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

net/publication/262209295

Polynomial Expansion Linear Detector Based Turbo Receiver for Block

Transmission Systems*This work was supported by China's 863 program
(NO.2003AA123310), national natural science...

Article in IEICE Transactions on Fundamentals of Electronics Communications and Computer Sciences · November 2005
DOI: 10.1093/ietfec/e88-a.11.3123

CITATIONS READS

0 133

3 authors:

Dongming Wang Xiqi Gao

Southeast University Southeast University
321 PUBLICATIONS 5,184 CITATIONS 592 PUBLICATIONS 18,315 CITATIONS

SEE PROFILE SEE PROFILE

You Xiaohu
Southeast University
1,163 PUBLICATIONS 21,827 CITATIONS

SEE PROFILE

All content following this page was uploaded by Xiqi Gao on 22 May 2015.

The user has requested enhancement of the downloaded file.

IEICE TRANS. FUNDAMENTALS, VOL.E88–A, NO.11 NOVEMBER 2005
3123

LETTER Special Section on Wide Band Systems

Polynomial Expansion Linear Detector Based Turbo Receiver for

Block Transmission Systems∗
Dongming WANG†a) , Xiqi GAO† , and Xiaohu YOU† , Nonmembers

SUMMARY A polynomial expansion linear detector (PELD) based onal matrix with x on its diagonal.
turbo receiver is proposed for single-carrier zero-padding block transmis-
sion (ZP-SCBT) systems over the single-input single-output (SISO) and
multiple-input multiple-output (MIMO) frequency selective channels. The
2. Linear System Models of ZP-SCBT Systems
performance is compared with the minimum mean square error (MMSE)
and match filter (MF) based turbo receivers. It is shown that the PELD- At the transmitter, a stream of information bits is encoded,
based turbo receiver provides a good trade-off between performance and bit-interleaved, and mapped to symbols. The symbols are
complexity compared with the other two alternatives.
key words: polynomial expansion linear detector, turbo receiver, block
grouped into blocks with each block having K symbols, and
transmissions then, L (L is larger than the maximum channel order) zeros
are added to the tail of each block to combat the multipath
effect [4]. Typically, to maintain the transmission rate, a
1. Introduction
large block size K is chosen. At the receiver, with the defi-
nition of the vectors and matrices
Recently, turbo processing has received much attention.
T
One of the promising low-complexity approaches is MMSE- s
sT1 , sT2 , · · · , sTn , · · · , sTN ,
based turbo detector. It is proposed by Wang and Poor [1]
T
for turbo multiuser detection and lately is extended to turbo r
r1T , r2T , · · · , rmT , · · · , r TM ,
equalization by Tüchler [2]. Due to the matrix inversion, the
T
complexity of MMSE-based turbo equalization is still a ma- n
nT1 , nT2 , · · · , nTm , · · · , nTM ,
 
jor concern for practical applications, especially for MIMO  H1,1 H1,2 · · · H1,N−1 H1,N 
frequency selective channels. Match-filter (MF) [3] based  H2,1 H2,2 · · · H2,N−1 H2,N 
turbo receiver does not involve any matrix inversion, but the H =  .. .. .. .. ..
 ,

 . . . . . 
performance is much worse than that of MMSE-based turbo 
equalization. H M,1 H M,2 · · · H M,N−1 H M,N
 
In this work, our objective is to develop soft-decision-  hm,n (0) 0 
based symbol detection scheme with low polynomial com-  . 
 .. hm,n (0) 
plexity. We propose a polynomial expansion linear detector  
 .. .. 
based turbo receiver. Simulation results are presented and Hm,n =  hm,n (L) . . 
discussed in coded SISO and MIMO ZP-SCBT systems. It  hm,n (L) hm,n (0) 
 
is shown that proposed receiver significantly outperforms  .. .. 
MF-based turbo detector and the performance is very close  . . 
to that of MMSE-based turbo receiver. 0 hm,n (L)
Notation: Vectors are column vectors and are denoted
received signal vectors can be expressed as
in lower case bold: x. Matrices are upper case bold: A. I
denotes the identity matrix. ei denotes the ith unit vector. r = Hs + n. (1)
(·)T and (·)H represent transpose and Hermitian transpose,
respectively. E(·) denotes the expected value operator. The Here, sn denotes the block at transmit antenna n, and it is
covariance operator is denoted as Cov(·). diag{x} is a diag- a K × 1 vector. Correspondingly, rm and nm denote the
received signals and noise at receive antenna m. rm is a
Manuscript received March 10, 2005. (K + L) × 1 vector. nm is a vector of independent zero-
Manuscript revised May 22, 2005.
Final manuscript received July 27, 2005. mean complex Gaussian noise entries with variance σ2 /2
†
The authors are with the National Mobile Communication Re- per real component. The channel matrix Hm,n is a (K+L)×K
search Lab., Southeast University, Nanjing, 210018, P.R. China. Toeplitz matrix.
∗
This work was supported by China’s 863 program
(NO.2003AA123310), national natural science foundation of 3. Polynomial Expansion Linear Detector
China (NO. 60390540) and a foundation of excellent doctoral dis-
sertation of Southeast University (NO. YBJJ0403).
a) E-mail: [email protected] The output of the zero-forcing (ZF) detector of (1) can be
DOI: 10.1093/ietfec/e88–a.11.3123 given by

Copyright 
c 2005 The Institute of Electronics, Information and Communication Engineers
 IEICE TRANS. FUNDAMENTALS, VOL.E88–A, NO.11 NOVEMBER 2005
3124

s̃ = (H H H)−1 H H r. (2) (LLR) form.

For the kth symbol (k = 1, 2, . . . , KN), the output of the
The problem which arises with the ZF detector is the high soft interference cancellation based on PELD can be given
computation complexity O(N 3 K 3 ) resulting from the matrix by
inversion. This can be approximated with the PELD (see [5]
and references therein). In general, the PELD can be written k W H (r − H s̄) + ρk s̄k .
ŝk = eH H
(4)
as
where ρk = eH k W Rek , s̄
E(s). Let ŝ = [ ŝ1 , ŝ2 , . . . , ŝKN ] ,
T
Nstage
ρ = [ρ1 , ρ2 , . . . , ρKN ] , and we express (4) as matrix-vector
T
W =γ I − γR i ≈ R−1 , (3) form
i=0

where Nstage denotes the stage of polynomial expansion, ŝ = W H H r − W Rs̄ + diag (ρ) s̄. (5)
 
 R1,1 · · · R1,N−1 R1,N 

In ZP-SCBT systems, to maintain the transmission rate, the
 .. .. .. ..  block size K is much larger than L, and Rm,n is a sparse
R = H H = 
H  . . . .  Toeplitz matrix. In equation (5), W H H r is only com-
 RN−1,1 · · · RN−1,N−1 RN−1,N 
  puted at the initial iteration. The matrix-vector multipli-
RN,1 · · · RN,N−1 RN,N cation W Rs̄ can be implemented with the complexity of
and Rm,n is a K × K Toeplitz matrix having [am,n (0), · · · , O(N(Nstage + 1)(2L + 1)) per symbol per iteration.
am,n (L)]T as its first column and am,n (0), · · · , am,n (−L) as Suppose the output of the detector (4) represents the
its first row. One of the special cases is Nstage = 0, and output of an equivalent AWGN channel having sk as its input
the detector is also called match filter. The normalization symbol, we may write
factor γ is introduced to ensure convergence of the sum in
ŝk = µk sk + ñk , (6)
(3) and is bounded by γ < 2/λmax (R). To maximize the
convergence speed, the optimal value of γ is [5] where µk is the equivalent amplitude of the output signal
γopt = 2/[λmax (R) + λmin (R)]. and ñk ∼ N(0, σ2ñ,k ) is a complex Gaussian noise sample.
We may compute the parameters µk and σ2ñ,k as follows
Here, λmax (R) and λmin (R) represent the maximum and
minimum eigenvalues of R. Based on the Gershgorin cir- E [ ŝk | sk = α (d)]
cle theorem [6], the upper bound of the largest eigenvalue
k W H [E ( r| sk ) − H s̄ + Hek s̄k ]
= eH H
and the lower bound of the smallest eigenvalue can be given
by [5] = ρk α (d) = µk α (d) ,
  σ2ñ,k = Cov [ ŝk , ŝk | sk = α (d)]

 

 
 

NK

λmax (R)  max   [R]i, j  = R1 = R∞ , = eH
k WH
H
σ2 I + H(V − vk ek eH
k )H
H
HW ek
i∈(1,NK) 
 j=1 

    = eH
k W (σ R + RV R)W ek − vk ρk ,
2 2


 

 NK   
 

λmin (R)  max 
 
−

[R]  , 0
 
,


min 


[R] i,i i, j 

 

 where V
Cov(s, s) = diag(v1 , . . . , vKN ), d is an Mc × 1
i∈(1,NK)   
j=1, j
i vector of data bits (Mc is the number of bit per constellation
where [R]i, j denotes the (i, j)th entry of R, |·| denotes symbol), and α(d) denotes a constellation symbol which is
the complex modulus, and  · 1 and  · ∞ are the 1-norm obtained using Gray mapping.
and ∞-norm, respectively. Since only addition and mod- Without the existence of a priori information, we as-
ulus are involved, the complexity of this eigenvalues es- sume s̄ = 0, V = I (Assuming the mean energy per symbol
timation is about N 2 (2L + 1) complex multiplications and E s = 1). Let Lak,i , i = 1, 2, . . . , Mc be the a priori informa-
N 2 K (2L + 1) additions (according to the Toeplitz structure tion provided by the channel decoder for the kth symbol sk .
of Rm,n , the number of additions can be further reduced). Then, the mean and the variance of sk can be computed as
Therefore the overall order of the complexity of the PELD
s̄k = α (d) P [sk = α (d)],
is not increased.
d∈S

4. PELD-Based Turbo Receiver vk = |α (d)|2 P [sk = α (d)] − | s̄k |2 ,

d∈S
Similar to [1], [2], [7], the turbo receiver under considera-
tion is an iterative receiver which performs turbo detection where S is the set of 2 Mc bit vector d. The probability of the
symbol sk is the function of the a priori information Lak,i [2]
by passing extrinsic messages on the code bits between a
soft-input soft-output detector and a soft-input soft-output

Mc   
decoder. Here, we will focus on the soft-input soft-output 1 1 a
P [sk = α (d)] = 1 + d̃i tanh Lk,i ,
detector. The extrinsic messages are in log-likelihood ratio i=1
2 2
LETTER
3125


∆ +1, di = 1
where d̃i = . The extrinsic information de-
−1, di = 0
livered by the detector is [7]


exp −| ŝk − µk α (d) |2 /σ2ñ,k + 12 d˜T[i] l[i]
d∈Si,1
Lek,i = log 

exp −| ŝk − µk α (d) |2 /σ2ñ,k + 12 d˜T[i] l[i]
d∈Si,0
(7)

where Si,1 = {d|di = 1}, Si,0 = {d|di = 0}. d˜[i] denotes the
subvector of d˜ obtained by omitting its ith element d̃i , and
l[i] denotes the vector of all the a priori information values,
also omitting Lak,i .
The extrinsic information of the detector will be de- Fig. 1 BER curves of 1 × 1 system with QPSK transmission.
interleaved to be used as the input to decoder. The turbo
detection and decoding algorithms are described in detail in
[1], [2], [7].

5. The Comparison with the Known Schemes

In [2], [7], the block-wise MMSE-based turbo receiver is

proposed for block transmission systems. The output of the
soft detector can be expressed as
 −1  
ŝ = σ2 I + RV H H r − Rs̄ + diag(ρ)s̄, (8)

where
 −1
ρk = eH
k σ I + RV
2
Rek . (9)

Then, the computational complexity is about O(N 2 K 2 ) per Fig. 2 BER curves of 2 × 2 system with QPSK transmission.
symbol.
A simplification of the soft detector is obtained by re-
placing the MMSE filter by the matched filter, and produc-
ing the estimate of s as

ŝ = H H r − Rs̄ + diag (ρ) s̄, (10)

where ρk = eH k Rek . Here we can see that MF-based soft de-

tector is a special case of PELD-based detector with Nstage =
0.

6. Simulation Results

We evaluate the performance of the turbo receivers for con-

volutional coded SISO and MIMO (2 × 2) systems. The
channel model is according to the channel Model A speci-
fied by ETSI for HiperLAN/2. Data block length has been Fig. 3 BER curves of 2 × 2 system with 16-QAM transmission.
chosen to K = 128 and each block is padded by 8 zeros.
We choose a rate 1/2, memory 2 convolutional code with
generator (1 + D2 , 1 + D + D2 ), the length of information Figures 1, 2, 3 illustrate the BER performance before
bits Lc = 128, 256, 512 for 1 × 1 system with QPSK, 2 × 2 and after 5 iterations of turbo detection in 1 × 1 with QPSK,
system with QPSK and 2 × 2 system with 16-QAM, respec- 2 × 2 with QPSK and 2 × 2 with 16-QAM, respectively. Ob-
tively. The coded bits are permuted with a random inter- viously, the MMSE-based turbo receiver has the best per-
leaver. For each block we perform 5 iterations over the de- formance. The error floor of MF-based turbo receiver is
tection loop. The decoding is performed according to the very serous. However, PELD-based turbo receiver provides
maximum a posteriori algorithm [8]. much better performance than MF-based turbo receiver, and
 IEICE TRANS. FUNDAMENTALS, VOL.E88–A, NO.11 NOVEMBER 2005
3126

it also outperforms the conventional noniterative MMSE-

based receiver (i.e., an MMSE detector followed by a soft References
decision decoder). In both of the 1 × 1 and 2 × 2 systems
with QPSK transmission, the performance of PELD-based [1] X. Wang and V. Poor, “Iterative (turbo) soft interference cancellation
and decoding for coded CDMA,” IEEE Trans. Commun., vol.47, no.7,
turbo receiver follows that of MMSE-based turbo receiver pp.1046–1061, July 1999.
very closely even with one-stage polynomial expansion. For [2] M. Tüchler, A. Singer, and R. Koetter, “Minimum mean square error
2 × 2 system with 16-QAM transmission, compared with equalization using a priori information,” IEEE Trans. Signal Process.,
MMSE-based turbo receiver, the performance loss of the vol.50, no.3, pp.673–683, March 2002.
PELD-based turbo receiver with three-stage polynomial ex- [3] J. Boutros and G. Caire, “Iterative multiuser joint decoding: Unified
pansion (Nstage = 3) is no more than 1 dB at BER = 10−4 . framework and asymptotic analysis,” IEEE Trans. Inf. Theory, vol.48,
no.7, pp.1772–1793, July 2002.
[4] Z.D. Wang and G.B. Giannakis, “Wireless multicarrier communica-
7. Conclusions tions: Where Fourier meets Shannon,” IEEE Singnal Process. Mag.,
vol.47, no.3, pp.29–48, May 2000.
We propose a low complexity PELD-based turbo detector [5] G.M.A. Sessler and F.K. Jondral, “Rapidly converging polynomial ex-
for SISO and MIMO block transmission systems over fre- pansion multiuser detector with low complexity for CDMA systems,”
Electron. Lett., vol.38, pp.997–998, 2002.
quency selective channels. It leads to drastic complexity re- [6] C.D. Meyer, Matrix analysis and applied linear algebra, Society for
duction w.r.t the MMSE-based turbo detector but still offers Industrial and Applied Mathematics, 2001.
excellent performance. [7] D. Wang, J. Zhao, X. Gao, and X. You, “Space-time turbo detection
and decoding for MIMO block transmission systems,” IEEE Int. Conf.
Acknowledgment Commun. (ICC’04), pp.2914–2918, Paris, France, June 2004.
[8] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear
codes for minimizing symbol error rate,” IEEE Trans. Inf. Theory,
The authors would like to thank the anonymous reviewer for vol.20, no.3, pp.284–287, March 1974.
their helpful comments and thorough review.

View publication stats