Coded Orthogonal Frequency Division Multiplex

BERNARD LE FLOCH, MICHEL ALARD, MEMBER, IEEE, AND CLAUDE BERROU, MEMBER, IEEE

Technological evolution and the ever-increasing demand for the beneficial consequences (through the use of a single
higher-quality services give broadcasters a strong incentive to frequency network (SFN)) to the economy and engineering
completely digitize their broadcasting networks. This digitization, of terrestrial TV broadcasting networks.
which is already well advanced in many program production areas
and transmission links, now has to be extended to complete the Section I1 gives an overview of the principles of coded
last link in the broadcast chain; i.e.,from broadcast transmitter to multicarrier transmission systems. Section 111, which covers
consumer receivers. a broader search for an optimum OFDM signal shaping,
It is therefore necessary to develop wholly new techniques for gives a prospective analysis of the possible options for
the broadcasting of digitally coded TV programmes. Thus an defining the orthogonal basis of elementary symbols form-
efficient baseband digital coding must be combined with a robust
digital modulation and channel coding scheme that can meet the ing the transmitted signal: The original signal configuration
requirements of every mode of broadcast reception. using a rectangular time-window shaping with guard inter-
This article presents the research work related to the coded val, as well as a new time-frequency shaping concept called
orthogonal frequency division multiplex (COFDM) technology, isotropic orthogonal transform algorithm (IOTA). Section
which has now been completed in the field of digital radio (DAB) IV deals with the choice of the channel coding technique.
[I], and which is under progress in the field of digital terrestrial
TV. Section V provides performance results as a function of
the tradeoff between power and spectrum efficiencies under
different transmission channel conditions, and Section VI
I. INTRODUCTION
gives a comparison of performances between COFDM and
Terrestrial broadcasting is faced with the problems of linear equalization of single carrier transmission. Section
echoes due to the phenomena of multiple propagation, and VI1 stresses the advantages of COFDM in terms of its suit-
saturation of the spectral resource. The intrinsic quality of ability with respect to broadcast engineering and spectrum
the signal, and the frequency planning, are two aspects management.
which are often considered as being independent. How-
ever the interference caused by long distance transmitters
11. GENERALPRINCIPLES OF COFDM
broadcasting the same program on the same frequency as
the local transmitter can be considered as artificial echoes, The COFDM technique is particularly suited to provide
and consequently the two aspects underlined above are in reliable reception of signals affected by strong distortions,
fact related to the same issue. The potentiality of a new as it can be the case for terrestrial broadcasting [2]. Mul-
broadcasting system will therefore depend on the strategy tipath propagation is known to limit the performance of
adopted regarding the echoes, whether they are natural or high bit-rate modulation schemes. The principle of COFDM
artificial. If the way of dealing with echoes is innovative and relies in splitting the information to be transmitted over a
systematically takes advantage of multipath propagation large number of carriers, in such a way that the signaling
instead of enduring it, the new perspectives opened up by rate on each of them becomes significantly lower than the
the digital system are really revolutionary. assumed channel coherence bandwidth. In other words, the
The following article goes over the principles of a mul- signal is conditioned to ensure that modulated symbols will
ticarrier modulation scheme developed with this goal in be much longer than the echo delay spread. Provided that a
mind. This scheme was initially designed and adopted for guard interval is inserted between successive symbols, mul-
the European Digital Audio Broadcasting project, but is tipath propagation does not generate anymore intersymbol
equally suitable for television broadcasting. This article interferences.
also highlights its advantages and gives prominence to However, in presence of strong echoes, some carriers
will suffer deep fades, due to destructive combination
Manuscript received July 1, 1994, revised March 2, 1995.
B. Le Floch is with the Centre Commun d’Etudes de Ttltdiffusion et of the various reflections, while others will be enhanced
Tiltcommunications (CCETT), 355 12 Cesson SCvignC Cedex, France. by constructive addition. The signal-to-noise ratio at the
M. Alard is with WAVECOM, 92130 Issy-les-Moulineaux, France. receiver input increases as soon as the signal power is
C. Berrou is with the Ecole Nationale Sup6rieure des Tiltcommunica-
tions de Bretagne (ENST Br), 29285 Brest Cedex, France. augmented by echoes separated by at least a delay equal
IEEE Log Number 9401 1374. to the inverse of the signal bandwidth. To benefit from this

0018-9219/95$04.00 0 1995 IEEE

982 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

.- - ~ ~ ~~

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 function of the signal basis

xm,n(t) =
.
e2i?rmt/r0 Rect(t - nTo). (2)
The signaling duration TOis chosen to be longer than the
time interval TO during which the carriers are orthogonal
(the carrier distance is equal to the inverse of TO). This
duration TOis the sum of the guard interval A and the
“useful symbol time” TO.
The Rect function is defined by
channel frequency
response Rect(t) =
if - To12 5 t < T0/2.
elsewhere (3)
time

Fig. 1. Principles of COFDM. Under the realistic assumption that the cumulated con-
tributions of the secondary lobes of the carriers spectra,

----
Transmitter arising from the time-limited symbol shaping, is negligible
outside the interval [ ~ , M / T Owhich
] . is the case because
a limited number of carriers on the spectrum edges are
not transmitted (in order to minimize adjacent channel
interference), it is possible to construct the signal from its
samples s ( ~ T o / M where
),
M-1
s(kTo/M) = ameZiTmklM. (4)
Fig. 2. Block diagram of a COFDM modem. m=O

In this equation, which holds for any COFDM symbol,

power increase, despite the fact that parts of the spectrum the time index n has been omitted without any loss of
are deeply faded, it is necessary to incorporate a powerful generality.
channel coding scheme in the system design [3]. Therefore, s(ICTO/M)is obtained by the discrete inverse
The role of coding, in conjunction with frequency and Fourier transform of .a{m
;::}
time interleaving, is to provide a link between bits trans- Under the assumption that the echoes remain shorter
mitted on separated carriers of the signal spectrum (see than the guard interval, the received signal observed over
Fig. l), in such a way that information conveyed by faded the “useful symbol time” TO is affected by intrasymbol
carriers can be reconstructed in the receiver, thanks to the interference only. This means that the modulation symbols
coding link which relates it to information conveyed by a, are transformed by the channel into Hmam,where Hm
well-received carriers. Coding and interleaving applied to is the channel frequency response at frequency index m.
OFDM can be seen as a tool to average local fadings over These terms H,am are recovered in the receiver using a
the whole signal bandwidth and over the time interleaving discrete direct Fourier transform; if s ’ ( t ) is the received
depth. Frequency selectivity, currently known to be a signal
disadvantage, is then turned into an advantage that can be M-1
called frequency diversity. This is the key feature which
H,a, = s’(k70/M)e-2i?rmklM. (5)
allows for SFN operation. k=O
The block diagram of the functions included in a COFDM
modem are presented in Fig. 2. It is important to notice that In order to achieve a coherent demodulation, &he channel
the tool used to build the signal at the transmitter and to frequency response H , must be recovered for each value of
analyze it at the receiver is the discrete Fourier transform m. This is generally achieved by inserting in the transmitted
[4], which allows for low cost implementation due to the signal a grid of time-frequency spread reference pilots [5],
existence of Fast algorithms. which comply with the sampling theorem with regard to
In the classical implementation of COFDM, the elemen- the frequency and time dispersion of the channel. Inter-
tary symbols constituting the signal are built of time-limited polation applied on these references allows estimation of
slices of sine and cosine waveforms. The transmitted signal the channel. The term H , includes a phase rotation and
can therefore be written in the following analytic form a modulus (p,), representing the attenuation factor of the
channel at a given frequency: Some carriers are enhanced
+CC M-1
by constructive interference, while others are faded by a
destructive combination of the echoes. After derotating the
n=-m m=O
samples at the Fourier transform output, the terms pmam
where am,n represents one of the elements of the modula- are processed by the channel decoder (desinterleaving and
tion alphabet conveyed by the carrier of index m during the maximum likelihood decoding), in which the attenuation
symbol time of index n,and ~ ~ , represents
~ ( t ) the generic factor pm is also taken into account as a measure of the

LE KOCH et al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 983

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 reliability of the information conveyed by the carrier of It is a well known result that any signal of duration T and
index m. bandwidth W can be represented, at least asymptotically,
by 2WT dimensions. Therefore, it is an essential property
111. DISCUSSION ON A BROADERSEARCH FOR AN of a modulation system to make the best use of all these
OPTIMUM OFDM ORTHOGONAL BASIS available space dimensions.
Hilbertian basis provides a powerful tool to design mod-
This section constitutes a prospective analysis of the
ulation systems which are optimum from this point of view.
optimality of the choice of the signal basis in order to
The transmitted signal can be written as
improve the system performance for the most difficult
channel conditions. As only the construction of this basis
is discussed hereunder and hence the discussion on the
coding is not included, the letter C of COFDM is generally where {zk} is a set of functions being a Hilbertian basis of
omitted. L2(R),and a k being real or complex values representing
the transmitted data. As { ~ k } is a basis of L2(R), no
A. Basic Principles space dimension is lost. Furthermore, if {xk} is a Hilbertian
basis, the distance properties of the coding scheme are not
1 ) OFDM and Hilbertian Basis: We consider first the
changed by the modulation process. This is an essential
case of an OFDM signal s ( t ) without any guard interval
property in a channel corrupted by gaussian noise, where
the performances are directly related to the Euclidian dis-
m,n
tance between transmitted signals. Last, but not least, a
Hilbertian basis is also desirable because it greatly sim-
where the coefficients am,n take complex values represent- plifies the receiver design, for we can simply recover the
ing the transmitted encoded data, and the basic functions transmitted data by taking benefit of the orthogonality of
zm,+(t)are obtained by translation in time and frequency the basic functions
of a prototype function z ( t )

Zm,n(t) = e2inmvot z(t - 7170) with UOTO = 1 (7)

ak = (s I Zk) = s, S(t)ZE(t) dt. (12)

and A last criterion for designing such an Hilbertian basis is

z ( t )=
{1
q if It1 5 70/2
elsewhere
related to the transmission channel itself. In a dispersive
channel, the transmitted signal is spread in time and in
frequency. In practice, we can also include in the channel
We can easily give a graphical representation of this set of model other distortions of the signal due to the receiver im-
functions in the time-frequency plane by their first order plementation, such as local oscillator offset and phase noise,
moments synchronization errors or jitter. We expect the transmitted
signals to be distorted as little as possible in such a channel.
The optimum solution to meet this requirement is to use
basic signals which are localized in time and in frequency
(9) with the same time-frequency scale as the channel itself.
More precisely, if AT and Au are respectively the channel
The coordinates of the basic functions form a two- delay spread and frequency spread, we can define a time
dimensional (2D) lattice in the time-frequency plane. The frequency scale TO and uo so that
density of this lattice is equal to unity, that is UOTO = 1.
This set of functions is orthonormal and form a Hilbertian
basis of L2(R). It must be noted that UOTO = 1 is a
necessary condition (while not sufficient) in order to 3) About the Existence of Other Hilbertian Bases: The
obtain a Hilbertian basis [7]. The projection of a given OFDM scheme previously described seems so straight-
signal on this basis can be thought as partitioning the forward that we could reasonably expect that other bases
signal in subsequences of duration TO. each of them being might exist, based on a similar approach. Another example
represented by the associated Fourier series. This is a is obtained by inverting the time and frequency axes. In
first step toward localization in both time and frequency this case, the prototype function is the Fourier transform of
domain, compared with conventional Fourier analysis, the initial window, i.e., a (sinz)/z function. This is in fact
which provides a perfect frequency localization, but on the asymptotic case of a system using nonoverlapping
the other hand loses the time information. frequency bands for elementary channels, which can
2) Hilbertian Basis and Modulation Schemes: The inter- be separated by rectangular filters. We could call this
est in Hilbertian basis for digital modulation can be approach the “zero roll-off approach.” Unfortunately,
understood from the Shannon theorem, which can be other extensions are not obvious. Even though other
written as functions having the same orthogonality properties as
the rectangular window and the (sinz)/z window have
number of bits per dimension = (log2(l+S/N))/2. (10) been described in the past [6], they are not satisfactorily

984 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 localized in the time and frequency domains, and, Now, we consider the situation where the QAM modula-
in addition, they are asymmetrical. Furthermore the tion of each carrier is replaced by an Offset QAM (OQAM)
Balian-Low-Coifman-Semmes theorem [7] shows that if scheme. This corresponds to another well known parallel
z(t) is a prototype function having the same orthogonality transmission scheme [9]-[ll]. In this case also, a link
properties as the rectangular window, then AtA f = CO, A t between the modulation scheme and the Hilbertian basis
and Af being the second order moments of the prototype can be established, but for this we need a redefinition of
function, defined as the signal space.

c At2 = Jt’(lz(t)((’ dt
Af’ = Jf211x(f)112df*
Therefore, we cannot expect to find a prototype function so
(14)
We have defined up to now the signal space as the Hilbert
space L’(R) with the standard inner product

(E I Y) = J,Z(t)Y*(t) dt. (17)

that both z ( t ) and X ( f ) decrease faster than ltJ-3/2 and
If I-’/’ respectively. This does not preclude that “good” Another way to define this Hilbert space is to use another
prototype functions might exist from the engineering point inner product
of view, but this leaves in our mind few chances to find
them. (z I Y)R = %le J,z(t)y*(t)dt. (18)
Another classical basis of L2(R) having good properties
regarding time and frequency localization was introduced Of course, the associated norm is the same in both cases
by Gabor [SI. The prototype function is a normalized
gaussian function
1141= J, ll~(t)1I2dt- (19)

A first example of a Hilbertian basis of L2(R) with this

new inner product is obtained with the prototype function
z(t) defined by its Fourier transform
It can be demonstrated that the set of functions zm,+(t)
forms a basis, the basic functions being perfectly localized
in the time-frequency plane. However, this basis is no
longer orthonormal. Furthermore, using Daubechies termi- If we consider z ( t )and its Fourier transform, we can notice
nology [7], this set of functions is not even a frame, which that X(f ) is strictly band-limited and that z ( t )decreases as
means that we cannot link the norm of a signal in L’(R) ltl-’. This is a much better result than the theoretical limit
to the norm of its coordinates over the basis by a relation derived from the Balian-Low-Coifman-Semmes theorem.
such as Let us define a set of basic functions zm,*(t) obtained by
translation in time and frequency of this prototype function
~ 1 1 ~ i1 1 ~ ll(s I zm,n)Il’ I~ l l s 1 1 ~ (16)
jm+n Zirrmvotz(t - n70)
m,n zm,n(t) = e
where A and B are two positive constants independent with
of s. This situation is very far from the usual case (for VoTo = 1/2. (21)
instance with a conventional QAM modulation) where
It can be easily shown [12] that this set of functions is
the minimum Euclidian distance between the transmitted
a Hilbertian basis. This result can be extended to any
signals is equivalent to the minimum Euclidian distance of
even function whose Fourier transform satisfy the following
the code. For digital transmission,where Euclidian distance
between signal is a major concern, this is obviously a major conditions:
drawback. X(f) = 0 If I 2 vo
Coming to this point, it might seem that if we want to
construct a modulation scheme making an optimal use of
{ +
lX(f>l2 1X2(f - v0)l2 = l/vo
if
if If I < vo
(22)

which corresponds to a half-Nyquist filter.

the signal space dimension, based on elementary signals Then, an OFDWOQAM signal s(t) can be written as in
which are correctly localized in both time and frequency (6) where the coefficients am,ntake real values representing
domains, then the OFDM approach is more or less a unique the transmitted encoded data.
solution. However, as will be explained below, such is not
the case. C. The OFDMMSK Approach
We will define the OFDM/MSK (minimum shift keying)
B. The OFDM/OQAM Approach
in the same way as an OFDWOQAM modulation, but using
In standard OFDM, the transmitted data are complex. a different prototype function. The name OFDM/MSK has
Each carrier is QPSK modulated (without any filtering), been chosen because each “carrier” can be considered as
or more generally QAM modulated. Therefore, we shall MSK modulated. The prototype function can be written as
call it OFDWQPSK (as in the DAB system [l], [3]) or
OFDWQAM, so that it can be differentiatedfrom the other
approaches introduced hereafter.

LE K O C H er al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 985

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 In fact, this modulation is nothing else than the dual Therefore, z , can be taken as the prototype function of
of OQAM. We have just inverted the time and the fre- a Hilbertian basis. Furthermore, Fz, = zlla. A special
quency axis. Therefore, the prototype function is time- case of particular interest is the function 3 = 21, which is
limited instead of being band-limited. The main difference therefore identical to its Fourier transform. This function is
is the implementation of the receiver, where the number represented on Fig. 3. Fig. 4 shows the ambiguity function
of tap of the input filter is greatly reduced compared with of 3, compared to the ambiguity function of a gaussian
OFDM/OQAM. function, which is perfectly isotropic in the time-frequency
plane. The top of these two functions are nearly identical,
D. The Isotropic Orthogonal Transform the main difference being at the base of the ambiguity
Algorithm (IOTA) Approach function. We consider the Hilbertian basis {3m,n} defined
All the modulation systems described previously are as
based on well known signal processing techniques: Short-
Time Fourier Transform for OFDWQAM, para-unitary
QMF filter bank for OFDWOQAM, and Lapped Orthogo-
nal Transform for OFDM/MSK. In all these examples, or-
thogonality between basic functions is obtained by straight-
forward methods, making use either of the time or fre-
We can alternatively choose
quency limitation of the prototype function. A different
approach is detailed here. m + n even
We consider here a Hilbertian basis defined according to
an OFDWOQAM scheme
{ ( P ~ =, 0~
( P ~ =, r~/ 2 m + n odd. (31)

xm,n(t)= j m + n e 2 i ~ m v ox(t
t Using this Hilbertian basis, we can define a new transform
- m-0) (24) that we shall name IOTA because of the nearly isotropic
with properties of the prototype function. The IOTA transform of
VoTo = 112. a signal s ( t ) having a Fourier transform S(f) is defined as
We shall normalize the time and frequency scales, taking
vo = TO = 1/& Let us define 0 as the orthogonalization
operator which transforms a function x into a function y,
am,n = Re
J s(t)S;,,dt

where
21/4x(u)
Y(U) = (25)
JC, llx(u - wm2
The effect of the operator 0 is to orthogonalizethe function
y along the frequency axis. The inverse transform being given by
Let us define the (Woodward) ambiguity function of the
function y as

A,(T, v ) = J' y(t + 7/2)y*(t - ~ / 2 ) e - ~dt.~ ~(26)

" ~

This function can be seen as a 2D autocorrelation function

in the time frequency plane.
The effect of the operator 0 can be seen directly on the
ambiguity function:
(33)
AY(O,m&) = 0 , m # 0. (27)
We use this inverse transform to define an OFDMDOTA
Similarly, F - l O F orthogonalizes the function y along the modulation, where the coefficients am,n of the signal s ( t )
time axis ( F being the Fourier transform operator). This take real values representing the transmitted encoded data.
can also be seen directly on the ambiguity function
A,(nJZ, 0 ) = 0 , n # 0. (28) E. Pe$ormance Analysis of OFDM Schemes
In multipath propagation, the channel can be described
This orthogonalization method is attributed to Henri
by a sum of elementary paths defined by a delay T, a
Poincark. More recently, it has been used in wavelet theory.
frequency offset v, an amplitude p and a phase cp. There-
Let us consider now the function z , = F - l O F O ~ , ,
fore, the sensitivity of any modulation scheme to multipath
with x, = (2a)1/4e-rau2 . Then it can be demonstrated
propagation can be evaluated from its sensitivity to a static
that
delay and frequency offset. It can be demonstrated that for
A Z a ( n & , m h )= 0, ( m , n )# (0,O). (29) any OFDM scheme defined by a Hilbertian basis and a

986 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 I I

II time frequency
6.4
2 1 0 1 2 3 (a)

4 2 0 2 4 6 Fig. 4. (a) Linear representation of the IOTA ambiguity function,

(b) logarithmic representation of the IOTA ambiguity function.
(b)

Fig. 3. (a) Linear representation of the IOTA function, (b) loga- demonstrated that this parameter is always less than unity.
rithmic representation of the IOTA function.
If the prototype function is real, this parameter is simply
related to its second order moments by the relation:
prototype function x(t), the intersymbol variance can be
written as 5 = 1/4rAtAf (36)

I = (1 - (%e[A,(.r,v ) ] ) ’ ) 0 2 (34)
where A, is the ambiguity function of x defined in Section
111-D, o being the variance of the transmitted data. In the
Consequently, in the case of OFDWQAM, 5 = 0. If we
case of an even prototype function, the ambiguity function
consider the normalized intersymbol interference function
is real, and we can simply write
shown in Fig. 5 , we find that the sensitivity to a static delay
V ) = (1 - A ~ ( T
I(T, v ), ) c 2 . (35) is very high. The reason for this is that A f is infinite. This
does not occur for a frequency shift, because At has a
Therefore, the behavior of any OFDM scheme regarding finite value.
multipath is closely related to the ambiguity function of its This high sensitivity to a static delay, and therefore
prototype function. We can even appreciate this sensitivity to multipath propagation, seems to be a major drawback
with a single parameter by considering the opening of of OFDWQAM. But fortunately there is a way out. As
the cone which is tangent to the normalized intersymbol described in Section 11, in the case of the OFDWQAM, we
interference function. This opening can be measured by the can extend the prototype function with a “guard interval.”
area of the surface 5 corresponding to the intersection of the In such a situation, the system is totally insensitive to
cone and the plane of maximum intersymbol interference, the delay as long as this delay is smaller than the guard
i.e., I(T,v) = 0 2 .Using the Heisenberg inequality, it can be interval. Fig. 6 shows the effect of the guard interval on

LE K O C H er al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 987

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 frequency
0 I/ , I
I
, I
lS O5 O d5 -l -15
time

Fig. 5. OFDWQAM intersymbol interference function without Fig. 7. OFDM/IOTA intersymbol interference function.
guard interval.

same E parameter. Therefore, OFDMMSK can be preferred

to OFDWOQAM in most applications.
c
The parameter is even improved in OFDMAOTA and
equals to 0.977 in this case. As we know, the theoretical
value is 1. This value cannot be reached, for it corresponds
to a gaussian prototype, which does not generate an Hilber-
tian basis. Therefore, it is our conviction (while not yet
demonstrated) that OFDMAOTA is an absolute optimum.
The corresponding intersymbol interference function is
0 4 i
shown in Fig. 7, and has a nearly circular symmetry.
1 From the implementation point of view, however, the
02;
simplest scheme is the OFDWQAM: no prefiltering is re-
0 , , I
I / 1
frequency
quired. In the case of OFDMMSK, the prefiltering is quite
15 1 05 o -OS .1 .,5
simple. In the case of OFDMDOTA, the prototype function
time
has a fast decreasing property (from a mathematical point of
Fig. 6. OFDWQAM intersymbol interference function in pres- view). Therefore, it can be easily truncated. OFDWOQAM
ence of a guard interval. leads to more complex prefiltering, because the prototype
function decreases slowly.
the intersymbol interference function. Of course, the guard Compared with OFDWQAM: OFDWOQAM,
interval does not come for free: first, there is some loss due OFDM/MSK and OFDMOTA are more complex in
to the power transmitted during the guard interval, secondly, term of FFT by a factor of 2, because of the density of the
there is a loss of some space dimensions (i.e., channel lattice of the basic functions. Also the channel estimation
is more complex. Finally, the choice between these various
capacity). Thus for a given overall spectrum efficiency,
methods greatly depends on the channel parameters, the
we need more bits per dimension, and this leads also to
spectrum efficiency which is needed and the cost of the
a penalty with respect to the Shannon limit.
receiver. For the time being, the OFDWQAM scheme is
The price to be paid greatly depends on the duration
preferred for its simplicity of implementation in hardware
of the guard interval which is needed and the spectral (no prefiltering required and simple FFT).
efficiency. As an example, if we want an overall spectrum
efficiency of 4 b/s/Hz with a guard interval of 1/4 of the I v . CHOICE OF THE CODE
useful symbol, we need a coding scheme transmitting 2.5
bits per dimension instead of 2, which represents a 3 dB A. Introduction
loss with respect to the Shannon limit. Adding the power Convolutional channel coding has become widespread
wasted in the guard interval itself, the price to pay is 4 dB. in the design of digital transmission systems. One major
In the case of OFDWOQAM, there is no guard interval, reason for this is the possibility of achieving real-time
and such a technique is not even feasible. The parameter decoding without noticeable information losses, thanks to
c equals 0.876. Therefore, OFDWOQAM is not far from the well known soft-input Viterbi algorithrfi [ 131. Moreover,
having optimum properties regarding multipath. However, the same Viterbi decoder may serve for various coding
its prototype function decreases relatively slowly in the time rates by means of puncturing [14], allowing the same
domain. This leads to a larger number of taps in the filter silicon product to be used in different applications. Two
of the receiver. This problem is solved with OFDM/MSK, kinds of convolutional codes are of practical interest: non-
which uses a finite prototype function, while keeping the systematic convolutional (NSC) and recursive systematic

988 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 10
! Gain

NSC code W I
Tu-code
8 ( v = 3 . 3 iter )
L
V=8
6 '-6 Simple
k'
v=4 Code .
4 v= 2
data dk v
> 'k
2

RSC code 1 - 1 0
Bandwidth
Expanslon
0.5 1 1.5 2 2.5 3 Rale(dB)

Fig. 9. Coding gains at BER = lom6 of simple NSC codes

(3-b quantization [17]) and turbo code (see Section IV-B) with
k' v = 3, 48 x 32 interleaving and three decoding iterations (4-b
quantization), for rates 1/2,2/3,
3/4,and 4/5in a Gaussian channel,
Fig. 8. Nonsystematic convolutional (NSC) and recursive sys- with QPSK modulation.
tematic convolutional (RSC) codes with memory code v = 3 and
polynomials 15, 17.
code) and a convolutional code (the inner code) in a serial
structure. Another-all convolutional-concatenated code,
convolutional (RSC) codes. Fig. 8 shows both types of named turbo code [18], [19], has been proposed in the
codes built with the same parameters: code memory v = recent past, which offers good performance and reasonable
+
3 (constraint length K = v 1 = 4), polynomials 15, 17, decoding complexity.
and rate R = 112. Though RSC codes have the same free
distance d f (i.e., the minimum Hamming distance between
two different sequences at the output of the encoder, with
the same starting and ending states) as NSC codes and B. Turbo Codes
exhibit better performance at low signal to noise ratio A turbo-encoder is built from a parallel concatenation
(SNR) and/or when punctured [15], only NSC codes have of two elementary RSC codes, separated by a nonuniform
actually been considered for channel coding, except in interleaver (defined below). Fig. 10(a) depicts such a struc-
trellis coded modulations (TCM) [ 161. ture with v = 3, leading to a basic R = 113 encoder.
For a given rate, the error-correcting power of convolu- Parallel concatenation, unlike the serial one, enables the
tional codes, measured as the coding gain at a certain binary whole encoder, and therefore the whole decoder, to run
error rate in comparison with the uncoded transmission, with a single clock. RSC codes have been chosen instead
grows more or less linearly with code memory v. Fig. 9 of NSC codes for two reasons. First and mainly, parallel
(Simple Code [17]) shows the achievable coding gains concatenation can only be built with systematic codes.
for different rates, and corresponding bandwidth expansion Secondly, for a global coding rate less than or equal to
rates, by using NSC codes with v = 2, 4, 6, and 8, for 112, both elementary codes have to be punctured and, as
a BER of For instance, with R = 112, each unit mentioned above, RSC codes are superior, especially at low
added to v gives about 0.5 dB more to the coding gain, SNR. When a concatenated coding scheme is considered,
up to v = 6; for v = 8, the additional gain is lower. giving anyhow a steep decrease in the BER curve, it is
Unfortunately, the complexity of the decoder is not a linear favorable to have performant codes at low SNR. Nonuni-
function of v and grows exponentially as v * 2'. The factor form interleaving consists at once in scattering data, as with
2' represents the number of states processed by the decoder usual interleaving, and giving the maximum disorder in
and the multiplying factor v accounts for the memory part the interleaved data sequence. The latter property makes
complexity (metrics and survivor memory). Other technical redundancy generation by the two encoders as diverse as
limitations like the interconnection constraint in the decoder possible. The decoder, according to the natural scheme in
layout make the value of 6 a practical upper limit for v, Fig. 10(b), uses each elementary decoder by feeding back
for most applications. results from one to the other. Data processed by each
In order to obtain high coding gains with moderate decod- elementary decoder are
ing complexity, concatenation has proved to be an attractive -symbols XI, (or Xn), representing information data,
scheme. Classically, concatenation has consisted in cascad- with additive noise,
ing a block code (the outer code, typically a Reed-Solomon -its own redundancy Y l k (or Y ~ I ,with
) , additive noise,

LE K O C H ef al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 989

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 -the extrinsic piece of information zk (or Z,), provided
by the other decoder. 'k

The output of the global decoder is the output of the

elementary decoder DEC 1 or DEC2, after desinterleaving,
and provides the logarithm of likelihood ratio (LLR) about
dk. The feedback or extrinsic information zk (or 2,) is
calculated as the output of the related decoder, reduced by
'3,
the contributions of inputs X and 2, at the corresponding uniform
time, IC or n. It provides an estimate of data dk (or d,), hav-
ipg a noise weakly correlated with the one which corrupts
interleaving
>- 2
',

symbol XI, (or x n ) , and the larger the interleaving size,

the lower the correlation between these noises. Parameter
y takes into account the noise variance of symbols X (or
Y) and 2, which are different.
The natural parallel version of the global decoder can
be advantageously replaced by the serial structure of
Fig. 1O(c), for there is only one piece of extrinsic informa-
tion to compute and one interleaver. In this case, a second
parameter (p) is necessary to compensate for unequal noise
variances of Xan and Y2k. Of course, the actual decoder
cannot be built as schematically described in Fig. 1O(c),
because of the different latencies, and the iterative process-
ing that must be carried out. One iteration corresponds to
passing through both decoders DECl and DEC2.
The basic idea for turbo-decoding relies on this observa-
tion: without the feedback loop, decoder DEC2 takes benefit
of the whole redundancy information, whereas decoder
DECl just uses redundancy Y1k; hence, the decoding
scheme is not optimal. The role of extrinsic information
is then to compensate for this dissymmetry.
Turbo codes with rates higher than 1/3 are easily obtained
by puncturing symbols Y1k and/or Y&. For high rates (R2
5/6), puncturing xk may be preferable. The best results
achieved with turbo codes have been reported in [18].
By using a very large interleaving matrix (256 x 256)
with appropriate nonuniform reading [ 191, and optimum
decoding for v = 4 elementary codes, a BER of was
obtained with additive white gaussian noise (AWGN) for
R = 1/2 at EI,/N, = 0.7 dB, after 18 iterations.
In practice, lower sizes for interleaving must be consid- Fig. 10. Turbo-encoder and decoders with natural parallel or se-
ered, typically 32 x 32 or 48 x 32. Furthermore, optimum rial structure.Int: interleaving,De: desinterleaving.LLR: logarithm
decodinp has to be replaced by soft-output Viterbi decoding of likelihood ratio about d k . (a) Turbo-encoder, (b) turbo-decoder
with parallel structure, and (c) turbo-decoder with serial structure.
[20] for acceptable complexity. Lastly, the number of iter-
ations for a monolithic integration in silicon is necessarily
short: 3 or 4. All this leads to some degradation relatively to present case, it is suitable to high speed operation. In fact,
the optimum case. Fig. 9 again shows the gain achievable in such a circuit, the critical path is the memory access,
with a monolithic turbo-encodeddecoder VLSI circuit, by and data rates up to 200 Mb/s in CMOS 0.6 p technology
comparison with the simple convolutional codes. It is built may be achieved by using parallelism, with one iteration
with two v = 3 identical codes, a 48 x 32 interleaving per circuit to avoid dissipation problems.
matrix and decoding is performed in three iterations. It can
be observed by extrapolationthat the error-correctingpower
is comparable to that of a simple code with v = 12. The C. Synchronization of Turbo-Decoders
material complexity of this circuit being about that of a v = Like any concatenated decoder, turbo-decoders have to
7 simple code (including all memories and synchronization be node and time synchronized. Because a turbo code is
functions), the gain in complexity given by the turbo code, a systematic code, synchronization may directly rely on
relatively to the simple code, is roughly 32. data frame synchronization, if any. In other cases, internal
Because a turbo-decoder is like a pipelined cascade of synchronization without loss of rate, is proposed. It is based
Viterbi decoders with a short number of states, 8 in the on an inverting pattern method and consists in replacing

990 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 redundancy Y1k at the encoder output by its inverse at BER I I I I
predetermined instants, according to a flag sequence. De-
coder DECl tries to locate this flag by comparing its binary
decision on Y1k with the corresponding input, which is not
used by the decoder. All combinations are scanned, which
may take some time, depending on the coding rate and
the type of modulation chosen. After synchronization, a
supervision function in DEC2, using the pseudo-syndrome
method [21], is in charge of detecting any out-of-synchro
situation in order to engage a new synchronization process.

D.Turbo Codes Compared to Classical Concatenation

Since a turbo code is a concatenated code, it is realistic to
compare its performance to that of a classical concatenated
code. Such a comparison is made with a particular example
in Fig. 11. The turbo code is the same as mentioned above:
v = 3, 3 iterations, interleaving 48 x 32. Its rate is
R = 314. The classical concatenated code uses a NSC
inner code with U = 6 and R = 415. The outer code is
a (204,188) Reed-Solomon (RS) code with 8-b symbols.
Its correction capacity is t = 8 symbols in the block of
204. The global coding rate of this concatenated code is
slightly lower: 0.737. For the latter code, three interleaving
depths (the number of blocks of 204 symbols processed
by the interleaver) are considered: I = 1 (no interleaving
between outer and inner codes), 2 and 4. In all cases,
the petformance is given taking practical limitations into
consideration (quantization, truncation length of Viterbi
decoders, etc. ...).
For an equivalent material complexity and comparable 3 4 5
latencies (5 or 6 kb) of the two encoders/decoders (turbo EbINo
coding brings no latency), we have to consider the case
Fig. 11. Comparison of a turbo code (same as Fig. 9, R = 0.75)
I = 2 for the classical concatenated code and then the and a classical serial concatenated code ( R = 0.737) performance
performance of the turbo code is better for the practical in a Gaussian channel (I:interleaving depth).
values generally sought for the BER. As allowed by the
principle of turbo codes, the slope of the curve may
Gaussian channel. Curve (a) is for a 16 QAM associated
even be increased by cascading two identical circuits, thus
with the same turbo code as presented above (v = 3,
performing six iterations. In this case, the performance of
3 iterations, interleaving 48 x 32, R = 314, integrated
the turbo code is always better, whatever the interleaving
circuit limitations). All information bits are protected and
depth of the classical concatenated code, but at the expense
of a doubled latency. decoding is carried out in a similar way as the pragmatic
Since puncturing a turbo-encoder leads to coding rates approach proposed by Viterbi et al. [23]. Indeed, the same
of the form mln, m, and n > m being any integers, its turbo-encoderldecoder, with puncturing, may be used in
association with various modulations is simple especially association with various modulations and different spectral
with respect to clock generation. For instance, this 314 rate efficiencies. It just needs specific transcoding between the
code may be directly combined with a 16-QAM modulation encoder and the modulator, and between the demodulator
(see next section) with a modulation rate exactly one third and the decoder, to realize adaptation from binary symbols
of the data rate. This is not so simple for the classical to nonbinary symbols, and vice-versa. Curve (b) gives the
concatenated code. theoretical performance of the @-state TC- 16QAM, which
is less favorable for BER lower than despite a greater
E. Turbo Codes Compared to TCM decoding complexity. Another, still more complex possible
Thanks to feedback information zk which acts as a transmitting scheme (curve (c)) associates the 64-state TC-
diversity factor, turbo codes are really robust codes when 16QAM and the RS (204,188) code with interleaving depth
associated with high spectral efficiency modulations andor I = 2. The spectral efficiency is slightly lower (2.76
when used on fading channels. Particularly, it has been b/Hz/s). The theoretical performance is comparable to the
shown [22] that turbo codes offer much better protec- turbo code solution. These results may be extended to
tion than TCM on both Gaussian and Rayleigh channels. various comparisons that can be made between TCM and
Fig. 12 gives the example of a 3 b/Hz/s transmission on a noncoded modulations with channel turbo-coding, whatever

LE K O C H et al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 991

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 BER I
lo-’
I I I I I I Bit Error Ratio , Gaussian channel
Selective channel
5

lo-*
5

1 CM ratio
Fig. 13. Illustration of the principle of additivity of the echoes.

by equipment limitation is not considered here) shows

performances which depend on two factors:
-the carrier-to-noise ratio ( C / N )
-the characteristics of the distortions.
When the channel is perfect (this is an unrealistic situation),
the BER performance is referred to as the Gaussian channel
performance. When the channel is strongly distorted, in-
cluding echoes that can be as powerful as the direct path, the
6 7 8 9
EblNo BER performance is referred to as the frequency selective
channel (worst case) performance.
Fig. 12. Comparison of a turbo code and a TCM performance in In the COFDM system, the difference between these two
a Gaussian channel. Curve (a): 16 QAM with R = 3/4 turbo code
(same as Fig. 9), spectral efficiency: 3 b/Hz/s. Curve (b): 64-state BER performances depends on the strength of the code
TC-16QAM (theoretical), spectral efficiency: 3 b/Hz/s. Curve (c): used.
64-state TC-16QAh4 with (204,188) RS code (I = 2), spectral The principle of additivity of the echoes can be inter-
efficiency: 2.76 b/Hz/s.
preted in the following way (see Fig. 13): starting from the
Gaussian channel (only one direct path), with a BER of
the spectral efficiency. In a Rayleigh channel, turbo codes lo3, the addition of an echo as strong as the direct path
may outclass TCM by more than 8 dB [22]. This is mainly will lead to the following two consequences:
because all bits are protected by the code, which moreover -a 3 dB increase of the C / N ratio,
offers a good performance despite puncturing. -a change from the Gaussian performance curve to the
Thus turbo codes represent a suitable solution for channel Frequency Selective performance curve.
coding in most digital communication systems. With cost- If the code used provides a distance between the two curves
effective VLSI circuits, good performance (and often the lower than 3 dB, then the BER will be decreasing when this
best at the present time) can be obtained for various spectral first echo is added; this is the case in the Eureka 147 DAB
efficiencies. For digital terrestrial television, a specific system.
cascadable VLSI turbo-decoder has been designed with 16- If the code used provides a distance between the two
state soft-outputViterbi decoders and a 64 x 32 interleaving curves higher than 3 dB (this may be the case in the TV
matrix. The performance it offers is given in the following broadcasting application where the need for a high spectrum
section. efficiency makes the coding performance demand more
challenging), the BER will then be increased by this first
v . SYSTEM PE%FORMANCES
“0 dB echo.”
The COFDM system has the potential to use an echo When afterwards further echoes are added, they will
to improve its performance. This has often been shown in all contribute to increasing the C / N ratio. However, the
demonstrationsby adding a 0 dB echo with 20 ps delay for performance curve will remain at worse the frequency
example to a transmission in white Gaussian noise which selective one, and therefore the BER will decrease with the
already includes a direct path and another 0 dB echo with insertion of further echoes. In that sense, COFDM benefits
15 ps delay. The power of the second echo (the 20 ps one) from the power addition of echoes to improve performance,
is used by the COFDM system to reduce the bit error rate as was observed in laboratory measurements using real time
(BER)! How is it possible? hardware.
Any digital transmission system in presence of linear The COFDM system will operate under the above men-
channel distortions (nonlinear distortion usually caused tioned condition as long as the multipath delay is contained

992 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 within the extent of the guard interval. However, an echo
exceeding the guard interval duration will not lead to an A
A
abrupt degradation of the performances: it will remain
graceful. An echo of power W delayed by r will present a
constructive component C and an interfering component I .
It can be shown that the powers of these two components
are given by
C = W . f ( ~ ) and I = W . (1 - f ( r ) ) (37)

where
f ( r )= 1 when T 5 A, (no interfering component)
f ( ~=
) (1 - s)~,
r-A
when A 5 r 5 TO I Gaussian channel Frequency-selective channel
I I I
I I I I I I
and 0 5 10 15 20 25 30
) 0 when T > 7'0
f ( ~= Carrier to noise ratio (dB)

where TOrepresents the total length of the COFDM symbol,

and A represents the duration of the guard interval, as
presented in Section 11. In presence of a received signal
constituted by i paths of powers W;, with individual
constructive component Ci and interfering component Ii,
the global C / I ratio then equals Fig. 14. COFDM performances.

of single carrier modulations is the handling of echoes

This formula allows the computation of the C / I ratio in and the resulting intersymbol interference (1%): unlike
the case where echoes are longer than the guard interval. COFDM, the symbol length is very small (typically 0.15
Relating it to the Bit Error Rate versus C / N curve permits ps), and echoes may occur, that are much longer than
to quantify the effect of echoes with very long excess delay a symbol (i.e., up to tens of ps in an urban area, up
on the system performance, as it can be shown that the to more than 100 ps in a Single Frequency Network).
interfering component has the characteristic of a Gaussian This means that the received signal samples are actually a
noise. combination of many consecutive symbols. The coefficients
Fig. 14 gives the performance of the COFDM system in of this combination reflect the channel impulse response.
terms of possible tradeoffs between the available bit rate in The comparison is limited to linear equalization, which is
an 8 MHz channel, and the carrier to noise ratio necessary to often considered for long term echoes mitigation. However,
obtain an unimpaired picture quality. These results include it is known that techniques such as Decision Feedback
the overheads due to the guard interval, the frequency guard Equalization are also considered for their efficiency in the
bands and the references for receiver synchronization, so presence of strong echoes.
that the effective spectrum efficiency is reduced by a factor Linear equalization roughly consists in inverting this
314. The performances are given for two types of codes: combination by combining several consecutive received
usual convolutional codes (v = 6) and turbo codes (v = samples with the proper coefficients. Since the coefficients
4, 3 iterations). depend on the channel response, such channel response
The Shannon limit is the theoretical limit for error-free must hence be evaluated. This may be achieved through a
digital transmission over a Gaussian channel (channel only training sequence (start-up), followed by a blind adaptation
affected by additive white Gaussian noise). of the coefficients (tracking). From a theoretical point of
The figure has to be understood in the following way: the view, the performance of ideal linear equalization can be
left end of the segment gives the performance of the coded assessed by considering frequency domain equalization,
modulation in the Gaussian channel; the right end of the where the transfer function of the equalizer is given by
segment gives the performance in the most difficult channel the formula
(the frequency-selective channel), in which full notches due
(39)
to echoes as strong as the main signal may appear.
where H ( v ) is the channel transfer function, and U; and CT:
VI. COFDM VERSUS LINEAR EQUALIZATION represent the variance of the noise and the data symbols,
OF CODEDSINGLECARRIER SCHEMES respectively.
Alternative solutions to the COFDM technique are known This equalizer is known to achieve the Minimum Mean
as single carrier modulation techniques. The main problem Square Error criterion, thus minimizing the combined effect

LE FLOCH et al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 993

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 BER BER

1 o-2

Fig. 15. Compared performances of COFDM and coded single Fig. 17. Compared performances of COFDM and coded single
carrier scheme under channel profile P O . carrier scheme under channel profile P6.

BER
and P6, is approximately 1 dB, which confirms that
the performance dispersion of turbo codes between
a “Gaussian-like” and a “Rayleigh-like” channel is
rather limited; this is a good indication of the intrinsic
system A
system E code performance.
-despite the use of the same code associated to the
Carriar to N o h same modulation in the case of optimal linear equal-
ratio (dB) ization of the single carrier scheme, the performance
15 16 17 18 19 20 21 22 improvement between PO and P3 is of the order of 10
Fig. 16. Compared performances of COFDM and coded single dB, and still a little more than 2 dB improvement is
carrier scheme under channel profile P3. achieved with profile P6. This shows that the statistics
of the maximum likelihood decoder inputs are of
utmost importance and strongly condition the decoding
of additive noise and residual intersymbol interference. In
performance. In the case of COFDM, the maximum
order to compare the performances of such equalization
likelihood decoder is able to benefit from full channel
applied to a single carrier modulation scheme, with those
knowledge thanks to its weighted inputs (i.e., metrics
of COFDM, the following hypotheses are retained
of the Viterbi decoder) that represent a reliable image
-system A is a COFDM system using a 64-QAM rate of the channel distortions.
2/3 turbo coded (v = 4, 3 iterations) modulation on -the performance differences between systems A and
each carrier; the receiver applies a coherent demod- B, at BER, are equal to 11 dB, 2.5 dB, and 1
ulation with an assumed perfect knowledge of the dB for channel profiles PO, P3, and P6, respectively.
transmission channel.
-system B is a single carrier scheme using the same
VII. BROADCASTING NETWORKARCHITECTURE
64-QAM rate 2/3 turbo coded modulation; the re-
ceiver applies the linear equalization defined by the The following paragraph examines COFDM in terms of
above transfer function G(v) (that assumes also perfect a “broadcasting system” and shows that the new horizons
knowledge of the transmission channel), and coherent opened up by this technique are even more attractive when
demodulation is applied. looking at the implementation of a broadcasting system as
a whole.
These two systems are passed through the same transmis-
sion channel, which is defined by the following impulse A. Gap-Fillers
response, containing one direct path and a single echo
Since the system has been designed to take advantage of
h(t)= S ( t ) +a .S(t - T). (40) echoes, it is possible to operate the system in the presence of
echoes which are created voluntarily. The initial application
The parameters T is chosen to be much larger than the of this idea provides the means of eliminating residual
inverse of the signal bandwidth (about 100 times), so that shadowed areas by using passive reflectors or small active
the simulation results are independent of the echo phase. relays without having to change the carrier frequency for
Three values of the parameter a have been retained: 1, the reflected or relayed signal. The signal is picked up at
I/&, 112. This corresponds to an echo attenuation of 0 dB, a location where reception conditions are satisfactory. It
3 dB, and 6 dB, respectively, and the three channel profiles is then reamplified and rebroadcast at the same frequency
are denoted by PO, P3 and P6. The results obtained are toward the shadowed area that is to be suppressed. These
plotted in Figs. 15-17. “gap-fillers’’ do not require any additional frequency and
The following conclusions can be derived from these their intrinsic simplicity ensures very low costs. If we
results: consider that the total number of relays necessary to provide
-in the case of COFDM, the performance improve- national TV coverage in a country such as France adds up
ment between PO and P3, as well as between P3 to several thousand units, it is easy to see the benefits of

994 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 such an approach. Since 1988, full-scale tests around 900
a dense nehvork
MHz have been carried out in Rennes, France, and have
successfully shown the feasibility of these “gap-fillers.’’
u e f i received pawer
wasted power
B. Dense Networks and Power EfJiciency of the System
If one imagines a network of terrestrial transmitters
distributed over a given territory, all of them time-
synchronized and broadcasting the same signal on the
same frequency, then the useful power received at the
input of the receiver is the sum of the incoming power
from each transmitter. The various incoming signals are coverage area

seen as echoes of the same signal and combine positively Fig. 18. Power efficiency of dense networks.
if their temporal spread is compatible with the selected
duration of the guard interval. In other words, COFDM
enables constructive overlapping of the various areas of
transmitter coverage. The linking, by this technique, of
the largest possible number of interdependently operating
transmitters offers many advantages:
-The broadcasting infrastructure is less expensive since
it avoids the need for excessively powerful transmitters tlu pnr.rt riImlion inlhenorv/uhm in6uIongrrrn
whose cost increases in relation to the area to be I program = I frcquemy
I program = 9frcquc~ics!
covered.
-It enables the launch of a new service with mini- (a) (b) (C)
mum initial investment, then a gradual expansion and Fig. 19. (a) Frequency planning for conventional analog or digital
systems, (b) introduction of a digital service in an analog context,
improvement in the area served. and (c) frequency efficiency of a single-frequency network.
-It makes more efficient use of the power transmitted
as shown qualitatively in the diagram below (Fig. 18).
-Unlike analog systems which cater for extremely vari- same program from geographicallyadjacent transmitters on
able signal to noise ratios, digital systems are very two distinct frequencies. If we consider that the 45 8-MHz
sensitive to a threshold effect within which there is channels in the UHF band allocated to television in France
marked quality loss and beyond which the quality is no are totally saturated by five national channels, COFDM can
longer improved. By enabling a more accurate desig- provide the means of increasing this number nine fold. If
nation of the area of coverage, and greater uniformity we also consider a rate of 24 Mb/s in each channel, then
with respect to the received power, dense networks the combination of COFDM and source coding provides
can considerably reduce this drawback, which is spe- a gain in terms of spectral occupation in a ratio of 9 x
cific to digital systems. Indeed, the concept could be 4 = 36, assuming that 6 Mb/s can provide a standard
summarized as: “putting the right power at the right quality picture. Fig. 19 shows the dramatic overall spectral
place.” efficiency achieved by a system based on COFDM.
-It improves control of co-channel interference at the
edge of the coverage area, and dramatically reduces the VIII. CONCLUSION
frequency reuse distances. This is a possible solution to
The tremendous improvements of the integrated circuits
the problem of spectrum sharing in dense urban areas.
technology today allows the real-time implementation of
-In addition to the time and frequency diversities pro-
communication systems which could hardly have been sim-
vided by COFDM, it brings spatial diversity into
transmission. ulated with the most powerful computers one decade ago.
This is the case of the COFDM system, which relies on the
An experimentalnetwork comprising two transmitters oper-
conjunction of very basic principles of the communication
ating at 60 MHz, which was brought into service around the theory.
town of Rennes, confirmed the fact that the overall coverage
In order to jointly resolve the problems of multipath
area served by two transmitters operating simultaneously is propagation in terrestrial broadcasting and of optimization
significantly greater than the sum of the coverage areas for of the spectrum usage through Single Frequency Networks,
each of the transmitters operating independently. this technique uses many tools which are now common to
the 1990’s engineer: Time-frequency transforms, maximum
C. Single Frequency Networks and Spectral likelihood Viterbi decoding, multidimensional filtering. The
EfJiciency of the System result obtained is not only an efficient modulation scheme
The concept of a network of transmitters operating syn- matched to high bit-rate transmission in a dispersive chan-
chronously on the same frequency conflicts with conven- nel, it also offers a promising approach to terrestrial broad-
tional techniques which require the broadcasting of the casting networks. Today, there are good reasons to believe

LE FLOCH et al.: CODED ORTHOGONAL FREQUENCY DIVISION MULTIPLEX 995

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.
 that the COFDM technology will be used to ensure the IEEE Int. Con$ on Commun.(ICC ’93),pp. 1064-1070, Geneva,
migration from analog to digital networks, for radio first, May 1993.
[ 191 C. Berrou and A. Glavieux, “Turbo-codes: General principles
and for television afterwards. and applications,” Proc. 6th Int. Tirrenia Workshop on Digital
The ruggedness of the system has been verified by real- Commun., pp. 215-226, Tirrenia, Italy, Sept. 1993.
[20] C. Berrou, P. Adde, E. Angui, and S . Faudeil, “A low com-
time implementation of a TV COFDM modem offering a plexity soft-output Viterbi decoder architecture,” Proc. IEEE
bit-rate of 21 Mb/s and able to work in a selective channel Int. Con$ on Commun. (ICC ’93),pp. 737-740, Geneva, May
affected by echoes as strong as the main path, with 1 dB 1993.
[21] C. Berrou and C. Douillard, “Pseudo-syndrome method for
implementation margin compared to the theoretical result. supervising Viterbi decoders at any coding rate,” Electron. Len.,
Depending on the time pressure of the standardization vol.. 30, no. 13, June 1994.
process, as well as on the tradeoff between performance [22] S . Le Goff, A. Glavieux, and C. Berrou, “Turbo-codes and
high spectral efficiency modulation,” Proc. IEEE Int. Conf: on
and receiver cost, different options of implementation can Commun. (ICC ’94),pp. 645-649, New Orleans, May 1994.
be selected, ranging from the classical implementation of [23] A. J. Viterbi, E. Zehavi, R. Padovani, and J. K. Wolf, “A prag-
COFDM to the most sophisticated design described in this matic approach to trellis-coded modulation,” IEEE Commun.
Mag., vol. 27, no. 7, pp. 11-19, July 1989.
article under the name of IOTA.

REFERENCES
Bernard Le Floch was born in Quimper,
[ l ] C. Dosch, P. Ratliff, and D. Pommier, “First public demon- France, in 1958.
strations of COFDM/MASCAM: A milestone for the future of He joined the CCETT (France Telecom and
radio broadcasting,” EBU Rev.-Tech. no. 232, Dec. 1988. Tklkdiffusion de France joint research center
[2] D. C. Cox and R. P. Leck, “Distributions of multipath delay for telecommunications and broadcasting) in
spread and average excess delay for 910 MHz urban mobile 1985. In 1987 he worked on digital audio
radio paths,” IEEE Trans. Vehic. Tech., vol. VT-26, Nov. 1977. broadcasting developments within the EUREKA
[3] D. Pommier and Yi. Wu, “Interleaving or spectrum-spreading 147 DAB project, and in 1988 became Head
in digital radio intended for vehicles,” EBU Rev.-Tech. no. of the Digital Modulation and Channel Coding
217, June 1986. Department. Since 1990his work has focused on
[4] S. B. Weinstein and P. M. Ebert, “Data transmission by fre- the development of digital terrestrial television,
quency division multiplexing using the discrete Fourier trans- and has been participating in the European RACE dlTb project as
form,” IEEE Trans. Commun. Tech., vol. COM-19, Oct. 1971. chairman of the Channel Coding and Modulation group. He is now Head
[5] J. F. Hklard and B. Le. Floch, “Trellis coded orthogonal of the CCETT division in charge of cable, satellite, and terrestrial digital
frequency division multiplex for digital video transmission,” broadcasting.
Proc. Globecom 91.
[6] E. Jensen, T. Hoholdt, and J. Justesen, “Double series repre-
. IEEE Trans, on Inf Theorv. vol.
sentation of bounded signals.”
-
34, July 1988.
[7] I. Daubechies, “The wavelet transform, time-frequency local- Michel Alard (Member, IEEE) was born in
ization and signal analysis,” IEEE Trans. Inform. Theorv. vol. Plouknan, France, in 1951. He graduated from
36, Sept. 1996 the Ecole Polytechnique, Paris, in 1976 and the
_181_ D. Gabor, “Theorv of communication.” J. Inst. Elect. Ena.. Ecole Nationale Sup6rieure des Telecommuni-
-.
London, vol. 93, io. 3, pp. 429-457, 1946. cations, Paris, in 1978.
[9] R. W. Chang, “Synthesis of band-limited orthogonal signals for From 1979 to 1982 he was with the Socittt
multichannel data transmission,” Bell Syst. Tech. J., vol. 45, pp. Francaise de Production and Tkltdiffusion de
1775-1796, Dec. 1986. France, where he was involved with TV broad-
101 B. R. Saltzberg, “Performance of an efficient parallel data casting. From 1982 to 1988 he was with the
transmission system,” IEEE Trans. Commun. Tech., vol. COM- CCETT, where his work focused on satellite TV
15, pp. 805-811, Dec. 1967. broadcasting and digital audio broadcasting. He
111 B. Hirosaki, “An analysis of automatic equalizers for orthogonal has developed and promoted the COFDM approach as the modulation and
multiulexed OAM svstems.” IEEE Trans. Commun.. vol. COM- channel coding scheme for the DAB EUREKA 147 project, for which
28, J h . 1986 . he was issued several patents. In 1988 he joined Matra Communication,
121 -, “A maximum likelihood receiver for an orthogonallv and was engaged in the development of the GSM mobile communication
multiplexed QAM system,” Proc. IEEE Int. Cont on C&mun:, system, either on base station and mobile phones. In 1993 he co-founded
vol. 1, 1984. WAVECOM, Paris, of which he is now President. His current interest is the
[13] G. D. Fomey, “The Viterbi algorithm,” Proc. IEEE, vol. 61, development and licensing of new technologies for mobile communication.
pp. 268-278, Mar. 1973.
[141 J. B. Cain, G. C. Clark, and J. M. Geist, “Punctured con-
volutional codes of rate (n - l ) / n and simplified maximum
likelihood decoding,” IEEE Trans. Inform. Theory, vol. IT-25,
pp. 97-100, Jan. 1979. Claude Berrou (Member, IEEE) was born in
[ 151 P. Thitimajshima, “Les codes convolutifs rkcursifs Penmarc’h, France, in 1951. He received the
systkmatiques et leur concathation,” dissertation, I’Universit6 electrical engineering degree from the Institut
de Bretagne Occidentale, no. 284, Brest, France, 1993. National Polytechnique, Grenoble, France, in
[ 161 G. Ungerboeck, “Channel coding with multileveVphase sig- 1975.
nals,” IEEE Trans. Inform. Theory, vol. IT-28, pp. 55-67, Jan. In 1978 he joined the &ole Nationale
1982. Su@rieuredes Telecommunciations de Bretagne
[17] Y. Yasuda, K. Kashiki, and Y. Hirata, “High-rate punctured (France Telecom University), where he is cur-
convolutional codes for soft-decision Viterbi decoding,” IEEE rently a Professor of Electronic Engineering. His
Trans. Commun., vol. COM-32, Mar. 1984. research focuses on joint algorithms and VLSI
[18] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shan- implementations for digital communications,
non error-correcting coding and decoding: Turbo-codes,’’ Proc. especially error-correcting codecs and synchronization techniques.

996 PROCEEDINGS OF THE IEEE, VOL. 83, NO. 6, JUNE 1995

- ~- ___

Authorized licensed use limited to: UNIVERSITY OF STRATHCLYDE. Downloaded on November 20,2020 at 22:10:38 UTC from IEEE Xplore. Restrictions apply.