2158 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO.

4, JULY 2007
Iterative Methods for Cancellation of Intercarrier
Interference in OFDM Systems
Andreas F. Molisch, Fellow, IEEE, Martin Toeltsch, Member, IEEE, and Sameer Vermani
Abstract—We consider orthogonal frequency-division multi-
plexing systems with intercarrier interference (ICI) due to in-
sufficient cyclic prefix and/or temporal variations. Intersymbol
interference (ISI) and ICI lead to an error floor in conventional
receivers. We suggest two techniques for the equalization of ICI.
The first, called “operator-perturbation technique” is an iterative
technique for the inversion of a linear system of equations. Alter-
natively, we show that serial or parallel interference cancellation
can be used to drastically reduce the error floor. Simulations show
that, depending on the SNR and the origin of the ICI, one of the
schemes performs best. In all cases, our schemes lead to a drastic
reduction of the bit error rate.
Index Terms—Cyclic prefix (CP), intercarrier interference
(ICI), interference cancellation, intersymbol interference (ISI),
iterative algorithms, orthogonal frequency-division multiplexing
(OFDM).
I. INTRODUCTION
O
RTHOGONAL frequency-division multiplexing
(OFDM) transmits data simultaneously on a number of
subcarriers, which, together, form OFDM symbols [1]–[4].
The frequencies of the subcarriers are chosen in such a way
that the signals on different subcarriers are orthogonal to each
other, even though their spectra overlap. This principle causes
an OFDM symbol to be much longer than a data symbol in a
single-carrier system (for equal data rates) and, thus, makes the
system more robust to the delay dispersion of the channel. An
efficient implementation of OFDM [with or without a “cyclic
prefix” (CP)] can be obtained by a blockwise inverse fast
Fourier transform (IFFT) of the input signal. For these reasons,
OFDM is particularly suitable for high-data-rate transmission
and has received great attention for wireless LANs (IEEE
802.11a and g standard [5]), as well as wireless personal
area networks (multiband-OFDM standard ECMA-368 [6])
and fixed wireless access (IEEE 802.16, WiMax [7]). In
Manuscript received November 28, 2004; revised March 18, 2006 and
May 24, 2006. Parts of this work were presented at IEEE PIMRC2000, London,
U.K., and IEEE ICC 2001, Helsinki, Finland. The works of A. F. Molisch and
M. Toeltsch were supported, while at the Technical University of Vienna, by
the Austrian Science Fund (FWF) under Contract FWF525/P12984-PHY. The
review of this paper was coordinated by Prof. Z. Wang.
A. F. Molisch is with the Mitsubishi Electric Research Laboratories,
Cambridge, MA 02139 USA, and also with the Department of Elec-
troscience, Lund University, SE-22100 Lund, Sweden (e-mail: Andreas.
Molisch@ieee.org).
M. Toeltsch is with the SYMENA Software and Consulting Ltd., A-1040
Vienna, Austria (e-mail: Martin.Toeltsch@symena.com).
S. Vermani is with Qualcomm Inc., San Diego, CA 92121-1714 USA
(e-mail: svermani@qualcomm.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2007.897628
order to avoid intersymbol interference (ISI) and intercarrier
interference (ICI), the CP, i.e., a copy of the last part of the
OFDM symbol, must be prepended to the transmitted symbol;
alternatively, a zero-suffix can be used [8], [9]. The duration
of this prefix has to be at least as large as the maximum
excess delay of the channel in order to retain orthogonality
between the subcarriers; furthermore, it is required that the
channel does not change during the transmission of one OFDM
symbol. If those conditions are fulfilled, then the receiver can
compensate for the channel distortions by a one-tap equalizer
in the frequency domain.
In many practical systems, the orthogonality between the
subcarriers is destroyed by two effects.
1) Temporal variations of the channel: Those variations lead
to ICI (loss of orthogonality between the subcarriers) due
to the Doppler shift associated with the temporal varia-
tions. Temporal variations of the channel are unavoidable
in wireless systems—even if both the transmitter and the
receiver are stationary, moving scatterers lead to Doppler
shifts of the incoming waves. An important issue in that
context is the correct estimation of the channel at any
given time; this issue is discussed, e.g., in [10]–[12] and
is considered to be perfectly solved for the remainder
of this paper. Similarly, oscillator frequency offset and
frequency drift can lead to ICI; also, this problem is
assumed to be solved for the purpose of this paper (see,
e.g., [13] and [14]).
2) Insufficient length of the CP: If the CP is shorter than
the maximum excess delay of the channel, ISI occurs,
i.e., each OFDM symbol affects the subsequent symbol;
this effect can be eliminated by decision feedback [15],
[16] and will be disregarded henceforth. Furthermore, the
insufficient CP also leads to ICI, as the original symbols
cannot be reconstructed by means of a one-tap equalizer
alone. An insufficient duration of the CP can arise for
various reasons. A system might consciously shorten or
omit the CP in order to improve the spectral efficiency. In
other cases, a system was originally designed to operate
in a certain class of environments (and, thus, a certain
range of excess delays) and is later on deployed also
in other environments that show a larger excess delay.
Finally, for many systems, the length of the CP is a com-
promise between the desire to eliminate ISI and to retain
spectral efficiency—in other words, a CP should not be
chosen to cope with the worst-case channel situation, as
this would decrease the spectral efficiency in the typical-
case situation.
0018-9545/$25.00 © 2007 IEEE
MOLISCH et al.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2159
Irrespective of the reasons, ICI (and ISI) can lead to a con-
siderable performance degradation in many systems and must
be combated. A large range of techniques has been developed
in recent years and can be classified as follows.
1) Optimum choice of the carrier spacing and OFDM sym-
bol length: By trading off ICI caused by temporal varia-
tions and by delay dispersion, the bit error rate (BER) can
be minimized without requiring any structural changes
in either the transmitter or receiver [17]. In a related
approach, the basic pulseshape for the signaling of the
subcarriers can also be modified [18], [19].
2) Self-interference-cancellation techniques: In this ap-
proach, the information is modulated not just onto a
single subcarrier but onto a group of them, which leads to
a strong reduction in the self-interference. The technique
was suggested in [20] and later extended in [21]–[23].
This technique is very effective for the mitigation of ICI
but leads to a reduction of the spectral efficiency of the
system.
3) Temporal equalizers: In these techniques, the equaliza-
tion is done in the time domain, before the FFT at the
receiver (with a possible feedback that undergoes back-
and-forth transformations) [24], [25]. Reference [26]
found the coefficients for a temporal filter that max-
imizes the SINR for both conventional OFDM and
multiple-input–multiple-output (MIMO)-OFDM. An ef-
ficient scheme called RISIC, based on tail cancellation
and cyclic reconstruction, was first suggested in the study
in [15] and improved in the study in [27]. In a similar
approach, channel shortening can be used to make the
effective impulse response shorter than the CP [28].
4) Techniques that depend on the presence of unused sub-
carriers (null carriers): In this approach, redundancy is
placed in the frequency domain by having more carriers
than is strictly necessary. Similarly to the principle of the
CP, which allows elimination of the ISI by redundancy in
the time domain, the redundancy in the frequency domain
allows elimination of the ICI [29]–[31].
5) Forward-error correction [32]: The capability of any
forward-error-correcting code can be used to eliminate
the errors caused by the ICI, and correlative coding [33]
can improve the ICI as well.
An area of particular interest has been linear equalization in
the frequency domain. The impact of the channel can be de-
scribed by a transfer-function matrix, which is inverted explic-
itly or implicitly. Equalization of ICI due to temporal variations
of the channels has been studied in [11], [12], and [34]–[39];
most of the equalization strategies suggested in these papers
rely on the special structure of the transfer-function matrix in
the case of pure Doppler-induced ICI (i.e., with sufficient guard
interval). Reference [38] also provides a matched-filter bound
for time-varying frequency-selective channels for the case that
the CP is sufficiently long. Reference [40] showed that the
elimination of ISI due to time variations is dual to the temporal
equalization in single-carrier systems. Equalization of ICI due
to insufficient guard intervals, possibly in combination with
decision-feedback equalization (DFE), was studied in [16];
moreover, various algorithms have been considered for the
matrix inversion. Direct inversion of the channel matrix [either
according to a zero-forcing or an minimum mean-squared error
(MMSE) criterion] is computationally too expensive. For wire-
line communications, Bogucka and Wesolowski [41] advocated
the use of the LMS algorithm to find the coefficients in the
frequency domain; however, this approach is not feasible in
time-variant wireless channels.
In this paper, we therefore suggest an iterative inver-
sion technique for the general case of time-varying delay-
dispersive channels with insufficient CP. We also present a
serial-interference-cancellation (SIC), as well as a parallel-
interference-cancellation (PIC) scheme that shows rapid con-
vergence and lower error floor for some ranges of delay spread.
Since 2000, when we first suggested these algorithms in our
conference papers [42], [43], interest in this approach has
greatly increased and various related schemes have been pub-
lished. Reference [44] uses a Gauss–Seidel iteration for the
inversion of the channel matrix; [45] and [46] suggested SIC
schemes and analyzed their performance.
The remainder of this paper is organized the following way:
in Section II, we describe the system model as well as the chan-
nel characteristics determining the system behavior. Section III
is dedicated to the “operator-perturbation technique” (OPT)
for the efficient inversion of matrices. Next, we investigate
interference-cancellation techniques, covering both PIC and
SIC. Section V gives numerical results for both the OPT and
the interference cancellation. A summary and the conclusions
wrap up this paper.
II. SYSTEM MODEL
This section briefly summarizes the model we use for OFDM
transceiver and channel (further details can be found in [2]–[4]).
OFDM divides the data stream into blocks of length N. The
subcarrier modulation is done by means of an inverse discrete
Fourier transform (IDFT) of the data blocks (typical length
N = 64 . . . 1024 carriers, but up to 8000 carriers are possible
for digital video broadcasting (DVB). Let the block X
(i)
k
be the
ith block of the transmission, with k indexing the N complex
modulation symbols. The transmit signal in the time-domain
x
(i)
[n] can then be written as
x
(i)
[n] =
1
N
N−1

k=0
X
(i)
k
e
j

N
nk
, −G ≤ n < N (1)
where G is the length of the CP; for the case of the CP
being absent, we simply set G = 0. The channel performs the
convolution with the impulse response of the channel h[n, l],
which consists of L multipath components 0 ≤ l < L and is
time-variant, as reflected by the double indexing. The received
signal y
(i)
[n] (without noise added by the channel) is then
y
(i)
[n] =
_
h ∗ x
(i)
_
[n]. (2)
As the CP is discarded by the receiver, we henceforth consider
only 0 ≤ n < N. The first L −G−1 samples of y
(i)
[n] are
2160 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 4, JULY 2007
distorted by ISI from the previous symbol x
(i−1)
; thus, y
(i)
consists of two parts:
y
(i)
[n] =
L−1

l=0
h[n, l]x
(i)
[n −l]σ[n −l + G]
+
L−1

l=G+1
h[n, l]x
(i−1)
[n−l+ G+N](1−σ[n−l+ G])
=y
(i,i)
[n] + y
(i,i−1)
[n] (3)
where σ is the Heaviside function, and y
(i,i)
[n] and y
(i,i−1)
[n]
are the contributions to y
(i)
[n] stemming from the ith and
i −1th block, respectively. To reconstruct the sent symbols,
the receiver performs a DFT on y
(i)
[n], Y
(i)
k
= DFT¦y
(i)
[n]¦
k
,
where k again is the subcarrier index. Throughout this paper,
we assume perfect synchronization of carriers and blocks, and
further, if not otherwise stated, absence of noise.
We now write the DFT of (3) in the receiver [15] as (4),
shown at the bottom of the page, where
˜
N = N + G.
Y
(i,i)
k
and Y
(i,i−1)
k
are, again, the parts of Y
(i)
k
that originate
from the own symbol i and the previous symbol i −1, respec-
tively. If we use (1), these parts can then be written as
Y
(i,i)
k
=
N−1

m=0
X
(i)
m
H
(i,i)
k,m
(5)
Y
(i,i−1)
k
=
N−1

m=0
X
(i−1)
m
H
(i,i−1)
k,m
(6)
where
H
(i,i)
k,m
=
1
N
N−1

n=0
L−1

l=0
h[n+i
˜
N, l]e
j

N
(nm−lm−nk)
σ[n−l+G]
(7)
and
H
(i,i−1)
k,m
=
1
N
N−1

n=0
L−1

l=G+1
h[n + i
˜
N, l]e
j

N
(nm−lm−nk)
¦1 −σ[n −l + G]¦ . (8)
This can be written in a compact form as a vector–matrix-
product
Y
(i)
=Y
(i,i)
+Y
(i,i−1)
=H
(i,i)
X
(i)
+H
(i,i−1)
X
(i−1)
(9)
where Y
(i,i−1)
is the ISI term, and Y
(i,i)
contains the desired
data disturbed by ICI. Note that if X
(i−1)
was detected suc-
cessfully (and the channel is completely known, as we always
assume in this paper), then Y
(i,i−1)
can be computed and
subtracted from Y
(i)
. This procedure can either be done in the
frequency domain, as we showed here, or in the time domain,
as in [15]. Due to this reason, ISI is not considered in the
remainder of this paper. For simplicity, we henceforth drop the
index of the block number superscript
(i)
.
The above formulation is general in the sense that it includes
the ICI due to the time variations and the insufficient guard
interval.
III. OPERATOR-PERTURBATION TECHNIQUE (OPT)
After ISI cancellation, the data vector still contains interfer-
ence fromits own symbol, the ICI. To equalize this interference,
we have to compute the data vector from (in the absence
of noise)
Y = HX. (10)
A straightforward solution is the inversion of this equation,
namely
⇒X
i
= H
−1
Y
i
. (11)
A matrix inversion require O(n
3
) operations, where n is the
size of the square matrix. In the next section, we will introduce a
technique (OPT) that only needs O(n
2
) operations per iteration.
A. Standard OPT
The OPT is an iterative method to efficiently approximate
and invert linear or nonlinear operators. It was originally intro-
duced by Cannon in [47] and is well known in the astrophysics
community. For matrices, it is also known as Jacobi-Iteration
Y
(i)
k
=
N−1

n=0
_
y
(i,i)
[n] + y
(i,i−1)
i
[n]
_
e
−j

N
nk
=
N−1

n=0
L−1

l=0
h[n + i
˜
N, l]x
(i)
[n −l]σ[n −l + G]e
−j

N
nk
. ¸¸ .
Y
(i,i)
k
+
N−1

n=0
L−1

l=G+1
h[n + i
˜
N, l]x
(i−1)
[n −l + G + N] (1 −σ[n −l + G]) e
−j

N
nk
. ¸¸ .
Y
(i,i−1)
k
(4)
MOLISCH et al.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2161
(see [48, ch. 10.1]), which has also been used in the context
of interference cancellation in code-division multiple access
(CDMA) [49]–[51]. We write (10) as
Y =
´
HX− (12)
where
´
H is the approximate operator whose inverse is easy to
compute, and is the deviation from the exact solution. For
example,
´
H could consist of the diagonal elements of H and
zero off-diagonal elements and would thus be trivial to invert.
Alternatively,
´
Hcan be taken as a banded matrix, i.e., a matrix
that contains only the main diagonal and a few off-diagonals of
the matrix H. The solution of (11) is found by the following
iteration:
X
(0)
=
´
H
−1
Y (13)

(0)
=(
´
H−H)X
(0)
(14)
X
(i+1)
=
´
H
−1
_

(i)
+Y
_
(15)

(i+1)
=(
´
H−H)X
(i+1)
(16)
where subscript (i) indexes the ith iteration. After initialization
[steps (13) and (14)], steps (15) and (16) are repeated until a
criterion is reached. Substituting step (15) in step (16) gives
X
(i+1)
= X
(i)
+
´
H
−1
_
Y−HX
(i)
_
. ¸¸ .
e
(
X
(i))
(17)
where e
(X
(i)
)
is the error of the old solution X
(i)
. If X
(i)
converges to X
(∞)
, then
X
(∞)
=
´
H
−1
_

(∞)
+Y
_

(∞)
=(
´
H−H)X
(∞)
X
(∞)
=
´
H
−1
_
(
´
H−H)X
(∞)
+Y
_
=X
(∞)
+
´
H
−1
_
Y−HX
(∞)
_
⇒ 0 =
´
H
−1
_
Y−HX
(∞)
_
. (18)
This means that if det¦
´
H
−1
¦ ,= 0, then we get the solution
Y−HX
(∞)
= 0 (19)
and therefore, X
(∞)
= X.
Step (16) of the iteration procedure requires a vector–matrix-
product that needs O(n
2
) operations. This step determines the
overall performance. The question of how many iteration loops
are necessary for a sufficient accuracy will be treated in the next
section.
B. Acceleration of Convergence
Auer [52] and Ng [53] developed methods to improve the
convergence speed of linear iterative schemes. These tech-
niques have been invented in astrophysics and chemical physics
but, astonishingly, do not seem to be widely known in mobile
radio. Clearly, linear iterative schemes are only linearly con-
vergent. An improvement of convergent speed can be achieved,
when we use more information than the last solution X
i−1
to
calculate the new X
i
:
X
(i)
= α
0
X
(i−1)
+
M

m=1
α
m
X
(i−1−m)
. (20)
α
0
= 1 −

α
m
is a normalization factor. The coefficients α
m
are determined as the solutions of the following set of linear
equations [53]:
M
acc
α = b
acc
(21)
where
M
acc
i,j
=

l
1
[X
N−1,l
[
[∆X
n,l
−∆X
n−i,l
]
[∆X
n,l
−∆X
n−j,l
] (22)
b
acc
i
=

l
1
[X
N−1,l
[
∆X
n,l
[∆X
n,l
−∆X
n−i,l
] (23)
where ∆X
n,l
is the lth component of the vector X
(n)

X
(n−1)
. For a Mth order acceleration, we must supply M + 2
successive estimates of the solution. Auer [52] gives a pseudo-
code for a second-order acceleration. Unfortunately, there is
no clear optimal choice of M; the results of the study in [52]
indicate that the utilization of higher orders than M = 2 usually
do not significantly improve the acceleration. This was also
validated by our own investigations; therefore, we implemented
the second-order acceleration only (M = 2).
According to the study in [52], the choice for M = 2 requires
the last four successive samples of the sequence of approxi-
mations of the solution. Following this algorithm, after every
fourth iteration in the approximation of Y =
´
HX, one acceler-
ation step is inserted. Then, the next four standard iterations
are done followed by another acceleration, etc. In principle,
this acceleration scheme is a very general method that can be
applied to every linear iterative approximation algorithm.
C. Further Modifications
We can further exploit the finite-alphabet properties of the
modulation: As we know that the solution of X = H
−1
Y must
lie in the set ¦−1, +1¦ for BPSK, we can include a BPSK slicer
after step (15) of the OPT algorithm. This slicer decides on the
components of X
i+1
and forces them to +1 or −1. Such an
“OPT with decision” technique can also be interpreted as a PIC
technique, which we will discuss in the next section.
IV. ICI CANCELLATION
In the last 15 years, efficient techniques have been developed
for the suppression of cochannel interferers in CDMA systems
[54]. Closer inspection of (5) and (7) reveals that the ICI
2162 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 4, JULY 2007
is mathematically equivalent to cochannel interference (even
though it is stemming from other subcarriers). Thus, the tech-
niques developed for CDMA can be reused for our application.
The basic idea behind OFDMinterference cancellation is that
the decisions of the symbols on the subcarriers are improved
iteratively. These improvements are due to the determination
and subtraction of the interference of all the other subcarriers,
based on the decisions of the previous iteration. A proper
initialization has to be done before entering the iteration loop.
We distinguish between PIC and SIC, depending on the order
in which the decisions of the subcarrier symbols are done.
A. Parallel Interference Cancellation (PIC)
For initialization, we again calculate a first estimate X
(0)
=
(diagH)
−1
Y, similar to the OPT technique. Here, diagH is
the diagonal of Honly and can be inverted with computational
effort O(n). The interference I
i
in step i is
I
(i)
= HX
(i−1)
−diagHX
(i−1)
. (24)
The updated interference-canceled received symbol vector Y
(i)
is simplified to
Y
(i)
= Y−I
(i)
. (25)
Hence, the updated estimate of the transmitted symbol on
subcarrier k is
(X
k
)
(i)
= (Y
k
)
(i)
/H
kk
(26)
where we equalize the interference-canceled received symbol
by the channel response on the frequency of subcarrier k. This
algorithm is called “parallel,” because (25) and (26) can be
applied to all channels k = 0 . . . n −1 simultaneously by a
vector–matrix calculus. Due to the vector–matrix multiplica-
tion, the computational complexity of the algorithm is O(n
2
).
These steps are identical to those of the OPT technique, with
the special assumption of
´
H = diagH.
Before entering the next round of iterations, the estimated
symbols can be processed in a nonlinear fashion. In the most
simple case, they could be forced to the nearest symbol of the
used alphabet. More generally
(
ˆ
X
k
)
(i)
= f
_
(X
k
)
(i)
_
(27)
where f() is the nonlinear decision function. A good choice
for BPSK is the hyperbolic tangent function [55] (
ˆ
X
k
)
(i)
=
tanh(c (X
k
)
(i)
) because far-off estimates are forced to ±1
and thus reduced in their impact on the next cancellation
stage, while estimates with small amplitudes are more or less
unchanged. The factor c in the hyperbolic tangent function
controls the slope near zero. Small c (e.g., c < 1) are appro-
priate if the estimates are not yet reliable. Large c (c > 10)
are appropriate for accurate estimates, e.g., if the signal-to-
interference ratio (SIR), which is defined in Section IV-B, is
large. That is the case for good channels with small delay
spread.
Performance can be further improved if c is increased from
one iteration step to the next. In the beginning, decisions are
still unreliable so that small c is appropriate. At the end of
the iteration, the decision function approximates the signum
function, and all decisions are in the set ¦±1¦. This technique
stems from an optimization technique called “simulated an-
nealing” and avoids the convergence of the MSE to a local
minimum [55].
The error vector and the MSE in the receiver are defined as
=Y−H
ˆ
X
i
MSE =||. (28)
The iteration stops after the MSE falls below a predefined
threshold or the number of iterations reaches a predefined
maximum number.
The most computationally expensive part of the algorithm is
the calculation of the interference (24)—it is a matrix–vector
product and is the reason for the O(n
2
)-complexity of this
method. For each subcarrier, the interference of all other subcar-
riers is computed. Due to the fact that the influence of two sub-
carriers decreases with increasing frequency distance, we can
restrict the interference calculation to the neighboring carriers.
It is also noteworthy that the optimum PIC scheme makes use
of the log-likelihood ratio from the MAP decoder [56].
B. Serial Interference Cancellation (SIC)
The main difference between SIC and PIC is that the update
steps (25) and (26) are not done in parallel for all subcarriers
but sequentially one channel after the other. In contrast to
PIC, an initialization is not required. After each estimate and
decision, the new interference is determined. This technique,
suggested by us in [43], was also used in [45], which proposed
simultaneous cancellation of several symbols in order to reduce
processing delays.
Let us first define a SIR for the subchannel k as
SIR
k
=
[H
kk
[
2
N−1

l=0
l=k
[H
kl
[
2
. (29)
We then order the subchannels by their SIR. This ensures that
high reliable channels (i.e., with high SIR) are processed before
the weak channels. In each iteration step (outer loop), the
subchannels are processed in the order of their corresponding
SIR (inner loop) that was calculated beforehand for the actual
channel situation. For each subchannel k of iteration step i, the
interference
(I
k
)
(i)
=

l∈X
l
decided in
previous iteration
and l=k
H
kl
(
ˆ
X
l
)
(i−1)
+

l∈X
l
decided in
current iteration
H
kl
(
ˆ
X
l
)
(i)
(30)
is computed and subtracted. The computational effort is again
O(n
2
) due to the necessity to sum up the off-diagonal elements
of the channel matrix and the computation of the interference
MOLISCH et al.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2163
terms. Note that the interference is calculated from the actual
estimate
ˆ
X
(i)
compared to Section IV-A, where it is calculated
from the previous estimate X
(i−1)
.
ˆ
X
(i)
contains all the deci-
sions of subchannels that were previously made because they
have higher SIR. The other channels were decided during the
last iteration step. The rest of the algorithm is similar to PIC.
The major advantage of SIC compared to PIC is that, within a
single iteration step, the impact of estimated symbols on low-
SIR carriers is considered immediately.
V. SIMULATION RESULTS
In the following, we present numerical results to demonstrate
the effectiveness of our proposed techniques. We investigate
1) the OPT, without any nonlinear processing (decisions) and
using Auer’s acceleration of convergence; 2) the PIC technique
with progressively increasing values of the parameter c, and
3) SIC. These cases will be simply called OPT, PIC, and SIC,
henceforth. For the OPT and the SIC (outer loop), we use a
predetermined number of iterations, namely ten. For the PIC,
we iterate until the MSE between subsequent iteration rounds
falls below a threshold of 10
−3
, or the number of iterations
becomes 30.
1
For the SIC and the PIC, we use the tanh mapping
function; we set c = 0.5 in the first iteration. For the PIC, c
increases by 1.0 in each iteration step, while for the SIC, it
increases by 2.45. Furthermore, we use a depth of five, i.e., four
off-diagonals, for the OPT.
We show the BER of an uncoded OFDM system for vari-
ous propagation channels. The channels are characterized by
their scattering functions P
s
(ν, τ), which is defined via the
relationship

_
−∞

_
−∞
E ¦h

(t, τ)h(t
/
, τ
/
)¦ exp [2πj(νt −ν
/
t
/
)] dtdt
/
= P
s
(ν, τ)δ(ν −ν
/
)δ(τ −τ
/
) (31)
assuming that the wide-sense stationary uncorrelated-scattering
assumption [57] is fulfilled. Here, ν is the Doppler frequency.
We furthermore assume that the statistics of the amplitudes are
zero-mean complex Gaussian, resulting in the familiar Rayleigh
fading. For most of the examples, we analyze a static channel
so that the channel is characterized by its power-delay profile
(PDP) only
PDP(τ) = E
_
[h(τ)[
2
_
. (32)
We first analyze the impact of the shape of the PDP on
the effect of the bit-error probability (BER) for the different
equalization techniques. It is well known that, for unequalized
single-carrier systems, the BER is proportional to the squared
magnitude of the rms delay spread [58], [59]. For OFDM
systems with CP, however, we need to consider only the part of
the PDP that exceeds the length of the CP. We can thus expect
1
In the simulations that follows, the MSE threshold was not reached, so that
the number of iterations was always 30.
Fig. 1. BER as a function of the rms delay spread for a two-spike profile.
OFDM system with 64 tones and CP with 12 samples length.
Fig. 2. BER as a function of the rms delay spread for a two-spike profile.
OFDM system with 64 tones and no CP.
a different dependence of the BER for different PDPs. Figs. 1
and 2 show the BER for a two-spike profile
PDP(τ) = δ(τ) + δ(τ −2τ
rms
) (33)
where τ
rms
is the rms delay spread normalized to the OFDM
symbol duration; the simulations use 100 channel realizations,
in each of which 100 OFDM symbols are transmitted (sam-
ple simulations with 500 channel realizations gave the same
results). We find that the BER is lower in an equalized system
without CP than in a conventional system with CP, even if the
CP is large enough to eliminate all ICI. This fact, which is
astonishing at first glance, can be explained by the frequency
2164 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 4, JULY 2007
Fig. 3. BER as a function of the rms delay spread for a rectangular PDP.
OFDM system with 64 tones and no CP.
diversity inherent in delay dispersion. A conventional system
with CP discards all information “hidden” in the first G re-
ceived samples. For an OFDM system without equalizer, this
information is detrimental, since it leads to ICI and, thus,
interference on the different tones. However, a system with
one of our equalizers can actually make good use of this
information—essentially, it carries information about each bit
on several tones simultaneously and is, thus, less sensitive to
fading. We also observe that, for very large delay spreads, all of
the equalization techniques start to break down. Finally, we note
that the SIC and PIC perform slightly better than the OPT. The
“choppiness” (nonmonotonicity) of the curves stems from the
idiosyncrasies for the channel models, i.e., for a given channel
model, some delay spreads can be equalized more easily than
others (this was confirmed by runs with more channel realiza-
tions). Finally, we find that, if there is a CP and no equalizer, the
impact of ICI is noticeable for all rms delay spreads larger than
0.094 (0.5∗12/64) for the choice of parameters in Fig. 1. Figs. 3
and 4 show the BER for rectangular and exponential PDPs,
respectively. Note that, for those profiles, the maximum rms
delay spread considered here is smaller: We investigate only
cases where the maximum excess delay is smaller or equal to
one symbol duration (as the exponential PDP extends to infinity,
we require that the “significant” part of the PDP lasts only one
symbol duration).
We also investigate the impact of using different approximate
operators
´
H in the OPT. Fig. 5 shows the BER as a function
of the delay spread of a rectangular PDP when
´
H contains
two, eight, 32, and 64 off-diagonals of H. We see that there
is hardly any difference in the resulting BER, which shows
that good results can be achieved with a
´
H that requires little
numerical effort for its inversion. We also investigate the impact
of the acceleration procedure for the OPT. Fig. 6 shows how the
acceleration steps improve of the MSE.
We analyze next the performance in a scenario that does not
show delay dispersion (single-tap channel) but rather only time
Fig. 4. BER as a function of the rms delay spread for an exponential PDP.
OFDM system with 64 tones and no CP.
Fig. 5. BER as a function of delay spread for OPT with different numbers of
diagonals accounted for in
´
H.
variance of the channel, following a “classical” Jakes Doppler
spectrum. We plot the BER versus the rms Doppler spread
(normalized to the inverse of the OFDM symbol duration) for
the case of a classical Jakes Doppler spectrum
S
D
(ν) ∝
1
_
ν
2
max
−ν
2
(34)
at an SNR of 30 dB; see Fig. 7. Once again, the OFDM symbol
length being considered here is 64 samples. It is noteworthy
that the OPT performs better than the interference-cancellation
technique in this case. Fig. 8 shows the impact of using different
numbers of off-diagonal elements in the iterations. We also note
that, in the Doppler case, the simple but effective approach of
[60] can be used.
MOLISCH et al.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2165
Fig. 6. Impact of the acceleration of convergence onto the MSE. Red stars:
Acceleration steps. Blue line: With acceleration. Black line: Without acceler-
ation. Simulation parameters: OFDM block size: 16 samples; two-spike PDP,
maximum excess delay: One sample; mean spike powers: [1,0.1]; no CP.
Fig. 7. BER as a function of the normalized Doppler spread for Jakes Doppler
spectrum. Single-tap PDP, 64 carriers.
In order to analyze the usefulness of our algorithms in
comparison to an OFDM system without any equalizer, we
performed simulations of a system similar to the IEEE 802.11
standard. The system bandwidth was taken to be 20 MHz, and
the carrier frequency was taken to be 5 GHz; we used the IEEE
802.11n channel models [61] which define both the delay and
angular dispersion; from this, the scattering function can be
derived. We approximate the standardized system by taking a
64-carrier OFDM system (the actual 802.11a system has 48
data carriers as well as pilots and null carriers); the velocity was
taken to be 100 m/s (which is much higher than can usually
be anticipated for 802.11a systems), making this case suffer
from ICI due to Doppler, as well as delay spread. If the CP
is 16 samples long, as defined in the 802.11a standard, then
ICI is negligible (simulation results not shown here for space
Fig. 8. BER as a function of Doppler spread for OPT with different numbers
of diagonals accounted for in
´
H.
Fig. 9. BER as a function of SNR for an 802.11-like OFDM system with 64
carriers and eight samples CP. Performance is analyzed in a channel model F of
the 802.11n channel models with 100-m/s velocity.
reasons). However, if the CP is only eight samples long, which
was recently suggested as an option for the 802.11n standard,
ICI becomes noticeable. Fig. 9 plots the BER for various
interference-cancellation techniques versus SNR in comparison
with a case where no equalizer is present in the OFDM receiver.
We find that, while the unequalized receiver shows an error
floor of 10
−3
, the OPT, as well as the PIC and SIC, does not
exhibit a floor at SNRs up to 30 dB.
VI. SUMMARY AND CONCLUSION
In this paper, we have presented several algorithms for
the equalization of OFDM systems that suffer from ICI. We
2166 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 56, NO. 4, JULY 2007
considered all sources of ICI, namely, an insufficient du-
ration of the guard interval, as well as temporal variations
of the channel. ISI can be eliminated by decision-feedback
algorithms. We first established a general system model to
represent all ICI contributions as off-diagonal elements of a
channel matrix that multiplies the transmit symbols. We then
investigated various schemes for the efficient inversion of this
process.
We first considered a linear-equalization approach, where the
inversion of the channel matrix is done iteratively by means of
OPT. This approach can be improved by adding an “acceler-
ation of convergence scheme.” Furthermore, introducing deci-
sions between the iteration steps leads to a PIC scheme, where
the interference from neighboring tones (based on preliminary
decisions) is subtracted in several stages. Alternatively, a SIC
scheme first decides about the tone with the strongest SINR,
subtracts its interference from the neighboring tones, decides
for the next-best tone, and so on.
Simulations demonstrated that all three schemes lead to a
drastic reduction of the error floor created by ICI and can even
lead to performance that is better than with long CP, while,
at the same time improving spectral efficiency. Our simulation
results show that the low-complexity equalization techniques
can be effectively used in practical OFDM systems and give
guidelines for their implementation.
ACKNOWLEDGMENT
The authors would like to thank Prof. O. Edfors for critical
reading of the manuscript. They would also like to thank the
anonymous reviewers for the helpful comments.
REFERENCES
[1] L. J. Cimini, “Analysis and simulation of a digital mobile channel us-
ing orthogonal frequency division multiplexing,” IEEE Trans. Commun.,
vol. COM-33, no. 7, pp. 665–675, Jul. 1985.
[2] T. May and H. Rohling, “OFDM,” in Wideband Wireless Digital
Communications, A. F. Molisch, Ed. Englewood Cliffs, NJ: Prentice-
Hall, 2000.
[3] L. Hanzo, M. Muenster, B. Choi, and T. Keller, OFDM and MC-CDMA
for Broadband Multi-User Communications; WLANs and Broadcasting.
Hoboken, NJ: Wiley, 2003.
[4] A. F. Molisch, Wireless Communications. Piscataway, NJ: IEEE Press,
2005.
[5] B. O’Hara and A. Petrick, The IEEE 802.11 Handbook: A Designer’s
Companion, 2nd ed. Piscataway, NJ: IEEE Std. Pubs., 2005.
[6] A. Batra et al., Multi-Band OFDM Physical Layer Proposal, 2003.
Document IEEE 802.15-03/267r2.
[7] F. Ohrtmann, WiMax Handbook. New York: McGraw-Hill, 2005.
[8] S. Weinstein and P. Ebert, “Data transmission by frequency-division mul-
tiplexing using the discrete Fourier transform,” IEEE Trans. Commun.,
vol. COM-19, no. 5, pp. 628–634, Oct. 1971.
[9] B. Muquet, Z. Wang, G. B. Giannakis, M. de Courville, and P. Duhamel,
“Cyclic prefixing or zero padding for wireless multicarrier transmis-
sions?” IEEE Trans. Commun., vol. 50, no. 12, pp. 2136–2148,
Dec. 2002.
[10] T. Muenster and L. Hanzo, “Second-order channel parameter estimation
assisted cancellation of channel variation-induced inter-subcarrier inter-
ference in OFDM systems,” in Proc. PIMRC, 2001, pp. 1–5.
[11] Y. Mostofi, D. C. Cox, and A. Bahai, “ICI mitigation for mobile OFDM
receivers,” in Proc. IEEE Int. Conf. Commun., 2003, pp. 3351–3355.
[12] A. Gorokhov and J. P. Linnartz, “Robust OFDM receivers for disper-
sive time-varying channels: Equalization and channel acquisition,” IEEE
Trans. Commun., vol. 52, no. 4, pp. 572–583, Apr. 2004.
[13] P. H. Moose, “A technique for orthogonal frequency division multiplexing
frequency offset correction,” IEEE Trans. Commun., vol. 42, no. 10,
pp. 2908–2914, Oct. 1994.
[14] M. J. Fernandez-Getino-Garcia, O. Edfors, and J. M. Paez-Borollo, “Fre-
quency offset correction for coherent OFDM in wireless systems,” IEEE
Trans. Consum. Electron., vol. 47, no. 1, pp. 187–193, Feb. 2001.
[15] D. Kim and G. L. Stuber, “Residual ISI cancellation for OFDM with ap-
plications to HDTV broadcasting,” IEEE J. Sel. Areas Commun., vol. 16,
no. 8, pp. 1590–1599, Oct. 1998.
[16] Y. Sun and L. Tong, “Channel equalization using one-tap DFE for wireless
OFDM systems with ICI and ISI,” in Proc. 2nd IEEE Workshop Signal
Process. Adv. Wireless Commun., 1999, pp. 146–149.
[17] H. Steendam and M. Moeneclaey, “Analysis and optimization of the
performance of OFDM on frequency-selective time-selective fading chan-
nels,” IEEE Trans. Commun., vol. 47, no. 12, pp. 1811–1819, Dec. 1999.
[18] W. Kozek and A. F. Molisch, “Nonorthogonal pulseshapes for multicar-
rier communications in doubly dispersive channels,” IEEE J. Sel. Areas
Commun., vol. 16, no. 8, pp. 1579–1589, Oct. 1998.
[19] S. Beaver and T. Stromer, “Optimal OFDM pulse and lattice design for
doubly dispersive channels,” in Proc. 35th Asilomar Conf. Signals, Syst.,
Comput., 2001, pp. 1763–1767.
[20] J. Armstrong, “Analysis of new and existing methods of reducing intercar-
rier interference due to carrier frequency offset in OFDM,” IEEE Trans.
Commun., vol. 47, no. 3, pp. 365–369, Mar. 1999.
[21] Y. Zhao and S. G. Haggman, “Intercarrier interference self-cancellation
scheme for OFDM mobile communication systems,” IEEE Trans. Com-
mun., vol. 49, no. 7, pp. 1185–1191, Jul. 2001.
[22] K. Sathananthan, R. M. A. P. Rajatheva, and S. B. Slimane, “Cancellation
technique to reduce intercarrier interference in OFDM,” Electron. Lett.,
vol. 36, no. 25, pp. 2078–2079, Dec. 2000.
[23] A. Seyedi and G. J. Saulnier, “General self-cancellation scheme for mit-
igation of ICI in OFDM systems,” in Proc. IEEE Int. Comf. Commun.,
2004, pp. 2653–2657.
[24] R. Schur, J. Speidel, and R. Angerbauer, “Reduction of guard interval
by impulse compression for DMT modulation on twisted pair cables,” in
Proc. IEEE Global Telecommun. Conf., 2000, pp. 1632–1636.
[25] S. Chen and T. Yao, “Low complexity ICI cancellation for OFDM systems
in doubly-selective fading channels,” in Proc. IEEE Int. Conf. Commun.,
2004, pp. 2535–2538.
[26] A. Stamoulis, S. N. Diggavi, and N. Al-Dhahir, “Intercarrier interfer-
ence in MIMO OFDM,” IEEE Trans. Signal Process., vol. 50, no. 10,
pp. 2451–2464, Oct. 2002.
[27] J. Kim, J. Kang, and E. J. Powers, “Abandwidth efficient OFDMtransmis-
sion scheme,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process.,
2002, pp. III-2329–III-2332.
[28] J. Zhang, W. Ser, and J. Zhu, “Effective optimisation method for channel
shortening in OFDMsystems,” Proc. Inst. Electr. Eng.—Communications,
vol. 150, no. 2, pp. 85–90, Apr. 2003.
[29] S. Trautmann, T. Karp, and N. J. Fliege, “Frequency domain equalization
of DMT/OFDM systems with insufficient guard interval,” in Proc. IEEE
Int. Conf. Commun., 2002, pp. 1646–1650.
[30] S. Chen and T. Yao, “FEQ for OFDM systems with insufficient CP,” in
Proc. IEEE Pers., Indoor, Mobile Radio Commun., 2003, pp. 550–553.
[31] C. J. Park and G. H. Im, “Efficient DMT/OFDM transmission with in-
sufficient cyclic prefix,” IEEE Commun. Lett., vol. 8, no. 9, pp. 576–578,
Sep. 2004.
[32] K. Sathananathan and C. Tellambura, “Forward error correction codes
to reduce intercarrier interference in OFDM,” in Proc. IEEE Int. Symp.
Circuits Syst., 2001, pp. 566–569.
[33] Y. Zhao, J. D. Leclercq, and S. G. Haggman, “Intercarrier interference
compression in OFDM communication systems by using correlative cod-
ing,” IEEE Commun. Lett., vol. 2, no. 8, pp. 214–216, Aug. 1998.
[34] L. Vandendorpe, “Performance of fractionally spaced linear and decision-
feedback equalizers for multitone systems,” in Proc. IEEE Globecom,
1996, pp. 36–40.
[35] W. G. Jeon, K. H. Chang, and Y. S. Cho, “An equalization technique
for orthogonal frequency-division multiplexing systems in time-variant
multipath channels,” IEEE Trans. Commun., vol. 47, no. 1, pp. 27–32,
Jan. 1999.
[36] Y.-S. Choi, P. J. Voltz, and F. Cassara, “On channel estimation and de-
tection for multicarrier signals in fast and frequency selective Rayleigh
fading channel,” IEEE Trans. Commun., vol. 49, no. 8, pp. 1375–1387,
Aug. 2001.
[37] M. Nakamura, M. Fujii, M. Itami, K. Itoh, and A. H. Aghvami, “A study
on an MMSE ICI canceller for OFDM under Doppler-spread channel,”
in Proc. IEEE Symp. Pers., Indoor, Mobile Radio Commun., 2003,
pp. 236–240.
MOLISCH et al.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2167
[38] X. Cai and G. B. Giannakis, “Bounding performance and suppressing in-
tercarrier interference in wireless mobile OFDM,” IEEE Trans. Commun.,
vol. 51, no. 12, pp. 2047–2056, Dec. 2003.
[39] P. Schniter, “Low-complexity equalization of OFDM in doubly selective
channels,” IEEE Trans. Signal Process., vol. 52, no. 4, pp. 1002–1011,
Apr. 2004.
[40] M. Guillaud and D. T. M. Slock, “Channel modeling and associated inter-
carrier interference equalization for OFDM systems with high Doppler
spread,” in Proc. IEEE ICASSP, 2003, pp. IV-237–IV-240.
[41] H. Bogucka and K. Wesolowski, “Frequency-domain echo cancellation in
digital multicarrier modulation systems,” IEEE Trans. Commun., vol. 48,
no. 2, pp. 333–342, Feb. 2000.
[42] M. Toeltsch and A. F. Molisch, “Efficient OFDM transmission without
cyclic prefix over frequency-selective channels,” in Proc. PIMRC, 2000,
pp. 1363–1367.
[43] M. Toeltsch and A. F. Molisch, “Equalization of OFDM-systems by inter-
ference cancellation techniques,” in Proc. ICC, 2001, pp. 1950–1954.
[44] Q. Xuerong and Z. Lijun, “Interchannel interference cancellation in wire-
less OFDM systems via Gauss–Seidel method,” in Proc. Int. Conf. Com-
mun. Technol., 2003, pp. 1051–1055.
[45] J. Cai, J. W. Mark, and X. Shen, “ICI cancellation in OFDM wireless
communication systems,” in Proc. IEEE Global Telecomm. Conf., 2002,
pp. 656–660.
[46] B. Xu, C. Yang, and S. Mao, “A multi-carrier detection algorithm for
OFDM systems without guard time,” in Proc. IEEE Int. Conf. Commun.,
2003, pp. 3377–3381.
[47] J. Cannon, “Non-local perturbation techniques,” in Methods in Radiative
Transfer, W. Kalkofen, Ed. Cambridge, U.K.: Cambridge Univ. Press,
1984.
[48] G. Golub and C. van Loan, Matrix Computations, 2nd ed. Baltimore,
MD: Johns Hopkins Univ. Press, 1989.
[49] H. Elders-Boll, H. D. Schotten, and A. Busboom, “Efficient implemen-
tation of linear multiuser detectors for asynchronous CDMA systems by
linear interference cancellation,” Eur. Trans. Telecommun., vol. 9, no. 5,
pp. 427–438, 1998.
[50] L. K. Rasmussen, T. J. Lim, and A. L. Johansson, “A matrix-algebraic
approach to successive interference cancellation in CDMA,” IEEE Trans.
Commun., vol. 48, no. 1, pp. 145–151, Jan. 2000.
[51] D. Guo, L. K. Rasmussen, S. Sun, and T. J. Lim, “A matrix-algebraic
approach to linear parallel interference cancellation in CDMA,” IEEE
Trans. Commun., vol. 40, no. 1, pp. 152–161, Jan. 2000.
[52] L. Auer, “Acceleration of convergence,” in Numerical Radiative Transfer,
W. Kalkofen, Ed. Cambridge, U.K.: Cambridge Univ. Press, 1987.
[53] K. C. Ng, “Hypernetted chain solutions for the classical one-
component plasma up to gamma = 7000,” J. Chem. Phys, vol. 61, no. 7,
pp. 2680–2689, Oct. 1974.
[54] S. Verdu, Multiuser Detection. Cambridge, U.K.: Cambridge Univ.
Press, 1998.
[55] D. Raphaeli, “Iterative cochannel interference cancellation in synchro-
nous CDMA on a frequency selective channel,” in Proc. 5th IEEE Conf.
Int. Univ. Pers. Commun., 1996, pp. 336–340.
[56] S. Suyama, H. Suzuki, and K. Fukawa, “An OFDM receiver employing
turbo equalization for multipath environments with delay spread greater
than the guard interval,” in Proc. IEEE Veh. Technol. Conf.—Spring, 2003,
pp. 632–636.
[57] P. Bello, “Characterization of randomly time-variant linear channels,”
IEEE Trans. Commun., vol. COM-11, no. 4, pp. 360–393, Dec. 1963.
[58] J. Chuang, “The effects of time delay spread on portable radio communi-
cations channels with digital modulation,” IEEE J. Sel. Areas Commun.,
vol. SAC-5, no. 5, pp. 879–888, Jun. 1987.
[59] A. F. Molisch, Ed., Wideband Wireless Digital Communications. New
York: Prentice-Hall, 2000.
[60] L. Rugini, P. Banelli, and G. Leus, “Simple equalization of time-varying
channels for OFDM,” IEEE Commun. Lett., vol. 9, no. 7, pp. 619–621,
Jul. 2005.
[61] V. Erceg, L. Schumacher, P. Kyritsi, D. S. Baum, A. F. Molisch, and
A. Y. Gorokhov, “Indoor MIMO WLAN channel models,” in Proc. Stan-
dardization Drafts IEEE 802 Meeting, Dallas, TX, Mar. 2003.
Andreas F. Molisch (S’89–M’95–SM’00–F’05)
received the Dipl.Ing., Dr.Techn., and Habilitation
degrees from the Technical University of Vienna,
Vienna, Austria, in 1990, 1994, and 1999,
respectively.
From 1991 to 2000, he was with the Technical
University of Vienna, most recently as an Associate
Professor. From 2000 to 2002, he was with the Wire-
less Systems Research Department, AT&T (Bell)
Laboratories Research, Middletown, NJ. Since then,
he has been with Mitsubishi Electric Research Lab-
oratories, Cambridge, MA, where he is now a Distinguished Member of
Technical Staff. He is also a Professor and Chairholder for radio systems
at Lund University, Lund, Sweden. He has done research in the areas of
surface-acoustic-wave filters, radiative transfer in atomic vapors, atomic line
filters, smart antennas, and wideband systems. His current research inter-
ests are multiple-input–multiple-output (MIMO) systems, measurement and
modeling of mobile-radio channels, cooperative communications, and ultra-
wideband (UWB) systems. He has authored, coauthored, or edited four books
(among them is the recent textbook Wireless Communications, Wiley–IEEE
Press), 11 book chapters, some 100 journal papers, and numerous conference
contributions.
Dr. Molisch is an Editor of the IEEE TRANSACTIONS ON WIRELESS
COMMUNICATIONS, Guest Editor of a recent Special Issue on UWB (in the
IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS). He has been a
member of numerous Technical Program Committees (TPCs), Vice Chair of the
TPC of Vehicular Technology Conference (VTC) 2005 spring, General Chair of
ICUWB 2006, TPC Chair for Chinacom 2007, and Wireless Symposium Chair
for Globecom 2007. He has participated in the European research initiatives
“COST 231,” “COST 259,” and “COST 273,” where he was Chairman of
the MIMO channel working group. He was Chairman of the IEEE 802.15.4a
channel model standardization group and is also Chairman of Commission C
(signals and systems) of the International Union of Radio Scientists. He is an
IEEE Distinguished Lecturer and recipient of several awards.
Martin Toeltsch (M’00) received the Dipl. Ing.
(M.S.) degree in communications engineering and
the Dr.Tech. degree from the Technische Universität
Wien (TU-Wien), Vienna, Austria, in 1998 and 2002,
respectively. In his doctoral thesis, he investigated
high-resolution evaluations of directionally resolved
radio-channel measurements.
From 1992 to 1998, he worked in the field of
signal and image processing for medical applica-
tions. From 1998 to 2002, he was a member of the
Mobile Communications Group at the Institut für
Nachrichten-und Hochfrequenztechnik, TU-Wien. His research focused on sig-
nal processing, orthogonal frequency-division multiplexing and smart-antenna
radio-channel measurements, as well as propagation and channel modeling. In
2002, he founded SYMENA Software and Consulting Ltd., Vienna, where he
develops smart-antenna radio network planning and optimization tools.
Sameer Vermani received the B.Tech. degree (with
the institute silver medal) in electrical engineering
from the Indian Institute of Technology, Delhi, India,
and the M.S. degree in electrical engineering from
the University of California, Berkeley, in 2006.
He was the recipient of the Vodafone U.S. Founda-
tion Fellowship for graduate study at Berkeley. He is
currently with Qualcomm Inc., San Diego, CA. His
research interests span the physical layer of wireless-
communication systems. In particular, he has done
work in the areas of multiple-input–multiple-output
and orthogonal frequency-division multiplexing systems.

The subcarrier modulation is done by means of an inverse discrete Fourier transform (IDFT) of the data blocks (typical length N = 64 . An area of particular interest has been linear equalization in the frequency domain. (2) As the CP is discarded by the receiver. The remainder of this paper is organized the following way: in Section II. Reference [38] also provides a matched-filter bound for time-varying frequency-selective channels for the case that the CP is sufficiently long.e.. various algorithms have been considered for the matrix inversion. Section III is dedicated to the “operator-perturbation technique” (OPT) for the efficient inversion of matrices. 4) Techniques that depend on the presence of unused subcarriers (null carriers): In this approach. 3) Temporal equalizers: In these techniques. (i) 2π −G ≤ n < N (1) where G is the length of the CP. 2) Self-interference-cancellation techniques: In this approach. Equalization of ICI due to insufficient guard intervals. OFDM divides the data stream into blocks of length N . [25]. the basic pulseshape for the signaling of the subcarriers can also be modified [18]. II. as well as a parallelinterference-cancellation (PIC) scheme that shows rapid convergence and lower error floor for some ranges of delay spread. Reference [26] found the coefficients for a temporal filter that maximizes the SINR for both conventional OFDM and multiple-input–multiple-output (MIMO)-OFDM. we investigate interference-cancellation techniques. Similarly to the principle of the CP. redundancy is placed in the frequency domain by having more carriers than is strictly necessary. l]. for the case of the CP being absent. which consists of L multipath components 0 ≤ l < L and is time-variant. [19]. 5) Forward-error correction [32]: The capability of any forward-error-correcting code can be used to eliminate the errors caused by the ICI. which leads to a strong reduction in the self-interference.MOLISCH et al. 1) Optimum choice of the carrier spacing and OFDM symbol length: By trading off ICI caused by temporal variations and by delay dispersion. This technique is very effective for the mitigation of ICI but leads to a reduction of the spectral efficiency of the system. interest in this approach has greatly increased and various related schemes have been published. moreover. A summary and the conclusions wrap up this paper. based on tail cancellation and cyclic reconstruction. was studied in [16]. Equalization of ICI due to temporal variations of the channels has been studied in [11]. The transmit signal in the time-domain x(i) [n] can then be written as x(i) [n] = 1 N N −1 k=0 Xk ej N nk . Let the block Xk be the ith block of the transmission. however. We also present a serial-interference-cancellation (SIC). this approach is not feasible in time-variant wireless channels. Reference [44] uses a Gauss–Seidel iteration for the inversion of the channel matrix. [12]. with k indexing the N complex modulation symbols. the equalization is done in the time domain. covering both PIC and SIC. the information is modulated not just onto a single subcarrier but onto a group of them. Bogucka and Wesolowski [41] advocated the use of the LMS algorithm to find the coefficients in the frequency domain. which is inverted explicitly or implicitly. channel shortening can be used to make the effective impulse response shorter than the CP [28]. The channel performs the convolution with the impulse response of the channel h[n. In a similar approach. with sufficient guard interval). . but up to 8000 carriers are possible (i) for digital video broadcasting (DVB). most of the equalization strategies suggested in these papers rely on the special structure of the transfer-function matrix in the case of pure Doppler-induced ICI (i. In a related approach. A large range of techniques has been developed in recent years and can be classified as follows. possibly in combination with decision-feedback equalization (DFE). Direct inversion of the channel matrix [either according to a zero-forcing or an minimum mean-squared error (MMSE) criterion] is computationally too expensive. [45] and [46] suggested SIC schemes and analyzed their performance. ICI (and ISI) can lead to a considerable performance degradation in many systems and must be combated. the redundancy in the frequency domain allows elimination of the ICI [29]–[31]. Since 2000. the bit error rate (BER) can be minimized without requiring any structural changes in either the transmitter or receiver [17]. we simply set G = 0. was first suggested in the study in [15] and improved in the study in [27]. 1024 carriers. The received signal y (i) [n] (without noise added by the channel) is then y (i) [n] = h ∗ x(i) [n]. we describe the system model as well as the channel characteristics determining the system behavior.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2159 Irrespective of the reasons. which allows elimination of the ISI by redundancy in the time domain. The impact of the channel can be described by a transfer-function matrix. and correlative coding [33] can improve the ICI as well. The first L − G − 1 samples of y (i) [n] are . Reference [40] showed that the elimination of ISI due to time variations is dual to the temporal equalization in single-carrier systems. [43]. before the FFT at the receiver (with a possible feedback that undergoes backand-forth transformations) [24]. Next. S YSTEM M ODEL This section briefly summarizes the model we use for OFDM transceiver and channel (further details can be found in [2]–[4]). when we first suggested these algorithms in our conference papers [42]. as reflected by the double indexing. The technique was suggested in [20] and later extended in [21]–[23]. we henceforth consider only 0 ≤ n < N . Section V gives numerical results for both the OPT and the interference cancellation. An efficient scheme called RISIC. and [34]–[39]. . we therefore suggest an iterative inversion technique for the general case of time-varying delaydispersive channels with insufficient CP. In this paper. For wireline communications.

or in the time domain. and Y(i. again. the data vector still contains interference from its own symbol. (8) 2π The OPT is an iterative method to efficiently approximate and invert linear or nonlinear operators. O PERATOR -P ERTURBATION T ECHNIQUE (OPT) After ISI cancellation.i−1) [n] are the contributions to y (i) [n] stemming from the ith and i − 1th block. y (i) consists of two parts: L−1 This can be written in a compact form as a vector–matrixproduct Y(i) = Y(i. l]x(i) [n − l]σ[n − l + G]e−j N nk 2π Yk N −1 L−1 (i.m (i. thus. (11) 1 N N −1L−1 ˜ h[n+iN . JULY 2007 distorted by ISI from the previous symbol x(i−1) . Due to this reason. the parts of Yk that originate Yk from the own symbol i and the previous symbol i − 1. VOL. To reconstruct the sent symbols.i−1) [n] e−j N nk 2π = n=0 l=0 ˜ h[n + iN .i−1) is the ISI term.m (i. NO.i−1) = H(i. III.m = (i.2160 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY. l]ej n=0 l=0 2π N (nm−lm−nk) σ[n−l+G] (7) A matrix inversion require O(n3 ) operations. where n is the size of the square matrix. The above formulation is general in the sense that it includes the ICI due to the time variations and the insufficient guard interval.i) [n] + yi N −1 L−1 (i. Throughout this paper. If we use (1). Note that if X(i−1) was detected successfully (and the channel is completely known. we henceforth drop the index of the block number superscript (i) . In the next section. Yk = DFT{y (i) [n]}k .i) contains the desired data disturbed by ICI. we have to compute the data vector from (in the absence of noise) Y = HX. To equalize this interference.i−1) can be computed and subtracted from Y(i) . (i. as in [15].i−1) (i) and Yk are.i) +Y(i. respectively.i) (5) Yk where (i. it is also known as Jacobi-Iteration Yk (i) N −1 = n=0 y (i. we will introduce a technique (OPT) that only needs O(n2 ) operations per iteration. Standard OPT and Hk. l]x(i−1) [n − l + G + N ] (1 − σ[n − l + G]) e−j N nk (4) Yk (i. the ICI. and further. ISI is not considered in the remainder of this paper. where k again is the subcarrier index. We now write the DFT of (3) in the receiver [15] as (4). It was originally introduced by Cannon in [47] and is well known in the astrophysics community.i) [n] and y (i.i) [n] + y (i. l]x(i−1)[n−l+ G+N ](1−σ[n−l+ G]) (3) = y (i.i) Yk N −1 = m=0 N −1 (i) Xm Hk. (i) the receiver performs a DFT on y (i) [n].i−1) = 1 N N −1 L−1 n=0 l=G+1 ˜ h[n + iN .i−1) . then Y(i.i−1) (6) A straightforward solution is the inversion of this equation.i) · X(i) + H(i. ˜ shown at the bottom of the page. 4.i−1) = m=0 (i−1) Xm Hk. and y (i.i) Hk.i−1) [n] where σ is the Heaviside function. if not otherwise stated. respectively.i) (i.m (i. This procedure can either be done in the frequency domain. 56. l]ej N (nm−lm−nk) × {1 − σ[n − l + G]} . we assume perfect synchronization of carriers and blocks.i) + n=0 l=G+1 2π ˜ h[n + iN . as we always assume in this paper). (10) y (i) [n] = l=0 h[n. l]x(i) [n − l]σ[n − l + G] L−1 + l=G+1 h[n. A. where N = N + G. namely ⇒ Xi = H−1 · Yi . For simplicity. as we showed here. absence of noise. For matrices. these parts can then be written as (i.i−1) · X(i−1) (9) where Y(i.

then we get the solution Y − HX(∞) = 0 (19) and therefore. there is no clear optimal choice of M . After initialization [steps (13) and (14)].l [∆Xn. +1} for BPSK. Such an “OPT with decision” technique can also be interpreted as a PIC technique. therefore. For a M th order acceleration. do not seem to be widely known in mobile radio. The question of how many iteration loops are necessary for a sufficient accuracy will be treated in the next section. Step (16) of the iteration procedure requires a vector–matrixproduct that needs O(n2 ) operations. The solution of (11) is found by the following iteration: X(0) = H−1 Y (0) niques have been invented in astrophysics and chemical physics but. this acceleration scheme is a very general method that can be applied to every linear iterative approximation algorithm.MOLISCH et al. Further Modifications We can further exploit the finite-alphabet properties of the modulation: As we know that the solution of X = H−1 Y must lie in the set {−1.l | ∆Xn. IV. the choice for M = 2 requires the last four successive samples of the sequence of approximations of the solution.l ] (17) (X(i) ) where e(X(i) ) is the error of the old solution X(i) . For example. linear iterative schemes are only linearly convergent. According to the study in [52]. Unfortunately. These tech- In the last 15 years. a matrix that contains only the main diagonal and a few off-diagonals of the matrix H. when we use more information than the last solution Xi−1 to calculate the new Xi : M X(i) = α0 X(i−1) + m=1 αm X(i−1−m) . Following this algorithm. Then. (20) α0 = 1 − αm is a normalization factor.j = = (H − H)X(0) (i) X(i+1) = H−1 (i+1) +Y (15) (16) bacc = i l 1 [∆Xn..1]). Clearly.l | (22) (23) = (H − H)X(i+1) × [∆Xn. Substituting step (15) in step (16) gives X(i+1) = X(i) + H−1 Y − HX(i) e |XN −1. C. Closer inspection of (5) and (7) reveals that the ICI . etc.l ] 1 l where subscript (i) indexes the ith iteration. If X(i) converges to X(∞) . This step determines the overall performance. Auer [52] gives a pseudocode for a second-order acceleration. ICI C ANCELLATION This means that if det{H−1 } = 0.l ] |XN −1. This was also validated by our own investigations. 10. which has also been used in the context of interference cancellation in code-division multiple access (CDMA) [49]–[51]. the next four standard iterations are done followed by another acceleration. i.e. An improvement of convergent speed can be achieved. Alternatively. We write (10) as Y = HX − (12) where H is the approximate operator whose inverse is easy to compute.l − ∆Xn−i. B. one acceleration step is inserted. astonishingly. (18) where ∆Xn. we can include a BPSK slicer after step (15) of the OPT algorithm. then X(∞) = H−1 (∞) (∞) +Y = (H − H)X(∞) X(∞) = H−1 (H − H)X(∞) + Y = X(∞) + H−1 Y − HX(∞) ⇒ 0 = H−1 Y − HX(∞) .l − ∆Xn−j. the results of the study in [52] indicate that the utilization of higher orders than M = 2 usually do not significantly improve the acceleration. and is the deviation from the exact solution. In principle. we must supply M + 2 successive estimates of the solution. steps (15) and (16) are repeated until a criterion is reached. which we will discuss in the next section. after every fourth iteration in the approximation of Y = HX.l − ∆Xn−i. H can be taken as a banded matrix. X(∞) = X. Acceleration of Convergence Auer [52] and Ng [53] developed methods to improve the convergence speed of linear iterative schemes. H could consist of the diagonal elements of H and zero off-diagonal elements and would thus be trivial to invert. efficient techniques have been developed for the suppression of cochannel interferers in CDMA systems [54]. ch. This slicer decides on the components of Xi+1 and forces them to +1 or −1. The coefficients αm are determined as the solutions of the following set of linear equations [53]: Macc α = bacc where (21) (13) (14) acc Mi.l is the lth component of the vector X(n) − X(n−1) .: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2161 (see [48. we implemented the second-order acceleration only (M = 2).

c < 1) are appropriate if the estimates are not yet reliable. if the signal-tointerference ratio (SIR). The most computationally expensive part of the algorithm is the calculation of the interference (24)—it is a matrix–vector product and is the reason for the O(n2 )-complexity of this method. which is defined in Section IV-B. decisions are We then order the subchannels by their SIR. This technique stems from an optimization technique called “simulated annealing” and avoids the convergence of the MSE to a local minimum [55]. the subchannels are processed in the order of their corresponding SIR (inner loop) that was calculated beforehand for the actual channel situation. These improvements are due to the determination and subtraction of the interference of all the other subcarriers. We distinguish between PIC and SIC. 4. A good choice ˆ for BPSK is the hyperbolic tangent function [55] (Xk )(i) = tanh(c · (Xk )(i) ) because far-off estimates are forced to ±1 and thus reduced in their impact on the next cancellation stage.g. . NO. These steps are identical to those of the OPT technique. the decision function approximates the signum function. . an initialization is not required. and all decisions are in the set {±1}. Thus. Let us first define a SIR for the subchannel k as SIRk = |Hkk |2 N −1 l=0 l=k where we equalize the interference-canceled received symbol by the channel response on the frequency of subcarrier k. (25) The iteration stops after the MSE falls below a predefined threshold or the number of iterations reaches a predefined maximum number. similar to the OPT technique. the techniques developed for CDMA can be reused for our application. Due to the fact that the influence of two subcarriers decreases with increasing frequency distance. In contrast to PIC. we can restrict the interference calculation to the neighboring carriers. Before entering the next round of iterations. which proposed simultaneous cancellation of several symbols in order to reduce processing delays. 56. For each subcarrier. This ensures that high reliable channels (i.. the new interference is determined. (28) The updated interference-canceled received symbol vector Y(i) is simplified to Y(i) = Y − I(i) . This algorithm is called “parallel. they could be forced to the nearest symbol of the used alphabet.. A. while estimates with small amplitudes are more or less unchanged. It is also noteworthy that the optimum PIC scheme makes use of the log-likelihood ratio from the MAP decoder [56]. Serial Interference Cancellation (SIC) The main difference between SIC and PIC is that the update steps (25) and (26) are not done in parallel for all subcarriers but sequentially one channel after the other. A proper initialization has to be done before entering the iteration loop. At the end of the iteration. the computational complexity of the algorithm is O(n2 ). Parallel Interference Cancellation (PIC) For initialization. After each estimate and decision. The basic idea behind OFDM interference cancellation is that the decisions of the symbols on the subcarriers are improved iteratively. VOL. with high SIR) are processed before the weak channels. diagH is the diagonal of H only and can be inverted with computational effort O(n). the interference (Ik )(i) = ˆ Hkl (Xl )(i−1) + l∈Xl decided in previous iteration and l=k ˆ Hkl (Xl )(i) l∈Xl decided in current iteration (30) is computed and subtracted. For each subchannel k of iteration step i.g. depending on the order in which the decisions of the subcarrier symbols are done. Performance can be further improved if c is increased from one iteration step to the next. Large c (c > 10) are appropriate for accurate estimates. Due to the vector–matrix multiplication. In the beginning. we again calculate a first estimate X(0) = (diagH)−1 · Y.e. (29) |Hkl |2 where f (·) is the nonlinear decision function. with the special assumption of H = diagH. JULY 2007 is mathematically equivalent to cochannel interference (even though it is stemming from other subcarriers). The interference Ii in step i is I(i) = HX(i−1) − diagH X(i−1) . The error vector and the MSE in the receiver are defined as ˆ = Y − H · Xi MSE = . In each iteration step (outer loop). In the most simple case. is large.. the updated estimate of the transmitted symbol on subcarrier k is (Xk )(i) = (Yk )(i) /Hkk (26) B. the interference of all other subcarriers is computed. The factor c in the hyperbolic tangent function controls the slope near zero. n − 1 simultaneously by a vector–matrix calculus. Small c (e. This technique. based on the decisions of the previous iteration. The computational effort is again O(n2 ) due to the necessity to sum up the off-diagonal elements of the channel matrix and the computation of the interference . the estimated symbols can be processed in a nonlinear fashion. That is the case for good channels with small delay spread. suggested by us in [43]. Here. More generally ˆ (Xk )(i) = f (Xk )(i) (27) . Hence.” because (25) and (26) can be applied to all channels k = 0 . (24) still unreliable so that small c is appropriate. e.2162 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY. was also used in [45].

For OFDM systems with CP. even if the CP is large enough to eliminate all ICI. while for the SIC. Note that the interference is calculated from the actual ˆ estimate X(i) compared to Section IV-A. which is defined via the relationship ∞ ∞ Fig. and SIC.45.0 in each iteration step. we analyze a static channel so that the channel is characterized by its power-delay profile (PDP) only PDP(τ ) = E |h(τ )|2 . Furthermore. four off-diagonals. We show the BER of an uncoded OFDM system for various propagation channels. 2) the PIC technique with progressively increasing values of the parameter c.. τ ). or the number of iterations becomes 30. The rest of the algorithm is similar to PIC. can be explained by the frequency . in each of which 100 OFDM symbols are transmitted (sample simulations with 500 channel realizations gave the same results). E {h∗ (t. without any nonlinear processing (decisions) and using Auer’s acceleration of convergence. The other channels were decided during the last iteration step. for the OPT. so that the number of iterations was always 30. we iterate until the MSE between subsequent iteration rounds falls below a threshold of 10−3 . OFDM system with 64 tones and no CP. BER as a function of the rms delay spread for a two-spike profile. 1. however. and 3) SIC. the impact of estimated symbols on lowSIR carriers is considered immediately. we set c = 0. the MSE threshold was not reached. namely ten. We find that the BER is lower in an equalized system without CP than in a conventional system with CP. for unequalized single-carrier systems. we use the tanh mapping function. within a single iteration step. τ )h(t . These cases will be simply called OPT. OFDM system with 64 tones and CP with 12 samples length. This fact. It is well known that. we use a depth of five.e. resulting in the familiar Rayleigh fading. Here.1 For the SIC and the PIC. We furthermore assume that the statistics of the amplitudes are zero-mean complex Gaussian. PIC. [59]. henceforth. c increases by 1. For most of the examples. a different dependence of the BER for different PDPs. we use a predetermined number of iterations. Figs. i. τ )δ(ν − ν )δ(τ − τ ) (31) assuming that the wide-sense stationary uncorrelated-scattering assumption [57] is fulfilled. We can thus expect 1 In the simulations that follows. where it is calculated ˆ from the previous estimate X(i−1) . V. ν is the Doppler frequency. The major advantage of SIC compared to PIC is that. 2. it increases by 2. we present numerical results to demonstrate the effectiveness of our proposed techniques. BER as a function of the rms delay spread for a two-spike profile. X(i) contains all the decisions of subchannels that were previously made because they have higher SIR. For the PIC. We investigate 1) the OPT. where τrms is the rms delay spread normalized to the OFDM symbol duration. we need to consider only the part of the PDP that exceeds the length of the CP. The channels are characterized by their scattering functions Ps (ν. For the OPT and the SIC (outer loop). S IMULATION R ESULTS In the following. which is astonishing at first glance. the simulations use 100 channel realizations.5 in the first iteration. 1 and 2 show the BER for a two-spike profile PDP(τ ) = δ(τ ) + δ(τ − 2τrms ) (33) We first analyze the impact of the shape of the PDP on the effect of the bit-error probability (BER) for the different equalization techniques.MOLISCH et al. τ )} exp [2πj(νt − ν t )] dtdt −∞ −∞ = Ps (ν. the BER is proportional to the squared magnitude of the rms delay spread [58].: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2163 terms. For the PIC. (32) Fig.

However. interference on the different tones. the maximum rms delay spread considered here is smaller: We investigate only cases where the maximum excess delay is smaller or equal to one symbol duration (as the exponential PDP extends to infinity. i. Finally. 8 shows the impact of using different numbers of off-diagonal elements in the iterations. OFDM system with 64 tones and no CP. eight. 1. diversity inherent in delay dispersion. Figs. A conventional system with CP discards all information “hidden” in the first G received samples. this information is detrimental. For an OFDM system without equalizer. It is noteworthy that the OPT performs better than the interference-cancellation technique in this case. 3 and 4 show the BER for rectangular and exponential PDPs. we find that. BER as a function of delay spread for OPT with different numbers of diagonals accounted for in H. Fig. Fig. Finally. 7. we require that the “significant” part of the PDP lasts only one symbol duration).e. Fig. Fig. some delay spreads can be equalized more easily than others (this was confirmed by runs with more channel realizations). which shows that good results can be achieved with a H that requires little numerical effort for its inversion. 32. JULY 2007 Fig. a system with one of our equalizers can actually make good use of this information—essentially. 5. We also observe that. 3. less sensitive to fading. OFDM system with 64 tones and no CP. We see that there is hardly any difference in the resulting BER. Note that. for those profiles. We also note that. the simple but effective approach of [60] can be used. BER as a function of the rms delay spread for a rectangular PDP. 56. if there is a CP and no equalizer. Once again. see Fig.094 (0. thus. We analyze next the performance in a scenario that does not show delay dispersion (single-tap channel) but rather only time Fig. 4. We also investigate the impact of the acceleration procedure for the OPT. in the Doppler case. respectively. The “choppiness” (nonmonotonicity) of the curves stems from the idiosyncrasies for the channel models. 4.. the impact of ICI is noticeable for all rms delay spreads larger than 0. all of the equalization techniques start to break down. thus. since it leads to ICI and. variance of the channel. for a given channel model. BER as a function of the rms delay spread for an exponential PDP. it carries information about each bit on several tones simultaneously and is. 6 shows how the acceleration steps improve of the MSE. following a “classical” Jakes Doppler spectrum. We also investigate the impact of using different approximate operators H in the OPT. for very large delay spreads. 5 shows the BER as a function of the delay spread of a rectangular PDP when H contains two. NO. VOL.2164 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY. we note that the SIC and PIC perform slightly better than the OPT. and 64 off-diagonals of H. . the OFDM symbol length being considered here is 64 samples. We plot the BER versus the rms Doppler spread (normalized to the inverse of the OFDM symbol duration) for the case of a classical Jakes Doppler spectrum SD (ν) ∝ 1 2 νmax − ν2 (34) at an SNR of 30 dB.5∗12/64) for the choice of parameters in Fig.

7. We . Blue line: With acceleration. the OPT. we used the IEEE 802. as well as the PIC and SIC.11 standard. does not exhibit a floor at SNRs up to 30 dB. 9. Performance is analyzed in a channel model F of the 802. VI. as defined in the 802. The system bandwidth was taken to be 20 MHz. Fig.0. making this case suffer from ICI due to Doppler. If the CP is 16 samples long. Black line: Without acceleration. We find that.11-like OFDM system with 64 carriers and eight samples CP. Fig.1]. maximum excess delay: One sample.11n channel models with 100-m/s velocity. Fig. and the carrier frequency was taken to be 5 GHz.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2165 Fig. two-spike PDP.11a system has 48 data carriers as well as pilots and null carriers). S UMMARY AND C ONCLUSION In this paper. We approximate the standardized system by taking a 64-carrier OFDM system (the actual 802.11n standard. Red stars: Acceleration steps. no CP. BER as a function of Doppler spread for OPT with different numbers of diagonals accounted for in H.MOLISCH et al. which was recently suggested as an option for the 802. 6. 9 plots the BER for various interference-cancellation techniques versus SNR in comparison with a case where no equalizer is present in the OFDM receiver. from this. BER as a function of the normalized Doppler spread for Jakes Doppler spectrum. as well as delay spread. if the CP is only eight samples long. BER as a function of SNR for an 802. Simulation parameters: OFDM block size: 16 samples. Impact of the acceleration of convergence onto the MSE. we performed simulations of a system similar to the IEEE 802. we have presented several algorithms for the equalization of OFDM systems that suffer from ICI. mean spike powers: [1. 64 carriers. 8. then ICI is negligible (simulation results not shown here for space Fig.11a standard.11n channel models [61] which define both the delay and angular dispersion. while the unequalized receiver shows an error floor of 10−3 . the velocity was taken to be 100 m/s (which is much higher than can usually be anticipated for 802.11a systems). In order to analyze the usefulness of our algorithms in comparison to an OFDM system without any equalizer. Single-tap PDP. the scattering function can be derived. ICI becomes noticeable. reasons). However.

2001. 1999. 2003. pp.” Furthermore. H... Zhao and S. Conf. Weinstein and P. Electr. [24] R. III-2329–III-2332. IEEE Int. Apr. 2002. Bahai. pp. “Forward error correction codes to reduce intercarrier interference in OFDM. Tong.” IEEE Trans. [17] H. Armstrong.” Proc. vol. “Effective optimisation method for channel shortening in OFDM systems. O. no. Muenster. [14] M. 566–569. IEEE Int. 2000. Kozek and A. [16] Y.. and T. Stamoulis. no. 1996. and A. IEEE Int. S. vol. O’Hara and A. and S. Chen and T. We then investigated various schemes for the efficient inversion of this process. G. “Data transmission by frequency-division multiplexing using the discrete Fourier transform. Adv. NJ: IEEE Std. M. 2001. Petrick. Ohrtmann.” IEEE Trans. Karp. Rohling. Pubs.. Speech. Yao.” IEEE Commun. 2136–2148. 2. “Intercarrier interference compression in OFDM communication systems by using correlative coding. “Nonorthogonal pulseshapes for multicarrier communications in doubly dispersive channels. pp. Z. Sun and L. Haggman. Lett. pp. vol.. 52. [28] J. Commun. Lett.. de Courville. “Optimal OFDM pulse and lattice design for doubly dispersive channels. 2002. 2000. Im. Slimane. H. vol. Moose. “An equalization technique for orthogonal frequency-division multiplexing systems in time-variant multipath channels. R. Choi. 47.. [32] K. Al-Dhahir. Symp. [36] Y.” in Proc. Jan.” in Proc. H. pp. 2078–2079. B. Seyedi and G. “OFDM. pp. Dec. pp.11 Handbook: A Designer’s Companion. 2535–2538. pp. H. M. Aug. Kim.. Apr. . Oct. 7. 4. J.15-03/267r2. Areas Commun. 2003. S. vol. 50. K. 1998. and F. [33] Y. IEEE Int. Park and G. Signals. pp.” IEEE Trans. “Frequency offset correction for coherent OFDM in wireless systems. vol. Angerbauer. Ser.. A. “A study on an MMSE ICI canceller for OFDM under Doppler-spread channel. We first established a general system model to represent all ICI contributions as off-diagonal elements of a channel matrix that multiplies the transmit symbols. Fujii. [2] T. vol. and S. “Frequency domain equalization of DMT/OFDM systems with insufficient guard interval. 42. 2908–2914.” in Proc. and E.. 1811–1819.” IEEE J. pp. 2005.” IEEE Trans.. VOL. M. pp.. pp. Wireless Communications. Hanzo. “Efficient DMT/OFDM transmission with insufficient cyclic prefix. pp. WiMax Handbook.” IEEE Trans. no. 1998. Jul. PIMRC. IEEE Globecom. no. 1185–1191.. and R. 85–90. Conf.. Jul. 2003. pp. Mar. Signal Process. Stuber. 12. vol. no. 1632–1636. M. “Robust OFDM receivers for dispersive time-varying channels: Equalization and channel acquisition. pp. 550–553. 1646–1650. 2001. OFDM and MC-CDMA for Broadband Multi-User Communications. 1999.” IEEE Trans. NO. Inst. J.” in Proc. Beaver and T. pp. Tellambura. pp. vol. “Performance of fractionally spaced linear and decisionfeedback equalizers for multitone systems. pp. Electron. “Analysis and simulation of a digital mobile channel using orthogonal frequency division multiplexing. “Low complexity ICI cancellation for OFDM systems in doubly-selective fading channels. Aug. D. J. an insufficient duration of the guard interval. Simulations demonstrated that all three schemes lead to a drastic reduction of the error floor created by ICI and can even lead to performance that is better than with long CP. Molisch. Haggman. Acoust. IEEE Symp. “A technique for orthogonal frequency division multiplexing frequency offset correction. Linnartz. pp. IEEE Pers.” IEEE Trans.. Comput. “Cyclic prefixing or zero padding for wireless multicarrier transmissions?” IEEE Trans. and J. Jeon. J. O. vol. Our simulation results show that the low-complexity equalization techniques can be effectively used in practical OFDM systems and give guidelines for their implementation. [25] S. no. 36. “Analysis and optimization of the performance of OFDM on frequency-selective time-selective fading channels. “Residual ISI cancellation for OFDM with applications to HDTV broadcasting. Gorokhov and J. Edfors. Rajatheva. [37] M. vol. B.” Electron. [3] L. Commun.. 236–240.. J. ACKNOWLEDGMENT The authors would like to thank Prof. as well as temporal variations of the channel. J. They would also like to thank the anonymous reviewers for the helpful comments. pp. J. Multi-Band OFDM Physical Layer Proposal. Commun. Commun. Chen and T. IEEE Int. IEEE Int. NJ: PrenticeHall. Eng. Paez-Borollo. [31] C. pp. 2001. [34] L. [15] D. Document IEEE 802. 2nd ed. no. no. Yao. vol. Zhao. and N.” IEEE Trans. Molisch. 56. 8. Ed.. Conf. 150. WLANs and Broadcasting. “ICI mitigation for mobile OFDM receivers. Keller. [7] F. 8.—Communications. Wireless Commun. Mobile Radio Commun. Oct...” IEEE Trans.” IEEE Trans. 2nd IEEE Workshop Signal Process. “A bandwidth efficient OFDM transmission scheme.” IEEE Trans. B. J. Voltz. COM-19. Areas Commun. “Intercarrier interference self-cancellation scheme for OFDM mobile communication systems. Cox. 10.. 3351–3355. Moeneclaey. Cimini. pp. 3. “FEQ for OFDM systems with insufficient CP. Leclercq. Aghvami. 12. F. M. 2653–2657. Muenster and L. Hoboken. Sathananathan and C. 1375–1387.. G.” in Wideband Wireless Digital Communications. Lett. Commun. 1–5. 16. Trautmann. Itami.” IEEE Commun. vol. W. vol. 27–32. May and H. A. 2003. Kim and G. vol.” in Proc. JULY 2007 considered all sources of ICI. and P. pp. 4. This approach can be improved by adding an “acceleration of convergence scheme. 187–193. no. 576–578. 2004. Commun.” IEEE J.” in Proc. Signal Process. “General self-cancellation scheme for mitigation of ICI in OFDM systems. Giannakis. Fliege.. Speidel. Duhamel. pp. Itoh. Syst. Chang. decides for the next-best tone. no.. Consum. Powers. no. Edfors for critical reading of the manuscript. Englewood Cliffs. 1994. a SIC scheme first decides about the tone with the strongest SINR. P. 47. [6] A. We first considered a linear-equalization approach. Commun. ISI can be eliminated by decision-feedback algorithms..” in Proc. “On channel estimation and detection for multicarrier signals in fast and frequency selective Rayleigh fading channel. no. and Y. Cho. Fernandez-Getino-Garcia. [8] S. pp. 2000. [26] A. Batra et al. Piscataway.. Sathananthan. Molisch. Zhu. COM-33. 35th Asilomar Conf. L. Conf. 8. M. NJ: Wiley. Commun. pp.2166 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY. 10. Indoor. no. 2005.” in Proc. while. 36–40. Pers. 365–369. Alternatively. where the inversion of the channel matrix is done iteratively by means of OPT. 1590–1599. 628–634. no. 2001. 49. “Reduction of guard interval by impulse compression for DMT modulation on twisted pair cables. and J. K. J. 16. 2451–2464. no. pp. and N. no. [27] J. G. vol. Wang. Circuits Syst. 25. G. “Second-order channel parameter estimation assisted cancellation of channel variation-induced inter-subcarrier interference in OFDM systems. 50. introducing decisions between the iteration steps leads to a PIC scheme. Sel. [23] A. Diggavi. 146–149. Stromer.. Choi. Mostofi.. [4] A. Commun. 1999. P. 2004. 665–675. R EFERENCES [1] L. [9] B. pp. F. Cassara. namely. Sep. [22] K. [35] W. Dec. “Channel equalization using one-tap DFE for wireless OFDM systems with ICI and ISI. 47. Mobile Radio Commun. 214–216. Oct. [30] S. 1971. 7. Dec. vol.. “Cancellation technique to reduce intercarrier interference in OFDM. The IEEE 802. [29] S. 2005. vol. 8. where the interference from neighboring tones (based on preliminary decisions) is subtracted in several stages. Commun. and A. 8. [19] S. Commun. 9. 1579–1589. Zhang. [5] B. [10] T.. Feb. “Intercarrier interference in MIMO OFDM. 2004. Indoor. pp. 47. 1763–1767. 572–583. pp. [18] W. Conf. Commun. Sel. Muquet. at the same time improving spectral efficiency. [13] P. 49.” in Proc. NJ: IEEE Press. 2. 5. Commun. pp. 1. C. New York: McGraw-Hill. “Analysis of new and existing methods of reducing intercarrier interference due to carrier frequency offset in OFDM. Steendam and M. D. and so on. no.. Hanzo. [20] J. P. 1998. T. Ebert. Kang. 2003. 1985. J.-S. Piscataway. 2004.” in Proc. F. Comf. Vandendorpe.. IEEE Global Telecommun. subtracts its interference from the neighboring tones.” in Proc. 2003. 2002. 1999. 2001.. Oct. [11] Y. Saulnier.. [21] Y. Schur.” in Proc. no. [12] A.” in Proc. Nakamura. 1. N. Oct. 2002. Commun.

pp. 1994. 333–342. pp.. From 2000 to 2002. Schotten. K. Mao. Sel.. “A matrix-algebraic approach to linear parallel interference cancellation in CDMA.. He was the recipient of the Vodafone U. T. Jul. Trans. 1051–1055. 336–340. [53] K. Guo. Commun. In particular. His current research interests are multiple-input–multiple-output (MIMO) systems. Lett. Commun. [44] Q. 2000. 1989. [43] M. Molisch. “Indoor MIMO WLAN channel models. 9. Conf. Areas Commun. Sameer Vermani received the B. Sweden. Dec. “Acceleration of convergence. pp. 2003. Leus. Austria. AT&T (Bell) Laboratories Research. 2000. “ICI cancellation in OFDM wireless communication systems. Bogucka and K. “Frequency-domain echo cancellation in digital multicarrier modulation systems. [59] A. Phys. where he develops smart-antenna radio network planning and optimization tools. 11 book chapters. Dec. 61. vol. 9. no. Chem. in 2006.: Cambridge Univ. no. Schniter. 1974. 1950–1954. no. Ed. pp. [46] B. IEEE Veh. vol. His research focused on signal processing. Cambridge..” “COST 259. 2003.Techn. New York: Prentice-Hall.” J. Commun.K. pp. 2002. D. 3377–3381. some 100 journal papers. orthogonal frequency-division multiplexing and smart-antenna radio-channel measurements. He was Chairman of the IEEE 802.K. P. 2047–2056. Telecommun. Rasmussen. W. Shen.S. 2000. coauthored. 2005. Baum. Molisch is an Editor of the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. Conf.. Cambridge. Andreas F. 1963. Banelli. Cai and G.” IEEE Trans. pp. IEEE Int. 1002–1011. B. He has participated in the European research initiatives “COST 231. U. and A. F. pp. Ed. and S. Dr. respectively.—Spring. and X. T.) degree in communications engineering and the Dr. and wideband systems.S.” IEEE Trans.. IV-237–IV-240. measurement and modeling of mobile-radio channels. “Equalization of OFDM-systems by interference cancellation techniques. Lim. pp. he founded SYMENA Software and Consulting Ltd. J. J. Johansson. PIMRC.” and “COST 273. Baltimore. Busboom. From 1991 to 2000. 1996. where he is now a Distinguished Member of Technical Staff. .. In 2002. Wiley–IEEE Press). degree in electrical engineering from the University of California. Press. U. [61] V. 152–161. Standardization Drafts IEEE 802 Meeting. SAC-5. Signal Process. Oct. Molisch. Rugini. Multiuser Detection. pp. Toeltsch and A. Xuerong and Z.. H. and ultrawideband (UWB) systems. [40] M. “Iterative cochannel interference cancellation in synchronous CDMA on a frequency selective channel. [47] J. 2004. Middletown. “Channel modeling and associated intercarrier interference equalization for OFDM systems with high Doppler spread. 1987. “Low-complexity equalization of OFDM in doubly selective channels. he was with the Wireless Systems Research Department. 1363–1367.” in Proc. Technol. Dallas. Verdu. Jan. S. P. and 1999.” IEEE J. [60] L. ICC. Commun. he was with the Technical University of Vienna.. and K. Cambridge.. in 1998 and 2002. pp..” Eur. and numerous conference contributions. Jun. Chuang. Press. Dr. Giannakis. 7. Commun.. 2003. C. [41] H. 1. Xu. degree from the Technische Universität Wien (TU-Wien). [54] S. Guillaud and D. Guest Editor of a recent Special Issue on UWB (in the IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS). Y. and T. 48. TU-Wien. Vienna. Since then. Slock. Int. he worked in the field of signal and image processing for medical applications.” in Proc. no. Lund. he was a member of the Mobile Communications Group at the Institut für Nachrichten-und Hochfrequenztechnik. 2003. 52. W.” where he was Chairman of the MIMO channel working group. [50] L. degree (with the institute silver medal) in electrical engineering from the Indian Institute of Technology. 7. Cai.MOLISCH et al. L. no. [48] G. Wesolowski.. 632–636. Vienna. He has authored. vol. He is also a Professor and Chairholder for radio systems at Lund University. His research interests span the physical layer of wirelesscommunication systems. Kalkofen. pp.” IEEE Trans. NJ. [39] P. 879–888. Bello. 4. in 1990. J. cooperative communications. IEEE ICASSP. 51.Tech. Molisch. TPC Chair for Chinacom 2007. vol. he has done work in the areas of multiple-input–multiple-output and orthogonal frequency-division multiplexing systems. Mark. Austria. 2nd ed. Suyama. COM-11. Commun. Erceg. Matrix Computations. Yang. Kyritsi. Ng. Conf. [56] S. 5th IEEE Conf. MA. and Habilitation degrees from the Technical University of Vienna. Berkeley.: ITERATIVE METHODS FOR CANCELLATION OF ICI IN OFDM SYSTEMS 2167 [38] X. In his doctoral thesis. F. 2000. H.K. Rasmussen. Vienna. vol. Toeltsch and A. “A matrix-algebraic approach to successive interference cancellation in CDMA. Jan. pp.Tech. He has been a member of numerous Technical Program Committees (TPCs). pp.” IEEE Trans.4a channel model standardization group and is also Chairman of Commission C (signals and systems) of the International Union of Radio Scientists. Ed. (M. and A. Sun. no. vol. Gorokhov. 4. Mar. Technol. pp. no. 1998. atomic line filters. Apr. India. L. U. 360–393. 1. pp.15. vol. 2003. 48. and Wireless Symposium Chair for Globecom 2007.” IEEE Trans. TX.” IEEE Commun. [49] H. “An OFDM receiver employing turbo equalization for multipath environments with delay spread greater than the guard interval. 145–151. [58] J. From 1998 to 2002. [52] L. 656–660.. [45] J. he has been with Mitsubishi Electric Research Laboratories. F. or edited four books (among them is the recent textbook Wireless Communications..: Cambridge Univ.S. Foundation Fellowship for graduate study at Berkeley. Ing. “Interchannel interference cancellation in wireless OFDM systems via Gauss–Seidel method. Vice Chair of the TPC of Vehicular Technology Conference (VTC) 2005 spring. 427–438. [51] D. Int. 2. pp. 2001. respectively. He has done research in the areas of surface-acoustic-wave filters. Univ. as well as propagation and channel modeling. [42] M. “Non-local perturbation techniques. Molisch (S’89–M’95–SM’00–F’05) received the Dipl. Raphaeli. 1998. 1984. Suzuki. Delhi. “A multi-carrier detection algorithm for OFDM systems without guard time. Conf. no. San Diego.. K. vol. W. Fukawa. and A.” in Methods in Radiative Transfer. F.” in Proc. D. “Bounding performance and suppressing intercarrier interference in wireless mobile OFDM. “Efficient OFDM transmission without cyclic prefix over frequency-selective channels. 2680–2689. Molisch. General Chair of ICUWB 2006. MD: Johns Hopkins Univ. L.. Lim. Schumacher. vol. “Simple equalization of time-varying channels for OFDM. He is an IEEE Distinguished Lecturer and recipient of several awards. smart antennas. Press. he investigated high-resolution evaluations of directionally resolved radio-channel measurements. From 1992 to 1998. “The effects of time delay spread on portable radio communications channels with digital modulation. He is currently with Qualcomm Inc. radiative transfer in atomic vapors. IEEE Global Telecomm. 40. S. CA. [55] D. A.” in Proc. 5.” in Proc.” in Proc.” in Proc. “Efficient implementation of linear multiuser detectors for asynchronous CDMA systems by linear interference cancellation.” IEEE Trans. “Hypernetted chain solutions for the classical onecomponent plasma up to gamma = 7000. Press. Golub and C. no. van Loan. most recently as an Associate Professor. Kalkofen. “Characterization of randomly time-variant linear channels. Pers. Martin Toeltsch (M’00) received the Dipl. Cambridge. and the M. no. Auer.” in Proc.” in Numerical Radiative Transfer. 2003. and G. Wideband Wireless Digital Communications. [57] P. vol. Lijun. M. 2000. Cannon.: Cambridge Univ. 5. pp.” in Proc. 619–621.Ing. Commun. C. Elders-Boll. 1987. pp. Commun.. 12. Feb.