IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 50, NO.
6, JUNE 2002
On the Comparison Between OFDM and Single
Carrier Modulation With a DFE Using a
Frequency-Domain Feedforward Filter
Nevio Benvenuto, Senior Member, IEEE, and Stefano Tomasin, Student Member, IEEE
Abstract—Most comparisons between single carrier and multi-
carrier modulations assume frequency-domain linear equalization
of the channel. In this paper we propose a new frequency-domain
decision feedback equalizer (FD-DFE) for single carrier modula-
tion, which makes use of a data block transmission format sim-
ilar to that of the orthogonal frequency-division multiplexing with
cyclic prefix (OFDM). The scheme is a nonadaptive DFE where the
feedforward part is implemented in the frequency domain, while
feedback signal is generated by time-domain filtering. Through
simulations in a HIPERLAN-2 scenario, we show that FD-DFE
yields a capacity very close to that of OFDM. This result is also
confirmed by analytical derivations for a particular case. Further-
more, when no channel loading is considered, FD-DFE performs
closely to OFDM for the same averaged frame error rate in a coded
transmission. Design methods of the FD-DFE are investigated and
a reduced complexity technique is developed, with the result that
FD-DFE and OFDM have a similar computational complexity in
signal processing.
Index Terms—Decision feedback equalizers, signal processing,
wireless LAN.
I. INTRODUCTION
W IRELESS broadband communications systems are char-
acterized by very dispersive channels. To face this phe-
nomenon, two modulation techniques can be used: single car-
rier (SC) modulation with broadband equalization, or multi-
carrier modulation with orthogonal frequency-division multi-
plexing (OFDM). Indeed, there are already numerous applica-
tions of OFDM which range from digital audio and video broad-
casting DAB and DVB standards [1], [2] to wireless LAN stan-
dards [3], [4].
There is a long on-going dispute over the advantages of the
two modulations [5]–[8]. The OFDM system (also called, for
wired transmission, discrete multi-tone, DMT) makes use of a
cyclic prefix (CP) and a one tap per subchannel equalizer, de-
signed according to zero forcing (ZF) or minimum mean square
error (MMSE) criteria. Up to now, the comparison between SC
Paper approved by C. Tellambura, the Editor for Modulation and Signal De-
sign of the IEEE Communications Society. Manuscript received December 8,
2000; revised June 29, 2001 and October 29, 2001. This paper was presented
in part at the IEEE 8th Symposium on Communications and Vehicular Tech-
nology, Benelux, Delft, The Netherlands, October 2001.
N. Benvenuto is with the Dipartimento di Elettronica e Informatica, Univer-
sità di Padova, 35131 Padova, Italy (e-mail: nb@dei.unipd.it).
S. Tomasin is with PACman Group, Philips Research Laboratory (Natlab),
5656 AA Eindhoven, The Netherlands, on leave from the Dipartimento di
Elettronica e Informatica, Università di Padova, 35131 Padova, Italy (e-mail:
tomasin@ieee.org).
Publisher Item Identifier S 0090-6778(02)05544-7.
0090-6778/02$17.00 © 2002 IEEE
947
and OFDM systems has been mainly limited to SC with a fre-
quency domain linear equalizer FD-LE [5], [6]. Among others,
we recall the simplified LE proposed by Huemer et al. [9],
which makes use of a data block transmission structure with a
cyclic prefix. The better performing decision feedback equal-
izer (DFE) was usually not considered in the comparison, since
it has a much greater complexity both in design and in signal
processing [7]. Zervos and Kalet [10] have already proved that,
at high SNRs, an SC system with a time-domain (TD) uncon-
strained length DFE and OFDM have the same capacity. How-
ever, the authors do not provide any reduced complexity DFE.
Berberidis and Palicot [11] instead proposed a frequency do-
main DFE (FD-DFE) for SC modulation, which is suitable only
for channels with a very long impulse response.
In this paper, we present a new FD-DFE for SC systems which
makes use of a data block transmission structure similar to CP.
The scheme is a nonadaptive DFE where the feedforward part
is implemented in the frequency domain, while feedback signal
is generated by time-domain filtering. While gaining the bene-
fits of a DFE in terms of performance, the new scheme has also
the advantage of using the discrete Fourier transform (DFT) to
reduce complexity of signal processing. For the design of the
FD-DFE we propose an efficient algorithm, which directly op-
erates in the frequency domain and is optimal in terms of mean
square error.
FD-DFE and OFDM are compared in two situations. The
first comparison is in terms of achievable bit rate ( ) [12].
This corresponds to a scenario where, for a given transmis-
sion channel, channel loading is performed on both systems.
By computer simulations, we conclude that OFDM and SC
FD-DFE yield a similar . This result is also confirmed
in a particular case by analytical derivations. When adaptive
modulation is considered for OFDM, FD-LE has a worse per-
formance than OFDM [6], while FD-DFE has a performance
close to OFDM. In the second scenario no channel loading is
assumed and performance is compared in terms of frame error
rate. Czylwik has already noted that FD-LE significantly out-
performs OFDM with fixed modulation [6]. Here we observe
that FD-DFE yields an additional 3-dB gain over FD-LE.
Considering that an SC system has a better peak to average
transmit power ratio than OFDM, the new equalizer makes the
SC scheme a valid alternative to OFDM in broadband wireless
systems.
We should say that in time-varying channels this approach, as
well as OFDM, may not be suitable. In fact, adaptive equaliza-
tion is hard to combine with the block filter structure. Moreover,
948
with respect to traditional SC systems, latency is increased, due
to block operation, and bandwidth efficiency is lowered, due to
the inclusion of the PN extension.
The paper is organized as follows. In Section II the new data
block transmission structure is presented and compared with
the cyclic prefix structure of the OFDM. In Section III the new
FD-DFE system is described, while in Section IV equalizer de-
sign methods are derived. Section V describes the two scenarios
for the performance comparison made in Section VI. Finally,
Section VII contains the conclusions.
II. LINEAR AND CIRCULAR CONVOLUTION
Frequency-domain equalization is based on the equivalence
of the convolution of two sequences in the time domain and
the product of their Fourier transforms. However, in general,
the Fourier transform can not be performed on the entire re-
ceived signal and frequency domain equalization must be per-
formed on blocks, using the DFT. In this case, the equivalence
between the time-domain convolution and the frequency-do-
main multiplication holds only if the transmitted signal forces
the linear convolution with the channel-impulse response to be
circular [13]. We recall that, given the sequences
and , their convolution (limited to the first
samples)
(1)
is circular on blocks of size if
(2)
In this case the element by element product of the -size DFT
of and of the first samples of yields the -size
DFT of samples . signal
Note that, to simplify the notation, in Section II the received
signal (1) does not take into account the channel noise.
We will now examine two possible block transmission for-
mats which force the convolution of the sent data with
the transmission channel to be circular. The first method,
denoted cyclic extension, is commonly used in OFDM systems.
For the FD-DFE systems instead, we will introduce a new trans-
mission format, the PN extension.
A. Cyclic Extension
The data signal is cyclically extended [9], [14], before trans-
mission. If is the th
block of data symbols, the transmitted block of length
becomes
(3)
where the first symbols coincides with the last symbols of
block . At the receiver, each block of samples
is split into two parts, where the first samples are discarded,
while DFT is performed on the remaining -size block . the optimization of the peak to average power ratio.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 50, NO. 6, JUNE 2002
From (3) we see that (2) is verified for on blocks of size
.
Note that the interference on the first samples of
is generated by the symbols .
Hence, if we apply a scalar DFE [15, pp. 621–22] to remove
this interference, the feedback filter should already know the
last symbols of the current data block, which is impossible.
Elaborate DFE, operating on a per-block basis have been
proposed [16], [17] and they can be used also for a cyclic
extended transmission. However, even if they can give a better
performance than OFDM, their signal processing and design
complexity is much greater (see Section VI-C). In this paper
we limit ourselves to much simpler solutions.
B. PN Extension
An alternative way to force the linear convolution (1) to be-
come circular is to extend each data block with a fixed sequence
of symbols [18], for example, a pseudonoise (PN) [15, p. 724]
sequence . The new data block of sym-
bols is
(4)
where now the last symbols are the PN sequence. Moreover, a
PN extension is transmitted before the first data block. As indi-
cated in Fig. 1, while the cyclic extension copies part of the in-
formation data at the beginning of each block, the PN extension
philosophy is to interleave information blocks with a known PN
sequence. Since in a continuous transmission
(5)
we have (2) verified for blocks of size of the
. Hence, if we indicate the -size DFT of
with
(6)
the DFT of with and the DFT of with
, is the multiplication of and
, element by element.
Now, the PN extension is suitable for use with a DFE, since in
a continuous transmission, the interference on the first sym-
bols of each block is generated by the PN extension sym-
bols, which are fixed and known at the receiver.
Note that the cyclic extension and the PN extension have ap-
proximately the same bandwidth efficiency , for a
transmission with numerous blocks. Indeed, the PN extension
format includes one more block of symbols than the cyclic
extension format, at the beginning of the transmission.
For both extensions, a higher yields a greater bandwidth
efficiency, but also a higher latency. In general, for time-varying
channels, must be chosen so that the channel could be con-
sidered static on each frame. Furthermore, the PN extension can
be reduced to zero-padded block transmission [17]. The choice
of the sequence may be done according to other criteria, such as
BENVENUTO AND TOMASIN: ON THE COMPARISON BETWEEN OFDM AND SINGLE CARRIER MODULATION WITH A DFE 949

(a)

(b)
Fig. 1. Data block structure for (a) cyclic extension and (b) PN extension.

III. FREQUENCY DOMAIN DFE

From (1) the received signal (baseband equivalent) is

(7)

where is generated as in (4), while is a zero-mean com-

plex Gaussian white noise with variance .
The model in (7) refers to a receiver where the timing phase
corresponds to the first significant sample of the channel-im-
pulse response. It does not imply that the channel is causal.
The FD-DFE structure is shown in Fig. 2. The feedforward
(FF) filter of the FD-DFE operates in the frequency domain on
blocks of size , while the feedback (FB) filter operates in the
time domain. In fact, the frequency-domain filtering must be
performed on a per-block basis, while the feedback section must
be fed with the previous detected symbols whose decision is per-
formed in the time domain. Note that also for block equalization
[17], the feedback filter must be a strictly upper triangular ma-
Fig. 2. FD-DFE equalizer for single carrier systems.
trix.
As shown in Fig. 2, after the DFT of the received sam-
ples, the FD feedforward filter, with coefficients , where, from (7) and (8)
, is applied to yield the block signal
with elements

(8) (12)

Through the inverse DFT, block is then transformed in the In (11) we note that if then the condition (2) holds
time domain to give for the useful part of , which can be written as a linear
convolution, and
(9)

Let us define the impulse response of the cascade of the

equivalent discrete-time channel impulse response and the (13)
feedforward equalizer as the IDFT 1 of the product of
and , i.e. Note that in (13) the last samples of each block have not
been considered, since they are a noisy and interfered version of
the PN sequence, which is already known. On the other hand, if
(10) then (13) does not hold and depends also on
, . In this case the
assuming it has a support of coefficients. Then, each feedback filter will not be able to cancel all the interference.
block can be written as the circular convolution of As mentioned previously, the feedforward section operates
and plus a noise term, namely in the frequency domain, while the feedback section operates in
the time domain. Let’s indicate the estimated data sequence with
and the extended estimated sequence with . From (4)
it holds
(11)
1To simplify the notation, the normalization factor 1=P has been included in (14)
the sequence fG g.
950 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 50, NO. 6, JUNE 2002

Then, if , , are the coefficients of Indicating with and , the real and imaginary parts
the FB filter, the signal at the input of the decision element is of , respectively, since is a strictly convex function of
and , its minimum is attained when the gradient is
set at zero, namely

(15)
Note that, as indicated in Fig. 2, for each block the first
data symbols, which initialize the feedback part of the DFE, (20)
should coincide with the PN symbols . Let [where denotes
Since the FD-DFE operates on a per-block basis, error prop- the transpose] and let’s define the following matrix and
agation through the FB filter is limited to one block. This is an the column vector
important advantage over the time-domain DFE (TD-DFE), [19,
p. 365], when the system operates at low values. More-
over, it has a reduced complexity when compared with other fre-
quency-domain equalizers [11], [13], that need -size DFTs. (21)
On the other hand, as mentioned in the Introduction, it also has
all drawbacks of block-based systems in time-varying channels. (22)
Note also that the FD-DFE scheme is an efficient realization
of a block DFE since the DFT, the feedforward equalization and
the IDFT can be combined into one single matrix multiplication. then from (20) we obtain the linear system of equations with
unknowns .
IV. DFE DESIGN Since is a Toepliz matrix, a reduced complexity algo-
rithm can be used to solve the problem. Additionally, observe
We start from the FD channel estimate which can be obtained that the elements of both and can be computed as the
directly as in [20] or by taking the -size DFT of the channel -size DFT of , and FFT can be used. As a matter of
impulse response estimate. fact, -size DFTs can be considered.
Zero Forcing FD-DFE: According to the ZF criterion, all Minimum MSE FD-DFE: According to the MMSE criterion,
interferers must be canceled by the feedback part. If the support the coefficients of the FF and FB filters are chosen to minimize
of is , (13) holds true and interference can be the sum of the power of the filtered noise, and the power of the
canceled by the feedback filter. Hence, let’s set . The residual interference. In particular, the mean square error at the
zero-forcing condition can be expressed as detector is given by
for (16)
and only coefficients , , can be chosen freely. (23)
Once the coefficients are known, by selecting
and By assuming that the past decisions are correct and that
(17) and , from (13) and (15) we obtain

the feedback filter cancels all residual interferers. In turn, from

(10) coefficients of the feedforward filter can be com-
puted as

(24)
(18)
Now, we rewrite (24) in the frequency domain. Firstly, we in-
where and assuming and using troduce the -size DFT of the FB filter
(17).
Here, the coefficients , , are chosen
to minimize the power of the filtered noise, which, from (10)
and under the condition (16), can be written as (25)
Moreover, from (10) the gain of the useful data at the decision
point can be written as

(19) (26)
BENVENUTO AND TOMASIN: ON THE COMPARISON BETWEEN OFDM AND SINGLE CARRIER MODULATION WITH A DFE 951

Hence, from (24), (25) and (26), according to the minimum mitter. Note also that in a wireless mobile scenario the channel
mean square error criterion, the functional to be minimized is loading must be updated whenever the transmission channel
changes significantly. Therefore we considered a second sce-
nario where channel loading is not performed and the constel-
lations are fixed. Now the performance is driven by modulation
and coding structure. The symbol interleaver plays also a sig-
(27) nificant rule for OFDM [5], since it reduces burst errors due to
where is the power of the signal . adjacent subchannels experiencing similar fading. We compare
Due to the -extension structure, the FB filter is not able the performance of coded OFDM and SC systems in terms of
to cancel more than interferers, hence, we must impose that frame error rate ( ).
. Here we consider the case . In order to
A. Channel Loading and Coding Scenario
compute the design of the FF and FB filters, we write the func-
tional only as a function of . In particular, we In the first scenario, performance of OFDM and SC systems
observe that, given the feedback filter, by applying the gradient are compared under the hypothesis of perfect channel loading,
method to (27), the feedforward filter is given by with a continuous varying constellation size and an optimum
coding. Channel loading is widely used in wired transmissions
(28) [12]. More recently it has also been considered for wireless ap-
plications since it can ease the problem of slowly time-varying
channels by exploiting the variation of the signal quality [21].
where indicates the complex conjugate. Inserting now (28)
Hence, we evaluate here the performance of the proposed sys-
in (27) we obtain
tems using also as a performance measure the achievable bit rate
( ), [12]. For the OFDM system, let us define the signal
to distortion plus noise ratio at the decision device of the th
subchannel

(32)
(29)
using (25). where is the -size DFT of . The is defined as
We define and the fol-
lowing matrix and the column vector as (33)

where is the symbol rate.

For SC modulation, by assuming and ,
(30) we define the signal to distortion plus noise ratio at the decision
device as
(31)
(34)
As in the ZF design, by applying the gradient method to mini-
mize , we obtain the linear system of equations with where is given by (19) or (24) and the value of varies ac-
unknowns . We note that the complexity cording to the equalizer: for ZF FD-DFE [see (16)],
of this method is similar to the ZF. Once the FB filter is deter- while for MMSE FD-DFE is given by (26). Note that we
mined, the FF filter is given by (28) and (25). are neglecting here the error propagation phenomena due to the
Note that the MMSE solution will reduce to the ZF solution DFE since we do not simulate the channel loading and, there-
when . fore, we do not know the constellation and the code of the trans-
mitted signal.
V. CHANNEL LOADING AND CODING Even if OFDM and FD-DFE have the same redundancy, for
the SC modulation channel loading is performed by choosing
To compare the performance of SC and OFDM systems, we only one constellation instead of as for the OFDM . Hence,
consider two scenarios. for SC modulation the is given by
In the first scenario, the transmitter knows the channel im-
pulse response and maximizes the capacity of the system by se- (35)
lecting the constellation for each subchannel of the OFDM and
for the unique channel of the SC modulation. This operation is where .
named channel loading [21]. In the Appendix we show analytically that assuming the
In some situations channel loading is not possible or unsuit- number of the feedback coefficients , then the
able. For example, it can not be used in a broadcast transmission of a ZF FD-DFE is very close to that of an OFDM. By
and it is not suitable (especially for OFDM) when it requires a simulations we verify that the ABR of MMSE FD-DFE is very
considerable amount of information to be sent back to the trans- close to that of an OFDM even when .
952 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 50, NO. 6, JUNE 2002

B. Coding Scenario
In the coding scenario, we consider a convolutional encoder
with rate , free distance and BPSK constellation. Let us
indicate with the number of codewords with distance
, and with the Bhattacharyyaa parameter
[15] for an AWGN channel with BPSK modulation.
For an SC system, to partially consider the problem of error
propagation in the DFE, under the assumptions and
, we introduce

(36)

where is the detected symbol in correspondence of the equal-

izer output. Indeed, is similar to and of (19)
and (24), with the addition that now the error propagation phe-
nomena of the DFE are considered. The signal to distortion plus
noise ratio at the decision device is defined as
Fig. 3. Complementary CDF of the achievable bit rate for different equalizer
(37) NR =
structures. S 12 dB.

where is estimated by simulations. VI. PERFORMANCE COMPARISON

A good approximation of the with frames of length
is given by [22]
(38)
where is a correction factor to fit simulated data, is given by
(37) and for the HIPERLAN-2 code.
We observe that even with the new definition (36), in the an-
alytical results of (38), we are neglecting the correlation among
errors at the input of the decoder, due to the error-propagation
phenomenon of the DFE. In any case, as analyzed in [19, pp.
335–370], the output error of the DFE is approximately white,
and in our setup an interleaver is inserted before decoding (see
Section VI) to further decorrelate errors. For these reasons, at
detector input the resulting distortion can be approximated as
white noise.
By simulations we first derived values of FER for lower
SNRs. The value of the parameter in (38) was then derived
by fitting the curve. Lastly, the curves were extended to higher
SNRs using (38).
For an OFDM system we use the analytical method proposed
by Nanda and Rege [22] in a different system. Let’s indicate
with the length of the shortest error event of the code. Let us
define the partial terms
(39)
where . A good estimate of the FER
can be obtained by
(40)
where the function is defined in (38).
In our analysis we assumed a frame length of . OFDM have a very close performance. Additionally, observe
We have considered a HIPERLAN-2 scenario with a symbol
rate of MHz. The systems’ performance was ob-
tained by simulation on a typical environment considered for
HIPERLAN-2, i.e., a no line of sight multipath channel. The
channel has a Rayleigh characteristic with an exponential de-
caying power profile with an rms delay spread of 100 ns, ac-
cording to channel model in [23]. The data block size is
and , (option II of model [23]). The inter-
leaver of the HIPERLAN-2 standard [3] is included. Frequency
and time synchronization are assumed perfect. The channel im-
pulse response is assumed known at the receiver and time in-
variant over the transmission of a block. Schemes considered
are OFDM, FD-LE (as described in [9]) and FD-DFE designed
both with a ZF and with an MMSE method. for all
DFE schemes.
A. Channel Loading and Coding Scenario
With reference to the scenario with channel loading, we as-
sume that at the input of the feedback past decisions are correct,
since we do not simulate the channel loading and therefore do
not know the constellation.
The complementary cumulative distribution function (cCDF)
of the is used to compare the different systems in Fig. 3.
The ratio between the average power of the signal (with respect
to the channel variations) and the power of the noise ( ),
is assumed to be 12 dB. It has already been shown that when
channel loading is performed, the OFDM exhibits a sharp in-
crease of performance over FD-LE [6]. This is confirmed by
Fig. 3. However, we observe that OFDM and MMSE FD-DFE
yield a very close performance (see also the Appendix). The
MMSE FD-DFE instead has better performance than the ZF
FD-LE, for the usual ZF’s problem of neglecting the presence
of noise in the design.
For the various systems, the mean versus is
shown in Fig. 4. The same observations made for Fig. 3 can be
repeated here. In particular we note that MMSE FD-DFE and
BENVENUTO AND TOMASIN: ON THE COMPARISON BETWEEN OFDM AND SINGLE CARRIER MODULATION WITH A DFE 953

Fig. 4. Mean ABR for different equalizer structures as a function of the Fig. 6. Mean frame error rate, in the presence of coding, for different equalizer
SNR. structures as a function of the SNR.

Fig. 5. Complementary CDF of the frame error rate, in the presence of coding,
for different equalizer structures. SNR=12 dB.
Fig. 7. Mean frame error rate, in the absence of coding, for different equalizer
structures as a function of the SNR.
that the feedback part of the DFE becomes more relevant as the
increases. Finally, the performance of uncoded systems is shown in
Fig. 7, in terms of averaged for various values of .
B. Coding Scenario In this case, the performance of FD-LE is much better than
With reference to the coding scenario, the BPSK modulation OFDM, while the MMSE FD-DFE gains about 3 dB over
is considered and the standard HIPERLAN-2 con- FD-LE.
volutional code [3] is used for all systems. The analytical eval-
uation of for both OFDM and FD-DFE fitted the simula- C. Computational Complexity
tion results for . For the DFEs in the presence of coding, Computational complexity of the system, in terms of the
the hard-detected coded data are used as input for the feedback number of complex multiplications per output sample, is re-
filter. Hence, error propagation phenomena are considered. ported in Table I for OFDM and SC FD-DFE. We also included
Fig. 5 shows the cCDF of the for the different coded the complexity of the block DFE [17], denoted as B-DFE. We
systems. We observe that the SC system with an MMSE have considered that a -size DFT requires
FD-DFE performs similarly to OFDM. Again, MMSE FD-DFE complex multiplications. For the OFDM system we have also
outperforms the MMSE FD-LE. included the IDFT of the transmitter. The channel estimation
The mean for various values of is shown in was not considered because both OFDM and FD-DFE need the
Fig. 6. The MMSE FD-DFE has a better performance than ZF same estimate of the channel frequency domain response which
FD-DFE at low s. For dB we see that OFDM can be obtained by known techniques (see [20] and references
outperforms MMSE FD-DFE by about 0.5 dB. therein). We see that, when compared to OFDM, FD-DFE
954 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 50, NO. 6, JUNE 2002

TABLE I SC benefits of a better peak-to-average power ratio than OFDM

COMPUTATIONAL COMPLEXITY OF THE SYSTEM, IN TERMS OF NUMBERS OF and requires fewer adaptations when operating in a variable rate
COMPLEX MULTIPLICATIONS PER OUTPUT SAMPLE, OF OFDM AND SC
SYSTEMS WITH BLOCK AND FD EQUALIZATION mode, SC FD-DFE is a suitable system for both (slow) time-
varying channels and broadcast channels.

APPENDIX
As mentioned in the Section I, Zervos and Kalet [10] have
already proven that for an unconstrained length ZF DFE and
high s, OFDM and SC FD-DFE have the same capacity.
Here we show that if , the s of ZF FD-DFE
and of OFDM are very close. The same conclusion is valid for
MMSE FD-DFE, because of its intrinsic superior performance
over ZF FD-DFE.
Now we show that by proving
TABLE II
COMPUTATIONAL COMPLEXITY, IN TERMS OF NUMBERS OF COMPLEX that and at the same time
MULTIPLICATIONS, DUE TO THE EQUALIZER DESIGN .
As observed by Bingham [24], comparing (33) to the water
pouring formulas, we conclude that if the number of subchan-
nels is sufficiently high, the is the highest pos-
sible for any system with the same bandwidth. Therefore,
we conclude that
(41)
Now, we show that under the condition then
(42)
From (33) and (35) we obtain

(43)

(44)
has an increased complexity due to the additional FB filter.
Moreover, as already mentioned, the B-DFE has a significantly We now derive the expression of in the ideal case
higher complexity than both OFDM and FD-DFE. . Indeed, this hypothesis gives an upper bound on performance
Table II, instead, compares the computational complexity of since we are assuming that the feedback can completely elim-
the equalizer design. Given the channel frequency response, the inate the interference even if and hence .
equalizer design for OFDM is carried out by one complex divi- Actually, when , (13) should be rewritten as
sion per subchannel. For the FD-DFE, firstly, matrix (21)
and vector (22) or matrix (30) and vector (45)
(31) can be computed through DFTs. Then a -size linear
system of equations must be solved, whose complexity is on
the order of multiplications. and since , (15) becomes
Indeed, the filter design for OFDM is much less complex than
that for FD systems. However, when operating in a variable rate (46)
mode with channel loading, other factors as the number of pa-
rameters that must be fed back to the transmitter should be con- because in the first symbols at most interferers
sidered. In fact, while for the FD-DFE only one constellation can be subtracted. In this case the mean square error
size must be fed back to the transmitter, for OFDM constel-
lation sizes, one for each subchannel, are needed. In this case, (47)
complexity of channel loading must be inserted.
depends on the symbol index as . In partic-
VII. CONCLUSIONS ular, from (45) we can infer that . 2 When
A DFE using a frequency domain feedforward filter for single all interferers can be canceled and a ZF solution exists.
carrier systems has been proposed which yields an very Now we will compute the value of that will give an upper
close to OFDM. Moreover, also when the constellation is fixed, bound on the performance of the SC FD-DFE.
FD-DFE performs closely to OFDM. In terms of complexity, 2The authors are grateful to the anonymous reviewer who provided the deriva-
FD-DFE requires a more complex filter design. Considering that tions of the above equations when N >L .
BENVENUTO AND TOMASIN: ON THE COMPARISON BETWEEN OFDM AND SINGLE CARRIER MODULATION WITH A DFE
From (19) we minimize
only under the constraint , i.e., . 1, Singapore, Nov. 1995, pp. 98–102.
The functional to be minimized becomes
(48)
where is the Lagrange multiplier. Applying the gradient
method to (48) we obtain
(49)
Therefore (44) yields
(50)
Since , we ob-
tain
(51)
Now, the arithmetic mean is lower bounded by the geometric
mean [25] and it holds true
(52)
Therefore, we conclude that (42) holds true.
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers for
their insightful comments and suggestions that contributed to
improve the quality of this paper.
REFERENCES
[1] European Telecommunications Standards Institute, “Digital audio
broadcasting (DAB); DAB to mobile portable and fixed receivers,”,
France, ETS 300 401, Feb. 1995.
[2] European Telecommunications Standards Institute, “Digital video broad-
casting (DVB); Framing structure, channel coding and modulation for
11/12 GHz satellite services,”, France, ETS 300 421, Dec. 1994.
[3] European Telecommunications Standards Institute, “Broadband radio
access networks (BRAN); HIPERLAN Type 2; Physical (PHY) layer,”,
France, TR 101 475 V.1.1.1, Apr. 2000.
[4] D. Andelman and D. Stopper, “Variable constellation multitone mod-
ulation (VCMT)—Proposal for high capacity physical layer,”, IEEE
802.14a HI_PHY Study Group, June 1998.
[5] H. Sari, G. Karam, and I. Jeanclaude, “Transmission techniques for dig-
ital terrestrial TV broadcasting,” IEEE Commun. Mag., vol. 33, no. 2,
pp. 100–109, Feb. 1995.
[6] A. Czylwik, “Comparison between adaptive OFDM and single carrier
modulation with frequency domain equalization,” in Proc. VTC’97, vol.
2, Phoenix, AZ, May 1997, pp. 865–869.
[7] T. J. Willink and P. H. Wittke, “Optimization and performance evaluation
of multicarrier transmission,” IEEE Trans. Inform. Theory, vol. 43, pp.
426–440, Mar. 1997.
[8] A. Gusmão, R. Dinis, R. Conȩicão, and N. Esteves, “Comparison of two
modulation choices for broadband wireless communications,” in Proc.
VTC’00—Spring, vol. 2, Tokyo, Japan, May 2000, pp. 1300–1305.
[9] M. Huemer, L. Reindl, A. Springer, and R. Weigel, “Frequency domain
equalization of linear polyphase channels,” in Proc. VTC’00—Spring,
vol. 3, Tokyo, Japan, May 2000, pp. 1698–1702.
955
[10] N. A. Zervos and I. Kalet, “Optimized decision feedback equalization
versus optimized orthogonal frequency division multiplexing for high-
speed data transmission over the local cable network,” in Proc. ICC’89,
Boston, MA, June 1989, pp. 1080–1085.
[11] K. Berberidis and J. Palicot, “A frequency domain decision feedback
equalizer for multipath echo cancellation,” in Proc. Globecom’95, vol.
[12] J. Cioffi, “A multicarrier primer,” T1E1.4/91-157, Nov. 1991.
[13] A. V. Oppenheim and R. W. Schafer, Digital Signal Pro-
cessing. Englewood Cliffs, NJ: Prentice-Hall, 1975.
[14] S. B. Weinstein and P. M. Ebert, “Data transmission by frequency-di-
vision multiplexing using the discrete Fourier transform,” IEEE Trans.
Commun. Technol., vol. COM-19, pp. 628–634, Oct. 1971.
[15] J. G. Proakis, Digital Communications, 3rd ed. New York: McGraw-
Hill, 1995.
[16] A. Klein, G. K. Kaleh, and P. W. Baier, “Zero forcing and minimum
mean-square-error equalization for multiuser detection in code-division
multiple access channels,” IEEE Trans. Veh. Technol., vol. 45, no. 2, pp.
276–287, May 1996.
[17] A. Stamoulis, G. B. Giannakis, and A. Scaglione, “Block FIR deci-
sion-feedback equalizer for filterbank precoded transmissions with blind
estimation capabilities,” IEEE Trans. Commun., vol. 49, pp. 69–83, Jan.
2001.
[18] T. Walzman and M. Schwartz, “Automatic equalization using the dis-
crete frequency domain,” IEEE Trans. Inform. Theory, vol. IT-19, pp.
59–68, Jan. 1973.
[19] E. A. Lee and D. G. Messerschmitt, Digital Communication. Boston,
MA: Kluwer Academic, 1988.
[20] O. Edfors, M. Sandell, J. van de Beek, S. K. Wilson, and P. O. Bör-
jesson, “OFDM channel estimation by singular value decomposition,”
IEEE Trans. Commun., vol. 46, pp. 931–939, July 1998.
[21] T. Keller and L. Hanzo, “Adaptive multicarrier modulation: A conve-
nient framework for time frequency processing in wireless communica-
tions,” Proc. IEEE, vol. 88, pp. 611–640, May 2000.
[22] S. Nanda and K. M. Rege, “Frame error rates for convolutional codes on
fading channels and the concept of effective E =N ,” IEEE Trans. Veh.
Technol., vol. 47, pp. 1245–1250, Nov. 1998.
[23] D. Lacroix and D. Castelain, “A study of OFDM parameters for high
data rate radio LAN’s,” in Proc. VTC’00—Spring, vol. 2, Tokyo, Japan,
May 2000, pp. 1075–1079.
[24] J. A. C. Bingham, “Multicarrier modulation for data transmission: An
idea whose time has come,” IEEE Commun. Mag., vol. 28, no. 5, pp.
5–14, May 1990.
[25] M. Abramowitz and I. A. Stegun, Eds., Handbook of Mathematical
Functions. New York: Dover, 1972.
Nevio Benvenuto (S’81–M’82–SM’88) received
the Laurea degree in electrical engineering from the
University of Padova, Padova, Italy, in 1976, and
the Ph.D. degree in electrical engineering from the
University of Massachusetts, Amherst, in 1983.
From 1983 to 1985, he was with AT&T Bell Lab-
oratories, Holmdel, NJ, working on signal analysis
problems. He spent the next three years alternating
between the University of Padova, where he worked
on communication systems research, and Bell Labo-
ratories, as a Visiting Professor. From 1987 to 1990,
he was on the faculty at the University of Ancona. He was on the faculty at the
University of L’Aquila from 1994 to 1995. Currently, he is a Professor with the
Electrical Engineering Department, University of Padova. His research interests
are in the areas of voice and data communications, digital radio, and signal pro-
cessing.
Stefano Tomasin (S’99) was born in Venice, Italy,
in 1975. In 1999 he received the Laurea degree in
telecommunications engineering from the University
of Padova, Italy, where he is now pursuing the Ph.D.
degree.
From October 1999 to January 2000 he was with
the IBM Research Laboratory, Zurich, Switzerland,
working on signal processing for magnetic recording
systems. In the academic year 2001–2002 he did
a traineeship at Philips Research, Eindhoven, the
Netherlands, working on signal processing for
multicarrier system in mobile applications.