SOFT DECISION FEEDBACK EQUALIZATION FOR PRE-CODED FIR-MIMO SYSTEMS
Zhi Tian
Department of Electrical and Computer Engineering
Michigan Technological University
Houghton, MI 49931, USA
ABSTRACT
In this paper, it is demonstrated that decision feedback
equalization (DFE) based on soft-decisions is able to out-
perform hard-decision DFE, with an attractive computational
cost that is polynomial in the number of symbols to be recov-
ered. A group of soft-decision feedback equalizers (SDFE)
with probabilistic data association are developed for multi-
input multi-output (MIMO) wireless communication systems
operating in time dispersive channels. Block transmissions
with zero-padded precoding are employed to guarantee chan-
nel identifiablity. Computer simulations indicate that SDFE
offers close-to-optimal error probability performance with-
out incurring exponential complexity. Further improvements
on these SDFE schemes promise enhanced performance at a
reduced computational cost.
Keywords: soft decision feedback equalization, MIMO sys-
tems, probabilistic data association.
1. INTRODUCTION
Enormous increase in bandwidth efficiency is promised by
the use of MIMO systems in wireless radio frequency links
[1]. In the case of frequency selective channels, orthogonal
frequency division multiplexing (OFDM) can be deployed
to transmit signals over frequency flat subcarriers. Alterna-
tively, channel-induced inter-symbol interference (ISI) can
be mitigated using high-speed equalization techniques, among
which decision feedback equalization (DFE) has been proved
very effective at low computational cost. Equalization in
MIMO systems is more challenging, however, because of
the need for signal detection in the presence of both mul-
tiple access interference (MAI) as well as ISI. MIMO DFEs
have been derived in [2], [3], [4]. In these schemes, tenta-
tive decisions on past symbols or symbols from MAI sources
are made by quantizing properly derived decision statistics.
Such a hard-decision based approach may suffer from catas-
trophic error propagation, and in most cases, incurs a non-
trivial performance gap relative to an optimal maximum like-
lihood (ML) detector in terms of bit error rate (BER) perfor-
mance.
0-7803-7551-3/02/$17.00 ©2002 IEEE
In this paper, we take on a soft decision approach to nar-
row the gap between DFE and ML detection for MIMO sys-
tems operating in frequency selective channels. We are in-
spired by the low-complexity soft-decision CDMA detector
introduced in [5]. It uses a multistage detection structure,
but replaces the intermediate finite alphabet symbol deci-
sions by soft decisions, expressed via their associated prob-
abilities. Such a soft-decision structure leads to significant
computation reduction when MAI is approximated to obey a
single-mode Gaussian probability distribution, an idea orig-
inated from the probability data association (PDA) filter [6].
The PDA detector provides near-optimal performance, with
overall complexity O(K 3 ), where K is the number of users.
We adopt this soft-decision approach and develop soft deci-
sion feedback equalization (SDFE) schemes for FIR-MIMO
channels in the quest for low-complexity equalizers with BER
performance competitive to that of ML detection.
The paper is organized as follows. In Section 2, soft
decision feedback equalization using PDA filtering is de-
rived for a single-input single-output (SISO) frequency se-
lective channel. The development of the SDFE scheme for a
SISO system (SISO-SDFE) not only provides an essential
component for building MIMO-SDFE, but also illustrates
a key feature of SDFE: the importance of block transmis-
sions with zero-padded (ZP) precoding. PDA filtering re-
quires a decorrelating preprocessor, which is infeasible for
rank-deficient channels. With ZP precoding, however, the
equivalent channels are guaranteed to be invertible [7], mak-
ing PDA-based SDFE applicable to all channel conditions.
In Section 3, the soft-decision equalization approach is ap-
plied to frequency-selective MIMO channels. Depending on
the desirable tradeoff between performance and complex-
ity, three soft-decision based algorithms are developed for
a general FIR MIMO model: (1) joint soft-decision equal-
ization and multiuser detection; (2) soft-decision feedback
equalization with hard-decision nulling-canceling multiuser
detection; (3) hard-decision feedback equalization with soft-
decision PDA multiuser detection. Computer simulations are
performed in Section 4 to compare the BER performance of
the proposed MIMO-SDFE schemes with that of ML detec-
tion, and a summary is given in Section 5.
594
 2. BLOCK SISO SDFE number of elements in b. To avoid this, we assume that b i ei
conditioned on y ~ is approximated as a uni-modal Gaussian
It has been argued that block equalization offers important random vector with matched mean and covariance, as sug-
advantages over serial equalization in terms of guaranteed gested in a PDA filter [5], [6]. Using the conditional means
channel identifiability and less performance degradation from
E fb jy~ g = 1
 + ( 1)
(1  ) = 2 1, we have
bj = j j j j
noise enhancement or error propagation [7]. We take the the following equalities
block equalization structure and start with a linear, disper-
sive, and noisy digital communication SISO system.
E fb e jy~ g = y~
gi =
X
(2j 1)ej


Denote h = [h(0);


; h(L)] as the symbol-spaced tap
i i (4)
j 6=
i

coefficients in the channel response, with L being the chan-
Covfb e jy~ g
Ri = i i
nel memory length. For block processing, we group the in- X

put time-serial b(n) into blocks of size N and define the k th = 4j (1 j )ej eTj + Rv~v~ (5)


transmitted block to be b k = [b(kN );


; b(kN + N 1)]T . j 6=
i

gi )T Ri 1 (bi ei
 1 (b e
To eliminate the inter-block interference, a zero-padding pre- exp 2 gi )
P r(bi ei jy
i i
~) =
(2 )N=2 jRi j1=2
coder can be used to insert guarding bits between successive (6)


data blocks. To do so, let s k = Pbk , where a guard-inserting P r(ei jy
= exp 2giT Ri 1 ei : (7)
i ~) 

=
matrix P = [ITN 0T(K N )N ]T is constructed by choosing 1 i P r( ei jy~)
K  N + L. The vector s k is a block of K symbols to
be transmitted over the channel, resulting in a received data Decision on bi can be made from  i accordingly. This ap-


block yk = [y (kK );


; y (kK+K 1)]T . The input-output proach is summarized into the following procedure for soft-
model of the block transmission system with ZP precoding decision DFE:
is given by [7]
yk = Hbk + vk ; (1) SISO-SDFE:

where H 2 RK N is a tall Toeplitz matrix made from h. 1. Initialize the probabilities as  j = 0:5; j = 1;


; N .
We assume that the noise vector v k has zero mean and tem-
2. For i = 1 : N , update g i and Ri from (4) and (5), and
poral covariance R vv . In this model, ISI only exists inside
update i based on (7) as follows:
each transmitted block, and does not propagate to adjacent
blocks. This enables the use of block-by-block processing 1
i = 1 :
1 + exp 2giT Ri 1 ei
without loss of optimality, at the expense of transmission re-  (8)
dundancy. We will henceforth drop the block index k for
notational simplicity.
3. Repeat (2) if necessary until all  i converge.
To construct a soft-decision feedback equalizer (SDFE),
we first perform a zero-forcing (ZF) operation on the re- 4. Make decisions on the block of symbols by
ceived block sequence y to obtain: 
1; i  0:5
^bi = i = 1;


; N:
y~ = H+ y = b + H+ v; (2) 1; i < 0:5;
(9)

where H+ denotes the Moore-Penrose psuedo-inverse of the Refinements discussed in [5] can be used to reduce the
channel matrix H, which is guaranteed to be invertible due computational cost of the above procedure. The overall com-
to the ZP precoding. Similar to [5], we write y
~ as plexity of SISO-SDFE is on the order of O(N 3 ).
X
N

y~ = bi ei + bj ej + v
~; (3) 3. MIMO SDFE
j =1 6=
;j i

We now consider a frequency-selective faded MIMO link

where bi is the ith element of b, e i is the ith column of the
N  N identity matrix I N , and the covariance of the noise
~ = H+ v becomes Rv~v~ = H+ Rvv (H+ )T . For
vector v
simplicity, a binary transmission is assumed, b k 2 f1gN .
In a soft decision approach, the finite alphabet symbol
bi is treated as a binary random variable, and is associated


with a probability  i = P r(bi = 1jy ~ ). Direct evaluation
of i from (3) incurs a computational cost exponential in the
with Ni inputs and No outputs. MIMO communications
arise in three situations: (i) multiple transmit and receive
antennas with a single user; (ii) multiple receive antennas
and a single transmit antenna with multiple-access users; (iii)
combination of both (i) and (ii). We will adopt a general
MIMO system model that describes all three cases. This
model considers each user at each transmitter as one input,
while each output refers to the signal received at each receive

595
antenna. Accordingly, the number of inputs is determined by
the number of transmit antennas and the number of multiple
access users as well. For convenience, we make no distinc-
tion among the three terms inputs, transmit antennas, and
users.
Similar to Section 2, a zero-padding precoder is employed
at each input. The channel response matrix between the sym-
bol block at the ith input and the j th output takes the same
form as the H matrix in (1), with each element denoted by
h(i;j ) (l), l = 0; 1;


; L, where L is the maximum length
of all the No Ni channel impulse responses. Each element in
(1) is then augmented into a block element, that is, y (i) and


b(i) are augmented into y(i) = [y (1) (i);


; y (No ) (i)]T 2
Joint SDFE/SDFD for FIR-MIMO channels:
The same as SISO-SDFE, except that (1) is replaced by (10),
and N is enlarged to N N i .
3.2. SDFE with nulling-canceling MUD
When multiple antennas are employed for high data-rate trans-
missions, the number of transmit and receive antennas N i
and No could be large. In [8], a MIMO system with a wire-
less spectral efficiency higher than 20 bps/Hz is demonstrated
using Ni = 8 and No = 12. In such applications, complex-
ity is heavily dependent on the multiuser detection schemes

RNo 1 and b(i) =
[b(1)(i);


; b(Ni ) (i)]T 2 RNi 1 , re-

used. The nulling-canceling (NC) algorithm is an attractive
low-complexity method that has been widely deployed for
spectively. In analogy to (1), the baseband equivalent MIMO the BLAST systems [8]. We propose to use the nulling-
signal model for block processing can be expressed in a canceling technique for multiuser detection over equalized
matrix-vector form as follows flat fading channels, coupled with SDFE within each user.
The overall receiver complexity of such
a MIMO system
is
y = Hb + v
reduced to the maximum of O N 3 Ni2 and O N 3 No2 .
(10)
2 3
H(0) 0 To make the MIMO system model amenable to single-
6 .. .. 7
6 . . 7 transmitter soft-decision equalization, we first diagonalize


H =
6
6 H(L)



7
H(0) 7 (11) the MIMO channel matrix H into a block-diagonal form.
6 7
6 .. .. 7 This can be achieved by premultiplying H with a row trun-
4 . . 5

cating matrix T0 = [IN No 0N No (K N )No ], followed by a
0 H(L) series of N No  N No matrices Ti in the form of
2 3
h(1 1) (l)



;
h(Ni ;1) (l) 2 3

6 7 IiNo 0




where H(l) = 4 .. ..
. . 5:
6 0 INo 7
h(1;No ) (l)


h(Ni ;No ) (l)

6
6
7
7
H(1)H+ (0)
.. ..
Ti = 6 . . 7;
6 7
Efficient soft-decision feedback equalization has been de- 6 .. 7
4 . I(N 1) 5
veloped for a dispersive SISO system. MIMO systems present
H(L)H+ (0)
i No

greater challenges because of the existence of capacity limit- 0

ing multiple access interference and the significant increase
T


T T y = I
H(0) b + Tv;
y~ = N 1 1 0} N (12)
in computational burden. While the soft decision approach | {z | {z }
promises near-optimal BER performance without incurring =
T =
H
intractable complexity, the hard-decision approach is attrac-
tive for its computational simplicity, which is typically on where the operator
denotes the Kronecker product. The
the order of O(N 2 ) in the SISO case. To balance the trade- transformed channel matrix H  is block diagonal. Such a
offs between performance and complexity for various MIMO block diagonalization process is equivalent to the decorre-
system setups, we develop three symbol detection and equal- lating preprocessor in SISO-SDFE. The non-diagonal matrix
ization schemes that blend the soft-decision and hard-decision inside each block, i.e., H(0), indicates the existence of non-
methods in different ways. orthogonal multi-access interference to be suppressed by a
MUD technique, namely nulling-canceling.
3.1. Joint soft-decision equalization and detection
The SISO-SDFE algorithm in Section 2 can be extended to The rest of the algorithm is described as follows:
the MIMO model in (10) in a straightforward manner. Such
an approach considers multiuser detection and channel equal- NC-MUD with SDFE for FIR-MIMO channels:
ization simultaneously, and treats both using soft-decision
feedback. The number of elements in the symbol block b is 1. Initialization for NC-MUD:
now N Ni , and
the computational complexity is on the order
of O (N Ni )3 accordingly. G1 = H+ (0);
i = 1:

596
 2. Recursion: 3.3. SDFD with hard-decision DFE

(a) User ordering: In MIMO system setups where the block size N is signifi-
cantly greater than the number of transmit and receive anten-
ki = arg min
2f 1



j= k ; ;ki 1g
jj(G ) jj2 :
i j (13) nas, the computational task on multiuser detection becomes
less stringent than block equalization. For such cases, we
may combine soft-decision feedback detection (SDFD) with
(b) Nulling: the conventional hard-decision feedback

equalizers. The over-
all complexity is now O N 2 Ni3 .
wki = (Gi )ki ; (14) Both zero-forcing and MMSE based hard-decision feed-
Wki = IN
wki ; (15) back equalization can be used. The following gives a hybrid
rki = Wki y~ i 2 RN 1 :
solution using hard-decision ZF-DFE and the soft-decision
(16)
PDA [5] method for MUD. It takes advantage of the block
Toeplitz structure of the MIMO channel matrix in (11).
(c) SISO-SDFE: note that r ki contains signal com-
ponents from the k i th transmitter only, SDFD-MUD with ZF-DFE for FIR-MIMO channels:

X
rki = el b(l ki ) + ej b(jki ) + v~ ;
N
For n = 1 : N
(17)
j =1 6=
;j l 1. Hard-decision feedback equalization of previously de-
tected symbols of all users:
where Rv~v~ = Wki TRvv TH WkTi .

y(n) X1
n

Apply Steps 1-4 of SISO-SDFE to Eqn. (17) to y (n) = H(m;n) b^ (m): (20)
(k )
obtain ^bj i , j = 1;


; N . m =1
(d) Cancellation: With correct past decisions, y
 (n) = H(0)b(n)+v(n).

X 2. Soft-decision PDA multiuser detection of symbols trans-

^b(ki ) (H
N

y~ i+1 = y~ i  )(j
j
j =1
Gi+1 = H+
(>ki ) (0);
i = i + 1:
As in [8], fk1 ;


; kNi g denotes a permutation of the in-
tegers 1;


; Ni specifying the order in which components
of the transmitted symbol vector b (i) are to be processed,
(Gi )j is the j th row of Gi , (H  )ki denotes the ki th column

of H, and H(>ki ) denotes the matrix obtained by zeroing
columns k1 ;


; ki of H.
Some observations on the NC-MUD/SDFE algorithm are
in order:
 Structurally, NC-MUD blends with SISO-SDFE nicely.
After the zero-forcing nulling in Step 2(b), the vec-
tor rki happens to possess the “decorrelated structure”
(17) required by the ensuing SISO-SDFE processor.
 Due to the preprocessing step (12), the structure of the
noise covariance matrix is changed and the ordering
algorithm in Step 2(a) is no longer optimal. Refine-
ments through noise whitening or other optimized user
ordering criteria must be devised in order to improve
the detection performance.
1) +
Ni ki ; (18) mitted at the nth time, starting with a preprocessor:

H+ (0)y(n) = b(n) + v~ (n):
(19) y~ (n) = (21)
Apply Steps 1-4 of SISO-SDFE to (21) to obtain the
^ (n) = [^b(1) (n);


; ^b(Ni ) (n)]T .
Ni 1 symbol vector b
End n.
4. SIMULATIONS
In the first example, we consider a single-user FIR fading
channel with a symbol-spaced memory length of L = 4.
Each FIR tap coefficient is Rayleigh faded. It is assumed to
remain constant over each block of N = 12 symbols, but is
allowed to change from block to block. SISO-SDFE is com-
pared against the conventional zero-forcing hard-decision feed-
back equalizer (DFE), zero-forcing linear filter (ZF-only),
and the optimal ML detector implemented using the sphere
decoding algorithm [9]. The BER performance versus signal-
to-noise ratio (SNR) of these receivers is depicted in Figure
1. It is shown that the performance of SISO-SDFE is close
to that of sphere decoded ML. The performance gap is negli-
gible in the low SNR region, but becomes more pronounced
for high SNRs. The hard-decision based DFE and ZF re-
ceivers are computationally simple (O(N 2 )), but suffer from
approximately 2 to 4 dB performance losses.

597
 0 0
10
−1
10
BER
−2
10
−3 soft−decision DFE
10 hard−decision DFE
ZF−only
Sphere decoding
−4
10
0 5 10 15
SNR (dB)
Fig. 1. FIR-SISO fading channel: L = 4; N = 12.
The second example studies multiuser detection and chan-
nel equalization in a frequency selective MIMO channel with
Ni = 2 transmitters and No = 4 receivers. The mem-
ory lengths of these 8 channels are assumed to be L = 8.
Over every block of N = 12 transmitted symbols, the tap
coefficients of each link are generated independently from
Rayleigh distributions. The BER curves of the three soft-
decision related algorithms developed in Section 3 are plot-
ted in Figure 2 with reference to that of the sphere decoded
ML. The suboptimal NC-MUD/SDFE and SDFD/ZF-DFE
algorithms exhibit performance gaps of about 2 dB from the
ML solution, but they are much more tractable in terms of
the computational load. Comparably, the joint SDFE/SDFD
algorithm offers closer to optimal BER performance, at the
expense of higher complexity.
5. SUMMARY
The soft-decision approach with probabilistic data associa-
tion is utilized to construct three types of low-complexity,
high-performance receivers for frequency selective MIMO
channels. These receivers balance between the performance
advantage offered by the soft-decision approach and the low
computational cost of the conventional hard-decision feed-
back approach. Their BER performance is close to but still
away from that of the optimal ML algorithm, with overall


complexity no greater than O (N  max(Ni ; No ))3 . Fur-
ther performance improvement can be expected with prudent
refinements over the current work.
10
−1
10
BER
−2
10
Joint SDFD / SDFE
−3
NC−MUD / SDFE
10
SDFD / ZF−DFE
Sphere decoded ML
−4
10
20 −8 −6 −4 −2 0 2 4 6 8 10
SNR (dB)
Fig. 2. FIR-MIMO fading channel: N i = 2, No = 4, L = 4,
N = 12.
6. REFERENCES
[1] G. J. Foschini, M. J. Gans, “On the limits of wireless commu-
nications in a fading environment when using multiple anten-
nas,” Wireless Personal Comm., no. 6, pp. 315-335, 1998.
[2] A. Duel-Hallen, “Equalizers for multiple-input multiple-
output channels and PAM systems with cyclostationary input
sequences,” IEEE J. Selected Area in Communications, vol.
10, pp. 630-639, Sept. 1992.
[3] N. Al-Dhahir, A. H. Sayed, “The finite-length multi-input
multi-output MMSE-DFE,” IEEE Trans. Signal Processing,
vol. 48, no. 10, pp. 2921-2936, 2000.
[4] A. Lozano, C. Papadias, “Layered space-time receivers for
frequency-selective wireless channels,” IEEE Transactions
on Communications,vol. 50, no. 1, pp. 65-73, Jan. 2002.
[5] J. Luo, K. R. Pattipati, P. K. Willett, F. Hasegawa, “Near-
optimal multiuser detection in synchronous CDMA using
probabilistic data association,” IEEE Communications Let-
ters, vol. 5, no. 9, pp. 361-363, 2001.
[6] Y. Bar-Shalom, X. R. Li, Estimation and Tracking: Princi-
ples, Techniques, and Software, Artech House, 1993.
[7] Z. Wang, G. B. Giannakis, “Wireless multicarrier communi-
cations - where Fourier meets Shannon,” IEEE Signal Pro-
cessing Magazine, pp. 29-48, May 2000.
[8] G. D. Golden, G. J. Foschini, R. A. Valenzuela, P. W. Wol-
niansky, “Detection Algorithm and Initial Laboratory Results
using the V-BLAST Space-Time Communication Architec-
ture,” Electronics Letters, Vol. 35, No. 1, pp. 14-15, 1999.
[9] U. Finkce and M. Pohst, “Improved methods for calculating
vectors of short length in lattice, including a complexity anal-
ysis,” Mathematics of Computation, vol. 44, pp. 463–471,
Apr. 1985.
598