58 IEEE COMMUNICATIONS LETTERS, VOL. 19, NO.

1, JANUARY 2015

TimeFrequency Joint Sparse Channel Estimation for MIMO-OFDM Systems

Wenbo Ding, Fang Yang, Senior Member, IEEE, Wei Dai, Member, IEEE, and Jian Song, Senior Member, IEEE

AbstractThis letter proposes a time-frequency joint sparse nificantly increase as the number of transmit antennas becomes
channel estimation for multiple-input multiple-output orthogonal large, which decreases the spectral efficiency. Recently, some
frequency division multiplexing (MIMO-OFDM) systems under greedy compressive sensing (CS) methods [5], [6] have been in-
the framework of structured compressive sensing (CS). The pro-
posed scheme first relies on a pseudorandom preamble, which is troduced for channel estimation in MIMO systems, which could
identical for all transmit antennas, to acquire the partial common improve the spectral efficiency but requires the knowledge of
support by utilizing the sparse common support property of the the accurate channel sparsity in prior [7].
MIMO channels. Then, a very small amount of frequency-domain In this letter, we propose a spectrum-efficient OFDM scheme,
orthogonal pilots are used for the accurate channel recovery. Sim- namely time-frequency training OFDM (TFT-OFDM), for
ulation results show that the proposed scheme demonstrates better
sparse MIMO systems based on a time-frequency joint channel
performance and higher spectral efficiency than the conventional
MIMO-OFDM schemes. Moreover, the obtained partial common estimation method under the framework of structured CS (SCS)
support can be further utilized to reduce the complexity of the CS [8]. First, by exploiting the sparse common support property
algorithm and improve the signal recovery probability under low of the MIMO wireless channels [9], the time-domain preamble
signal-to-noise-ratio conditions. is used to acquire the partial common support and sparsity of
Index TermsMIMO-OFDM, structured compressive sensing, the channels, while the exact channel recovery will depend
sparse common support, time-frequency joint channel estimation, on a small number of frequency-domain pilots by using SCS.
spectrum-efficient. Second, the overhead of both preamble and pilots in the pro-
posed scheme is far less than the conventional MIMO-OFDM
I. I NTRODUCTION
systems [3], [4], and hence higher spectral efficiency can be
achieved. Third, the partial common support obtained can be
O RTHOGONAL FREQUENCY division
plexing (OFDM) and multiple-input multiple-output
(MIMO) have attracted considerable interests from both
multi-
further utilized to reduce the complexity of the standard CS
algorithms as well as improve the channel recovery probability
academia and industrial due to the outstanding capability to under low signal-to-noise ratio (SNR) conditions.
combat multipath fading and achieve high spectral efficiency Notation: The lowercase and uppercase boldface letters de-
[1]. Hence, MIMO-OFDM has been widely adopted in the note vectors and matrices, respectively. ()T , ()H , ()1 , () ,
latest communication standards, such as LTE, IEEE 802.11, diag(), and   p denote the transpose, Hermitian transpose,
etc., and it is also considered as a key technique for future matrix inversion, matrix pseudo inversion, diagonal matrix, and
wireless communications [2]. l p norm operations, respectively. Finally, represents the sub-
As one of the major challenges for practical MIMO-OFDM matrix comprised of the columns of the matrix .
systems, accurate channel estimation is essential to guarantee
the system performance [2]. The conventional channel estima- II. S YSTEM M ODEL
tion methods for MIMO-OFDM systems can be categorized
into two types: the time-domain preamble based [3] and the A. MIMO Channel Model With Sparse Common Support
frequency-domain orthogonal pilot aided [4] methods. How- In an Nt Nr MIMO system consisting of Nt transmit and Nr
ever, the overhead of the required preamble or pilots will sig- receive antennas, the L-length channel impulse response (CIR)
associated with the p-th transmit antenna and a certain receive
Manuscript received September 23, 2014; revised October 31, 2014; antenna1 during the i-th symbol can be modeled as [4]
accepted November 14, 2014. Date of publication November 20, 2014; date
 
of current version January 7, 2015. This work was supported by the National (p) (p) (p) (p) T
Natural Science Foundation of China under Grants 61401248 and 61471219. hi = hi,1 , hi,2 , , hi,L . (1)
The associate editor coordinating the review of this paper and approving it for
publication was P. Salvo Rossi.
W. Ding is with the Department of Electronic Engineering and the Research
Due to the physical properties of outdoor electromagnetic prop-
Institute of Information Technology, Tsinghua National Laboratory for Infor- agation, the CIRs in wireless communications usually contain
mation Science and Technology, Tsinghua University, Beijing 100084, China a few significant paths, i.e., the CIRs are sparse [7], [10].
(e-mail: dwb11@mails.tsinghua.edu.cn).
F. Yang and J. Song are with the Department of Electronic Engineering and
Moreover, when the scale of the antenna array is negligible
the Research Institute of Information Technology, Tsinghua National Labo- compared to the signal wavelength in the commonly used
ratory for Information Science and Technology, Tsinghua University, Beijing collocated MIMO systems, channels associated with different
100084, China, and also with Shenzhen City Key Laboratory of Digital TV
System, Shenzhen 518057, China (e-mail: fangyang@tsinghua.edu.cn; jsong@
transmit-receive antenna pairs share the similar path arrival
tsinghua.edu.cn). times.2 Therefore, the corresponding CIRs in MIMO systems
W. Dai is with the Department of Electrical and Electronic Engineer-
ing, Imperial College London, London SW7 2AZ, U.K. (e-mail: wei.dai1@
1 As the processing for channel estimation is identical for every receive an-
imperial.ac.uk).
Color versions of one or more of the figures in this paper are available online tenna, the receive antenna index will be omitted for simplification from now on.
2 This assumption could be invalid in the distributed MIMO system where the
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LCOMM.2014.2372006 distance of the transmit antennas is comparable to the signal wavelength.

1089-7798 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.
DING et al.: TIMEFREQUENCY JOINT SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 59

Fig. 1. The frame structure of the proposed MIMO-TFT-OFDM systems.

usually share a sparse common support in spite of the different

path gains [9].

B. TFT-OFDM Signal Structure for MIMO Systems

As shown in Fig. 1, unlike conventional MIMO-OFDM
schemes where the training information for channel estimation
only exists in either time [3] or frequency domain [4], the Fig. 2. Illustration of the partial common support information acquisition by
the time-domain preamble in a 4 4 MIMO system.
proposed MIMO-TFT-OFDM has both time-domain preamble
and frequency-domain pilots for every symbol. The i-th TFT-
information. Hence, the preamble length here only needs to be
OFDM symbol for the p-th transmit antenna consists of an
no less than the channel length, i.e., M L.
M-length preamble c = [c1 , c2 , , cM ]T and an N-length OFDM
(p) (p) (p) (p) T (p)
At the receiver, we adopt the received preamble c i without
block xi = [xi,1 , xi,2 , , xi,N ] = FH x i , where F is the N interference removal to correlate with the local preamble c to
N discrete Fourier transform (DFT) matrix. The preamble is obtain the partial sparse common support,
an identical pseudo random sequence for all transmit antennas,
Nt
which has good autocorrelation property and can be used for 1 1
hi
(p)
(p) z i = c c i = + vi , (3)
partial common support acquisition. x i denotes the frequency- MNt Nt p=1
domain OFDM block which contains a small amount NP of
(p) NP
pilots on the location of D(p) = {dn }n=1 and NP (Nt 1) where is the circular correlation operation, vi denotes the
null pilots (zeros) on the pilot location reserved for the other interference including the channel noise and the inter-block
transmit antennas. For the (p + 1)-th transmit antenna, the pilot interference (IBI), and the correlation result z i is the rough
location D(p+1) follows the following pattern, estimate of the superposition of all the CIRs associated with
the Nt transmit antennas.
(p+1) (p) (1)
dn = dn + 1 = dn + p, 1 n NP . (2) Due to the channel sparse common support property and the
good correlation property of the preamble, the partial com-
Instead of the equally spaced comb-type pilots in the con- mon support information T0 = {l : zi,l 2 a}L1
l=0 can be ob-
ventional MIMO-OFDM systems [4], the pilot location D(1) = tained by appropriately selecting the threshold a = 3zi 2 /L =
(1) NP 1/2
{dn }n=1 is randomized to guarantee the restricted isometry 3(Ll=1 |zi,l |2 ) /L [11], as shown in Fig. 2, where the ITU Ve-
property of the observation matrix in the CS [8]. hicular B (ITU-VB) channel model [10] (L = 161) with a SNR
of 10 dB is considered. Meanwhile, the channel sparsity level
which is essential to the CS algorithm can also be estimated
III. T IME -F REQUENCY J OINT C HANNEL E STIMATION
by adding a compensation factor, i.e., K = T0 0 + b. Such
FOR S PARSE MIMO S YSTEMS
information is used to reduce the complexity of the standard
In this section, by fully exploiting the time-frequency train- CS algorithm for practical applications, which will be detailed
ing feature of MIMO-TFT-OFDM scheme and the channel in Step 3.
sparse common support property, we proposed the time-
frequency joint sparse channel estimation based on the sparsity
B. Step 2: Cyclicity Reconstruction of OFDM Block
adaptive simultaneous orthogonal matching pursuit (SA-SOMP)
algorithm. The proposed channel estimation is composed of In the proposed MIMO-TFT-OFDM scheme, the standard
three steps as follows. cyclic prefix (CP) [4] is replaced by the known preamble to
improve the spectral efficiency while sacrificing the cyclicity
property of the received OFDM block under multipath fading
A. Step 1: Partial Common Support Acquisition
channels. Hence, cyclicity reconstruction is required to obtain
Different from the conventional time-domain estimation for the pilot for accurate channel recovery.
MIMO systems where the preamble length M has to be no For single input and single output (SISO) systems, the cyclic-
less than Nt times of the channel length L for accurate channel ity reconstruction of the OFDM block is achieved by the mature
estimation, i.e., M Nt L [3], the proposed scheme utilizes overlapping and adding operation (OLA) [11], [12]. Then the
the preamble merely to acquire the partial channel support cyclicity reconstruction of the OFDM block in MIMO systems
60 IEEE COMMUNICATIONS LETTERS, VOL. 19, NO. 1, JANUARY 2015

can be implemented by extending the OLA operations above to T0 0 )Nt NP (L + K 2 )), which is less than that of the standard
every transmit antenna. After that, the orthogonal pilots can be SOMP algorithm, O (KNt NP (L + K 2 )). Moreover, the channel
extracted for the accurate channel estimation in the following sparsity level can be well estimated by the partial common
Step 3. support information, making the algorithm adaptive to variable
channel sparsity under different conditions.
C. Step 3: Accurate Channel Estimation Based on SA-SOMP
Algorithm 1 Sparsity adaptive SOMP (SA-SOMP).
After cyclicity reconstruction, the received OFDM block in
the frequency domain can be written as,
Inputs:
Nt   1) Initial channel common support T0
diag
(p) (p)
y i = x i FL hi + wi , (4)
p=1
2) Channel sparsity level K;

3) Noisy measurements M = Ui ;
where FL is the N L partial DFT matrix containing the first L
4) Observation matrix = F(1) .
columns of F. We only focus on the received frequency-domain
Output: The K-sparse estimate H = Hi .
pilots located at D(p) , and (4) can be simplified as
Initial Configuration:
(p)
ui = F(p) hi + w
i ,
(p) (p)
1 p Nt , (5) 1: T0 ;
2: H (0) 0; H (0) M;
T
(p) (p) (p) (p)
where ui = [y (p) /x (p) , y (p) /x (p) , , y (p) /x (p) ] , and 3: R M H (0) ;
i,d1 i,d1 i,d2 i,d2 i,dN i,dN
P P Iteration:
F(p) is the NP L partial DFT matrix with the (n + 1, k + 1)-th 4: for k = 1 : K T0 0 do
(p)
entry being exp( j2dn k/N)/ N. 5: P H R;
From the designed pilot pattern according to (2), the recur- 6: {arg max j pi, j 1 };
i
sive representation of F(p) can be written as (k) 0; H (k) M;
7: H
F(p+1) = F(p) = F(1) p , (6) 8: R M H (k) ;
9: end for
where is the diagonal matrix given by 10: H H (k) .
 T 
2 2
= diag 1, e j N , , e j N (L1) . (7)
After the estimate H of Hi has been obtained, the CIR
associated with the p-th transmit antenna can be restored as
(p)
Stacking the noisy measurements ui into one matrix Ui =
1
(1) (2) (N )
[ui , ui , , ui t ], a matrix formulation of (5) for all antennas h (p) = ( p1 ) h p , (9)
can be derived as
where h p is the p-th column of H.
Note that the inversion of a
(1)
i,
Ui = F Hi + W (8) 1
diagonal matrix ( p1 ) only involves a low complexity.

where Hi = [hi , hi , , Nt 1 hi t ] contains all the CIR

(1) (2) (N )

information which needs to be estimated and W i denotes the IV. S IMULATION R ESULTS
(p)
noise. Fortunately, the hi usually share the same sparse sup- This section investigates the performance of the proposed
port and moreover, the matrix p1 , p N will not change MIMO-TFT-OFDM scheme. The simulation parameters are
(p)
the sparse support or the absolute value of hi . For multiple summarized as follows: the system bandwidth is fS = 8 MHz
sparse signals Hi formulated in (8), the SOMP algorithm [8] located at the central frequency of 835 MHz; the OFDM block
developed from the standard OMP algorithm [5] could improve length and the preamble length are N = 4096 and M = 256,
the signal recovery performance by taking the correlation of respectively; the 6-tap ITU-VB channel model specified by
the sparse signals into account. However, it requires the known 3GPP [9] is considered. The threshold is set according to [11]
signal sparsity level K a priori for the recovery performance and the compensation factor is b = 2.
guarantee, which is impractical in real systems. Moreover, its To evaluate the performance gain of the proposed SA-SOMP
computational complexity is high due to the high-dimensional algorithm compared with conventional SOMP and OMP algo-
matrix inversion in every iteration. rithms, Fig. 3 presents the correct CIR recovery probability
Based on the basic principle of SOMP [8], we propose the when different numbers of pilots NP are used under the static
SA-SOMP algorithm to estimate all of the sparse signals in the ITU-VB channel with the fixed SNR of 20 dB in a 4 4
sparse matrix Hi , which is described in Algorithm 1. The pro- MIMO system. Here, the correct recovery is defined as the
posed algorithm exploits the obtained partial common support estimation mean square error (MSE) is lower than 102 . It
information in Step 1 for more accurate initial configuration can be seen from Fig. 3 that by utilizing the obtained partial
and hence reduces the iteration number in the conventional common support information and the sparse common support
SOMP. The SA-SOMP algorithm has the complexity of O ((K property, the required number of pilots in the SA-SOMP is less
DING et al.: TIMEFREQUENCY JOINT SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 61

TABLE I
S PECTRAL E FFICIENCY C OMPARISON

matrix of F(1) . It could be seen that for the mobile ITU-VB

channel, the MSE performance of the proposed TFT-OFDM
scheme performs 5 dB and 7 dB better than the preamble
and comb-type pilot based schemes, respectively. Besides, the
spectral efficiency of the three different schemes has been
Fig. 3. Comparison of the CIR recovery probabilities at the SNR = 20 dB. specified in Table I, and the proposed MIMO-TFT-OFDM
scheme demonstrates much higher spectral efficiency than the
two conventional counterparts.

V. C ONCLUSION
In this letter, a spectrum-efficient MIMO-TFT-OFDM alter-
native to the standard MIMO-OFDM scheme is proposed by
designing a time-frequency joint sparse channel estimation with
high accuracy under the framework of SCS. The proposed SA-
SOMP algorithm has better performance and lower complexity
than the standard SOMP and OMP algorithms. Simulation
results show that the proposed MIMO-TFT-OFDM scheme
demonstrates better performance, more robustness and higher
spectral efficiency than the conventional time- and frequency-
domain based MIMO-OFDM schemes.

R EFERENCES
Fig. 4. Comparison of the MSE performance among different schemes under [1] G. Stuber et al., Broadband MIMO-OFDM wireless communications,
the time-varying ITU-VB channel in a 4 4 MIMO system. Proc. IEEE, vol. 92, no. 2, pp. 271294, Feb. 2004.
[2] J. Ketonen, M. Juntti, and J. Cavallaro, Performance-complexity compar-
ison of receivers for a LTE MIMO-OFDM system, IEEE Trans. Signal
than that in the SOMP algorithm and far less than that in the Process., vol. 58, no. 6, pp. 33603372, Jun. 2010.
standard OMP algorithm, e.g., NP = 15 could ensure the correct [3] S. K. Mohammed, A. Zaki, A. Chockalingam, and B. S. Rajan, Highrate
space-time coded large-MIMO systems: Low-complexity detection and
CIR recovery probability. In practice, considering the tradeoff channel estimation, IEEE J. Sel. Topics Signal Process., vol. 3, no. 6,
between the robustness of the CIR recovery performance and pp. 958974, Dec. 2009.
the pilot overhead, we choose the number of pilots for each [4] H. Minn and N. Al-Dhahir, Optimal training signals for MIMO OFDM
channel estimation, IEEE Trans. Wireless Commun., vol. 5, no. 5,
transmit antenna to be NP = 25 for the proposed scheme. pp. 11581168, May 2006.
Fig. 4 illustrates the MSE performance of different schemes [5] J. A. Tropp and A. C. Gilbert, Signal recovery from random measure-
under the time-varying ITU-VB channel with the mobile speed ments via orthogonal matching pursuit, IEEE Trans. Inf. Theory, vol. 53,
no. 12, pp. 46554666, Dec. 2007.
of 90 km/h in a 4 4 MIMO system. The conventional time- [6] W. Dai and O. Milenkovic, Subspace pursuit for compressive sensing
and frequency-domain channel estimation methods based on signal reconstruction, IEEE Trans. Inf. Theory, vol. 55, no. 5, pp. 2230
preamble and comb-type pilots are also evaluated for com- 2249, May 2009.
[7] N. Wang, G. Gui, Z. Zhang, T. Tang, and J. Jiang, A novel sparse channel
parison. The preamble based scheme adopts the preamble of estimation method for multipath MIMO-OFDM systems, in Proc. IEEE
length M = 1024 for channel estimation, and each preamble is VTCFall, Sep. 2011, pp. 15.
followed by only one OFDM block to combat the time-varying [8] J. A. Tropp, A. C. Gilbert, and M. J. Strauss, Algorithms for simultaneous
sparse approximation Part I: Greedy pursuit, Signal Process., vol. 86,
channel [3]. The comb-type pilot based scheme uses NP = 256 no. 3, pp. 572588, Mar. 2006.
embedded orthogonal pilots for each transmit antenna, and [9] Y. Barbotin, A. Hormati, S. Rangan, and M. Vetterli, Estimation of sparse
hence the overall pilot overhead is 1024 [4]. A 256-length MIMO channels with common support, IEEE Trans. Commun., vol. 60,
no. 12, pp. 37053716, Dec. 2012.
CP is added in both of the two schemes to avoid the IBI [10] L. Dai, J. Wang, Z. Wang, P. Tsiaflakis, and M. Moonen, Spectrum-
and the average power of the preamble or the pilot are set and energy-efficient OFDM based on simultaneous multi-channel recon-
equal to that of the transmitted data for all the tested schemes. struction, IEEE Trans. Signal Process., vol. 61, no. 23, pp. 60476059,
The MSE of the ideal channel estimation is illustrated as the Dec. 2013.
[11] X. Zhou, F. Yang, and J. Song, Novel transmit diversity scheme for TDS-
benchmark, which shows the optimal estimation performance OFDM system with frequency-shift m-sequence padding, IEEE Trans.
when the exact location of nonzero channel taps is known Broadcast., vol. 58, no. 2, pp. 317324, Jun. 2012.
a priori. In addition, the performance of the linear LS based [12] Z. Gao, C. Zhang, Z. Wang, and S. Chen, Priori-information aided
iterative hard threshold: A low-complexity high-accuracy compressive
estimation method with the same pilot number is provided for sensing based channel estimation for TDS-OFDM, IEEE Trans. Wireless
comparison, which turns out to fail due to the underdetermined Commun., to be published.