Вы находитесь на странице: 1из 7

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.

1, JANUARY 2008 65

Iterative Channel Estimators in V-BLAST OFDM Systems


JoonBeom Kim, Student Member, IEEE, Gordon L. Stüber, Fellow, IEEE, and Ye (Geoffrey) Li, Fellow, IEEE

Abstract— An iterative pilot-symbol aided modulation (PSAM) systems have been proposed by Sandell et al. [6] and Valenti et
channel estimation approach is proposed for vertical Bell Lab- al. [7]. Since the iterative channel estimators (ICEs) in [6], [7]
oratories layered space-time (V-BLAST) orthogonal frequency are intended for SISO systems, they have the drawback that the
division multiplexing systems operating on frequency-selective
fading channels. Since the signals at the receive antennas are the interference from other transmit antennas cannot be removed
superposition of signals from multiple transmit antennas, accu- when applied to V-BLAST OFDM systems. Furthermore,
rate channel estimates are crucial for good error performance. since the channel estimation error due to interference between
Furthermore, the time selectivity of the fading channels leads to layers in V-BLAST is proportional to the number of transmit
inter-carrier interference (ICI). While ICI can be ignored for slow antennas, the ICE should account for its own estimation error
fading channels, it should be mitigated for fast fading channels.
This paper proposes an ICI mitigation scheme for time-varying to avoid interference enhancement.
channels. We also propose an iterative channel estimator with This paper proposes an ICE for MPSK V-BLAST OFDM
low-complexity. Simulation results demonstrate the usefulness of systems operating on frequency-selective fading channels. Our
the proposed algorithm on frequency-selective fading channels. approach uses correlations that depend not only on the channel
Index Terms— Adaptive estimation, interference suppression, statistics but also on the a priori information of the transmitted
iterative methods, MIMO systems, parameter estimation, Wiener symbols. Thus, the proposed ICE is more general than the
filtering. approach in [7], and has an inherent interference cancelling
ability [8]. Furthermore, the proposed ICE is shown to be
I. I NTRODUCTION robust on fast fading channels, and significantly improves
the bit error rate (BER) performance when compared to

T HE vertical Bell Laboratories layered space-time (V-


BLAST) system proposed in [1] has received consid-
erable attention due to its high spectral efficiency. Since the
conventional non-iterative PSAM estimators. However, the
computational complexity of the proposed ICE is significantly
greater than that of PSAM techniques due to the required
signals at the various receive antennas are the superposition of inversion of an autocorrelation matrix. Furthermore, unreliable
signals from multiple transmit antennas, the V-BLAST system a priori information from the channel decoder degrades the
requires accurate channel estimates to cancel the inherent performance of the proposed ICE. Hence, we also propose
interference between layers. a low-complexity (LC)-ICE that exploits the most reliable a
For multiple-input multiple-output (MIMO) orthogonal fre- priori information in an efficient manner.
quency division multiplexing (OFDM) systems, a training- The remainder of this paper is organized as follows. Sec-
based channel estimation scheme [2], [3] is often suggested, tion II briefly describes the system model. Section III presents
especially for quasi-static channels that change little over a our proposed ICE and LC-ICE. Sections IV and V describe
data block. However, when applied to rapidly time-variant our inter-carrier interference (ICI) mitigation technique for fast
fading channels, a training-based channel estimation scheme fading channels, and iterative processing with a serially con-
has the disadvantage of providing outdated channel estimates. catenated coded system, respectively. Simulation results and
Moreover, since the training symbols must be inserted with conclusions are provided in Sections VI and VII, respectively.
greater frequency as the rate of channel time variations in-
creases, the system throughput can be significantly reduced II. S YSTEM M ODEL
due to the training overhead. Consider the V-BLAST OFDM system with nt transmit
Pilot-symbol assisted modulation (PSAM) schemes for and nr receive antennas (nr ≥ nt ) shown in Fig. 1, which
channel estimation [4], [5] are popular in single-input single- is given in the top of the next page. The sequence of
output (SISO) systems due to their simplicity and minimum source bits is partitioned into blocks and encoded with a
mean square error (MMSE) optimality. Iterative PSAM chan- rate-R convolutional code to produce the code sequence
nel estimators for binary phase-shift keying (BPSK)-SISO c −1
{c̃lj }N
l=0 , where j = 1, · · · , nt and Nc is the number
Manuscript received August 15, 2006; revised January 9, 2007 and July
of bits in each block. The coded bits are bit-interleaved
4, 2007; accepted July 18, 2007. The associate editor coordinating the and mapped onto an M -ary symbol constellation. After-
review of this letter and approving it for publication was C. Xiao. Prepared wards, known pilot symbols are inserted into the sym-
through collaborative participation in the Collaborative Technology Alliance bol stream to produce the length-K sequence X j (n) =
for Communications & Networks, sponsored by the U.S. Army Research
Laboratory under Cooperative Agreement DAAD19-01-2-0011. The U.S. [Xj (n, 0), · · · , Xj (n, K −1)]T ∈ C K×1 . The symbol vectors
Government is authorized to reproduce and distribute reprints for Government X(n) = [X T1 (n), X T2 (n) · · · , X Tnt (n)]T are then passed to
purposes notwithstanding any copyright notation thereon.
The authors are with the School of Electrical and Computer Engi-
an inverse fast Fourier transform (IFFT) to generate the time-
neering, Georgia Tech, Atlanta, GA 30332 USA (e-mail: {jbkim, stuber, domain sequence
liye}@ece.gatech.edu). (nt )H
Digital Object Identifier 10.1109/TWC.2008.060583. x(n) = F̃ X(n), (1)
1536-1276/08$25.00 
c 2008 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.
66 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 1, JANUARY 2008

Y1 (n, k) (Pilot)
X1 (n, k)
(1,1) ûm
 hn (k; l) −1 1
ENC F −1 F DEC
um c̃l1 cl1 (nr ,1)
Channel
1
hn (k; l) λ(cl1 )
.. .. Estimator
. . (ICI λe (c̃l1 )
.. .. OFDM ..
. . Mitigation) .
Mod.
&
Ynr (n, k)
Xnt (n, k) MMSE
(nt ,1) λ(clnt )
hn (k; l) Detector ûm
 −1
−1 nt
ENC F F DEC
(n ,n )
um
nt
hn t r (k; l)
c̃lnt clnt
 λe (c̃lnt ) 
(Pilot) Mapper

Mapper

Fig. 1. V-BLAST MIMO-OFDM system with proposed receiver.

(M)
where F̃ = I M ⊗ F (⊗ denotes the Kronecker in the OFDM time-frequency lattice are pt OFDM symbols
product), I M is an M × M identity matrix, [F ]p,q = in the time direction and pf sub-carriers in the frequency
√1 exp{−j2πpq/K}, p, q = 0, 1, · · · , K − 1, and [A]p,q is direction. The transmitted OFDM symbols that contain pilot
K
the (p, q)-th element of matrix A. After appending a length-G sub-carriers have the form
cyclic prefix (CP), the nt OFDM symbols are applied to their
respective transmit antennas for transmission. At the receiver, X j (αl ) = [P j , Xj (αl , nt ), · · · , Xj (αl , pf − 1), P j ,
  
after removing the CP and applying the FFT to y(n), the (pf −nt ) symbols
received signal vector at epoch n is · · · , Xj (αl , K − 1)]T , (5)

Es
Y (n) = H(n)X(n) + W (n), (2) where αl = l · pt , l = 0, 1, · · · , Nt − 1, Nt = N/pt is the
nt
number of OFDM symbols that contain pilot sub-carriers in
(nr )
where W (n) = F̃ w(n) is a zero-mean complex Gaussian a block of N OFDM symbols, and P j is an nt -dimensional
random vector with covariance matrix N0 I nr K . The channel row vector in which all elements are zero except for the jth
matrix H(n) is given by element which has complex value Pj .
⎡ ⎤
H (1,1) (n) · · · H (1,nt ) (n)
⎢ .. .. .. ⎥
H(n) = ⎣ . . . ⎦, (3) III. C HANNEL E STIMATION
(nr ,1) (nr ,nt )
H (n) · · · H (n) An ICE can be implemented by calculating tentative channel
estimates based on both a priori information of the transmitted
where H (i,j) (n) ∈ C K×K , i = 1, · · · , nr and j = 1, · · · , nt ,
signals and the initial channel estimates, which are calculated
whose elements are given by
as described in Section III-A. Furthermore, channel fading
[H (i,j) (n)]p,q induces ICI [9], [10]. The ICI can be neglected for slow fading
⎧ K−1 L−1 channels, meaning that the off-diagonal elements of H (i,j) (n)
⎪   (i,j)
⎪ 1
⎪ hn (m; l)e−j2πpl/K p = q (p = q in (4)) are small. However, the ICI cannot be ignored

⎨ m=0 l=0
K

=  L−1
K−1  (i,j) , (4) for fast fading channels. Under such conditions, the ICI must
1

⎪ h (m; l)· be mitigated in the channel estimation process as described
⎪ m=0 l=0 n

K
⎩ in Section IV. Without loss of generality, we will derive
e−j2π(p−q)m/K e−j2πql/K p= q
our ICE for slow fading channels. However, for fast fading
(i,j)
where hn (m; l) = h(i,j) (n(K + G) + m; l) is the lth tap channels, the proposed ICE can be similarly implemented
(L ≤ G) of the channel between the jth transmit antenna after ICI mitigation. In our approach, we require the following
and ith receive antenna at epoch m of the nth OFDM assumptions:
symbol. Note that the diagonal and off-diagonal elements of 1) The transmitted symbols are statistically independent
H (i,j) (n) correspond to the desired signal and the undesired and of equal energy, i.e.,
ICI, respectively. E[X(n)X H (n)] = I nt K and |Xj (n, k)|2 = 1.
2) The second order statistics of the channels are avail-
A. Pilot Structure able at the receiver.
In the proposed system, pilot symbols are transmitted on 3) The channels  are spatially uncorrelated, i.e.,
(i,j) (i ,j  )∗ 
distinct sub-carriers for each transmit antenna, while zeros are E[hn (k; l)hn (k ; l)] = R̄(k−k  ; l)δi−i δj−j  .
transmitted on the sub-carriers that are reserved to transmit pi- 4) The lth and l th channel taps are uncorrelated, i.e.,
(i,j) (i,j)∗
lot symbols from other antennas. The pilot sub-carrier spacings E[hn (k; l)hn (k  ; l )] = R̄(k − k  ; l)δl−l .

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 1, JANUARY 2008 67

(q) (q) (q) (q)


Z j (n, k) = [Z1,j (n, k), Z2,j (n, k), · · · , Znr ,j (n, k)]T ∈ C nr , (12)
(q−1) (q−1) (q−1) (q−1)
X̂ j (n, k) = [X̂1 (n, k), · · · , X̂j−1 (n, k), 0, X̂j+1 (n, k), · · · , X̂n(q−1)
t
(n, k)]T (13)

(i,j)
H̃ (n, k) = [H̃ (i,j) (n + P, k + Q), · · ·, H̃ (i,j) (n + P, k − Q), · · ·, H̃ (i,j) (n − P, k + Q), · · ·, H̃ (i,j) (n − P, k − Q)]T(15)

(q) (q−1)
A. Conventional PSAM Channel Estimator where Z j (n, k) and X̂ j (n, k) are given in the top of
(q−1)
Based on the channel estimator in [4] and an assumed this page where j = 1, · · · , nt , and Ĥ (n, k) is the nr ×
separability of the channel correlation function in time and nt estimated channel matrix for the kth sub-carrier at epoch
frequency [11], nt ×nr conventional PSAM channel estimators (q−1)
n and the (q − 1)th iteration. Vector X̂ j (n, k) is the a
can be implemented to produce the initial channel estimates. priori information vector of the symbols transmitted from all
For brevity, only the channel estimator in frequency direction transmit antennas, except for the jth transmit antenna at the
is described here. First, define the tentative channel estimate (q − 1)th iteration. Hereafter, for mathematical convenience,
vector we drop the subscript (q).
 To calculate the nr × nt refined channel estimates at the
V (i,j) (αl ) = [V (i,j) (αl , j), V (i,j) (αl , pf + j), · · · ,
kth sub-carrier and epoch n, assume that the previous channel
V (i,j) (αl , (Nf − 1)pf + j)]T , (6)
estimate Ĥ(n, k) is perfect. Then, the residual received signal
(i,j)
where V (n, k) = Pj∗ (n, k)Yi (n, k),
Pj (n, k) is the known associated with the jth transmit and the ith receive antenna,
pilot symbol, and Nf = K/pf is the number of pilot sub- which is the ith element of Z j (n, k) in (11), is given by
carriers in an OFDM symbol that contains pilot sub-carriers. E s
After simple manipulation, the PSAM channel estimate for the Zi,j (n, k) = H i (n, k)[X (n, k) − X̂ j (n, k)] + Wi (n, k),
nt
kth sub-carrier at epoch αl is given by (14)
(i,j) (i,j)
Ĥ (αl , k) = cH
j V (αl ), (7) where H i (n, k) is the ith row vector of H(n, k) and i =
where the weighting vector cj is given by [12] 1, · · · , nr . Define a soft tentative channel estimate vector
associated with the jth transmit antenna and the ith receive
cj = R−1 r j . (8) antenna as (15), which is given in the top of this page,
(i,j)
The autocorrelation matrix R and the crosscorrelation vector where H̃ (i,j) (n, k) = X̂j∗ (n, k)Zi,j (n, k) and H̃ (n, k) ∈
r j in (8) are given by (P̄ Q̄)×1
C with P̄ = 2P + 1 and Q̄ = 2Q + 1, where P̄ and
Q̄ are interpolation orders in time and frequency directions,
R = E[V (1,1) (αl )V (1,1)H (αl )] (9a)
respectively. Then, the refined channel estimate Ĥ (i,j) (n, k)
(i,j) (i,j)∗
r j = E[V (αl )H (n, k)], (9b) can be expressed as the linear interpolation [12]
respectively, where the H (i,j) (n, k) are the principal diago-
(i,j)
nal elements of H (i,j) (n) (p = q in (4)). Since the pilot Ĥ (i,j) (n, k) = M (i,j)H (n, k)H̃ (n, k), (16)
symbol spacing is same for all transmit and receive antenna
pairs, the autocorrelation matrix in (9a) is identical for all where M j (n, k) ∈ C (P̄ Q̄)×1 . From orthogonality principles,
pairs i, j, where j = 1, · · · , nt , i = 1, · · · , nr . However, the matrix M (i,j) (n, k) satisfies
the crosscorrelation vectors in (9b) are distinct, because the (i,j)
M opt (n, k) = [R(i,j) (n, k)]−1 r(i,j) (n, k), (17)
pilot symbol positions are distinct for the different transmit
antennas. Hence, the weighting vector cj is distinct for each where the autocorrelation matrix and crosscorrelation vector
transmit antenna. are given by
(a) (i,j) (i,j)H
R(i,j) (n, k) = Rj (n, k) = E[H̃ (n, k)H̃ (n, k)]
B. Two-Dimensional Iterative Channel Estimator
(18a)
From (2), the received signal for the kth sub-carrier at epoch (a) (i,j)
(i,j) (i,j)∗
n is defined as r (n, k) = rj (n, k) = E[H̃ (n, k)H (n, k)],

Es (18b)
Y (n, k) = H(n, k)X(n, k) + W (n, k), (10)
nt (a)
respectively. Note that = comes from the fact that the
where X(n, k) ∈ C nt , Y (n, k) ∈ C nr , and H(n, k) ∈ channels associated with all transmit and receive antenna pairs
C nr ×nt . We now proceed to calculate the channel estimate are uncorrelated and statistically identical. Before calculating
vector associated with the jth transmit antenna. The residual (18), the a priori covariance matrix is first evaluated
received signal vector associated with the jth transmit antenna H
and the kth sub-carrier at epoch n and the qth iteration is Φj (r − r , s − s ) = E[X̃ j (n + r, k + s)X̃ j (n + r , k + s )],
 (19)
(q) Es (q−1) (q−1)
Z j (n, k) = Y (n, k) − Ĥ (n, k)X̂ j (n, k), 
nt where X̃ j (n, k) = X(n, k) − X̂ j (n, k), |r|, |r | ≤ P ,
(11) and |s|, |s | ≤ Q. Since the elements in X(n, k) are all
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.
68 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 1, JANUARY 2008



⎪ 1 − |X̂j (n + r, n + s)|2 m = j, r = r and s = s

1 m = j, r = r and s = s
[Φj (r − r , s − s )]m,m = , (20)

⎪ 0 m = j, r = r or s = s

X̂j (n + r, n + s)X̂j∗ (n + r , n + s ) m = j, r = r or s = s

independent, the elements of Φj (r − r , s − s ) are, after D. Comparison for Computational Complexities


some mathematical manipulation, given in the top of this The computational complexity of the PSAM approach [4]
page, where it is assumed that |Xj (n, k)|2 = 1. Note that is O(nr · nt (Nf3 + Nt3 + (2Nf Nt + 2Nt N )K)), where Nf =
E[Xj (n, k)] = X̂j (n, k), which will be discussed in Sec- K/pf , Nt = N/pt , and N is the number of OFDM symbols
tion V-B. Hence, using (14), (15), and (20), the elements of in each block. The cubic terms arise from the matrix inversion,
the autocorrelation matrix in (18a) are and the (2Nf Nt + 2Nt N )K terms arise from interpolation in
Rj (r − r , s − s ) the frequency and time directions, respectively. The additional
Es ∗ computational complexity, over the PSAM approach, that is
= trace(Φj ) · X̂ (n + r, k + s)X̂j (n + r , k + s ) · required to implement the full-complexity ICE is O(((P̄ Q̄)3 +
nt j
3(P̄ Q̄)2 + 4P̄ Q̄)nt N K(NI − 1)) ≈ O((P̄ Q̄)3 nt N K(NI −
R̄(r − r , s − s ) + |X̂j (n + r, k + s)|2 N0 δr−r δs−s , 1)), where NI (> 1) is the number of iterations. Note that
(21) NI = 1 implies that only the initial PSAM channel estimates
are used. Similarly, the additional computational complexity
where trace(A) is the sum of the elements on the principal
required to implement the LC-ICE is O(L3th nt N K(NI − 1)).
diagonal of A, and
Note that the additional computational complexity required
R̄(r − r , s − s ) to implement the LC-ICE, O(L3th nt N K(NI − 1)), is much
= E[H (i,j) (n + r, k + s)H (i,j)∗ (n + r , k + s )] smaller than the additional computational complexity required
to implement the full complexity ICE, O((P̄ Q̄)3 nt N K(NI −
1 
K−1
= 2 J0 [2πfd,max(Ts (r − r ) + T (m − m ))] 1)).
K m,m =0


L−1 IV. ICI M ITIGATION T ECHNIQUES

· σl2 e−j2π(s−s )l/K , (22) Perfect channel knowledge is necessary to cancel the ICI.
l=0 However, since perfect CSI is never available, the ICI cannot
where fd,max is the maximum Doppler frequency, T is the be completely removed. For this reason, we propose to use a
time-domain sample period, Ts = (K + G)T , and 2-D filter to mitigate the ICI. The received signal vector Y (n) in
isotropic scattering is assumed. The elements of the cross- (2) is
correlation vector in (18b) are 
Es
Y (n) = [ H d (n)X(n) + H o (n)X(n)] + W (n),
rj (r, s) = |X̂j (n + r, k + s)|2 · R̄(r, s). (23) nt      
desired signal ICI
(25)
C. Low-Complexity Iterative Channel Estimator
The computational complexity of the two-dimensional ICE where Y (n) = [Y T1 (n), Y T2 (n) · · · , Y Tnr (n)]T and
is prohibitive, thus motivating our LC-ICE. In the two- Y i (n) = [Yi (n, 0), · · · , Yi (n, K − 1)]T . The second term
dimensional ICE, unreliable a priori information from the in (25) is the ICI due to channel time variations. To mitigate
channel decoder results in channel estimation errors. For this the ICI, we generate the weighted vector
reason, we propose an LC-ICE that orders the (P̄ Q̄) coordi-
Y̌ (n) = B H (n)Y (n), (26)
nates of the ICE and uses only the most reliable positions
to generate the channel estimates. To implement the LC-ICE, where B(n) ∈ C nr K×nr K . By orthogonality principles, the
we exploit not only the correlation coefficient, but also the a optimal weight B opt (n) is
priori information. Define a weight indicator function (WIF)
as B opt (n) = [E{Y (n)Y H (n)}]−1 E{Y (n){H d (n)X(n)}H }.

nt (27)
ξ(αQ̄ + β) = |R̄(r, s)|2 |X̂j (n − r, k − s)|2 , (24)
Since it is assumed that all channels are uncorrelated and
statistically identical, the term E{Y i (n)Y H
j=1
i (n)} in (27) is
where α = r + P , β = s + Q, (|r| ≤ P , |s| ≤ Q), and R̄(r, s) 
Gi i = i
is given by (22). Then, we order the WIF with ξ(i0 ) ≥ ξ(i1 ) ≥ E{Y i (n)Y H i (n)} = , (28)
0K×K otherwise
· · · ≥ ξ(iLth ) ≥ · · · ≥ ξ(iP̄ Q̄−1 ), where il = 0, 1, · · · , P̄ Q̄ −
1. We use the a priori information corresponding to the first where Gi is a K × K matrix. The elements of matrix Gi
Lth < P̄ Q̄ positions (with indices l = 0, 1, · · · , Lth − 1) to become (29), which is given in top of the next page, where p =
estimate the channel for the kth sub-carrier at epoch n. By 0, · · · , K−1. The term E{Y (n){H d (n)X(n)}H } in (27) can
using these Lth positions, the LC-ICE can be implemented in be similarly derived. Then, the optimal weight B opt (n) can
a fashion similar to the ICE. be rewritten as (30), which is given in the top of the next page,
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 1, JANUARY 2008 69

Es    
K−1 K−1 K−1 L−1

E [Yi (n, r + p)Yi∗ (n, r)] = 2
R̄(k − k  ; l)ej2π(m−r)(k−k )/K e−j2πpk/K + N0 δp = (Es + N0 )δp , (29)
K m=0  k=0 k =0 l=0

Es  t n
(1,j)H (n ,j)H
B opt (n) = Bdiag[E(H (1,j) (n)H d (n)), · · · , E(H (nr ,j) (n)H d r (n))], (30)
nt (Es + N0 ) j=1

where Bdiag(·) is a block diagonal matrix. The elements in


(30) are 10 0
 
(i,j)H
E [H (i,j) (n)]p,q [H d (n)]q,q = nt = nr = 2

1 
K−1
J0 (2πfd T (m − m ))e−j2π(p−q)m/K ,(31)
K2 10 -1
m,m =0

where p, q = 0, 1, · · · , K − 1.

NMSE
V. T URBO P ROCESSING WITH A S ERIALLY 10 -2
C ONCATENATED S YSTEM
A. MMSE V-BLAST Detector and Soft-Output Demapper Conventional PSAM, fd,maxTs = 0.04
The output of the MMSE detector is P=Q=1, fd,maxTs = 0.04, 2 Iterations
10 -3 P=Q=2, fd,maxTs = 0.04, 2 Iterations
−1 H
Θ(n, k) = [Ψ(n, k) + ΔI nt ] Ĥ (n, k)Y (n, k), (32) Conventional PSAM, fd,maxTs = 0.09
P=Q=1, fd,maxTs = 0.09, 2 Iterations
where for a zero-forcing detector Ψ(n, k) =
H P=Q=2, fd,maxTs = 0.09, 2 Iterations
Ĥ (n, k)Ĥ(n, k) and [Δ]j,j = N0 /Es + δj , and where
δj results from the imperfect channel estimates in (14). 10 -4
The channel estimation error δj can be calculated from the 2 4 6 8 10 12 14
previous channel estimates, viz., Eb /N0
−1
δj = (nt − 1)[Es − r H
j (n, k)Rj (n, k)r j (n, k)]. (33) Fig. 2. NMSE versus Eb /N0 with various normalized fading rates fd,max Ts
in TU channels when nt = nr = 2.
Given the signals Θ(n, k), the demapper calculates the log-
likelihood ratio (LLR) of the coded bits as

p[Xj (n, k)|Θ(n, k)]
Xj :clj =1
in Fig. 1. We use the maximum squared Euclidean weight
λ(clj ) = log  , (34) (MSEW) symbol mapper [13] that provides the best error-
p[Xj (n, k)|Θ(n, k)]
floor performance in the high SNR region. The expectation of
Xj :clj =0
the code bits E[clj ] and the corresponding a priori probabilities
where l = 0, · · · , q − 1 and j = 1, · · · , nt . in (36) are grouped by the mapping rule to yield the a priori
transmitted symbol
B. Channel Decoder and Bit-to-Symbol Mapper
E[Xj (n, k)]= X̂j (n, k) (37)
The LLRs of the coded bits in the channel decoder are   
0 q−1
 c −1
= T (cj )Pr T (cj = (cj , · · · cj )) ,
n n
p[c̃lj |{λ(c̃lj )}N
l=0 ] cn
j :Xj ∈M
c̃lj =1
c −1
λ(c̃lj |{λ(c̃lj )}N
l=0 ) = log  c −1 where M is the MPSK symbol alphabet, T (cnj ) is the symbol
p[c̃lj |{λ(c̃lj )}N
l=0 ]
c̃lj =0 corresponding to cnj , where each element of cnj is given
by (36b). The a priori transmitted symbol in (37) and the
= λa (c̃lj ) + λe (c̃lj ), (35)
estimated channel matrix in (16) are used for the input to the
where λa (c̃lj ) is a priori information on c̃lj , evaluated from ICE in (20), (21), (23), and (24). Assuming that the code bits
(34), and λe (c̃lj ) is extrinsic information, which is used as a are independent,
 the a priori
q−1 probability of the transmitted
priori information for the ICE. The a priori probability and symbol is Pr T (cnj ) = i=0 Pr(cnj ).
the expectation of c̃lj are given by [6], [7]
   VI. S YSTEM PARAMETERS AND S IMULATION R ESULTS
1 1 c −1
Pr[c̃lj ] = 1 + c̃lj tanh λe [c̃lj |{λ(c̃lj )}N ] ,(36a) In our simulation, typical urban (TU) channels are generated
2 2 l=0
with fd,max Ts = 0.006 through 0.12 [14]. A bandwidth
E[c̃lj ] = tanh(λe (c̃lj )/2), (36b)
of 500 kHz is divided into 64 sub-channels. The symbol
respectively. In case of non-binary modulation the code bits duration is 128 μs, and a 14 μs guard interval is used. A rate-
must be mapped onto the transmitted symbols as shown 1/2 recursive systematic convolutional code with generator
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.
70 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 1, JANUARY 2008

10 -1 10 -1
Conventional PSAM P=Q=1,2 Iterations, fd,max T s = 0.04
P=Q=1,2 Iterations P=Q=2,2 Iterations, fd,max T s = 0.04
P=Q=2,2 Iterations LC-ICE,2 Iterations, fd,max Ts = 0.04
10 -2 P=0,Q=9,2 Iterations P=Q=1,2 Iterations, fd,max T s = 0.09
P=9,Q=0,2 Iterations
P=Q=2,2 Iterations, fd,max T s = 0.09
10 -2
LC-ICE,2 Iterations, fd,max Ts = 0.09
BER

BER
fd,max Ts = 0.09
10 -3

10 -3
10 -4

nt = nr = 2 nt = nr = 2 fd,max Ts = 0.04
-5 -4
10 10
0.00 0.02 0.04 0.06 0.08 0.10 6 8 10 12 14 16 18 20 22
Normalized fading rate (fd,maxTs ) Order (Lth )

Fig. 3. BER versus normalized fading rate fd,max Ts in TU channels at Fig. 4. BER versus order Lth of the LC-ICE with fd,max Ts = 0.04 and
Eb /N0 = 12 dB and nt = nr = 2, where fd,max is the maximum Doppler 0.09 in TU channels when Eb /N0 = 10 dB and nt = nr = 2. For reference,
frequency. Ts is the duration of the extended OFDM symbol. the error performance of the ICE with P = Q = 1 and P = Q = 2 is shown.

polynomial (1, 021/037) is used. The bit-interleaved coded


the limit of the two-dimensional Nyquist sampling rate. Con-
bits are mapped onto a QPSK constellation with a MSEW
sequently, for fast fading channels the one-dimensional ICE
mapper [13]. Each OFDM block consists of N = 64 OFDM
 in the frequency direction (P = 0) and the two-dimensional
symbols. The received bit energy-to-noise ratio is Eb /N0 = ICE yield more robust error performance than either the
nt Eb /N0 , where Eb is bit energy per transmit antenna. Two- conventional PSAM or the one-dimensional ICE in the time
dimensional Nyquist sampling requires that pf < K/L and direction (Q = 0).
pt < 1/(2fd,maxTs ). The values of pt and pf are set at 4 and
Fig. 4 compares error performance of the LC-ICE with that
8, respectively.
of the ICE when fd,max Ts = 0.04 and 0.09 as a function of
The normalized mean square error (NMSE) for the channel
the order Lth . For reference, the error performance of the ICE
estimates is given by
  with P = Q = 1 (or P̄ Q̄ = 9) and P = Q = 2 (or P̄ Q̄ = 25)

E |H (i,j) (n, k) − Ĥ (i,j) (n, k)|2 is shown. Since the large computational complexity of the two-
n,k,i,j dimensional ICE results from the matrix inversions, which are
NMSE  =    . (38)
E |H (i,j) (n, k)|2 required for every position and transmit antenna as described
n,k,i,j in (18) and (20), it can be substantial. However, as shown
Fig. 2 plots the NMSE for different normalized fading rates, in Fig. 4, the performance of the LC-ICE with Lth = 14 is
i.e., fd,max Ts = 0.04 and 0.09 when P = Q = 1 and 2, degraded only slightly from that of the full complexity ICE
respectively. The NMSE of the proposed ICE is more than with P = Q = 2.
2 dB better than that of the conventional PSAM estimator at Fig. 5 shows the BER of the proposed ICE and LC-ICE
higher fading rates. Thus, the overall NMSE of the proposed when fd,maxTs = 0.04 and 0.12. As shown in Fig. 5 (a), the
ICE is better than that of the conventional PSAM approach. As ICE in the frequency direction or the two-dimensional ICE has
mentioned before, since the proposed ICE uses nearby symbol at least 0.5 dB better performance than the ICE in the time
positions to estimate the channel rather than just the pilot direction, which is due to the fact that correlation between
symbol positions, it makes better use of the channel correlation samples in the time direction reduces rapidly as the fading
characteristics. rate increases. For reference, the error performance of a linear
Fig. 3 compares the BER performance of two-dimensional interpolator with perfect channel coefficients is shown, which
and one-dimensional ICEs at Eb /N0 = 12 dB against the yields an ICI-free signal in (11) after the first iteration. If the
normalized fading rate for a TU channel. In the lower normal- two-dimensional ICE with P = Q = 2 is employed, the error
ized fading rate region (< 0.03), the bit-wise interleaver is the performance is degraded 0.5 dB from that of the interpolator
main factor for the performance improvement in the proposed with perfect channel knowledge in the high SNR region. As
ICE. However, at higher normalized fading rate region the shown in Fig. 5 (b), the ICE with our proposed ICI mitigation
system performance degrades because of the reduced channel scheme has at least a 0.5 dB performance gain over the ICE
correlation and because the pilot symbol spacing approaches without ICI mitigation at the fifth iteration.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 1, JANUARY 2008 71

10 0
2 × 2 MIMO system. This also results in the error floor at the
Conventional PSAM fifth iteration in the lower BER region.
P=Q=1,2 Iter
P=Q=2,2 Iter
10 -1 P=0,Q=7,2 Iter VII. C ONCLUSIONS
P=0,Q=9,2 Iter
P=7,Q=0,2 Iter
This paper proposed an ICE and LC-ICE that exploits
P=9,Q=0,2 Iter a priori information and the channel correlation statistics.
LC-ICE,Lth =14,2 Iter Simulations have demonstrated that the BER performance
10 -2
of the proposed two-dimensional ICE is at least 2.5 dB
BER

better than the non-iterative conventional PSAM estimator.


Furthermore, the system performance of the proposed ICE
10 -3
is robust even if the pilot symbol spacing approaches the
limit of the two-dimensional Nyquist sampling rate. However,
the computational complexity of the proposed ICE increases
10 -4 significantly over the conventional PSAM approach because
the inverse of the channel autocorrelation matrix is needed
nt = nr = 2 to calculate the channel estimates. However, our proposed
10 -5 LC-ICE can achieve a similar error performance as the full
2 4 6 8 10 12 14 16 complexity ICE, but with significantly reduced complexity.
Eb /N0
D ISCLAIMER
(a) fd,maxTs = 0.04
The views and conclusions contained in this document
are those of the authors and should not be interpreted as
10 0 representing the official policies, either expressed or implied,
of the Army Research Laboratory or the U. S. Government.

10 -1 R EFERENCES
[1] G. J. Foschini,“Layered space-time architecture for wireless commu-
nication in a fading environment when using multi-element antennas,”
Bell Labs. Tech. J., pp. 41-59, Autumn 1996.
10 -2 [2] T. L. Marzetta, “BLAST training: Estimating channel characteristics for
BER

high capacity space-time wireless,” in Proc. 37th Annu. Allerton on


Conf. Communication, Control, and Computing, Sept. 1999.
[3] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for
10 -3 n t =n r =2 , W/O ICI filter,1 Iter MIMO OFDM systems in mobile wireless channels,” IEEE Trans.
n t =n r =2 , W/O ICI filter,5 Iter Signal Processing, vol. 51, pp. 1615-1624, June 2003.
n t =n r =2 , With ICI filter,1 Iter [4] J. K. Cavers, “An analysis of pilot symbol assisted modulation for
n t =n r =2 , With ICI filter,5 Iter Rayleigh fading channels,” IEEE Trans. Veh. Technol., vol. 40, pp. 686-
10 -4 n t =n r =3 , W/O ICI filter,5 Iter 693, Nov. 1991.
n t =n r =3 , With ICI filter,5 Iter [5] O. Edfors, M. Sandell, J. J. van de Beek, S. K. Wilson, and P. O.
n t =3, n r =4 , W/O ICI filter,5 Iter Börjesson, “OFDM channel estimation by singular value decomposi-
n t =3, n r =4 , With ICI filter,5 Iter P =Q=2 tion,” IEEE Trans. Commun., vol. 46, pp. 931-936, July 1998.
[6] M. Sandell, C. Luschi, P. Strauch, and R. Yan, “Iterative channel
10 -5 estimation using soft decision feedback,” in Proc. IEEE Globecom 1998,
2 4 6 8 10 12 14 16 pp. 3728-3733.
[7] M. C. Valenti and B. D. Woerner, “Iterative channel estimation and de-
Eb /N0 coding of pilot symbol assisted Turbo codes over flat-fading channels,”
(b) fd,max Ts = 0.12 IEEE J. Select. Areas Commun., vol. 19, pp. 1697-1705, Sept. 2001.
[8] X. Wang and V. Poor, “Iterative (Turbo) soft interference cancellation
Fig. 5. BER versus Eb /N0 with normalized fading rates fd,max Ts = 0.04 and decoding for coded CDMA,” IEEE Trans. Commun., vol. 47, pp.
and 0.12 in TU channels. 1046-1061, July 1999.
[9] M. Russell and G. L. Stüber, “Interchannel interference analysis of
OFDM in a mobile environment,” in Proc. Vehicular Technology Conf.,
pp. 820-824, July 1995.
Finally, since the total transmit power is constant in (2), [10] A. Stamoulis, S. N. Diggavi, and N. Al-Dhahir, “Intercarrier interference
in MIMO OFDM,” IEEE Trans. Signal Processing, vol. 50, pp. 2451-
the transmit power for each transmit antenna is reduced as the 2464, Oct. 2002.
number of transmit antennas increases. However, the ICI at [11] Ye (Geoffrey) Li, L. J. Cimini, Jr., and N. R. Sollenberger, “Robust
each receive antenna maintains the same power level. This channel estimation for OFDM systems with rapid dispersive fading
channels,” IEEE Trans. Commun., vol. 46, pp. 902-915, July 1998.
implies that the error performance degradation due to ICI [12] S. M. Kay, Fundamentals of Statistical Signal Processing:Estimation
becomes more severe as the number of transmit antennas Theory. Prentice-Hall, vol. 1, 1993.
increases. This error performance degradation can be reduced [13] J. Tan and G. L. Stüber, “Analysis and design of interleaver mappings
for iteratively decoded BICM,” IEEE Trans. Wireless Commun., vol. 4,
by using more receive antennas as shown in Fig. 5 (b). pp. 662-672, Mar. 2005.
Note that since the pilot spacing does not satisfy the Nyquist [14] G. L. Stüber, Principles of Mobile Communications. Kluwer Academic
sampling rate, an error floor exists at the first iteration for the Publishers, 2nd ed., 2000.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 4, 2009 at 06:54 from IEEE Xplore. Restrictions apply.

Вам также может понравиться