Anda di halaman 1dari 10

2042 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO.

7, JULY 2009

Iterative MAP Equalization and Decoding in


Wireless Mobile Coded OFDM
Daniel N. Liu, Member, IEEE, and Michael P. Fitz

Abstract—Orthogonal frequency division multiplexing beginning of each OFDM symbol to eliminate intersymbol-
(OFDM) system suffers extra performance degradation in interference (ISI), orthogonality among the subcarriers ensures
fast fading channels due to intercarrier interference (ICI). inexpensive hardware implementations and makes OFDM the
Combining frequency domain equalization and bit-interleaved
coded modulation (BICM), the iterative receiver is able to prominent candidate for high data rates applications such
harvest both temporal and frequency diversity. Realizing that as: digital video broadcasting (DVB) [1], wireless local area
ICI channels are intrinsically ISI channels, this paper proposes a networks (WLAN) [2] and worldwide interoperability for
soft-in soft-out (SISO) maximum a posteriori (MAP) equalizer by microwave access (WiMAX) [3]. However, OFDM transmis-
extending Ungerboeck’s maximum likelihood sequence estimator sion in a wireless mobile environment with rapid channel
(MLSE) formulation to ICI channels. The SISO MAP equalizer
employs BCJR algorithm and computes the bit log-likelihood variation over each symbol period severely corrupts the or-
ratios (LLR) for the entire received sequence by efficiently thogonality among each subcarrier and give rises to intercarrier
constructing a trellis that takes into account of the ICI channel interference (ICI) [4]–[7] (and the references therein). The
structure. A reduced state (RS) formulation of the SISO MAP information theoretic studies of ICI bounds [4]–[7] spawned
equalizer which provides good performance/complexity tradeoff two lines of work: one which considered ICI suppression
is also described. Utilizing the fact that ICI energy is clustered
in adjacent subcarriers, frequency domain equalization is in the un-coded system [7]–[13] and one which considered
made localized. This paper further proposes two computational ICI suppression in the coded system [14]–[19]. This work is
efficient linear minimum mean square error (LMMSE) based concerned with the latter research paradigm where suppressing
equalization methods: recursive q-tap SIC-LMMSE equalizer ICI in the coded system.
and recursive Sliding-Window (SW) SIC-LMMSE equalizer
respectively. Simulations results demonstrate that the iterative The optimal receiver calls for maximum-likelihood (ML)
SISO RS-MAP equalizer achieves the performance of no ICI joint equalization and decoding using a hypertrellis which
with normalized Doppler frequency fd Ts up to 20.46% in is constructed from both the outer channel code and the
realistic mobile WiMAX environment. ICI channel trellis structure. This is clearly computational
Index Terms—Iterative decoding, turbo processing, intercar- infeasible. Owning to the concept of iterative turbo processing
rier interference, MAP equalization, OFDM. [20], a practical decoding strategy performs iterative process-
ing between two separate entities: front-end soft-in soft-out
I. I NTRODUCTION (SISO) equalizer and outer channel decoder. Huang et al. [17]
proposed a reduced complexity ML equalizer which works
O VER the past decade, there has been tremendous effort
to improve the performance and robustness of orthogo-
nal frequency division multiplexing (OFDM) system in wire-
for relative low normalized Doppler frequency by considering
only the main tap. Kim and Pottie [15] presents an one-shot
q-tap linear minimum mean square error (LMMSE) equalizer
less mobile environment. OFDM transmission enjoys a simple
with a (2q + 1) × K observation window, where K is the total
one-tap equalization in time-invariant channels due to the fact
number of subcarriers. While the iterative counterpart SIC-
that the Fourier basis forms an orthogonal eigenbasis for the
LMMSE equalizer can be found in [18] with a further reduced
channels. Provided that guard interval (GI) is inserted to the
window of (2q + 1) × (2q + 1). With q being small, the com-
Paper approved by A. Anastasopoulos, the Editor for Wireless Communi- plexity of O(K), but suffers extra performance degradation.
cation of the IEEE Communication Society. Manuscript received September Schniter [16] proposes an O(K) SIC-LMMSE equalizer at the
24, 2007; revised April 23, 2008.
D. N. Liu was with UnWiReD Laboratory, the Department of Elec- expense of time-domain preprocessing. In [19], Peng and Ryan
trical Engineering, University of California Los Angeles, Los Angeles, suggest two-stage equalizer: trellis-based symbol detector pre-
CA 90095, USA. He is now with the Electrical Engineering Department, ceded with ICI-suppressing demapper which the complexity is
University of Southern California, Los Angeles, CA 90089, USA (e-mail:
danielnl@usc.edu). O(Kq 2 + K23Mc ), which Mc is number of bits per symbol.
M. P. Fitz was with UnWiReD Laboratory, the Department of Electrical In this paper, an “optimal” symbol by symbol maximum
Engineering, University of California Los Angeles, Los Angeles, CA 90095, a posteriori (MAP) equalizer is deireved by constructing a
USA. He is now with Northrop Grumman Corp., Redondo Beach, CA 90278,
USA (e-mail: Michael.Fitz@ngc.com). trellis which taking into account of ICI channel structure [21]–
This work was supported in part by the grant from STMicroelectronics [23]. Once the trellis is defined, the MAP equalizer employs
Inc. with a matching grant from the University of California Discovery BCJR forward-backward algorithm [24] to compute the a
Program under Grant COM-0310142. The material in this paper was presented
in part at the IEEE International Symposium on Personal, Indoor and posteriori probability (APP) of each coded bit. Moreover,
Mobile Radio Communications, Athens, Greece, September 2007, in part at following the footsteps of Cai and Giannakis [7], recursive
Forty-Fifth Annual Allerton Conference On Communication, Control, and algorithms to update LMMSE equalizer coefficients for both q-
Computing, Monticello, Illinois, September 2007 and in part at IEEE Global
Communications Conference, Washington D.C., USA, November 2007. tap SIC-LMMSE equalizer and Sliding-Window SIC-LMMSE
Digital Object Identifier 10.1109/TCOMM.2009.07.070476 equalizer are derived which further reduce complexity without
0090-6778/09$25.00 
c 2009 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
LIU and FITZ: ITERATIVE MAP EQUALIZATION AND DECODING IN WIRELESS MOBILE CODED OFDM 2043

compromising performance. Section III derives the “optimum” MAP equalizer and RS-
By utilizing the property of ICI channel energy distribution, MAP equalizer. This section also introduces the low com-
the “optimal” front-end MAP equalizer can be realized effi- plexity recursive q-tap SIC-LMMSE equalizer and SWSIC-
ciently. The ICI channel is intrinsically a ISI channel which LMMSE equalizer and analyzes the computational complexity
is well studied by Forney and Ungerboeck [21], [22]. The of proposed algorithms. Section IV shows numerical example
received symbol sequence in frequency domain are correlated about deployment of proposed algorithms over WiMAX sys-
by ICI, symbol decisions ought to be based on the entire tem [3]. Section V concludes the paper.
received sequence. The fact that straight forward likelihood
computation grows exponentially with the message length
II. S YSTEM M ODEL
(i.e. total number of subcarrier K) has warranted a retreat to
symbol-by-symbol decisions [17], linear filtering [15], [18] or Fig. 1 depicts a single-input single-ouput coded OFDM
localized trellis-based soft demapper [19]. Due to majority of system. A set of K-coded QAM “frequency-domain” symbols
symbol energy concentrate in vicinity of the desired symbol, T
d = [d(1) . . . d(K)] forms the input to IFFT. This paper
the ICI channel memory length q is much less than K, that further assumes the average symbol energy Es ≡ E|d(k)|2 =
is: 1 and symbols are equally likely chosen from a complex
q  K. (1) constellation D with cardinality |D| = 2Mc . The time domain
waveform at the output of IFFT x(t) is then given by
The crucial observation in (1) cuts the intimidating likelihood
computation from O(2Mc K ) to O(2Mc 2q ) and creates a way 1 
K

to accumulate likelihood metrics recursively in the trellis x(t) = √ d(k)ej Ts kt , −Tg ≤ t ≤ Ts (2)
Ts k=1
[23]. In this paper, not only the “optimal” MAP equalizer is
derived, but also a Reduced-State MAP (RS-MAP) equalizer where Ts , and Tg are the OFDM symbol duration, and guard
inspired by [25] is offered which further reduce the likelihood interval length, respectively. Thus, one OFDM symbol block
computation to O(2Mc ). in time is Tb = Tg + Ts . This paper assumes a time-varying
Based on the fact that most of a symbol’s energy is (TV) wireless multipath channel with a impulse response:
distributed over neighboring subcarriers, a low complexity
mp −1
recursive q-tap SIC-LMMSE equalizer is derived. A direct 
computation of the LMMSE equalizer coefficient bears the h(t, τ ) = hn (t)δ(τ − τn ), (3)
complexity of O(K 3 ) by inverting a K × K channel matrix n=0

H has justified a localized LMMSE equalizer implementation where τ0 ≤ τ1 ≤ · · · ≤ τmp −1 with τn being the tap-
[7], [15]. As shown in [7], the q-tap equalizer coefficient can delay on nth tap, and hn (t) is the randomly time-varying tap
be computed recursively by traversing the diagonal of H with gains: hn (t) = αn (t)ejθn (t) , respectively. Moreover, h(t, τ ) is
a (2q + 1) × K observation window. Also incorporating the modeled as wide sense stationary uncorrelated-scattering (WS-
a priori information feed back from outer channel decoder, SUS) channel and the tap gains {hn (t)} are complex Gaussian
m p −1
the q-tap SIC-LMMSE equalizer is really just a generaliza- with zero mean and variance σn2 , where n=0 σn2 = 1. The
tion of q-tap LMMSE equalizer. In this paper, it’s shown autocorrelation of the WSSUS channel is
that the q-tap SIC-LMMSE equalizer can also be updated
recursively as done in the un-coded system. Comparing to E[h(t, τ1 )h(t + Δt, τ2 )† ] = Rh (Δt)φτ (τ1 )δ(τ1 − τ2 ), (4)
conventional method which requires matrix inversion with
complexity O(32 · q 2 K 2 ), the proposed recursive equalizer with the assumption that the angle of arrival of the received
only has complexity O(20 · qK 2 ) but without sacrificing signal waveform is a uniformly distributed random variable,
performance. E[h(t, τ1 )h(t+Δt, τ2 )† ] is separable in time and delay. In (4),
The novelty of Sliding-Window SIC-LMMSE (SWSIC- Rh (Δt) is the normalized time-correlation function and φτ (·)
LMMSE) equalizer lies in the equalization process. The com- is the delay power spectrum with σn2 ≡ φτ (τn ). The received
plexity of q-tap SIC-LMMSE equalizer is O(K 2 q) [7] and waveform y(t) depends on x(t) via,
remains too high for efficient hardware implementation. By mp −1

restricting the observation window into (2q + 1) × (4q + 1), y(t) = hn (t)x(t − τn ) + n(t) (5)
SWSIC-LMMSE equalizer computes the filter coefficients by n=0
sliding a smaller window along the diagonal of H. Recogniz-
ing the content of current observation window is only different where n(t) is AWGN. To recover the original frequency
from the previous one by last row and column, a recursive way domain message on mth subcarrier, FFT is performed
to update equalizer coefficients is derived in SWSIC-LMMSE  Ts
˜ 1 2π
equalizer. Different from the conventional Sliding-Window d(m) = √ y(t)e−j Ts mt dt. (6)
(SW) type of equalizer with complexity O(Kq 3 ). Recursive Ts 0
SWSIC-LMMSE equalizer consumes only O(Kq 2 ) and more Clearly, (6) can be rewritten as
importantly without compromising performance. With q  K 
and recursive updating, SWSIC-LMMSE truly achieves a ˜
d(m) = hm,m d(m) + hm,k d(k) +ñ(m) (7)
further reduced computational complexity of O(K). k=m
  
In Section II system model and notations are introduced. ICI

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
2044 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 7, JULY 2009

x(t) GI d mapper c BCC b


IFFT π
insertion μ encoder

Channel π(LA ) LA
h(t, τ ) π

y(t) GI d̃ LE π −1 (LE ) b̃
FFT Equalizer π −1 SISO BCC
removal decoder
AWGN n(t)

Fig. 1. A baseband equivalent model for coded OFDM system.

0
where hm,k is the (m, k)th element of H ∈ CK×K and fdTs = 2.27%
defined as -10 fdTs = 4.55%
mp −1 −j 2π kτn  Ts
 e Ts 2π
fdTs = 6.82%
hm,k ≡ hn (t)e−j Ts (m−k)t dt, (8) -20

Normalized ICI Power, dB


fdTs = 13.64%
n=0
T s 0
fdTs = 40.92%
-30
and ñ(m) is i.i.d. complex Gaussian random variables with
zero mean and variance N0 /2 per dimension. Realizing the -40
orthogonality among subcarriers
 Ts -50

e−j Ts (m−k)t dt = Ts δ(m − k), (9)
0 -60

one can easily see that k=m hm,k d(k) = 0 if hn (t) remains
-70
constant over Ts . On the other hand, ICI is produced when
hn (t) is varying. In a mobile OFDM environment over a -80
TV channel, ICI can be characterized by normalized Doppler -150 -100 -50 0 50 100 150
Frequency Spacing
frequency fd Ts , where fd is the maximum Doppler frequency.
Intuitively, one can think of the normalized Doppler frequency Fig. 2. Normalized ICI power distribution with K = 256.
as a maximum cycle change of the TV channel per OFDM
symbol in a statistical sense. In [4]–[7], the explicit mathe-
matical expression for ICI power is derived. As an example the desired subcarrier. For example, more than 98% of the
of the significance of ICI considered, RH (Δt) = J0 (2πfd Δt) symbol energy clusters in the neighboring 5 subcarriers for
where J0 (·) is the zeroth-order Bessel function of first kind, fd Ts = 13.64%. Due to the fact that normalized ICI power
the normalized ICI power of subcarrier m caused by subcarrier decreases exponentially as Δf increases, it is safe to assume
k is [7] that
 1
1 hm,k = 0, iff |m − k| > q (12)
γm,k ≡ 2 Sn (f )sin2 (πfd Ts f )
π 0 where q can be used as a parameter to trade-off performance

1 1 and complexity. Collecting (7) into matrix notation and em-


× + df, (10)
(fd Ts f + (m − k))2 (fd Ts f − (m − k))2 ploying (12),
where Sn (f ) is the normalized Doppler spectrum, and total d̃ = H̃d + ñ (13)
ICI power for subcarrier m is
with H̃ ∈ CK×K is a banded matrix with both upper and

K
lower band width equal to q[26].
γm = γm,k . (11) Approaching ML performance with reasonable complexity
k=1,k=m relies on iterative processing between equalization and decod-
Examining (10) and (11) reveals the symbol energy leakage ing. Analogous to a turbo decoder, the inner soft-in soft-out
caused by ICI is highly concentrated in the neighboring (SISO) equalizer and outer channel decoder can be regarded
subcarriers. Fig. 2 demonstrates the distribution of normalized as two elementary “decoders” [27] in a serial concatenation
ICI power as a function of frequency spacing (i.e. Δf = |m − architecture. As shown in Fig. 1, coded and interleaved
k|/Ts ) for different values of normalized Doppler frequency bit sequence c = [ c1 c2 . . . cKMc ]T is mapped into
fd Ts . As a result of increased fd Ts , more and more symbol frequency domain coded symbols d. The SISO equalizer
energy leaks into the neighboring subcarriers and produces takes the channel observation d̃ and a priori log-likelihood
increased ICI levels. It is evident from Fig.2 that most of ratio (LLR) LA (cl ) to computes the extrinsic information
the symbol energy concentrate around the neighborhood of LE (cl ) for each of KMc bits per received d̃. With cl = +1

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
LIU and FITZ: ITERATIVE MAP EQUALIZATION AND DECODING IN WIRELESS MOBILE CODED OFDM 2045

⎧ ⎧ ⎡ ⎤⎫⎫

⎪ ⎪
⎪ ⎪
⎪ ⎪
⎨K
1 ⎨ ⎢ 
m+q 2q
⎥⎬⎪⎬
P [d̃|d] ∼ exp R d(m)† ⎢2 h † ˜ − gm,m d(m) − 2
d(k) g d(m − k)⎥ (19)
⎪ N ⎪ ⎣ k,m m,m−k ⎦⎪⎪

⎩m=1 0 ⎪⎩ k=m−q, k=1, ⎭⎪
⎪ ⎭
1≤k≤K q>0

⎧ ⎧ ⎡ ⎤⎫⎫

⎪ ⎪
⎪ ⎪⎪
⎪ ⎪
1 ⎨⎢ ⎨ 
m+q
⎥⎬⎬ 2q

λm (Sm−1 , Sm ) = exp †⎢
R d(m) ⎣2 † ˜
hk,m d(k) − gm,m d(m) − 2 gm,m−k d(m − k)⎥ (21)
⎪ N ⎪ ⎦⎪⎪ ,

⎩ 0 ⎪⎩ k=m−q, k=1, ⎪⎭
⎭ ⎪
1≤k≤K q>0

representing a binary one and cl = −1 representing a binary where (·)† means complex conjugate-transpose, R(·) denotes
zero, LA (cl ) from outer channel decoder is defined as real part and gm,k is the (m, k)th element of G respectively.
P [cl = +1] The straightforward approach of computing P [d̃|d] via (18)
LA (cl ) ≡ ln (14) for all data sequence allowed is clearly impracticable due
P [cl = −1]
to computational complexity grows exponential in K. The
where l = 1, . . . , KMc . The a posteriori LLR LD (cl |d̃) for second term in (18) which refers to as energy correction term
bit cl , conditioned on received vector d̃ is similarly defined for each postulated data sequences d is essentially a sum of
as   individual entities like d(m)† gm,k d(k). Ungerboeck made a
P cl = +1|d̃ keen observation[22], [23]: d† Gd for each postulated data
LD (cl |d̃) ≡ ln   (15) sequence d only has to sum over where gm,k = 0. Following
P cl = −1|d̃
Ungerboeck’s formulation, (18) can be calculated via (19)
 
much more efficiently. Examining (19) closely reveals the like-
where P cl = ±1|d̃ is the a posteriori probability (APP)
lihood computation can not only be accumulated recursively,
of bit cl . “New” (extrinsic) information learned at the equal- more important the correlation length among the postulated
ization stage can easily be separated from a posteriori LLR data sequence d has been shurnk from K to 2q. Substituting
LD (cl ) by subtracting off the a priori LLR LA (cl ). That is, (19) into (17) and apart from constant multiplicative terms,
LE (cl ) = LD (cl |d̃) − LA (cl ). (16) the APP for coded symbols is
In view of (16), extrinsic information LE (cl ) is then fed into 
K

the outer channel decoder as a priori information on the coded P [d|d̃] ∼ λm (Sm−1 , Sm )P (Sm |Sm−1 ), (20)
bit cl . m=1

where λm (Sm−1 , Sm ) is the channel transition probability as


III. I TERATIVE E QUALIZATION AND D ECODING
defined in (21) and the a priori probability is
A. Reduced-State MAP Equalizer

Mc
Maximum a posteriori (MAP) equalizer optimally com- P (Sm |Sm−1 ) = P [cMc (m−1)+l ]. (22)
putes the APP for each coded and interleaved bit. Realizing l=1
the fact that there is a one-to-one mapping between the
coded and interleaved bit sequence c and the sequence of It is clear that the trellis constructed via (20) still has 2Mc 2q
frequency domain coded symbols d, the APP for the coded states, because the state Sm−1 at any “time” is being given
and interleaved sequence would indeed the same as the APP by the 2q most recent inputs:
for coded symbols: P [c|d̃] = P [d|d̃]. Assuming bits are Sm−1 =
independent because of interleaving and from Baye’s rule,  
 Mc d(m − 2q), d(m − 2q + 1), . . . , d(m − 1) ,
P [d̃|d] K l=1 P [cMc (m−1)+l ]
P [d|d̃] = m=1
. (17) where each input has 2Mc possibilities.
P [d̃]
The number of states in MAP equalizer is reduced through
It is helpful to define the state at (m−1)th subcarrier for some decision feedback [25]. By introducing past decision into the
finite q as Sm−1 = (d(m − 2q), . . . , d(m − 1)). channel metric computation (21), the number of trellis state
Due to the band diagonal structure of H̃, the likelihood is reduced to 2Mc . In state transition (Sm−1 , Sm ), RS-MAP
function P [d̃|d] can be computed efficiently. Owning to the equalizer makes hard decision on the last 2q − 1 symbols at
Gaussian-noise assumption and apart from constant propor- state Sm−1 . That is:
tionality, P [d̃|d] becomes
  Ŝm−1 =
1 2  
P [d̃|d] ∼ exp − d̃ − H̃d ˆ − 2q), d(m
ˆ − 2q + 1), . . . , d(m − 1) . (23)
N0 d(m
⎧ ⎡ ⎤⎫
⎨ 1 ⎬ Realizing path histories contain the surviving state sequences
∼ exp ⎣2R(d† H̃† d̃) − d† H̃† H̃ d⎦ (18)
⎩ N0    ⎭ leading to the current state, the hard decisions are obtained as
≡G feed back from path history associated with state Sm−1 .

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
2046 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 7, JULY 2009

To produce APP, MAP equalizer employs the well-known The optimal LMMSE equalizer coefficient wm is obtained by
BCJR forward-backward algorithm [24]. The a posteriori LLR solving (24) and is given by
of the coded and interleaved bits LD (cl |d̃) can be computed !−1
N0 †
as wm = I2q+1 + Hm Δm Hm hm , (27)
Es
LD (cl |d̃) ∼
 where the covariance matrix Δm is
(Ŝm−1 ,Ŝm )∈+1 αm−1 (Ŝm−1 )λm (Ŝm−1 , Ŝm )βm (Ŝm ) " 2 2 2 2
#
ln  , σd(1) σd(m−1) σd(m+1) σd(K)
(Ŝm−1 ,Ŝm )∈−1 αm−1 (Ŝm−1 )λm (Ŝm−1 , Ŝm )βm (Ŝm ) Δm = diag ,··· , , 1, ,··· , ,
Es Es Es Es
where αm (Ŝm ) and βm−1 (Ŝm−1 ) are recursively computed 2
and σd(k) , k = 1, 2, . . . , K with k = m, is the transmit symbol
as
 variance and generally can be computed as,
αm (Ŝm ) = αm−1 (Ŝm−1 )λm (Ŝm−1 , Ŝm ) 
2
σd(k) = ¯ 2 P [d(k) = d].
|d − d(k)| (28)
Ŝm−1 ∈Ŝ
 d∈D
βm−1 (Ŝm−1 ) = βm (Ŝm )λm (Ŝm−1 , Ŝm ) In (28), P [d(k) = d] is a priori symbol probability can be
Ŝm ∈Ŝ calculated from a priori information fed back from outer
with Ŝ denotes all the possible states. The cardinality of such channel decoder respectively [29].
a set |Ŝ| is 2Mc as shown in (23). This paper further assumes There is an interesting way to recursive update the q-
α0 (Ŝ0 ) is initialized to some known auxiliary starting state tap LMMSE filter wm in (27). In (26), it is crucial to
and βK (ŜK ) is equal probable among Ŝ. observe that the channel matrix Hm is only different from
the previous version, Hm−1 , in a simple row, i.e., Hm−1 =
$ %† $ † %†
B. Low Complexity Recursive q-tap SIC-LMMSE Equalizer vm−1 H̄†m−1 and Hm = H̄m−1 vm . Thus,
there ought to be a way to recursively update the current
Exploiting the fact that most of the ICI caused symbol equalizer coefficients, wm . The recursive update relationship
energy leakage is only distributed over the neighboring sub- between wm−1 and wm hinges upon rewriting wm as
carriers, the SIC-LMMSE equalizer requires a filter length !−1
much less than the number of total subcarriers K. In general, N0
wm = I2q+1 + Hm (Δm−1 + Ψm ) H†m hm
the SIC-LMMSE equalizer/detector consists of three distinct Es
stages of processing. These three three stages include: soft ⎛ ⎞−1
interference cancellation, LMMSE equalizing/filter and LLR ⎜N ⎟
computation respectively. This paper focuses on the devel- ⎜ 0 ⎟
= ⎜ I2q+1 + Hm Δm−1 H†m +Hm Ψm H†m ⎟ hm
opment of LMMSE equalizer stage and leaves details of ⎝ Es ⎠
  
other two stages to references [28], [29]. The q-tap SIC- R̃−1
m
LMMSE equalizer only takes observation of ±q subcarriers , -−1

m
around the desired subcarrier m. Given the observation model, = R̃−1
m + ψk,k hk h†k hm , (29)
the LMMSE equalization process is precisely stated in the k=m−1
following optimization problem:   
Rm
ˆ
minimize E|d(m) − d(m)| 2
(24) where
ˆ
subject to d(m) †
= wm dm ,
Ψm = Δm − Δm−1
where dm ∈ C(2q+1)×1 represents the channel observation af- " 2 2
#
σd(m−1) σd(m)
ter “parallel” soft interference cancellation. Thus, dm linearly = diag 0, . . . , 0, − 1, 1 − , 0, . . . , 0 .
depends on d̃m ∈ C(2q+1)×1 via Es Es


K Examining (29) reveals that R−1
m is a series of rank-one mod-

dm = d̃m − ¯ (25) ifications to R̃−1
m . Applying the degenerate matrix inversion
d(k)hk
k=1,k=m lemma, it has been shown that Rm can be obtained through
 T recursive updates from R̃m [29] via the following relationship:
where d̃m = d(m˜ − q), . . . , d(m
˜ + q) , d(k) ¯ is the symbol
mean calculated from a priori information and hm is the mth R(i)
m = Rm
(i−1)

. /. /†
column of the channel matrix Hm ∈ C(2q+1)×K respectively. ψi,i
. / R(i−1) hi R(i−1) hi , (30)
m m
As shown in Fig. 3, Hm is a sub-matrix of H in (8) and (i−1)
1 + ψi,i h†i Rm hi
defined as
⎡ ⎤ (0)
hm−q,1 · · · hm−q,K with initialization of Rm = R̃m . Thus, wm can be re-
⎢ .. .. .. ⎥ cursively calculated from R̃m . Obviously, now the question
⎢ . . . ⎥
⎢ ⎥ becomes: how to obtain R̃m from Rm−1 ? Since R−1m−1 , which
Hm = ⎢ ⎢ h · · · h ⎥.
⎥ (26)

m,1 m,K
⎥ is defined as
.. .. ..
⎣ . . . ⎦ N0
hm+q,1 · · · hm+q,K R−1m−1 = I2q+1 + Hm−1 Δm−1 H†m−1
Es

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
LIU and FITZ: ITERATIVE MAP EQUALIZATION AND DECODING IN WIRELESS MOBILE CODED OFDM 2047

Hm ∈ C (2q+1)×K
as in (26)

Hm ∈ C (2q+1)×(4q+1)
as in (36)

H ∈ C K×K as in (8)

Fig. 3. An illustration of recursive q-tap SIC-LMMSE equalizer and recursive SWSIC-LMMSE equalizer.

only differs from R̃−1


m by the first row and first column, there [30] only need to compute once for initialization, the recursive
should be a way to compute R̃m from Rm−1 recursively algorithm cycles through step 3 to 7 updating equalizer
as well. With this observation, it is useful to partition the coefficients from one to other. Due to the observation window
Hermitian matrix R−1m−1 into spans the entire H with length K as in (26), this translate to

a quite costly operation of O(Kq) in updating b̃m (i.e. step 5
θm−1 b†m−1
R−1
m−1 = , (31) of Table 1 in [30]). Thus, the over all computation complexity
bm−1 Am−1 of q-tap SIC-LMMSE equalizer is O(K 2 q).
where Am−1 ∈ C2q×2q , bm−1 ∈ C2q×1 and θm−1 is scalar
C. Low Complexity Recursive Sliding-Window SIC-LMMSE
respectively. Similarly, R̃−1
m is partitioned into

Equalizer
−1 Am−1 b̃m Further utilizing the structure of H in (8), Sliding-Window
R̃m = , (32)
b̃†m θ̃m SIC-LMMSE equalizer offers a more flexible detecting strat-
† egy. It is evident that the data subcarriers d correlates among
with b̃m = H̄m−1 Δm−1 vm and θ̃m = vm Δm−1 vm + N Es .
0

−1 each other explicitly through the energy correction term gm,k


Moreover, let the inverse of Rm−1 be partitioned the same
as shown in (19). If the ICI channel memory has length
way as in (31)

q, the correlation length among each data subcarriers is at
p p†m−1 most 2q. Instead of stretching the decision feedback obser-
Rm−1 = , (33) vation window to cover the entire H as in (26), SWSIC-
pm−1 Pm−1
LMMSE equalizer employs a much smaller window of size
where Pm−1 ∈ C2q×2q , pm−1 ∈ C2q×1 and p is a scalar. (2q + 1) × (4q + 1) as depicted in Fig. 3. More specifically,
It can be shown that the inverse of Am−1 is the Schur the channel matrix Hm ∈ C(2q+1)×(4q+1) in SWSIC-LMMSE
complement of p in Rm−1 [7], [26]. That is: equalizer becomes
⎡ ⎤
pm−1 p†m−1 hm−q,m−2q · · · hm−q,m+2q
A−1
m−1 = Pm−1 − . (34) ⎢ .. .. .. ⎥
p ⎢ ⎥
⎢ . . . ⎥
To this end, it is clear that R̃m is the just inverse of R̃−1
m by Hm = ⎢ ⎢ h m,m−2q · · · h m,m+2q
⎥.
⎥ (36)
definition and can be directly constructed from (32) ⎢ .. .. .. ⎥
⎣ . . . ⎦
−1

Am−1 + ηm um u†m ηm um hm+q,m−2q · · · hm+q,m+2q


R̃m = , (35)
ηm u†m ηm
The channel observation after “parallel” SIC dm ∈ C(2q+1)×1
. /−1
changes to
with um = −A−1 m−1 b̃m and ηm = θ̃ m − b̃†
A−1
m m−1 b̃m

m+2q
respectively. Therefore, R̃m can be computed from Rm−1 dm = d̃m − ¯
d(k)hk
through A−1m−1 as shown in (35). Table 1 in [30] summarizes k=m−2q,k=m
the recursive q-tap SIC-LMMSE equalizer algorithm.
with hk being the kth column of Hm in (36). The optimal
Being able to recursively update the LMMSE filter coeffi-
equalizer coefficient wm can still be computed using (27) but
cients, the computational complexity for q-tap SIC-LMMSE
with a minor modification of the covariance matrix Δm as
equalizer remains manageable. Retaining only the dominant
terms, the computational complexity of recursive q-tap SIC- Δm =
" 2 #
LMMSE equalizer is shown in right column of Table 1 in σd(m−2q) 2
σd(m−1) 2
σd(m+1) 2
σd(m+2q)
[30] in terms of real-number addition (RA) and real-number diag ,··· , , 1, ,··· , .
Es Es Es Es
multiplication (RM). Realizing steps 1 and 2 of Table 1 in

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
2048 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 7, JULY 2009

,
" #
-−1
N0 Θm βm Λm 0 Θ†m γm
wm = I2q+1 + 2 hm (37)
γ †m β †m α†m
σd(m+2q)
Es αm 0 Es
⎡ 2 2
⎤−1
σd(m+2q) σd(m+2q)
N0
Es I2q + Θm Λm Θ†m + β m β †m Θm Λm γ m + α†m β m
=⎣ 2
σd(m+2q)
Es Es
σ 2
⎦ hm
γ †m Λm Θ†m + Es αm β†m N0
Es + γ †m Λm γ m + d(m+2q)
Es |αm |2
  
≡Rm

,
" 2
#
-−1
γ †m−1 α†m−1 β †m−1
σd(m−1−2q)
N0 αm−1 0
Rm−1 = I2q+1 + Es . (39)
Es βm−1 Θm 0 Λm−1 γm Θ†m

!−1 !−1
N0 N0
Rm−1 → A−1
m−1 → I2q + Θm Λm−1 Θ†m → I2q + Θm Λm Θ†m → A−1
m → Rm (41)
Es Es
     
Bm−1 Bm

Realizing Hm in (36) only differs from previous channel Probably the most non-trivial link in (41) is the update from
observation Hm−1 by last row and last column, wm can be Bm−1 to Bm . From (41), Bm only differs from Bm−1 via a
computed in the following partitioned form as shown in (37) diagonal matrix Λm . Examining both Λm and Λm−1 reveals

−1 the fact that the two diagonal matrices only differ at two
Am bm
wm = hm . (38) places. That is:
b†m am
Similarly, Rm−1 can also be partitioned exactly the same way Ψm ≡ Λm − Λm−1
" #
shown in (39) as 2
σd(m−1) 2
σd(m)

−1 = diag 0, . . . , 0, − 1, 1 − , 0, . . . , 0 . (43)
am−1 b†m−1 Es Es
Rm−1 = . (40)
bm−1 Am−1
The crucial observation made in (43) enables an expression
with Am−1 , bm−1 and am−1 each defined as of Bm in terms of B−1
m−1 and offers a chance doing recursive
N0 update. Bm can be rewritten as
Am−1 ≡ I2q + Θm Λm−1 Θ†m
Es !−1
2 N0 †
σd(m−1−2q) Bm = I2q + Θm (Λm−1 + Ψm ) Θm
+ βm−1 β †m−1 , Es
Es , -−1
2 
m

σd(m−1−2q) −1
= Bm−1 + ψk,k θ k θ k
bm−1 ≡ Θm Λm−1 γ m−1 + α†m−1 β m−1 ,
Es k=m−1
2
σd(m−1−2q)
am−1 ≡ |αm−1 |2 + γ †m−1 Λm−1 γ m−1 . where ψk,k denotes (k, k)th element of Ψm and θ k is the kth
Es column of Θm ∈ C2q×4q . Following (30) and [29], Bm can
The LMMSE coefficients for SWSIC-LMMSE equalizer be obtained via recursive update on Bm−1 via:
can be also updated recursively. The heart of such recursive
update lies in the fact that Rm can be recursively computed B(i)
m = Bm
(i−1)

. /. /†
from Rm−1 without performing the actual inversion. The ψi,i
/ B(i−1)
. m θi B(i−1)
m θi , (44)
whole recursive update can be visualized in (41). Assuming (i−1)
1 + ψi,i θ†i Bm θ i
the knowledge of Rm−1 and it can be partitioned as shown
in (33), A−1
m−1 is then computed via (34). Realizing the fact (0)
with initialization of Bm = Bm−1 . Observing Am in (38) is
that Am−1 is just a rank-one modification to B−1 m−1 , that is: also a rank-one modification to B−1
m , that is:
, -−1
2
σd(m−1−2q) , 2
-−1
−1 −1 †
Am−1 = Bm−1 + β m−1 βm−1 . −1 −1
σd(m+2q) †
Es Am = Bm + βmβm .
Es
Thus, by applying the degenerate matrix inversion lemma [26],
Bm−1 is calculated as Therefore, A−1
m can be updated from Bm as
2
σd(m−1−2q)
A−1
m = Bm −
Bm−1 = A−1
m−1 2 −
σd(m−1−2q)
Es
 −1 × 2
σd(m+2q)
1+β†m−1 Am−1 β m−1 Es †
 −1 
Es
† 2
σd(m+2q)
(Bm βm ) (Bm βm ) . (45)
Am−1 β m−1 A−1
m−1 β m−1 . (42) 1+ Es β†m (Bm βm )

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
LIU and FITZ: ITERATIVE MAP EQUALIZATION AND DECODING IN WIRELESS MOBILE CODED OFDM 2049

TABLE I TABLE II
R ECURSIVE S LIDING -W INDOW SIC-LMMSE E QUALIZER A BBREVIATION FOR L EGENDS

Recursive Algorithm Computational Complexity MnUC2 MMSE w/ n taps equalizer, Un-Coded QPSK
1. Form R−1 ∼ 64q 3 RA + 64q 3 RM DnUC2 DFE + PIC w/ n taps equalizer, Un-Coded QPSK
m−1 as in (40).
2. Find Rm−1 by NICO4 no ICI, Coded OFDM 16QAM w/ Soft-Output
directly inverting as in (40). ∼ 48q 3 RA + 48q 3 RM RMnI4 Reduced-State MAP equalizer, n Iterations 16QAM
3. Compute A−1 ∼ 16q 2 RA + 24q 2 RM RSLmnI4 Recursive SIC-LMMSE equalizer w/ m taps,
m−1 as shown in (34).
n Iterations, 16QAM
4. Calculate Bm−1 as in (42). ∼ 32q 2 RA + 40q 2 RM
5. Obtain Bm via RSWSLmnI4 Recursive Sliding-Window SIC-LMMSE equalizer
w/ m taps, n Iterations, 16QAM
recursive update as shown in (44). ∼ 64q 2 RA + 80q 2 RM
6. Compute A−1 m as in (45). ∼ 32q 2 RA + 40q 2 RM
7. Construct Rm as shown in (46). ∼ 48q 2 RA + 56q 2 RM
D. Complexity Analysis
3500
10-tap SIC-LMMSE Equalizer
To demonstrate the effectiveness of the proposed equalizers,
Recursive 10-tap SIC-LMMSE Equalizer the actual number of floating-point operations (FLOP) per
SWSIC-LMMSE Equalizer w/ q = 10
3000
Recursive SWSIC-LMMSE Equalizer w/ q = 10 one OFDM block is measured and plotted. Fig. 4 illustrates
Reduced-State MAP Equalizer, 16QAM
Reduced-State MAP Equalizer, 64QAM the complexities of the various proposed front-end equalizers
2500 in terms of MegaFLOPs per second versus different number
MegaFLOPs/Second

of frequency subcarriers. The complexity of the RS-MAP


2000 equalizer is mainly a function of: number of trellis states
which is |Ŝ| = 2Mc and number of transitions to/from each
1500 state, equal to 2Mc . Thus, the overall complexity for RS-
MAP equalizer is O(16 · 22Mc K). Being able to recursively
1000 compute the equalizer coefficients in either q-tap SIC-LMMSE
equalizer or SWSIC-LMMSE results in an order of mag-
500 nitude savings in computational complexity as compared to
conventional methods. In particular, at K = 512, recursive
0
0 200 400 600 800 1000 1200 10-tap SIC-LMMSE equalizer only spends 67.2 MegaFLOPS
K, Number of subcarrier which is only 7.5% of computational effort of conventional
SIC-LMMSE equalizer; while, the recursive SWSIC-LMMSE
Fig. 4. Complexity of different proposed equalizers vs. number of subcarrier.
equalizer with q = 10 offers a computational savings by a fac-
tor of 5.2 comparing to the non-recursive ones. For example, at
K = 1024, recursive SWSIC-LMMSE equalizer with 10-tap
Finally, the construction of Rm (38) can be accomplished with only spends 44.2 MegaFLOPs/Second; while the conventional
the availability of A−1 −1
m . Given the partition of Rm in (38), SWSIC-LMMSE equalizer bears 229.4 MegaFLOPs/Second.
Rm can be directly constructed as This translates to a complexity saving of ∼ 81%.
−1

Am + ηm um u†m ηm um
Rm = , (46) IV. E XAMPLE OF A PPLICATION T O T HE W I MAX S YSTEM
ηm u†m ηm
This section presents computer simulation results of the pro-
 −1 posed front-end equalizers with application to IEEE 802.16e
with um = −A−1
and ηm = am −
m bm b†m A−1
m bm
mobile WiMAX standard. The number of subcarriers is as-
respectively. Table I gives a summarization of the recursive
sumed to be K = 256 and the length of guard interval is
Sliding-Window SIC-LMMSE equalizer.
Tg = Ts /4. The outer channel code is the de facto standard 64-
The computational complexity of SWSIC-LMMSE equal-
state rate-1/2 binary convolutional code (BCC) with polynomi-
izer is further reduced due to a smaller observation win-
als in octal notation (133, 171)8. With different BCC code rate
dow size. Retaining only the dominant terms, the computa-
and constellation mappings (i.e. QPSK, 16QAM and 64QAM),
tional complexity of Sliding-Window SIC-LMMSE equalizer
variety of spectral efficiencies which range from R = 1 bit
is shown in the right column of Table I in terms of real-
per channel use (BPCU) to R = 5 BPCU can be achieved.
number addition (RA) and real-number multiplication (RM).
This Chapter considers the COST typical urban (TU) channel
The recursive SWSIC-LMMSE equalizer only performs step
model [31] with independent Rayleigh faded rays. The ray’s
1 and 2 once for initialization and cycles step 3 through step
relative power and delay are:
7 to update each equalizer coefficient. Being able to recursive
calculate the equalizer coefficient, SWSIC-LMMSE equalizer PdB =
takes only O(432q 2 ) floating-point operations (FLOP) to [−4, −3, 0, −2.6, −3, −5, −7, −5, −6.5, −8.6, −11, −10],
detect each frequency domain subcarrier. On the contrary,
τμs = [0, 0.1, 0.3, 0.5, 0.8, 1.1, 1.3, 1.7, 2.3, 3.1, 3.2, 5] .
conventional schemes cost O(224q 3 ) FLOPs because of ma-
trix inversion. For example, the recursive SWSIC-LMMSE It is further assumed that perfect timing synchronization and
equalizer with 10-tap uses only 19% of computational effort perfect channel state information (PCSI) are used for the
as compare to conventional matrix inversion schemes. With iterative receiver.
q  K, the computational complexity for SWSIC-LMMSE Fig. 5 illustrates the performance of different purposed
equalizer is O(K). equalizers in the coded OFDM system. In the simulation the

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
2050 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 7, JULY 2009

0 0
10 10

-1 -1
10 10

-2 -2
10 10
BER

BER
-3 -3
10 10

M1UC2:[CG03] M1UC2:[CG03]
D1UC2:[CG03] D1UC2:[CG03]
M5UC2:[CG03] M5UC2:[CG03]
-4 -4
10 D5UC2:[CG03] 10 D5UC2:[CG03]
M25UC2:[CG03] M25UC2:[CG03]
D25UC2:[CG03] D25UC2:[CG03]
NICO4 NICO4
RM1I4 RM1I4
-5 -5
10 RM4I4 10 RM4I4
RSL51I4 RSL51I4
RSL54I4 RSL54I4
RSWSL51I4 RSWSL51I4
RSWSL54I4 RSWSL54I4
-6 -6
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Eb/N0, dB Eb/N0, dB

Fig. 5. BER Performance for reduced-state MAP equalizer, recursive SIC- Fig. 7. BER performance for reduced-state MAP equalizer, recursive SIC-
LMMSE equalizer and recursive SWSIC-LMMSE equalizer, K = 256, LMMSE equalizer and recursive SWSIC-LMMSE equalizer, K = 256,
fd Ts = 6.82%, 16QAM, 64-state rate-1/2 BCC with COST-TU channel. fd Ts = 20.46%, 16QAM, 64-state rate-1/2 BCC with COST-TU channel.

0
10

0
10

-1
10
-1
10

-2
10
-2
10
BER

-3
10 -3
10
BER

M1UC2:[CG03]
D1UC2:[CG03]
M5UC2:[CG03] -4
-4
D5UC2:[CG03] 10 M1UC2:[CG03]
10
M25UC2:[CG03] D1UC2:[CG03]
D25UC2:[CG03] M5UC2:[CG03]
NICO4 D5UC2:[CG03]
RM1I4 -5 M25UC2:[CG03]
10
-5
RM4I4 D25UC2:[CG03]
10
RSL51I4 NICO4
RSL54I4 RM1I4
RSWSL51I4 -6 RM4I4
10 RSL51I4
RSWSL54I4
-6 RSL54I4
10
0 5 10 15 20 25 30 RSWSL51I4
RSWSL54I4
Eb/N0, dB -7
10
0 5 10 15 20 25 30
Eb/N0, dB
Fig. 6. BER performance for reduced-state MAP equalizer, recursive SIC-
LMMSE equalizer and recursive SWSIC-LMMSE equalizer, K = 256,
fd Ts = 13.64%, 16QAM, 64-state rate-1/2 BCC with COST-TU channel. Fig. 8. BER performance for reduced-state MAP equalizer, recursive SIC-
LMMSE equalizer and recursive SWSIC-LMMSE equalizer, K = 256,
fd Ts = 40.92%, 16QAM, 64-state rate-1/2 BCC with COST-TU channel.

number of subcarriers K = 256, the normalized Doppler


frequency fd Ts = 6.82%, 16QAM modulation, and the
standard 64-state BCC with rate-1/2 are used. Curves with
RS-MAP equalizer with iteration achieves the no ICI bound
up to 4 iterations are shown. Legends are abbreviated and
as indicated in Fig. 7.
the full meanings can be found in Table II. Being able to
harvest diversity provided by both of the ICI-channel and outer Fig. 8 shows the performance of different purposed equal-
channel code, the iterative receiver even provides a gain of izers in the coded OFDM system with normalizer Doppler
∼ 0.6 dB at BER of 10−4 compared to no ICI. frequency of fd Ts = 40.92%. For RS-MAP equalizer, an
Fig. 6 and Fig. 7 show the performance of different pur- error floor develops at BER = 10−4 due to the residual ICI.
posed equalizers in the coded OFDM systems with higher Realizing the RS-MAP equalizer is essentially a 1-tap decision
normalizer Doppler frequency of fd Ts = 13.64% and fd Ts = feedback sequence estimator (DFSE) [25], it is certainly
20.46% respectively. As the normalized Doppler frequency possible to remove the error floor by incorporating higher
increases, the performance of single-shot decoding pass de- taps at the cost of overall system complexity. On the other
grades due to higher ICI in the system. On the other hand, hand, being able to take advantage of the temporal diversity
as SNR increases the system becomes interference-limited, introduced by ICI, the LMMSE based equalizers exhibit a
so that performance gain by doing iterations increases. At strong immunity to intercarrier interference and provides a
normalized Doppler frequency fd Ts as high as 20.46%, the gain of ∼ 1.0 dB at BER of 10−4 compared to no ICI.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.
LIU and FITZ: ITERATIVE MAP EQUALIZATION AND DECODING IN WIRELESS MOBILE CODED OFDM 2051

V. C ONCLUSION [19] F. Peng and W. E. Ryan, “A low-complexity soft demapper for OFDM
fading channels with ICI," in Proc. IEEE Wireless Commun. Networking
Iterative equalization and decoding of the wireless mobile Conf., 2006, pp. 1549-1554.
coded OFDM system is considered. A RS-MAP equalizer [20] J. Hagenauer, “The turbo principle: tutorial introduction and state of
along with two computational efficient recursive LMMSE the art," in Proc. International Symposium Turbo Codes Related Topics,
based equalizer are derived which provides excellent perfor- Brest, France, Sept. 1997, pp. 1-11.
[21] G. D. Forney, “Maximum-likelihood sequence estimation of digital
mance and complexity trade-off. The performance of the dif- sequences in the presence of intersymbol interference," vol. 18, pp. 363-
ferent proposed equalizer has been evaluated over the COST- 378, May 1972.
TU channel model with application to WiMAX. The results [22] G. Ungerboeck, “Adaptive maximum-likelihood receiver for carrier-
modulated data-transmission systems," IEEE Trans. Commun., vol. 22,
suggest that iterative processing at the receiver end allows full pp. 624-636, May 1974.
exploitation of both temporal and frequency diversity available [23] A. M. Tonello, “MIMO MAP equalization and turbo decoding in
in a spectrally efficient system. interleaved space-time coded systems," IEEE Trans. Commun., vol. 51,
pp. 155-160, Feb. 2003.
[24] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of
R EFERENCES linear codes for minimizing symbol error rate," IEEE Trans. Inform.
[1] ETSI EN 300 744 v1.5.1 (2004-2006), “Digital video broadcasting Theory, vol. 20, pp. 248-287, Mar. 1974.
(DVB); framing structure, channel coding and modulation for digital [25] M. V. Eyuboǧlu and S. U. H. Qureshi, “Reduced-state sequence estima-
terrestrial television," European Broadcasting Union, Tech. Rep., 2006. tion with set partitioning and decision feedback," IEEE Trans. Commun.,
[2] IEEE Std. 802.11a-1999, “Part 11: wireless LAN medium access control vol. 36, pp. 13-20, Jan. 1988.
(MAC) and physical layer (PHY) specification: high speed physical layer [26] G. H. Golub and C. F. Van Loan, Matrix Computations. The John
in the 5 GHz band," IEEE-SA Standards Board(1999-09-16), Tech. Rep., Hopkins University Press, 1989.
1999. [27] C. Berrou and A. Glavieux, “Near optimum error correcting coding
[3] IEEE Std. 802.16e-2005 and IEEE 802.16-2004/Cor1-2005, “Part 16: and decoding: turbo-codes," IEEE Trans. Commun., vol. 44, no. 10, pp.
air interface for fixed and mobile broadband wireless access systems," 1261-1271, Oct. 1996.
IEEE-SA Standards Board(2006-02-28), Tech. Rep., 2006. [28] X. Wang and H. V. Poor, “Iterative(turbo) soft interference cancellation
[4] M. Russell and G. L. Stüber, “Interchannel interference analysis of and decoding for coded CDMA," IEEE Trans. Commun., vol. 47, pp.
OFDM in a mobile environment," in Proc. IEEE Veh. Technol. Conf., 1046-1061, July 1999.
1995, pp. 820-824. [29] D. N. Liu and M. P. Fitz, “Low complexity affine MMSE detector
[5] P. Robertson and S. Kaiser, “The effects of doppler spreads in OFDM(A) for iterative detection-decoding MIMO-OFDM systems," IEEE Trans.
mobile radio systems," in Proc. IEEE Veh. Technol. Conf., 1999, pp. Commun., vol. 56, pp. 150-158, Jan. 2008.
329-333. [30] ——, “Recursive LMMSE-based equalization in wireless mobile coded
[6] Y. G. Li and L. J. Cimini, “Bounds on the interchannl interference of OFDM," in Proc. IEEE Int. Symp. Personal, Indoor Mobile Radio
OFDM in time-varying impairments," IEEE Trans. Commun., vol. 49, Commun., 2007.
pp. 401-404, Mar. 2001. [31] Commission of the European Communities, “Digital Land Mobile
[7] X. Cai and G. B. Giannakis, “Bounding performance and suppressing Radio Communications - COST 207, Final Report," Office Official
intercarrier interference in wireless mobile OFDM," IEEE Trans. Com- Publications European Communities, Luxembourg, Tech. Rep., 1989.
mun., vol. 51, pp. 2047-2056, Dec. 2003.
[8] J. Armstrong, P. M. Grant, and G. Povey, “Polynomial cancellation Daniel N. Liu (S’03-M’08) received the B.S.E.E.
coding of OFDM to reduce intercarrier interference due to Doppler degree (magna cum laude) from the University of
spread," in Proc. IEEE Global Telecommunications Conf., vol. 5, 1998, California Los Angeles, Los Angeles, in 2003. He
pp. 2771-2776. received the M.S. degree in electrical engineering in
[9] Y. Zhao and S.-G. Häggman, “Intercarrier interference self-cancellation 2005, and the Ph.D. degree in electrical engineering
scheme for OFDM mobile communication systems," IEEE Trans. Com- in 2008 from the University of California Los An-
mun., vol. 49, pp. 1185-1191, July 2001. geles (UCLA), Los Angeles. He was a recipient of
[10] W. G. Jeon, K. H. Chang, and Y. S. Cho, “An equalization technique the UCLA Alumni Scholarship in 2001 and Frank
for orthogonal frequency-division multiplexing systems in time-variant Young scholarship in 2002. Since April 2008, he
multipath channels," IEEE Trans. Commun., vol. 47, pp. 27-32, Jan. has been a Post Doctoral Research Associate in
1999. the Electrical Engineering Department, University
[11] Y.-S. Choi, P. J. Voltz, and F. A. Cassara, “On channel estimation and of Southern California, Los Angeles. His research interests are in the area of
detection for multicarrier signals in fast and selective Rayleigh fading physical layer communication theory, information theory and coding theory
channels," IEEE Trans. Commun., vol. 49, pp. 1375-1387, Aug. 2001. with particular focus on: detection and estimation theory, equalization, channel
[12] A. Gorokhov and J.-P. Linnartz, “Robust OFDM receivers for dispersive estimation and space-time coding theory.
time-varying channels: Equalization and channel acquisition," IEEE
Trans. Commun., vol. 52, pp. 572-583, Apr. 2004. Michael P. Fitz received his B.E.E. degree (summa
[13] Y. Mostofi and D. C. Cox, “ICI mitigation for pilot-aided OFDM mobile cum laude) from the University of Dayton, Dayton,
systems," IEEE Trans. Wireless Commun., vol. 4, pp. 765-774, Mar. Ohio, in 1983 and his MS and Ph.D. degrees in
2005. electrical engineering from the University of South-
[14] Y. H. Kim, I. Song, H. G. Kim, T. Chang, and H. M. Kim, “Performance ern California in 1984 and 1989, respectively. From
analysis of a coded OFDM system in time-varying multipath Rayleigh 1983-1989 he worked as a communication systems
fading channels," IEEE Trans. Veh. Technol., vol. 48, pp. 1610-1615, engineer for Hughes Aircraft and TRW Inc. In 1989
Sept. 1999. he ventured into academia and was faculty at Purdue
[15] S. Kim and G. J. Pottie, “Robust OFDM in fast fading channels," in University, the Ohio State University (OSU), and the
Proc. IEEE Global Telecommun. Conf., 2003, pp. 1074-1078. University of California Los Angeles.
[16] P. Schniter, “Low-complexity equalization of OFDM in doubly selective Dr. Fitz is currently employed at the Northrop
channels," IEEE Trans. Signal Processing, vol. 52, pp. 1002-1011, Apr. Grumman Corporation as a senior systems engineer working on satellite
2004. communications. Dr. Fitz’s research is in the broad area of statistical commu-
[17] D. Huang, K. B. Letaief, and J. Lu, “Bit-interleaved time-frequency nication theory and experimentation. Dr. Fitz is the author of the textbook,
coded modulation for OFDM systems over time-varying channels," Fundamentals of Communications Systems. He was a recipient of the 2001
IEEE Trans. Commun., vol. 7, pp. 1191-1199, July 2005. IEEE Communications Society Leonard G. Abraham Prize Paper Award in
[18] R. Chen, Y. Xu, H. Zhang, and H. Luo, “Iterative ICI mitigation method the Field of Communications Systems. Prof. Fitz’s research group at UCLA
for MIMO OFDM systems," Institute Electronics, Info. Commun. En- currently is interested in the theory of space-time modems and operates an
gineers Trans. Comm., vol. E89-B, No. 3, pp. 859-866, Mar. 2006. experimental wireless wide area network and a space-time coding testbed.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on August 2, 2009 at 04:18 from IEEE Xplore. Restrictions apply.

Anda mungkin juga menyukai