Explorer les Livres électroniques
Catégories
Explorer les Livres audio
Catégories
Explorer les Magazines
Catégories
Explorer les Documents
Catégories
ENST
Soutenue le 22 Octobre 1999 devant le jury composé de
Je voudrais remercier ici tous ceux qui m’ont aid et encourag pendant les trois annes
de thse. Tout d’abord j’exprime ma reconnaissance Monsieur Philippe Loubaton, Pro-
fesseur l’Universit de Marne la Valle, pour l’honneur qu’il m’a fait en acceptant de
prsider le jury.
Je remercie mes amis et collgues pour les motivations et les encouragements qu’il
m’ont donns. Je voudrais citer spcialement Nicolas, Karim, Amal, Ammar, Bahram,
v
REMERCIEMENT
Digitally signed by
vi ENST
Contents
Remerciement v
Résumé xvii
Notations xix
Acronyms xxi
Abstract xxiii
1 Work Context 1
1.1 Definitions and system model . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Overview and contribution of thesis . . . . . . . . . . . . . . . . . . . . 9
vii
CONTENTS
viii ENST
CONTENTS
BIBLIOGRAPHY 137
ENST ix
CONTENTS
x ENST
List of Figures
3.1 The evolution of the likelihood as a function of iterations, for the it-
erative detection of two uncoded users over the Gaussian channel: the
received signal is yt = A − B + νt . . . . . . . . . . . . . . . . . . . . . . 28
3.2 4 PAM non symmetrical modulation as a sum of two BPSK modulations. 28
3.3 Union upper bound on the PEP of two users jointly detected over the
Rayleigh fading channel: there is no ambiguity. . . . . . . . . . . . . . 29
3.4 Iterative decoding scheme of K = 2 users over the Gaussian channel. . 33
3.5 Multiple access of two users with convolutional codes at common rate
R = 12 . The code (4, 7) that has the word with all ones is compared to
the code (23, 33). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.6 Coded multi-user with the convolutional code of rate R = 14 and con-
straint length 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.7 Coded multi-user with the convolutional code of rate R = 18 and con-
straint length 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.8 Convolutional codes of common rate R = 14 , and constraint length 5
over the Rayleigh fading channel. . . . . . . . . . . . . . . . . . . . . . 37
3.9 Two users using the Golay(24,12) code at rate 12 . . . . . . . . . . . . . 41
3.10 Multi-block codes using the code BCH(31, 6) over the Gaussian channel. 41
xi
LIST OF FIGURES
xii ENST
LIST OF FIGURES
5.18 Average SER versus Eb/N0 . Θ2 with 16-QAM over quasi-static fading,
N = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.19 Average BER of ST codes Θ4,2 , ΘUDS,4 with BPSK modulation over
quasi-static fading, M = 4, N = 1, l = 2. . . . . . . . . . . . . . . . . . . 108
5.20 Average BER of the ST code Θ4,3 with BPSK modulation over quasi-
static fading, M = 4, N = 1, l = 3. . . . . . . . . . . . . . . . . . . . . . 108
5.21 Θ2 ST code with 4-QAM over a fast fading channel N = 1. . . . . . . . 113
5.22 Θ3,2 ST code with 4-QAM over a fast fading channel, N = 1. . . . . . . 113
5.23 Θ4,2 and Θ4,3 ST codes with BPSK over a fast fading channel, N = 1. . 114
5.24 Θ2 concatenated with outer substream codes. . . . . . . . . . . . . . . 114
5.25 Concatenated Θ2 with BCH(16, 5). N = 2, 4-QAM over quasi-static
fading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.26 Concatenated Θ2 with BCH(32, 6). N = 2, 4-QAM over quasi-static
fading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.27 Concatenated system over a pure fast fading channel. Outer code Golay(24, 12),
inner code Θ2 , N = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.28 Concatenated system over a realistic fading channel. Outer code Golay(24, 12),
inner code Θ2 , N = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
ENST xiii
LIST OF FIGURES
xiv ENST
List of Tables
5.1 First row of the optimal rotation matrices in dimensions 8 and 12. . . . 101
xv
LIST OF TABLES
xvi ENST
Résumé
Le premier point nous a conduit à proposer un schéma de décodage itératif dont les
performances s’approchent de celles de décodage à maximum de vraisemblance après
un petit nombre d’itérations. L’algorithme proposé est optimal dans le sens où tous les
interféreurs co-canal sont pris en compte à chaque itération. Nous avons testé le schéma
de décodage itératif sur des codes convolutifs et des codes en blocs, les avantages et
inconvénients de chaque système ont été montrés et expliqués. Nous avons proposé
une méthode efficace pour associer des valeurs souples aux décisions dures prise par
l’algorithme “Fossorier-Lin“ de décodage des codes en blocs.
Le second point a été traité sur deux types de canaux:
xvii
RÉSUMÉ
Pour le premier type de canaux, nous avons proposé une nouvelle classe des codes multi-
utilisateurs basée sur des constellations multi-dimentionnelles tournées. Les codes pro-
posés sont decodable d’une faon unique sur le canal additif non bruité, atteignant ainsi
la capacité de ce canal, puisqu’ils ne comportent pas de l’information redondante. Sur
un canal bruité, les performances des codes proposés surpassent presque toujours celles
de TDMA. La concaténation avec des codes extérieurs mono-utilisateur permet un
rehaussement des performances du système.
Pour le deuxième type de canaux, nous avons proposé une nouvelle classe des codes
espace-temps basée sur des constellations construites sur les corps des nombres algébriques.
Ces codes vérifient les critères de la construction des codes espace-temps sur des canaux
à évanouissement quasi-statique et rapide. Ici encore, les codes proposés n’ajoutent pas
de la redondance; ainsi, dans un système multi-antennes, nous avons pu construire des
codes qui rálisent une diversité maximale sur le canal à évanouissement, et qui ont une
grande efficacité spectrale.
La nouvelle classe des codes espace-temps peut être représentée comme un réseau de
points, ce qui permet d’appliquer l’algorithme de décodage sphérique. Cet algorithme
permet d’obtenir les performances de décodage à maximum de vraisemblance pour une
complexité polynomiale en la dimension du système multi-antennes, et indépendant de
la taille de constellation utilisée; donc on peut arriver à des efficacités spectrales très
élevée tout en gardant des bonnes performances et une complexité raisonnable. Le
décodage sphérique est limité en dimension réelle, néanmoins, pour un bon compro-
mis complexité/performances, on peut dépasser cette limite en combinant décodage
sphérique et élimination d’interférence par la combinaison optimale des différentes
versions du signal reu (maximum ratio combiner). Comme résultat supplémentaire,
on peut assigner des valeurs souples aux décisions dures prise par l’algorithme sous-
optimal.
Dans un environment multi-antennes, en combinant codage conjoint avec une con-
caténation des codes mono-utilisateur et des codes espace-temps aux émetteurs; et en
effectuant un décodage itératif aux récepteurs, nous avons montré que de très bonnes
performances peuvent être réalisées avec une complexité de décodage modérée, et une
grande efficacité spectrale. Ce qui fait de ce système un candidat prometteur pour la
troisième génération de radio-mobile (UMTS).
xviii ENST
Notations
Throughout this work the notations of the matrices and the vectors are set in boldface,
for example M , y are matrices or vectors depending on the context, while M and y
are scalars.
√
• i= −1.
• Pr{A|B} is the conditional probability of the event A knowing that the event B
has occured.
• y T is the transpose of y.
xix
NOTATIONS
xx ENST
Acronyms
ML Maximum Likelihood.
xxi
ACRONYMS
ST Space-Time.
UD Uniquely Decodable.
ZF Zero Forcing.
xxii ENST
Abstract
The push toward higher data rates and higher quality in wireless communications is
expected to continue. Thus a system designer should take into account:
• Low error probability at the receiver for both data and voice transmissions.
The first objective of this work is to simplify joint detection algorithms by the use
of iterative decoding methods. Hence the multi-user receiver becomes simpler to im-
plement, while one keeps the robustness and the good performance of ML-detection.
If the joint detection allows high quality of service at the receiver, it implies restrictions
on the data rates in order to make enough room for the multitude of users.
Joint multi-user coding permits an enhancement of the total amount of transmitted
information by taking into consideration the existence of other users on the transmission
medium at the transmitter side.
Joint multi-user coding in multiple-input multiple-output channels allows more en-
hancement in the quantity of transmitted information, due to the huge capacity of
these channels.
Representing the multi-antenna system by a lattice sphere packing and applying the
universal lattice decoder yields the performance of the ML detection at low complexity
independently of the spectral efficiency of the system. Hence, a very high spectral
efficiency is reached in a system that exploits the full diversity, which allows one to
appreciate the capacity of the multi-antenna environment.
Joint multi-user coding and decoding in a multi-antenna system makes the goal
of transmitting high data rates at low error probability and moderated complexity
feasible.
xxiii
ABSTRACT
xxiv ENST
Chapter 1
Work Context
The new elements added to the single user problem are: interference and coopera-
tion among users. We will explain in the sequel the system model, and the different
components in the transmission chain used in the following Chapters.
1
CHAPTER 1. WORK CONTEXT
2. One sender transmits to several receivers, like in the down link of the mobile-
radio system. In this scheme, the difference with the broadcast channel [24] is
that each user is interested in a different part of the received signal.
Source 1 Transmitter 1
Source 2 Transmitter 2
Estimated Data
Channel Receiver
Source K Transmitter K
Figure 1.1: Multi-user communication via the same medium to one receiver.
1. K Sources: Each one is modeled by a stochastic process, in this work only se-
quences of i.i.d (independent and identically distributed) binary random variables
are considered. This is the case of most band-limited sources (such as audio and
image sources after sampling at a frequency higher than twice times the band
of the source, and supposing that the source coding is done perfectly so that all
dependencies among the samples are omitted).
1
The distance inter-antenna should be at least half the wavelength in order to make the N channels
independent. This scenario is imaginable with mobile personal computers.
2 ENST
1.1. DEFINITIONS AND SYSTEM MODEL
2.1) Channel code (outer code): adds redundancy to the source stream to
protect it from errors caused by the channel and guarantee a reliable
information transmission. Although the first proposed codes were not
linear [83], the greatest amount of work in communication and informa-
tion theory is done over linear codes [33, 62, 44]. The linearity means
that the sum of two codewords belongs to the code.
There are two classes of linear codes:
(a) Linear block codes: the main class of linear block codes used in
the thesis is the class of binary BCH (Bose, Chaudhrui, and Hoc-
quenghem) codes.
(b) Convolutional codes.
2.2) Multi-user code (K-user code): a collect of single user codes, each one
takes into account the presence of the other codes over the transmission
medium, and has the role of protecting the concerned user from other
users interference. In concatenated systems, sometimes it is referred
to as inner code [38], since it is the nearest code to the channel in the
transmission chain.
2.3) Interleaver: a permutation of the given sequence X = (x1 , x2 , . . . , xn ),
that is used eventually when it is desired to introduce randomness
into correlated stochastic process (like time diversity over frequency-
selective Rayleigh fading channel, independent a posteriori probabil-
ities in iterative decoding schemes, etc). The inverse operation (de-
interleaver) is performed at the receiver. There exist different types of
interleavers:
(a) Matrix interleaver: X is read by the interleaver line by line then
it is output column by column.
(b) Quotient interleaver: let p be a prime number with n, let q%n
denotes the rest of the Euclidean division of q by n, then the in-
terleaver output is Y = (xp%n , x2p%n , . . . , xnp%n ), where the index
0 equals to the index n.
(c) Random interleaver: is simply a random ordering of the elements
of X, and is generated in simulations by choosing a seed S to
activate the random generator of the computer, then by choosing
ENST 3
CHAPTER 1. WORK CONTEXT
(b) Phase shift keying (PSK): the baseband presentation of this mod-
ulation is
2π √
M-PSK = {e(m·i M ) , m = 0, . . . , M − 1, i = −1}
0 1
BPSK
01
00 4-PSK 11
01 11
10
4-QAM
00 10
4 ENST
1.1. DEFINITIONS AND SYSTEM MODEL
3.1) The multiple access Gaussian channel (Fig. 1.4): one can consider both
real and complex channels depending on the used modulation. The real
Gaussian channel assigns to the real inputs x1 (t), x2 (t), . . . , xK (t) at the
instant t an output y(t) = x1 (t) + x2 (t) + . . . + xK (t) + w(t), where w(t)
is an additive white Gaussian noise (AWGN); white refers to the fact
that the random variables w(t), t ≥ 0 are independent, in opposition
to colored noise. Gaussian states that w(t) at each instant t follows
the Gaussian (Normal) probability law
1 z2 − (x−µ) 2
ENST 5
CHAPTER 1. WORK CONTEXT
channel
x1(t)
x2(t) y(t)
received
signal
xK(t)
transmitted
signals noise
w(t)
BS2
BS1
BSK
6 ENST
1.1. DEFINITIONS AND SYSTEM MODEL
Y = (w1 + iw2 ) · X + ν
= eiθ · α · X + ν,
ENST 7
CHAPTER 1. WORK CONTEXT
8 ENST
1.2. OVERVIEW AND CONTRIBUTION OF THESIS
is bounded by
PEP ≤ κ · (SNR)−L ,
where L is called the diversity order or gain, while the coding gain
is defined by the coding advantage over an uncoded system with the
same diversity order, and equals to κ−L .
For an overview of fading channels, diversity techniques and list of
references see [12].
4. Receiver: the receiver task is to extract an estimate of the desired user from the
received signal. This can be done by:
The receiver tolerance of the BER and delay transmission depends on the trans-
mitted information : a file containing computer data is tolerated to arrive with a
big delay, but its BER should be less than 10−5 . While an audio transmission in
the mobile-radio system could have a BER around 10−3 , but a very small delay.
ENST 9
CHAPTER 1. WORK CONTEXT
user decoding, joint multi-user coding, and space-time codes in a multi-antenna envi-
ronment.
Chapter 4 deals with enhancing the spectral efficiency of the multi-input single-
output channels (MISO), by joint multi-user coding. A construction of new codes by
rotated constellations is given for 2 . . . 7 users.
Over the noiseless channel, the proposed K-user code profits from the algebraic struc-
ture of the considered number field to separate K sources by one sensor. Achieving
thus the capacity of the noiseless adder channel. A soft-input hard-output decoding
algorithm is then proposed. Over the Gaussian channel, the new class of codes is
compared to TDMA. Studying the different parameters in the proposed code shows
their influences in the obtained gain compared to TDMA. Then we present the system
obtained by concatenation with single user codes, where the special case of very low
rate single user codes is studied.
10 ENST
1.2. OVERVIEW AND CONTRIBUTION OF THESIS
Contributions:
• An iterative decoding method for a multi-user coded system, that achieves the
single user performance.
• Assigning soft values to the hard decisions obtained by the Fossorier-Lin decoding
algorithm.
ENST 11
CHAPTER 1. WORK CONTEXT
12 ENST
Chapter 2
I nformation theory gives the upper limits of the data transmission rate over a given
channel; and the lower limits of the data compression rate for a given source [24, 83]
(Fig. 2.1).
Data compression Data transmission
limit limit
1
0
0
1 1
0
0
1
min I(X;Xs) max I(X;Xc)
It also suggests means of achieving these limits. However, due to the decoding
complexity of the proposed random codes which achieve these limits when the code
length goes to infinity, information theory had opened a new area of research to find well
structured codes that approach the upper limits with a reasonable decoding complexity
[33, 62, 44]. It was amazing to find that, almost 50 years after Shannon proposed his
capacity formulas and the random codes achieving it, the turbo codes [11, 79, 9], which
are the closest known codes to the capcity limits with reasonable decoding complexity,
have a random code likewise weight distribution [60].
In the sequel, we present the information theoretical viewpoint of the multiple ac-
cess problem, which should illuminate our approach to cancel the interference and to
13
CHAPTER 2. INFORMATION THEORY AND STATE OF THE ART
perform cooperation among users. Then we give the state of the art of the three areas
treated in the thesis: joint multi-user decoding, joint multi-user coding, and space-time
codes in a multi-antenna environment.
where Rj , j = 1 . . . K is the rate of the user j. They also showed that the maximum
achievable sum rate Csum is asymptotically equal to 12 log2 (πeK/2) (Fig. 2.2).
It is worth to notice that the capacity ≈ 12 log2 (πeK/2) is proportional to log2 the
size of the alphabet obtained by the sum of all users [92]. So one can increase the
capacity by two ways
2. Transmitting at different powers, so that the size of the sum of signals increases:
Example 2.1 {−A, +A}+{−A, +A} = {−2A, 0, +2A} the alphabet sum has
a size equals to 3. While for A = B, {−A, +A} + {−B, +B} = {−A − B, −A +
B, +A − B, +A + B} it has a size equals to 4
14 ENST
2.1. INFORMATION THEORY VIEWPOINT
2.8
2.6
2.4
2.2
per channel use
1.8
sum
C
1.6
1.4
1.2 Asymptotical C
sum
exact C
sum
1
1 2 3 4 5 6 7 8 9 10
number of users
Figure 2.2: Maximum achievable rate in the K-user binary adder channel (BAC).
where Ri is the rate of user i, and C is the channel capacity of a single user as a
function of signal-to-noise ratio SNR :
1 P
C(P/σ 2 ) = log2 (1 + 2 ) [bits per dimension]. (2-3)
2 σ
Note that when all users have the same rate the previous inequalities become [24]
K
K
1 Pi
Ri ≤ log2 1 + i=1
, (2-4)
i=1 2 σ2
which is the bisecting line in the above region. This last equation shows the surprising
fact that the sum rate can be as large as the rate achieved by a single user who sends
with a power equal to the sum of the powers. In practical systems without multi-user
detection, the achievable rates fall far behind the capacity.
In Fig. 2.3 we plotted the capacity region of two users. The vertices A and B are
known to be achievable by “successive interference cancellation” (SIC) [10] as follows:
user one considers user two as noise, so he can achieve a rate of C(P1 /(σ 2 + P2 )) error
free. So after decoding the first user, the receiver can subtract its contribution from
the received signal, a procedure that allows the second user to transmit at C(P2 /σ 2 )
error free.
The application of this rule in real situations encounters two problems:
ENST 15
CHAPTER 2. INFORMATION THEORY AND STATE OF THE ART
R2
A
C(P2/(No/2)) C
C(P2/(P1+No/2)) B
R1
D
C(P1/(P2+No/2)) C(P1/(No/2))
• The second user is not a Gaussian noise, to apply the central limit theorem the
number of active users should be ≥ 10.
• There is no error free single user code that achieves the capacity of the single user
channel; even the well known turbo codes [11] are 0.7 dB far from the Shannon
capacity (2-3), and due to the big slope of the BER curve as a function of Eb/N0 ,
this class of codes is very sensitive to the changes of the noise variance caused by
the other user.
1. Time division multiple access (TDMA): the frame duration is divided into non-
overlapping slots, each slot is assigned to a different user, and there is a slot
reserved for signaling. Each user can use the entire frequency bandwidth during
his time slot. TDMA cellular includes North American Digital Cellular, Global
System for Mobile Communications (GSM) [65], and Personal Digital Cellular
(PDC).
16 ENST
2.1. INFORMATION THEORY VIEWPOINT
3. Code division multiple access (CDMA): assigns to each user a unique code se-
quence that distinguishes him from the multitude of active users simultaneously
transmitting through the entire bandwidth. It is referred to as direct-sequence
CDMA (DS-CDMA) [99]. CDMA has been used in military satellite commu-
nication system, after it has gone commercial [61], [82], the Telecommunication
Industry Association (TIA) has adopted it, the standard IS-95 [2] was proposed
to replace IS-54 TDMA standard. CDMA technology has been proposed for the
third generation of the mobile-radio standard UMTS (Universal Mobile Telephone
System) [22].
These methods are not exclusive, one can combine FDMA/CDMA to obtain Frequency-
Hopping CDMA. There are other strategies of multiple access such as space division
multiple access (SDMA). Rimoldi and Urbank [78], proposed a rate splitting scheme,
where all users but one split their code rates into two rates depending on the total
allocated power, such that the new set of rates forms a vertex in the capacity region of
a higher dimension space, and a vertex is known to be achievable via single user coding
and successive cancellation.
Theoretical potential
• TDMA (FDMA): because of the duality time/frequency all the results on TDMA
are applied to FDMA.
Let αj denote the slot duration of user j. Pj , P̂j its average and peak power
respectively1 . In a normalized system we have Kj=1 αj = 1. The capacity region
of this system is given by
Rj ≤ αj C(P/σ 2 ), ∀j = 1 . . . K, (2-5)
1
In order to compare different multiple access systems, the total transmitted power is fixed in all
of them. In the case of TDMA(FDMA) the transmitted power can not exceed a certain value due to
physical limitations when the number of users increases.
ENST 17
CHAPTER 2. INFORMATION THEORY AND STATE OF THE ART
2. Power limitation.
When Pˆj = ∞, the sum rate of TDMA can reach any point on the bisecting line
(2-4). The other extreme case, which is commonly considered as time sharing
in the communication community2 , is when P̂j = Pj . Which, in the case of two
users, is given by the region under the dashed line connecting C and D in Fig.
2.3: C presents the situation where user 2 occupies the whole period of time, and
D is achieved by reversing the role of user 1 and 2.
Note that in TDMA (FDMA) there is no room for cooperation among users
at the receiver side, the cooperation can be done only at the transmitter by
choosing rates proportional to the allocated power, which is very difficult to
realize physically.
• CDMA, non cooperative: when each user considers the other active users on the
channel as white noise, and is independently demodulated and decoded, his rate
is bounded by
Pj
Rj ≤ C . (2-6)
σ2 + k=j Pk
1
Let the number of users K −→ ∞, then the sum of the rates is bounded by 2 log 2
=
0.721 bits/s. This limit renders the CDMA system slightly less efficient than
TDMA(FDMA), but the big advantage of CDMA is that there is no dedicated
channels, which makes cooperative decoding schemes possible at the receiver.
• CDMA cooperative [98]: the cooperation implies that the receiver knows the code
structure of the other users so he can decode each and every one and cancel their
contribution from the received signal, which incidentally increases the complex-
ity. The cooperative scheme follows the same principle of successive cancellation
to achieve the vertices in the capacity region
(Fig. 2.3):
the receiver decodes
user 1, which transmits at a rate R1 ≤ C σ2 +P1 P , then user one is coded
k>1 k
and modulated then subtracted
from the
received signal which allows user 2 to
transmit at a rate R2 ≤ C σ2 +2 P , and so on. The sum rate of this scheme
P
k>2 k
could equal that of TDMA without power constraint (2-4), with the difference
that no synchronization among users is needed and there is no limitation on peak
power since all users transmit during the whole period.
2
The comparisons done in Chapter 4 are based on this definition of TDMA.
18 ENST
2.1. INFORMATION THEORY VIEWPOINT
RX
TX RX
TX RX
Transmitter
Receiver
TX RX
TX RX
RX
The physical limitation of the wireless channel presents the underlying challenge be-
hind the use of multi-antenna array. Consider the system with M transmit and N
receive antennas (Fig. 2.4). Telatar in [91], obtained capacity and error exponent [33]
expressions for multiple transmit-receive antenna system in the presence of Gaussian
noise, where he made the assumption of independent fading at each channel use. Fos-
chini and Gans [30] derived the expression of the outage capacity under the assumption
that fading is quasi-static. A particular layered space-time architecture [39] was shown
to have the potential to achieve a substantial fraction of capacity. The main conclusion
of their work is that the capacity of a multi-antenna system exceeds by far the single-
antenna system and grows at least linearly with the number of transmit antennas M
as long as the number of receive antennas N ≥ M.
For capacity computation purpose, the main difference between flat and quasi-static
fading is that the fading seen as stochastic process in time is no more ergodic in the
latter case [91], which implies that the channel capacity does not equal the maximum
of mutual information (see [24] Chapter 8 and [70]). This impairment of the channel
leads to a new definition of capacity more suited for the wireless channel. Let H be
the N × M transfer matrix of the fading channel.
ENST 19
CHAPTER 2. INFORMATION THEORY AND STATE OF THE ART
Definition 2.1 Outage capacity: when the Shannon capacity in the sense of maxi-
mum rate error free is equal to zero no matter how long the code is, the outage capacity
is proposed as a trade off between supportable rates and outage probability. Namely,
given a rate R, and a power constraint P , the outage probability Pout (R, P ) is defined
as follows: for any rate r ≤ R and ∀ > 0, there exists a code satisfying the power
constraint P , for which the error probability Pe ≤ for all but a set of quasi-static
fading H whose total probability is less than Pout .
Let ρ be the average signal to noise ratio per receive antenna (independent of M).
Then for each realization of H the capacity is given by [31]:
ρ †
C = log2 det I N + · HH . (2-7)
M
This huge increase is because in the fast fading case, the system has infinite time
diversity.
1. M = N = 1: no diversity
C = log2 1 + ρ · χ22 . (2-8)
20 ENST
2.1. INFORMATION THEORY VIEWPOINT
M
M
ρ ρ
log2 1 + · χ22(N −i+1) < C ≤ log2 1 + · χ22Ni , (2-11)
i=1 M i=1 M
Note that:
Remark 2.1 For the capacity computation, it is always assumed that the transmit-
ter knows the distribution law of the channel, and the receiver knows the CSI. When
assuming that the transmitter knows the CSI too, it can be proved that the capacity en-
hancement obtained by transmit diversity is equivalent to that of receive diversity [19].
Furthermore, the multiple-input multiple-output channel (MIMO) can be transformed
into an equivalent system of min(M, N) single-input single-output channels (SISO) by
using linear operations at both the receiver and the transmitter.
ENST 21
CHAPTER 2. INFORMATION THEORY AND STATE OF THE ART
22 ENST
2.2. STATE OF THE ART
channels is provided.
Our investigation on the K-user codes for the real adder channel is motivated by
concerns of simplicity and symmetrical construction for each user code. We believe
that construction of multi-user codes on multi-dimensional modulations by rotations
is first proposed in this dissertation.
ENST 23
CHAPTER 2. INFORMATION THEORY AND STATE OF THE ART
Our approach using rotated constellations was motivated by exploiting the transmis-
sion diversity without loss in spectral efficiency. It consists in forming a matrix of a
rotated multi-dimensional constellation over space-time which verifies the criterion in
[90].
24 ENST
Chapter 3
I n a pioneering paper [96] in 1986 Verdú has established the theoretical limits of
the maximum likelihood (ML) multi-user detector over the Gaussian multiple ac-
cess channel. It is well known that the conventional detector suffers from multiple
access interference (MAI), and the near-far effect. Multi-user detection copes with
both problems and allows the multitude of users to reach the single user performance.
The inconvenience of the ML multi-user detection is its huge complexity.
Iterative decoding schemes that we apply, look very attractive to achieve the same task
as an ML detector at high SNR, with lower complexity [84, 63].
Consider a system of many senders S1 , . . . , SK and one received signal r being the
sum of the K transmitted signals, attenuated or not attenuated by independent fadings,
and disturbed by an AWGN. The receiver is assumed to know the power of all the users,
and the channel parameters. By using this knowledge, the desired information of Sk
has to be extracted. In the sequel, we assume that the BPSK modulation is used as
well as a perfect synchronization among all the users.
The Chapter is organized as follows: in Section 3.1 we study the uncoded system
of two users over the Gaussian and the Rayleigh fading channels. An iterative detec-
tion scheme is proposed and studied for the uncoded case. In Section 3.2, an iterative
decoding method is proposed and tested on convolutional coded systems over the Gaus-
sian and the Rayleigh fading channels. Section 3.3 deals with block coded systems and
shows the advantages/disadvantages of block codes compared to convolutional codes in
the proposed iterative decoding method. General observations, and rules of thumb are
given in Section 3.4; then elements of analysis of the block coded two users are detailed
in Section 3.5. Before concluding the Chapter, we give in Section 3.6 a comparison of
25
CHAPTER 3. JOINT MULTI-USER DECODING
where at , bt ∈ {−1, +1} are the transmitted symbols of user 1 and user 2, and νt is an
AWGN with variance σ 2 = N20 .
Definition 3.1 Ambiguity, and p-ambiguity: in a multi-user system using the BPSK
modulation over the Gaussian channel, an ambiguity occurs when the sum of two or
more users is equal to zero. The ambiguity in a coded multi-user system means that the
multi-user code, which is the collection of different single user codes, is not uniquely
decodable (UD) [20]. In a coded multi-user system, we say that a p-ambiguity occurs
when there are ambiguities over p positions in the codeword.
In an uncoded system over the Gaussian channel, the ML detector can not reveal
the ambiguity over the different users, nevertheless, at high SNR, it can separate them
when they transmit at different power. In the sequel, we develop an iterative detection
scheme for two uncoded users.
The log likelihood value Lv(·) [43] of the random variable at is defined by
Pr{at = −1|yt }
Lv(at ) = log . (3-2)
Pr{at = +1|yt}
Assuming that at , bt in (3-1) are independent, the following relations between the two
likelihoods were derived (see appendix 3-A).
2Ayt Lv(bt ) AB
Lv(at ) = − 2 − 2arc tanh tanh tanh
σ 2 σ2
Lv(bt )
1 + tanh( σA2 yt ) · tanh( AB
σ2
) · tanh( AB
σ2
− 2
)
+ log Lv(bt )
. (3-3)
1 − tanh( σA2 yt ) · tanh( AB
σ2
) · tanh( AB
σ2
+ 2
)
26 ENST
3.1. UNCODED TWO USERS
2Byt Lv(at ) AB
Lv(bt ) = − 2 − 2arc tanh tanh tanh
σ 2 σ2
Lv(at )
1 + tanh( σB2 yt ) · tanh( AB
σ2
) · tanh( AB
σ2
− 2
)
+ log Lv(at )
. (3-4)
1 − tanh( σA2 yt ) · tanh( AB
σ2
) · tanh( AB
σ2
+ 2
)
The first term in (3-3) and (3-4) represents the situation when the user is alone. For
different signal-to-noise ratios between the two users, applying the previous relations
is equivalent to subtracting the most powerful user from the received signal. With
SNR of the same order for both users, the third term in (3-3) and (3-4) shows the
ambiguity
over
the two users, which is the worst case in an iterative decoding scheme.
If tanh σ2 ≈ 1 then for A B, the third term in (3-3) and (3-4) equals to the
AB
opposite of the first one, thus after the initialization of the likelihoods to zero, they
stick to this value during the iterations as shown in Fig. 3.1 (the dotted line).
Fig. 3.1 presents the evolution of the log likelihoods of the two users as a function
of the number of iterations for a given realization of the AWGN. It is shown that for
a big difference of SNR (5 dB) the iterative scheme separates the two users, while for
SNR of the same order the iterative scheme fails to reveal the ambiguity over the two
users.
Definition 3.2 The pairwise error probability is the probability of decoding the word
E knowing that the word C was transmitted, when these two words are considered to
be the only words of the code.
One expects no ambiguity over the Rayleigh fading channel, because if the two users
are affected by two independent fades, then the probability that the two fadings are
equal is zero. In the following the PEP is derived.
ENST 27
CHAPTER 3. JOINT MULTI-USER DECODING
1
Lv(bt ),
) Eb2=0 dB
The loglikelihood : log(Pr(bit=0|observation)/Pr(bit=1|observation))
0
Lv(bt ), Lv(at ) for A=B
−1
−2
−3
−4
−5
Lv(a ),
) Eb1=5 dB
t
−6
0 5 10 15 20 25 30
Number of iterations
Figure 3.1: The evolution of the likelihood as a function of iterations, for the iterative
detection of two uncoded users over the Gaussian channel: the received signal is yt =
A − B + νt .
where αt , βt are two independent Rayleigh fadings of variance σA2 , σB2 respectively. For
the user with amplitude A, the conditional error probability knowing that −A is trans-
mitted is equal to the conditional error probability knowing that +A is transmitted,
and is given by
1
Pr{e| − A} ≤ (Pr{e| − A − B} + Pr{e| − A + B}) . (3-6)
2
−2A 0 2A
× × × A=B
−A − B −A + B A−B A+B
× × × × A>B
Figure 3.2: 4 PAM non symmetrical modulation as a sum of two BPSK modulations.
Considering the 4 PAM non symmetrical modulation {−A−B, −A+B, A−B, A+B}
28 ENST
3.1. UNCODED TWO USERS
No No No No
0
10
−− : The single user union bound
−1
10
Union Bound on the PEP
−2
10
−3
10
0 2 4 6 8 10 12 14 16 18 20
Eb/No in dB, the two users transmit at the same power
Figure 3.3: Union upper bound on the PEP of two users jointly detected over the
Rayleigh fading channel: there is no ambiguity.
We present in Fig. 3.3 a union upper bound on the error probability for a joint
detection of the two users over a the Rayleigh fading channel. There is a difference of
3 dB between the two users bound and the single user one, but there is no ambiguity.
Although iterative detection using the joint probability of an uncoded two users only
improves the performance of the system slightly, it allows one to understand the main
ENST 29
CHAPTER 3. JOINT MULTI-USER DECODING
problem of such a scheme: the p-ambiguity is resolved below by using error control
codes with interleavers.
K
rt = Ak · akt + νt , (3-9)
k=1
Our decoding algorithm is an iterative SIC type using the soft-input soft-output Bahl
et al. MAP decoding algorithm at each stage [5].
We do not derive here the algorithm parameters [5]. We just give their expressions and
note where the modification was done. Let rt = (rt,1 , . . . , rt,n ) be the received signal at
time t, with n being the number of outputs of the convolutional code1 ( n1 is its rate).
r t1 denotes the received sequence from instant 1 to instant t; the algorithm computes
the following parameters
αt (m) = Pr{St = m; r t1 } = αt−1 (m ).γt (m , m) Forward (3-10)
m
βt (m) = Pr{r τt+1 |St = m} = βt+1 (m ).γt+1 (m, m ) Backward (3-11)
m
γt (m , m) = Pr{St = m; r t |St−1 = m }
= pt (m|m )qt (X|m , m)R(r t , X) Branch-metric, (3-12)
X
where St is the trellis state at time t, τ is the length of the received sequence r.
pt (m|m ) is the transition probability in the trellis from the state m to the state m.
1
The different outputs are transmitted serially in time, but for computation purpose they are noted
in a parallel way. In the sequel, when there is no confusion, we note rt,j by rj .
30 ENST
3.2. CONVOLUTIONAL CODED SYSTEMS
We suppose that when decoding the user i, the receiver wants to take into consideration
the available information on the other active users on the channel. Assuming that the
K transmitters are independent, then the channel transition probability of the user i,
at the position j of the received signal, is given by
p{rj |aij } = p{rj |a1j , . . . , aK 1
j } · Pr{aj } · · · Pr{aj }.
K
(3-14)
akj ∈{−Ak ,+Ak }, k=1...K, k=i
When the receiver knows the code structure of all active users, like a base station in
the up link mobile-radio communication, it can profit from this knowledge to weight the
contributions of the other users in (3-14) when decoding the user i. The conventional
detector computes p{rj |aij } without taking into accounts the other users. It considers
them as an AWGN, which is equivalent to use the relation (3-14) with Pr{akj } =
0.5, k = 1 . . . K, k = i.
If we suppose that the receiver knows (by a genius aid!) that the user k has his jth
symbol equal to −Ak , so it is well known that subtracting the contribution of the kth
user from the received signal at the instant j improves the performance. Doing so is
equivalent to put Pr{akj = −Ak } = 1 and Pr{akj = +Ak } = 0 in (3-14). On the other
hand, if the receiver only knows that the probability of akj = −Ak is greater than the
probability of akj = +Ak , then it is wiser to give the metrics containing akj = −Ak more
weight than those containing akj = +Ak .
In general, the decoder performs better when a non-uniform distribution of the a priori
probabilities is available at its input, because it uses the a priori as a second “source”
of beforehand information about the received sequence; the first source being the code
structure [9].
2
If there are no parallel branches then it will be only one output for each transition.
ENST 31
CHAPTER 3. JOINT MULTI-USER DECODING
Remark 3.1 We call the proposed iterative scheme a “SIC type” algorithm, but
improperly. While no subtraction of the decoded users from the received signal is per-
formed, the way the algorithm weights the kth user contribution to the received signal
is equivalent to subtract it when its a posteriori probabilities are relevant.
To clarify our multi-user decoder, we take K = 2. The receiver performs the following
recursion (Fig. 3.4):
3
In the first iteration, when Pr{akj } = 0.5, k = k0 . . . K it is not necessary to sum over these values
of k in (3-14).
32 ENST
3.2. CONVOLUTIONAL CODED SYSTEMS
Interleaver 2
Interleaver 1
Data 2
Coder 2 Interleaver 2 Mapper
AWGN
Decoder 2 Deinterleaver 2
a posteriori 2
Figure 3.4: Iterative decoding scheme of K = 2 users over the Gaussian channel.
1. De-interleave the received signal r τ1 , and the a priori of the second user with
respect to the first interleaver.
2. Decode the first user by the Bahl et al. MAP decoding algorithm using the
channel transition probability
3. Replace the a priori of the first user by the a posteriori computed by the decoding
algorithm. Then interleave the received signal and the new a priori of the first
user with respect to the first interleaver.
We present in Fig. 3.5 the performance of an iterative decoding of two users using the
GSM [1] convolutional code (23, 33) of rate 1/2, and the systematic convolutional code
with polynomial generators (4, 7), which represents a special case where a convolutional
code of rate 1/2 has the word (1 . . . 1) (see Section 3.4). The SNR of the second user
is fixed to 5 dB for both codes, while the SNR of the first user varies from 0 to 5 dB.
It is shown that for the GSM code, the BER of the first user in the multiple access
scheme is close to the single user case starting from SNR1 = 0 dB, and SNR2 = 5 dB.
Fig. 3.6 presents the average BER of two and three users versus the common SNR
of all users. In simulations over the Gaussian channel, we counted at least 100 errors
ENST 33
CHAPTER 3. JOINT MULTI-USER DECODING
0
10
code(4,7), user 1, Eb2/No = 5dB
code(4,7), user 1 alone
code(4,7), user 2, Eb2/No = 5dB
-1
code(23,33), user 1 alone
10 code(23,33), user 2, Eb2/No = 5dB
code(23,33), user 1, Eb2/No = 5dB
-2
10
BER
-3
10
-4
10
-5
10
0 1 2 3 4 5
Eb1/No in dB
Figure 3.5: Multiple access of two users with convolutional codes at common rate
R = 12 . The code (4, 7) that has the word with all ones is compared to the code
(23, 33).
34 ENST
3.2. CONVOLUTIONAL CODED SYSTEMS
0 0
10 10
User alone
Two users Eb1=Eb2
Three users, Eb1=Eb2=Eb3
User alone
-1 -1 Four users,Eb1=Eb2=Eb3=Eb4
10 10
Average BER
Average BER
-2 -2
10 10
-3 -3
10 10
-4 -4
10 10
0 1 2 3 4 1.0 2.0 3.0 4.0
Eb/No Eb/No
Figure 3.6: Coded multi-user with the Figure 3.7: Coded multi-user with the
convolutional code of rate R = 14 and con- convolutional code of rate R = 18 and con-
straint length 5. straint length 5.
ENST 35
CHAPTER 3. JOINT MULTI-USER DECODING
for each user as a stop criterion. The computed BER’s of all users were very close, so
one can consider the average BER as the BER of each user. We present in Fig. 3.7 an
iterative decoding of four users transmitting at the same power, using convolutional
code of rate 1/8 and memory 4.
In all the simulations, the interleaver size is fixed to 1000 bits at the input of the
convolutional code (which makes it n×1000 symbols at the output). The convolutional
codes of a given rate and a given constraint length were chosen to have maximum free
distance [76].
We notice that when the number of users increases, the iterative scheme without
interleavers shows a severe degradation which we interpreted from the phenomenon
where the erroneous paths resulted from one decoder are passed entirely to the next
one. So the error paths permeate among users and bound the performance.
We present in Fig. 3.8 an iterative decoding of two and three users with or without
interleavers over the Rayleigh fading channel. All users use the same convolutional
code of rate 1/4 and memory 4. All users are transmitting at the same power, and the
variances of the different fadings are normalized to 1.
The class of block codes used in the sequel is the binary BCH codes.
36 ENST
3.3. BLOCK CODED SYSTEMS OVER THE GAUSSIAN CHANNEL
0
10
-1
10
Average BER
-2
10
-3 User alone
10
Two users with interleavers
Three users with interleavers
Two users without interleavers
Three users without interleavers
-4
10
0 1 2 3 4 5
Eb/No in dB
Figure 3.8: Convolutional codes of common rate R = 14 , and constraint length 5 over
the Rayleigh fading channel.
ENST 37
CHAPTER 3. JOINT MULTI-USER DECODING
x = uG. (3-16)
r = y + ν = (r1 , . . . , rn ), (3-17)
38 ENST
3.3. BLOCK CODED SYSTEMS OVER THE GAUSSIAN CHANNEL
This algorithm is substantially less complex than the block decoding algorithm based
on the trellis structure of block codes [32]. In addition, it approaches the ML decoding
performance for relatively small values of the reprocessing order l depending on the
code dimension k, and the considered BER.
π(l + 1) PB , (3-19)
where π(l + 1) is the probability that there are l + 1 or more errors in the first k posi-
tions of the decoded word (before reprocessing), and PB is the block error probability
of maximum likelihood decoding. The inequality (3-19) guarantees a sufficient condi-
tion for order-l reprocessing to provide practically optimum error performance at the
considered BER.
ENST 39
CHAPTER 3. JOINT MULTI-USER DECODING
We have checked by simulation that, for k ≤ 16, a second order reprocessing provides
soft values very close to those obtained by exhaustive decoding. For these values of
k first order reprocessing in hard-output decoding is near optimum, while to obtain
near exhaustive decoding soft values, second order reprocessing is needed. However,
simulations show that iterative decoding converges to its best performance for soft
values obtained by first order reprocessing even though they are a bit different from
those obtained by exhaustive decoding.
Computing soft values in this way, induces that the computation of the metric should
be exponential, since we sum over all the codewords treated in the reprocessing stage.
While for the algorithm without soft decisions, the computed metric is only the Eu-
clidean distance. One can remedy this computations increase by choosing for the
0 1
likelihood (3-18) only the closest codewords to Cmin , Cmin among the word treated
within the reprocessing stage. Doing so, one computes only the Euclidean distance of
0 1
Cmin , and Cmin to the received sequence, because taking the logarithm of the ratio of
two exponentials is equal to the subtraction of the two exponents. This last way of
assigning soft values to the decoding algorithm gives small degradation compared to
summing over all the codewords treated by the algorithm, at high SNR.
Computation cost: Assigning soft values will add lj=1 kj comparisons and addi-
tions operations plus k divisions and storages, to the computations cost in [32].
Approximately, for short codes n ≤ 32, or medium codes 32 < n ≤ 64 with rate
R ≥ 0.6, the computation cost for near optimum hard decisions decoding is obtained
after second order reprocessing, and its computation cost is o(k 2 ) [32]. While for longer
codes higher reprocessing order is required to achieve optimum performance, most of
the coding gain is achieved within the first two reprocessing stages.
Compared to the Chase algorithm [21, 77], this algorithm presents the advantage of
less complexity and always having codewords with bit j equal to zero or one, j = 1 . . . k,
among the codewords treated by the algorithm.
In all the simulations first order reprocessing was used. We present in Fig. 3.9 the
two users multiple access scheme at the common rate of 1/2. It is shown that for block
40 ENST
3.3. BLOCK CODED SYSTEMS OVER THE GAUSSIAN CHANNEL
0
10
0 10
user alone
User alone
two users, Eb2/No = 5dB
10
-1 Two users
-1
10 Three users
Four users
BER of user 1
Average BER
-2
10
-2
10
-3
10
-3
10
-4
10
-4
-5 10
10 0 1 2 3 4 5
0 1 2 3 4 5
Eb/No , All users transmit the same power
Eb1/No in dB
ENST 41
CHAPTER 3. JOINT MULTI-USER DECODING
code at the rate 1/2 the ambiguity still exists after coding and interleaving. There is
a difference of 1.2 dB between the two users case and the single user case. We present
in Fig. 3.10 an iterative decoding of two, three, and four users using the block code
BCH(31, 6). It is shown that the rate 6/31, in the block coding case, is not small
enough to separate completely four users over the Gaussian channel, while it does for
three and two users.
• One observes that the BER in a multi-user scheme is close to the single user BER
only after some threshold EbT (e.g 1.5 dB for two users and 2 dB for three users in
Fig. 3.6). One can interpret this as follows: in the first iteration user 1 considers
the others as noise, so his SNR is given by No +2R×Eb1 . This results in a BER
Ebi i≥2
Pe1 which should be small enough to allow the second user to obtain “reliable”
information about the first one, so that the last one could be subtracted6 from
the received signal.
• One observes that Rsum = K · R ≈ const. [98], which is due to the fundamental
relationship between power and bandwidth efficiency on the AWGN channel. For
a small number of users using short codes, the above relationship between power
and bandwidth efficiency could be interpreted as follows: for a given SNR, the
number of ambiguity situations increases combinatorially with K; consider the
scheme of three users, where all users are transmitting at the same power. If the
received signal is A + ν, where
A is the common amplitude of the used BPSK
3
modulation, then there are 2 = 3 possible situations of ambiguity: (1, 1, −1) ,
(1, −1, 1), or (−1, 1, 1).
6
In a soft way: since the exponent in (3-14), that subtracts the correct value of the first user from
the received signal, is multiplied by a higher value than that subtracting the wrong value.
42 ENST
3.4. PERFORMANCE ANALYSIS: GENERAL OBSERVATIONS AND RULES-OF-THUMB
Simulation results have shown that common rates of 12 , or 13 with memory greater
than 3 are sufficient to obtain a BER close to the single-user scheme for two users
multiple access scheme. While for the same memory, a common rate of 14 is needed
for three users, and R ≤ 15 for four users.
• In the simulations, only a small number of iterations was required to reach the
stable performance (it varies from 10 iterations for low SNR (EbT dB) to 3 for
high SNR (EbT +2 dB)). It is shown that in opposition with the “turbo decoding”
scheme of a single user [9], for high SNR, the number of necessary iterations
decreases. This is due to the used metric (3-14), which, for high SNR, is equivalent
to subtracting the most powerful user from the received signal when decoding the
next users.
• Practical convolutional codes (Fig. 3.5, code(23,33)) are better than block codes
(see Fig. 3.9) at the rate 12 because they do not have the word with all ones
(1 . . . 1), which implies that every codeword Ci has its inverse C̄i = (1 . . . 1) +
Ci in the code, which increases the ambiguity cases7 . It seems that for code
rates smaller than 12 , the existence of the word with all ones does not affect the
performance of the iterative scheme. This is explained by:
• Interleavers : the results of subsection 3.2.2 allow us to conjecture that the role
of independent interleavers, associated with the same code output, is to make
the different users as orthogonal as possible. The orthogonality is guaranteed by
two points:
– Minimize the number of p-ambiguity cases over the Gaussian channel. Note
that when p is greater than the correction capacity of the code [60], the ML
decoding fails to reveal the ambiguity over the users even over a noiseless
adder channel. Thus the interleaver role is to make p as small as possible
over every codeword.
7
All the block codes used in the simulations have the word (1 . . . 1), and in general all the practical
known cyclic block codes have this word (ex. BCH, RM, QR) [60], since their generator polynomial
is not divisible by 1 + X, while the practical convolutional codes do not have it: an example of
convolutional code that has this word was given in Fig. 3.5.
ENST 43
CHAPTER 3. JOINT MULTI-USER DECODING
Consider the received signal (3-1), with a normalized noise variance 1/2. The prob-
ability distribution of aj , bj , during the semi-iteration l, are denoted by plj , qjl . The
performance analysis consists in first computing a union upper bound of the error
probability in subsection 3.5.1, and then expressing a relationship between the two
probability distributions during the semi-iteration l + 1 in subsection 3.5.2.
Taking the logarithm and excluding the common term from the previous metric, the
following quantity should be maximized
n
n
Lv l (bj ) − 4Brj
CMi =2 rj (2cij −1)A+ log cosh 2AB(2cij − 1) + , (3-21)
j=1 j=1 2
ql
where Lv l (bj ) = log 1−qj l . By using the central limit theorem, it can be shown that
j
CMi is a Gaussian random variable. By using the linearity of the code, it can be
assumed that the all zero word is transmitted. The PEP for a codeword Ci is equal to
Pr{CMi ≥ CM1 }.
In the first semi-iteration, one initializes Lv(bj )1 = 0, ∀ j. By doing algebraic manipu-
lations, it can be shown (see appendix 3-C) that after the first semi-iteration, the PEP
is bounded by
1 wi (A − B)2
Pr{CMi > CM1 } ≤ exp − ,
2 2n(2A2 B 2 + A2 + B 2 ) + 2n(1 − 2n)(2B 2 − log 2)2
(3-22)
44 ENST
3.5. ELEMENTS OF ANALYSIS OF BLOCK CODED 2 USERS OVER THE AWGN CHANNEL
Note that over the symbols where qj ≈ 1 − qj , we have θj ≈ −4AB, which, in a scheme
where both users have the same order of SNR A ≈ B, is subtracted from 4A2 . Over
the other symbols where the decoding was done with good reliability, the corrective
term Lv l (bj ) is added to the term 4A2 − 4AB, so it allows a tighter bound on the PEP.
i=n
0 1
i=n
Lv l (bi )
Lv l+1
(aj ) ≈ 4 (Cmini − Cmini )ri A − log cosh( )
i=1 i=1 2
0 Lvl (bi )−4Brj
i=n
cosh(2AB(2Cmini − 1) + 2
)
+ log l
Lv (bi )−4Bri
, (3-25)
1
i=1 cosh(2AB(2Cmini − 1) + 2
)
κ
where Cmini denotes the bit i of the closest codeword to the received signal with bit
j = κ, with κ ∈ {0, 1}.
In the above equation, the first term represents the situation when the user is alone.
The third term represents the interaction between the two users. It is equivalent to the
third term in equation (3-3) of the uncoded scheme. The second term depends only of
ENST 45
CHAPTER 3. JOINT MULTI-USER DECODING
For example, for K = 2, using a common code with memory 4, and rate R = 12 : to
reach the single user performance, simulations show that interleavers of length LI =
2000 symbols are needed. LI increases as a function of K and the memory of the code
since the ambiguity increases with K and the length of error paths increases with the
memory of the code. LI decreases as a function of R.
46 ENST
3.7. CONCLUSIONS
3.7 Conclusions
We studied in this Chapter a method for iterative successive interference cancellation
using the joint probability of all the users. The proposed scheme is said to be optimal
among the iterative decoding schemes, because all the users are taken into consider-
ation by the joint probability measure (3-14). Iterative decoding schemes look very
attractive, especially over the Rayleigh fading channel where the ambiguity does not
exist. The case of two users over the Gaussian channel shows the importance of coding
and interleaving to avoid ambiguity. Using coded systems with or without interleavers
over the Rayleigh fading channel shows the role of interleavers associated with the same
encoder in making different users as orthogonal as possible.
However, the proposed algorithm requires perfect synchronization, and perfect knowl-
edge of the channel.
Both convolutional and block codes were used in the discussed system. While for the
rate 12 practical convolutional codes perform better than block codes, the complexity
of the Fossorier-Lin algorithm is much smaller than the Bahl et al. MAP algorithm,
and block codes do not require long interleavers to achieve their best performance.
8
For noiseless channels both schemes have the same performance at high rates [84]
ENST 47
CHAPTER 3. JOINT MULTI-USER DECODING
0
10
-1
10
-2
Average BER
10
-3
10
-5
10
0 1 2 3 4 5
Eb/No, All users have the same power
48 ENST
3.7. CONCLUSIONS
ENST 49
CHAPTER 3. JOINT MULTI-USER DECODING
Appendix of Chapter 3
3-A The likelihoods of uncoded two users over the
Gaussian channel
Compute the a posteriori of the random variable at by using Bayes relation yields
(y +A−B)2
(− t )
e 2σ 2
+ (y +A−B)2 (yt −A−B)2
· Pr{bt = 1|yt}. (3-A-4)
(− t )
e 2σ 2 + e−( 2σ 2
)
50 ENST
3-B. UNION BOUND OF UNCODED TWO USERS OVER THE RAYLEIGH FADING CHANNEL
ENST 51
CHAPTER 3. JOINT MULTI-USER DECODING
Let
2
δ1t 1
Et2 = + 2,
8No 2σA
δ2 1
Ft2 = 2t + 2 ,
8No 2σB
2 8No 8No
λt + 1 = 1 + 2 2 1+ 2 2
2σA δ1t 2σB δ2t
2 2
A σA B 2 σB2
SNR1t = , SNR2t = .
No No
−Lv l (bj )
E[CM1 ] = 2nA2 + nE[log(cosh( + 2B 2 bj + 2Bηj ))], (3-C-11)
2
52 ENST
3-C. THE MEANS AND THE VARIANCES OF THE MATCHED FILTERS
where the expectation E[·] is made over the Gaussian noise and the two values of the
second user,
−Lv l (bj )
E[log(cosh( + 2B 2 bj + 2Bηj ))] =
2
+∞
−Lv l (bj )
(1 − qj ) log(cosh( + 2B 2 + 2Bt)) · exp(−t2 )dt (3-C-12)
−∞ 2
+∞
−Lv l (bj )
+qj log(cosh( − 2B 2 + 2Bt)) · exp(−t2 )dt,
−∞ 2
and compute
j=n
CM1 = 2nA − 2A2
(Bbj + ηj )
j=1
j=n
−Lv l (bj )
+ log(cosh( + 2B 2 bj + 2Bηj )), (3-C-15)
j=1 2
yields
−Lv l (bj )
E[(CM1 )2 ] = 4n2 A4 + 4nA2 (B 2 + 1/2) + nE[log2 (cosh( + 2B(Bbj + ηj )))]
2
−Lv l (bj )
+ 4n2 A2 E[log(cosh( + 2B(Bbj + ηj )))]
2
−Lv l (bj )
− 4nAE[(Bbj + ηj ) log(cosh( + 2B(Bbj + ηj )))]. (3-C-16)
2
Let
+∞
F (x, y) = logν (cosh(x + yt)) · exp(−t2 )dt (3-C-17)
−∞
exp(−x2 /y 2 ) +∞ ν
= ( ) log (cosh(u)) · exp(−u2 /y 2 + 2xu/y)du
y −∞
2 2
exp(−x /y ) +∞ ν
= 2( ) log (cosh(u)) cosh(2xu/y 2) · exp(−u2 /y 2 + 2xu/y)du,
y 0
ENST 53
CHAPTER 3. JOINT MULTI-USER DECODING
for ν = 1, 2.
exp(u)
For high SNR9 , we make the approximations cosh(u) ≈ 2
and cosh(2xu/y) ≈
exp(2|x|u/y)
2
, which results in
+∞
u − |x|
F (x, y) ≈ (u − log 2)ν exp −( )du
0 y
+∞
≈ −|x|
(yt + |x| − log 2)ν exp(−t2 )dt. (3-C-18)
y
Let
+∞
G(x, y) = t · log(cosh(x + yt)) · exp(−t2 )dt (3-C-19)
−∞
exp(−x2 /y 2) +∞ u − x
= ( ) ( ) log(cosh(u)) · exp(−u2 /y 2 + 2xu/y)du
y −∞ y
−x 2 exp(−x2 /y 2) +∞
= ( ) log(cosh(u)) cosh(2xu/y 2) · exp(−u2 /y 2 + 2xu/y)du
y y 0
1 2 exp(−x2 /y 2 ) +∞
+ ( ) u log(cosh(u)) sinh(2xu/y 2) · exp(−u2 /y 2 + 2xu/y)du.
y y 0
exp(u)
Making the approximations cosh(u) ≈ 2
, sinh(2xu/y) ≈ sign(x) exp(2|x|u/y)
2
, gives
−x +∞
G(x, y) ≈ (yt + |x| − log 2) exp(−t2 )dt
y −|x|
y
+∞
+ sign(x) −|x|
(yt + |x|)(yt + |x| − log 2) exp(−t2 )dt.
y
54 ENST
3-C. THE MEANS AND THE VARIANCES OF THE MATCHED FILTERS
and
and
E[CMi ] ≈ (n − wi )(2A2 + 2(2B 2 − log 2)) + wi (8AB − 2A2 − 2 log 2). (3-C-26)
−Lv l (bj )
E[(CM1 )2 ] = 4n2 A4 + 4nA2 (B 2 + 1/2) + nE[log2 (cosh( + 2B(Bbj + ηj )))]
2
−Lv l (bj )
+ 4n2 A2 E[log(cosh( + 2B(Bbj + ηj )))]
2
−Lv l (bj )
− 2nAE[(Bbj + ηj ) log(cosh( + 2B(Bbj + ηj )))], (3-C-27)
2
and it is approximated by
which gives
ENST 55
CHAPTER 3. JOINT MULTI-USER DECODING
and
E[CMi ] ≈ (n − 2wi)(2A2 )
Ll (bj ) Ll (bj )
+ 2 qj |2B 2 + | + (1 − qj )(|2B 2 − | − log 2)
j/ci j=0
2 2
Ll (bj )
2
+ qj (|4AB + 2B + | − log 2)
j/ci j=1
2
Ll (bj )
+ (1 − qj )(| − 4AB + 2B 2 − | − log 2). (3-C-30)
2
Finally, the mean of the difference is given by
Ll (bj ) Ll (bj )
E[CM1 − CMi ] ≈ 4wi A + 2
(qj (|2B + 2 2
| − |4AB + 2B + |)
j/ci j=1
2 2
Ll (bj ) Ll (bj )
+ (1 − qj )(|2B 2 − | − | − 4AB + 2B 2 − |)).(3-C-31)
2 2
56 ENST
Chapter 4
M any works have addressed the K-user coding over both noiseless and noisy chan-
nels [53, 20, 54, 50]. In [20] a class of multi-user codes has been proposed, and
proved to approach asymptotically the capacity of the channel in the sense that their
ratio goes to 1 when the number of users goes to infinity. The construction method by
iteration on the number of users and the code length of the difference matrix [20] has
inspired all the subsequent works on the K-user codes [54], and [50]. In a recent work
[47], another class of iteratively constructed codes was proved to reach the capacity in
the sense that the difference between the sum rate and the theoretical capacity tends
to zero when the number of users goes to infinity, which gives substantially higher rates
than [20].
57
CHAPTER 4. JOINT MULTI-USER CODING
Considering the sum of the three corresponding BPSK modulated signals gives a received
signal belonging to the constellation (see Fig. 4.1):
CS = {(3 − 1), (1 − 1), (1 − 3), (−1 − 3), (11), (−11), (−1 − 1), (−3 − 1)}.
which has a minimum squared Euclidean distance of 4 (it has a minimum Hamming
distance of 1 −→ 1-decodable). The sum rate of this 3-user code is 1.5 bits, which is
.3113 bits far from the capacity (2-1). Over the noiseless adder channel, the code is
uniquely decodable. Over the noisy channel, this scheme privileges users 1 and 2, since
their proper minimum distances are greater than that of the third user.
1
0.5
−0.5
−1
−1.5
−2
−2.5
−3
−3 −2 −1 0 1 2 3
Our investigation on the K-user codes for the real adder channel is motivated by
concerns about simplicity and symmetrical construction for each user code.
An example of the application of multi-user coding in the MISO mobile-radio cellular
system [57] is the use of 7-user code in a cluster where all users share the same frequency
band, and then repeat this set of codes over the other clusters (Fig. 4.2).
The Chapter outline is as follows: In Section 4.1, we present the system over which
we intend to build K-user codes. In Section 4.2, we consider some information theory
in order to make comparisons with the proposed K-user code. In Section 4.3 we give
the code construction, and study the influence of the different parameters of the code
on its performance. In Section 4.4, we propose a detection algorithm over the noiseless
adder channel which profits from the algebraic structure of the code. Unfortunately,
the considered algorithm is not robust in the presence of additive noise (Section 4.5).
Before concluding the Chapter, we study the concatenated system of inner and outer
codes. A special emphasis is given for the scheme of outer codes with very low rates.
58 ENST
4.1. SYSTEM MODEL
Code 3
Code 6
Definition 4.2 We define the residual interference loss (RIL) to be the ratio be-
tween the minimum squared Euclidean distance of the normalized constellation sum
d2m (S) and the minimum squared Euclidean distance of the normalized original QAM
constellation, which equals 4,
d2m (S)
Ls = 10 log10 . (4-1)
4
The goal of the designer is to find a K-user code that minimizes2 Ls.
1
For the QAM modulation, D should be even. Thus every two components of the modulated vector
represent a point in the two-dimensional QAM constellation. When D is odd, one can consider the
scheme of every component belonging to the BPSK modulation.
2
The K-user code is called uniquely decodable if d2m (S) > 0.
ENST 59
CHAPTER 4. JOINT MULTI-USER CODING
q 4 16 32 64 128 256
Table 4.1: Asymptotic gain, compared to the BPSK, of the modulation q-QAM.
Rotated constellations were first used in [14] to obtain diversity over the Rayleigh
fading channel. In [15, 37, 36] algebraic tools were introduced to build such rotations.
Here, we use rotated constellations to map different users on the real adder channel.
The capacity of the noiseless adder channel is bounded by log2 Sa , where Sa is the size
of the input alphabet [92]. Moreover, when sending K independent symbols from the
q-QAM modulation over the noiseless adder channel, the alphabet at the channel input
could have at most q K different symbols. Therefore, the maximum achievable sum rate
in this scheme is K log2 q bits per two dimensions.
Over the Gaussian channel, the system using q-QAM modulation with average power
per user equal to P as defined above is compared to a TDMA system of K users
described as follows:
each user transmits during K1 of the allocated time duration using the 2K log2 q -QAM
modulation with average power per user equals to min (P̂ , KP ), where P̂ is the power
peak allowed by the physical limitations. P̂ = P in the definition of TDMA commonly
used in the communications community, which we consider in the body of this Chapter.
For the modulation q-QAM, at high SNR, the asymptotic gain per symbol, with
respect to the BPSK, is given by [76]
3
gq = .
2q − 1
Table 4.1 lists the asymptotic gain for the modulation q-QAM with respect to the
BPSK: one looses approximately 3 dB per each added bit.
60 ENST
4.3. K-USER CODE CONSTRUCTION
K
j−2 iφ
r = x1 + e2 xj , xj ∈ q-QAM. (4-3)
j=2
00
11
00
11 000000000000
111111111111
111111111111111
000000000000000 000000000000
111111111111
000000000000000
111111111111111
000000000000000
111111111111111 000000000000
111111111111
000000000000000
111111111111111 000000000000
111111111111
000000000000000 111111111111
111111111111111 000000000000
000000000000
111111111111
000000000000000
111111111111111
000000000000000
111111111111111 000000000000
111111111111
000000000000000
111111111111111 000000000000
111111111111
000000000000000 111111111111
111111111111111 000000000000
000000000000
111111111111
000000000000000
111111111111111
000000000000000
111111111111111 000000000000
111111111111
00
11
00
1100000000000
11111111111
000000000000000
111111111111111 000000000000
111111111111
0011111111111
1100000000000 000000000000
111111111111
00000000000
11111111111 000000000000
111111111111
000000000000
111111111111
00000000000
11111111111
00000000000
11111111111 000000000000
111111111111
11111111111 00
11
00000000000 111111111111111
00
11
00000000000
11111111111 000000000000000
00000000000
11111111111 000000000000000
111111111111111
000000000000000
111111111111111
00000000000
11111111111
00000000000 111111111111111
11111111111 000000000000000
00000000000
11111111111 000000000000000
111111111111111
00000000000
11111111111 000000000000000
111111111111111
000000000000000
111111111111111
00000000000
11111111111
00000000000 111111111111111
11111111111 000000000000000
00000000000
11111111111 000000000000000
111111111111111
00000000000
11111111111 000000000000000
111111111111111
000000000000000
111111111111111
00
11
00000000000
11111111111
00111111111111111
11 000000000000000
In this system, where all users have the same modulation constellation q-QAM, the
first user sends his symbols without rotation. The second user rotates his symbols by
φ, the third user by 2φ, the fourth by 4φ and so on (Fig. 4.3). The proposed code is
constructed as if each user has a different received carrier phase, which equals twice the
preceding user arrival angle. Thus the receiver could separate the multitude of users
without ambiguity.
In the following we note φo,K the optimal angle for the K-user code (4-2) in the sense
of minimizing Ls (4-1).
ENST 61
CHAPTER 4. JOINT MULTI-USER CODING
2
φo,2 = arg (max{ min {||Us − Us || }}), (4-4)
φ Us =Us
which is equivalent to
φo,2 = arg (max{ min
||(U1 − U1 ) + (U2 − U2 )||})
φ or U2 =U2
U1 =U1 ,
where
Z1 ∈ C = {0, ±2, ±2i, ±1 ± i}, Z2 ∈ eiφ C.
2
2i
C
iφ
// e C
1.5
iφ
e (1+i) //
1 1+i
0.5
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Figure 4.4: Constellations C, and eiφ C: the maximization of d2m (S) for two users over
the rotation angle φ.
62 ENST
4.3. K-USER CODE CONSTRUCTION
Fig. 4.4 presents C and eiφ C. By using the symmetry of the constellation C, one
notices that the minimum distance between C and eiφ C occurs between the point
eiφ (1 + i) and its closest neighbors. The maximum of the minimum distance occurs
when the triangle (1 + i, eiφ (1 + i), 2i) is bilateral, which is the situation for φo,2 = π/6.
This angle gives d2m (S) = 1.0718, which yields Ls = −5.719 dB. This scheme has a .74
dB gain over TDMA using 16-QAM (see table 4.1).
4
4−QAM
iπ/6
e 4−QAM
3 Cs
0
y
−1
−2
−3
−4
−3 −2 −1 0 1 2 3
x
Remark 4.1 For two users, TDMA could be seen as a special case of the proposed
scheme. The scheme of two users in time sharing using the modulation 4-QAM is
equivalent to the scheme of the two constellations transmitted simultaneously
C1 = {1 + i, −1 − i},
and
π
C2 = ei 4 C1 ,
where d2m (S) = 8, Ls = 0 dB3 , and the spectral efficiency of the system is 1 bit per
dimension.
3
The comparison is done with the original constellation C1 with minimum squared Euclidean
distance equals to 8.
ENST 63
CHAPTER 4. JOINT MULTI-USER CODING
1 4 0
There are three parameters in the aforementioned K-user code (4-2): K, q, and the
dimension D of the considered q-QAM constellation. In the following we show the
empirical effect of increasing these parameters on Ls.
Fig’s. 4.6, and 4.7 present the distributions of the points in the constellation’s sum
in dimension 2 for K = 3, 4. Notice the repetition of the same pattern of groups of
points in these figures.
Fig. 4.8 presents the evolution of d2m (S) as a function of the rotation angle φ varying
from 0 to π/4 with a discrete step of 0.0014 . Notice that for all the values taken by φ,
(d2m (S), K = 2) ≥ (d2m (S), K = 3) ≥ (d2m (S), K = 4) for this class of K-user codes.
4
For the QAM constellation, d2m (S) is symmetrical with respect to the angle π/4.
64 ENST
4.3. K-USER CODE CONSTRUCTION
0
y
−1
−2
−3
−4
−4 −3 −2 −1 0 1 2 3 4
x
0
y
−1
−2
−3
−4
−5
−5 −4 −3 −2 −1 0 1 2 3 4 5
x
ENST 65
CHAPTER 4. JOINT MULTI-USER CODING
1.4
K=2 users
K=3 users
K=4 users
1.2
0.8
dm(S)
2
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
φ
Figure 4.8: d2m (S) as a function of the rotation angle φ for K = 2, 3, 4 users, and
D = 2.
66 ENST
4.3. K-USER CODE CONSTRUCTION
q 4 16 32 64 128
Table 4.3 lists the values of d2m (S), for q = 4, 16, 32, 64, 128 for K = 2, D = 2.
For q = 16, the rotated scheme has spectral efficiency of 8 bits per two dimensions,
and has a loss of −11.4 dB compared to the single 16-QAM. While, for the same
efficiency of 8 bits per two dimensions, the TDMA system using 256-QAM has a loss
of −12.1706 = −22.3130 + 10.1424, compared to 16-QAM (see table 4.1). Thus there
is a gain of 0.77 dB compared to TDMA. For q > 16, TDMA is better than the rotated
constellations scheme for D = 2.
where M 1 is fixed to be the optimal two-dimensional matrix that maximizes d2m (S)
for D = 2. Thus M 4 depends only on one parameter λ. Respecting the orthogonality
ENST 67
CHAPTER 4. JOINT MULTI-USER CODING
1 λ
a= , b = aλo,2 , c = , d = cλo,2 ,
U 1 + λ2o,2 Uλo,2
where
λ2o,2 + λ2 + λ2o,2 λ2 1
U= , λo,2 = √ .
λo,2 3
1.4
K=2, q=4, D=2
K=2, q=4, D=4
1.2
0.8
d2 (S)
m
0.6
0.4
0.2
0
0 1 2 3 4 5 6
λ
68 ENST
4.4. DETECTION OVER THE NOISELESS ADDER CHANNEL
Definition 4.3 An algebraic number field K = Q(θ) is the set of all possible alge-
braic combinations of an algebraic number θ with the rational numbers of Q. θ is a
root of an irreducible polynomial mθ over Q which is called the minimal polynomial.
θ could be real or complex. If mθ is monic then θ is called integral over the ring of
integers Z.
Let n be the degree of this polynomial, it is also the degree of the algebraic number
field. mθ has n distinct roots in C called the conjugates of θ: θ1 , . . . , θn . (1, θ, · · · , θn−1 )
forms a basis for K. K is called an extension of degree n over Q.
√ √
Let F denote either the field R of real numbers, or the field C of complex numbers
(depending on θ). It is proved that K is dense in F n , thus it is possible to uniquely
represent each element of K, by means of the canonical embeddings, with a point in
6
Some of the results below are only useful in Chapter 5, but for a concise review of algebraic
number theory we put all the required results in one Section.
ENST 69
CHAPTER 4. JOINT MULTI-USER CODING
n−1
n−1
σk : x = xj θj → σk (x) = xj θkj ,
j=0 j=0
Definition 4.5 One can define an algebraic norm over K by N(φ) = σ1 (φ) × · · · ×
σn (φ). One has, N(φ) ∈ Q, or ∈ Z, for φ ∈ K, and φ ∈ OK respectively.
We call the algebraic number field totally real if all the roots of the minimal poly-
nomial are real, the case in which we are interested in this work. Let F n = Rn , and con-
sider an integral basis (w1 , w2 , . . . , wn ). The n vectors vj = (σ1 (wj ), σ2 (wj ), . . . , σn (wj )) ∈
F n , for j = 1 . . . n, are linearly independent. So they define a full rank lattice (a discrete
group) of F n with generator matrix
σ1 (w1 ) σ2 (w1 ) . . . σn (w1 )
σ1 (w2 ) σ2 (w2 ) . . . σn (w2 )
G= .
.. .. .. ..
. . . .
σ1 (wn ) σ2 (wn ) . . . σn (wn )
√
σ1 = e : θ → θ, the identity
70 ENST
4.4. DETECTION OVER THE NOISELESS ADDER CHANNEL
σ2 = ρ : θ → −θ
σ3 = σ : θ → θ̄
σ4 = σρ : θ → −θ̄. (4-6)
forms a Galois group with respect to the function composition law [55]. The basis
{1, θ, θ2 , θ3 } is integral. The lattice of R4 defined by the following generator matrix has
a full rank
1 1 1 1
θ −θ θ̄ −θ̄
G4 = .
θ2 θ2 θ̄2 θ̄2
θ3 −θ3 θ̄3 −θ̄3
4.4.2 Detection, K = 2, D = 2
Let us write the received signal r as
Q(iθ) is of degree 4 over Q, thus in developing (4-7), we obtain a complete linear system
of 4 equations and four unknowns over Z:
x1 = a + ib
x2 = c + id, (4-8)
ENST 71
CHAPTER 4. JOINT MULTI-USER CODING
with a, b, c, d integers from PAM constellations over Z. Replacing in (4-7) and sepa-
rating real and imaginary parts yield
1 1
r1 = (r) = a − d + (c )θ
2 2
1 1
r2 = (r) = b + c + (d )θ. (4-9)
2 2
Over Q(θ), {1, θ} forms a base, thus in (4-9) there is a unique representation of r1 , r2
in the base {1, θ}. The detection algorithm for K = 2, D = 2, ∀ q is based upon the
representation of r1 , r2 in the base {1, θ} which yields a unique solution for a, b, c, d.
We need the following definition to describe the detection algorithm
(x) = |x − [x]|2 ,
4. Replace the values obtained for c, d in (4-7) and compute the values of a, b.
√
The detection algorithm has a linear complexity in q, since the minimization is done
separately over the different PAM constituents of the used q-QAM constellation. It
provides optimal detection over the noiseless adder channel because of the following
lemma
72 ENST
4.4. DETECTION OVER THE NOISELESS ADDER CHANNEL
4.4.3 Detection, K = 3, D = 2
For the 3-user code over the noiseless adder channel, the received signal (4-2) could be
written as
Let
x1 = a + ib, x2 = c + id, x3 = e + if,
r = c + (a + e) cos φ + (b − f ) sin φ
+ i (d + (b + f ) cos φ + (e − a) sin φ) . (4-13)
θ = cos φ, and θ̄ = sin φ, are two algebraic numbers of degree n ≥ 3 over Q, because
the code is uniquely decodable by construction. Since {1, θ, θ̄} is a free set over Q(θ),
the following linear system has a unique solution over Z
1. r ← e−iφ r.
ENST 73
CHAPTER 4. JOINT MULTI-USER CODING
Over the noiseless adder channel, this algorithm succeeds in separating the users with
considerably less complexity than the ML detection. The complexity of the proposed
algorithm depends on the size of the considered DPAMq,2 . For example, over the q-
√
QAM modulation obtained by the same q-PAM over real and imaginary parts, one
√
considers in the detection two minimizations over a DPAMq,2 of size 2 q − 1. This
√
induces 2(2 q − 1)2 computations of the function (x) (4.6), while the ML-detection
computes q 3 Euclidean distances or cross correlations.
yields
Proposition 4.1 In the above K-user code, over the noiseless adder channel the
real and imaginary parts are uniquely decodable separately.
we have
K
j−2 iφ
K
j−2 iφ
x1 + e2 xj = x1 + e2 xj .
j=2 j=2
Suppose that in equation (4-16) one has one solution for r1 and two solutions for r2 ,
then combining the two equations one obtains two different solutions for (4-15) and thus
74 ENST
4.4. DETECTION OVER THE NOISELESS ADDER CHANNEL
K||q 4 16 32 64
Table 4.4: The complexity of the ML detection given in the number of computed
Euclidean distance.
K||q 4 16 32 64
3 18 98 213 450
Table 4.5: The complexity of the algebraic detection given in the number of computed
function (x).
two different solutions for (4-3) which contradicts the hypothesis of uniquely decodable.
Applying the same principle of the algebraic detection on r1 − (xK−1 ) and r2 −
(xK−1 ), one should perform one minimization of the function (x) over at most
2(K − 3) variables belonging to the constellation PAM and two variables belonging
√
to DPAMq,2 . Thus, the complexity of the algorithm is at most (q K−3 + 1)(2 q − 1)2
computations of the function (x) (4.6). This complexity is to be compared with
q K computations of Euclidean distances, which is the complexity of the ML-detection.
Besides, the computation of the function (x) is less complex than the computation
of the Euclidean distance. In Tables 4.4, and 4.5, we list the complexity of the ML and
the algebraic detections for different values of K and q. For example for K = 3, and
q = 64, the complexity of the ML detection is 582 times that of the algebraic detection
algorithm.
The values of the rotation angles were obtained through an exhaustive search by the
computer search.
In order to separate K users in one algebraic equation (4-3), the algebraic number eiφ
should be at least of degree ≥ 2K so that cos φ could be of degree ≥ K. For example,
ENST 75
CHAPTER 4. JOINT MULTI-USER CODING
2 2π/12 −5.719 2
3 2π/20 −10.9 4
4 2π/41 −17.3 20
5 2π/83 −22.42 41
6 2π/180 −29.12 24
Table 4.6: Optimum rotations in a close form on 4-QAM in the dimension D = 2, and
the degrees of the corresponding algebraic number fields.
for K = 2, eiπ/6 is of degree9 Φ(12) = 4 over Q, and cos π6 is of degree 2. For K = 3, the
closest algebraic number to ei.312 in close form is ei2π/20 which has a degree Φ(20) = 8
π
over Q, and thus cos 10 is of degree 4. The rotation ei2π/20 in the 3-user scheme gives
an RIL of Ls = −10.9 dB. Table 4.6 shows the closest algebraic number in close form
to the optimal rotations. It also lists their degrees over Q and their RIL’s.
9
Φ(·) is the Euler function giving the number of integers less than N and prime with it ([13] p.
363).
76 ENST
4.5. DETECTION OVER THE GAUSSIAN CHANNEL
Estimate Decoder 1
Source 1
Joint
Detection
Estimate Decoder 2
Source 2
Figure 4.10: Concatenating the K-user code with error control codes.
Consider the system in Fig. 4.10: the mapper is normalized in a way that Eb1 =
Eb2 = 1, and N0 varies. Each couple of coded bits btj , bt+1 j , j = 1, 2 of user j are
2
mapped to the point xj = (2bj − 1, 2bj − 1) in R , x2 is then rotated by π6 . By
t t t+1 t
joint detection we mean that the detector takes into consideration the form of the
received signal r= xt1 +xt2 +νt , where νt is a two-dimensional white noise, component-wise
independent, with variance N20 on each dimension. The detector provides Pr{btj = 0|rt }
and Pr{btj = 1|rt } to the decoder j at each time t. The last one considers these
reliability values as a priori and performs soft-input decoding.
ENST 77
CHAPTER 4. JOINT MULTI-USER CODING
In contrast with iterative multi-user decoding (Chapter 3), in this scheme the de-
coders use single user metric, and there are no iterations among them.
• The use of very low rate codes implies operation over a very noisy channel in term
of signal-to-noise ratio Es/N0 , as well as a very low symbol signal-to-interferer
ratio Es/I. Since the transformation done on each user renders the sum uniquely
decodable, multiplying by a very low code rate yields a system with practically
no interferers, operating on a very low signal-to-noise ratio.
78 ENST
4.7. CONCLUSIONS
1e-01
1e-03
1e-04
1e-05
1e-06
-0.20 0.10 0.40 0.70 1.00 1.30 1.60 1.90 2.20 2.50 2.80
Eb/No in dB
4.7 Conclusions
We presented in this Chapter a new class of K-user codes for the real adder channel.
The considered code does not belong to the family of K-user code over the binary
adder channel (BAC) [53, 20, 54, 50]. Its main characteristics are:
• The construction of the code is simply done by rotating the constellation QAM.
• Over the noiseless adder channel, the proposed K-user code is uniquely decodable
without any redundancy, therefore it achieves the capacity of the noiseless adder
channel with K log2 q bits per two dimensions.
• Over noisy channels, the absence of redundancy in the considered class of uniquely
decodable codes induces a penalty as a decrease in the minimal squared distance
of the resulted constellations sum.
We have proposed an algebraic detection scheme for some cases over the noiseless adder
channel, but the general case over the noisy channel still lacks an efficient soft-input
soft-output decoding algorithm.
ENST 79
CHAPTER 4. JOINT MULTI-USER CODING
Over the noisy channels, in the 2-dimensional space, the proposed scheme gives better
performance than TDMA. Increasing the three parameters K, q, D in the proposed K-
user code yields almost always more improvement on TDMA.
The drawback is the need of carrier synchronization and the detection complexity
over the noisy channel.
Concatenation with very low rate turbo code improves drastically the performance of
the system, since the resulted system is equivalent to a SISO channel operating at low
average symbol to noise energy ratio Es/N0 .
In a cellular system, it seems that joint decoding is always necessary and prevalent
on joint coding. While the latter cancels the MAI originating from the other users in
the same cluster, multi-user decoding cancels the MAI of all active users transmitting
in the same frequency band [46].
The proposed K-user code fakes a multi-antenna environment when each user goes
k−2
through a different fading ei2 φ such that the sum of all users at the receive antenna
is uniquely decodable. In the next Chapter, K-user codes for a multi-antenna system
are explored.
80 ENST
Chapter 5
R ecently, the field of space-time (ST) coding has attracted interest as means of
exploiting the inherent transmit diversity in a multi-antenna system [91], and
[30]. The channel is assumed to be static1 over the code length in time, meaning that
there is no time diversity in the channel. It is the role of the code to transform the
inherent transmit diversity into time diversity . In a system of M transmit antennas,
and N receive antennas, the maximum time diversity induced by the use of ST codes
is equal to M, and the total space-time diversity is MN [30].
1
This channel model fits the indoor rich scattering wireless channel [45].
81
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
... . .
.... ...
.... .... .... .......
- TX
.... ...... ..
a1 .... ......
Quasi-static
RX
Modulation a2 - TX
...
....
... ..
......
...
.
....
.
....
..
.... ..
......
....
.
...
- ?
Data and Vector RX
- Encoding
...
....
Rayleigh fading
.
....
.
.... ...
.... Detection -
Sink
- TX -
..
.... .... ..
.... ...
RX
.... .
....
channel
.
...
... ...
...
6
- TX
.... .
.......... .... ....
a4 .. ......
RX
r1 = H · a + ν, (5-1)
2
In a multi-user scenario, where each substream represents a different user, this assumption is not
valid. However, for a unified frame work it is always assumed that the total radiated power is indepen-
dent of M . Furthermore, in a multi-user context single substreams codes are called interchangeably
single user or outer codes, and space-time codes are called K-user or inner codes (Section 5.4).
82 ENST
5.1. UNCODED SYSTEM
The permutation (5-2) is set to the identity. The detection is performed as follows:
• Step 1: yk1 = w k1 r 1 .
• Step 2: Slice yk1 to obtain âk1 = Q(yk1 ) the nearest point to yk1 in the constellation
in the sense of minimum Euclidean distance.
where (H)k1 denotes the k1 -th column of H. Steps 1 − 3 are performed over the
rest of S
E[|aki |2 ]
ρki = , (5-4)
σ 2 ||wki ||2
3
The two methods are equivalent for high SNR, which is often the case in the V-BLAST detection
scheme due to the absence of code inter-substreams.
4
The Moore-Penrose pseudoinverse of a matrix is defined for an over-determined linear system,
which means that there are more equations than unknowns. That is why N should be greater than
M in the V-BLAST architecture.
ENST 83
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
where the expectation is taken over the constellation points. Thus the descending
order of ρki is equivalent to ordering w ki in ascending values. It is intuitive that a SIC
type algorithm that subtracts hard values5 at each iteration, performs better if the
subtracted interferers are ordered in descending order of their post-detection SNR. By
doing so, V-BLAST with optimal ordering statistics is described as follows:
• Initialization: i ← 1. G1 = H + .
• Recursion:
w ki = (Gi )ki
y ki = w ki r i
âki = Q(yki ) (5-5)
ri+1 = r i − âki (H)ki
Gi+1 = H+
k i
i ← i + 1,
V-BLAST limits
84 ENST
5.1. UNCODED SYSTEM
Improving on V-BLAST
The optimal order version of the algorithm recalls the structure of the Fossorier-Lin [32]
decoding algorithm of block codes: second order statistics are ordered in descending
order and a “hard” detection is performed on the symbols. When comparing the two
algorithms, one may want to perform a reprocessing of the detected vector in the V-
BLAST. The difference between the two schemes is the absence of the code structure
in the matrix H in the latter case.
The computed nulling vectors w ki are stored for reprocessing. An Or -order repro-
cessing is performed as follows:
ENST 85
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
0
10
−1
10
−2
10
−3
10
−4
10 V−BLAST optimal order alone, M=N=2.
Second order repro. V−BLAST optimal order, M=N=2.
V−BLAST optimal order alone, M=N=4.
Second order repro. V−BLAST optimal order, M=N=4.
−5
10
0 5 10 15 20 25
SNR per bit
Figure 5.2: V-BLAST alone and V-BLAST with second order reprocessing, M TX and
N RX over a Rayleigh fading channel.
The detector computes yj and âj (5-5). Then yj , j = 1, . . . , M, are passed to the soft-
input decoder. The decoder does not use the channel-state information (CSI), since
the CSI has been exploited by the detector to inverse the channel effect. The decoder
considers yj as a disturbed version of the transmitted symbol aj . The multitude of
transmitted substream signals are separated thanks to the rich scattering multi-path
channel. However, in the situation, where the decoders could not exploit directly the
multi-antenna CSI, implies severe degradation in performance compared to iteratively
decoded substreams schemes.
Iterative decoding
Independent random interleavers are added after the modulators in order to break error
paths resulted from the single substream decoders in the iterative decoding scheme (Fig.
86 ENST
5.1. UNCODED SYSTEM
RX
TX RX
coder 1 Decoder 1
TX RX
coder 4 Decoder 4
RX
5.4)7 .
At the iteration ι, the soft estimate computed here ãιj is [84], [63]
√
ãιj = (2prj − 1) + (2pij − 1) −1,
where prj = Pr{(aj ) = +1|Observationι } (5-7)
and pij = Pr{(aj ) = +1|Observationι }.
The a posteriori probabilities are initialized to 0.5. The interleavers size is 5 times the
codeword length, and the number of iteration is 3.
Simulation results
ENST 87
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
TX RX
Iterative Decoding
Data
Stream
TX RX
Interference Interference
Cancellation Cancellation
soft estimate of substream 2
88 ENST
5.1. UNCODED SYSTEM
0 0
10 10
−1
10
−1
10
Average Bit Error Rate
−2
10
−3
10
−3 −4
10 10
0 2 4 6 8 10 12 0 2 4 6 8 10 12 14 16
Eb/N0 in dB Eb/N0 in dB
complex Gaussian random variables of variance 0.5 per dimension. H is fixed during
the burst length which is 5 times the codeword length.
The average BER is plotted as a function of the transmitted Eb/N0 in dB; power
computation is done as follows: the q-QAM modulation is normalized such that the
average symbol energy Esav = 1, then multiplied by a factor Aq depending on the
average SNR per bit by the formula:
Rout
Aq = log2 (q) · Eb/N0 · , (5-9)
M
where Rout is the outer code rate.
Fig. 5.6 shows the performance of the outer codes BCH(16, 5) and Golay(24, 12),
where the Fossorier-Lin decoding algorithm with first order reprocessing is used. The
scheme using the code BCH(16, 5) has a spectral efficiency of 2.5 b/s/Hz, and iterative
decoding gives average BER of 10−3 at Eb/N0 = 12 dB. While for the same Eb/N0 ,
the scheme with the Golay code, which has a spectral efficiency of 4 b/s/Hz gives BER
equal to 10−2 .
Although in a rich scattering environment, one can separate the multitude of sub-
streams by a multi-substream detector cascaded by single substream decoders, one
ENST 89
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
observes a big difference in performance between this scheme and the iterative decod-
ing one. In the latter case, the decoder has access to the CSI of the N receive antennas,
which exploits better than the detector in the former case. For example, for M = 1, a
maximum ratio combiner cascaded by a decoder, allows the last one to exploit the N
received versions of the signal. While for M > 1, a detector which cancels the interfer-
ences with other substreams, gives the decoder its decision on the detected substream
without allowing it to profit form the N versions of the received substream disturbed
by other signals.
In Fig. 5.7, the average BER is plotted for different iterations. After the third
iteration no substantial improvement in the iterative decoding scheme can be observed.
The ML performance of the single substream (M = 1 transmitter using the code
BCH(16,5), and N = 2 receivers) is reached due to the multi-path structure that helps
the decoder to separate the different substreams.
where Es is the average transmitted energy per symbol. d2 (X, E) defines a distance
between X and E, it is given by
N
d2 (X, E) = H j AH j † , (5-11)
j=1
where H j is the j-th row of the transfer matrix H, and A(X, E) = B(X, E)B † (X, E)
90 ENST
5.2. SPACE-TIME CODES, CONSTRUCTION CRITERION AND PERFORMANCE
Simplifying (5-10) and averaging over the Rayleigh fading variables gives
r −N
Pr X → E ≤ λi (Es/4N0 )−rN . (5-12)
i=1
Thus a diversity gain of r · N and a coding gain of (λ1 λ2 · · · λr )1/r is achieved, where r
is the rank of A, λi , i = 1 . . . r are the eigenvalues of A. Finally minimizing the PEP
is equivalent to
• The rank criterion: the rank r of B is called the diversity gain of the code. In
order to achieve the maximum diversity M · N, the matrix B has to be full rank
for any pair of codewords (X, E).
• The determinant criterion: if the diversity gain of the code is r, then the mini-
mum of the r-th root of the sum of determinants of all r × r principal cofactors of
A(X, E) taken over all the codewords pairs (X, E) must be maximized (coding
1
2
gain). In the sequel we always refer to the coding gain by (δmin ) r whether B is
square or not.
• Apply the rank criterion without redundancy (constellations from algebraic num-
ber fields).
In the rest of this Section we give some examples of the first category, in the next
Section we examine in detail the second category of ST codes.
ENST 91
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
0 00 01 02 03
1
1 10 11 12 13 2 0
2 20 21 22 23 3
3 30 31 32 33
In Fig. 5.8 is depicted a trellis code that transform the 2 transmit antenna diversity to
2 time diversity for the 4-PSK modulation. The code has one register, one input from
the 4-PSK modulation (thus it has 4 states), and two outputs for the two transmit
antennas. For example if the register state is the signal 0, and the input is 1, we
transmit the signal 1 on the first antenna, and the signal 0 on the second, and so on.
This code verifies the rank criterion because over two periods of time, the matrix B
of the code has a full rank over all codeword pairs (see [90] for demonstration), and the
decoding is simply done by a Viterbi algorithm. However, the rate of this code is 1/2.
92 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
Space
6
TX 1
S
S (
x1 x2 )
TX 2
S
S (;
x2 x1
)
- Time
τ1 = h1 x1 + h2 x2 + ν1
x̃1 = h∗1 τ1 + h2 τ2∗ = |h1 |2 + |h2 |2 x1 + noise
x̃2 = h∗2 τ1 − h1 τ2∗ = |h1 |2 + |h2 |2 x2 + noise
ENST 93
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
We aim to transmit complex symbols with two transmit antennas, over two slots of
time such that the space-time matrix B(X, E) (5.2.1 ) has full rank, i.e. the minimum
over all codeword pairs of the absolute value of its determinant is nonzero.
Choose a complex number field of degree 4 over Q(i), let
α −σ(α)
Θ2 = ,
σρ(α) ρ(α)
This code has no redundancy, over 2 time periods we transmit 4 symbols with two
antennas. Let
C4,q = {x = x1 − x2 , x1 = x2 ∈ q-QAM4 },
denote the set of all the difference of vectors x1 − x2 with component from q-QAM
4 4
constellation. For q = 4, the size of C4,4 is |C4,4 | = 4 (42 −1) = 32640. The determinant
of B(X, E) is given by8
94 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
δ = 2b21 − 2b22 (θ2 + θ¯2 ) + b23 (θ4 + θ¯4 ) − b24 (θ6 + θ¯6 )
+ 2b1 b3 (θ2 + θ¯2 ) − 2b2 b4 (θ4 + θ¯4 )
= 2b21 − 8b22 + 12b23 − 40b24 + 8b1 b3 − 24b2 b4 . (5-13)
Proposition 5.1 In the above example, the determinant of the matrix Θ2 could be
written as a trace of a norm, and thus is an integer. Over C4,4 it has a minimum
absolute value of 8.
Proof: Consider the group G = {e, σ, ρ, σρ}, where e is the identity, σ maps θ to θ̄,
and ρ maps θ to −θ. ρ and σ are commutative because G is isomorphic to Z22 (the
Cartesian product of the group of integers modulo 2). The determinant of Θ2 is
Lemma 5.1
√ √
σ : Q( 2) −→ Q( 2)
√ √
2 → − 2.
√ √
Proof: One has 2 ∈ Q(θ), and its representation in the basis {1, θ, θ2 , θ3 } is 2 =
−2 + θ2 . It yields
√
σ( 2) = σ(−2 + θ2 )
= σ(−2) + σ(θ2 )
√
= −2 + θ2 = −2 + θ̄2 = − 2. (5-15)
The preceding arguments guarantee that δ is integer, but we can not guarantee that
it is non zero over all the constellations. Only over constellations of small size like the
4-QAM or the BPSK where the size of C4,q is not very large we can compute its value
by exhaustive search, which gives 8.
In this special case the determinant is a non zero integer (when all bj are not zeros
simultaneously, since it is a trace of a norm), thus we have a full diversity system. But
ENST 95
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
we are interested in finding θ, and thus the algebraic number field K, that maximizes the
minimum square of the absolute value of the determinant over the group of rotations,
thus we maximize the coding gain.
Because of the resemblance between this optimization problem and the one of max-
imizing the minimum product distance dP,min , in rotated constellations that achieve
maximum diversity order over the Rayleigh fading channel, we follow the Hadamard
construction in [16].
has
2
= min{| det (Θ2 )|2 },
δmin
C4,4
maximized over the group of rotations. This optimization problem has many param-
eters that grows rapidly with the dimension. To solve this optimization problem, we
follow the iterative construction used for the Hadamard matrices.
where M 1 is fixed to be the optimal two-dimensional matrix that maximizes the min-
imum product distance [16] 9 . Doing so, the matrix M 4 is then determined by one
variable λ. Respecting the orthogonality and unit constraints, the considered matrix
family is given by
1 λ
a= , b = aλo,2 , c = , d = cλo,2 ,
U 1 + λ2o,2 Uλo,2
9
This is a logical choice of M 1 , since in the square absolute value of the determinant, the product
distance in the 2-dimensional space intervenes.
96 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
where √
λ2o,2 + λ2 + λ2o,2 λ2
1± 5
U= , λo,2 =
λo,2 2
λo,2 is the optimum parameter that maximizes the minimum product distance in di-
mension 2 [16].
In dimension 4, we can compute δ as follows:
which can be computed by an exhaustive search through the points of the finite con-
stellation C4 built on 4-QAM.
2
In Fig. 5.10 is depicted the values of δmin as a function of λ, with steps of 0.005. We
2
also plotted the following upper bounds on δmin :
16 · |ad + bc|2 , (2, 0, 0, 0)
2
δmin ≤ 16 · |(a2 + d2 − c2 − b2 ) − 2(ad + bc)|2 , (2, 0, 0, 2) (5-17)
16 · |(a2 + d2 − c2 − b2 ) + 2(ad + bc)|2 , (0, 2, 2, 0)
2
We identify the two highest peaks of δmin at the intersection of the first and the
second upper bound
1 √ √
λo,4 = ( 2 − 1) 50 + 10 5 = 0.3523511.
10
and at intersection of the first and the third upper bound
1 √ √
λo,4 = ( 2 + 1) 50 + 10 5 = 2.0536527.
10
It is worth noticing that these same values of λ maximize the minimum product
2
distance in [16], although the upper bounds on δmin and dP,min are not the same. Over
the 4-QAM constellation, the optimum values are δmin,o = 16
2
10
, and dP,min = 16
40
.
The choice of real rotations is motivated by the fact that the resulting rotated con-
stellation will have full spatial diversity [36], [16]. A property that is useful for building
space-time codes good for both quasi-static and fast fading channels (see subsection
5.3.6).
ENST 97
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
16
δ2
min
2
14 16|ad+bc|
2 2 2 2 2
16|(a +d −c −b )+2(ad+bc)|
16|(a2+d2−c2−b2)−2(ad+bc)|2
12
δ2 , minimum square determinant
10
6
min
0
0 0.5 1 1.5 2 2.5 3
λ
2
Figure 5.10: Maximizing δmin over a family of rotations constructed in a Hadamard
method.
1.6
δ2
min
dP,min
1.4
1.2
1
P,min min
, δ2
0.8
d
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3
λ
2
Figure 5.11: The peaks of dP,min are bounded by δmin in 4 dimensions.
98 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
Higher dimensions
• Dimension 6, M = 3, l = 2:
y1 −y4
Θ3,2 = y2 y5 , (5-18)
y3 y6
with
U= 1 + λ2 (λ23,o − λ + 2),
where λo,3 = [2 cos(4 · π/7)]−1 = −2.24698 is the optimal value maximizing dP,min
in the dimension 3.
2
We computed the values of δmin and dP,min over
C6,4 = {x = x1 − x2 , x1 = x2 ∈ 4-QAM6 },
ENST 99
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
0.7
δ2
min
d
P,min
0.6
0.5
0.4
P,min min
d , δ2
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
λ
2 2
Figure 5.12: δmin and dP,min in dimension 6. The peaks of dP,min are bounded by δmin .
where the vectors from C6,4 of norm greater than 3 were skipped. The optimal
2 2
value of δmin is obtained for λo,6 = .455 and equal δmin,o = 0.6; notice that (Fig.
5.12) for this value of λ one has a local optimum of dP,min = .045. The optimal
2
value of dP,min = .116 is obtained for λo,6 = .53, for this value δmin = 0.54 which
2
is close to δmin,o .
2
Fig. 5.12 shows the evolution of δmin and dP,min as a function of λ varying from 0
to 1 with a discrete step of 0.005. Notice that the peaks of dP,min are bounded by
2
δmin . This empirical observation is also verified in dimension 4 (see Fig. 5.11).
10
When M = l, ΘM,l is referred to by ΘM
100 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
Table 5.1: First row of the optimal rotation matrices in dimensions 8 and 12.
• M = 4, l = 2:
y1 y2 y3 y4
ΘT4,2 = ,
(5-19)
−y5 y6 −y7 y8
• M = 4, l = 3,
y1 −y5 y9
y2 y6 y10
Θ4,3 = , (5-20)
y3 −y7 y11
y4 y8 y12
ENST 101
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
16 = l · M,
y1 −y5 −y9 y13
y2 y6 −y10 −y14
Θ4 = , (5-21)
y3 −y7 y11 −y15
y4 y8 y12 y16
For the case M > l, one can consider unequal diversity schemes (UDS)12 , by coding
the first l antennas with the corresponding Θl code, and then transmitting uncoded
symbols on the M − l remaining antennas during l periods of time.
Θl
a1 ... al
ΘUDS,M,l = , (5-22)
.. .. ..
. . .
a(M −l−1)l+1 . . . a(M −l)l
One can build another scenario of UDS in the case of multi-user communication
where each user disposes of l transmit antennas. The proposed UDS scheme says that
each user encodes his symbols by Θl ST code. The rank criterion is still verified in
such a scheme since each ST code Θl is a full rank matrix.
In a system of p users where the number of total transmit antennas equals to M = p · l,
11
For higher dimensions than 12, the construction method on the Hadamard family [16] becomes
intractable. However, one can construct rotations on the cyclotomic number fields that possess a good
minimum product distance along with full spatial diversity [8].
12
In comparison with unequal error protection in the case of error correcting codes
102 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
Proposition 5.2 The matrix A(X, E) (see subsection 5.2.1) of the ST code ΘUDS,M
has a rank equal to l. In additions, the l-th root of the sum of the determinants of all
1
2
l × l principal cofactors of A equals to (δmin,l ) l , where δmin,l is the minimum of the de-
terminants of Θl taken over all codeword pairs (X, E) in the corresponding modulation.
To prove the determinant criterion, one notices that each l × l principal cofactor Aco of
A is a product of two sub-matrices B co and B †co of B and B † : Aco = B co B †co . Hence
the determinant of Aco is a positive number because it equals to the absolute value of
†
the determinant of B co . On the other hand we know that Aco,j = Θjl Θjl , j = 1 . . . p
2
are l × l principal cofactors of A with a common minimum determinant of δmin,l , thus
1
2
the coding gain is greater than (δmin,l ) l since the components of X − E can be all
zeros.
To prove the inverse inequality one computes the minimum of the sum of the determi-
nants of all l × l principal cofactors of A in the special case where X − E has all its
ENST 103
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
components equal to zeros except the last l2 symbols. This minimum equals to δmin,l
2
,
which proves the determinant criterion of the code ΘUDS,M .
In the next subsection we give an example for M = 3, and l = 2, where the first four
symbols are coded by Θ2 and the remaining two symbols are sent uncoded on the third
antenna during two time periods. Another example of UDS schemes is given for two
users transmitting simultaneously, where each user disposes of two transmit antennas
and encodes his symbols by the code Θ2 . Comparisons are made with the ST code
Θ4,2 .
Fig. 5.13 shows the symbol error rate (SER) over a quasi-static Rayleigh fading
channel as a function of Eb/N0 of the modulation 4-QAM for the codes Θ2 , G2 and the
uncoded system; for one receive antenna. It is shown that the two curves of Θ2 and
G2 have the same slope and thus the same diversity order. G2 has better coding gain.
Fig. 5.14 shows the average symbol error rate for the Θ2 , G2 and the uncoded 4-QAM
for two receivers. In Fig. 5.15 we present the average BER for the same scheme.
In Fig. 5.16 we show the performance of the ST code Θ3,2 obtained by a rotation in
dimension 6, in a scheme of three transmitters, using the modulation 4-QAM and one
receiver. Fig. 5.17 presents the performance of each one of the 6 symbols when using
unequal diversity scheme by coding 4 symbols with Θ2 , and sending the remaining
two symbols uncoded on the third transmit antenna within two periods of time; the
scheme contains one receiver. Notice that even the uncoded symbols have enhanced
their performance in this scheme. One can understand better this phenomenon if one
supposes that the ML detection is performed by a SIC type algorithm: when the first
four coded symbols have small error rates, then subtracting them from the received
signal would allow the remaining two uncoded symbols to have better performance. It
is noticed that the average symbol error rate taken over all the coded symbols is almost
like the average shown in Fig. 5.16 in the scheme Θ3,2 .
104 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
0
10
uncoded, 4 bits/s/Hz
-1
10 Algebraic code, 4 bits/s/Hz
G2 code, 2 bits/s/Hz
-2
10
Average symbol error rate
-3
10
-4
10
-5
10
-6
10
-7
10
10 20 30 40
SNR per bit
ENST 105
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
0 0
10 10
−3 −3
10 10
−4 −4
10 10
−5 −5
10 10
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
SNR per bit SNR per bit
−1
10 −1
10
Average symbol error rate
−2
10 −2
10
−3
10 −3
10
−4 −4
10 10
10 15 20 25 30 35 10 15 20 25 30 35
SNR per bit SNR per bit
Figure 5.16: ST code Θ3,2 with 4-QAM Figure 5.17: ΘUDS,3,2 with 4-QAM over
over quasi-static fading, N = 1. quasi-static fading, N = 1.
106 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
0
10
−2
10
Average symbol error rate
−3
10
−4
10
−5
10
−6
10
0 5 10 15 20 25 30
SNR per bit
Figure 5.18: Average SER versus Eb/N0 . Θ2 with 16-QAM over quasi-static fading,
N = 2.
The symbol error rate of the modulation 16-QAM is presented in Fig. 5.18 for two
receive antennas for Θ2 , G2 ST codes and the uncoded system. When the size of
the constellation is increased, we notice that the coding disadvantage of Θ2 is a bit
degraded to 6 − 7 dB compared to G2 .
Fig. 5.19 shows the performance of the algebraic ST codes Θ4,2 and ΘUDS,4 with
BPSK modulation using 4 transmit antennas and one receive antenna over a quasi-
static Rayleigh fading channel. In this multi-user scenario we maintain the total
launched power independent of M.
Fig. 5.20 presents the performance of Θ4,3 with the BPSK modulation over a quasi-
static fading with one receive antenna.
Remark 5.1 One notices that, for a scheme of 4 bits/s/Hz, the uncoded system of
4 transmit antennas using the uncoded BPSK modulation is worse than a scheme of 2
transmit antenna with uncoded 4-QAM modulation which has also 4 bits/s/Hz (see Fig.
5.13). This is due to the way we normalize the power on all transmit antennas: if one
increases the total spectral efficiency by increasing the number of transmit antennas then
the radiated power per bit on each antenna is lowered such that the total radiated power
is still constant. While, when increasing the spectral efficiency by increasing the size of
the modulation, the radiated power per bit on each antenna is still constant. However,
ENST 107
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
0 0
10 10
−2
10
Average bit error rate
−2
10
Average BER
−3
10
−3
10
−4
10
−4
10
−5
10
−5 −6
10 10
10 15 20 25 30 35 10 15 20 25 30 35
SNR per bit SNR per bit
Figure 5.19: Average BER of ST codes Figure 5.20: Average BER of the ST code
Θ4,2 , ΘUDS,4 with BPSK modulation over Θ4,3 with BPSK modulation over quasi-
quasi-static fading, M = 4, N = 1, l = 2. static fading, M = 4, N = 1, l = 3.
108 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
all the comparisons that we have made are valid since in a given multi-antenna scheme
using a certain modulation, we fix the SNR per bit for all the compared systems (ΘM
ST codes, uncoded symbols, and G2 ).
5.3.3 Analysis
The coding gain of the code Θ2 is Gc = δmin,o 2
= √410 , while it equals 4 for the code
G2 . At the same diversity order there is a loss of 10 log10 √110 = −5 dB, which we can
see in the simulations at high SNR. When the size of the modulation increases beyond
2 bits/symbol, the determinant criteria is no longer guaranteed. Nevertheless, coding
information symbols by Θ2 always improves over the uncoded system (see Fig. 5.18).
This is because guaranteeing the determinant criteria over the 4-QAM modulation
means that the decoder corrects all error patterns of magnitude less than that of the
4-QAM. The error events of higher magnitude have small probabilities of occuring.
For the other ΘM,l ST codes, we notice that the slopes of performance curves are
proportional to l (see Fig. 5.19 where the curve decreases l = 2 orders of the power of
10 within 10 dB, and Fig. 5.20 where the curve decreases l = 3 orders within 10 dB).
Thus, simulation results confirm the theory that ΘM,l has a diversity order of lN in a
scheme of M transmit and N receive antennas.
r = [r1 , r2 ] = HΘ2 + ν
y1 −y3
= [h1 , h2 ]
+ν
y2 y4
⇐⇒
h1 h2 0 0
rT = V ec(Θ2 ) + ν T
0 0 h1 h2
h1 h2 0 0
= M [a1 , a2 , a3 , a4 ]T + νT
4
0 0 h1 h2
ENST 109
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
= H [a1 , a2 , a3 , a4 ]T + ν T , (5-25)
where V ec(Θ2 ) transforms the matrix Θ2 into one vector containing the first then
the second column of Θ2 . In the above relation M 4 is the rotation matrix M 4 with
the third line multiplied by −1. Thus H , the new 2 × 4 transfer matrix, consists in
correlated complex Gaussian random variables; because Gaussian random variables are
fixed under unit transformation. Writing the space-time codes in this form shows how
they transform the inherent transmit diversity M into time diversity expressed here
by increasing the receive diversity N −→ lN. Coding disadvantage is shown by the
correlations between lines of the new transfer matrix H .
5.3.5 Decoding
The price to pay for the bandwidth lossless full diversity space-time codes is the ML
decoding complexity. Since all the information is contained in every transmitted sym-
bol one can not separate the metrics as in other less efficient codes. In the situation
of M = l = 2, ML decoding is feasible, but for higher dimensions it becomes rapidly
intractable since its complexity grows exponentially with M and l.
In this subsection we will prove that ΘM codes verify the aforementioned criteria
and present simulation results over the fast fading channel.
110 ENST
5.3. SPACE-TIME CODING WITHOUT LOSS IN SPECTRAL EFFICIENCY
Channel model
In the literature, fast fading is modeled by the transfer matrix H N ×M (t), where all the
hj,k (t) are independent complex Gaussian random variables.
Remark 5.2 When the space-time codes ΘM are used over fast fading that changes
during an inner codeword, it is still possible to represent them by their equivalent un-
coded system 5.3.4. For example, M = 2, N = 1, the equivalent uncoded system is
given by
h1 h2 0 0
rT =
· M [a1 , a2 , a3 , a4 ]T
4 + νT (5-26)
0 0 h3 h4
= H [a1 , a2 , a3 , a4 ]T + ν T ,
Design criteria
For both fast and quasi-static fading, the criterion in 5.2.1 become
• The product/determinant criterion: let V (X, E) denote the set of time instances
1 ≤ t ≤ l, such that x1t x2t · · · xM 1 2
t = et et · · · et and let
M
M
||X t − E t ||2 = |xjt − ejt |2 .
j=1
Then to achieve the most coding advantage in a rapid fading environment, the
minimum of the products
||X t − E t ||2 ,
t∈V (X ,E )
ENST 111
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
taken over distinct codewords must be maximized. As for the determinant crite-
rion, it is the same as in 5.2.1.
For ΘM codes, V(X, E) = {1 · · · l}, and by construction again, the minimum
product distance
l
M
dP,min = min |xjt − ejt |,
X =E t=1 j=1
is maximized. But for positive real numbers z1 , z2 , · · · , zM , maximizing Mj=1 zj
M 2
will maximize j=1 zj . Thus the space-time codes from algebraic constellations
verify the product criterion. As for the determinant criterion, it is verified as long
2
as the conjecture of having δmin maximized for the same rotation that maximizes
dP,min is true.
Thus, ΘM space-time codes are proved to be suitable for both fast and quasi-static
fading channels.
Remark 5.3 Notice that the construction criteria over quasi-static and over fast
fading are independent; i.e. one can find a code that realizes the quasi-static fading
criteria but not the fast fading one, and vice versa. For example, consider the following
matrices
1 1
B1 =
,
0 1
and
1 2
B2 =
.
1 2
The rank of B 1 is 2, but the number of positions where its columns are different is 1.
While this number equals 2 for the matrix B 2 , its rank is 1. Consider the ST code G2
with the modulation BPSK. G2 verifies the rank criteria, but the number of positions
where the columns of its matrix B (see Section 5.2) are different is equal to 1. If
we consider the codewords X = [x1 , x2 ], and E = [x2 , x1 ] then the matrix B in this
codewords pair equals to
x1 − x2 x∗1 − x∗2
B(X, E) =
,
x2 − x1 x∗1 − x∗2
112 ENST
5.4. CONCATENATING SPACE-TIME CODES WITH OUTER SUBSTREAM CODES
0 0
10 10
uncoded 4 bits/s/Hz uncoded 6 bits/s/Hz
algebraic code 4 bits/s/Hz algebraic code 6 bits/s/Hz
−1
10
−1
10
−2
10
Average symbol error rate
−3
10
−3
10
−4
10
−4
10
−5
10
−6 −5
10 10
10 15 20 25 30 35 10 15 20 25 30 35
SNR per bit SNR per bit
Figure 5.21: Θ2 ST code with 4-QAM Figure 5.22: Θ3,2 ST code with 4-QAM
over a fast fading channel N = 1. over a fast fading channel, N = 1.
In Fig. 5.21 is depicted the symbol error rate of Θ2 ST code using the 4-QAM
modulation over fast fading channel. Fig. 5.22 shows the symbol error rate of Θ3,2
ST code using the 4-QAM modulation over fast fading channel. It is noticed in both
figures that the performance is enhanced over the fast fading compared to quasi-static
fading (Fig.’s 5.13, 5.16).
Fig. 5.23 shows the average BER of Θ4,2 and Θ4,3 ST codes with BPSK modulation
over a fast fading channel, for N = 1. Notice that one obtains more enhancement of
the performance over fast fading when l increases.
Consider a double coded system by outer substream codes and inner inter-substream
space-time ΘM codes (Fig. 5.24). At the receiver inter-substreams coded system is
replaced by its equivalent uncoded system (5.3.4), then iterative decoding is performed
ENST 113
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
0
10
uncoded 4 bits/s/Hz
algebraic code,M=4,l=2
algebraic code M=4,l=3
−1
10
−2
10
Average bit error rate
−3
10
−4
10
−5
10
−6
10
10 12 14 16 18 20 22 24 26 28 30
SNR per bit
Figure 5.23: Θ4,2 and Θ4,3 ST codes with BPSK over a fast fading channel, N = 1.
TX
in 2x2 Space-Time matrix
Rotation and arrangement
TX
114 ENST
5.4. CONCATENATING SPACE-TIME CODES WITH OUTER SUBSTREAM CODES
on the coded substreams as in 5.1.2. The interleaver size is 5 times the outer codeword
length, a period on which the Rayleigh fading is fixed, then changes randomly. The
number of iterations is chosen to be 3, since after the third iteration, no substantial
improvement in performance occurs. Power computation is done like in (5-9) where
Rout is replaced by Rout · Rin the product of the inner and outer codes.
0 0
10 10
2 coded substreams without ST code Two coded substreams without ST code
4 coded substreams with ST code Four coded substreams with ST code
−1
10
−1
10
Average Bit Error Rate
−2
10
−3
10
−3
10
−4
10
−4 −5
10 10
0 2 4 6 8 10 12 0 2 4 6 8 10 12 14 16
Eb/N0 in dB EbN0 in dB
In Fig. 5.25 average BER is plotted as a function of Eb/N0 for the outer code
BCH(16, 5), the decoding algorithm is the Fossorier-Lin with first order reprocessing.
The ST code Θ2 is compared to a system without space-time code, at the same spectral
efficiency. Iterative decoding is also used for the second system.
Fig. 5.26 shows the performance of the outer code is BCH(32, 6). For both outer codes,
the diversity order is multiplied by 2 when using the ST code Θ2 .
In this subsection the performance of the concatenated system is shown over a pure
fast fading and then over a “realistic” fading model.
ENST 115
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
We present in Fig. 5.27 the average BER of the concatenated system using Θ2 as an
inner code, and Golay(24, 12) as an outer code for all substreams with M = N = 2.
The channel changes randomly every time instant t. This iterative scheme over a pure
fast fading is compared to the scheme with the same components except the space-
time code. The number of iteration is 4. The independent random interleaver length
is 5 × 12 = 70 symbol periods.
Channel model since the reality is somewhere between the two extreme cases of
fast and quasi-static fading, we use the following model to fit the real fading channel
(without Doppler effects).
“Realistic” fading model: if the outer code is C(n, k) then the fading are supposed
to be fixed over an outer codeword n bits (n/2 symbols for the 4-QAM modulation),
then it changes randomly. The interleaver size is 5 times the outer codeword length
(5n/2 symbols).
Now, random interleavers associated with the modulators outputs over the Rayleigh
fading channel have two different roles:
Break error paths (words in the case of block codes) resulted from substreams de-
coders.
116 ENST
5.5. CONCLUSIONS
0
0 10
10
1st symbol
without ST code, iteration 1 2nd symbol
without ST code, iteration 4 3rd symbol
with ST code, iteration 1
4th symbol
with ST code, iteration 4
−1
−1 10
10
average BER
BER
−2 −2
10 10
−3 −3
10 10
0 1 2 3 4 5 6 0 1 2 3 4 5 6 7 8
SNR per bit SNR per bit
Figure 5.27: Concatenated system over Figure 5.28: Concatenated system over
a pure fast fading channel. Outer code a realistic fading channel. Outer code
Golay(24, 12), inner code Θ2 , N = 2. Golay(24, 12), inner code Θ2 , N = 2.
5.5 Conclusions
In this Chapter, we have studied a multi-antenna system over a quasi-static Rayleigh
fading channel, in four different cases
4. Concatenated codes.
In the first case we commented on the limitations of the well known V-BLAST detec-
tion algorithm. Although it succeeds in separating the different substreams, and thus
reaches an astonishing throughput, it does not fully exploit the receive diversity. For
this purpose, iterative decoding schemes of coded substreams give better performance
than a detector cascaded by decoders. In the first two schema the inherent transmit
diversity of the multi-antenna system was not exploited. To do so, space-time codes
transform transmit diversity into time diversity.
ENST 117
CHAPTER 5. SPACE-TIME CODES FOR MULTI-USER APPLICATIONS
We have presented a new class of bandwidth lossless space-time codes from algebraic
constellations. These codes realize the rank and determinant criteria of the space-time
codes design for multi-dimensional QAM constellations. Their performance is indepen-
dent from the size of the original QAM (4, 16, 32, . . .), thus very high throughput can
be reached in a system that fully exploits receive and inherent transmit diversity. The
new class of ST codes is proved to be suitable for both fast and quasi-static fading.
The drawback of these codes is the complexity of their ML decoding: to remedy this
last problem we have studied the concatenated systems, and their iterative decoding
schemes that approach the ML performance at high SNR for reasonable complexity
(the number of iterations times single substream decoder complexity).
Finally, we have shown simulations over a pure fast fading channel and over a realistic
Rayleigh fading channel of the concatenated system, where the expected improvement
was presented and explained by the combination of independent types of diversity.
The obtained performance/SNR is very suitable for real wireless communications with
power constraint and mobility effect.
The multi-antenna system and the algebraic ST codes are akin to lattice codes, thus
applying the universal lattice decoder suggests itself. This problem is examined in the
next Chapter.
118 ENST
Chapter 6
119
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
6.1.1 Definitions
120 ENST
6.1. LATTICE REPRESENTATION OF MULTI-ANTENNA ARCHITECTURE
The independence of the receive antennas and the affecting fades of each substream
presupposes that the transfer matrix H (5-1) has a full rank almost always; i.e. the
event of having two or more dependent columns in H is negligible with respect to the
probability measure.
r = HaT + ν, (6-3)
where a = (a1 , a2 , . . . , aM ) denotes the transmitted vector which belongs here to the
constellation QAM carved from Z(i). In the above equation ν is an N × 1 complex
column vector AWGN component-wise independent with a variance σ 2 per dimension.
Equivalently, one can write the system (6-3) as
r
= [(r T ) (r T )]
(H ) T
(H )
T
= u + ν
−(H T ) (H T )
= uM H + ν , (6-4)
The proof is a direct result from the properties of matrix product. Hence, we can
represent the multi-antenna environment by a lattice sphere packing, and one can apply
the universal lattice decoder in a multi-antenna system.
ENST 121
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
15
6
10
4
2 5
0 0
−2 −5
−4
−10
−6
−15
−5 0 5 −10 −5 0 5 10
10 10
5 5
0 0
−5 −5
−10 −10
−10 −5 0 5 10 −20 −10 0 10 20
Suppose that we have a lattice code carved from the n-dimensional lattice Λ defined
by its generator matrix M Λ . Over the Gaussian channel the received signal is given
by
r = uM Λ + ν, (6-6)
122 ENST
6.2. THE SPHERE DECODING ALGORITHM
where x = uM Λ . The lattice decoding algorithm enables to find the closest point of
the lattice constellation in terms of the metric (6-7) and thus realizes the performance
of the ML detection for real dimensions up to 32 [101], a set of dimensions for which
the exhaustive search becomes infeasible.
The principle of the algorithm is to search the closest lattice point to the received
√
signal within a sphere of radius C centered at the received signal (Fig. 6.2). The
algorithm starts by a ZF detection on the received signal by inverting the matrix M Λ :
ρ = rM −1
Λ .
ENST 123
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
The choice of C is very crucial to the speed of the algorithm. In practice, C can be
adjusted according to the noise (and eventually the fading) variance. When a failure
is detected, one can either declare an erasure on the detected symbol, or increase C.
The complexity of the algorithm is independent of the lattice constellation size, which
is very useful for high data rate transmission. In [29] Fincke and Pohst showed that if
d−1 is a lower bound for the eigenvalues of the Gram matrix GM , then the number of
arithmetical operations is
4dC
2 n−1
O n × 1+ , (6-8)
4dC
which, for a judicious choice of the radius C ≈ d−1 , is approximated by O (n6 ) arith-
metical operations.
Fig. 6.3 presents a flowchart of the sphere decoding algorithm. The function
Chol(GM ) computes a normalized Cholesky’s factorization of the Gram matrix GM
as follows
GM = RRT , (6-9)
The function x gives the smallest integer greater than x, and x gives the greatest
integer smaller than x.
In simulation we use the constellation q-QAM, with q = 4, 16. The average energy
124 ENST
6.3. SPHERE DECODING OF THE UNCODED MULTI-ANTENNA SYSTEM
- qGM MM T
:=
INPUT
:= Chol(G
r C M M ; 1 M)
= rM ; 1
?
2
d := C
Tn := C
S :=
?
i := n
Li :=
p?
bp Ti =qii + Si c
d; e
--
ui := Ti =qii + Si
# #
?
" !"# !
i := i + 1 ui := ui + 1
no
6 ? i := i ; 1
6P
i = n? ui > Li ?
Si ;1 := ;1 =+
n
qi;1 j j
" !
i j i
yes
yes
? ?no 6
;
OUPUT
u 2
^ d^
i > 1?
yes
- Ti;1 := ? T
i := i
; qii (Si
ui
; 2
ui ) > 0
#
i
2
d^ := Tn
?
; 1 11
T +q
no
(S1 ; u1 )2
" !
? u^ := u
no
2
d^ < d ?
2
yes
- 2
d := d^
Tn := d^
2
2
ENST 125
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
where Esav is the average symbol energy of the q-QAM when Eb = 1. For example,
for q = 4 and q = 16, Esav equals 2 and 10, respectively.
0 0
10 10
V−BLAST optimal order V−BLAST optimal order detection
Sphere decoding Sphere decoding
−1
10
−2 −1
10 10
Average symbol error rate
−3
10
−4 −2
10 10
−5
10
−6 −3
10 10
10 11 12 13 14 15 16 17 18 6 7 8 9 10 11 12
SNR per bit SNR per bit
Fig. 6.4 presents the average symbol error rate of a scheme of 8 transmit antennas
and 8 receive antennas using the 16-QAM modulation. A symbol error rate of 10−5
is attained at SNR ≈ 17.8 dB for a total throughput of 32 bits/s/Hz. It is worth
mentioning that, in this case, ML detector by exhaustive search requires 168 (8 × 8)
matrix multiplications and 168 (1 × 8) vector norm computations, which is infeasible.
3
This normalization is useful to have a lattice representation by transmitting symbols from Z(i).
However, if the amplitudes of the transmitted symbols are not integers, one can always perform the
normalization at the receiver.
126 ENST
6.4. SPHERE DECODING OF ALGEBRAIC SPACE-TIME CODES
Fig. 6.5 shows the block error rate of sphere decoding and V-BLAST detection using
a 4-QAM modulation over a quasi-static Rayleigh fading, with a burst length of 100
symbols.
The lattice sphere packing representation of the ST code ΘM,l is done in two steps
2. Representing the resulted uncoded system by its lattice form (6-4): n = 2lM.
h11 h12 y1 −y3
r=
+ ν,
(6-12)
h21 h22 y2 y4
4
see Section 5.3.
ENST 127
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
r [(r T ) (r T )]
=
T T
(H ) (H )
= u + ν
T T
−(H ) (H )
= uM H + ν , (6-15)
where u = [(a) (a)] ∈ Z2lM , and ν = [(ν T ) (ν T )] ∈ R2lM . The dimension
increase for the lattice representation is M = 2 −→ 2lM = 8.
One can alleviate the computational complexity in the case of ST codes ΘM,l by
precomputing the inverse of the corresponding rotation matrix (see Chapter 5), and by
expressing the inverse of the transfer matrix of the corresponding equivalent uncoded
system by the inverse of the transfer matrix of the ST coded system. In the Example
6.1 one has
H −1 = M 4 −1 H −1
hi11 0 hi21 0
hi21 0 hi22 0
−1
= M 4 , (6-16)
0 hi11 0 hi21
0 hi21 0 hi22
128 ENST
6.4. SPHERE DECODING OF ALGEBRAIC SPACE-TIME CODES
The variance of the transmitted symbols, the transfer matrix and the AWGN are
computed like in Section 6.3, because the ST code ΘM,l does not add redundancy.
Comparisons are done with uncoded systems of (M, N) transmit-receive antennas.
The limitation on n implies that lM ≤ 16.
0 0
10 10
−1 −1
10 10
−2 −2
10 10
Average symbol error rate
−3 −3
10 10
−4 −4
10 10
−6 −6
10 10
0 5 10 15 20 25 30 0 2 4 6 8 10 12 14 16 18
SNR per bit SNR per bit
Fig. 6.6 shows the average symbol error rate of SD and V-BLAST optimal order
detection for the uncoded system M = N = 2, and for the algebraic ST code Θ2 . In
both schema 16-QAM constellation were used. One notices that the SD decoding gives
the ML detection results (see Fig. 5.18).
Fig. 6.7 presents the average symbol error rate of sphere decoding and V-BLAST
detection of 4 substreams using the modulation 4-QAM in a scheme of 4 receivers, and
ENST 129
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
total throughput of 8 bit/s/Hz. It also shows the SD of the ST code Θ4,2 and Θ4,3 of
the same scheme with the same throughput over a quasi-static fading of length l = 2
and l = 3 symbols . One notices that the slope of the sphere decoding curve is close to
4 (within 10 dB, the curve decreases 4 orders of the power of 10: at 8 dB ≈ 210−2 , and
at 18 dB ≈ 6.5e−6 ) which is the diversity expected by the ML detection upper bound
[90]. The slopes of the ST codes curves are greater than that of SD of the uncoded
system: the scheme of ST code ΘM,l has a total diversity of lN.
yields
r 1 = u1 M 1 − u2 M 2
r 2 = u1 M 2 + u2 M 1 , (6-17)
1.
r sd1 = (r 1 + û2 M 2 )M T1 + (r 2 + û2 M 1 )M T2 . (6-18)
3.
r sd2 = (−r 1 − û1 M 1 )M T2 + (r 2 + û1 M 2 )M T1 . (6-19)
5. Iterate.
130 ENST
6.5. SUB-OPTIMAL SCHEME, SPHERE DECODING WITH SIC
The maximum ratio combiner done in (6-18) and (6-19) by optimal combination of
r 1 and r 2 aims to allow the sphere decoder to exploit the receive diversity after the
SIC operation. The complexity of the sub-optimal algorithm is O ((n/2)6 ) arithmetical
operations5 , which permits decoding of schemes with complex dimensions up to 32.
One can assign soft decisions to the hard detected symbols as a by-product of this
sub-optimal algorithm as follows:
The probability of the symbol aj = s is set to the number of times aj = s during the
iterations over the total number of iterations6 .
0
10
−1
10
Average symbol error rate
−2
10
−3
10
−4
10
0 5 10 15 20 25 30
SNR per bit
Fig. 6.8 shows the performance of the proposed sub-optimal scheme over a multi-
antenna system with 4 transmitters and 4 receivers using the 4-QAM modulation. We
notice that one gains almost 5 dB after the second iteration compared to V-BLAST
optimal order detection alone. However, the sub-optimal scheme does not increases the
diversity order exploited by the receiver. This small improvement is logical because
when performing a SIC operation on uncoded substreams (without outer codes), an
erroneous detected data will propagate from one iteration to another (see Chapter 3).
5
The initialization by V-BLAST demands O (n/2)4 arithmetical operations, besides it depends
linearly on the constellation size.
6
This method is akin to the one used in the Gibbs algorithm [34].
ENST 131
CHAPTER 6. SPHERE DECODING OF SPACE-TIME CODES
Remark 6.1 This SIC method is similar to the reprocessing method combined with
V-BLAST proposed in Section 5.1 (subsubsection Improving on V-BLAST). We have
seen there that one obtained little improvement by the proposed method for small di-
mensions because we were limited in the dimension by the ML detection. The difference
here is that we are less limited in dimensions.
6.6 Conclusions
We have studied in this Chapter the sphere decoding of the V-BLAST multi-antenna ar-
chitecture over the Rayleigh fading channel, and then the application of this algorithm
to decode the algebraic ST codes ΘM,l . The algorithm reaches the ML performance
with a computational complexity of O (n6 ). It is independent of the size of the con-
stituent QAM constellation. Hence, a very high throughput could be reached along
with ML performance.
The huge difference in performance between the sphere decoding and the V-BLAST
detection algorithm [39, 104] over the uncoded system, shows the limitations of the lat-
ter algorithm in taking advantage of the receive diversity, and allows one to appreciate
the huge capacity of the multi-antenna systems [91].
The sphere decoding solves the intricate problem of decoding the algebraic ST codes
ΘM,l up to dimensions lM ≤ 16. It was shown by the simulation results and the
properties of these codes (see Chapter 5) that one can verify the full diversity in a
multi-antenna system without adding any redundancy, and perform the ML decoding
with low computational complexity.
The combination of sphere decoding with a SIC based on maximum ratio combiner
allows one to increase the complex dimensions lM up to 32 for a reasonable complexity.
Besides, one can assign soft values to the decisions taken by the sub-optimal SIC algo-
rithm. Although the performance is largely degraded compared to the ML detection,
this scheme is still interesting in high dimensions since it outperforms the classical SIC
algorithm like the V-BLAST.
132 ENST
Chapter 7
M ultiple access problem concerns the efficient use of a common transmission medium
by several users. Of the four principle problems in this scheme, Namely: the
channel impairment, the transmission delay, the BER of each user, and the average
overall amount of transmitted information; the present thesis focuses on the last two
issues.
After outlining the characteristics of the multiple access channel, and reviewing the
information theory viewpoint, we have studied an iterative decoding scheme for the
uncoded system first, then for coded systems. The convergence of the iterative method
was shown by simulations for both convolutional and block codes.
We provided elements of analysis of the block coded two users system over the Gaus-
sian channel. It shows how the convergence of the iterative method is done: by using
the “useful” information of the most “reliable” decoded bits, and by neglecting the
“useless” information of the erroneous decoded bits.
Studying the iterative method over the Rayleigh fading channel illuminates the role of
independent random interleavers associated with the encoders outputs in preventing
the error paths from permeating among the users in the iterative decoding process,
and by diminishing the number of ambiguity situations over the Gaussian channel.
133
CHAPTER 7. CONCLUSIONS AND PERSPECTIVES
To remedy the last problem, we have studied joint multi-user coding, which sup-
poses that each transmitter takes into consideration the existence of other users on the
medium. All the active transmitters are designed to minimize their mutual interference
in order to coexist and to transmit at high data rates.
We have proposed a new class of multi-user codes based on rotated multi-dimensional
constellations. Over the noiseless adder channel, the proposed code is proved to be
uniquely decodable, without adding any redundancy. Thus, the proposed class of codes
achieves the capacity of the noiseless adder channel. We have proposed a sub-optimal
detection algorithm which profits from the algebraic structure of the code. The alge-
braic detection succeeds in separating the multitude of users over the noiseless adder
channel. The main characteristics of the considered code are: the simplicity of coding;
the symmetry of the residual MAI to all users; it is codeword and frame asynchronous.
Over the noisy channel, minimizing the MAI, by maximizing the minimum squared
Euclidean distance of the constellations sum, outperforms TDMA in most cases. The
concatenation with error control codes with very low rate further enhances the perfor-
mance of the system.
The concatenation with outer codes, and the application of an adapted version of
134 ENST
7.2. PERSPECTIVES
7.2 Perspectives
Of course, some open problems remain:
ENST 135
CHAPTER 7. CONCLUSIONS AND PERSPECTIVES
– The efficient decoding of the K-user codes in the presence of noise (modified
sphere decoding [27]).
– Coded modulation in the K-user code; an extensive study of the influence
of the concatenated outer code on the K-user code.
• Multi-antenna systems
136 ENST
Bibliography
[1] “ETSI-GSM-05.03.”
[8] J.-C. Belfiore, X. Giraud, and J. Rodriguez, “Linear Labeling for Joint Source
Channel Coding,” submitted to IEEE Trans. on Information Theory, 2000. See
also the Proc. of ISIT’2000, Sorrento, Italy, June 2000.
137
BIBLIOGRAPHY
[9] S. Benedetto and G. Montorsi, “Unveiling Turbo Codes: Some Results on Parallel
Concatenated Coding Scheme,” IEEE Trans. Inform. Theory, vol. 42, pp. 409–
428, Mar. 1996.
[14] K. Boullé and J.-C. Belfiore, “Modulation Schemes Designed for the Rayleigh
Fading Channel,” in CISS’92, Mar. 1992.
[15] J. Boutros, E. Viterbo, C. Rastello, and J.-C. Belfiore, “Good Lattice Constel-
lations for both Rayleigh and Gaussian channel,” IEEE Trans. Inform. Theory,
vol. 42, pp. 502–518, Mar. 1996.
[16] J. Boutros and E. Viterbo, “Signal space diversity: A power and bandwidth effi-
cient diversity technique for the Rayleigh fading channel,” IEEE Trans. Inform.
Theory, vol. 44, pp. 1453–1467, July 1998.
[17] J. J. Boutros, Lattice Codes for Rayleigh Fading Channels. PhD thesis, École
Nationale Suérieure des Télécommunications (ENST) de Paris, May 1996.
[18] L. Brunel and J. Boutros, “Euclidean Space Lattice Decoding for Joint Detection
in CDMA Systems,” in ITW’99, (Kruger National Park, South Africa), June
1999.
[20] S.-C. Chang and J. Edward J. Weldon, “Coding for T-User Multiple-Access
Channels,” IEEE Trans. Inform. Theory, vol. 25, pp. 684–691, Nov. 1979.
138 ENST
BIBLIOGRAPHY
[21] D. Chase, “A class of algorithm for decoding block codes with channel measure-
ment information,” IEEE Trans. Inform. Theory, vol. IT-18, pp. 170–182, Jan.
1972.
[22] S. Chia, “The Universal Mobile Telecommunication System,” IEEE Comm. Mag.,
vol. 30, pp. 54–62, Dec. 1992.
[23] J. Conway and N. Sloane, Sphere Packing, Lattices and Groups, 3rd ed. Springer-
Verlag, NY, 1998.
[26] G. da Silva and F. Assis, “Using space-time TCM with rotated constellations for
transmission over mobile communications channels,” Electronics Letters, vol. 35,
pp. 454–456, Mar. 1999.
[29] U. Fincke and M. Pohst, “Improved methods for calculating vectors of short
length in a lattice, including a complexity analysis,” Mathematics of Computa-
tion, vol. 44, pp. 463–471, Apr. 1985.
ENST 139
BIBLIOGRAPHY
[32] M. P. C. Fossorier and S. Lin, “Soft-decision decoding of linear block codes based
on ordered statistic,” IEEE Trans. Inform. Theory, vol. IT-41, pp. 1379–1396,
Sept. 1995.
[34] G. Gasella and E. George, “Explaining the Gibbs sampler,” The American Statis-
cian, vol. 46, pp. 167–174, 1992.
[36] X. Giraud, E. Boutillon, and J.-C. Belfiore, “Algebraic Tools to Build Modulation
Schemes for Fading Channels,” IEEE Trans. Inform. Theory, vol. 43, pp. 938–
952, May 1997.
[37] X. Giraud and J.-C. Belfiore, “Constellation Matched to the Rayleigh Fading
Channel,” IEEE Trans. Inform. Theory, vol. 42, pp. 106–115, Jan. 1996.
[38] D. L. Goekel, Performance Limits and Optimal Resource Allocation for Coded
Multi-User Communication Systems. Technical report, University of Michigan,
Ann Arbor, Michigan 48109, Aug. 1996.
[41] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Product. Aca-
demic Press, Inc, 1980.
[42] J.-C. Guey, M. P. Fitz, M. R. Bell, and W.-Y. Kuo, “Signal design for transmit-
ter diversity wireless communication systems over Rayleigh fading channels,” in
Proc. of VTC’96, Atlanta, GA, Apr. 1996.
[43] J. Hagenauer, E. Offer, and L. Papke, “Iterative Decoding of Binary Block and
Convolutional Codes,” IEEE Trans. Inform. Theory, vol. 42, pp. 429–445, Mar.
1996.
140 ENST
BIBLIOGRAPHY
[46] A. A. Hassan and J.-C. Belfiore, “Enhancing the spectrum efficiency of radio
cellular systems,” in DSPCS’99, (Perth, Australia), Feb. 1999.
[47] B. L. Hughes and A. B. Cooper, “Nearly Optimal Multiuser Codes for the Binary
Adder Channel,” IEEE Trans. Inform. Theory, vol. IT-42, pp. 387–398, Mar.
1996.
[48] N. Ibrahim and G. K. Kaleh, “Iterative Decoding and Soft Interference Can-
cellation for the Gaussian Multiple Access Channel,” in URSI-ISSSE, vol. 1,
(Piza-Italy), pp. 156–161, Sept. 1998.
[49] N. Ibrahim and G. Kaleh, “A Turbo Coding Scheme for Spread-Spectrum Sig-
naling,” in ICT’98, vol. 1, June 1998.
[50] A. B. C. III and B. L. Hughes, “Coding for Improved Tactical Channel Sharing,”
in 19th Army Science Conference, June 1994.
[53] T. Kasami and S. Lin, “Coding for a Multiple-Access Channel,” IEEE Trans.
Inform. Theory, vol. IT-22, pp. 129–137, Mar. 1976.
[56] T.-W. Lee, M. S. Lewicki, M. Girolami, and T. J. Sejnowski, “Blind Source Sep-
aration of More Sources Than Mixtures Using Overcomplete Representations,”
IEEE Signal Processing Letters, vol. 6, pp. 87–90, Apr. 1999.
ENST 141
BIBLIOGRAPHY
[59] H. H. J. Liao, Multiple access channel. PhD thesis, Dep. Elec. Eng, University
of Hawaii, Honolulu, 1972.
[60] S. Lin and J. D.J. Costello, Error Control Coding Fundamentals and Applications.
Englewood Cliffs, NJ: Prentice-Hall, 1983.
[65] M. Mouly and M.-B. Pautet, The GSM System for Mobile Communications.
Michel Mouly and Marie-Bernadette Pautet, 1992.
[66] D. Muder, “Minimal trellises for block codes,” IEEE Trans. Inform. Theory,
vol. 34, pp. 1049–1053, Sept. 1988.
142 ENST
BIBLIOGRAPHY
[73] R. Podemski, Analyse Spectrale des Turbo-Codes. PhD thesis, École Nationale
Supérieure des Télécommunications (ENST) de Bretagne, 1998.
[74] M. Pohst, “On the computation of lattice vectors of minimal length, successive
minima and reduced basis with applications,” in ACM SIGSAM, vol. 15, pp. 37–
44, Bull, 1981.
[79] P. Robertson, “Illuminating the Structure of Code and Decoder of Parallel Con-
catenated Recursive Systematic (Turbo) Codes,” in Globcom ’94, vol. 3, Nov.
1994.
ENST 143
BIBLIOGRAPHY
[90] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for high data
rate wireless communications: Performance criterion and code construction,”
IEEE Trans. Inform. Theory, vol. 44, pp. 744–765, Mar. 1998.
[92] G. Ungerboek, “Channel Coding with Multilevel Phase Signal,” IEEE Trans.
Inform. Theory, vol. IT-28, pp. 55–67, Jan. 1982.
144 ENST
BIBLIOGRAPHY
[97] S. Verdú, Multiuser Detection. The Pitt Building, Trumpington Street, Cam-
bridge, CB2 IRP United Kingdom: The press syndicate of the university of
Cambridge, 1998.
[98] A. J. Viterbi, “Very Low Rate Convolutional Codes for Maximum Theoretical
Performance of Spread-Spectrum Multiple-Access Channels,” IEEE J. Selec. Ar-
eas on Commun., vol. 8, pp. 641–649, May 1990.
[101] E. Viterbo and J. Boutros, “A universal lattice code decoder for fading channel,”
IEEE Trans. Inform. Theory, vol. 45, pp. 1639–1642, July 1999.
[103] J. H. Winters, J. Salz, and R. D. Gitlin, “The Impact of Antenna Diversity on the
Capacity of Wireless Communication Systems,” IEEE Trans. Commun., vol. 42,
pp. 1740–1751, Feb. 1994.
ENST 145
Index
a posteriori, 3, 9, 10, 32, 33, 46, 47, 87 capacity of multi-antenna system, 19,
a priori, 31–33, 77 132, 134
capacity of TDMA, 17
algebraic, 69 capacity of the Gaussian multiple ac-
algebraic combination, 69 cess channel, 14
algebraic constellation, 81, 112, 118, capacity of the noiseless K-user adder
134 channel, 14
algebraic detection, 75, 79, 134 capacity of the single user Gaussian
algebraic dimension, 76 channel, 15, 16
algebraic equation, 75 capacity region, 15, 18
algebraic manipulation, 44, 74 outage capacity, 20
algebraic norm, 70, 94 CDMA, 17, 18, 24, 49, 122, 136
algebraic number, 69, 73, 75, 76 cellular, 1, 16, 17, 58, 59, 80
algebraic number field, 69–71, 76, 81, central limit theorem, 7, 16, 44, 49
91, 93, 94, 96, 134 channel, 5, 25, 47, 57, 58, 60, 79–81, 83,
algebraic number theory, 69 86, 90, 97, 110, 111, 116, 133–135
algebraic ST codes, 107, 119, 127, 132 binary adder channel, 79
algebraic structure, 10, 58, 134 channel code, 3
algebraic tools, 60, 135 channel transition probability, 31, 33
ambiguity, 26, 27, 29, 30, 36, 42, 43, 45–47, Gaussian channel, 21, 23, 25, 26, 28,
61, 133 30, 31, 33, 36, 38, 40, 42, 43, 47,
AWGN, 5, 7, 16, 25–27, 30, 38, 42, 47, 83, 60, 77, 122, 133
87, 121, 126, 129 mobile-radio channel, 7
multi-path channel, 86
BER, 1, 9, 16, 33, 38, 39, 42, 43, 49, 78, multiple access channel, 5, 133
89, 90, 104, 115, 116, 133 noiseless adder channel, 5, 43, 58, 60,
block codes, 3, 10, 23, 25, 36, 38, 39, 43, 69, 71–74, 79, 134
47, 85, 116, 133, 134 noisy channel, 47, 57, 58, 78–80, 134
Rayleigh fading channel, 3, 8, 20–23,
capacity, 14, 15, 20, 49, 57, 58, 60, 79, 134 25, 27, 29, 36, 47, 60, 76, 82, 85,
capacity of CDMA, 18 96, 104, 107, 110–113, 116–118, 121,
capacity of correction, 43, 46 122, 133
capacity of FDMA, 17 rich scattering channel, 10, 110
146
INDEX
complexity, 11, 13, 25, 26, 40, 46, 47, 49, Hadamard, 67, 96, 98–101
72, 74, 75, 77, 80, 85, 110, 118, Hamming, 57, 58
119, 122, 124, 127, 128, 131–133,
information bits, 32, 38
135
information theory, 3, 13, 14, 58, 60, 133
convolutional codes, 3, 23, 25, 30, 33, 36,
interference, 1, 3, 7, 8, 22, 43, 78, 87, 90,
38, 43, 46, 47, 78, 133
134
cooperation, 1, 14, 18
ISI, 22
CSI, 9, 10, 86, 90, 110, 123
MAI, 22, 25, 134
delay, 1, 9, 133, 135 RIL, 59, 76, 79
diversity, 8, 11, 60, 85, 111, 116, 118, 130 interleaver, 3, 30, 33, 36, 38, 43, 46, 47, 78,
diversity gain, 9, 91–93 86, 87, 115, 116, 133
diversity order, 9, 21, 96, 104, 109, 115, iterative decoding, 3, 10, 11, 22, 25, 27, 33,
131 36, 40, 42, 47–49, 81, 86, 89, 90,
full diversity, 90, 95, 110, 132, 134 110, 113, 115, 117, 118, 133–135
receive diversity, 10, 20, 21, 23, 84,
lattice, 70, 119, 120, 123, 124
110, 116, 117, 131, 132
lattice code, 120, 122
spatial diversity, 76, 97, 111
lattice decoder, 121, 123
time diversity, 3, 20, 81, 92, 116, 117
likelihood, 26–28, 39, 40, 44–46
transmit diversity, 20, 21, 23, 24, 81,
92, 110, 116–118 MAP, 9, 30, 33, 47
transmit-receive diversity, 21 mapper, 4, 77
unequal diversity scheme, 102, 104 maximum ratio combiner, 11, 84, 90, 93,
130–132, 135
error paths, 36, 44, 46, 86, 116, 133
minimum distance, 38, 39, 58, 59, 63, 120
error probability, 1, 8, 20, 28, 29
ML, 9, 11, 22, 23, 25, 26, 39, 43, 74, 75,
fading, 3, 7, 8, 10, 19, 25, 27–29, 36, 47, 80, 77, 85, 90, 104, 110, 118, 123, 126,
83, 84, 91, 96, 111, 115, 116, 118, 129, 132
124 MMSE, 9, 82, 83
fast fading, 20, 97, 110–113, 115, 116, mobile-radio, 1, 7, 9, 17, 58
118, 134 modulation, 4, 38, 82, 93, 102, 103
flat fading, 8, 19 BPSK, 4, 25, 26, 28, 31, 38, 42, 59, 60,
quasi-static fading, 8, 19, 20, 82, 85, 104, 107, 112
86, 97, 104, 107, 111–113, 116–118, coded modulation, 136
130, 134 PAM, 4, 28
FDMA, 17 PSK, 4, 62, 92
Fossorier-Lin, 10, 11, 38, 44, 47, 85, 89, QAM, 4, 59–61, 63, 67, 74, 85, 87, 89,
115, 134 93, 94, 104, 107, 113, 116, 126, 131
multi-antenna, 2, 10, 14, 19, 80, 81, 84, 86,
Gram matrix, 120, 121, 123, 124 90, 117, 119, 121, 132, 134–136
ENST 147
INDEX
multi-user, 1, 2, 11, 22, 24–26, 35, 40, 42, transmitted information, 1, 9, 49, 133, 135
57, 78, 82, 102, 107 turbo code, 13, 16, 22, 78, 80, 135
multi-user code, 3, 11, 23, 26, 134
UMTS, 17, 136
multi-user coding, 10, 23, 58, 134
uncoded system, 25, 26, 82, 104, 109, 111,
multi-user decoding, 10, 32, 36, 78, 80,
113, 119, 127, 128, 133
134, 135
uniquely decodable, 26, 57–59, 73–75, 78–
multi-user detection, 15, 22, 25, 49,
80, 134
133
multiple access, 1, 16, 17, 33, 40, 76, 78 ZF, 9, 82, 83, 123
SIC, 15, 22, 30, 32, 47, 81–86, 104, 131, 132
single user, 1, 10, 26, 29, 32, 33, 36, 38, 42,
43, 46, 47, 64, 77, 78, 82, 134
SNR, 11, 15, 20, 25–27, 33, 38–40, 42, 43,
47, 60, 83, 84, 89, 109, 116, 118,
126
soft values, 11, 38–40, 44, 84, 134, 135
space-time, 19, 24, 81, 94, 96
ST codes, 10, 11, 21, 23, 81, 82, 90–94,
97, 104, 107, 109–113, 115–119, 127,
134–136
sphere decoding, 11, 110, 119, 122, 127,
130, 135
sub-optimal, 10, 26, 46, 49, 130–132, 134
synchronization, 18, 25, 47, 80, 82
system model, 1, 59, 121
148 ENST