IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO.

4, APRIL 2005 677

Multiple-Access Interference Suppression for

MC-CDMA by Frequency-Domain Oversampling
Brandon Hombs, Student Member, IEEE, and James S. Lehnert, Fellow, IEEE

AbstractA technique for interference suppression in mul- but they require substantial knowledge about the interfering sig-
ticarrier code-division multiple-access (MC-CDMA) systems is nals and spreading sequences. This is especially true when the
proposed which exploits the structural differences in signals spreading sequences are aperiodic, because the optimal detector
that arrive at the receiver with Doppler shifts or carrier offsets.
Oversampling the received signal in the frequency domain and changes significantly from bit to bit [4]. Due to the complexity
properly combining the samples provides the interference sup- of MUD algorithms, there has been research on excising in-
pression. Frequency-domain oversampling is accomplished by terference using single-user detectors that exploit structural
using a time extension of the conventional MC-CDMA signal or differences in the signals of different users. Techniques that
unconsumed portions of the cyclic prefix. Furthermore, a receiver
have been successful in direct-sequence code-division multiple
structure is introduced that despreads and combines groups of
samples so that a linear minimum mean-squared error solution access (DS-CDMA) suppress MAI using limited knowledge
for combining the groups is easily found. This combining scheme about the interfering signals. Advantages provided by these
increases the signal-to-interference ratio experienced by the techniques relative to multiuser detectors are lower complexity,
desired user in the MC-CDMA system. In addition, the receiver lower requirements for system-wide knowledge, and possibil-
performs well in severe nearfar scenarios when there is sufficient
frequency separation between the signals of the desired user and
ities for simple adaptive implementations. Since the symbol
an interferer. Numerical results show that the proposed receiver matched filter is the optimal receiver for the additive white
significantly outperforms the conventional MC-CDMA receiver in Gaussian noise (AWGN) channel, the techniques for MAI
many channel realizations. suppression often focus on processing the output of the symbol
Index TermsCode-division multiple access (CDMA), Doppler matched filter. In [5], the authors oversample the output of
effect, interference suppression. a symbol matched filter in the time domain and combine the
samples to maximize several different performance metrics in
a DS-CDMA system utilizing long pseudorandom spreading
I. INTRODUCTION
sequences. These techniques are especially effective when

R ECENTLY, there has been significant interest in multicar-

rier and code-division multiple-access (CDMA) systems.
Because efficient fast Fourier transform (FFT) techniques and
there are relatively few users occupying the channel, or when
there is severe nearfar interference, because the interference
is strongly cyclostationary and should not be modeled as a
high-speed digital signal processors are available, multicarrier wide-sense stationary Gaussian process. In addition, the op-
systems are now being developed for high-data-rate communi- timal combining solution does not change from bit to bit as it
cations. The convergence of multicarrier and CDMA systems is does for MUD.
based on the application of spreading codes to orthogonal fre- Although interference suppression for DS-CDMA has
quency-division multiplexing (OFDM). The use of multicarrier been well studied, techniques for interference suppression
code-division multiple access (MC-CDMA) was first proposed in MC-CDMA have not been thoroughly explored. With
in [1] and [2], where spreading sequences are used in the fre- increasing mobility of users and the desire to increase the
quency domain instead of the time domain. Since the appear- processing gain in MC-CDMA systems, frequency shifts of
ance of [1] and [2], there has been significant work on improving users that result from Doppler shifts or oscillator inaccuracies
the performance of these systems for different channel condi- may soon become significant, and could even approach the
tions. An overview and comparison of MC-CDMA techniques subcarrier frequency spacing. Here, we show that large perfor-
is given in [3]. mance enhancements are possible if the structural differences
In CDMA systems, multiple-access interference (MAI) is of interfering users caused by frequency shifts at the receiver
one of the main sources of interference. Multiuser detection are exploited. The results for DS-CDMA in [5] can serve as
(MUD) techniques can successfully reduce the effects of MAI, a good tool for understanding the time-domain dual of the
proposed system that we explore in this paper.
In this paper, we propose and examine the interference-sup-
Paper approved by J. Wang, the Editor for Equalization of the IEEE Com- pression capability of a new MC-CDMA receiver that oversam-
munications Society. Manuscript received July 18, 2003; revised May 12, 2004.
This paper was presented in part at the 38th Asilomar Conference on Signals, ples the received signal in the frequency domain by exploiting
Systems, and Computers, Pacific Grove, CA, November 2004. the unconsumed portion of the cyclic prefix, or by using a mod-
The authors are with the School of Electrical and Computer Engi- ified transmission scheme. The linear minimum mean-squared
neering, Purdue University, West Lafayette, IN 47907 USA (e-mail:
[email protected]; [email protected]). error (MMSE) solution for combining samples is found using an
Digital Object Identifier 10.1109/TCOMM.2005.844964 effective algorithm with low complexity. The solution is greatly
0090-6778/$20.00 2005 IEEE
678 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 4, APRIL 2005

simplified by combining samples in groups before finding the of each symbol to samples. Here, samples are used
MMSE solution. as a cyclic prefix or guard interval for channel equalization at
The best results are obtained when there are significant fre- the receiver, and the value of depends on the length of the
quency shifts in the signals received from different users. The channel response. The received signal contains
results are best applied to an uplink MC-CDMA system, since samples which contribute to the decision statistic, where is
pseudorandom codes have good cross-correlation properties called the number of excess samples.
and each user experiences a different channel. Advantages The multiplicative window, , used in the time domain
of the proposed receiver over receivers performing MUD are to limit the support to one symbol, corresponds to convolution
lower complexity and fewer requirements for information about of the Fourier transform of with the Fourier transform of
the interfering users signals. If small changes to the trans- . However, for fixed corresponds to a sequence
mitted signal are made, significant improvements in system of impulses in the frequency domain. Hence, transmitter win-
performance can be achieved with minor increases in receiver dowing can be interpreted as centering the Fourier transform of
complexity. the window, given by
The remainder of the paper is organized as follows. In
Section II, the structure of the transmitter is examined. Next,
(2)
in Section III, the channel model is explained, and a receiver
structure that maximizes the signal-to-interference ratio (SIR)
for a particular user is developed. Combining and detection at each subcarrier and multiplying by the appropriate element of
methods are described in Section IV. Numerical examples the spreading sequence and the information symbol. With this
with a discussion of the optimization of the multiple-access framework, can be viewed as the frequency-domain dual
capability are provided in Section V. Finally, in Section VI, of the time-domain chip waveform in DS-CDMA.
bit-error rate (BER) simulations are presented. Conclusions are The th user transmits the complex baseband signal
drawn in Section VII. given by

II. TRANSMITTER MODEL

We consider a standard multicarrier transmitter implemented (3)
with discrete Fourier transform (DFT) techniques. The trans-
mitter is similar to an OFDM transmitter. However, in the where is the transmission power, , and
MC-CDMA system, the same information symbol is sent on all is a pulse that satisfies the Nyquist criterion for zero
subcarriers with different elements of the spreading sequence intersymbol interference (ISI) (e.g., square-root raised-cosine)
used on each subcarrier. A constant symbol time and band- when processed by a matched filter and sampled at . In addi-
width is assumed for all systems considered in this paper so tion, . Since the pulses in (3) are generated
that a fair comparison is ensured. The spreading codes used at the rate , the subcarriers are separated by in the
for the MC-CDMA system are modeled as independent and frequency domain upon transmission.
identically distributed (i.i.d.) random variables. The th user For clarity, we note that the structure of the transmitted sig-
is assigned the spreading sequence with chosen nals can be modified slightly to allow for virtual carriers, as de-
from the set with equal probability. The scribed in [6], by setting several subcarriers to zero before taking
th user transmits the i.i.d. information sequence with the inverse discrete Fourier transform (IDFT) of the spreading
chosen from the set with equal probability and sequence, as shown in (1). In this case, the total number of sub-
independently of . carriers (active and virtual) is still , although the actual pro-
Because the spreading sequence is used in the frequency do- cessing gain is the number of active subcarriers. In all cases, the
main in an MC-CDMA system, it is convenient to define the active carriers are all in the flat part of the spectrum of .
time-domain sequence associated with the frequency-domain Since virtual carriers lead to notational complexity, we assume
information symbol. The th transmitted information symbol for the remainder of the paper that no virtual carriers are used. In
generates the time-domain sequence addition, the subcarrier separation can be decreased, and a high
data rate can still be maintained if subsets of the subcarriers are
used for different bits, as is done in [7]. This is desirable because
(1) smaller subcarrier spacing implies that less frequency offset is
needed to take advantage of the structural differences in signals.
Note that for fixed is periodic in with period , Since this is a logical and simple extension, for clarity, we now
where is the processing gain. Thus, the frequency-domain consider only one bit per multicarrier symbol.
representation of is simply an impulse at each sub-
carrier, located at multiples of in the frequency domain,
III. CHANNEL AND RECEIVER MODEL
multiplied by the corresponding element of the spreading
sequence and the th information bit. Before transmission, the After transmission, each users signal experiences a time
sequence is windowed by to limit the support delay and frequency shift as it propagates through the channel
HOMBS AND LEHNERT: MULTIPLE-ACCESS INTERFERENCE SUPPRESSION FOR MC-CDMA 679

Fig. 1. Receiver model. The numbers in brackets are used to identify elements of vectors or sequences. The final decision statistic is ^
b for detection of b .

to the receiver. The impulse response of the channel for the th effect is even smaller, because the active subcarriers are all in
user is given by the flat part of the spectrum of [6].
After filtering, the signal is sampled at time intervals of
(4) and the sampled signal is denoted by . Next, the samples
are windowed so that only samples associated with one symbol
remain. The receiver window, denoted by , has support
where , and are the time delay, carrier offset, over samples, and has a spectrum given by
and path gain experienced by the th user in the th channel
path, respectively. This impulse response is constructed using
the channel convention given in [8]. When , this model (6)
adequately describes a system where users are mobile and
experience Doppler shifts relative to other users. In addition, if To maintain orthogonality among subcarriers, conventional
, it is useful for modeling wireless local area networks, systems use a rectangular window of length . However, it
where users may have oscillator differences and experience is shown in [9] that using excess samples can provide useful
multipath fading. If there are a total of simultaneous trans- information for demodulation and detection in OFDM systems.
mitters, the received signal is given by Furthermore, longer windows that satisfy a dual version of the
Nyquist condition for zero ISI can be chosen to maintain zero
(5) inter-carrier interference (ICI) at the receiver. Note that the
need for zero ICI is not as important in MC-CDMA systems as
it is in OFDM systems, because the subcarriers in MC-CDMA
where is a complex AWGN process with i.i.d. real and are all modulated by the same information symbol. Therefore,
imaginary parts, each with power spectral density . compensation can be made for a window that induces ICI at the
Without loss of generality, in the remainder of the paper, we receiver, because the relationship between subcarriers is com-
assume that the receiver detects the information transmitted by pletely specified by knowledge of the desired users spreading
the first user. All other signals are considered interference. We sequence and the windows used at the transmitter and receiver.
also assume that the receiver can accurately track the channel re- The combined effect of the windows at the transmitter and
sponse of the desired user so we can set and receiver is that the spectrum of the received signal is a linear
without loss of generality. Furthermore, we make the assump- combination of shifted versions of their combined frequency
tion that the system is quasi-synchronous, as described in [3]. response given by , where represents
This ensures that the signal is analyzed over a single symbol periodic convolution as described in [10]. This is because the
interval for all users. With this assumption, the receiver can de- transmitter sends the periodic discrete-time sequence described
modulate the signals of all users by taking one FFT after sam- in (1), which is then multiplied by both and . The
pling, thereby decreasing complexity. The cyclic prefix can desired users part of the received signal spectrum after receiver
be chosen based on the delay spread of the channel and the windowing is , where
amount of timing ambiguity allowed by the quasi-synchronous
assumption. The cyclic prefix length is chosen so that
, where is the timing ambiguity allowed (7)
by the quasi-synchronous timing system.
Fig. 1 depicts the receiver structure. We proceed by de- is the desired users spectrum and is the transfer function
scribing the received signal at each block in the receiver. First, of the channel experienced by user 1. For the desired user, we
the received signal is filtered by the matched filter to the trans- can write
mitted pulse, which has impulse response given by ,
where the superscript represents complex conjugation. We
assume that the frequency shift experienced by any user is
small when compared to the total bandwidth occupied by the
transmitted signal. In this case, the effect of the frequency (8)
shift on each users signal is negligible, as it is filtered by
. If virtual carriers are used at the transmitter, the where only has one argument, because .
680 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 4, APRIL 2005

The discrete-time Fourier transform (DTFT) of the windowed the optimal combining coefficients for all samples. We
signal corresponding to the th information symbol is given by later establish that the reduction of the optimization problem to
a smaller set of coefficients still provides significant MAI sup-
(9) pression. After the FFT in Fig. 1, a vector is formed for each
group of samples. The th element of the th vector is given
by
The spectrum is sampled by taking the FFT of the sequence
. In contrast to conventional multicarrier
(10)
systems, the use of excess samples allows for sampling in the
frequency domain at intervals closer than the carrier spacing.
Hence, an FFT larger than is computed. Since for and for group indexes
has support over samples, a -point FFT is suffi- . To represent (7) in matrix form, we define the
cient to completely specify its spectrum. Nevertheless, for con- matrix , where the superscript signifies that this matrix
venience and to aid understanding, we zero pad is associated with the first user, and the subscript signifies that
so that the length of the FFT is a multiple of . If we this matrix is used for the th group of samples. This matrix
zero pad to samples, where is an integer, there will be is used to account for the effect of the transmitter and receiver
samples in the frequency domain, a subset of which cor- windows. The th element of this matrix is given by
respond exactly to the frequencies of the subcarriers.
Notice that in this MC-CDMA system only samples (11)
are needed to successfully despread the signal at the receiver
because there are subcarriers. A conventional OFDM-type for ; and .
receiver will discard the extra samples to restore orthogonality Notice that is simply a Toeplitz matrix that shows how to
among subcarriers at the receiver [11]. We define as combine components of the spreading sequence to determine
the excess time because it is the ratio of the extra time that what value a sample in the frequency domain would have in an
is not needed for despreading divided by the time needed for interference-free environment. We let be an -dimensional
despreading.1 For fair comparison, the length of the interval, , vector of the spreading sequence used for the current informa-
is held constant for all systems examined in this paper so that tion symbol transmitted by user 1, and let be an
all use the same symbol time. To increase the processing gain diagonal matrix that specifies the channel effects. The diagonal
of a system, the excess time is decreased, and vice versa. For elements of are given by
example, consider the case when . This results
in twice as much spacing between carriers when compared (12)
to conventional MC-CDMA with the same symbol time (i.e.,
putting a null carrier between each active carrier). It is shown with and . The despreading
in the following that as excess time increases, processing gain vector for the th group is defined as .
decreases, but gains at the receiver are still realized, because The vector contains despreading values in the frequency
using the excess time results in beneficial structure of the domain. The values correspond to frequencies separated by
interfering signals. with the first element corresponding to the frequency
In conventional MC-CDMA, the frequency-domain samples . The despreading operation is shown in Fig. 2.
are despread by multiplying each sample by the complex con- The samples contained in are despread and combined ac-
jugate of the element of the spreading sequence that was mod- cording to , where the superscript stands for the
ulated onto the corresponding subcarrier. This is described in Hermitian transpose. Hence, can be viewed as a decision
[13]. However, in the proposed receiver, there are more samples statistic generated by the th group of samples.
than carriers, and some samples are not centered at the carrier
frequencies. Despite this difference, each sample has a known IV. COMBINING AND DETECTION
contribution from the desired user with the only ambiguity re-
The sample values from the group despreader are com-
sulting from the multiplication by . Hence, each sample is
bined to make bit decisions based on the linear MMSE com-
despread by its value when noise and interference are absent
bining criterion to maximize the SIR. The vector of combining
and . This is a form of discrete matched filtering or
coefficients is denoted by , and the goal is to minimize
maximal ratio combining. The samples of the received signal
, where is the vector containing the components .
spectrum after computing a zero-padded FFT are ,
To find the optimal choice of , the autocorrelation matrix of
for . Hence, the samples can be despread
the received vector after despreading must be found to solve the
by multiplying the th sample by .
Wiener-Hopf equations for the linear MMSE solution. The final
For convenience, the system can be described in vector form.
decision statistic is , and the users information
We drop the symbol index, , because the processing is over only
symbol is decided based on the sign of this statistic.
one symbol. Since there are samples in the frequency do-
main and there are subcarriers, there are samples between
A. Statistics of the Received Signal
each subcarrier. The samples are grouped into groups of
samples for combining. This simplifies the problem of finding The received vector , after despreading, can be decomposed
into three components
1Later, we show that excess time in MC-CDMA has a role similar to excess
bandwidth in DS-CDMA systems [12]. (13)
HOMBS AND LEHNERT: MULTIPLE-ACCESS INTERFERENCE SUPPRESSION FOR MC-CDMA 681

Since a zero-padded FFT is used, there are samples of

to use for processing. The noise component on the th
branch of the despreader is
(17)
where the th element of the th vector is

(18)

for and . Finally, the

autocorrelation matrix of is given by
(19)
where is the vector of components . A matrix form, sim-
ilar to (11), can be found for describing the noise components.
In (16), is viewed as a signal with period , since
only has support over samples. In this case, is discrete
in frequency, and the windowing can be interpreted as a convo-
lution of the DTFTs of and . Since is modeled
to have period , the FFT can be used to find the spectrum. In
Fig. 2. This figure illustrates the operation of the despreader. Brackets are used this case, the FFT of leads to frequency-domain samples
to index elements of vectors (i.e., p [0] is the first element of vector p ). that are i.i.d. Gaussian random variables with zero mean and
variance per dimension (assuming the FFT is normalized
in power). The values of the frequency-domain samples due to
where is the contribution of the desired user, is the con- noise before the despreader can be found by using an
tribution of the interferers, and is the contribution from white matrix , with the th element given by
noise. Since these components are mutually uncorrelated, the
autocorrelation matrix of can be found from the individual (20)
autocorrelation matrices. The autocorrelation of is given by
for ; and .
(14) This matrix is used to find the correlation of the white noise
components. Note that, after the FFT, the noise samples are
where is the autocorrelation matrix of is the auto-
close enough in frequency that there can be correlation between
correlation matrix of , and is the autocorrelation matrix of
samples. After despreading, the th element of the noise au-
.
tocorrelation matrix is given by
Since the contribution of the desired user to is
, where is the Euclidean norm of the (21)
vector, the th element of the autocorrelation matrix at
the output of the despreader is given by where is the trace of the matrix.
The final task is to find the autocorrelation matrix of the MAI
(15) term after the despreader. It is assumed that the frequency shift
of the interferer is negligible when compared with the band-
To calculate the autocorrelation matrix of the noise and width of . Hence, when the signal is sampled, the sam-
interference terms, the structure of the components after the ples correspond to the subcarriers that the interferer transmitted
despreader in Fig. 1 must be determined. We focus first on shifted in frequency by .
the autocorrelation matrix of the white noise term. After the The DTFT of the th users signal for the th bit is
matched filter and sampler, there are samples from filtered , where is given by
white noise with zero correlation between samples. Hence, the
samples are i.i.d. complex Gaussian random variables with mean
zero and variance per dimension. The real and imaginary
parts are independent. The samples can be represented in the (22)
frequency domain as the DTFT of i.i.d. Gaussian random
variables multiplied by the window . The spectrum of and is defined as in (8), but for the th user. To main-
these samples is given by tain the form described in (7) and (8), the frequency shift of the
th user is accounted for in the term , and the channel
frequency selectivity is accounted for in the term . The
(16) frequency samples before the despreader can be evaluated with
a matrix similar to . We define
where the components are the white noise components after
(23)
matched filtering and sampling.
682 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 4, APRIL 2005

for ; and .
The spreading sequence for the th symbol transmitted by user
is defined as the -dimensional vector, . The interference
term for the th branch of the despreader is

(24)

where indexes the branch of the de-

spreader. The autocorrelation matrix of is given by

(25)

After averaging, the th element of the autocorrelation ma-

trix for the MAI can be expressed as
Fig. 3. SIR for the case P = 64 1
; f T = 02: ; E =N = 15 dB, and an
interferer that has 30 dB more power than the desired user.

(26) In addition, it can be shown that the SIR resulting from the linear
MMSE combining solution is the maximum SIR over all linear
filters in (27). Hence, the MMSE solution is optimal for BER as
B. MMSE Combining long as the CGA holds.
To find the optimal combining vector so that the decision
statistic can be formed, the solution to
V. NUMERICAL EXAMPLES
(27)
Numerical examples are now presented. It is assumed that
is required. Since the autocorrelation matrix of the received each user experiences a frequency shift in the channel, but no
vector after despreading is available, the optimal can be found multipath fading. The performance of this system is evaluated
using the WienerHopf equations [14]. The optimal combining for several different scenarios and compared with conventional
vector is given by MC-CDMA. The trade-off is examined between processing gain
and excess time while maintaining data rate. In each case, con-
(28)
ventional MC-CDMA corresponds to the case in which excess
The SIR of the decision statistic with MMSE combining is given time is equal to zero.
by The first example illustrates the effectiveness of the receiver
at overcoming severe interference from one interfering user. The
interferer has 30 dB more power than the desired user. The SIR
(29) is examined as the frequency shift of the interferer changes. In
addition, the effects of choosing different values for excess time,
, are illustrated. To ensure that the symbol period is held
A key figure of merit is the BER of the system. The conditional constant for each system realization, is set at 64. In all exam-
Gaussian Approximation (CGA) [15] can be used to evaluate the ples, we use a rectangular window of length for both the trans-
BER of the system. Although [15] is focused on DS-CDMA, the mitter and receiver windows, defined by and , and
results still hold for MC-CDMA. Similar results to those of [15] 15 dB, where is the energy per bit of the desired
can be developed and applied in the frequency domain instead of user.
the time domain. Since the output of the filter is approximately
Fig. 3 shows the performance of the receiver for the case of
Gaussian for modestly high processing gains, conditioned on
one interferer with a relative frequency shift of .
the carrier offset and channel state, we use the equation
The performance becomes dramatically better as more fre-
(30) quency-domain samples are used and there is some excess time.
In fact, even with a small excess time, the gain is substantial.
to predict the error performance, where Fig. 3 shows that for the given frequency shift it is best to use
. When this approximation holds, the fewer subcarriers and use some excess time with at least three
maximization of the SIR is a good optimization criterion times oversampling at the receiver. Note that the oversampling
because the -function is a monotonically decreasing function. factor can be chosen at the receiver without any impact on
HOMBS AND LEHNERT: MULTIPLE-ACCESS INTERFERENCE SUPPRESSION FOR MC-CDMA 683

Fig. 4. SIR for the case P = 64;  = 16 (i.e., =N = 1=3); E =N = 15 Fig. 6. SIR for the case P = 64; E =N = 15 dB, 10 interferers with equal
dB, and an interferer that has 30 dB more power than the desired user. power, and random frequency shifts.

The second example shows the performance for a system with

11 users, all with equal received power. The frequency shift is
modeled as a uniform random variable on for
each user, and the performance is averaged over 10 000 real-
izations. Fig. 6 shows that even when there are several users in
the system, the processing gain can be reduced and a good SIR
can still be maintained if the information provided by the ex-
cess time is used. The standard CDMA equation for evaluation
of SIR predicts that the SIR for the conventional receiver would
be2

(31)

where is the received energy per bit of the th user. The

results can be checked with this equation when excess time is
zero (since this is the conventional MC-CDMA case). For the
example, for all , and
dB. Hence, (31) predicts that the conventional MC-CDMA
Fig. 5. SIR for the case P = 64; M = 3; E =N = 15 dB, and an interferer system should have dB. This matches well with
with 30 dB more power than the desired user. the results shown in Fig. 6 when excess time is zero.
The final example shows the performance for a single inter-
ferer with 10 dB more power than the desired user. Once again,
the choice of processing gain and excess time. The only limiting the frequency shifts of the users are modeled as uniform random
factor in choosing is the amount of processing power that is variables on and the performance is averaged
available for performing the FFT at the receiver. over 10 000 realizations. The interference power is the same in
Fig. 4 illustrates the performance as the relative frequency this example and the previous example, but the performances
shift of the interferer changes for a given excess time. The per- differ because the interference has significantly more structure.
formance is best when the frequency shift of the interfering user Fig. 7 shows that there is the potential for a 3 dB increase in SIR
is large. Also, Fig. 4 shows that for this choice of and , if the system is designed to utilize excess time. The gain is not
there are diminishing returns in oversampling at rates beyond as dramatic in the 11-user case because the interference appears
those that result when . more like a wide-sense stationary Gaussian random process as
Fig. 5 shows the effect of . Notice that, even for a small , a result of the Central Limit Theorem. The smaller gain of the
there are substantial returns in SIR, but with excess time equal proposed approach is expected when the total power of the inter-
to zero, there is no gain from oversampling. This is because the ference is divided among many interferers, because the optimal
desired users signal and the interfering users signal occupy the
2This assumes the use of random spreading and a sinc pulse for the chip
same signal space, no matter what the frequency shift, since both
waveform (rectangular transmitter and receiver windows). This is explained in
are just linear combinations of shifted versions of , which terms of the interference function in [12] for DS-CDMA systems, but the fre-
is an aliased sinc function in this case. quency-domain dual holds for MC-CDMA.
684 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 4, APRIL 2005

Fig. 7. SIR for the case P = 64; E =N = 15 dB, one interferer with 10 dB Fig. 8. Plot showing that the CGA is an accurate predictor for the BER.
more power than the desired user, and random frequency shifts. For K = 11, the carrier offsets for the interfering users are 0; 0:1=T ;
0:2=T ; . . . ; 0:9=T with all powers equal to that of the desired user. For
K = 2, the carrier offset of the interferer is 0:3=T with 30 dB more power
receiver in white Gaussian noise is the conventional matched than the desired user.
filter receiver.
The results of these evaluations show that there is a trade-off
between excess time and processing gain. A similar trade-off
is examined in [12] between processing gain and excess band-
width for DS-CDMA. The interference suppression capability
from the processing gain comes from the separation of the sub-
spaces (or code spaces) occupied by the users. The interference
suppression capability from excess time comes from the fact that
the frequency shifts create enough difference in the structure of
the signals that the receiver can exploit the differences.

VI. SIMULATION AND COMPARISON

Here, we show simulation results for and
with every users path gain equal to 1 so that the CGA applies.
The performance is shown for a system with 11 users and
, as well as for two users and . Using the SIR derived
above and the CGA, we plot the predicted BER and the BER
obtained from simulation in Fig. 8. The results match up very Fig. 9. BER for the case P = 64; E =N = 15 dB, and one interferer with
well because the processing gain is high enough for the Central 30 times more power than the desired user. Several different systems are shown
Limit Theorem to be useful. This figure validates the use of SIR for comparison.
as a performance metric.
Next, the performance of the proposed system is simulated than any other system, as long as there are small carrier differ-
in a multipath channel in which users may have small oscil- ences, but with the cost of requiring more system-wide knowl-
lator differences that cause frequency offsets. For this example, edge and computational complexity. However, as the frequency
a multipath channel with paths and an exponential offsets become more severe, the performance of MUD quickly
power delay profile is used. The fading coefficients are mod- degrades because it does not compensate for the frequency off-
eled as complex Gaussian random variables with normalized sets of the interfering users when calculating the MMSE solu-
power, , that is, the total average tion. The sensitivity to carrier offsets could be remedied by using
channel gain is 1. For this simulation, three times oversampling an (frequency-domain) asynchronous linear MMSE MUD, but
is used for all systems with nonzero excess time. As this would further increase the complexity of the receiver.
before, for all systems. A comparison of the perfor- Fig. 9 shows the performance of these systems and demon-
mance of the proposed system to a system that uses a linear strates the robustness of the proposed system to the near-far
MMSE MUD [16] with no excess time, 64 carriers, and perfect problem without any knowledge of the interfering users car-
channel knowledge is given. As expected with the assumption rier offset or spreading sequence. A frame size of 1000 bits is
of perfect channel knowledge, the MUD performs much better used for all cases, and the performance is averaged over 10 000
HOMBS AND LEHNERT: MULTIPLE-ACCESS INTERFERENCE SUPPRESSION FOR MC-CDMA 685

channel realizations for each value of . In addition to using samples in the frequency domain. Furthermore, the excess time
the theoretical autocorrelation matrix derived in Section IV, the information provides significant improvement in near-far situ-
performance is simulated using the unbiased estimator of the ations, which is a very important scenario in CDMA systems
autocorrelation matrix. The estimate is given by where users are highly mobile or where users may require dif-
ferent rates.
(32) A new transmission scheme has been proposed that can lead
to an interference suppression capability by a receiver. In ad-
dition, the MMSE combining solution has been developed for
Here, is the current symbol in the frame that is being detected, a simple despreading technique. The BER performance is well
so (32) is the standard approximation for the autocorrelation approximated using the CGA and the derived SIR. Simulations
matrix given the previous samples. To get initial estimates of of the system with multipath fading show that gains are realized
the autocorrelation matrix, 10 training symbols are used before at the receiver with perfect channel knowledge.
data detection. Otherwise, the initial choice of causes an
error floor, due to the relatively small frame size. An alternative
to this training would be to estimate the autocorrelation over REFERENCES
one frame, and then to use that estimate to go back through the
received vectors to make decisions. This would be appropriate [1] N. Yee, J. P. M. G. Linnartz, and G. Fettweis, Multi-carrier CDMA in
indoor wireless networks, in Proc. PIMRC, Sept. 1993, pp. 109113.
when the latency of the detector is not a concern. [2] A. Chouly, A. Brajal, and S. Jourdan, Orthogonal multicarrier tech-
A complexity comparison can be made between the proposed niques applied to direct sequence spread spectrum CDMA systems, in
system and a system with a linear MMSE MUD. The complexity Proc. GLOBECOM, vol. 3, Houston, TX, Nov. 1993, pp. 17231728.
of MMSE MUD is much higher than the proposed system be- [3] A. C. McCormick and E. A. Al-Susa, Multicarrier CDMA for future
generation mobile communication, Electron. Commun. Eng. J., vol. 14,
cause it requires knowledge of every users spreading sequence, pp. 5260, Apr. 2002.
and spreading sequences change from bit to bit. In addition, [4] S. Buzzi and H. V. Poor, Channel estimation and multiuser detection
the number of users occupying the channel must be known, in long-code DS/CDMA systems, IEEE J. Select. Areas Commun., vol.
19, no. 8, pp. 14761487, Aug. 2001.
and accurate estimates of the received powers of all users and [5] T. F. Wong, T. M. Lok, J. S. Lehnert, and M. D. Zoltowski, A linear
the AWGN in the channel must be available. Furthermore, the receiver for direct-sequence spread-spectrum multiple-access systems
linear MMSE MUD must continually estimate the channel for with antenna arrays and blind adaptation, IEEE Trans. Inf. Theory, vol.
44, no. 2, pp. 659676, Mar. 1998.
all users. This estimation must accurately determine the channel [6] H. Sari, G. Karam, and I. Jeanclaude, Transmission techniques for dig-
[17]. After correlating with all the possible spreading sequences ital TV broadcasting, IEEE Commun. Mag., vol. 33, pp. 100109, Feb.
and compensating for all subcarrier gains, the MUD must com- 1995.
pute a matrix inverse, which changes from bit to bit, [7] J. Namgoong, T. F. Wong, and J. S. Lehnert, Multiple access interfer-
ence suppression for OFDM systems based on Doppler shifts, in Proc.
to jointly minimize the effects of MAI and noise. This simula- IEEE Military Communications Conf., Los Angeles, CA, Oct. 2002, pp.
tion assumes that the MUD does not track the carrier offset of 567571.
each user individually, since that would significantly increase [8] S. Benedetto and E. Biglieri, Principles of Digital Transmission, with
Wireless Applications. New York: Kluwer, 1999.
the computational complexity in finding the MMSE solution. [9] S. H. Mller-Weinfurtner, Optimum Nyquist windowing in OFDM re-
In contrast, the proposed receiver is a decentralized ceivers, IEEE Trans. Commun., vol. 49, no. 3, pp. 417420, Mar. 2001.
single-user receiver. Hence, its complexity is independent [10] J. G. Proakis and D. G. Manolakis, Digital Signal Processing: Princi-
ples, Algorithms, and Applications, 3rd ed. Upper Saddle River, NJ:
of the number of users, and it does not require any more
Prentice Hall, 1995.
system-wide knowledge than a conventional MC-CDMA re- [11] B. Muquet, M. de Courville, P. Duhamel, and G. Giannakis, OFDM
ceiver. The proposed receiver requires despreading vectors. with trailing zeros versus OFDM with cyclic prefix: Links, comparisons
An matrix inverse of the autocorrelation matrix must be and application to the HIPERLAN/2 system, in Proc. Int. Commun.
Conf., June 2000, pp. 10491053.
computed. However, as the numerical evaluation section shows, [12] J. H. Cho and J. S. Lehnert, An optimal signal design for band-limited
the performance of the new scheme is essentially the same for asynchronous DS-CDMA communications, IEEE Trans. Inf. Theory,
all values of . Therefore, the main complexity difference vol. 48, no. 5, pp. 11721185, May 2002.
[13] R. V. Nee and R. Prasad, OFDM for Wireless Multimedia Communica-
between the proposed receiver and a basic MC-CDMA receiver tions. Boston, MA: Artech House, 2000.
is a larger FFT size (which can simply be implemented with [14] S. Haykin, Adaptive Filter Theory, 3rd ed. Upper Saddle River, NJ:
-point FFTs) and more despreading vectors. Prentice Hall, 1996.
[15] T. M. Lok and J. S. Lehnert, An asymptotic analysis of DS/SSMA com-
munication systems with random polyphase signature sequences, IEEE
VII. CONCLUSIONS Trans. Inf. Theory, vol. 42, no. 1, pp. 125136, Jan. 1996.
[16] S. Verdu, Multiuser Detection. New York: Cambridge Univ. Press,
The results show that there is a potential for significant MAI 1998.
suppression in MC-CDMA systems when the users experience [17] T. T. Chen and J. S. Lehnert, Multiuser decision-feedback receivers
frequency shifts. These shifts can arise from Doppler or from os- for asynchronous CDMA systems over a mismatched Rayleigh fading
channel, IEEE Trans. Commun., vol. 48, no. 9, pp. 15631570, Sept.
cillator inaccuracies in transmitters. If no excess time is used, all 2000.
users are in the same subspace and nothing is gained from extra [18] T. F. Wong, T. M. Lok, and J. S. Lehnert, Asynchronous multiple-ac-
processing of the frequency-domain data. However, if some ex- cess interference suppression and chip waveform selection with ape-
cess time is used (or available from uncorrupted portions of riodic random sequences, IEEE Trans. Commun., vol. 47, no. 1, pp.
103114, Jan. 1999.
the cyclic prefix), then any frequency shift can lead to an im- [19] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Pro-
provement in SIR that can be achieved by processing the extra cessing. Englewood Cliffs, NJ: Prentice Hall, 1989.
686 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 53, NO. 4, APRIL 2005

Brandon P. Hombs (S98) received the B.S. degree James S. Lehnert (S83-M84-SM95-F00) re-
in electrical engineering from Oklahoma Christian ceived the B.S. (highest honors), M.S., and Ph.D.
University, Oklahoma City, in 2000. Following that, degrees in electrical engineering from the University
he received the M.S.E.C.E. degree from Purdue of Illinois at Urbana-Champaign in 1978, 1981, and
University, West Lafayette, in 2002, where he is cur- 1984, respectively.
rently working toward the Ph.D. degree in electrical From 1978 to 1984, he was a Research Assistant
engineering. at the Coordinated Science Laboratory, University of
His research interests include digital communi- Illinois, Urbana, Illinois. He has held summer posi-
cations with an emphasis on spread-spectrum and tions at Motorola Communications, Schaumburg, IL,
multicarrier systems. During the summer of 2004, in the Data Systems Research Laboratory, and Harris
he worked on MIMO design and implementation as Corporation, Melbourne, FL, in the Advanced Tech-
a research intern at MIT Lincoln Laboratory, Lexington, MA. nology Department. He is currently a Professor in the School of Electrical and
In 2001, Mr. Hombs was a recipient of the National Science Foundation Computer Engineering at Purdue University, West Lafayette, IN. His current
Graduate Research Fellowship. He is currently a GAANN Fellow at Purdue research work is in communication and information theory with emphasis on
University. spread-spectrum communications.
Dr. Lehnert was a University of Illinois Fellow from 1978 to 1979, and an
IBM Pre-Doctoral Fellow from 1982 to 1984. He is currently a Purdue Uni-
versity Faculty Scholar. He has served as Editor for Spread Spectrum for the
IEEE TRANSACTIONS ON COMMUNICATIONS and as Guest Editor for the IEEE
JOURNAL ON SELECTED AREAS IN COMMUNICATIONS.