i is the ith OFDM block in ỹi = FN×N yi = FN×N FH N×N FN×N hN + w̃
i i
the frequency domain. In contrast to the conventional TDS-  i 
h
OFDM, the OFDM block in TFT-OFDM contains not only = diag(x̃i )FN×N + w̃i , (8)
0N−L
the traffic data, but also a small number J of pilots denoted
as x̃i | , where  is the pilot location set and represented as = diag(x̃i )F(L) hi + w̃i

 = {P0 , P1 , · · · , PJ−1 } , (2) where FN×N FH N×N = diag(x̃ ), F(L) is the matrix comprising
i

the first L columns of the FN×N , and w̃i = FN×N wi denotes the
where 0 ≤ P0 < P1 < · · · < PJ−1 ≤ N − 1 can be assumed AWGN in the frequency domain. Then the received frequency-
without loss of generality. By the use of CS and sparse chan- domain pilots ỹi | within yi can be represented by
nel nature, the pilot number J could be reduced significantly
(far less than L, around 1% of N) and hence the spectral effi- ỹi | = diag(x̃i | )F(L) hi + w̃i | , (9)
ciency loss is negligible. The power of the pilots could be
boosted for better CE performance, which is similar to that where F(L) represents the sub-matrix comprising the rows of
in the traditional pilot-aided CE method [17]. The difference the F(L) on the set .
is that the pilots are equally spaced in conventional schemes, In the proposed scheme, the time-domain PN sequence and
while the pilots are randomly located in the proposed scheme the frequency-domain pilots are jointly exploited to perform
to ensure good CE performance based on CS. The reason will the CE, which will be addressed in detail in the next section.
be discussed later in Section IV.
In wireless communications, the discrete-time CIR hi = III. C OMPRESSIVE S ENSING BASED C HANNEL
[h0 , hi1 , . . . , hiL−1 ]T of length L comprising S resolvable prop-
i E STIMATION
agation paths can be modeled as [18] For (9), CS theory has proved that the target signal hi of

S−1 large dimension L can be exactly recovered by a very small
hil = αs δ [l − τs ], 0 ≤ l ≤ L − 1, (3) number of observations J if hi is sparse, i.e., the number
s=0 of nonzero entries of the target signal is much smaller than
where αs and τs denote the path gain and the normalized path its dimension [19], [20]. Fortunately, numerous theoretical
delay of the sth path, respectively. The path delay set T is analysis and experimental results have verified that wireless
defined as channels are sparse in nature, i.e., in the CIR model (3),
the dimension of the CIR L may be large, but the number
T = {τ0 , τ1 , · · · , τS−1 } , (4) of the active paths S with significant gains is usually small,
where 0 ≤ τ0 < τ1 < · · · < τS−1 ≤ L − 1 can be assumed i.e., S  L, especially in the wireless wideband communica-
without loss of generality. tions [1], [21]. This indicates that the sparse channel can be
If the channel is exactly known at the receiver, the IBI of the recovered by a very small amount of frequency-domain pilots
PN sequence on the OFDM block can be completely removed. based on CS.
Then, by using the idea of the classical overlap and add (OLA) Many efficient signal recovery algorithms have been devel-
algorithm [16], the cyclicity reconstruction of the OFDM block oped for CS. Among them, the subspace pursuit (SP) algorithm
can be obtained and hence the CP effect can be restored [6]. is a widely used CS algorithm due to its robustness to noise,
This process will be described in Section III-B in detail. where the most significant S components of the original sparse
The time-domain OFDM signal are identified in an iterative manner [22]. However,
T block after cyclicity reconstruction
yi = yi0 , yi1 , · · · , yiN−1 can be represented as it requires the priori knowledge on the sparsity level of the
signal, and has relatively high complexity.
yi = hiN + wi , (5) In this section, by fully exploiting the joint time-frequency
⎡ i ⎤
i
x0 xN−1 ··· x1i signal feature of TFT-OFDM, we propose a CS-based CE
⎢ x i i ··· x2i ⎥ method with the auxiliary information based SP (A-SP) algo-
⎢ 1 x0 ⎥
=⎢ . .. .. .. ⎥ , (6) rithm. Unlike the conventional TS-based or pilot-aided CE
⎣ .. . . . ⎦ schemes which depend on either time- or frequency-domain
i
xN−1 i
xN−2 · · · x0i N×N information [6], [17], the proposed CS-based CE method


hi firstly utilizes the PN-based correlation in the time domain
hiN = , (7)
0N−L to acquire the auxiliary channel information, and then the
316 IEEE TRANSACTIONS ON BROADCASTING, VOL. 60, NO. 2, JUNE 2014

Fig. 4. Illustration for the OLA-based OFDM block cyclicity reconstruction.

practical CS-based signal recovery algorithm with improved

performance as well as reduced complexity in Step 3.

B. Step 2: Cyclicity Reconstruction of the OFDM Block

Fig. 3. Coarse path delay estimation for auxiliary channel information
acquisition under ITU-VB channel with SNR of 10 dB. The cyclicity reconstruction of the OFDM block is achieved
by firstly subtracting the IBI caused by the PN sequence
from the received OFDM block, then adding the received PN
frequency-domain pilots are used for the final exact CIR esti- sequence and finally subtracting the first part of linear convo-
mation based on CS. The specific procedure of this method lution outputs between the PN sequence and channel [6]. This
can be divided into three steps: 1) PN-based coarse path delay process is based on the idea of OLA algorithm [16] and is
estimation; 2) Cyclicity reconstruction of the OFDM block; briefly illustrated in Fig. 4.
3) Exact CIR estimation using A-SP. In this step, the IBI caused by the PN sequence is obtained
by computing the linear convolution between the local PN
A. Step 1: PN-Based Coarse Path Delay Estimation sequence and the estimated CIR obtained in the preceding
symbol. Under the slow time-varying channels, which can be
Based on the good auto-correlation property of the PN
assumed in many wireless broadcasting systems [26], the esti-
sequence [23], the received PN sequence di is directly cor-
mated CIR obtained in the preceding symbol can be used for
related with the locally known PN sequence c to acquire the
i the IBI removal in the current symbol. In fact, the received
coarse channel estimate h̄ as
PN sequence contains not only the useful part which is the IBI
i 1 caused by the OFDM block, but also the useless part that is
h̄ = c ⊗ di = hi + v, (10)
M the linear convolution between the PN sequence and channel.
where v denotes the AWGN as well as the effect of inter- Hence, the useless part should be removed after the received
ference caused by the preceding OFDM block. As shown PN sequence is added to achieve the cyclicity reconstruction.
in Fig. 3, where the ITU Vehicular B (ITU-VB) channel
model [24] with the signal-to-noise ratio (SNR) of 10 dB is C. Step 3: Exact CIR Estimation Using A-SP
considered, although the coarse path delay estimation is not After Step 2, the pilots can be extracted from the OFDM
accurate due to the existence of IBI, the good auto-correlation block after cyclicity reconstruction for the final accurate CE.
property of the PN sequence ensures that the auxiliary chan- Based on the basic idea of classical SP algorithm [22], we
nel information necessary for the following A-SP algorithm propose the A-SP algorithm, whereby the auxiliary channel
could be obtained. Such information includes the locations of information obtained in Step 1 is exploited to improve the CE
the most significant taps and the approximate channel sparsity performance and lower the computational complexity. The pro-
level. posed A-SP algorithm is described by pseudo code as shown
i
The path gains in h̄ are discarded directly, and only the in Algorithm 1.
path delays of the most significant taps (partial path delays) Compared to the classical SP algorithm, the proposed A-SP
are retained in the initial coarse path delay set algorithm has quite similar procedure, but differs from SP in
   L−1
T0 = l:h̄il 2 ≥ pth l=0 , (11) the following three aspects:
1) Initial Configuration. In the SP algorithm, the ini-
where pth is the power threshold configured according to [25]. tial approximation of the support T0 is set to the S indices
Then, the channel sparsity level S can be approximated by corresponding to the largest magnitude entries in the vector
S = S0 + a = T0 0 + a, (12) H r because no prior knowledge of the signal is avail-
able. However, in the proposed A-SP scheme, by exploiting
where S0 = T0 0 denotes the initial channel sparsity level, the obtained partial path delays T0 in Step 1, the initial
and a is a number used to combat the interference effect, approximation can be directly configured as T0 ← T0 .
since some low-power active paths may be treated as noise 2) Significant Entry Identification. Unlike SP where the S
i
in h̄ . This auxiliary channel information is essential for most significant entries of the vector H r are identified in each
DING et al.: COMPRESSIVE SENSING BASED CHANNEL ESTIMATION FOR OFDM SYSTEMS UNDER LONG DELAY CHANNELS 317

TABLE I
Algorithm 1 Auxiliary information based SP (A-SP) S PECTRAL E FFICIENCY C OMPARISON
Inputs:
1) Initial coarse path delay set T0 , channel sparsity level
S, initial channel sparsity level S0 ;

2) Noisy measurements m = ỹi | ;

3) Observation matrix  = diag(x̃i | )F(L) .

i
Output: The S-sparse CIR estimation ĥ .
1: T0 ← T0 ;
†
2: r ← m − T0  0 m;
T
3: k ← 0; algorithm
  only
 requires a very small number of observations
4: while k < S − S0 do O Slog2 L S [22]. In fact, for the most common broadcast-
5: k ← k + 1; ing channels with six active paths [1], the number J = 36 of
p ← H r; the embedded pilots is enough
  for
 good channel recovery per-
6:
 
7: T̃k ← T̃k−1 ∪ sup(pS−S0 ); formance (note that Slog2 L S = 6 × log2 256 6 ≈ 32).
8: q ← † k m; In the typical system configuration of M = 256 and N =
T̃
9: Tk ← sup(qS ); 4096 [10], the spectral efficiency of the proposed CS-based
10: r ← y − Tk †Tk m; TFT-OFDM scheme is 93.29%, which is only 0.83% less than
11: end while that of the conventional TDS-OFDM system.
12: Actual path delay acquisition: T ← Tk ; Moreover, the constant 36-symbol transmission parame-
13: Then the signal model (9) can be simplified as ter signaling (TPS) used in the practical DTMB systems
can be regarded as the pilots once after being successfully
detected [28], which indicates that even the negligible spectral
ỹi | = T hi |T +wi | . (13) efficiency loss can be avoided.

14: Final least squares (LS): B. CRLB of the Channel Estimator

 −1 H i According to the signal model (13) in which  the AWGN

i
ĥ |T =†T ỹi | = H
T T T ỹ | . (14) vector w̃| is subject to the distribution CN 0, σ 2 IJ , where
IJ is the identity matrix of size J×J, the conditional probability
density function (PDF) of ỹi | with the given hi |T is
  
iteration, we remain the S0 most significant entries unchanged,


 1 1   i i 
p ỹi |  hi |T =  J/2 exp − 2 ỹ | − T h |T 2 ,
and identify the next S − S0 most significant ones instead. 2π σ 2 /A 2σ /A
3) Iteration Number. The required number of iterations is (15)
reduced from S in SP to S − S0 in A-SP, so the computational
complexity can be reduced. where A denotes the average boosted power of the pilots.
Then, we can derive the CRLB of the unbiased estima-
IV. P ERFORMANCE A NALYSIS OF CS-BASED TFT-OFDM tor (14) by using the vector estimation theory [29], which is
given as
This section addresses the performance analysis of the
 i   Sσ 2
proposed scheme in terms of the spectral efficiency, the  
Cramér-Rao lower bound (CRLB) of the proposed CE method, CRLB = E ĥ |T − hi |T  = . (16)
2 JA
and the computational complexity as well.
Compared with the channel estimator in DPN-OFDM sys-
tems, where the best mean square error (MSE) performance
A. Spectral Efficiency is σ 2 , the channel estimator based on A-SP can achieve much
Table I compares the spectral efficiency of different TDS- better MSE performance, since S is much smaller than J, and
OFDM schemes with the ideal OFDM system without any the boosted power A is usually larger than 1.
overhead [27]. The length of the PN sequence is M = 256 Note that if the observation matrix T does not have
and the pilot number embedded in TFT-OFDM is J = 36. orthogonal columns, the CRLB (16) cannot be achieved by
The conventional TDS-OFDM scheme has the highest spec- the practical channel estimator. However, due to the random
tral efficiency, but the iterative interference cancellation is positions of the pilots used in TFT-OFDM and the random
required, which results in high complexity and performance locations of active paths of wireless channels, the matrix
loss. In the DPN-OFDM scheme, the PN sequence is dupli- T has imperfect but approximate orthogonal columns (the
cated to avoid the interference from the preceding OFDM requirement of near orthogonality is equivalently similar to the
block to the second PN sequence, but the spectral efficiency is restricted isometry property (RIP) of the observation matrix in
significantly reduced. In the proposed scheme, the embedded the CS theory [30]). Thus, the CRLB can be approximately
pilots would somehow reduce the spectral efficiency, but the approached, which will be validated by the simulation results
penalty is very small, since the proposed A-SP signal recovery in Section V.
318 IEEE TRANSACTIONS ON BROADCASTING, VOL. 60, NO. 2, JUNE 2014

TABLE II
M ULTIPATH C HANNEL PARAMETERS FOR S IMULATION

C. Computational Complexity
In the proposed CS-based CE method, the M-point cir- Fig. 5. MSE performance comparison under the ITU-VB channel with the
cular correlation in Step 1 could be efficiently imple- channel length smaller than the GI length.
mented by M-point FFT, so the corresponding complexity
is O Mlog2 (M) /2 . In Step 2, it requires the complexity
of O Mlog2 (M) /4 + 3M for the cyclicity reconstruction
operation.
In fact, the main computational burden of the proposed
method is the A-SP algorithm used to acquire the actual
path
 delays in
 Step 3. Each iteration has the complexity of
O J L + S2 , and  the overall
 complexity of SP comprising
S iterations is O JS L + S2 [22]. However, as has been dis-
cussed in Section III-C, only S − S0 iterations are required by
A-SP, since some of the locations of the significant taps have
been detected already,
 the complexity
  of the proposed A-SP
is reduced to O J (S − S0 ) L + S2 . So the complexity of
A-SP is lower than that of SP.

V. S IMULATION R ESULTS
This section investigates the performance of the CS-based
CE for TFT-OFDM under typical broadcasting channels. The
signal bandwidth is 7.56 MHz locating at a central fre- Fig. 6. MSE performance comparison under the SARFT-8 channel with the
channel length close to the GI length.
quency of 760 MHz. The OFDM block length N is 4096,
and the GI length M is 256. The low-density parity-check
(LDPC) code with code rate of 0.6 and code length of
7488 in DTMB is adopted [1]. The well-known iterative scheme, the last G = 36 samples of the IBI-free region are
decoding algorithm called belief propagation (BP) is used used for CE, while the number of frequency-domain pilots
with the maximum iteration number of 30 [31]. The mod- J = 36 (around 1% of the total sub-carrier number N = 4096)
ulation schemes 256 QAM for the static channel and 16 with the boosted power A = 2 is utilized in the proposed
QAM with a receiver velocity of 60 km/h are both consid- scheme.
ered to evaluate the support for UHDTV and mobile services, Figs. 5–7 present the MSE performance comparison of the
respectively. proposed scheme with the conventional DPN-OFDM and CS-
The multipath channel parameters used for simulations are based TDS-OFDM schemes under three different channels
listed in Table II. The typical six-tap ITU-VB channel model with different channel lengths. It can be seen from Fig. 5 that
(S = 6, L = 152) is adopted for performance evaluation. under the ITU-VB channel, the MSE performance of the pro-
Furthermore, the State Administration of Radio, Film, and posed scheme enjoys a significant SNR gain of 4 dB and 10 dB
Television 8 (SARFT-8) channel model (S = 6, L = 241) with compared to those of CS-based TDS-OFDM and DPN-OFDM,
a very strong echo path close to the GI length [1] and the mod- respectively, when the target MSE of 10−3 is considered. If
ified ITU-VB channel (S = 7, L = 303) with an extremely the channel length L is fairly close to the GI length M, see
long path delay exceeding the GI length are also adopted the SARFT-8 channel considered in Fig. 6, the MSE perfor-
to evaluate the system performance under severely fading mance of the proposed scheme is 7.5 dB better than that of
channel with long delays which may exist for broadcasting DPN-OFDM, while the recent CS-based TDS-OFDM cannot
systems, especially in SFN. In the CS-based TDS-OFDM work due to the reduced size of the IBI-free region.
DING et al.: COMPRESSIVE SENSING BASED CHANNEL ESTIMATION FOR OFDM SYSTEMS UNDER LONG DELAY CHANNELS 319

Fig. 7. MSE performance comparison under the modified ITU-VB channel Fig. 9. BER performance comparison under the ITU-VB channel with the
with the channel length larger than the GI length. channel length smaller than the GI length.

Fig. 8. Snapshot of the channel estimate of the proposed CS-based CE Fig. 10. BER performance comparison under the SARFT-8 channel with the
method under the modified ITU-VB channel with a long path delay L = 303 channel length close to the GI length.
when SNR is 10 dB.

In the extreme case that the actual channel length L is larger Figs. 9–11 compare the LDPC-coded bit error rate (BER)
than the GI length M (i.e., L > M), the conventional DPN- performance of the proposed scheme with the conventional
OFDM and CS-based TDS-OFDM schemes will fail to work DPN-OFDM and CS-based TDS-OFDM schemes under the
completely due to the limitation of the PN length, while the three considered channels with different channel lengths. The
proposed scheme can still have good MSE performance, as BER performance with the ideal CE is also presented as the
shown in Fig. 7. This can be explained by Eq. (9) that the benchmark for comparison. Since the schemes have different
maximum CE length of the proposed scheme is only limited spectral efficiencies, for fair comparison, the energy per bit
by the column number of F(L) , which could be N at most, to noise power spectral density ratio (Eb /N0 ) is more suitable
i.e., L ≤ N. In this way, the maximum channel delay spread for the performance evaluation than the traditional SNR. As
of the proposed CE method is extended from M to N, as long shown in Fig. 9, under the ITU-VB channel, the BER per-
as the cyclicity reconstruction of the OFDM block is near- formance of the proposed scheme enjoys a significant Eb /N0
perfect. Fig. 8 shows a snapshot of the channel estimate of gain of about 1 dB and 2.5 dB compared with CS-based TDS-
the proposed CS-based CE method under the modified ITU- OFDM and DPN-OFDM, respectively, when the target BER
VB channel with a long path delay L = 303 (larger than the of 10−4 is considered. It can be observed from Fig. 10 that
GI length M = 256) when SNR is 10 dB. It is clear that the under the SARFT-8 channel, the BER performance of the pro-
proposed scheme can acquire very exact CIR estimation com- posed scheme is about 2 dB better than that of the DPN-OFDM
pared with the actual channel, including both the path delays scheme, while the CS-based TDS-OFDM scheme cannot work
and path gains. due to the IBI-free region is severely contaminated by the long
320 IEEE TRANSACTIONS ON BROADCASTING, VOL. 60, NO. 2, JUNE 2014

[7] K. Yan, F. Yang, C. Pan, and J. Song, “Reception quality prediction

in a single frequency network for the DTMB standard,” IEEE Trans.
Broadcast., vol. 58, no. 4, pp. 629–636, Dec. 2012.
[8] L. Vangelista et al., “Key technologies for next-generation terrestrial dig-
ital television standard DVB-T2,” IEEE Commun. Mag., vol. 47, no. 10,
pp. 146–153, Oct. 2009.
[9] J. Fu, J. Wang, J. Song, C. Pan, and Z. Yang, “A simplified equaliza-
tion method for dual PN-sequence padding TDS-OFDM systems,” IEEE
Trans. Broadcast., vol. 54, no. 4, pp. 825–830, Dec. 2008.
[10] L. Dai, Z. Wang, and Z. Yang, “Compressive sensing based time domain
synchronous OFDM transmission for vehicular communications,” IEEE
J. Sel. Areas Commun., vol. 31, no. 9, pp. 460–469, Sep. 2013.
[11] S. Li, J. Xiong, L. Gui, and Y. Xu, “A generalized analytical solution
to channel estimation with intersymbol interference cancelation and co-
channel interference cancelation for single input single output/multiple
input single output digital terrestrial multimedia broadcasting systems,”
IEEE Trans. Broadcast., vol. 59, no. 1, pp. 116–128, Mar. 2013.
[12] J. Xiong, L. Gui, H. Liu, and P. Cheng, “On channel estimation and
equalization in 2x1 MISO TDS-OFDM based terrestrial DTV systems,”
IEEE Trans. Broadcast., vol. 58, no. 1, pp. 130–138, Mar. 2012.
[13] L. Dai, Z. Wang, and Z. Yang, “Time-frequency training OFDM with
high spectral efficiency and reliable performance in high speed envi-
Fig. 11. BER performance comparison under the modified ITU-VB channel ronments,” IEEE J. Sel. Areas Commun., vol. 30, no. 4, pp. 695–707,
with the channel length larger than the GI length. May 2012.
[14] C. Pan, L. Dai, and Z. Yang, “Unified time-frequency OFDM trans-
mission with self interference cancellation,” IEICE Trans. Fund., vol.
channel. Fig. 11 shows that when the actual channel length is E96-A, no. 4, pp. 807–813, Apr. 2013.
[15] J. Wang et al., “A general SFN structure with transmit diversity for TDS-
larger than the GI length, the two conventional schemes fail to OFDM system,” IEEE Trans. Broadcast., vol. 52, no. 2, pp. 245–251,
work, while the proposed scheme still performs well. In addi- Jun. 2006.
tion, the actual BER curve is very close to that of the ideal [16] B. Muquet, Z. Wang, G. Giannakis, M. D. Courville, and P. Duhamel,
“Cyclic prefixing or zero padding for wireless multicarrier trans-
case, which indicates the good performance of the proposed missions?” IEEE Trans. Commun., vol. 50, no. 12, pp. 2136–2148,
scheme. Dec. 2002.
[17] Z. Tang, R. Cannizzaro, G. Leus, and P. Banelli, “Pilot-assisted time-
varying channel estimation for OFDM systems,” IEEE Trans. Signal
VI. C ONCLUSION Process., vol. 55, no. 5, pp. 2226–2238, May 2007.
We have proposed a CS-based CE method for TFT-OFDM [18] B. Yang, K. Letaief, R. Cheng, and Z. Cao, “Channel estimation for
OFDM transmission in multipath fading channels based on parametric
in this paper. The MSE performance of this method outper- channel modeling,” IEEE Trans. Commun., vol. 49, no. 3, pp. 467–479,
forms the conventional schemes and is close to the CRLB by Mar. 2001.
simultaneously exploiting the time-domain PN sequence and [19] E. Candès and M. Wakin, “An introduction to compressive sampling,”
IEEE Signal Process. Mag., vol. 25, no. 2, pp. 21–30, Feb. 2008.
frequency-domain pilots. Simulation results show that the pro- [20] W. Bajwa, J. Haupt, A. Sayeed, and R. Nowak, “Compressed channel
posed scheme has a good BER performance in both static and sensing: A new approach to estimating sparse multipath channels,” Proc.
mobile scenarios and can well support the 256 QAM, espe- IEEE, vol. 98, no. 6, pp. 1058–1076, Jun. 2010.
[21] C. Berger, Z. Wang, J. Huang, and S. Zhou, “Application of compressive
cially when the maximum channel delay spread is fairly close sensing to sparse channel estimation,” IEEE Commun. Mag., vol. 48,
to or even larger than the GI length. Besides, by using the aux- no. 11, pp. 164–174, Nov. 2010.
iliary channel information, the proposed A-SP algorithm has [22] W. Dai and O. Milenkovic, “Subspace pursuit for compressive sens-
ing signal reconstruction,” IEEE Trans. Inf. Theory, vol. 55, no. 5,
lower complexity than the conventional SP algorithm. Thus, pp. 2230–2249, May 2009.
this scheme is expected to extend TDS-OFDM in the emerg- [23] B. Ai, J. Ge, and Y. Wang, “Symbol synchronization technique in
ing UHDTV applications under SFN with long channel delay COFDM systems,” IEEE Trans. Broadcast., vol. 50, no. 1, pp. 56–62,
Mar. 2004.
spread. Furthermore, for CP-OFDM system with time-domain [24] Guideline for Evaluation of Radio Transmission Technology for IMT-
preamble or TS, this scheme can also be applied. 2000, ITU-R M. 1225 Standard, Dec. 1997.
[25] F. Wan, W. Zhu, and M. Swamy, “Semi-blind most significant tap detec-
R EFERENCES tion for sparse channel estimation of OFDM systems,” IEEE Trans.
Circuits Syst. I, Reg. Papers, vol. 57, no. 3, pp. 703–713, Mar. 2010.
[1] Error-Correction, Data Framing, Modulation and Emission Methods for [26] X. Zhou, F. Yang, and J. Song, “Novel transmit diversity scheme for
Digital Terrestrial Television Broadcasting, ITU-R BT. 1306-6 Standard, TDS-OFDM system with frequency-shift m-sequence padding,” IEEE
Dec. 2011. Trans. Broadcast., vol. 58, no. 2, pp. 317–324, Jun. 2012.
[2] J. Song et al., “Technical review on Chinese digital terrestrial televi- [27] L. Dai, Z. Wang, and S. Chen, “A novel uplink multiple access scheme
sion broadcasting standard and measurements on some working modes,” based on TDS-FDMA,” IEEE Trans. Wireless Commun., vol. 10, no. 3,
IEEE Trans. Broadcast., vol. 53, no. 1, pp. 1–7, Feb. 2007. pp. 757–761, Mar. 2011.
[3] L. Dai, Z. Wang, and Z. Yang, “Next-generation digital television terres- [28] L. Dai, Z. Wang, C. Pan, and S. Chen, “Wireless positioning using TDS-
trial broadcasting systems: Key technologies and research trends,” IEEE OFDM signals in single-frequency networks,” IEEE Trans. Broadcast.,
Commun. Mag., vol. 50, no. 6, pp. 150–158, Jun. 2012. vol. 58, no. 2, pp. 236–246, Jun. 2012.
[4] B. Ai et al., “On the synchronization techniques for wireless OFDM sys- [29] S. M. Kay, Fundamentals of Statistical Signal Processing, Volumn I:
tems,” IEEE Trans. Broadcast., vol. 52, no. 2, pp. 236–244, May 2006. Estimation Theory. Upper Saddle River, NJ, USA: Prentice-Hall, 1993.
[5] L. He, F. Yang, C. Zhang, and Z. Wang, “Synchronization for TDS- [30] J. Haupt, W. Bajwa, G. Raz, and R. Nowak, “Toeplitz compressed sens-
OFDM over multipath fading channels,” IEEE Trans. Consum. Electron., ing matrices with applications to sparse channel estimation,” IEEE Trans.
vol. 56, no. 4, pp. 2141–2147, Nov. 2010. Inf. Theory, vol. 56, no. 11, pp. 5862–5875, Nov. 2010.
[6] J. Wang, Z. Yang, C. Pan, and J. Song, “Iterative padding subtraction of [31] J. Chen, A. Dholakia, E. Eleftheriou, M. Fossorier, and X. Hu,
the PN sequence for the TDS-OFDM over broadcast channels,” IEEE “Reduced-complexity decoding of LDPC codes,” IEEE Trans. Commun.,
Trans. Consum. Electron., vol. 51, no. 11, pp. 1148–1152, Nov. 2005. vol. 53, no. 8, pp. 1288–1299, Aug. 2005.
DING et al.: COMPRESSIVE SENSING BASED CHANNEL ESTIMATION FOR OFDM SYSTEMS UNDER LONG DELAY CHANNELS 321

Wenbo Ding received the B.S.E degree (with the Linglong Dai received the B.S. degree from
highest honor) from the Department of Electronic Zhejiang University, Zhejiang, China, in 2003, the
Engineering, Tsinghua University, Beijing, China, in M.S. degree (with the highest honors) from the
2011. He is currently pursing the Ph.D. degree with China Academy of Telecommunications Technology,
the DTV Technology Research and Development in 2006, and the Ph.D. degree (with the highest
Center, Tsinghua University. His current research honors) from Tsinghua University, Beijing, China,
interests are in the field of channel estimation and in 2011. From 2011 to 2013, he was a Post-
equalization for multimedia communication systems Doctoral Fellow at the Department of Electronic
as well as the power line communication and visible Engineering, Tsinghua University, and then in July
light communication. 2013, he became an Assistant Professor with the
same Department. His current research interests
include wireless communications with the emphasis on OFDM, MIMO, syn-
chronization, channel estimation, multiple access techniques, and wireless
positioning. He has published over 30 journal and conference papers. He
received the IEEE ICC Best Paper Award in 2013, the Outstanding Post-
Fang Yang received the B.S.E. and Ph.D. degrees Doctoral Fellow of Tsinghua University in 2013, the China Post-Doctoral
from the Department of Electronic Engineering, Science Special Foundation in 2012, the Excellent Doctoral Dissertation of
Tsinghua University, Beijing, China, in 2005 and Beijing in 2012, the Outstanding Ph.D. Graduate of Tsinghua University in
2009, respectively. Currently, he is as an Associate 2011, and the Academic Star of Tsinghua University in 2011.
Professor with the DTV Technology Research and
Development Center, Tsinghua University. His cur-
rent research interests are in the field of channel
estimation and interference cancellation for digital
wireless communication system, space-time coding, Jian Song received the B.Eng. and Ph.D. degrees
and diversity techniques, as well as the training in electrical engineering both from Tsinghua
sequence design. University, Beijing, China, in 1990 and 1995, respec-
tively, and was with the same university upon
his graduation. He has worked at the Chinese
University of Hong Kong, Hong Kong, China, and
the University of Waterloo, Waterloo, ON, Canada,
Changyong Pan is a Full Professor with the in 1996 and 1997, respectively. He has been a
Department of Electronic Engineering and Deputy Professor with the Hughes Network Systems, USA,
Director of DTV Research and Development Center for 7 years before joining with the faculty team in
of Tsinghua University, Beijing, China. He is Tsinghua in 2005. Currently, he is the Director with
actively involved in Chinese DTTB standard pro- Tsinghua’s DTV Technology Research and Development Center. He has been
cess as one of the major technical contributors, working in different areas of fiber-optic, satellite and wireless communica-
including algorithm development, system design as tions, as well as power line communications. His current research interests
well as hardware implementation, standard drafting, include the area of digital TV broadcasting. He has published over 110 peer-
especially in charge of all field trials and related reviewed journal and conference papers. He holds two U.S. patents and over
system-level research and development. He has been 20 Chinese patents. He is a fellow of IET.
a Member of Chinese ITU-R delegation group for
the Chinese DTTB standard, representing China to present in Geneva in DTV
area several times and is now the contact person for DTTB-related issues.
He is the key working group member for the serial national equipment stan-
dards related to the DTTB. He has authored or co-authored over 180 technical
papers and published five technical books. He holds 34 patents, and has won
2nd National Technical Invention Award twice, and is also the winner of
numerous other awards.

View publication stats