0 évaluation0% ont trouvé ce document utile (0 vote)

40 vues162 pagesJun 26, 2011

© Attribution Non-Commercial (BY-NC)

PDF, TXT ou lisez en ligne sur Scribd

Attribution Non-Commercial (BY-NC)

0 évaluation0% ont trouvé ce document utile (0 vote)

40 vues162 pagesAttribution Non-Commercial (BY-NC)

Vous êtes sur la page 1sur 162

Le grade de Docteur en Sciences de lcole Normale Suprieure de Cachan Spcialit : Traitement du signal

par

Mrouane Debbah

Soutenue le 15 octobre 2002 devant la commission dexamen Prsident Rapporteurs Examinateurs Pierre Duhamel Ezio Biglieri Giuseppe Caire Jean-Claude Belore Marc de Courville Pascal Larzabal Philippe Loubaton

Remerciements

Cette thse est le rsultat de trois annes de recherche au sein de Motorola Labs et du Laboratoire Systmes de communication de lUniversit de Marne La valle. Jadresse dans un premier temps tous mes remerciements aux rapporteurs de cette thse: Prof. Ezio Biglieri du Politechnico di Torino, pour qui jexprime la plus profonde admiration et dont les articles ont fortement inuenc ma vision de la thorie de linformation. Prof. Giuseppe Caire dEurecom dont le dynamisme et limagination nont eu de cesse dinspirer mes travaux. Je crois que cette thse naurait pu voir le jour sans laide prcieuse de mes deux directeurs de thse: Philippe Loubaton, pour toute sa patience, sa rigueur et son incroyable disponbilit couter tous ses thsards; je ne saurai oublier les heures de discussion passes dans ton bureau. Je tiens te remercier pour ton rle de directeur de thse modle. Marc de Courville, qui a su allier de faon exemplaire qualits scientiques et humaines. Je tiens te remercier pour toute ton aide et la gestion efcace de toutes mes linuxeries . De mme, je dois toute ma reconnaisance Walid Hachem, dont la rigueur et la culture scientique ont t dune grande aide durant ma thse. Et surtout, je tiens remercier toutes les personnes qui ont agrment ces trois annes de moments inoubliables: Samson et Walner, pour leur amiti et tous les fous rires que lon a eus; Christian (alias 21 et gestionnaire du campus), pour ton norme gentillesse; Jean-Louis pour ton dvouement et tous les plans de start-ups que tu proposes; les membres de lquipe Signal pour les communications pour leur amiti et leur accueil chaleureux. Je garde une norme admiration pour les membres de lquipe BSTL dont la qualit scientique et technique na pas dquivalent, Markus, mon binme et dont les conseils furent si prcieux, Laurent, mille merci pour ta disponibilit, Xavier pour ton dynamisme et bien sr Jean-Noel, Emilio, Selim, Patrick, Pietro, Alexandre, Vronique, Sbastien, Philippe, Jeremy, Sandro, Patrick Maill dont lencadrement de stage fut un vrai plaisir ansi que tous les autres membres de Motorola (dont la liste serait trop longue donner) savoir Nesrine, Sandrine, Soodesh, Vivien... Je remercie galement mes amis personnels, Nassim pour les soires de lhotel Delambre; Andr pour tous les bons moments passs, Ali, Farid et Hakim, mes amis denfance. Jadresse ma profonde gratitude Emmanule dont lapport scientique et moral a t si prcieux. Je te remercie pour ta comprhension et ainsi que ton humour Gad Malhien. Enn, je ddie cette thse mon pre, ma mre, ma soeur Imne et mes frres Amine et Mounir, Merci pour votre aide et votre soutien durant toutes ces annes. 29 aout 2002

ii

Rsum

Ce sujet de thse se situe dans le cadre de ltude des systmes mobiles de communication intra-batiments trs hauts dbits et sattache tudier des solutions bases sur des modulations LP-OFDM (Linear Preocded Orthogonal Frequency Division Multiplexing Code Division Multiple Access) connu encore sous le nom de MC-CDMA (Multi-Carrier Code Division Multiple Access) dans le contexte multi-utilisateur. Lune des difcults majeures pose par les systmes de transmission trs hauts dbits dans les batiments est due la selectivit en frquence du canal de propagation indoor. Ceci se traduit en pratique par la ncessit didentier et dgaliser un canal trs grand nombre de coefcients. Le grand avantage des modulations Multi-Porteuses est donc de remplacer letape dgalisation par des simples corrections de facteurs complexes scalaires grce lutilisation dun prxe cyclique et de Transformes de Fourier Rapides (FFT). Leur grand inconvenient est par contre li au fait que les symboles transmis au voisinage de frquences de FFT trs attnues par le canal sont potentiellement affects de taux derreur importants. Une solution dune nature un peu differente consiste modier le schma de modulation de faon transmettre plusieurs symboles autour de la meme porteuse: LP-OFDM. Le but de la thse a t dtudier les performances asymptotiques de tels schmas de transmission (quand le nombre N de K porteuses et le nombre K de symboles transmis augmentent mais le rapport N reste constant) sur des canaux ralistes, en supposant des stratgies de dcodages linaires et non linaires la rception. Ces rsultats asymptotiques, utilisant la thorie des probabilits libres, fournissent des outils thoriques trs intressants sur limpact de lorthogonalit de la matrice de prcodage ainsi que sur le choix optimal de redondance (rapport K ) dans des schmas cods. De faon surprenante, ces rsultats sinterprtent N facilement et dterminent les paramtres utiles considrer pour la mise en place de tels schmas.

Mots-Cls: OFDM, LP-OFDM, MC-CDMA, Probabilits Libres, Matrices Alatoires, Filtrage MMSE.

iii

Abstract

A multi-carrier OFDM (Orthogonal Frequency Division Multiplexing) system using a Cyclic Prex for preventing inter-block interference is known to be equivalent to a system based on multiple at fading parallel transmission channels in the frequency domain. In such a system, the information sent on some carriers might be subject to strong attenuations and could be unrecoverable at the receiver. The aim of this thesis is to investigate a transmission scheme known as Linear Precoded OFDM to overcome these drawbacks. Linear Precoding consists in spreading the information contained in a K-dimensional vector onto N carriers. Therefore , the K 1 infomation vector to be transmitted is multiplied by an N K matrix. The precoder is intended to increase the overall frequency diversity of the modulator so that unreliable carriers can still be recovered by taking advantage of the subbands enjoying a high signal to noise ratio. Using new tools, borrowed from the so-called free probability theory, the Signal to Interference plus Noise Ratio (SINR) at the output of linear and non-linear equalizers are derived when N and K/N assuming both ergodic and non ergodic channels. Based on the assumption of asymptotic Gaussian interference, the SINR is then used as a benchmark measure of performance for different criterias. These asymptotic results aim at adjusting the different parameters involved in the detection process : nature and impact of the orthogonality of the precoding matrix columns, optimal amount of redundancy of K the precoding matrix when coding is applied (ratio N , values of the coefcients of the non-linear equalizers). Interestingly, the results have very simple interpretations and determine the useful parameters to be considered in the design of these transmissions schemes. Finally, conclusions are drawn on the future research topics in connexion with our work.

Keywords: empirical eigenvalue distribution, FFT equalization, free probability, OFDM (Orthogonal Frequency Division Multiplexing), LP-OFDM, Linear precoding, MC-CDMA, Minimum Mean Square Error, random matrix, single carrier modulation.

iv

List of Symbols

ABBREVIATIONS ADC ADSL AWGN a.s BER CC CDMA COFDM CP DFE DMT ETSI IBI i.i.d. ISI FIR GI (I)FFT LP-OFDM MIMO ML MMSE MU MUD OFDM PAPR p.d.f. QPSK SC SNR SINR SIC SNR PIC ZF Analog to digital conversion Asymmetric digital subscriber line Additive white Gaussian noise Almost surely Bit error rate Convolutive coding Code Division Multiple Access Coded OFDM Cyclic prex Decision feedback equalizer Discrete Multitone Transmission European telecommunications standard institute Inter block interference Independent and identically distributed Inter symbol interference Finite impulse response Guard interval (inverse) Fast fourier transform Linear Precoded OFDM Multiple inputs multiple outputs Maximum likelihood Minimum mean square error Mobile unit Multi-user detection Orthogonal frequency division multiplexing Peak to average power ratio Probability density function Quadrature Phase shift keying Single Carrier Signal to Noise Ratio Signal to Interference plus Noise Ratio Successive Interference Cancellation Signal to noise ratio Parallel Interference Cancellation Zero forcing

vi OPERATORS (.)*(.) Convolution operator Multiplicative free convolution of measure and ( ) Complex conjuguation T () transposition operator ( )H Hermitian transposition operator Diag(X) Diagonal matrix whose diagonal elements are the components of vector X E(.) Statistical expectation operator NOTATIONS yN nN sK yk (n) WN,K HN hi FN T HISI HIBI D C R 2 r Cn received vector of size N 1 noise vector of size N 1 K-dimensional transmit vector kth Component of the N 1 received vector y N at time n Precoding matrix of size N K Diagonal frequency channel matrix of size N N Channel frequency response on the i th carrier Matrix performaing the Fourier Trasnform Sampling rate Inter-symbol interference channel matrix Inter-Block interference channel matrix size of the guard interval Complex numbers Real Numbers Noise variance n! r!(nr)! where n! is a Factorial.

vii

Contents

Remerciements 1 2 Introduction OFDM 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 OFDM principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 The pros and cons of OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . LP-OFDM 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 LP-OFDM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MMSE analysis of LP-OFDM 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 MMSE analysis of certain class of unitary matrices (K = N) 4.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 4.2.2 Uncoded scenario . . . . . . . . . . . . . . . . . . . 4.2.3 Coded scenario: Ergodic Channels . . . . . . . . . . 4.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . 4.3 MMSE analysis of certain class of matrices . . . . . . . . . 4.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Statement of the main results . . . . . . . . . . . . . 4.3.3 Numerical results . . . . . . . . . . . . . . . . . . . 4.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . 4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . Non-Linear Detection schemes for LP-OFDM 5.1 Introduction . . . . . . . . . . . . . . . . . . . 5.2 Successive Interference Cancellation Detector . 5.2.1 BLAST MMSE algorithm . . . . . . . 5.2.2 Performance issues . . . . . . . . . . . 5.3 Parallel Interference Cancellation Detector . . . 5.3.1 Iterated-Decision Multiuser Algorithm . 5.3.2 Performance issues . . . . . . . . . . . 5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i 1 15 15 15 24 27 27 27 29 39 39 41 41 42 48 60 61 61 62 67 80 80

81 . 81 . 82 . 82 . 83 . 90 . 90 . 91 . 100

viii

CONTENTS

A Useful Mathematical results 105 A.1 Probability theory results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 B Useful results on random matrix theory 109 B.1 Random matrix theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 B.2 An application example: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 C Useful results on free probability C.1 Algebraic Context . . . . . . . . . . . . . . . . . . . C.2 An Example of a Noncommutative Probability Space C.3 The Joint Distribution . . . . . . . . . . . . . . . . . C.4 Freeness . . . . . . . . . . . . . . . . . . . . . . . . C.5 Free Multiplication . . . . . . . . . . . . . . . . . . C.6 Free Probability and Random Matrices . . . . . . . . D Proof of Theorem 1 E Proof of Theorem 2 F Proof of theorem 3 G Proof of theorem 4 H Proof of theorem 6 I J Proof of proposition 3 Proof of proposition 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 113 114 114 115 116 117 121 127 129 131 135 137 139 141 143

K COFDM Bibliography

Chapter 1

Introduction

The need to reliably transmit data over channels subject to fadings has led to a wide range of research topics. Several transmission schemes were shown to approach the capacity of a general multipath channel, where one must contend with intersymbol interference (ISI) as well as with Gaussian noise (in this respect, see [12, 48, 27, 111, 87, 88, 91, 92]). In particular, it is well acknowledged that in a bandwidthlimited channel, ISI is a dominant impairment and practical techniques for combined coding, shaping and equalization are required to approach capacity. In this introduction, a brief overview of single user transmission schemes is provided. We give the conditions under which multi-carrier transmissions are optimal and show that the choice of a judicious transmission strategy depends on the channel state knowledge available at the transmitter or/and at the receiver. When the channel is known at the transmitter, Shannon showed that the optimum (capacity-achieving) transmitted spectrum may be computed by the well-known water-pouring method [50]. One method for transmitting data is to use multi-carrier modulation to split the channel up into many independent narrow subchannels. If the bands are sufciently narrow, then each subchannel may be regarded as an ideal (ISI-free) Gaussian channel and only trivial equalization is required. The capacity of the whole channel is achieved by appropriate bit rate and power allocation of each subchannel independently. A remarkable result of this strategy is that the coding scheme must be adapted to the channel frequency response; in other words, the signal power should be high when the channel SNR is high and low when the channel amplitude is low. An interesting analogy can be made with an architect building a house on a non-stable eld. Changing the whole eld by bringing new soil would be prohibitively too expensive. A low-cost strategy would be to put the houses strongest pillars on the most solid spots of the eld. If the soil of these spots is good enough, the stability of the house can be garanteed. Multi-carrier Modulation is not the only method to achieve capacity with perfect state channel at the transmitter. An alternative is to use single-carrier modulation with transmitter precoding techniques in combination with decision feedback equalization. In particular, the work of Ciof et al. [3, 31, 32] established that the performance of an optimized single carrier system with a decision feedback equalizer on a xed ISI channel provides the same performance as an optimized multi-carrier system when the channel is perfectly known at the transmitter. However, from these two methods, Multicarrier Modulation seems to be the only transmission scheme that can simply solve the combined equalization and coding problem. Multicarrier Modulation is optimal only when channel knowledge is available at the transmitter and its success is mostly due to its low complexity equalization. Orthogonal Frequency Division Multiplexing also known as OFDM [15] is a

1. I NTRODUCTION

particular form of Multi-carrier Modulation which benets from this aspect: OFDM subcarriers overlap in the frequency domain yielding a high spectral efciency (compared to FDMA schemes) while keeping low complexity FFT processing [42, 24]. Moreover, thanks the ingenious use of cyclic redundancy at the transmitter, OFDM can cope with high data rates and frequency selective channels. This was the source of motivation for using OFDM in various standards (IEEE.802.11a, HiperLAN/2, ADSL. . . ) and the advantages are acknowledged worldwide (simple equalization, FFT processing split at the transmitter and the receiver [1, 108, 86, 34, 33, 106, 85, 120, 29, 15, 94]) When no channel knowledge at the transmitter is available, the water pouring strategy is inappropriate and the transmitted information on one OFDM subchannel can be irremediably lost if a deep fade occurs. Various methods also known as diversity techniques [22] have been proved efcient to cope with these channel impairments. In particular, if the receiver can be provided with different replicas of the information which have been subject to independent fadings, an appropriate combination of the replicas can restore the information. There are several techniques by which the receiver can be provided with independently faded replicas of the same version of the emitted signal. In the case of time diversity, the replicas of the signal are transmitted at different times. In this case, the symbols will be affected by independent attenuations if the time interval between each of these transmissions is chosen greater than the coherence time of the channel. In the case of frequency diversity, the signal is sent on different frequency bands separated by at least the channel coherence bandwidth. In the case of antenna diversity, different replica of the signal are provided thanks to the use of multiple antennas at the transmitter and the receiver. In order to achieve a higher level of diversity, the antenna spacing must be greater than half the wavelength. In order to turn these diversity methods spectrally efcient, sophisticated techniques based on the use of coding are usually used. In such scenarios, coding (in combination with time, frequency and antenna diversity) is used to combat the effect of fading and the effect of stationary noise but may incur a nonnegligible delay. Recently, Giraud and Belore [22] and Boutros and Viterbo [21] introduced a spectrally attractive scheme for fading channels called signal space diversity. The goal is to design high modulation diversity constellations (fading codes) based on the Hamming distance criteria. They showed in particular that when coding on Rayleigh fading channels is performed, the Hamming distance criteria plays a central role (whereas maximizing the Euclidean distance is the commonly approved design criteria for Gaussian channels). Two approaches were proposed to construct high diversity multidimensional constellations: the rst is based on the design of high diversity lattice constellations by applying the canonical embedding to the ring of integers of an algebraic number eld [35, 7] the second is based on the design of rotation matrices with respect to classical signal constellation in such a way that any two signals (after rotation) achieve the maximum number of distinct components. It is important to point out that this approach is useless in non-fading channels since it provides only diversity gains. Their construction assume Maximum likelihood detection and are based on sophisticated concepts from algebraic number theory. For increasing the diversity order in Rayleigh fading channels, a commonly

1. I NTRODUCTION

approved approach is to increase the systems dimension but the optimization becomes non trivial. Moreover, the decoding complexity limits greatly the use of such coding schemes for high dimensions. In this thesis, we used different tools to study the signal space diversity topic, considering different receivers such as the MMSE, the Successive Interference Cancellation and Parallel Interference Cancellation receiver. The focus is on a single user scenario, in which, the information is linearly precoded before transmission. The role of linear precoding is to spread the information on various carriers. This procedure is simply expressed by the fact that a K dimensional transmit information vector is pre-multiplied by a N K matrix before transmission on a frequency selective fading channel. Since we have to keep in mind the decoding complexity versus performance dilemma issue, Linear Precoded OFDM (LP-OFDM) schemes are studied to benet from the simple frequency channel equalization nature (thanks to the introduction of the well-known cyclic prex). One of the main attractive features of Linear Precoded OFDM is that it exploits all the frequency diversity of the channel when the precoder is appropriately chosen. Previous works have already studied this scheme. In particular, Boutros et al. [21] and Belore et al. [51] proposed to optimize the coefcients of a square (K = N) precoding matrix by deriving an upper bound of the error of the maximum likelihood (ML) detector. More recently, Giannakis [119] generalized these results to rectangular matrices (K < N) considering limited diversity channels. However, all previous studies have only focused on Maximum likelihood detection and small dimensional matrices. But in the OFDM wireless Local Area Context, the number of carriers is high (e.g. HiperLAN/2 [64] has 64 carriers) and the high computational cost of the ML detector may prevent its use for implementation in nowadays receivers. Our aim is to study the performance of various redundant N K precoders with respect to the chosen equalizer under certain channel assumptions and try to resolve the complexity versus performance dilemma through: the use of suboptimum detectors (MMSE, non-linear detectors: Successive Interference Cancellation and Parallel Interference Cancellation) performing an asymptotic analysis (i.e when the dimension of the block length tends to innity) A brief description of the content of the thesis is provided below

Chapter 2

In this chapter, notations as well as a brief description of OFDM is provided. In particular, we show that the discrete time OFDM model is equivalent in the frequency domain to a set of parallel frequency at fading channels. Under the condition that the cyclic prex (a sort of cyclic guard interval) is longer than the channel impulse response length, the Ndimensional received vector at time n, y(n) = (y1 (n), , yN (n))T can be expressed as a function of the emitted symbol vector s(n) = (s1 (n), , sN (n))T : y(n) = HN s(n) + n(n) (n))T n(n)n(n)H (1.1)

where n(n) = (n1 (n), , nN is an additive white Gaussian noise such that E = 2 IN , and where HN = diag([h1 , . . . , hN ]) is the N N diagonal complex matrix bearing on its diagonal the channel frequency gains. Note that n denotes the time dependence but all the vectors are expressed in the frequency domain. We describe the advantages of OFDM (simple equalization, FFT complexity split at the transmitter and the receiver) and drawbacks (loss of information on faded carriers, lack of diversity,

1. I NTRODUCTION

peak to average ratio [107]). We introduce coded OFDM (COFDM [47, 128]) as a solution for overcoming these drawbacks and propose to analyze the performance of redundant precoders for a high diversity scheme called Linear Precoded OFDM (LP-OFDM). The linear precoding technique was initially introduced by Wornell in 1996 [122] and called Spread Response Precoding but we will use hereafter the terminology LP-OFDM in accordance with Giannakis [119].

Chapter 3

In this chapter, the LP-OFDM scheme is introduced as a way of increasing the diversity of plain OFDM schemes. The K-dimensional symbol vector s(n) = (s 1 (n), , sK (n))T (a white vector process with E s(n)sH (n) = IK ) is multiplied by an N K matrix WN,K where N K. The role of matrix WN,K is to introduce diversity (and redundancy when N > K). Therefore, the N dimensional received vector y(n) = (y1 (n), , yN (n))T can be expressed as: y(n) = HN WN,K s(n) + n(n) We also introduce in our study two channel models: Non-ergodic channel: The symbols are transmitted over a multi-path channel whose impulse response is given by c(t) = M1 l g(t l ). (l )(l=0,...,M1) and (l )(l=0,...,M1) are respectively l=0 the delays and channel gains. The gains assumed to be uncorrelated complex Gaussian variables. c(t) is assumed to be compactly supported. Although the following assumption is not compatible with the compact support assumption, we assume that c(t) is band-limited and denote by W the 1 corresponding bandwidth. Let T = W be the sampling rate, the equivalent discrete-time channel transfer function denoted h( f ) = L1 c(lT )ei2l f is given by h( f T ) = M1 l ei2l f G( f ) l=0 l=0 where W f W . L is the length of the channel order. Coefcients (h k )1,...N are given by: 2 2 k hk = h(e j2 N ). Ergodic channel: We assume that in fast fading environments, time and frequency interleaving is performed on the components of WN,K s(n). The purpose of interleaving is to scramble the precoded data stream in order to endow the coefcients (h k )1,...N with ergodic properties. In particular, if G( f ) is an ideal lter (G( f ) = 1 if f W and 0 elsewhere) then the sequence (h k )1,...N can be assumed to be an identically distributed ergodic sequence. Therefore, it is reasonable to consider the coefcients (hk (n))1,...N as random variables with the same probability distribution and 1 N f (|hk |2 ) = N+ N k=1 lim f (t)p(t)dt almost surely (1.2)

It is interesting to note that LP-OFDM is strongly related to other schemes and can be seen as: A combination of two schemes, OFDM and CDMA: Proposed in the early 90s [67, 82, 124, 83, 45] for downlink transmissions, Multi-carrier CDMA (MC-CDMA) is a multi-user access scheme based on an OFDM modulator. All the users share the same bandwidth and time transmission but are separated by different codes. This scheme is interesting for channel estimation and more robust in terms of diversity than other multi-access schemes (OFDM-FDMA or OFDM-TDMA). However, the multi-user interference generated by the use of multi-access codes in fading channels

1. I NTRODUCTION

can penalize the performance with standard receivers. In order to improve the performance, multiuser detection techniques are investigated but can greatly increase the mobile terminals complexity. LP-OFDM can be seen as a MC-CDMA scheme in which all the codes are assigned to the same user. A multi-dimensional constellation and other signal space diversity schemes on non-selective Rayleigh fading channels: Linear Precoding is a smart and simple method for mapping information to a new multi-dimensional constellation [22, 51, 18, 21]. When the precoder is orthogonal, the new multi-dimensional constellation has the same Euclidean distance as the previous one but a higher Hamming distance (which is crucial in Rayleigh fading channels). Previous works in this eld are also described. The reader can refer to: Belore et al.[22, 51], Boutros et al.[18, 21]: expressions of the pair-wise error probability, ML detection, optimization of the coefcients. Giannakis [119, 118]: Performance analysis of rectangular precoders in the single-user OFDM scenario Fazel [46], Fettweis [5], Kaiser [99], Lindner [67], Linnartz [125], Rohling [82, 83]: Performance analysis of coded OFDM-CDMA versus COFDM, use of suboptimal equalizers. Analogy with space-time coding schemes [97, 100] The use of precoding in the case of maximum likelihood detection has been already acknowledged in previous publications in order to increase diversity [122]. In the case K = N, usual tools rely on minimizing an upper bound of the pairwise error probability (PEP) (the probability of mistaking a vector x for another vector y of the constellation). These studies have shown that the PEP depends on two parameters: The diversity gain which characterizes the slope of the PEP. In this case, the diversity gain is achieved thanks to the use of linear precoding. The coding gain which is related to the ordinate at the origin of the PEP curve. This coding gain is achieved thanks to the use of redundancy. However, in the case of suboptimum equalizers with redundant (K < N) precoders (to our knowledge, this topic has not been studied before), no straightforward conclusion can be given on the coding versus diversity gain benet of linear precoders. The goal of the thesis is to study analytically this issue.

Chapter 4

The focus of this chapter is on the MMSE analysis of LP-OFDM. The use of the MMSE receiver is appealing since it has a simple implementation structure. Furthermore, this receiver takes into account the presence of noise, unlike the other type of linear receiver called the decorrelator. As a matter of fact, the decorrelator, which is the multiuser equivalent of the zero-forcing (ZF) equalizer, usually enhances the noise effect.

1. I NTRODUCTION

Considering the model of equation 1.2, the output of the MMSE detector is s(n) = [s1 (n), . . . , sK (n)]T and is given by s(n) = E s(n)yH (n) E(y(n)yH (n))

1

y(n)

1

y(n) .

Each component sk (n) of s(n) is corrupted by the effect of both the thermal noise and by the "multi user interference" due to the contributions of the other symbols {s l (n)}l=k . The component sk (n) has therefore the following form: sk (n) = wk sk (n) + k (n) where wk = wH HH HN WN,K WH HH + 2 IN k N N,K N

1

HN wk .

(1.3)

wk is the kth column of WN,K . It has been shown in [80] and [127] that the additive noise k (n) can be considered as Gaussian when K and N are large enough. Therefore, k (n) can be considered as a zero mean Gaussian noise of variance wk (1 wk ). In order to analyze rigorously the performance, the Signal to Interference plus Noise Ratio (SINR) at the output of the MMSE detector is considered. The SINR wk at the output k of the MMSE detector is dened as: wk = wk 1 wk

(1.4)

The SINR is then used as benchmark measure of performance for various criteria, in particular the bit error rate and the spectral efciency (which are directly linked to the SINR). In a rst step, we analyze the performance of certain types of square (K = N) unitary precoding matrices (since synchronization is perfectly controlled at the emitter and these type of matrices give the best performance) such as FFT, Walsh Hadamard, random unitary matrices when N . The asymptotic assumption is particularly appealing since the output SINR converges in this case to deterministic values, that can be analytically expressed. Furthermore, results indicate a rather fast convergence to the asymptotic limits even with nite size, demonstrating the practical importance of these limits. We show that in the uncoded case, with sufcient time and frequency interleaving, precoding effectively transforms, after MMSE equalization, an arbitrary Rayleigh fading channel into a non fading, Gaussian channel. Furthermore, such a precoding requires no additional power or bandwidth, and is attractive in terms of computational complexity [80, 125, 122]. In the coded case, a surprising result appears. The coded case topic has given rise to many publications since 1992. In particular, several papers have shown by simulations that there is a threshold coding rate of 1 above which Coded LP-OFDM outperforms COFDM. They observed that the impact 2 of this diversity gain can be considerably reduced when Convolutional Coding and suboptimum equalizers are used with respect to other schemes like COFDM. We analyze more rigorously this observation [36, 63, 46, 81, 50, 5] and give conclusions.

1. I NTRODUCTION

In a second step, for increasing the spectral efciency and overcoming the suboptimality of MMSE receivers, we propose to add redundancy (N > K) and optimize the precoders dimension. The performance of the MMSE receiver can still be characterized by its SINR, which, unfortunately in this case, has a non interpretable analytical expression. Indeed, the problem becomes analytically intractable even for systems of moderate nite size (except for some particular matrix structure). In order to overcome this difculty, we propose to use an attractive approach already used in the context of multi-users CDMA systems, where the precoder is modeled as a certain type of random matrix. Indeed, since 1999, several excellent papers have analyzed the performance of asymptotic CDMA schemes (by asymptotic, it is understood that the number of users tends to innity, the processing gain tends to innity while the ratio remains constant) [11, 44, 102, 113, 93, 117, 90]. They have in particular obtained interpretable formulas and shown that for the MMSE receiver, a trade-off exists with coding and spreading. In this chapter, we further extend these results to the LP-OFDM specic case. Two types of precoding matrices for ergodic and non-ergodic channels are analyzed: random i.i.d matrix: coefcients of the precoding matrix are modeled as i.i.d random variables. This precoder model is introduced only for deriving a reference bound on the performance. The i.i.d. case study is based on mathematical results related to the "limiting distribution of eigenvalues" of some large random matrices with independent and identically distributed entries (see e.g. [93],[102] ). Under certain conditions, we show that when N grows towards innity and K/N , the SINR wk at the output k of a MMSE equalizer converges almost surely to a value that is the unique solution of the equation: Ergodic case:

0

t t + 2 + 2

p(t) dt =

. +1

(1.5)

Non-ergodic case:

1 2 1 2

(1.6)

h( f ) is the equivalent baseband transfer function. Note that belongs to [0, [. It is remarkable that the SINR is shown to converge almost surely to the same value. For random i.i.d entries matrix, we show that an non-negligible amount of redundancy has to be added to the precoder in order to increase the spectral efciency. The spectral efciency with MMSE equalization (when considering Gaussian entries) is dened as: = log2 1 + (, 2 ) . The need for an optimum amount of redundancy can be easily understood by the previous formula. For increasing the spectral efciency, one can increase to 1 (no redundant precoding) but in this case (, 2 ) and therefore log2 1 + (, 2 ) decreases (since the Multi-user interference increases). On the other hand, if decreases to 0, (, 2 ) increases (as an extreme case, the reader can imagine a single symbol spread across N carriers; the redundancy factor is equal to while the SINR is equal to = spectral efciency.

1 N

1 2 1 2

|h( f )|2 d f) 2

1. I NTRODUCTION random Haar distributed isometric matrix: A N N random unitary matrix is said to be Haar distributed if its probability distribution is invariant by right multiplication by deterministic unitary matrices. We restrict our study to isometric matrices (obtained by extracting K < N columns from a Haar unitary matrix) since perfect synchronization is ensured at the transmitter and these type of matrices are able to provide better performance than i.i.d ones. The Haar distribution is introduced for calculus purpose. This type of matrix is more suited to our case, but it requires advanced mathematical concepts (due to the orthogonality of the column codes of the precoder). New tools, borrowed from the so-called free probability theory [115, 116] are used to analyze the SINR formula in equation(1.4). Free Probability Theory was rst introduced abstractly by D. Voiculescu around 1983 in order to study some problems of Von Neuman algebras, but some years later, around 1990, he realized that a concrete probabilistic model of free random variables is afforded by large independent random matrices. To our knowledge, the rst use of free probability in the telecommunication eld was made by Evans and Tse in the CDMA context [44], but for solving quite different problems. Mueller [73] and Biglieri [13] should also be considered as precursors for introducing free probability in the antenna eld. In the same vein as in the i.i.d case, we show that for all matrices chosen according to the previously mentioned distribution, when N and K/N 1, the SINR wk at the output k of the MMSE equalizer converges almost surely to a value that is the unique solution of the equation: Ergodic case:

0

Non-ergodic case:

1 2 1 2

t p(t) dt = . t + 2 (1 ) + 2 +1

(1.7)

(1.8)

In order to take into account the case > 1, results for N K precoders having orthogonal rows, called the co-isometric case, are given. These results are easy to derive since W N,K WH = I N,K (the scaling factor guarantees that the average power allocated to each component of s is 1) and yield: Ergodic case:

0

t p(t) dt = . 2 t + +1

(1.9)

p(t) is the probability density of the random variables (|h i (n)|2 )i=1,...,N . Non-ergodic case:

1 2

1 2

(1.10)

1. I NTRODUCTION

Note that > 1 is not at all an unrealistic case and can indeed (as shown later in this thesis) incur an improvement in the i.i.d case (see also [90] in the CDMA case) For < 1, the results of orthogonal spreading are closely related to the i.i.d case (only a factor 1 has changed in the denominator) and we can study the impact of the orthogonality of the codes. The random Haar distributed isometric precoders provide better performance then the i.i.d K case especially for high loads ( N close to 1). Moreover, simulations show that substituting random Haar distributed isometric matrices with practical N K Walsh-Hadamard precoders incur no loss. For > 1, the co-isometric and i.i.d precoders perform quite similarly. As an extreme case, when , 0 and formula (1.10) and (1.5) are equal. In this chapter, an explicit answer to the coding versus linear precoding trade-off issue in terms of spectral efciency is given. The optimum amount of redundancy is derived and our results are then sustained by simulations for Convolutional codes. However, even when the optimum redundancy factor is applied, the capacity of the whole system (i.e the system including the MMSE receiver) is still far away from the theoretical channel capacity. We extend therefore the analysis to non-linear detection techniques with reasonable complexity in order to achieve greater improvements

Chapter 5

Chapter 5 focuses on the analysis of non-linear detectors for LP-OFDM schemes. In particular, we argue that the single user LP-OFDM scenario is suited for these techniques by considering each frequency data stream as a separate user [112]. We focus our study on two type of detectors: Successive interference cancellation scheme (SIC):

At each step of the detection process, the contribution of a stream of symbols (or user) is estimated then subtracted from the received signal. Inter-carrier interference generated by the precoder is thus decreased gradually. SIC detectors have been shown to be capacity achieving ([110, 90]) schemes when coding is applied in the context of uplink CDMA and random spreading . We extend these results to the LP-OFDM context [75, 121] when using Haar random isometric precoders. The performance at each iteration is derived as well as the spectral efciency.

0 log2 (1 +

The spectral efciency of re-encoded successive cancellation is shown to be: = (x)) dx in bits/s/Hz. Hereby, the SINR (x) is solution of the following equation: random Haar distributed isometric matrix Ergodic Case:

0

t xt + 2 (1 x)(x) + 2

p(t) dt =

(x) . (x) + 1

(1.11)

Non-Ergodic case:

1 2

1 2

(1.12)

0

1. I NTRODUCTION

t xt + 2 ((x) + 2

p(t) dt =

(x) . (x) + 1

(1.13)

Non-Ergodic case:

1 2 1 2

(1.14)

As an example, for ergodic channels, assuming perfect retrieval of the symbols, the SINR at the nal step converges to the value (0) (for i.i.d and isometric precoding) which is solution of: (0) = (0) + 1

0

t p(t) dt 2 ((0) + 1)

and yields the performance of the single user bound in fading channels:

0

(0) =

t p(t) dt 2

In order to be efcient, for each substream k (1 k K) such as k = [xN + 1], coding on this substream at the rate log2 (1 + (x)) must be performed at the transmitter. In this case, we show that the spectral efciency of re-encoded successive cancellation is equal to the spectral efciency of the optimal detector: =

0

The coding strategy requires however knowledge of the SINR (x) (which depends on the channel attenuations in the non-ergodic case or only the channel statistics in the ergodic case) at the transmitter. In fast fading environment, such a strategy may not be implementable. Simulations show that the optimal spectral efciency is achieved for = 1 in the isometric case and for in the i.i.d case. Surprisingly, the i.i.d case incurs no loss of spectral efciency as long as . But the SIC algorithm requires the inversion of (K 1) matrices (corresponding to the K symbols to be detected) which makes the i.i.d case too complex for LP-OFDM. Unlike the MMSE case, when the MMSE SIC detector is implemented, choosing i.i.d codes instead of orthogonal is more a question of complexity than performance. Parallel Interference Cancellation scheme (PIC): in these schemes, all the symbols are decoded simultaneously and retrieved in order to decrease the non desirable interference. The process is iterated a certain number of times to enhance the accuracy of the estimated symbols. Several papers have also shown that PIC schemes perform very well in the LP-OFDM context [99]. However, they were unable to optimize the different parameters involved in the non-linear detection process. Based on previous publications in the CDMA context [28], we give an efcient manner to optimize the parameters while keeping a very simple implementation structure for:

1. I NTRODUCTION

11

Unitary (K = N) matrices: we derive the results for Walsh Hadamard, FFT and unitary random matrices without the need of free probability theory Haar random isometric matrices: we provide useful expressions based on the free probability theory Denote by lk the SINR corresponding to symbol k at iteration l, we show in these two cases that when N and K/N 1, lk converges almost surely to a value l that is the unique solution of the equation Ergodic Case:

0

|h( f )|2 l df = . l1 |2 )|h( f )|2 + 2 (1 |l1 |2 )(1 )l + 2 1 (1 |l1 |2 )l + 1 2 (1 | (1.16) l = 1 2Q l in the case of QPSK constellations. with Recall that Q(x) =

t2 1 e 2 x

2

1 2

dt.

In the case of i.i.d precoding, a quite similar result is obtained when N and K/N : Ergodic Case:

0

(1.17)

Non-Ergodic case:

1 2 1 2

with l = 1 2Q

|h( f )|2 l df = . (1 |l1 |2 )|h( f )|2 + 2 (1 |l1 |2 )l + 2 (1 |l1 |2 )l + 1 l in the case of QPSK constellations.

(1.18)

with l = 1 2Q

Using the same methodology, these results can be of course extended to all types of constellations. Formula1.15 differs from 1.17 only by a factor (1 ). We show that in the i.i.d case as well as in the isometric case, l is an increasing function of the number of iterations l.

12

1. I NTRODUCTION When orthogonal codes are used, performances close to the single user bound are achieved (with no coding) When i.i.d codes are used, remarkably, close to single user bound performance is still possible to achieve at the price of a greater number of iterations than the orthogonal case. Indeed, as l increases, l = 1 2Q l 1 and expressions 1.17 and 1.15 become similar. As for the SIC detector, the use of orthogonal spreading is more a question of complexity than performance.

Finally, a basic tutorial on free probability focusing in particular on the theorems in relation with our work is reported in the appendix [9, 115, 116, 68, 14, 58]. One major contribution of this thesis is the use of free probability in the optimization design of wireless schemes. Remarkably, these results have simple interpretations and only few parameters (noise variance, channel characteristics, ratio K/N, orthogonality of the codes) seem to impact the performances. These results are quite useful as systems become more and more complex and therefore depend on too many parameters. Moreover, we show that the results (and their conclusions) can be straightforwardly extended to MC-CDMA and multi-antenna schemes. Further research topics are also proposed. Note that all related works that we performed during this thesis can be found in the following bibliography.

M. Debbah, P. Maill, and M. de Courville. Multiresolution decoding algorithm for OFDMCDMA, Spread OFDM single user systems. Disclosure, Motorola, 2001. M. Muck, M. de Courville, and M. Debbah A new modulator and demodulator for high data-rate communications systems: Pseudo-Random Cyclic Prex Orthogonal Frequency Division Multiplexing . Disclosure, Motorola, 2002. M. Debbah, B. Muquet, M. de Courville, M. Muck, S. Simoens, P. Loubaton. A New MMSE Successive Interference Cancellation Scheme for Spread OFDM Systems In Vehicular Technology Conference, Tokyo, Japan, May, 2000. M. Debbah, M. de Courville, M. Muck, and P. Loubaton. A new Adjustable Hybrid Spread OFDM Modulator. In 5th International OFDM Workshop, Hamburg, Gemany, September 2000. M. Debbah, P. Loubaton, and M. de Courville. Spread OFDM Performance with MMSE equalization. In IEEE International Conference on Acoustics, Speech, and Signal Processing, Salt Lake City, USA, May 2001. M. Debbah, W. Hachem, P. Loubaton, and M. de Courville. Asymptotic Performance Analysis for Redundant Block Precoded OFDM Systems with MMSE Equalization. In 11th IEEE Statistical Signal Processing Workshop, Singapore, August 2001. M. Debbah, W. Hachem, P. Loubaton, and M. de Courville. MMSE Analysis of Certain Large Isometric Random Precoded Systems. In 2001 Information Theory Workshop, Cairns, Australia, September 2001.

1. I NTRODUCTION

13

P. Maill, M. Debbah, P. Loubaton, and M. de Courville. Asymptotic Analysis of Successive versus Parallel Interference Cancellation Schemes for Block Precoded OFDM . In Proceedings of the Allerton Conference, 2001. M. Debbah, P. Loubaton, and M. de Courville. Linear Precoders for OFDM wireless communications with MMSE equalization: facts and results. In Proceedings of Eusipco, European Sig. Proc. Conf., Toulouse, France, September 2002. M. Debbah, M. de Courville, and P. Maill. Multiresolution Decoding algorithm for WalshHadamard Linear Precoded OFDM. In 7th International OFDM Workshop, Hamburg, Germany, September 2002. M. Debbah, W. Hachem, P. Loubaton, and M. de Courville. MMSE Analysis of Certain Large Isometric Random Precoded Systems. Submitted to IEEE Trans. Information Theory, 2002.

14

1. I NTRODUCTION

15

Chapter 2

OFDM

2.1 Introduction

Recently, a worldwide convergence has occurred for the use of Orthogonal Frequency Division Multiplexing (OFDM) as an emerging technology for high data rates. In particular, three wireless local Area Networks systems have been normalized simultaneously in three regions of the world: the USA, EUROPE and JAPAN, respectively, IEEE 802.11a, ETSI BRAN HIPERLAN/2 and ARIB MMAC. They all occupy the 5 Ghz band, with several channels spaced by 20 Mhz offering data rates of up to 54 Mbit/s, adequate for the current (multimedia,. . . ) and future applications. OFDM is a particular form of Multicarrier transmission and is suited for frequency selective channels and high data rates. This technique transforms a frequency-selective wide-band channel into a group of non-selective narrowband channels, which makes it robust against large delay spreads by preserving orthogonality in the frequency domain. Moreover, the ingenious introduction of cyclic redundancy at the transmitter reduces the complexity to only FFT processing and one tap scalar equalization at the receiver

In this section, we will focus on the baseband discrete-time representation of OFDM. For a more general presentation based on orthogonal transmultiplexers or block precoding issues, the reader can refer to [2, 37, 118]. The history of multi-carrier modulation began more then 30 years ago. At the beginning, only analog design based on the use of orthogonal waveforms was proposed [85]. The use of discrete Fourier Transform (DFT) for modulation and demodulation was rst proposed in [120]. Only recently has it been nding its way into commercial use, as the recent developments in technology have lowered the cost of the signal processing that is needed to implement OFDM systems [43, 108, 106]. In this section, we rst give a brief overview of frequency selective channels [81, 12]

2. OFDM

c(t)x(t )d + n(t).

(2.1)

Frequency selectivity occurs whenever the transmitted signal x(t) occupies an interval bandwidth [ W , W ] 2 2 greater then the coherence bandwidth B coh of the channel c(t) (dened as the inverse of the delay spread Td [81]). In this case, the frequency components of x(t) with frequency separation exceeding B coh are subject to different gains.

Fig 2.2 presents a typical time impulse response c(t) of a channel. For usual high data rates schemes, the symbol rate T is small compared to Td (they are also called broadband signals) and the signals are therefore subject to frequency selectivity. The multipath-channel can be modeled by an impulse response given by c(t) = M1 l g(t l ) l=0 where g(t) is transmitting lter and Td is the duration of the multipath or delay spread. Here, the complex gains (l )l=0,...,M1 are the multi-path gains and the (l )l=0,...,(M1) are the corresponding time delays. The variance of each gain as well as the time delays are usually determined form propagation measurements. As a typical example, we give in table 2.1 and table 2.2 provide the delay prole of 1 indoor channels A and E. Let W denote the signal bandwidth and T = W the sampling rate. We will assume hereafter that the transmitting lter is supposed to be ideal (G( f ) = 1 for f [ W , W ] and 0 2 2 outside, G( f ) being the Fourier transform of the transmitting lter g(t)). One of the main concerns in transmissions schemes is to retrieve x(t) from eq (2.1). This operation is called equalization and the difculty of extracting x(t) is mostly due to the frequency selectivity behavior of the channel (with c(t) = (t) where (t) is the Dirac distribution, no equalization is required!). Moreover, the equalization task is all the more difcult to implement that the complexity of an equalizer grows with the channel memory. Therefore, the cost (in terms of complexity and power consumption) of such an equalizer could be prohibitively too high, especially in the case of high data rates communications. The main idea of OFDM transmissions is to turn the channel convolutional effect of equation (2.1) into a multiplicative one in order to simplify the equalization task. To this end, OFDM schemes add redundancy known as cyclic prex in a clever manner in order to circularize the channel effect. Based on the fact that circular convolution can be diagonalized in an FFT basis [78], the multi-path time domain

2.2. OFDM

PRINCIPLE

17

Tap Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Delay (ns) 0 10 20 30 40 50 60 70 80 90 110 140 170 200 240 290 340 390

Average Relative Power (dB) 0 -0.9 -1.7 -2.6 -3.5 -4.3 -5.2 -6.1 -6.9 -7.8 -4.7 -7.3 -9.9 -12.5 -13.7 -18.0 -22.4 -26.7

Table 2.1: Model A, corresponding to a typical ofce environment for NLOS conditions.

Tap Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Delay (ns) 0 10 20 40 70 100 140 190 240 320 430 560 710 880 1070 1280 1510 1760

Average Relative Power (dB) -4.9 -5.1 -5.2 -0.8 -1.3 -1.9 -0.3 -1.2 -2.1 -0.0 -1.9 -2.8 -5.4 -7.3 -10.6 -13.4 -17.4 -20.9

Table 2.2: Model E, corresponding to a typical large open space environment for NLOS conditions.

18

2. OFDM

channel is transformed into a set of parallel frequency at fading channels. Moreover, OFDM systems take benet from the low cost implementation structure of digital FFT modulators 1 . All these advantages make OFDM particularly suited for frequency selective channels.

MODULATOR DEMODULATOR

xN

D 1

modulation

Let us now focus on the conventional OFDM transceiver depicted in g.2.2. As a starting point, we will consider the noiseless transmission case. The incoming high rate information is split onto N rate sub-carriers . The data is therefore transmitted by blocks of size N: s(k) = [s 1 (k), . . . , si (k), . . . , sN (k)] where the index k is the block OFDM symbol number and the subscript i is for the carrier index. The block OFDM symbol is precoded by an inverse FFT matrix F H = F1 to yield the so-called time domain N N block vector x(k) = [x1 (k), . . . , xi (k), . . . , xN (k)]. At the output of the IFFT, a guard interval of D samples is inserted at the beginning of each block [x ND+1 (k), . . . , xN (k), x1 (k), . . . , xi (k), . . . , xN (k)]. It consists of a cyclic extension of the time domain OFDM symbol of size larger than the channel impulse response (D > L 1). The cyclic prex (CP) is appended between each block in order to transform the multipath linear convolution into a circular one. After Parallel to Serial (P/S) and Digital to Analog Conversion (ADC), the signal is sent through a frequency-selective channel. The channel can be represented by an equivalent discrete time model and its effects can be modeled by a linear Finite Impulse Response (FIR) ltering with Channel Impulse Response c N = [c1 , . . . , cL1 , 0, . . . , 0]. Usually, the system is designed so that D is smaller than N (D= N ) and greater than (L 1). One can no4 N tice that the redundancy factor is equal to N+D . On the one hand, in order to avoid spectrally inefcient N transmissions, N has to be chosen far greater than D (lim N N+D = 1: for a xed D, the redundancy factor tends to 1 as the the number of carriers increases). On the other hand, the FFT complexity per carrier grows with the size of N. Moreover, the channel should not change inside one OFDM symbol to 1 be able to circularize the convolution. Finally, the carrier spacing is related to the factor NT and reduces as N increases: there is no gain in terms of diversity for a xed channel by increasing N. The choice of N depends therefore on the type of channel (slow varying, fast fading, high diversity channel, impulse response length...) and the complexity cost one is able to accept.

reader should note that other schemes based on block redundant precoding are also able to circularize the linear convolution (zero-padded OFDM [76], pseudo-random cyclic prex [72])

1 The

xN k

xN k

sN k

rN

cp

rN k

demodulation

xN

D 1

yN k

sampling rate T

DAC

Ct

"

x3 k

r3 k

s1 k s2 k

x1 k x2 k

x1 k

r1 k r2 k

rD k

xN k

cp

y1 k

hN

P S

P S

r1 k

cp

h1

sk

xk

rcp k

rk

yk

2.2. OFDM

PRINCIPLE

19

N carriers

f f

1 NT

B A C) 8 7 9

6 3 &5 42

' %&$

B A @) 8 7 9

6 45 342

20

2. OFDM

At the receiver, symmetrical operations are performed: down conversion, Analog to Digital Conversion (ADC). The discrete time received signal with guard interval r CP has therefore the following expression:

CP r = with

(N+D)1

(N+D)1

xND+1 (k 1) . . . xN (k 1) + HIBI x1 (k 1) . . . xN (k 1)

(N+D)1

c0 . . . c HISI = L1 0 . . . 0 and

0 .. . .. .. .

.. . .. . ..

.. .. . .

.. ..

. . 0 cL1

HIBI = 0 c0 (N+D)(N+D)

0 . . . . . . . . .

0 0 cL1 . .. .. .. . . . . . . .. .. . . . . . .. . . . . .. . . . 0

c1 . . . cL1 0 . . . 0

(N+D)(N+D)

s1 (k) . . . . . . . . . sN (k)

N1

HISI represents inter-symbol interference generated by the frequency selective behavior of the channel inside an OFDM block at time k [xND+1 (k), . . . , xN (k), x1 (k), . . . , xi (k), . . . , xN (k)] while HIBI corresponds to inter-block interference between two consecutive OFDM block transmissions at time k [xND+1 (k), . . . , xN (k), x1 (k), . . . , xi (k), . . . , xN (k)] and at time (k1) [xND+1 (k1), . . . , xN (k1), x1 (k 1), . . . , xi (k 1), . . . , xN (k 1)]. Denote H(z) = L1 ck zk be the channel transfer function and let k=0 H = FN 1 cN = [H(0), H(e j2/N ), . . . , H(e j2N1)/N )]T = [h1 , . . . , hN ]T be its Fourier transform. At the receiver, in order to suppress the inter-block interference, the rst D samples of the received signal rCP are discarded.

2.2. OFDM

PRINCIPLE

21

rCP

D+1 (k)

. . . . . .

N1

c0 .. .

0 .. .

.. . .. .

cL1

r1 (k) . . . . . . rN (k)

N1

N1

The use of cyclic redundancy has thus enabled us to turn the linear convolution into a circular convolution. Since any circulant matrix is diagonal in the Fourier basis [53, 54, 26, 25, 59], it is very easy to diagonalize the channel effect by FFT processing at the receiver:

y1 (k) . . . . . . yN (k)

N1

= FN

r1 (k) . . . . . . rN (k)

(2.2)

N1

y1 (k) . . . . . . yN (k)

N1

c0 . . . c = FN L1 0 . . . 0

0 .. . .. .. .

cL1 .. .. . . .. .. . . .. .. . . .. .. . . 0 cL1

c1 . . . cL1 . . . 0 c0

NN

FH N

s1 (k) . . . . . . . . . sN (k)

(2.3)

N1

Since the circular convolution yields a multiplication in the frequency domain, the signal [s 1 (k), ..., sN (k)]

22

2. OFDM

is transmitted over N parallel at fading channels, subject each to a complex frequency attenuation h i h1 0 . . . . . . . . . 0 0 h s1 (k) y1 (k) 0 ... ... ... 2 . . .. . . . 0 ... ... . . 0 . . . . . (2.4) = . . . . ... ... .. ... ... . . . . . . . . . . ... . ... .. ... yN (k) N1 sN (k) N1 . . ... ... ... ... h . N NN In the case of noisy transmission, the time Gaussian noise vector [b 1 (k), . . . , bN (k)]T added is multiplied at the receiver by the FFT demodulator:

n1 (k) . . . . . . nN (k)

N1

= FN

b1 (k) . . . . . . bN (k)

N1

Since the statistics of a Gaussian vector does not change by orthogonal transform, [n 1 (k), . . . , nN (k)]T is a white gaussian vector with the same variance. We give hereafter the frequency equivalent model representation of OFDM.

Figure 2.5: OFDM frequency model. Hereafter, a short description of a standardized OFDM scheme known as HiperLAN/2 is provided in order to give the reader some basic knowledge of the common parameters used in a wireless network (number of carriers, constellations,...). HiperLAN/2 [65] is a 5 Ghz European Standard developed by ETSI with a physical layer based on OFDM. HiperLAN/2 is intended to provide wireless connectivity between PCs, laptops. . . either in an indoor or outdoor environment for pedestrian mobility. The cell

w v u x s

w v u t s E U HG T X RHY X S W V y

RI G i e h g p f P Hr P q P

w vu x

w vut

U HG T RRS G Q P HI G F

U HG T c G T HI b RPUUX a `

RI G i e h g f P T HT e I U X c d

w vu xs

w vut s

2.2. OFDM

PRINCIPLE

23

radius extends to 30 m in indoor environments or up to 150m outdoors. The lower frequency band, from 5.15 to 5.35 Ghz contains 8 channels spaced by 20 MHz while the upper band consists of 11 channels, from 5.470 to 5.725 Ghz. A typical centralized network consists of different Mobile terminals (MT) that communicate with their respective Access Points (AP) over the air interface. The important general characteristics are summarized below and in the table 2.3: sampling frequency : Fe =20MHz; adjacent channels spacing : 20MHz; FFT size : N = 64; K = 48 useful carriers and 4 pilots (on carriers : 7, 21; used for phase tracking); modulation of sub-carriers : BPSK, QPSK, 8PSK, 16QAM, 64QAM; guard interval of 16 samples (800 ns). Wireless LAN for indoor/campus/home environment OFDM modulation with a TDMA/TDD access scheme 19 channels (8 in the lower band, 11 in the upper one) with a bandwidth of 20 Mhz High bit rate on top of the PHY layer (6-54 Mbit/s) Quality of service support Automatic frequency allocation Convolutive code constraint length 7 punctured Modulation BPSK BPSK QPSK QPSK 16-QAM 16-QAM 64-QAM Code Rate Net rate on top of PHY 1/2 6 Mbit/s 3/4 9 Mbit/s 1/2 12 Mbit/s 3/4 18 Mbit/s 9/16 27 Mbit/s 3/4 36 Mbit/s optional 3/4 54 Mbit/s Byte per Symbol 3 4.5 6 9 13.5 18 27

24

2. OFDM

In the previous section, we showed how OFDM converts a frequency selective channel into a collection of at fading channels thanks to the use of cyclic prex. Such a strategy has immediate advantages. As previously stated, one of the attractive features of OFDM is that, for a certain delay spread, the complexity of an OFDM modem vs. sampling rate does not grow as fast as the complexity of a single carrier system with an equalizer (thanks to the use of redundancy). The reason is that when the sampling rate is reduced by a factor of two, an equalizer has to be made twice as long at twice the speed, so its complexity grows quadratically with the inverse of the sampling rate, whereas the complexity of OFDM grows only slightly faster than linear. This makes easier to implement modems, which have to handle data rates exceeding 20 Mb/s. In OFDM systems, only simple (scalar) equalization is performed at the receiver (whereas in the context of single carrier transmission, a matrix inversion is required). Indeed, provided that the impulse response of the channel is shorter then the Guard interval, each constellation is multiplied by the channel frequency coefcient and there is no Inter-Symbol Interference (ISI). However, the channel still has to be compensated by a multiplication of each FFT output by a single coefcient: si (k) = gi hi si (k) + gi (k)ni (k) The matrix equivalent equation is: s(k) = Gy(k) with g1 0 . . . 0 g 0 2 . .. . . . 0 G= . . ... ... . . . ... . . . ... ... .

. . . . . . gN

... ... 0 ... ... ... 0 ... ... .. . ... ... .. . ... ...

(2.5)

Among other schemes equalization schemes, zero forcing or MMSE equalization (which takes into account the noise enhancement) is performed at the receiver. ZF equalization: gi =

h i |hi |2

1 hi

MMSE equalization: gi =

h i |hi |2 +i 2

i 2 is the noise variance on the carrier i. Of course, the coefcients (h i (k))i=1,...,N can either be known or estimated. The channel attenuations can easily be determined in the frequency domain thanks to a learning sequence [17, 77] or by blind estimation methods [105, 103]. Some useful estimation (also called denoising estimation) methods exploit also the time structure of the channel (limited number of coefcients..). It is also possible to take into account the time and frequency autocorrelation function of the channel for turbo estimation [62]. In classical standardized systems such as HiperLAN/2,

2.3. T HE

OFDM

25

two OFDM consecutive blocks are transmitted at the beginning of each frame to estimate the channel after synchronization and before the useful transmitted data. Note that after equalization, the noise variance changes from carrier to carrier depending on the channel frequency response. The decoder has to be fed with these modied metrics. Finally, the spectral efciency is increased by allowing frequency overlapping of the different carriers (compared to FDMA systems).

However, the OFDM system also exhibits several weaknesses relative to its single-carrier counterparts. OFDM does not capitalize on channel diversity, which prohibits the use of plain OFDM schemes in fading environments. The diversity achieved by the OFDM system can be less than a singlecarrier system employing the same error control code in a signaling environment rich in diversity. Indeed, due to frequency at fading, the transmitted information on one OFDM subchannel can be irremediably lost if a deep fade occurs [108]. Moreover, the Rayleigh behavior of such fading can have a dramatic impact on the performance of uncoded OFDM schemes. Methods based on coding (convolutional codes, block codes, multidimensional constellations, turbo-codes [8, 114, 23, 19, 55, 104, 126]) are usually employed with the use of interleaving to combat fading. When no channel knowledge is available at the transmitter, interleaving is intended to send the information on different carriers. If the receiver can be provided with different replicas of the information, which have been subjected to independent fadings, an appropriate combination of the replicas can restore the information. In the case of time interleaving, the coded bits are sent at different times with interval distances greater than the coherence time of the channel. This method is particularly useful in fast fading environment. Otherwise, it incurs a non-tolerable delay in the transmission. Frequency interleaving is particularly suited for rich scattering environments. The coded bits are sent on different frequency bands separated by the coherence bandwidth. Finally, in the case of space interleaving, the coded information is sent on different antennas. A particular simple space-time coding schemes which benets from the space diversity is the well-known Alamouti scheme [4]. As a special case of coded diversity used in OFDM, COFDM schemes [128] is usually used in standards. In classical COFDM standardized systems [57, 33, 123], the input bit stream is rst scrambled to generate random equal distribution bits at the input of the encoder. The bits are then processed by a convolutional encoder of rate R and constraint length L. The bits are then frequency interleaved in one OFDM block (time interleaving across OFDM blocks is not performed in HiperLAN/2) and mapped into symbols that are sent to the OFDM modulator. The memory size of the encoder is of critical importance. Indeed, the encoder performs a kind of redundant spreading of the information on the channel by linking the various bits through the memory of the encoder (see section.4.2.3 for the bit error rate performance evaluation with respect to the channel memory in Rayleigh fading channels). This spreading of the information can achieve, in some cases, full diversity. At the receiver, symmetrical operation are performed. Metrics are derived with frequency de-interleaving. These metrics are then fed into the Viterbi decoder in order to retrieve the coded bits before descrambling.

26

2. OFDM

scrambler

convolutional coder

frequency interleaving

mapping

OFDM modulation

channel

gaussian noise Viterbi decoder frequency deinterleaving metrics computation OFDM demodulation

descrambler

Figure 2.6: COFDM scheme. The baseband transmitted signal can also exhibit signicant amplitude uctuations over time, generating a high input backoff ratio at the amplier of the transmitter. Usually, the power amplier introduces non-linear distortions which destroy the orthogonality between the carriers. This peak to average power ratio (PAPR) [107, 16] increase has drawn intense research lately in order to decrease the power consumption of the ampliers [107, 101]. Most of PAPR methods are based on a modication of the transmitted signal by a correction vector. It introduces some non-negligible complexity at the transmitter. The vector correction is added to the frequency domain symbols yielding a new constellation with better peak to average power ratio properties (Tone Insertion method of Tellado [101]. In [71], an overview as well as implementations structures are given. The OFDM model only applies when the channel length is effectively smaller than the cyclic prex. If this is not the case, the orthogonality between subcarriers is only approximative and some Inter-Carrier Interference appear (ICI): a symbol transmitted on a given subcarrier is polluted by those of adjacent subcarriers. This problem occurs in particular in ADSL context and is solved by using a shortening lter at the receiver [69] whose aim is to reduce the channel length (or more rigorously to concentrate most of the energy in the rst L taps) so as to minimize ICI. OFDM is also more vulnerable to frequency off-set as well as synchronization problems [61, 79, 70]. In the rst case, a frequency off-set yields inter-carrier interference and destroys the orthogonality between subcarriers. In the second case, synchronization errors incurs a phase shift on the estimated symbols. In [79], sensitivity of OFDM systems to Carrier Frequency Offset and phase noise is analytically analyzed. It is shown in particular that OFDM is orders of magnitude more sensitive to frequency offset and phase noise than single carrier modulations. Sensitivity increases with the constellation size. The higher sensitivity of OFDM with respect to single carrier is mostly affected by the N times longer duration of an OFDM block symbol and by the intercarrier interference due to loss of the carrier orthogonality.

27

Chapter 3

LP-OFDM

3.1 Introduction

A multi-carrier OFDM system [15] using a Cyclic Prex for preventing inter-block interference is known to be equivalent to multiple at fading parallel transmission channels in the frequency domain. In such a system, the information sent on some carriers might be subject to strong attenuations and could be unrecoverable at the receiver. This has motivated the proposal of more robust transmission schemes combining the advantages of CDMA with the strength of OFDM known as OFDM-CDMA [45] [124] [56], in which the information is spread across all the carriers by a pre-coding matrix at the transmitter. Usually, for complexity reasons, a Walsh Hadamard matrix is used (to benet from the Fast Walsh Hadamard Transform). This combination increases the overall frequency diversity of the modulator, so that unreliable carriers can still be recovered by taking advantage of the subbands enjoying a high Signal to Noise Ratio (SNR). Although originally proposed for a multiuser access scheme, this concept is extended to all single user OFDM systems and is referred in the sequel as Linear Precoded OFDM (LP-OFDM).

The baseband frequency domain block equivalent model of a LP-OFDM system is depicted in gure3.1. The receiver front-end is formed by a symbol-matched lter followed by sampling at the symbol rate. The input symbol stream is serial to parallel converted, then the resulting K-dimensional symbol vector s(n) = (s1 (n), , sK (n))T (a white vector process with E s(n)sH (n) = IK ) is multiplied by a N K matrix WN,K where N K. This N-dimensional vector x(n) = W N,K s(n) is parallel to serial converted, and the corresponding generated data stream is sent across a non selective Rayleigh fading channel. After serial to parallel conversion, due to intercarrier interference generated by the precoder, the N dimensional received vector y(n) = (y 1 (n), , yN (n))T can be expressed as a function of the emitted symbol vector s(n):

28

3. LP-OFDM

(3.1)

where n(n) is a additive white Gaussian noise such that E n(n)n(n)H = 2 IN , and where HN = diag([h1 , . . . , hN ]) is the N N diagonal complex matrix bearing on its diagonal the channel gains. The role of matrix WN,K is to introduce diversity so that it allows to transmit each component of s(n) over N subcarriers.

PRECODING MATRIX

DETECTION METHOD

In the remainder of this thesis, channel knowledge and perfect synchronization at the receiver is assumed. The following two cases will be studied: Non-Ergodic Channel:

Let W be the system bandwidth i.e the frequency components of the transmitted signal occupy the band [ W , W ]. We will assume that the system bandwidth is much larger than the coher2 2 ence bandwidth of the channel. In the context of OFDM, the symbols are transmitted over a M-path multi-path channel of impulse response c(t) = M1 l g(t l ). Here, the complex gains l=0 (l )l=0,...,M1 are centered Gaussian random signals and the () l=0,...,M1 are the corresponding time delays. Both the gains and time delays of the paths remain static during the transmission of one K-dimensional codeword (but can change from codeword to codeword). Since this time variation is irrelevant here, it is not mentioned. In the following, we assume that the transmit g(t) lter with Fourier transform G( f ) is designed such as: W W , ] 2 2 W W G( f ) = 0 for [ , ] 2 2 G( f ) = 1 for f [

and the receiver lter is matched to the transmit lter. Although the following assumption is not compatible with the compact support assumption, we assume that c(t) is band-limited and 1 denote by W the corresponding bandwidth. Let T = W be the sampling rate, the equivalent discrete-time channel transfer function denoted h( f ) = L1 c(lT )ei2l f is given by h( f T ) = l=0

klm tn klm j

vx wx

d i h ddgf e

vu wu rs qo p klm n

3.3. S TATE

OF THE ART

29

M1 l ei2l f G( f ) where W f W . L is the channel order length. The channel frequency l=0 2 2 k1 attenuation on carrier k is given by: h k = L1 cl ei2l N (1 k N) with cl = c(lT ). l=0 Ergodic Channel: We assume fast fading environments in which time and frequency interleaving is performed on the components of x(n) = WN,K s(n). The purpose of interleaving is to scramble the precoded data stream in order to provide the coefcients (h k )1,...N with ergodic properties. This strategy incurs a delay in order to retrieve all the components of W N,K s(n) before de-spreading . In this case, considering that G( f ) = 1, the coefcients (h i )i=1,...,N can be considered as random variables with the same probability distribution. Let p(t) be the probability density of the random variables (|hi |2 )i=1,...,N . As far interleaving and asymptotic results are concerned, it is reasonable to make the following ergodic assumption: For each continuous bounded function f : R R, 1 N f (|hk |2 ) = N+ N k=1 lim f (t)p(t) dt almost surely

(see section Appendix.A for the denition of the almost sure convergence). In other words, we assume that the empirical channel distribution converges weakly to the limiting distribution given by the probability density function p(t). As a typical example, if the coefcients {h i }i=1,...,N are complex Gaussian1 (resulting in a Rayleigh fading), then p(t) = e t which corresponds to a 2 distribution with two degrees of freedom. It is important to notice that random variables {h i }i=1,N are not assumed to be independent. We set E(|hi |2 ) = 1 and E(|si |2 ) = 1 so that 2 dened by E(nnH ) = 2 IN represents the inverse of the SNR at the receiver input. The input SNR is dened as: 1 2

The LP-OFDM scheme is related to different schemes and a substantial amount of publications on these themes have appeared since the early 90s: OFDM-CDMA:

This scheme also known as MC-CDMA has drawn a lot of research interest lately. This type of modulation was rst proposed in 1993. The scheme was developed by different researchers, namely by N. Yee, J-P. Linnartz, G. Fettweis and K. Fazel [124, 46]. The basic idea is to combine the advantages of OFDM with the robustness of CDMA. It allows to use the spectrum in an efcient way and to exploit frequency diversity [56]. The transmission scheme for the downlink mobile communication is depicted in g.3.2. The transmitter consists of a spreader and an

a complex Gaussian random variable, we mean a complex random variable whose real and imaginary parts are independent Gaussian random variables having the same variances.

1 By

30

3. LP-OFDM OFDM multiplexer. Each users information is encoded separately and modulated directly onto a spreading sequence of length N. Each column of W N,K represents the code allocated to each user. Each receiver comprises an OFDM demultiplexer, a demodulator and a de-spreader. Equation (3.1) represents the signal received after guard interval suppression and Fourier transformation. In the frequency domain, the received N 1 vector signal can be seen as resulting from a transmission over N parallel at fading channels. Diagonal entries of H N represent the frequency domain channel gains. The information dedicated to the specic user is decoded by a Viterbi decoder.

TX

Channel

Figure 3.2: OFDM-CDMA scheme. In [45], Fazel proposes the use Walsh-Hadamard code spreading in the downlink (base to mobile terminal) and gives the practical performance of various detectors such as conventional, maximum likelihood detection and iterative detection with 64 users. When coding is applied, Lindner [67] shows through simulations that there is a trade-off between the signal diversity achieved by the spreading in MC-CDMA and the coding diversity of OFDM schemes. In particular, in [63], Kaiser conducts simulation studies on the performance of MC-CDMA systems with MMSE equalization and makes a comparison with MC-TDMA and MC-FDMA. He shows that MC-CDMA outper1 forms MC-FDMA and MC-TDMA signicantly for high code rates R > 2 and especially in the 1 uncoded case. When considering code rates R 2 , there is no signicant difference (<1dB) between MC-FDMA, MC-TDMA and MC-CDMA in the case of full system load. However, if not all the users are active simultaneously, he shows that the results differ signicantly in favor of MC-CDMA even at low code rates R 1 . 2 Single carrier transmission When a square FFT matrix is used as a precoder, the system is equivalent to a single carrier transmission with cyclic prex since the effect of the inverse Fourier Transform modulator disappears. The comparison of single carrier and multi-carrier system has been studied extensively for xed intersymbol interference (ISI) channels. With no cyclic prex, [3] has established that the the performance of an optimized multicarrier system when the channel is known exactly at the transmitter is strictly equivalent to an optimized single carrier system with a decision feedback equalizer but at a lower complexity. For transmissions with no channel knowledge at the transmitter (case of interest here), there has been limited work on the comparison of single-carrier and multi-carrier systems. In [122], Wornell shows that spread response precoding effectively transforms an arbitrary fading channel into a non-fading, simple marginally white Gaussian noise channel. Furthermore, spread-response precoding requires no additional power or bandwidth, and is attractive in terms of computational complexity ,robustness and delay constraint. He also argues that

z y

z y

z y

S/P

WN

hN

yk nN k

detector

yN k

z y

z y z y

z y z y

z y

h1

n1 k

y1 k

sp k

Viterbi Decoder

P/S

RX for user P

3.3. S TATE

OF THE ART

31

n(t)

Precoder

Channel

noise

Adaptive lter

the use of precoding in conjunction with coding has the potential to substantially reduce the computational complexity at both the transmitter and the receiver for a given level of performance. He points out that the effects of fading are entirely controlled by the precoder and therefore only additive noise remains for coding to control, which can in turn be achieved with comparatively shorter codes. With cyclic prex, the work of Fettweis [5] is also to our knowledge a precursor in this eld and part of section 4 will be inspired by this work. In particular, Fettweis [5] makes a comparison based on the cut-off rate for a multi-carrier system and a linear equalized single carrier system with cyclic prex. For code rates close to one or for uncoded transmissions, the linearly equalized single carrier system substantially outperforms the COFDM system. For low to medium code rates COFDM and a linearly equalized single carrier system perform approximately identically. He argues that for low SNRs preferring COFDM over a linearly equalized single carrier system is an implementation issue. If less transmitted bits are to be used for coding or at high SNRs, a single carrier with an MMSE equalizer can perform signicantly better and should be the preferred solution. As the linear equalizer can be implemented in the frequency domain (thanks to the cyclic prex), the complexity of the single carrier system is approximatively the same as of the COFDM system.

Channel Encoder Symbol Mapping Guard Interval Insertion

Channel Decoder

IFFT

Equalizer

FFT

Multi-dimensional constellation

As pointed out in the introduction, high modulation diversity can be achieved by applying a certain rotation to a classical constellation in such a way that any two points achieve the maximum number of distinct components. LP-OFDM can be seen as a pure OFDM scheme in which the emitted symbol belong to a multidimensional set s.

~ }|

Channel Guard Interval Removal

p(t)

h(t)

f t

DFE equalization

Noise

~ }

~ }

sn

~ }

yn

sn

32

3. LP-OFDM

y = HN WN,K s + n = HN s + n where s = WN,K s To give a better idea of this new constellation, let us consider the transmission of QPSK constellation sk :

Figure 3.5: Constellation before precoding s. With a 64 64 Walsh Hadamard precoding matrix, the resulting multidimensional constellation has components spanning the following space S (as shown in gure.3.6):

S = [p,k ] = s

642p 28

p

64 Each real component sp,. , taken independently, has P(p,. ) = 264 prior probability. As N increases, s the probability distribution of each individual entry s p,k behaves like a gaussian distribution. Even though the entries tend to have a Gaussian-like behavior, the resulting constellations are not independent. The better performance of LP-OFDM with respect to OFDM cannot be explained by the Gaussian entry argument (which is a condition to achieve capacity). To conrm our point, if we take a closer look at the capacity formula, we see that unitary precoding has no effect on the capacity:

C=

3.3. S TATE

OF THE ART

33

Precoding enables only to increase the diversity order (minimum number of distinct components for any two distinct constellation points), a criteria which is crucial in fading channels. If ML detection is performed at the receiver (dependencies across the carriers are taken into account), the Rayleigh fading channel can indeed be converted into a Gaussian like behavior channel when the precoder is appropriately chosen (and for high dimensional matrices). Two and four-dimensional rotations were rst found in [51], while the search for good high dimensional rotations needs algebraic number theory [18]. An interesting feature of the rotation operation is that the rotated signal set has exactly the same performance than the non rotated when used on a pure AWGN channel, while great enhancements are achieved over Rayleigh fading channels. The effect of precoding deserves a more detailed explanation. For explanation sake, we will show here that a simple rotation increases the diversity order from L=1 to L=2. Let assume a 2 carriers OFDM noiseless transmission with symbols s i (n) = 1 or si (n) = 1 with real Rayleigh independent fading: y1 (n) y2 (n) h1 0 0 h2 s1 (n) s2 (n)

This leads to 4 possible values for the input vector s(n) = [s 1 (n), s2 (n)] (see g.3.7). The diversity order (minimum number of components for any two distinct point of the constellation) is L=1. Now, let us assume that the Rayleigh behavior of the channel fades on h 2 (i.e h2 = 0). As a result, we lose the information carried by y 2 (n) and as a consequence s2 (n). To avoid this, one strategy is to use precoding. Precoding, in this case, means an orthonormal transformation R of the vector s(n) over the OFDM channel. y1 (n) y2 (n) where cos() sin() sin() cos() h1 0 0 h2 s1 (n) s2 (n)

This is equivalent with a rotation in space. If = 0[ ], then it is possible from y1 (n) to get the 4 information on s(n) = [s1 (n), s2 (n)] (g.3.8). By applying this rotation, the number of distinct components of the precoded symbol has been increased without changing the Euclidean distance. In this case, the diversity order is L=2. We have in this case an error free transmission. To give a better idea, let us derive in the general complex case (in this respect, refer the Phd thesis of J. Boutros [23]) the parameters of interest in order to optimize the precoders coefcients considering maximum likelihood detection. Dene x = W N,K s and z = WN,K r be two vectors of the constellation. Dene P(x z) the probability event that received signal y be closer to z than to x (x and z are the vectors obtained after the effect of the precoding matrix). Assuming that x is transmitted, the detector selects z if the following metric inequality is respected m(z | y, h) < m(x | y, h) with m(z | y, h) = N | yi hi zi |2 . The pairwise probability is thus given by: i=1

34

3. LP-OFDM

(1,1)

(1,1)

(1,1)

(1,1)

(1,1)

(1,1)

(1,1)

(1,1)

3.3. S TATE

OF THE ART

35

P(x z | h) = P = P

i=1 N

| y i h i zi |2 | y i h i xi |2

i=1

i=1

2Re(hi (xi zi )n ) | hi |2 | xi zi |2 i

i=1

Since 2Re N h(i)(xi zi )n is a Gaussian random variable with zero mean and variance 2 2 N | i=1 i i=1 hi |2 | xi zi |2 , we obtain: N | h i |2 | x i z i |2 i=1 ) 22

t2 1 e 2 x

2

dt.

The pairwise error probability P(x z) is obtained by averaging over the fadings h i P(x z) = P(x z | h) f (h)dh (3.2)

where f (h) is the probability density function of the fading coefcients. Searching for optimal precoders: The question arises on the choice of a good precoder for a channel with known statistics. Following the works of Boutros [18], let us give an upper bound of (3.2) in the case of Rayleigh independent fading channels:

P(x z) = =

1 N 1 |xi zi |2 2 i=1 1 + 2

4

x The last inequality is obtained by using the fact that: Q( x) 1 e 2 . Therefore, for large signal 2 to noise ratios: 1 2 xi i =z 1

|xi zi |2 42

P(x z)

An interesting point is that the precoder has to be designed in order to maximize the term xi =zi | xi zi |2 = i/vi (sr)=0 | vi (s r) |2 (vi is the ith row of matrix WN,K ). This optimization is nontrivial and depends on the modulation of the constellation s. The optimization may rely on number theoretic approaches as well as numerical optimization. For a review of matrices maximizing this

36

3. LP-OFDM criteria in dimension 4, 5 and 8, the reader may consult [66]. In [119], Giannakis extended these results to rectangular precoding matrices and i.i.d as well as correlated Rayleigh fading channels. Based on the fact that when the channel has nulls close to or on the FFT grid, OFDM faces serious problems for recovering the symbols; he proposes to linearly precode symbols as an alternative to various error-control coding techniques. By also performing pairwise error probability analysis, he upper bounds the diversity order of OFDM transmissions over frequency-selective fading channels. In particular, he shows that for i.i.d FIR Rayleigh fading channels of order L, maximum diversity order is achieved if any N L rows of the precoder are linearly independent. It turns out that rectangular linear precoding can be designed to guarantee maximum diversity order irrespective of the symbol constellation with minimum redundancy. Several constructions are also provided for maximum-diversity achieving precoders. Remarkably, all these previous studies have focused mainly on ML detection and small dimensional matrices. In the context of independent at fading channels, one strategy for increasing the performance when the matrix coefcients are appropriately chosen is to increase the systems dimension. However, this increase in dimension yields an increase in the decoding complexity and annihilates the benet of such a strategy.

In this document, we took a different approach and studied the performance of various precoders suited for certain linear (MMSE) and non-linear (SIC and PIC) equalizers in terms of overall performance (BER) and spectral efciency. These equalizers are very attractive in terms of implementation due to the structure of our model (specic MIMO channel which is the product of a diagonal frequency channel matrix by a N K precoder). In the case of suboptimum equalizers, a difcult task appears since the study does not rely anymore on the optimization of exponent error bounds (as the previous ML analysis). We took the SINR (Signal to Interference plus Noise Ratio) as a benchmark measure of performance of our precoding design. Moreover, the analysis can be conducted in the asymptotic regime (N , K , K ) since usual N wireless OFDM systems often involve a high number of carriers ( 64). However, for a specic precoder, the optimization is hardly trivial and entails the use of sophisticated mathematical tools. In order to derive interpretable expressions of the SINR, new tools borrowed from the free probability theory have been used. The basic idea relies on the fact that the expression of the asymptotic SINR can be easily derived if we model the precoder as a random matrix subject to certain constraints (orthogonality of the columns,. . . ). Indeed, free probability is able to predict the behavior of the empirical eigenvalue distributions of products or sums of hermitian matrices. Theses results are striking (in terms of closeness to simulations results with reasonable matrix size) and give us means to derive the SINR by only a few meaningful parameters (noise variance 2 , limit value of the ratio ratio K/N, frequency model of the channel). In particular, we are able to optimize the optimal amount of redundancy in terms of spectral efciency for these equalizers. We note that this subject has given rise to many publications in the year 2001 (known as the coding in the Galois or the complex eld issue see [119],[38],. . . ). For these types of matrices, we are able to compare the gains with usual precoders as well as other equalizers. Finally, in our quest to have everything for nothing , we will derive and optimize a simple yet efcient equalizer for the LP-OFDM scheme. Note nally that the model proposed for describing the system is broad enough to capture a multiplicity of transmission schemes, in particular multi-antenna schemes. Extensive research has been conducted in this eld since the publication of Telatar [100] and of Foschini [49]. H N is a N N matrix with non zero entries representing the channel between N transmitting and N receiving antennas. The design of N K precoders is related to the design of linear space time codes. The optimization of space time codes done by Tarokh et al. [96] can be seen as an extension of the signal space diversity scheme of Boutros

OF THE ART

37

38

3. LP-OFDM

39

Chapter 4

4.1 Introduction

Previous studies have already focused on the optimum design of precoding matrices for LP-OFDM, also known as signal space diversity schemes . In particular, [51] and [21] considered the case where K/N = 1, i.e. WN,K is unitary. In their papers, the entries of H N are assumed independent and identically distributed, and proposed to derive an upper bound of the error probability for the Maximum Likelihood (ML) detector of s(n). They discovered that, at least for high Signal to Noise Ratios, W N,K has to be chosen in such a way that the minimum so-called Ldistance product of the constellation {WN,K si , i I} be maximum, where {si , i I} denotes the set of possibles values taken by vector s(n). These studies were mostly based on results borrowed from number theory. As far as limited diversity channels are concerned, Giannakis [119] generalized these results to rectangular matrices. For a channel of diversity order L, Giannakis [119] showed that the N K matrices should have any N L rows linearly independent to achieve full diversity. However, the high computational cost of the ML detector prevents its use in practical contexts: as a typical example, N is in the order of 64 in HiperLAN/2 which makes 264 possible set to check for BPSK constellations. Actually, as a rst approach and due to its lower complexity, MMSE detection will be considered in this section. The impact of the choice of W N,K , the inuence of the orthogonality of the columns of W N,K and of the channel characteristics, the use of coding or not and of parameter K/N on the performance of the MMSE receiver will be investigated. Based on the fact that the noise generated after MMSE equalization is asymptotically Gaussian, our analysis will focus mainly on the SINR as a benchmark measure of performance. Indeed, the Signal to Interference plus Noise Ratio (SINR) at the output of each component of the MMSE detector can characterize entirely the performance of the modulation scheme equipped with a MMSE receiver. It shall be pointed out in the following document that coding stands for standard encoding schemes while Linear Precoding concerns the N K precoding matrix. As the time index n is irrelevant, we will omit it from now on. Based on the model of Chapter 3, the output of the MMSE detector s = [s1 , . . . , sK ]T is given by s = E syH E(yyH )

1

y

1

(4.1) y. (4.2)

40

4. MMSE

ANALYSIS OF

LP-OFDM

Each component sk of s is corrupted by the effect of both the thermal noise and by the "multi-user interference" due to the contributions of the other symbols {s l }l=k . Let us now derive the expression of the SINR at one of the K outputs of the MMSE detector. Let w k be the column of WN,K associated to element sk (n), and UN,K the N (K 1) matrix which remains after extracting wk from WN,K . Dene AN,K = HN WN,K WH HH + 2 IN . N,K N The component sk after MMSE equalization has the following form: s k = wk s k + k where and k = wH HH (AN,K )1 HN WN,K [s1 , . . . , sk1 , 0, sk+1 , . . . , sK ]T + wH HH (AN,K )1 n k N k N It has been shown in [80] and [127] that the additive noise k can be considered as Gaussian when K and N are large enough. Therefore, in order to derive the performance from a BER and spectral efciency point of view, an asymptotical analysis of the SINR is derived since the additive noise can be considered Gaussian in this case. k is therefore an asymptotically zero mean Gaussian noise of variance V = E(| k |2 | HN , WN,K ). Knowing that UN,K UH = WN,K WH wk wH , we get: N,K N,K k V = wH HH (AN,K )1 HN UN,K UH HH (AN,K )1 HN wk + 2 wH HH (AN,K )1 (AN,K )1 HN wk k N N,K N k N = wH HH (AN,K )1 [HN WN,K WH wk wH + 2 IN HH ] (AN,K )1 HN wk N,K N k N k = wk (wk )2 = wk (1 wk ) wk = wH HH (AN,K )1 HN wk . k N (4.3)

The SINR wk at the output k of the MMSE detector can thus be expressed as: wk = = = E[| wk sk |2 | HN , WN,K ] E[| k |2 | HN , WN,K ]

(wk )2 wk (1 wk ) wk 1 wk

Writing HN WN,K WH HH = HN UN,K UH HH + HN wK wH HH and invoking the matrix inversion N,K N N,K N K N lemma, we get after some simple algebra another useful expression for this SINR (see e.g. [102]): wk = wH HH HN UN,K UH HH + 2 IN k N N,K N

1

HN wk .

(4.4)

For a xed K and N, it is extremely difcult to get insight on the performance of the MMSE receiver from the expressions (4.4) and (4.3). Therefore, in order to obtain interpretable expressions, we will focus our asymptotical analysis in the next two sections on certain type of precoding matrices:

4.2. MMSE

(K = N)

41

Constant modulus unitary (K = N) matrices Random i.i.d matrices Random Haar isometric matrices (Isometric in the sense where the matrix columns are orthogonal) As we will show later on, the asymptotical assumption is realistic and particularly appealing, since in this case, the output SINR converge to deterministic values for certain type of matrices, that can be analytically expressed. Furthermore, results indicate a rather fast convergence to the asymptotic limits with nite size, demonstrating the practical importance of these limits.

4.2.1 Introduction

This section focuses on unitary precoding matrices (K = N, no redundancy added) since synchronization is perfectly controlled, at the emitter. As W N,N WH = IN , the outputs of the MMSE lter is the vector s N,N

s = WH diag N,N

Such a detection strategy is motivated by its very simple implementation (scalar channel equalization followed by a matrix multiplication). The SINR wk is still given by: wk = and wk 1 wk

wk =

We will consider hereafter the study of two types of matrices: Unitary matrices with constant modulus entries. Matrices such FFT and Walsh Hadamard belong to this category. FFT matrix : These matrices are dened by their (p, k)th element such as

j2pk 1 FN = [e N ]0p,k<N N

Walsh Hadamard matrix: These matrices are dened in a recursive manner: W(1)=1 and for all n > 1 1 W(2n) = 2 W(n) W(n) W(n) W(n)

42

4. MMSE

ANALYSIS OF

LP-OFDM

Random Haar unitary matrices. We introduce these matrices for calculus purpose and their utility will be revealed in section 4.3. In order to explain how these random matrices W N,N are generated. Some notations and denitions need be introduced. Denote by U the multiplicative group of N N unitary matrices, and by a random N N unitary matrix. is said to be Haar distributed if the probability distribution of is invariant by left multiplication by constant unitary matrices 1 . Since the group U is compact, this condition is known to be equivalent to the invariance of the probability distribution of by right multiplication by constant unitary matrices. In order to generate Haar distributed unitary random matrices, let X = [x i, j ]1i, jN be a N N random matrix with independent complex Gaussian centered unit variance entries. The unitary matrix X(XH X)1/2 is Haar distributed. To see this, notice that for each constant unitary matrix U, UX(XH X)1/2 = UX[(UX)H UX]1/2 . Since the probability distribution of X and UX coincide, matrices X(X H X)1/2 and UX((UX)H UX)1/2 have the same distribution. The above equality thus implies that UX(X H X)1/2 and X(XH X)1/2 are identically distributed. There is another way for building Haar distributed unitary matrices that will be useful to our context. Instead of multiplying X by the inverse of the Hermitian square root of X H X, one can introduce the uniquely dened upper triangular matrix with positive diagonal elements Q(X) dened by XH X = Q(X)H Q(X) . The unitary matrix V(X) dened by V(X) = XQ(X)1 (4.5)

is also Haar distributed. To see this, we rst remark that for each constant unitary matrix U, the probability distribution of V(X) and of V(UX) coincide. But, it is obvious that Q(UX) = Q(X), so that UV(X) = V(UX). Therefore, the probability distribution of V(X) and of UV(X) coincide. Remark that the columns of V(X) are obtained by a Gram-Schmidt orthogonalization of the column space of X.

Based on an asymptotical analysis, performance results of the MMSE receiver in the uncoded case (section 4.2.2) and in the coded scenario (section 4.2.3) will be derived.

First let study the performance of OFDM and LP-OFDM with MMSE equalization in the uncoded case. These results will then be used to obtain upper bounds on the bit error probability in the coded case.

Uncoded OFDM, Rayleigh fading Let rst recall the performance of uncoded OFDM. The expression of the received signal on the k th carrier is yk = hk sk + nk where the additive noise sequence (n k )k=1,...,N is white and Gaussian circular

1 In

other words, the probability distribution of coincides with the so-called Haar measure on U .

4.2. MMSE

(K = N)

43

with variance 2 . Assuming that QPSK symbols are transmitted, the conditional bit error probability given hk is: POFDM|hk = Q | h k |2 2

t 2 /2 dt. x e

0

| h i |2 2 1 ] 1 + 22

2 | hi | e|hi | d | hi |

1 [1 2

For high SNR, POFDM and log10 (POFDM ) log10 (SNRdB ). Due to the lack of diversity, the 2 slope of the BER curve is one in this case.

Uncoded LP-OFDM: General results We derive the BER for Walsh Hadamard matrices, FFT matrices and random Haar unitary matrices. In order to obtain interpretable expressions, let us rst evaluate the behavior of the SINR when N +. wk =

N

wk , 1 wk

(4.6)

For Walsh Hadamard and FFT matrices i.e | w i,k |= surely towards = E|h|2 ( |h||h| 2 ) for ergodic channels. 2 + =

|h( f )|2 2 df 1 2 |h( f )|2 +2

1 2

1 , N

wk =

1 N

For random unitary matrices, we also obtain an almost sure convergence towards the same results (the convergence in the L2 sense is immediate while the almost sure convergence is a consequence of the results presented in section 4.3)

4. MMSE

ANALYSIS OF

LP-OFDM

In all the following, when considering ergodic channels, each random variable | h l |2 is assumed to be 2 distributed with two degrees of freedom (while | h l | has a Rayleigh distribution). Therefore, = E|h| =

u

| h |2 | h |2 +2 x ex dx. 2 0 x+

In the following, denote by Ei (x) = x e u du the so-called exponential integral function, and F(x) = 1 1 1 1 e x Ei ( 1 ). It is easily seen that coincides with F( 2 ). x x =

0

u eu du u + 2

2

= 1 2 e Ei (2 ) 1 = F( 2 ) The SINR has a similar behavior when N + and it is straitforward to show that = lim wk =

N+

almost surely. 1

(4.10)

OFDM

LP-OFDM

Channel Distortion

Noise Pollution

Figure 4.1: Comparison of LP-OFDM and OFDM schemes in the Ergodic channel case. This remarkable result explains the good performance of LP-OFDM over OFDM. In OFDM, the amplitude carrying symbol sk is a random variable with a Rayleigh distribution while in LP-OFDM after MMSE equalization, symbol sk is carried by the term which can be considered as deterministic when N is chosen large enough. LP-OFDM after MMSE equalization thus transforms the N parallel fading channels of OFDM into N equivalent Gaussian channels with signal to noise ratio 1 .Our calculus

n1

N 0

h1

n1

N 0 2

4.2. MMSE

(K = N)

45

show therefore that, at least when N is large enough, the bit error probability of uncoded LP-OFDM after MMSE equalization is deterministic and is given by PLP-OFDM = Q 1 (4.11)

The results provided in this document for QPSK symbols can be of course extended to all other types of complex constellations as well as real constellations.

10

0

SOFDM performance theoretical bound, BPSK 64 independant carriers, BPSK theoretical bound, QPSK 64 independant carriers, QPSK

10

BER

10

10

10

7 8 Eb/N0 [dB]

10

11

12

13

14

15

In gure 4.2, we compare the BER given by formula 4.11 with the BER evaluated by a Monte Carlo simulations for BPSK (the performance is derived according to the methodology applied for QPSK constellations) and QPSK constellations. A Walsh-Hadamard precoding matrix has been used. The BER in the LP-OFDM scenario is dened as: Pe = 1 N Q N k=1 wk

wk Q , this implies that Pe Q Since for each k, Q The number N of carriers is chosen equal to 64, and the coefcients (h k )k=1,...,N are assumed circular Gaussian and independent and identically distributed. We recall that the independence assumption can be justied when assuming that for Linear Precoded OFDM, the precoded symbols x(k) are time and frequency interleaved on independent carriers. However, in this case, a decoding delay incurs at the receiver to retrieve the N symbols x(k) and perform block equalization. No coding is applied to the scheme. Note that the formula (4.11) is in agreement with the numerical results. The theoretical curve as well as the real context curve match pretty well in both cases and conrms that the addition of a precoding unitary matrix annihilates the fadings of the channel and leads to a quasi-Gaussian performance of the OFDM scheme with no additional bandwidth expansion while preserving simple linear detectors.

4. MMSE

ANALYSIS OF

LP-OFDM

It is extremely difcult to derive the pdf (probability density function) of in the non-ergodic case. Indeed, the SINR is a random variable linked to the number of paths of the time channel impulse response by a non-trivial equation. The evolution of the SINR with respect to the number of independent taps is an interesting diversity related topic and we will only analyze numerically the effect of the number of independent taps on the SINR distribution, considering a unit average power channel. We assume that the discrete-time channel impulse response values c(lT ) = c l , 0 l L 1 are i.i.d complex centered Gaussian variables, and that the average power of the channel response is 1. For a nite L, the asymptotic value of the SINR as N is a random variable, depending on the channel. We rst estimate the asymptotic mean bit error rate for several values of L. To better understand the inuence of the diversity on the distribution of the SINRs, we estimate the density of the SINR for xed values of E b /N0 . Diversity inuence on the mean bit error rate: 500 independent channels were simulated for each value of L. The corresponding asymptotic values of the SINRs were estimated according to the following equation:

1 2

1 2

1 2

1 2

The mean of the corresponding bit error probabilities E Q( ) were plotted in Figure 4.3, considering a QPSK constellation. One can notice that the case L = 1 gives the performance of a classical OFDM system: the channel matrix is of the form H = h 0 IN , so the output of the MMSE detector is s = WH HH HN WN,N WH HH + 2 IN N,N N N,N N = = |h0 |2 WH (WN,N s + n) |h0 |2 + 2 N,N |h0 |2 (s + n ) |h0 |2 + 2

1

where n = WH n has the same distribution as n. The output of the MMSE detector is the same N,N as that of a OFDM system without linear precoding. This remark shows that LP-OFDM schemes are useless with respect to OFDM for one path fading channels. However, as the number of paths increases, the performance improves as shown by gure.4.3. Effect of the diversity on the distribution of the SINR: In order to study the inuence of the diversity, we estimated the density of the SINR obtained during the previous simulations, for a Eb xed value of No . The results are presented in Figure 4.4. The results stated in the ergodic case can be veried in gure 4.4: the distribution of the SINR converges to a Dirac distribution, which conrms the fact that when the number of taps increases the SINR converges almost surely to a deterministic value. It is also remarkable that the mean value of the SINR decreases as the diversity increases, so does the mean bit error probability decreases. However, since the variance decreases, this has no impact in terms of BER as highlighted by gure.4.3.

4.2. MMSE

(K = N)

47

10

10

BER 10

3

10

10

"Theory" MMSE, 1 tap "Theory" MMSE, 2 taps "Theory" MMSE, 8 taps "Theory" MMSE, 128 taps 0 2 4 6 8 Eb/N0 [dB] 10 12 14 16

Figure 4.3: Channel diversity effect on MMSE mean performance, QPSK constellation.

Density of the SINRs for 1 tap 0.6 mean value : 20.0 0.4 0.2 0 0 10 20 30 40 Density of the SINRs for 4 taps 50 0.4 0.2 0 0 0.6

50

0.6 mean value : 8.8 0.4 0.2 0 0 10 20 30 40 Density of the SINRs for 128 taps 50

0.6 mean value : 7.2 0.4 0.2 0 0 10 20 30 40 Density of the SINRs for 256 taps 50

Eb No

30

40

50

Figure 4.4: Channel diversity effect on the distribution of the SINRs for a MMSE detector,

48

4. MMSE

ANALYSIS OF

LP-OFDM

In this section, we have shown that LP-OFDM with MMSE detection gives great means of improvement over OFDM. However, every system is doomed to work with error correcting codes as an efcient mean to transmit reliable data. It is remarkable to observe that the great improvement brought by precoding with respect to OFDM vanishes when convolutional coding. A analysis of the performance of coded LP-OFDM with respect to OFDM is detailed next section.

In this section, we focus our study to the ergodic channel case with Rayleigh fading. Kaiser et al. [63] and Lindner et al. [67] showed by simulations that the diversity improvement induced by the precoding matrix is considerably reduced when coding is applied in the COFDM case. Indeed, the high redundancy of low code rates (which already performs a kind of spreading) in conjunction with the suboptimality of the MMSE receiver reduces the performance with respect to COFDM. This section is intended to analyze more rigorously this observation. Analytic expressions of upper bounds of the bit error probability provided by the MMSE receiver are derived for a Rayleigh fading channels when the number N of carriers is large enough. When coding and interleaving is applied, the inuence of the coding rate is studied on the performance of COFDM and 1 LP-OFDM. It is shown in particular that coded LP-OFDM outperforms COFDM if the input SNR = 2 is greater than a given threshold which only depends on the constellation size. This result is believed to be quite useful since it provides new insights for choosing between a COFDM or coded LP-OFDM modulation scheme in a given context. The results of Kaiser et al. [63] and Lindner et al. [67] can then Eb be all explained theoretically when considering the relation between the SNR , No , the constellation size and the code rate. Indeed, denote k the number of bits conveyed by the modulation symbol s i (n). Eb The relation between the input signal to noise ratio , No , k and the code rate R is : = k.R. Eb No (4.12)

E For a xed threshold , as R decreases, the threshold Nb increases. The high redundancy of low code o rates introduces a kind of information spreading in COFDM. This favors COFDM with respect to coded Eb LP-OFDM as characterized by the higher threshold in terms of No .

In this thesis, we will analyze two types of coding strategies with MMSE equalization for COFDM and coded LP-OFDM: Random coding Convolutional coding Random Coding The random coding approach brings out clearly and simply the relationships between error probability, data rate, constraint length and channel behavior. It provides a useful bound to the rates where ML decoding can be practically used. Therefore, one can compare modulation systems on the basis of their exponent error. Fettweis [5] already introduced the random coding approach to study single carrier with

4.2. MMSE

(K = N)

49

cyclic prex versus multi-carrier transmissions. However, he only focused on the outage probability for the two-ray fading channel (two paths) with QPSK as a fundamental modulation technique. We extend these results to any kind of modulation and show the impact of increasing the modulation size on the performance of LP-OFDM. In the case of LP-OFDM, we consider random coding on each frequency-stream as depicted by g 4.5. Indeed, after MMSE equalization, the noise is correlated across the carriers: taking into account the matrix correlation in the optimal decoding of the random codes would be prohibitively (in term of complexity) to expensive. Each substream is therefore encoded separately. We consider a constellation s of vectors of the Euclidean N-dimensional space C N , to be used for transmission over the fading channel. The mapper associates a k-tuple of input bits to a signal point s = (s(1), ..., s(N)) with s(i) A and card(A)=M. Let 2k be the total number of signal points in the constellation i.e s l (1 l 2k ). An interleaver is applied prior to channel ltering in the system model as in the uncoded case. Independence of the fading samples represents the situation where the components of the transmitted points are perfectly interleaved. In order to derive the probability of error with random coding for OFDM and LP-OFDM, we consider the general model case r(i) = a(i)x(i) + n(i) For OFDM, r(i) = yi and a(i) = hi is a random variable, | hi | having a Rayleigh distribution and variance 1. n(i) is a Gaussian noise with variance 2 = 2 both For LP-OFDM after MMSE equalization, r(i) = s i and a(i) is a constant term . n(i) is a Gaussian noise with variance 2 = (1 ) both

OFDM FLAT FADING CHANNEL

Encoder

Channel Distortion

Noise Pollution n1 N 0

Random Encoder

Random Encoder

LP-OFDM

Let P(sl sq /a) denote the pairwise probability i.e : the probability that when s l is transmitted, r is

nN

N 0

Random

S/P

hN

nN

N 0 2

P/S

Optimal Decoder

1

h1

n1

N 0 2

Optimal Decoder

1

Optimal Decoder

GAUSSIAN CHANNEL

4. MMSE

ANALYSIS OF

LP-OFDM

i=1 N

i=1 N i=1

i=1 N

i=1

i=1

Since = 2Re N a(i)(sl (i) sq (i))n(i) is Gaussian with zero mean and variance 2 = 22 N | i=1 both i=1 a(i) |2 | sl (i) sq (i) |2 . The conditional pairwise error probability is therefore given by: P(sl sq /a) = P > | a(i) |2 | sl (i) sq (i) |2

i=1 N

= Q

where .

1 Q(x) = 2

et

2 /2

dt

By upper bounding the Gaussian tail function by an exponential Q(x) < e 2 , we obtain: P(sl sq /a) e

x2

e

i=1

Thus, instead of attempting to nd a single set of 2 k coded waveforms for which we compute the error probability, the random coding approach considers the ensemble of M N distinct ways in which one can select 2k vertices from the M N available vertices [81] (in this approach, we take also into consideration the case of choosing 2k times the same code). In order to derive simple bounds, we will consider a M-PSK modulation, s(i) = e M with 0 m < M 1. Now, we can average P(sl sq /a) over all the ensemble of codes and over the channels. Since all the codes are equally probable, the component s l (i) is equally likely to be one of the M possibilities and is statistically independent from all the other components of vectors s l and sq . The mean value of the pairwise probability is thus bounded by:

j2m

2k

4.2. MMSE

(K = N)

51

Of course, we observe that the right side of the above equation is independent of the specic indices g l and q. The probability of error for a mapping g of k bits of information (S k ) into a random selection of k waveforms is upper-bounded by: 2

g P(Sk )

l=1l=q

2k

P(sl sq )

N

(p j) |a(i)|2 |sin( M )|2 2 both

g The probability of error is nally obtained by averaging P(S k ) over all k-bit information sequences g (a number equal to 2k ). Since P(Sk ) is independent of g:

Pe

1 2 g P(Sk ) 2k g=1

N M1 M1 |a(i)| p=0 j=0

2 |sin( (p j) )|2 M 2 both

g P(Sk )

2 k Ea

1 2 i=1 M

N M1 M1 |a(i)| p=0 j=0

2 sin2 ( (p j) ) M 22 both

(4.13)

Pe 2kNC0

= 2NC0 . (4.14)

Pe 2N(C0 C)

k where C = N is the code rate (not the coding rate!) while C0 is the cut-off rate. We see that the cut-off rate plays a central role. Indeed, when C < C0 , the probability of error Pe 0 as the code block length N . The cut-off rate gives us the limit in terms of the rate for a specic type of modulation. We will emphasize our analysis on the redundancy factor also known as the coding rate dene by: R = Nlogk2 (M)

C0 log2 (M) .

52

4. MMSE

ANALYSIS OF

LP-OFDM

For the OFDM channel, (ai )i=1,...,N are i.i.d random variables with a Rayleigh distribution and 2 both = 2 : (p j) |a(i)|2 sin2 ( M ) 1 1 M1 M1 2 both RO,OFDM = log Ea e log2 (M) 2 M 2 p=0 j=0 M1 M1 1 1 1 = log2 2 sin2 ( (p j) ) log2 (M) M p=0 j=0 M 1+ 2 For the LP-OFDM channel with MMSE equalization, a = and 2 both = (1 ) R0,LPOFDM = 1 log log2 (M) 2

( M 1 M1 M1 sin (1) e 2 M p=0 j=0 2 (p j) )

1 We have plotted in g 4.6, g 4.7, g 4.8 and g 4.9 R 0 versus the input SNR = 2 for respectively BPSK, QPSK, 8-PSK and 20-PSK. We observe that an SNR threshold appears in the rst three case for which coded LP-OFDM outperforms COFDM. This threshold depends on the type of constellation (1 dB for BPSK, 6.5 dB for QPSK and 12 dB for 8-PSK) and disappears as the size of the constellation increases (g.4.9). This interesting results is due to the fact that as the size of the constellation increases, the emitted multi-dimensional codes have good hamming distances properties. In this case, the precoder will have nearly no effect in terms of diversity. As an extreme case, one can take codes with independent Gaussian entries. Since the precoder is unitary, precoding will have no effect on the distribution of the Gaussian code vector. In this case, precoding is useless. In other words, results depend on the type of constellations one is going to use. If the transmission scheme uses BPSK or QPSK constellations, LP-OFDM pays off while for higher constellations, the precoding idea become less useful. Another remarkable result is that for QPSK constellations, coded LP-OFDM outperforms COFDM if the redundancy factor is higher than 0.55 while the inverse occurs for a redundancy factor less than 0.55. This result conrms the simulations of the papers [63, 67] with QPSK constellations.

4.2. MMSE

(K = N)

53

1 0.9 0.8 0.7 0.6 Redundancy 0.5 0.4 0.3 0.2 0.1 LPOFDM OFDM

1 0.9 0.8 0.7 0.6 Redundancy 0.5 0.4 0.3 0.2 0.1 LPOFDM OFDM

54

4. MMSE

ANALYSIS OF

LP-OFDM

1 0.9 0.8 0.7 0.6 Redundancy 0.5 0.4 0.3 0.2 0.1

3.5

2.5

4.2. MMSE

(K = N)

55

Convolutional coding with QPSK constellation We study in this section the performance of the Viterbi algorithm in the case of soft decision decoding according to the method used in [81] when convolutional coding is used at the transmitter. in (g.4.10), a COFDM and coded LP-OFDM is described.

Convolutive Encoder

Interleaver

Mapper

Convolutive Encoder

Interleaver

Mapper

TX

COFDM In classical standardized COFDM systems, the incoming input bit stream is rst convolutionally encoded with a code rate of R, interleaved and punctured. The resulting bits are then mapped onto a QPSK constellation for forming symbols that are distributed over all the carriers. It is assumed that a soft-output Viterbi algorithm is used to decode the transmitted bits. In order to specify on which signal the Viterbi algorithm operates, we denote (y(m)) mZ the sequence obtained after parallel to serial conversion of the received sequence, i.e. the scalar sequence . . . y 1 (m), y2 (m), . . . , yN (m), y1 (m + 1), y2 (m + 1), . . . , yN (m + 1), . . .. The same notation is applied to variables h, s and n so that we obtain: y(m) = h(m)s(m) + n(m). In COFDM systems equalization is performed by applying the real and imaginary part operators to h(m) sequence ( 2 |h(m)| y(m))mZ , the corresponding real sequence being nally de-interleaved. Thanks to the de-interleaver, the resulting real valued discrete-time signal is denoted by z(m) = (m)b(m) + u(m) where represents an i.i.d sequence of Rayleigh distributed random variables of second order moment 1, b is the BPSK sequence obtained by mapping to +1 and -1 the bits generated by the convolutional encoder, and u is an i.i.d. sequence of real Gaussian random variables of variance 2 . The Viterbi decoding algorithm processes sequence z. In order to evaluate upper bounds on the BER, one usually rst evaluates the probability PCOFDM (d) of deciding P1 instead of P0 where P0 is the path of the Viterbi algorithm trellis associated to the transmitted sequence and P1 is a path which differs by d bits from P0 . Following [81], let S1 = [b1 (1), . . . b1 (M)] be the sequence corresponding to path P1 and S0 = [b0 (1), . . . b0 (M)] the sequence corresponding to path P0 . For sake of simplicity and without loss of generality, assume that b1 (k) = b0 (k) for k = 1, . . . , d and b1 (k) = b0 (k) if d < k M. The probability of deciding P1 instead of

sK k

Coded LP-OFDM

Channel

S/P

sk

s1 k

WN

hN

yk nN k

MMSE

yN k detector sK k

h1

n1 k

y1 k

sN k

Convolutive Encoder

Interleaver

Mapper

S/P

hN nN k

yk hN

yN k

s1 k

COFDM

h1 n1 k

h1 y1 k

P/S

Viterbi Decoder

s1 k k

Viterbi Decoder

P/S

Viterbi Decoder

RX

56 P0 is: Pe = P

k=M k=1

4. MMSE

ANALYSIS OF

LP-OFDM

k=1 k=M k=1

k=M

= P 2 = P 2

(k)z(k)(b1 (k) b0 (k)) < 0 (k)((k)b0 (k) + u(k))((b1 (k) b0 (k))) < 0

k=M k=1

Since b0 (k)b1 (k) = 1 for k = 1, . . . , d, the probability of d errors is: PCOFDM (d) = P 2 (k)2 + (k)u(k)(b1 (k) b0 (k)) < 0

k=1 k=1 k=d k=d

As b1 (k) b0 (k) can take the two values -2 or 2, the previous expression simplies to: PCOFDM (d) = P (0, 2 2 ) > k = P (0, 1) > = Q

k=1 k=d k=d k=1

2 k

d 2 k=1 k 2

Where (0, 1) is a zero mean Gaussian variable with unit variance. The mean error probability is : PCOFDM (d) = E Q

0

Using that d 2 is 2 -distributed (2d degrees of freedom), we get that k=1 k PCOFDM (d) = Q y 2 1 yd1 ey dy. (d 1)!

d 2 k=1 k . 2

After some calculus (see Appendix K and [81] for the demonstration), the probability PCOFDM (d) can be written as PCOFDM (d) = 1 ] 1 + 22 d1 1 1 l . C2l1 1 2 1 + 2 (2( 2 + 2))l l=1 1 [1 2

4.2. MMSE

(K = N)

57

The evaluation of the probability of the above error events leads to the following BER upper bound: PCOFDM d PCOFDM (d) d=dmin

(4.15)

Here, dmin is the minimal distance of the code, is the number of input bits in the encoder and d is the number of incorrectly decoded information bits, for each possible incorrect path differing from the correct one by d bits.

Coded LP-OFDM with WH, FFT or random unitary matrix In the context of coded LP-OFDM, one Viterbi decoder is applied on each subband k and processes the real and imaginary parts of the signal at the output of the Wiener lter. Denote z k (m) the corresponding real symbol. According to section 4.2.2, z k (m) = bk (m) + uk (m) , bk is the BPSK sequence obtained by mapping the bits at the output of the convolutional encoder, and u k is a real Gaussian i.i.d. of variance 2 both = (1 ). Since is deterministic, the bit error probability in the pairwise comparison of two paths that differ by d bits is given by : d l=1 PCLP-OFDM (d) = P (0, 1) > d l=1 (1 ) 1 dF( 2 ) = Q 1 (1 F( 2 ))

1 We recall that = F( 2 ) (see equation (4.9)) and (0, 1) is zero mean Gaussian variable with unit variance. The overall error probability is thus bounded by:

PCSOFM

(4.16)

COFDM versus coded LP-OFDM. In the previous section, we have derived bounds to evaluate the bit-error probability knowing the transfer function of the code (i.e d and , the number of incorrectly decoded information bits, for each possible incorrect path differing from the correct one by a certain number of bits). However, the most important point of this section lies in the observation that for any rate R convolutional code, the difference l(d) = PCOFDM (d) PCLP-OFDM (d) is positive for all d > 2 if the input SNR is greater than a threshold. This is an important issue since formulas (4.15) and (4.16) in conjunction with the above properties of l(d) imply that for any code of rate R, coded LP-OFDM outperforms COFDM if Eb/No is higher than a threshold which depend only on R, and not of the particular chosen code. This claim is illustrated in gure 4.11 and gure 4.12:

58

4. MMSE

ANALYSIS OF

LP-OFDM

20 19 18 17 16 15 14 13 log(PCOFDM/PCSOFDM) 12 11 10 9 8 7 6 5 4 3 2 1 0 1 2 3 0

SOFDM versus COFDM d=3 d=4 d=5 d=6 d=7 d=8 d=9 d=10

7 8 lambda

10

11

12

13

14

15

12

11

10

lambda

10 11 distance d

12

13

14

15

16

17

18

4.2. MMSE

(K = N)

59

1 In g.4.11, we have plotted log(PCOFDM (d)/PCLP-OFDM (d)) versus the input SNR = 2 . We observe that for a xed distance d, there exists a threshold lim (d) such as l(d) < 0 for < lim (d) while l(d) > 0 for > lim (d).

The purpose of g.4.12 is to show that lim (d) is nearly independent of the distance d. For this purpose, the threshold lim (d) (corresponding to the solution of equation l(d)=0) versus various values of d = 3, . . . , 18 is plotted. For usual codes (d higher than 5), the threshold varies only by 0.6 db ( 7.5dB < lim (d) < 8.1dB for all d > 5). In practice, these results show that regardless of the code structure (i.e the expression of d in our previous analysis), coded LP-OFDM outperforms COFDM if the signal to noise ratio is greater than a threshold included in a tight interval. These results gives us tools to choose between coded LP-OFDM Eb and COFDM for a given couple (R, No ) and conrms that a trade-off between spreading and coding is to be considered when designing a coded LP-OFDM scheme .

E Indeed, an interesting feature is the relation between the code rate and the threshold Nb . Recall o limit that the number of bits conveyed by the modulation symbol s i is 2. The relation between the input signal E to noise ratio lim , Nb and the code rate R is : o lim

lim = 2R

Eb No

(4.17)

lim

These results are conrmed in the following table characterizing the for various code rates. limit R= 1 4 R= 1 3 R= 1 2 R= 2 3 R= 3 4 minimum 7.5dB 10.5dB 9.2dB 7.5dB 6.2dB 5.7dB maximum 8.1dB 11.1dB 9.8dB 8.1dB 6.8dB 6.3dB

Eb No

limit

in a condence interval

Table 4.1: Condence interval vs. code rate. Our analysis thus conrms the experimental results observed by simulation by Kaiser [63] and the previous theoretical discussion on the cut-off rate. Figures 4.13 plots the simulated performance of respectively two constraint length K = 7 convolutional encoders rate 1 and 2 . It conrms the theoretical behavior inferred from the board.4.1. For each 2 3 code, a threshold is observed which belongs to the condence interval of the table.4.1. For R = 2 , the threshold is: 6.5dB ( Condence interval 6.2dB-6.8dB) 3 For R = 1 , the threshold is: 8dB ( Condence interval 7.5dB-8.1dB) 2

60

4. MMSE

Convolutional Code (K=7) QPSK constellation

ANALYSIS OF

LP-OFDM

10

COFDM rate 2/3 coded SOFDM MMSE rate 2/3 COFDM rate 1/2 coded SOFDM MMSE rate 1/2

10

10 BER 10

10

10

7 8 Eb/N0 [dB]

10

11

12

13

14

15

4.2.4 Conclusion

In this section, we have addressed the problem of designing and analyzing the performance of an LPOFDM scheme equipped with a minimum mean square error receiver in the coded as well as in the uncoded scenario. In the uncoded case, we have quantied the improvement brought by linear precoding for Ergodic channels. In the non-ergodic case, the diversity improvement with respect to the number of taps has been studied; in particular, the results encourage the use of time and frequency interleaving as a mean of increasing the performance. In the coded case, theoretical tools for selecting between a COFDM or coded LP-OFDM modulation scheme in a given context have been provided. We have derived efcient bounds to evaluate the bit-error probability for COFDM and coded LP-OFDM schemes. It has been shown that coded LP-OFDM outperforms COFDM if the signal to noise ratio is greater than a threshold which depends on the type of constellation. More precisely, LP-OFDM is useless for high constellations since the latter benet already of good hamming (in the sense given by Boutros and Belore) distances. However, for constellations such as BPSK, QPSK, Linear Precoding enables to increase the diversity in a certain range of SNR which has been determined explicitly for random and convolutionnaly coding.

4.3. MMSE

61

4.3.1 Introduction

In the case K = N, it is extremely difcult to get interpretable expressions from equation 4.4. Such an expression depends explicitly on the coefcients of the specic precoder used. In order to provide useful expressions, several papers [113, 44, 102] have recently analyzed the behavior of the SINR at the output of the MMSE detector in the uplink CDMA case where the entries of W N,K are independent and identically distributed random variables (to be referred to in the sequel as the i.i.d. case) and the matrix H N (n) is reduced to IN and the various users can have different powers. The i.i.d model is relevant in the uplink CDMA context since there is no synchronization between users. In this case, it has been shown that this SINR converges almost surely towards a well dened deterministic value which does not depend on the particular realization of WN,K . These results have been used by several authors to better understand the performance of the chosen transmitter/receiver chain. In particular, Biglieri et al. showed in [11] how to K use these results to nd the optimum value of the ratio N = via the analysis of the system throughput. We also note that [74] considered the case where the columns of W N,K are independent and identically distributed Ndimensional random vectors uniformly distributed on the unit sphere of C N . In this section, we also study the behavior of the SINR at the output of the MMSE detector assuming a random matrix WN,K and taking benet of the corresponding results to discuss the choice of . However, we address more specically the case where W N,K is isometric (the column codes of the matrix are orthogonal). In the sequel, this topic will be called the isometric case. The design of an isometric precoding matrix is common in signal space diversity systems or precoded OFDM (with Walsh-Hadamard precoders), for it provides better results than the choice of an i.i.d. matrix. Moreover, we stress on the fact that in the LP-OFDM model, matrix H N is not reduced to identity. To our knowledge, the problem we address here has not been considered in previous works. For calculus purposes, the Linear Precoder is assumed to be an isometric matrix extracted from a Haar distributed random unitary matrix. The Haar distribution is selected because of the resulting properties to be detailed later. One of theses properties has already been studied in section 4.2.1. being one such matrix, its probability distribution is also invariant under right multiplication by any unitary matrices, hence this distribution coincides with the distribution of P for any permutation matrix P. This shows that the N K isometric matrices obtained by extracting any subset of K columns from have the same probability distribution. In the sequel, it will be assumed that W N,K is generated by extracting any K columns from a N N Haar distributed unitary matrix independent of H N . Note that for usual precoded OFDM systems, precoding matrices are not generated this way. Walsh-Hadamard matrices, whose entries are binary, cannot in particular be considered as realizations of Haar distributed random unitary matrices. However, simulations in the following section show that our theoretical evaluations t with Walsh-Hadamard linear precoded systems. Note also that Haar distributed random unitary and Walsh Hadamard matrices provide the same asymptotic performance if = 1. From a technical stand point, the i.i.d. precoding case study of [102] leans on mathematical results related to the "limiting distribution of eigenvalues" of some large random matrices with independent and identically distributed entries (cf. [93]). As for the isometric case, a considerably more involved material will be needed. The results given here rely on the so-called free probability theory initially developed by D. Voiculescu [117] in order to solve problems of operator algebra classication. At the end of the eighties, Voiculescu realized that this theory could also be used to analyze the eigenvalue distribution of sums or products of certain independent large random matrices. This will be the starting point of our analysis.

62

4. MMSE

ANALYSIS OF

LP-OFDM

In section 4.3.2, we present our main result (theorem 1), which asserts that the SINR at the output of the MMSE detector converges almost surely to a deterministic value. At the end of this section, an asymptotic result for the i.i.d. case is also given for reference. Section 4.3.3 conrms by simulations the fact that our asymptotic analysis allows to predict the performance of the MMSE detector for relatively small values of N. An evaluation of the whole system throughput (cf. [10]) is performed with respect to for a xed allocated bandwidth. The purpose of this computation is to determine the optimal amount of redundancy that should be spent on the linear precoder in a system where linear precoding is combined with convolutional encoding. This analysis is sustained by a performance comparison between the linearly precoded system and a system equipped with a classical convolutional forward error correcting code in addition to linear precoding, the overall coding rate being the same for both systems. We conrm by practical examples the fact that the redundancy tradeoff between linear precoding (through the choice of ) and coding (through the choice of the convolutional code rate) can be given priori by the throughput analysis. This discussion was mainly motivated by the recent paper [119]. As we believe that free probability and random matrix theory are important and promising tools, we propose to give in appendix B and C a comprehensive introduction to its most important aspects concerning our purpose. In section D, we show how to apply an important lemma of [6] (also used in [44]) to our particular situation, and nally use the almost sure asymptotic freeness result presented in the recent monograph [58] to prove theorem 1. Note that Evans and Tse were the rst to introduce Free Probability in the context of performance analysis [44] but for solving quite different problems in the CDMA context.

In this section, we derive the following result in form of a theorem: Theorem 1 : for all WN,K random Haar Isometric matrices chosen as above, when N and K/N 1, the SINR wk of the kth symbol at the output of the MMSE equalizer converges almost surely to a value that is the unique solution of the equation in the ergodic case: (p(t) is assumed to be compactly supported)

0

t t + 2 (1 ) + 2

p(t) dt =

. +1

(4.18)

1 2 1 2

(4.19)

Note that in the ergodic case, the SINR converges to a deterministic constant while in the non-ergodic case, the limiting SINR is random variable depending on the channel realization (but not the precoder realization). Bear in mind that function p(t) is here assumed to be compactly supported. This assumption is important on a technical point of view because the most powerful results of free probability theory (in particular the asymptotic freeness of independent large random matrices) need compactly supported measures. Although the usual probability distributions of the coefcients (h i )iN , like the Rayleigh or

4.3. MMSE

63

the Rice distributions, do not verify this assumption. This restriction is of course not very important in practice. In particular, the use of formula (4.18) with p(t) = e t as is the case of Rayleigh channels allows to predict quite well the performance of our precoded modulation scheme using MMSE detection. When the measure p(t)dt is replaced by the Dirac measure (1) at point t = 1, matrix H N is reduces IN . By direct computation, the receiver output SINR in this situation (called the Gaussian channel situation) is easily shown to be 1/2 . Only an outline of the main steps is provided below. The full proof can be found in Appendix D: Step 1: we rst show that wk converges almost surely to independent of the particular matrix realization. Recall that wk = wH BN wk with k BN = HH HN UN,K UH HH + 2 IN N N,K N

1

HN

As wk and BN are not independent, the proof is a bit complicated. Indeed, if w k and BN were independent, then it is possible to show [102] that wk 1 a.s tr(BN ) 0 N

In our case, the proof follows a different procedure. We rst show that wk 1 tr[(I UN UH )BN (I UN UH )] 0 N N N K

1 And then we use free probability results to show that NK tr[(I UN UH )BN (I UN UH )] converges N N to a deterministic constant . Indeed, since (I U N UH )BN (I UN UH ) can be approximated by N N a non commutative polynomial of UN UH and HH HN . As UN UH and HH HN are almost surely N N N N 1 asymptotically free (see appendix C for denitions and properties), this implies that NK tr[(I UN UH )BN (I UN UH )] converges to a deterministic constant . If we put the pieces together, we N N see that wk . Once we have determined the convergence of wk , the rest of the proof will explicit the limit . k Step 2: The previous step implies that wk = w +1 +1 . For a given N, there are K quantities k wk each corresponding to the choice of a particular column code in W N,K . Their sum over all the 1 columns of this matrix is K K wk . Hence, the limit value is given by : k=1 w N

1 tr (HN WN,K WH HH + 2 IN )1 HN WN,K WH HH N,K N N,K N K 1 = lim tr (AN,K + 2 IN )1 AN,K N N 1 1 N i 1 = lim = lim dN () 2 N N i=1 i + N + 2

K

lim

when AN,K = HN WN,K WH HH and (i )i=1,...,N denote the eigenvalues of AN,K and N is the probN,K N 1 ability measure on R+ dened by dN () = N N (i ). In the context of random matrices i=1 theory, N is called the empirical eigenvalue distribution of A N,K . step 3: by applying free probability results, we can show that N converges almost surely to a compactly supported measure , which can be derived explicitly. Therefore, converges almost t 1 surely to: t+2 d(t) . This implies that wk converges to a deterministic value 1 denoted and solution of 4.18 and (4.19).

64

4. MMSE

ANALYSIS OF

LP-OFDM

One of the key information provided by theorem 1 is that the SINR converges to a deterministic constant when N and K/N . Another interesting aspect is that when 0, the performance of the optimal detector is achieved with MMSE detection. Indeed, in this case, the following results are achieved: in the ergodic case:

0

(4.20)

1 2

1 2

(4.21)

Similar results were obtained in the quite different context of uplink CDMA with at fading in the paper of Verdu et al. [90]. It is shown that when the load is low ( the load in the CDMA context corresponds to the ratio of the number of users by the spreading length of the code), linear and optimal detectors perform similarly. This is due to the fact that the high redundancy introduced compensates the suboptimality of linear receivers. The gain of non-linearity is mostly achieved for high loads. Equations (4.18) and (4.19) have a unique solution. As a matter of fact, the left hand side of (4.18) is a positive decreasing function of which converges to 0 when +, while the right hand side of (4.18) is an increasing function of which is 0 at = 0. The case > 1 of course does not make sense in the isometric case. However, instead of using a precoding matrix whose columns are orthogonal, one may use if > 1 a matrix W N,K whose rows are orthogonal. In this context, we model the precoding matrix W N,K as a random matrix obtained rst by extracting N rows from a Haar distributed unitary K K matrix, and second by multiplying the resulting matrix by the scaling factor 1/2 . Therefore, WN,K satises WN,K WH = IN N,K (4.22) The scaling factor in (4.22) guarantees that the average power allocated to each component of s is 1 as is the case when < 1. Derivation of the MMSE output SINR in the asymptotic regime is similar to the case where = 1. To be precise, as WN,K WH = IN , wN is given by N,K wN =

k=1

|wN,k |2 |hk |2 + 2

|hk |2

(4.23)

These results are therefore easy to derive (no need of Free Probability Theory) since W N,K WH = I. N,K in the ergodic case:

0

t p(t) dt = . 2 t + +1

(4.24)

where p(t) be the probability density of the random variables (|h i (n)|2 )i=1,...,N . in the non-ergodic case:

1 2

1 2

(4.25)

4.3. MMSE

65

We should emphasize that > 1 is not at all an unrealistic case: even though in the LP-OFDM case, it may seem useless to precode with a N K precoder such as K > N, this may not be non-sense for other types of non-linear detectors (as it will be highlighted later in section 5) or in the case of a heavily loaded downlink CDMA scheme(in a conguration where the number of users in the system is greater than the spreading length). We also address the i.i.d. case in order to compare in the next section the performance of large isometric and i.i.d. precoders, and more specically the impact of the precoder columns code orthogonality. Theorem 2 : assume that the entries of W N,K are centered i.i.d. random variables of variance 1/N, Then, when N grows towards innity and K/N , the SINR wk at the output of a MMSE equalizer converges almost surely to a value 1 that is the unique solution of the equation in the ergodic case: assuming that p(t) is compactly supported and that for each bounded continuous function , 1 N 1 |hi |2 N+ N n=1 lim almost surely. Then, 1 is solution of

0

=E

1 |h|2

1 t

p(t) dt

t t + 2

1

+ 2

p(t) dt =

1 . 1 + 1

(4.26)

1 2 1 2

(4.27)

The proof is a direct consequence of corollary 1 in [44] and of the main result of [93]. Appendix E details the proof. Note that in the theorem is not restricted to [0, 1] but spans R + . This result deserves a few comments.

Remark: orthogonal versus i.i.d codes We now compare the two types of precoders in the ergodic case (the same came comparison can be made in the non-ergodic case). Recall the SINR for < 1 and for the case where the entries are: i.i.d:

0

t t + 2 t

1

+ 2

p(t) dt =

1 . 1 + 1 . +1

(4.28)

Orthogonal

0

t + 2 (1 ) + 2

p(t) dt =

(4.29)

0

4. MMSE

ANALYSIS OF

LP-OFDM

t t + 2 + 2

p(t) dt

t t + 2 (1 ) + 2

p(t) dt .

This implies that for a xed value of , the SINR in the i.i.d case is always smaller than in the orthogonal case. Moreover, the (asymptotic) performance of the MMSE receiver in the isometric case is all the better compared to the i.i.d. case when is close to 1. Conversely, the SINR in the i.i.d case is nearly equal to the SINR in the orthogonal case if is close to 0. In gure 4.14, we have plotted the solution in the i.i.d case and the one in the orthogonal case. The gap gain due to the choice of orthogonal codes appears clearly and is all the more important that is close to one. However, when 0, orthogonal codes and i.i.d codes perform similarly. In fact, once again, the high redundancy introduced compensates the use of completely random codes. The reader may also check that if p(t)dt is replaced by (1) (H N = IN ) the solution of (4.28) coincides with the value given in [102].

Figure 4.14: Comparison i.i.d and isometric precoding. For > 1, i.i.d:

0

t 1 p(t) dt = . t + 2 1 + 2 1 + 1

0

Orthogonal

t p(t) dt = . t + 2 +1

(4.30) (4.31)

4.3. MMSE

67

Thus, co-isometric precoding outperforms i.i.d precoding when > 1. However, as , 0 and the denominator of the right-handside of 4.30 and 4.31 become quite the same. This observation can be interpreted by the fact that the eigenvalue distribution of W N,K WH in the i.i.d and isometric case is the N,K same when Remark: comparison with the Tse-Hanly formula The so-called Tse-Hanly equation for the MMSE receiver [102, equation] gives the asymptotic SINR value for i.i.d. codes in a at fading channel, 1 which in our context amounts to hn = h for each n, or equivalently HN = hIN . As the powers allocated to the various components of vector s all coincide in our context, this equation is written 1 = . (4.32) 1 2 +1 + |h|2

1

i of s on the desired component k at the desired target SINR by K the number of users, while the coefcient

1 N

1 +1 1

+1 1 1 K N +1 1

. 1

the total amount of multiuser interference at the output of the MMSE receiver (the term

1 +1 1

1

(4.33) (4.34)

1 N = lim N N i=1

1

1 +1 + |hi |2

2

It is interesting to note that the right-handside of (4.33) coincides with an averaged version (on the amplitude of the channel coefcients) of the inverse of the sum of the multiuser effective interference 2 2 term and of the term |h|2 which represents the contribution of a thermal noise of variance |h|2 in a +1 1 at fading channel of complex gain h. This shows rst that the diversity provided by the precoder is of course due to the averaging on the values taken by |h| 2 in (4.33). More importantly, (4.33) also indicates that the important concept of multiuser effective interference introduced in [102] is still relevant if H N is not a multiple of IN . In the isometric case, we have: 1 = p(t) dt (4.35) 2 0 + t (1 )

1 +1 1+

1 N N i=1 lim

1

+1

2 |hi |2

(1 1+ )

(4.36)

For a target SINR , the performance of an isometric precoded system with noise variance 2 is equivalent to an i.i.d one with noise variance (1 ).

1+

In this section, we study in a more precise manner the inuence of coding ratio , the impact of the orthogonality of the precoder on the bit error rate (BER) and the spectral efciency for:

4. MMSE

ANALYSIS OF

LP-OFDM

The theoretical claims will then be conrmed with practical simulations using realistic matrix dimensions and convolutional coders. For the BER analysis, all the symbols will have their values in a QPSK constellation. We also assume perfect synchronization and channel knowledge at the receiver. In the ergodic case the coefcients (hk )k=1,...,N are assumed independent identically distributed. ERGODIC CHANNEL CASE The channel is assumed to be a Rayleigh fading channel, in other words p(t) = et strictly speaking. This hypothesis does not meet the technical assumption compact support required for validity. We shall nevertheless make use of formulas (4.18) and (4.26) to predict the asymptotic performances of our precoded systems. Let + eu Ei (x) = du u x be the so-called exponential integral function. For a Rayleigh fading channel, equations (4.18) and (4.26) will have the following forms: Isometric case Let r = (2 (1 ) + 2 )/. The SINR is solution of : 1 r er Ei (r) = Proof:: By writing Ei (x) = ex

+ eu 0 u+x du 0

+1

(4.37)

0 t 2 (1)+2

et dt dt

et

+1

1 2 (1 ) + 2 2 (1)+2 2 (1 ) + 2 1 e Ei ( )

i.i.d. case Denoting r = 2 (1 + 1)/. After a computation similar to that of the isometric case, the resulting SINR 1 is solution of : 1 r er Ei (r) = BER considerations 1 . 1 + 1 (4.38)

We will take the mean BER as measure of performance. It is dened by: Pe = Eh 1 K Q K k=1 wk

4.3. MMSE

69

10

Performance using QPSK constellation with independant rayleigh channel 256,256 random orthogonal codes 128,256 random orthogonal codes 64,256 random orthogonal codes theoretical performance 1/4 theoretical performance 1/2 theoretical performance 1 Gaussian channel

10

BER

10

10

10

7 8 Eb/N0 [dB]

10

11

12

13

14

15

wk Q

Figures 4.15 and 4.16 show the BER in the isometric case and in the i.i.d. case respectively for = 1, 1 , and 1 . The BER in the Gaussian channel case (i.e. H N = IN ) is also represented since, as 2 4 can easily be shown (cf.(4.18)), it represents the asymptotic performance when 0. We notice that there is an important performance gap between systems with = 1 and systems with = 1 . 2 This gap is clearly reduced when we go from = 1 to = 1 . In order to verify the practical 2 4 relevance of these evaluations, we also provide experimental results obtained through numerical simulations for N = 256. Although the chosen matrix size N is not extremely large, we observe that the theoretical curves match quite well the experimental ones for both isometric and i.i.d. cases. If we now compare gure 4.15 to gure 4.16, we notice that isometric precoding outperforms signicantly i.i.d. precoding. Moreover, the results are in accordance with our previous remarks in the sense that the performance gap between these two types of precoding is signicant for = 1, while for = 1 , it is less than 1 dB. The inuence of the number of carriers on the systems 4 performance is shown in gure 4.17. It can be noticed here that for N = 128 carriers, asymptotic analysis is fairly accurate. In g.4.18, we have plotted the performance of the Walsh Hadamard precoder for various sizes with respect to the theoretical results in the isometric case. A relative close match is achieved. One must not conclude that the theoretical results of the random Haar isometric matrices predict the performance of Walsh Hadamard codes (it is certainly not the case except for = 1!). Our claim is to show that isometric Haar precoders can do as well as standard Walsh Hadamard codes in terms of BER and can therefore be used as interesting precoders for their prediction/performance/optimization possibilities. In order to compare more accurately isometric to i.i.d. precoding, we also plot in gure 4.19 the SINRiid SINR loss ( SINRortho ) of i.i.d. precoders with respect to isometric ones for various values of the coding ratio . Notice that the gap is again important for = 1

70

4. MMSE

ANALYSIS OF

LP-OFDM

10

Performance using QPSK constellation with independant rayleigh channel 256,256 i.i.d codes 64,256 i.i.d codes 128,256 i.i.d codes theoretical performance 1/4 theoretical performance 1/2 theoretical performance 1

10

BER

10

10

10

9 10 11 Eb/N0 [dB]

12

13

14

15

16

17

18

19

20

10

Performance using QPSK constellation with independant rayleigh channel 256,256 random orthogonal codes 128,128 random orthogonal codes 64,64 random orthogonal codes theoretical performance 1 256,128 random orthogonal codes 128,64 random orthogonal codes 64,32 random orthogonal codes theoretical performance 1/2

10

BER

10

10

10

8 9 10 Eb/N0 [dB]

11

12

13

14

15

16

17

18

4.3. MMSE

71

10

Performance using QPSK constellation with independant rayleigh channel 256,256 walsh codes 64,256 walsh codes 128,256 walsh codes theoretical performance 1/4 theoretical performance 1/2 theoretical performance 1 Gaussian channel

10

BER

10

10

10

7 8 Eb/N0 [dB]

10

11

12

13

14

15

0.9

0.8

SINRiid/SINRortho

0.7

0.6

0.3

0.2 1

9 10 11 Eb/No [dB]

12

13

14

15

16

17

18

19

20

72

4. MMSE

ANALYSIS OF

LP-OFDM

Choice of the ratio : In this section, we try to address the problem of nding the value of the ratio that maximizes the systems throughput. As said above, the purpose of this analysis is to determine the amount of redundancy that should be spent on the linear precoder. The need for an optimum can be justied intuitively. On the one hand, at xed bandwidth, choosing the ratio close to 1 allows to use small rate error correcting codes but the effect of the "multi-user interference" signicantly decreases the SINR at the output of the MMSE detector. On the other hand, a small value of the ratio increases the SINR at the output of the MMSE detector, but higher rate error correcting codes are then needed. In the context of this report, the throughput also called spectral efciency is the total number of bit/s/Hz that can be reliably transmitted with our precoded system equipped with a MMSE receiver. Biglieri et al [11] have already used the spectral efciency as a mean for designing and answering the coding versus spreading issue in uplink CDMA with random i.i.d codes. They have shown in particular that a non-negligible amount of spreading should be performed when using the MMSE receiver. We look at an extension of such results with random Haar isometric matrices. Our throughput analysis will then be conrmed by simulations. The throughput , 2 is dened by , 2 = C , 2 where C , 2 is solution of the equation C , 2 = log2 1 + SINR(, 2 ) It is understood that Eb /N0 = (C2 )1 , see [81] for more details. In gure 4.20, Eb /N0 is xed to 10 dB and we plot the spectral efciency versus the ratio in the isometric (or co-isometric if > 1) and the i.i.d. cases. We observe that isometric precoding outperforms i.i.d precoding, especially for 0.30 < < 3. When , isometric and i.i.d precoders have similar performance and converge to a limit 2 . Figure 4.21 shows the behavior of the optimum value of the ratio (i.e. for which the throughput is maximum) with respect to Eb /N0 for both isometric/co-isometric and i.i.d. cases. In the isometric/co-isometric case, gure 4.21 shows rst that the optimum value of is always 1 and that nearly no redundancy should be spent on the precoder to maximize the throughput. Whereas, Eb in the i.i.d case, a signicant amount of redundancy is needed when N0 > 4dB. For instance, at Eb N0 = 10dB, the optimum value of the ratio is 0.63. Surprisingly, for low values of E b /N0 , a lack of redundancy should be introduced in the i.i.d case. Indeed, for E b /N0 . This result is conrmed in gure 4.22. The maximum throughput in the i.i.d case is achieved for while in the isometric case = 1. We next compare the throughput of the current systems to the maximum of the throughput in the Gaussian channel case (the case where H N = IN ), which is of course reached for = 1. This upper bound will be called the Gaussian channel bound in the following. It obviously coincides with the capacity of the standard Gaussian channel. In gure 4.23, we represent the optimum (w.r.t. the E ratio ) throughput versus Eb /N0 in the isometric and i.i.d case. At Nb = 15dB, 5.3bit/s/Hz can be 0

2 The

spectral efciency of i.i.d and co-isometric precoders converge as to the solution of the following equation = log2 (e). This result can be obtained by noting that

1 2

t 1 t+ Eb

Eb No .

term of the Taylor series of 4.3.3 together with equation 4.38 yields the result

No

4.3. MMSE

73

MMSE Haarprecoded MMSE i.i.dprecoded Gaussian channel single user bound limit MMSE i.i.d

4 alpha

10 9 8 7 6 alpha 5

optimal alpha (in the throughput sense) for isometric and orthognal matrix isometric case i.i.d case

10 11 12 Eb/No [dB]

13

14

15

16

17

18

19

20

74

4. MMSE

ANALYSIS OF

LP-OFDM

1 0.9 0.8 0.7 spectral efficiency bit/s/Hz 0.6 0.5 0.4 0.3 0.2 0.1 0

MMSE Haarprecoded MMSE i.i.dprecoded Gaussian channel single user bound limit MMSE i.i.d

4 alpha

10 9 8 7 6 capacity 5 4 3 2 1 0 1

optimal capacity isometric matrix i.i.d matrix single user (gaussian channel)

9 10 11 Eb/No [dB]

12

13

14

15

16

17

18

19

20

4.3. MMSE

75

optimal capacity alpha=2/3 alpha=3/4 alpha=8/10 alpha=1 single user (gaussian channel)

9 10 11 Eb/No [dB]

12

13

14

15

16

17

18

19

20

Eb N0

reliably transmitted using a random isometric spreading matrix ( = 0.95) while only 3.2bit/s/Hz can be transmitted using a random i.i.d spreading matrix ( = 0.68). Notice that the Gaussian channel bound is equal to 8 bit/s/Hz.

8 We also represent in gure 4.24 the throughput for various values of the ratio (1, 10 , 3 , 2 ) in the 4 3 3 isometric case. The Gaussian channel bound is also represented. One can notice that for 4 < < 1, 3 little can be gained by optimizing the ratio (which is not the case for < 4 ). As a conclusion, for the isometric precoder, one can choose = 1 without losing much of the spectral efciency

Performance of practical coding schemes: The performance of a system where Linear Precoding of rate is combined with classical Convolutional Coding of rate R is studied in order to conrm the spectral efciency analysis. In the context of coded LP-OFDM, one Viterbi decoder is applied on each subband k processing the real and imaginary parts of the signal output of the Wiener lter. In order to evaluate upper bounds on the BER, let rst evaluate the probability Pcoded LP-OFDM (d) of deciding P1 instead of P0 where P0 is the path of the Viterbi algorithm trellis associated to the transmitted sequence and P1 is a path which differs by d bits from P0 . Following [81, 40] and the previous section, the probability of that event, is given by Pcoded LP-OFDM (d) = E Q The overall error probability is thus bounded by: Pcoded LP-OFDM d Pcoded LP-OFDM (d) d=dmin M d()

(4.39)

76

QPSK constellation coding rate 1/2

4. MMSE

ANALYSIS OF

LP-OFDM

10

orthogonal codes CC 1/2 spreading 1 orthogonal codes CC 10/16 spreading 8/10 orthogonal codes CC 2/3 spreading 3/4 orthogonal codes CC 3/4 spreading 2/3

10

BER

10

10

10

5 Eb/N0 [dB]

Figure 4.25: Various convolutional coding rates for isometric precoding matrix. Here, dmin is the minimal distance of the code, M is the number of input bits in the encoder and d is the number of incorrectly decoded information bits, for each possible incorrect path differing from the correct one by d bits. The goal is to determine the optimum balance between the ratio and R, assuming a constant overall spectral efciency of 1 bit (in the case of QPSK constellations, we have 2 R = 1), The role of the ratio and R is revealed by eq (4.39) and shows that the performance is strongly related to the transfer function of the code. Indeed, when the ratio increases, () decreases but R also decreases yielding a greater minimum distance d min through the transfer function. Thus, the tradeoff depends on the transfer function of the code (through d and dmin ). The analytical expression of the transfer function can only be drawn for codes with a limited number of states. Since theoretical study of general codes is extremely difcult, we will conduct performance simulations of various convolutional codes. Note that convolutionaly coded systems with i.i.d random spreading has already been studied in [89] in the case of CDMA systems. In our simulations, different values 8 of the ratio (1, 10 , 3 , 2 ) are considered and each of them is associated to a convolutional code of 4 3 rate R(). In order to compare the corresponding systems at xed spectral efciency, we assume 1 that R() = 2 for each . Figure 4.25 illustrates the performance of the various coded schemes in the isometric case. Our throughput analysis is in accordance with gure 4.21 in the sense that the best results are obtained for 1. Figure 4.26 compares the performance of the i.i.d case for the same values of the ratio . The value 2 provides signicantly better performance in accordance 3 with the throughput analysis of gures 4.20 and 4.21 in the i.i.d. case. NON-ERGODIC CHANNELS: BER Considerations: This paragraph aims at showing that our analysis ts with usual precoding matrices for non-ergodic channels. Figure 4.27 and 4.28 plot the theoretical and simulated 1 (with Walsh Hadamard matrices) BER respectively for = 2 and = 1 considering different

4.3. MMSE

77

10

10

10 BER 10

10

random codes CC 1/2 spreading 1 random codes CC 10/16 spreading 8/10 random codes CC 2/3 spreading 3/4 random codes CC 3/4 spreading 2/3 10

5

7 8 Eb/N0 [dB]

10

11

12

13

14

15

Figure 4.26: Various convolutional coding rates for i.i.d precoding matrix. channel orders (L = 1, L = 2, L = 8). One can observe that the simulated and theoretical curves closely match for a realistic number of carriers (N=128). As the channel order increases, the BER improves considerably. As expected when the ratio decreases, the performance also increases for a given channel order. However, a loss in terms of spectral efciency occurs.

Spectral efciency Considerations: The purpose of this paragraph is to design the LP matrix with respect to the ratio K/N = in order to maximize the spectral efciency for non-ergodic channels. The spectral efciency with MMSE equalization is dened as (cf. [11]: (K, N) = 1 K N i=1 log2 (1 + wi ) log2 (1 + ())). In this case, since the SINR is a random variable depending only on the channel realization, the spectral efciency is also random. In our simulations, only the mean spectral efciency is considered. Figure 4.29 plots the optimum of the spectral efciency for different channel orders as well as the model A (cf. chapter 2). Nearly no redundancy ( > 0.92) should be spent on Linear Precoding whatever the diversity order L is. Note also that the gain in spectral efciency by optiE mizing the ratio increases with Nb as shown in g.4.30 but little can be gained when 0.7 < < 1 o E and Nb < 6dB o

Convolutive Coding schemes: g.4.31 plots the performance of various Convolutional Coding 8 3 rate schemes (, R) { (1, 1 ); ( 10 , 10 ); ( 3 , 2 ); ( 2 , 4 ) with diversity order 8. In that purpose, the 2 16 4 3 3 convolutional code specied in the HIPERLAN/2 standard is used. The mother code has a constraint length of 7 (133o, 171o) and rate 1 . The other rates are achieved by puncturing the coded 2 symbols. These gures suggest that the optimum is 1 and are in accordance with the spectral efciency analysis of g 4.29

78

4. MMSE

ANALYSIS OF

LP-OFDM

10

10

BER 10

3

10

10

simul 1 tap theory 1 tap simul 2 taps theory 2 taps simul 8 taps theory 8 taps simul Rayleigh theory Rayleigh 0 1 2 3 4 5 6 7 8 Eb/N0 [dB] 9 10 11 12 13 14 15 16

10

10

BER 10

3

10

10

simul 1 tap theory 1 tap simul 2 taps theory 2 taps simul 8 taps theory 8 taps simul Rayleigh theory Rayleigh 0 2 4 6 8 Eb/N0 [dB] 10 12 14 16

4.3. MMSE

79

1.02

0.96

0.94

0.92

0.9

N=128, 3 channels from 4 independent taps alpha=1/10 alpha=2/10 alpha=3/10 alpha=4/10 alpha=5/10 alpha=6/10 alpha=7/10 alpha=8/10 alpha=9/10 alpha=10/10

10

12

14

16

Figure 4.30: Spectral efciency for different and channels with 4 independent taps

80

Coding rate 1/2 with diversity order 8

4. MMSE

ANALYSIS OF

LP-OFDM

10

10

10

10 BER 10

10

orthogonal codes CC 2/3 LP 3/4 orthogonal codes CC 10/16 LP 8/10 orthogonal codes CC 1/2 LP 1 orthogonal codes CC 3/4 LP 2/3

10

10

6 Eb/N0 [dB]

10

12

Figure 4.31: Spectral efciency for different and channels with 8 independent taps

4.3.4 Summary

We have observed in this section that : in a system designed with random isometric transforms, the optimum trade-off between the redundancy of the linear precoder and the rate of the convolutional encoder favors values of close to 1 for ergodic and non-ergodic channels this is the converse with the i.i.d. case in which non negligible spreading redundancy is needed.

for the optimum values of the ratio , random isometric precoders systems outperform i.i.d precoders.

4.4 Conclusion

In this chapter, we have studied thoroughly the asymptotical performance of the MMSE receiver in the uncoded and coded case. When considering unitary matrices such as Walsh Hadamard, FFT and random unitary, LP-OFDM outperforms OFDM due to the spreading effect of the information in the uncoded case. When coding is applied, COFDM can outperform coded LP-OFDM, especially as the constellation size increases. We have also studied the optimal choice of the ratio K for random Haar N isometric matrices and shown that the optimal choice is near 1 for ergodic and non-ergodic channels. The use of orthogonality pays off with respect to completely i.i.d. codes. Finally, the choice of the precoders and our analysis has been validated by simulations as similar performance is achieved by standard Walsh Hadamard codes with realistic size matrices.

81

Chapter 5

5.1 Introduction

The performance of the MMSE receiver has been studied in the previous chapter. For certain type of precoding matrices with MMSE equalization, the amount of redundancy has been optimized with respect to the spectral efciency. The MMSE receiver was shown to provide important improvements but the results are still far from what the channel capacity is able to provide. The suboptimality of the MMSE receiver has raised considerable interest [99, 41] for non-linear multiuser detectors revealing important gains over the performance of the MMSE detector. The price for those improvements is of course an increased receiver complexity. Basically, all these non-linear methods borrow equalization schemes from the CDMA context. Indeed, the single user LP-OFDM context is suited for Multi-user detection techniques if one considers each encoded data substream at the transmitter as a separate user: the different symbols play here the role of users. In Multi-user detection, it is now well acknowledged that a great performance improvement can be achieved if the detector takes into account the structure (due to the codes and constellations used) of the Multi-user interference generated after MMSE equalization. This strategy increases the SINR by reducing the multi-user interference. In particular, the Gaussian noise behavior argument of the Multi-user interference is not employed anymore. Moreover, the complexity of the detector is considerably reduced in comparison with the exponentially behavior (in the number of users) of the Maximum likelihood detection. These multiuser detection techniques have known tremendous success lately in accordance with the numerous publications on this subject since 1990, especially by the pioneering book of Sergio Verdu [112]. This section aims at analyzing and deriving the theoretical asymptotic performance of two non-linear detection schemes for Linear Precoded OFDM, namely : the MMSE Successive Interference Cancellation (SIC) proposed in the multi-antenna context (known as BLAST) in [52, 121] and in the CDMA context [74, 75, 60] the iterated-decision multiuser detector introduced by Chan and Wornell in the uplink CDMA

82

5. N ON -L INEAR D ETECTION context [28], which uses Parallel Interference Cancellation (PIC)

SCHEMES FOR

LP-OFDM

More precisely, the asymptotic results aim at adjusting the different parameters involved in the two decoding schemes. Recent contributions in the CDMA context ([75, 28, 74]) have already studied the asymptotic performance of these detectors using random matrices W N,K when HN is reduced to IN and N , K and K . In this context, it is quite natural to modelize matrix W N,K as a random N matrix with i.i.d. entries. However, as pointed out in the previous section, in the LP-OFDM context, this matrix should be assumed isometric (the columns of the precoder are orthogonal) in order to achieve higher gains (and since synchronization is ensured at the transmitter). The isometric assumption as well as the specic structure of our model (diagonal matrix by a N K precoder) provides a challenging but at the same time complicated framework. Using, as before, tools from the free probability theory, the Signal to Interference plus Noise Ratio at each iteration of the non-linear detection process will be derived when N + and K/N 1 assuming both ergodic and non-ergodic fading channels. Note nally that all the analysis carried out in this section can be directly applied to the downlink MCCDMA context.

5.2.1 BLAST MMSE algorithm

The scheme was proposed by Foschini et al. in [52, 121] for MIMO multi-antenna transmission schemes. The algorithm relies on a sequential detection of the received block [121]. Recall that y = H N WN,K s + n At the rst step of the method, a MMSE equalization of matrix M N,K = HN WN,K is performed by a multiplication of y by matrix G1 = MH (MN,K MH + 2 I)1 . As pointed out in the previous section, N,K N,K at this step, all the symbols enjoy asymptotically the same 1 SINR ( K ) which only depends on the limit2 N ratio = K . Since there is no optimal choice for detecting the rst symbol, suppose that the algorithm N starts by decoding symbol sK . The estimated symbol can go through a decoder chain (Viterbi decoder) in order to improve the reliability of the detection process. Assuming a perfect decision (this is possible if the information sK has been encoded at a rate log 2 (1 + ( K )), the resulting estimated symbol s K is N subtracted from the vector of received samples in the following manner: r 2 = r1 sK mK (mi represents the ith column of MN,K and vector r1 = y). This introduces one degree of freedom for the next canceling vector choice which enables to reduce the noise plus interference inuence and yields an increase in the decision process reliability. The second step can be virtually represented by a completely new system of K 1 symbols (s 1 , ..., sK1 ) transmitted by an N (K 1) isometric precoding matrix W N,K1 on the same at frequency fading H 2 1 channel. Equalizing with matrix G2 = MH N,K1 (MN,K1 MN,K1 + I) , all the estimated symbols enjoy asymptotically an SINR of ( K1 ). Assuming that the information in s K1 has been encoded at N a rate log2 (1 + ( K1 )), one can reiterate the same process describe at the beginning. The advantage N K of such a scheme is that ( N ) ( K1 ): one is able therefore to convey much more information on N the second symbol (since the SINR increases) than with MMSE equalization. If one does not want to

the case of nite systems, the SINR of each symbol is different. Therefore, only the symbol with the best SINR is detected rst. There is an optimum ordering in the detection process (depending on the channel attenuations) which makes the analysis extremely difcult for nite dimensions. 2 In order to present the SIC MMSE algorithm, we denote the SINR by ( K ), although the notation does not make any N sense.

1 In

83

increase the rate, the performance in terms of BER increases. The complete detection algorithm can thus be summarized as follows:

initialization: i r1 G1 = = 1 y

1 MH (MN,K MH + 2 I) N,K N,K

(i)

i+1

where WN,Ki denotes the matrix obtained by suppressing columns (K, , K i + 1) of W N,K and (i) MN,Ki = HN WN,Ki . gi is the ith row of Gi (the MMSE ltering matrix) at step i. Once the symbol is detected (equation (5.4)), a decision is made modeled by operator D (equation (5.5)). In our case, D is the decision operator made after equalization. If no coding is applied, D will simply represent the hard decision operator (in the uncoded case, we will not study the soft-decision operator as it did not provide interesting enhancements)

In this section, we will derive the asymptotic SINR (x) for each symbol k = [xN + 1] ([] is the integer part operator) during the detection process. These results will then be used to study the performance of the SIC algorithm from a BER and spectral efciency point of view. We assume hereafter that the detected symbols at each iteration are perfectly retrieved. Before going further, recall the expression of the SINR at the output of the MMSE detector in the case of the transmission of K l linearly precoded symbols, when N , K l and Kl x 1. N

Kl Kl For each symbol s j , 1 j K l, the SINR w j was easily shown to be w j = 1 1Kl wj Kl wj H 2 wH HH HN WN,Kl WH j N N,Kl HN + IN

Kl where w j =

C C1 qC C1 qC C

84

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

Kl chapter 4 showed that when N and K l in such a way that Kl tends towards a constant, w j N converge to a certain limit, for all matrices W N,Kl haar distributed. More precisely, the following result holds:

Proposition 1 Let k be the number of symbols to be detected and N the numbers of carriers. When N and k/N x 1, the SINR k j (1 j k) at the output of the Blast MMSE equalizer with w random isometric Haar distributed matrices converges almost surely to a value (x) that is the unique solution of the equation: in the ergodic case:

0

t xt + 2 (1 x)(x) + 2

p(t) dt =

(x) . (x) + 1

(5.9)

1 2 1 2

(5.10)

One can notice that when x = , the performance of the MMSE receiver in section 4.3 is achieved in the ergodic as well as in the non-ergodic case. Since (x) is a decreasing function of x (shown in the previous section with x = ), perfect retrieval of the symbols at each step of the algorithm reduces the interference and yields an increase in terms of performance. This consist in the asymptotic justication of the BLAST MMSE algorithm. Note also that at the nal step, the single user bound is achieved. Indeed, when x 0, eq.(5.9) and eq.(5.10) yield: in the ergodic case:

0

(5.11)

1 2

1 2

(5.12)

We give also for comparison purpose the performance in the i.i.d case: Proposition 2 Let k be the number of symbols to be detected and N the numbers of carriers. When N and k/N x, the SINR k j (1 j k) of the at the output of the Blast MMSE equalizer with w random i.i.d matrices converges almost surely to a value (x) that is the unique solution of the equation in the ergodic case:

0

t xt + 2 (x) + 2

p(t) dt =

(x) . (x) + 1

(5.13)

85

1 2

1 2

(5.14)

Not also in this case that at the nal step (x=0), the single user bound is also achieved. In our case, the performance criteria considered is not the SINR of each symbol but the average SINR or more precisely the average BER and spectral efciency taken across all the symbols. The following theorem will be used in order to derive the asymptotic performance of the SIC or BLAST MMSE algorithm in terms of spectral efciency and bit error probability : Theorem 3 let f : R R be a continuous function. If K , N and K/N > 0, then

w 1 K f (iwi ) x=0 f ((x))dx where iwi = 1iiw with iwi = wH HH HN WN,i WH HH + 2 IN i=1 i N N,i N i and (x) solution of equation (5.9) or (5.10).

1 K

a.s

HN wi

The proof can be found in appendix F Spectral efciency Considerations: In our case, the spectral efciency is dened as the highest rate at which information can be sent with arbitrary low probability error with MMSE SIC detection. Our analysis (in which we consider perfect retrieval of the detected symbols) is justied if coding on each carrier k (k = [xN + 1]) of gure 5.2 is carried at exactly the spectral efciency rate (so that no errors occur with perfect decoding at the receiver) log 2 (1 + (x)). For xed K and N, the mean spectral 1 efciency of re-encoded successive cancellation is: (K, N) = N K log2 (1 + iwi ). Theorem 3 can be i=1 therefore applied to the continuous function f : y log 2 (1 + y) and yields the two following results: Spectral efciency for ergodic channels: the asymptotic values of the spectral efciency of reencoded successive cancellation is deterministic and equal to: = 0 log2 (1 + (x)) dx . Spectral efciency for non-ergodic channels: in this case, since the SINR is a random variable depending on the channel realization, the spectral efciency is random as well. An outage or mean spectral efciency of the random variable can be derived. The mean spectral efciency (the mean is taken with respect to the channel distribution) is dened by: mean = E 0 log2 (1 + (x)) dx An interesting issue is the gap gain of SIC schemes with respect to Maximum likelihood detection. In the CDMA at fading context, it is shown in [110, 90] that the spectral efciency of the SIC scheme is equal to the spectral efciency of the optimum receiver. This stems from the optimality of a successive cancellation non-linear receiver where each stage consists in an MMSE transformation which takes into account only those users not yet demodulated. The previous results can be extended in the case of LP-OFDM and shows the optimality of Successive MMSE Interference Cancellation. Theorem 4 : let SIC denote the spectral efciency of the SIC MMSE algorithm and opt the spectral efciency when optimal detection is performed. SIC (2 ) =

0

86

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

and opt (2 ) = HN WN,K WH HH 1 N,K N log2 det IN + N N 2 t = log2 (1 + 2 )d( )(t) lim

is the limiting eigenvalue distribution of W N,K WH and is the limiting eigenvalue distribution of N,K HN HH .The following equality holds: N opt (2 ) = SIC (2 ) The proof is based on the fact that: HN WN,K WH HH 1 N,K N log2 det IN + 2 N = 1 K log (1 + iwi ). N i=1 2 (5.15)

H H H Indeed, writing HN WN,K WH HH = HN WN,K1 WH N,K N N,K1 HN + HN wK wK HN and using a triangular decomposition:

Theorem 3 can be therefore applied to the continuous function f : y log 2 (1 + y) and yields theorem4. We also provide in Appendix G an interesting proof, specic to our case, based on free probability results and without the need of theorem 3. In gure 5.3, we have plotted the spectral efciency in the case of = 1 for the following various detectors, namely the MMSE detector, the SIC MMSE also known as blast algorithm. For = 1, notice 2 that opt (2 ) = E log2 (1 + |h|2 ) which will be referred as the single user bound. Simulations were performed assuming Rayleigh fading channel and random Haar isometric distributed matrices. SIC

R=log2 1

TX

Channel

RX

= (1 + wK )det IN +

1

H HN WN,K1 WH N,K1 HN

K N

sK k

yk

detector

sK k

S/P

R=log2 1

sk

WN

hN n N k

SIC

sk

1 N

s1 k

h1

n1 k

yk

s1 k

Viterbi Decoder

P/S

Viterbi Decoder

87

MMSE achieves the single user bound in a Rayleigh channel. The gap gain with respect to MMSE is also important as the energy per bit grows. This gap gain is also called the non-linearity gain. It gives us some basic knowledge on whether to use or not a non-linear detector for increasing the spectral efciency. We have also plotted the non-faded case as a comparison basis (called here the Gaussian case). The spectral 1 efciency of the Gaussian channel is dened as: ( 2 ) = E log2 (1 + 2 ) . We should insist on the fact that the coding rate must be adjusted on each substream k = [xN + 1] accordingly to the expected limiting SINR (x). A remarkable result of theorem 4 in the ergodic case is that the SIC MMSE is able to achieve the single user bound without the knowledge of the channel realization at the transmitter (for coding on each substream) but only the channel statistics. We believe that this is truly an important result as it does not require a channel feedback mechanism but only knowledge of the channel statistics. On the opposite side, in the non-ergodic case, the layers rates do depend on the fading realization. We should then pose the problem in terms of outage probability. In these terms, stripping is very suboptimal, in the sense that it does not achieve optimal outage probability, while full joint decoding without layering does achieve outage probability. The reason for this fact is simple: with layering, the outage is dominated by the rst layer, which is also the one enjoying the smallest diversity order as it is decoded based on the output of a linear MMSE lter (no decision feedback, since it is the rst layer). The suboptimality of layering in the non-ergodic case is very well known in the multiple antenna community.

9 Theoretical spectral efficiencies, alpha=1, Haardistributed precoding, Rayleigh channel gaussian channel SE Blast SE = faded single user SE MMSE SE

6 Spectral Efficiency

8 E /N [dB]

b 0

10

12

14

16

Figure 5.3: Spectral efciency (SE) of BLAST algorithm, Rayleigh channel (ergodic channel). In gure 5.4, we study the impact of the orthogonality of the precoders columns with respect to the i.i.d case at 10 dB in the ergodic case with i.i.d Rayleigh fading. Important conclusions can also be drawn: with orthogonal codes, a great improvement with respect to the MMSE receiver is achieved when is close to 1 (nearly 1.5 bit/s/Hz for = 1). When is low, there is no need into using such a detector. In this case, the high redundancy of the precoder is able to compensate the suboptimality of the MMSE detector. with i.i.d codes, a surprising result shows that there is no loss in terms of spectral efciency as long as ! Therefore, one can use a matrix with i.i.d entries and still achieve the best performance

88

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

possible. However, since the complexity of the SIC detector is related the number of inversions K, using i.i.d codes can be prohibitively to complex to use in this case. For 0 1, the SIC detector with i.i.d codes performs similarly as the MMSE detector with orthogonal codes. In this region, orthogonality pays off.

Unlike the MMSE case, the use of orthogonal precoders with SIC equalizers is not dictated by a performance criteria but rather by a complexity criteria. The complexity of the receiver in the i.i.d case is much larger than the orthogonal case for the same target of spectral efciency since one has to choose in the i.i.d case. For these two criteria to be fullled (complexity and performance) the best choice is thus an orthogonal precoder with = 1. However, there is no suboptimality in using i.i.d codes when a appropriate SIC detector is used.

Spectral efficiency of different systems with Eb/N0=10dB

MMSE Haarprecoded SIC Haarprecoded MMSE i.i.dprecoded SIC i.i.dprecoded Gaussian channel single user bound limit MMSE i.i.d

4 alpha

Figure 5.4: Spectral efciency versus for various detectors at 10dB, Rayleigh channel (ergodic channel).

BER Considerations: In this section, we analyze the BER. The constellation used is QPSK. Moreover, at each step of the SIC algorithm, perfect detection of the symbols is assumed. Even though such an assumption is not justied3 , the performance evaluation gives us a lower bound on the achievable performance. It is possible to take into account, under certain assumptions, the propagation of errors but this requires more involved materials already used in [30] and goes beyond the scope of this document. As the asymptotic contribution of the noise and of the multi-user interference at the output of the detector can be considered Gaussian [80], it is standard to associate to each asymptotic SINR the asymptotic as BER given by Q SINR (we consider here a QPSK constellation). For xed K and N, the mean BER 1 is: N K Q wi i . Theorem 3 can be therefore applied to the continuous function f : y Q( y) and i=1 yields the two following results:

this assumption, the nal step of the detection process achieves the performance of the single user bound while it is well established that due to the propagation of errors, such a result is not reached

3 With

89

BER for ergodic channels: the BER (mean of the bit error rates over all the carriers) is dened 1 as: BER = 0 Q (x) dx BER for non-ergodic channels: in this case, the SINR is a random variable depending on the channel realization. We can therefore dene a mean BER (with respect to the channel distribution) 1 (x) dx as: BER = E 0 Q Simulations have been conducted using a QPSK constellation (according to the theoretical results obtained) and assuming perfect synchronization and channel knowledge at the receiver.

10

0

10

theory MMSE equalizer simul MMSE equalizer perfect SIC performance simul SIC single user OFDM

10 BER 10

10

10

8 E /N (dB)

b 0

10

12

14

16

Figure (5.5) and (5.6) show the BER in the uncoded case respectively for independent Rayleigh channel attenuations and a realistic channel with 8 taps. Notice that only the random isometric Haar distributed case is illustrated. The theoretical curves closely match to the simulation results at high SNR, using a realistic number of subchannels (N=256). At low SNR, a little gap appears due to the fact that uncoded transmissions have been simulated. In this case, the symbols are not retrieved perfectly and yield propagation errors. Indeed, for performing a good interference cancellation, due to the underlying feedback mechanism involved in the successive detection, the MMSE BLAST algorithm should decode rst the reliable carriers enjoying a greater SINR and then the most corrupted ones. Unfortunately, all the carriers share the same SINR resulting in practice to marginal performance gain when applying successive interference cancellation approaches. The theoretical curve in gure 5.5 and 5.6 show also that even when ideal interference cancellation is performed (the best achievable BER), the performance gain in the uncoded case is poor in comparison to the great complexity of the algorithm. Indeed, the performance is mostly affected by the errors incurred at rst stage of the algorithm. If one does not change the constellation of the other symbols to be detected (in order to increase the rate), there is no use into having a greater SINR at the nal stages. As a conclusion, the SIC scheme reaches the expected performance only when perfect coding and decoding structures are applied. We also observe that the BLAST MMSE algorithm incurs no loss of

90

0

5. N ON -L INEAR D ETECTION

alpha=1, N=256, 8 taps

SCHEMES FOR

LP-OFDM

10

10

theory MMSE equalizer simul MMSE equalizer perfect SIC performance simul SIC single user OFDM

10 BER 10

10

10

8 E /N (dB)

b 0

10

12

14

16

spectral efciency and using orthogonal or i.i.d codes is more a question of equalization complexity than performance when coding is applied. The optimal complexity/performance solution is orthogonal precoding with = 1. However, even in this case, the complexity in terms of implementation is not negligible (N 1 matrix inversions). Moreover, the transmitter must encode each substream at a given rate to take full benet of the system, otherwise little gain is achieved as in the case of uncoded LPOFDM. For non-ergodic channels and coded schemes, the coding rate must be adjusted according to the channel realization, which is impossible if no channel knowledge is available at the transmitter. Power and bit loading strategies are far more interesting in this case. From a practical point of view, all these observations make the SIC algorithm unadapted to LP-OFDM. For ergodic channels, however, the SIC scheme and its analysis is interesting and may have a direct application to the multi-user scenario, namely MC-CDMA. From a theoretical point of view, the results are extremely important as they give answers on the impact of the choice of the precoder (orthogonality, matrix size,. . . ).

5.3.1 Iterated-Decision Multiuser Algorithm

This section provides a performance analysis of the iterated-decision multiuser detector proposed by Wornell et al. [28] in the uplink CDMA and depicted in gure 5.7. Note that [99] already proposed such a scheme in the LP-OFDM context but was unable to provide theoretical justications for the optimization of the different parameters involved in the detection process. The basic idea underlying this detection process is to retrieve simultaneously after equalization all the interfering symbols based on previous estimations. This reduces the multi-user interference for all the symbols and increases the SINR. This Parallel Interference Cancellation (PIC) scheme is in fact suited for LP-OFDM since at any step, there is no symbol enjoying a higher SINR than the other.

91

At each iteration, the algorithm is based on the use of two matrices: a forward matrix C l for equalization of the received block symbol and a backward matrix D l for retrieving the interfering symbols. The aim of this section is to carry an asymptotic analysis of this scheme in order to derive the optimum matrices Cl and Dl maximizing the SINR. These results will then be used to get insights on the behavior of the algorithm. Similarly to block DFE schemes [109] , at iteration l of the equalizer (l = 1, 2, 3, . . .), the received vector H H y is premultiplied by a K N matrix Cl = [cl1 | . . . |clK ]H , producing the K 1 vector yl = Cl y. Next, an estimate of the multi-user interference (based on previous estimates of the symbols) is subtracted H H from yl to produce sl , i.e. sl = yl Dl sl with Dl = [dl1 | . . . |dlK ]H a K K matrix. In order to only deal with interference, the diagonal elements of D l are forced to zero. If the estimates were accurate, the performance of a single symbol transmission (by a N 1 precoder) would be obtained. At the nal step, a bank of slicers applied to sl generates the K 1 vector of hard decisions sl . In the coded case, the symbols go through a decoding chain to increase the reliability of the process. However, the decoding chain may incur a delay not tolerable for real time applications. A great merit of this scheme is that the iterations (and therefore the complexity of the algorithm) can be stopped when the desired bit error rate is reached. The asymptotic study aims also at deriving the number of iterations required for a target BER.

In order to derive the calculus, some notations and assumptions need to be introduced. Simulations have shown [28] that s and sl1 can be considered as vectors of zero-mean uncorrelated symbols, each with energy 1, with normalized correlation matrix of the form E[sl1 ] s

H

l1 l1 = diag[l1 , l1 , . . . , K ] 1 2

Moreover, we assume that sl1 and n are uncorrelated. i i i l1 is a condence coefcient that is attributed to the symbol sl1 at iteration l 1. If sl1 has l1H l1 1. In the other case, l1 0. It been detected correctly and is close to s i , than E[si .si ] = i i i measures the closeness of sl1 to si .

92

5. N ON -L INEAR D ETECTION The slicer input for the ith symbol at the l th iteration can be expressed as si l = cli HN wi si + vli

H

SCHEMES FOR

LP-OFDM

H H H (i) where vli = cli HN WK s dli sl1 + cli n is assumed to be a complex-valued, asymptotically Gaussian (as N ), zero-mean, and uncorrelated with s i , whose variance is a function of cli and dli . Matrices Dl and Cl are derived in order to maximize the SINR during the l th iteration. The following result holds:

Proposition 3 The output SINR is maximized for the iterated-Decision algorithm with the following choice of matrices: Dl = l1 and Cl [2 I + HN WK (I l1 l1 )WH HN H ]1 HN WK K The proof can be found in the Appendix. I. At each iteration, D l expresses the fact that only nondiagonal elements (which correspond to interference) are retrieved. The interfering signals to be detected are not alterered. The special case l = 1 deserves some comment. During the rst pass, MUI subtraction is not performed because 0 = 0. Therefore, the matrix C1 reduces to the linear MMSE multiuser detector. Thus, the performance of the iterated-decision multiuser detector, after just one iteration, is identical to that of the linear MMSE multiuser detector. With the previous expression of matrices C l and Dl , let derive the SINR li dened as: li |cl HN wi |2 = i var(vli )

H H H

var(vli ) = wH HN H [2 I + (HN WN,K )(I l1 l1 )(HN WN,K )H ]1 HN wi i (1 |il1 |2 )(wH HN H [2 I + (HN WN,K )(I l1 l1 )(HN WN,K )H ]1 HN wi ) i As li = wH HN H [2 I+(HN WK )(Il1 l1 )(HN WK )H ]1 HN wi , then the expression of the SINR i for the ith symbol becomes li 1 (1 |l1 |2 )li i

H H

li =

(5.16)

5.3. PARALLEL I NTERFERENCE C ANCELLATION D ETECTOR The condence coefcient l1 is evaluated in a recursive manner. Indeed, let p i = Q i

1 part are independent, E(si si ) = 2E(sr i sr i ). Since sr i sr i = 1 2

93 i l1 be

1 1 the bit error probability and si = sr i + jsi i (sr i = 2 , si i = 2 ). Assuming that the real and imaginary

with probability p, E(sr i sr ) = 1 (1 pi ) + 1 pi . Therefore, E(si si ) = il1 = 1 2p = 1 2Q i 2 2 FFT, Walsh Hadamard and random unitary matrices (K = N, = 1)

i l1

For the case K = N and for certain type of unitary matrices such as FFT, Walsh Hadamard and random Haar unitary random matrices, expression 5.16 has a simple form in the asymptotic regime given by the following theorem: Theorem 5 : when N , the SINR li at iteration l at the output of the MMSE iterated Multiuser detector converges almost surely to a value l that is the unique solution of the equation in the ergodic case:

0

t (1 |l1 |2 )t + 2

p(t) dt =

l . (1 |l1 |2 )l + 1

(5.17)

with l = 1 2Q

1 2 1 2

|h( f )|2

df =

l (1 |l1 |2 )l +1

(5.18)

with l = 1 2Q

The theorem can be proved by recurrence on l. At each iteration, for a given channel and noise variance 2 , l1 can be also be computed straightforwardly at the receiver according to equation (5.17) and (5.18). At the rst iteration, 0 = 0 and the MMSE performance results of the previous section is achieved. The SINR i 1 is asymptotically the same for all the symbols. Therefore, each QPSK constellation has a probability of error given by: Q 1 and 1 = 1 I with 1 = E(si si 1 ) = 1 2Q 1 . In this case, 2 = wH HN H [2 I + (HN WK )(I 1 1 )(HN WK )H ]1 HN wi i i =

4 Results

H

N

for QPSK constellations have been given but the theorem can be generalized to any type of constellations

2

5. N ON -L INEAR D ETECTION

1 N

2

SCHEMES FOR

LP-OFDM

1 2 1 2

|h( f )|2

d f for non-ergodic channels. this justies equations 5.17 and 5.18 . If we suppose that at iteration l, the

Since 2 = i

SINR i 1 is asymptotically the same for all the symbols 1 i N, then it is easy to conduct the same reasoning as the case l = 1. Let rewrite matrices Cl and Dl in the case of FFT, Walsh Hadamard and random unitary matrices: Cl Dl

H

= WH diag[ N,N

= l1 [Al diag(Al 11 , . . . , Al NN )]

2 2 1

h h N 1 ,..., 2 ] 2 + (1 | l1 |2 ) | h1 |2 + (1 | l1 |2 ) | hN |2 ]WN,N .

N|

The iterated-decision multi-user detector seems therefore suited for implementation when K = N . Indeed, only scalar inversions and matrix multiplication are required in this case. Moreover, in the case of FFT and Walsh Hadamard precoders, one can make use of their Fast implementation structure. In that particular case, it is possible to check that the iterated-decision algorithm improves the performance of the decoding at each iteration as stated by the following proposition: Proposition 4 In the case = 1 and for a xed value of 2 , the series of the SINRs (l )lN is an increasing function. Since the SINR at each iteration ( l ) is bounded (by the single user bound one user = in the ergodic case, one user = in the non-ergodic case), the series of the SINRs 2 is solution of the implicit equation: converges to a value one user . in the ergodic case:

0 t p(t) 2

dt

1 2 1 2

|h( f )|2 d f

t (1 |1 2Q |2 )t + 2

p(t) dt =

(1 |1 2Q |2 ) + 1

(5.19)

1 2

1 2

(1 |1 2Q

|h( f )|2

|2 )|h( f )|2 + 2

df =

(1 |1 2Q |2 ) + 1

(5.20)

The proof can be found in appendix J. Proposition 4 highlights the benet of iterating and shows the increase in terms of SINR. One should note that is very close to the value achieved by the single user bound at low probability of errors. Indeed, for Q 103 , 1 |1 2Q |2 1 and the equation 5.19 becomes:

95

t p(t) dt = . 2

(5.21)

Isometric random matrices We consider in this paragraph isometric random matrices with the Haar distribution. The analysis of the isometric case is more sophisticated than the previous section and requires to use free probability in order to derive the results. Theorem 6 When N and K/N 1, the SINR li at iteration l at the output of the MMSE iterated Multiuser detector with isometric Haar distributed matrices converges almost surely to a value l that is the unique solution of the equation

0

l t p(t) dt = . (1 |l1 |2 )t + 2 (1 |l1 |2 )(1 )l + 2 (1 |l1 |2 )l + 1 l in the case of QPSK constellations and 0 = 0.

(5.22)

with l = 1 2Q

1 2

1 2

l |h( f )|2 df = . (1 |l1 |2 )|h( f )|2 + 2 (1 |l1 |2 )(1 )l + 2 (1 |l1 |2 )l + 1 l in the case of QPSK constellations.

(5.23)

with l = 1 2Q

The proof is detailed in appendix H For a given channel, the previous result shows that all the SINRs converge to the same deterministic value (depending on the channel realization in the non-ergodic case). An interesting point is that when the load is low, the iterated multiuser detector has no effect on the performance. Indeed, in this case, the single user bound is immediately achieved (even though a great lost of spectral efciency occurs): in the ergodic case:

0

(5.24)

1 2 1 2

(5.25)

96

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

Similar results were drawn in [90] in the CDMA context with other non-linear detectors. It is shown that when the load is low, all detectors achieve the same performance and that there is no need to use ML detection in this case. In other words, iterating has an effect only on highly redundant systems. We also give for reference purpose the performance in the case of i.i.d matrices: Theorem 7 when N and K/N 1, the SINR li at iteration l at the output of the MMSE iterated Multiuser detector with i.i.d random matrices converges almost surely to a value l that is the unique solution of the equation: in the ergodic case:

0

(5.26)

with l = 1 2Q

1 2

1 2

l |h( f )|2 df = . (1 |l1 |2 )|h( f )|2 + 2 (1 |l1 |2 )l + 2 (1 |l1 |2 )l + 1 l in the case of QPSK constellations.

(5.27)

with l = 1 2Q

The i.i.d case is also interesting to study. When 0, i.i.d and isometric precoders perform similarly. Moreover, for a xed , when l , l converges to the single user bound as shown by equation .5.26. In other words, after a certain number of iterations, i.i.d precoders perform very well for a xed .

BER Considerations: for the ergodic channel: Since the output SINR converges to a deterministic constant at each iteration, the BERl at iteration l is dened as: BERl = Q l

for the non-ergodic channel: In this case, the SINR is a random variable depending on the channel realization. We can therefore dene a mean BERl as : BERl = E Q l

SINR Evolution In Figure 5.8, the SINRs for the rst iterations and for an innite number of iterations are plotted for QPSK constellations and Rayleigh channel for ergodic channels. The results are in accordance with proposition 4. Indeed, as the number of iterations increases, the SINR converges to the single user bound when 2 .

Asymptotic values of the SINRs, iterateddecision algorithm 1st iteration 2nd iteration 3rd iteration Gaussian channel limit ID

97

16

14

12

10

SINR

8 2 1/sigma

10

12

14

16

Figure 5.8: Iterated-decision multiuser detector : values of the SINRs, Rayleigh channel, QPSK constellation, = 1.

Diversity Considerations Figure (5.9) and (5.10) plot the BER performance for various iterations of the algorithm assuming respectively independent Rayleigh channels and ETSI BRAN HiperLan/2 channel A (for N = 256 subcarriers). The single user bound is provided for a reference and corresponds N |h(i)|2 to HN = i=1N IN . One can notice that the single user bound can be achieved with only a few iterations Eb (3 iterations) and this with a reasonable decoding complexity. For sufcient high N0 (more than 6dB in our case), as the number of iterations increases, the reliability of the detected symbols increases leading to a faster convergence of the algorithm. We have also studied the case of a channel of 8 taps in gure 5.11. In all these cases, we still need a small number of iterations to converge to the optimum. As expected, Parallel interference cancellation schemes seem better suited when uniform spreading is performed at the transmitter side.

Impact of the orthogonality of the precoder In this section, we study the impact of the orthogonality of the precoder when used with the PIC algorithm. Remarkably, we observe that the PIC algorithm is really effective in this case and converges to the single user bound. In gure 5.12, we have plotted the performance of a random i.i.d precoder in a Rayleigh fading channel with = 1. We observe that in this case, after six iterations, the curves are close to the single user bound for high SNR. For low SNR, there is a mismatch due to the fact that in this region, the assumption that the estimated symbols are decorrelated from the receive symbols, does not hold. It affects more the i.i.d case than the isometric case since the probability of error for the rst step of the algorithm is higher in the i.i.d case then in the isometric case. However, using a non isometric precoder can still achieve a good performance. The main difference lies in the fact that the number of iterations needed to achieved the same target of BER is higher in the i.i.d case than in the orthogonal case. Therefore, for this detector, using i.i.d precoders in the uncoded case is not suboptimal in terms of performance but mostly in terms of complexity (since for the isometric case, the MMSE equalization already reduces a great gap which is not the case for the i.i.d case). A simple explanation can be provided by the fact that i.i.d precoders perform well at high SNR. The theoretical SINR after an innite number of iterations of the i.i.d precoder is solution of the following equation in

98

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

10

10

BER 10

3

10

10

simul MMSE theory MMSE simul PIC 2nd theory PIC 2nd simul PIC 3rd theory PIC 3rd simul PIC 5th theory PIC 5th single user simul OFDM 0 2 4 6 8 E /N [dB]

b 0

10

12

14

16

N=256, Channel A

10

10

BER 10

3

10

10

simul MMSE "Theory" MMSE Simul PIC 2nd Theory PIC 2nd Simul PIC 3th Theory PIC 3th Simul PIC 5th Theory PIC 5th single user 0 2 4 6 8 Eb/N0 [dB] 10 12 14 16

N=256, 8 taps

99

10

10

BER 10

3

10

10

simul MMSE theory MMSE simul PIC 2nd theory PIC 2nd simul PIC 3rd theory PIC 3rd simul PIC 5th theory PIC 5th single user simul OFDM 0 2 4 6 8 E /N [dB]

b 0

10

12

14

16

the ergodic case (the same reasoning applies in the non-ergodic case)

t (1 |1 2Q For Q |2 )t + 2 + 2 (1 |1 2Q |2 )

p(t) dt =

(1 |1 2Q |2 ) + 1 (5.28)

103 , 1 |1 2Q

t p(t) dt = . 2

(5.29)

which is the single user bound. Therefore, the algorithm converges rapidly once a BER threshold is reached. This threshold is achieved after only one or two iterations. It is important to note that the optimistic results of the analysis of the PIC algorithm is not an effect of the nite size of the simulated system but a shortcoming of the optimistic analysis based on the independence assumption. As matter a fact, uncoded PIC algorithms [95] can be analyzed using a different technique, based on statistical mechanics results and the replica method (a standard trick used by physicists to analyze macroscopic properties of spin glasses). In the coded case, PIC algorithms can be analyzed and derived in a unied framework via the use of factor graphs, as instance of the sum-product algorithm [20] in the CDMA context. In this case, the independance assumption can be justied.

100

5. N ON -L INEAR D ETECTION

N=256 , Rayleigh Channel 10

1

SCHEMES FOR

LP-OFDM

10

10

10

Simul MMSE Theory MMSE Simul 1st etape Theory, 1st etape Simul 2nd etape Theory, 2nd etape Simul 3rd etape Theory, 3rd etape Simul 4th etape Theory, 4th etape Simul 5th etape Theory, 5th etape Simul 6th etape Theory, 6th etape

BER

10

6 E /N [dB]

b 0

10

12

5.4 Conclusion

The Parallel Interference Cancellation algorithm has been shown to be very efcient. For this detector, our asymptotic analysis gave us means for deriving the theoretical expression of parameters involved in the detection process for a given channel and noise variance. Moreover: when orthogonal codes are used, performance close to the single user bound (with no coding) is achieved. The use of orthogonal codes is justied in this case by the very simple implementation structure of the PIC detector when = 1 (no matrix inversion required!) when i.i.d codes are used, remarkably, close to single user bound performance is still achievable at the price of a greater number of iterations then for the orthogonal case. However, the implementation requires matrix inversions at the receiver. As for the SIC detector, the use of orthogonal spreading for the PIC detector is more a question of complexity than a performance issue.

101

Conclusions

This thesis aims at studying the performance of various precoders in the context of Linear Precoded OFDM (LP-OFDM) schemes while keeping in mind the trade-off complexity versus performance issue with no channel knowledge (which is a realistic scenario for wireless communications) at the transmitter. LP-OFDM exploits all the diversity of the channel and the greatest challenge of such a scheme is to nd a simple and yet efcient decoding structure. Indeed, while OFDM requires no equalization at the receiver (only phase compensation of the channel), the precoding operation generates inter-carrier interference increasing therefore the complexity of the receiver. Even though detection techniques from the CDMA context can be used, the specic model of LP-OFDM (diagonal at frequency fading channel by a code matrix) provides means to have a simple decoding structure (as underlined in chapter 5). Based on the observation that the systems are designed with a high number of carriers, an asymptotic analysis of the LP-OFDM scheme has been carried out and was shown to apply when reasonable matrix size dimensions are used. This analysis is quite insightful on the parameters to be considered for a given channel/equalizer and the results can be easily exploited. In particular, we have shown that: All Haar distributed isometric matrices achieve the same performance asymptotically. The motivation of using such matrices nds its roots in the ability to provide an interpretable formula for the SINR through the use of the free probability theory. Remarkably, results indicate a rather fast convergence in our context to the asymptotic limits with nite size, demonstrating the practical importance of these limits. The isometric Haar distributed precoding matrices give also interesting reference bound measures since Walsh Hadamard and FFT matrices seem to perform similarly (and which are easy to implement through their fast algorithm version: Fast Fourier Transform or Fast Walsh Hadamard Transform, cf. chapter 4 and chapter 5). The optimum amount of redundancy is nearly one for ergodic and non-ergodic channels when MMSE equalization is performed with isometric precoders. These results encourage the use of short powerful codes in conjunction with square linear precoding matrices. In this case, the effects of fading are entirely controlled by the precoder and therefore only additive noise remains for coding to control. We sustain that this result is not at all obvious: there is no reason why the trade-off between coding versus linear precoding favors coding when suboptimum equalization is performed. Hence, when i.i.d linear precoding matrices are used, a non negligible trade-off occurs (as proved in chapter 4). Similar results were also obtained in the quite different context of at fading random CDMA schemes [90]: in the uplink CDMA scenario, the optimum spectral efciency is achieved for an amount of redundancy of nearly 0.75 with MMSE receivers at 10 dB. Moreover, MMSE equalization is extremely simple to perform with square unitary matrices (K =

102

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

N). As shown in chapter 4, only a scalar equalization/compensation of the channel (as in OFDM) is performed followed by a matrix despreading operation. This enables low cost mobile receivers and only adds a linear precoding operation to OFDM systems. As a consequence, the same frequency channel estimation methods of OFDM systems (HiperLAN/2 or IEEE802.11a) based on training sequences can apply. For usual constellations, there exists a threshold (depending on the rate and on the type of modulation used) for which LP-OFDM outperforms COFDM. As the modulation size increases, Coded OFDM schemes perform very well since the emitted code words have greater Hamming distances (number of distinct components)[18, 22, 51, 21] but the decoding complexity increases. This threshold has been established through a cut-off rate approach and a Convolutional Coding study (chapter 4). The effect of the number of channel taps has also been highlighted for a unit average energy channel. When the number of taps increases, the SINR after MMSE equalization becomes less random and the performance increase. This is a well known justication of the use of time and frequency interleaving (chapter 4). Finally, we have studied the asymptotic performance of two suboptimum iterative equalizers in chapter 5: The Successive Interference Cancellation (SIC) scheme has shown to incur no loss of spectral efciency: With orthogonal codes, a great improvement with respect to the MMSE receiver is achieved when is close to 1. When is small, there is no need in using such a detector. In this case, the high redundancy of the precoder is able to compensate the suboptimality of the MMSE detector. With i.i.d codes, a surprising result shows that there is also no loss in terms of spectral efciency as long as . However, since the complexity of the SIC detector is related to K (the number of inversions is related to K), using i.i.d codes can be prohibitively too complex in this case. For 0 1, the SIC detector with i.i.d codes performs similarly as the MMSE detector with orthogonal codes. In this region, orthogonality pays off. Unlike the MMSE case, the use of orthogonal precoders with SIC equalizers is not dictated by a performance criteria but more by a complexity criteria. The complexity of the receiver in the i.i.d case is much higher than the orthogonal one for the same target of spectral efciency since one has to make in the i.i.d case. For these two criterias to be fullled (complexity and performance) the best choice is an orthogonal precoder with = 1. Moreover, one has to keep in mind that the SIC scheme imposes a specic coding rate on each carrier dependent on the channel knowledge at the emitter. This condition cannot always be fullled. We also showed that in the uncoded case, the propagation of errors in the detection process yields to a very small performance gain in terms of BER. All these remarks show that the SIC detector is unattractive unless simple recursive matrix inversions are derived 5 . The Parallel Interference Cancellation algorithm has shown to be very efcient. For this detector, our asymptotic analysis gave us means to derive an expression of the parameters involved in the detection process for a given channel and noise variance. Moreover:

5 Multiresolution decoding algorithm for OFDM-CDMA, Spread OFDM single user systems , submitted patent on Multiresolution algorithm where we take benet of the tree structure of the Walsh Hadamard matrix to decode the transmitted bits, depending on the target BER, not detailed in this document

5.4. C ONCLUSION

103

When orthogonal codes are used, performance close to the single user bound (with no coding) is achieved. The use of orthogonal codes is justied in this case by the very simple implementation structure of the PIC detector when = 1 (no matrix inversion!) When i.i.d codes are used, remarkably, close to single user bound performance is still possible to achieve at the price of a greater number of iterations than the orthogonal case. However, the implementations require matrix inversions at the receiver. As for the SIC detector, the use of orthogonal spreading is more a question of complexity than performance. Finally, we should emphasize that the previous remarkable results based on the free probability theory have very simple interpretations and can be easily exploited by the engineering community. We give hereafter some research topics to be pursued: as already mentioned in the introduction, the results (and their conclusions) can be straightforwardly extended to MC-CDMA and Multi-antenna schemes. Multi-Carrier CDMA (MC-CDMA) downlink transmissions [45]: in this case, the elements of s(n) represent K different streams of symbols destined to K different users, N coincides with the number of subcarriers, and each column of W N represents the code allocated to each user. Vector WN s(n) is sent to an OFDM modulator, and equation (3.1) represents the signal received after guard interval suppression and Fourier transformation. In the frequency domain, the received N 1 vector signal can be seen as resulting from a transmission over N parallel at fading channels. Diagonal entries of H N represent the frequency domain channel gains for data frame n. The non-equal power is also quite relevant in the context of multiuser systems. The approach used in the present thesis cannot be generalized to this context because the calculation of the SINR involves explicitly the equal power assumption. The non-equal power case requires the use of more sophisticated tools as shown in [30]. Multi-antenna transmissions: HN is a N N matrix with non zero entries. Even though the channel is no more diagonal, our result concern only the limiting distribution of the eigenvalues and not specic to the matrix type. The columns of the channel matrix correspond to the number of transmitting antenna while the rows correspond to the number of receiving antennas. All linear space time codes can be studied with our model and the same analysis can be conducted with the different equalizers. As a typical example, recently [13] has showed, by using free probability theory that as the dimensionality of the system increases, codes based on the Euclidean distance perform quite similarly or better then the ones based on the minimization of the determinant and rank criteria [98]. The effect of the channel estimation on the robustness of our results is to be conducted. All our results assume perfect channel knowledge at the receiver as well as perfect synchronization. Moreover, the case of non-coherent detection has not been investigated and interesting results could be achieved in this direction. Efcient coding and decoding schemes are still a great concern for the PIC schemes. This thesis has only focused on the equalization part and not the coding/decoding structure of the PIC scheme.

104

5. N ON -L INEAR D ETECTION

SCHEMES FOR

LP-OFDM

The cut-off rate rate approach can be extended to Maximum likelihood LP-OFDM decoding with Haar distributed matrices and compare the results with the optimum coefcients of the number theory approach. In this document, the single-user scenario LP-OFDM scheme has been studied and greater work needs to be pursued in order to extend these results in the context of multi-user systems (effect of non-perfect synchronization between users, interference in ad hoc/centralized networks). The combination of LP-OFDM with multiple antennas is also a research topic to be investigated in terms of performance design (number of antennas, redundancy of LP-OFDM, . . . ).

105

Appendix A

This section provides general useful theorems and denitions in relation with the mathematical results developed in this thesis. For specic developments, the reader can consult [84]. At the end of each result, the chapter using the result is given.

A.1

Let (, A, P) be a probability space. Denition 1 A sequence f n of random variables is said to converge to a function f everywhere or pointwise if lim fn () = f ()

n

for every . Denition 2 A sequence f n of random variables is said to converge to a measurable function f almost everywhere or almost surely (usually abbreviated as a.s) if there exists a measurable set N with P(N)=0 such that: lim fn () = f ()

n

for every

Nc

Theorem 8 (Borel Cantelli Lemma) let {An } be a sequence of events and let A be the event consisting of the occurrence of an innite n=0 number of events An , n = 1, . . . Then if

n=1

P(An ) <

P(A) = 0

then

106

A. U SEFUL M ATHEMATICAL The Borel-Cantelli is used in the proof of theorem 1 in the Appendix D

RESULTS

Denition 3 A sequence f n of measurable functions is said to converge to a measurable function f in probability if lim P[ :| fn () f () | ] = 0

n

for every > 0. Denition 4 A sequence n of probability measures on the real line R with distribution functions Fn (x) is said to converge weakly to a limiting probability measure with distribution function F(x) (in symbols n or Fn F, F is also called the limit distribution) if

n

for every x that is a continuity point of F. Theorem 9 (Lvy-Cramer continuity theorem) The following are equivalent: n or Fn F For every bounded continuous function f(x) on R

n R

lim

f (x)dn =

f (x)d

If n (t) and (t) are respectively the characteristic functions of n and , for every real t,

n

The Lvy-Cramer continuity theorem is not used but knowledge of these results are necessary to understand the proof of theorem 1 in the Appendix D. Theorem 10 (The dominated Convergence Theorem) If for some sequence { f n } of measurable functions we have f n f almost surely and | f n () | g() for all n for some integrable function g, then fn dP f dP as n . The dominated convergence theorem is used in the proof of theorem 3 in the Appendix F Theorem 11 (Fubinis Theorem) Let f () = f (1 , 2 ) be a measurable function of on (1 2 , A1 A2 , P1 P2 ). Then f can be considered as a function of 2 for each xed 1 or the other way around. The functions g 1 (.) and h2 (.) dened respectively on 2 and 1 by: g1 (2 ) = h2 (1 ) = f (1 , 2 )

A.1. P ROBABILITY

THEORY RESULTS

107

are measurable for each 1 and 2 . If f is integrable then the functions g 1 (2 ) and h2 (1 ) are integrable for almost all 1 and 2 respectively. Their integrals G(1 ) = and H(2 ) =

1 2

g1 (2 )dP2

h2 (1 )dP1

are measurable, nite almost everywhere and integrable with respect to P1 and P2 respectively. Finally, f (1 , 2 )dP = G(1 )dP1 = H(2 )dP2

Conversely for a nonnegative measurable function f if either G or H, which are always measurable, has a nite integral, so does the other and f is integrable with its integral being equal to either of the repeated integrals, namely integrals of G and H. The Fubinis theorem is used in the proof of theorem 3 in the Appendix F Denition 5 (Innitely Divisible Distributions) A probability distribution is said to be innitely divisible if for every positive integer n, can be written as the n-fold convolution (n )n of some other probability distribution n Remark 1 We give hereafter some useful inequalities: Jensens Inequality If (x) is a convex function of x and f () and ( f ()) are integrable then: ( f ())dP Holders Inequality E (| XY |) E(| X | p 1 1 + =1 p q Schwarzs Inequality (case p=2) E (| XY |) E(| X | 2 ) Markovs Inequality P[| X | a] r > 0. E (| X |r ) ar

1 1

f ()dP

E(| Y | q

E(| Y | 2 )

108

A. U SEFUL M ATHEMATICAL

RESULTS

Some of these results are used in the proof of theorem 1 in the Appendix D. Denition 6 (Unitarily invariant) An N N self-adjoint random matrix A is called unitarily invariant if the probability measure of A as a random matrix is equal to that of the matrix VAV H for any unitary constant matrix V .

109

Appendix B

B.1 Random matrix theory

Random matrix theory have been a part of advanced multivariate statistical analysis since the end of the 1920s. Random matrices were rst proposed by Eugene Wigner in quantum mechanics where the statistics of experimentally measured energy levels of nuclei were explained in terms of the eigenvalues of random matrices. When Ts [102] and Verdu [113] introduced nearly simultaneously random matrices in 1999 to model uplink unfaded CDMA systems equipped with certain type of receivers, random matrix theory entered the eld of telecommunications. From that time, random matrix theory has known considerable interest to study the performance of multiple antenna schemes [73, 100] This section is intended to give a brief survey on random matrix theory. Our aim is to give a better understanding of matrix theory in order to apply the work of Ts and Verdu to i.i.d Linear Precoders for OFDM. Let us give rst some basic denitions. Let be a probability measure on R. Its Stieljes transform:

G (z) =

d(t) t z

is dened on C/R. When z lies in the upper half-plane C + = (z C : Im(z) > 0), the transform G (z) is an analytic function in C+ possessing the following properties: G (C+ ) C+ and | G (z) |

1 Im(z)

Let be compactly supported. Then G (z) is analytic in a neighborhood of . Since (z t) 1 = t k zk1 , it is obvious that G (z) has the following expansion at z = : k=0

110

B. U SEFUL

G (z) = z1 + mk ()zk1 .

k=0

where mk () = t k d(t)(k Z+ ) Writing z = x + iy, it is possible to recover from its Cauchy transform up to a factor . When is absolutely continuous with respect to the Lebesgue measure, its density f(x) is given by: 1 f (x) = + lim ImG (x + iy) y0 The random matrix theory in consideration here (see [68, 93] for more details) deals with the limiting distribution of random Hermitian matrices of the form A + W DW H . Here, W (N K), D(K K), and A(N N) are independent, with W containing i.i.d entries having nite second moments, T is diagonal with real entries, A is Hermitian and K/N > 0 as N . The behavior is expressed in terms H of the limiting distribution (see section A for the denition) function F A+W DW of the eigenvalues of H A + W DW H (i.e F A+W DW (x) is the proportion of eigenvalues of A + W DW H x). The remarkable H result of random matrix theory is the convergence in some sense, of F A+W DW (x) to a non random F. The papers vary in the assumption on T, W and A. We will only take here the case of interest. Theorem 12 Let A be a N N hermitian matrix, nonrandom, for which F A converges weakly as N D to a distribution function A. Let F converges weakly to a nonrandom probability distribution function denoted D. Suppose the entries of NW i.i.d for xed N with unit variance (sum of the variances of the real and imaginary parts in the complex case). Then the eigenvalue distribution of A+W DW H converges weakly to a deterministic F. Its Stieljes transform G(z) satises the equation:

G(z) = GA z

dT() 1 + G(z)

(B.1)

We give hereafter the eigenvalue distribution of matrices dened by: WN,K WH + 2 I N,K where NWN,K is a N K matrix with i.i.d entries with zero mean and variance one and N such as K xed. This example is intended to show to what extent (in terms of the matrix dimension) N theoretical and practical results t. Moreover, we give the basic machinery that the reader can use to derive any limiting eigenvalue distribution of matrices dened as in theorem 12.

B.2. A N APPLICATION

EXAMPLE :

111

Denote A = 2 I and D = IK,K . In this case, dA(x) = (x 2 ) and dD(x) = (x 1). Applying theorem 12, the following result is obtained:

G(z) = G2 I z = G2 I (z = =

( 1)d 1 + G(z)

2 z +

G2 I (z) is the Cauchy transform of the eigenvalue distribution of matrix 2 I. The solution of the second order equation B.5 yields: 1 1 + 1 2 z) 2 z) 2( 2 2(

G(z) =

f () =

[1 ]+ (x) + (2 ) 0

2 1 ( 2 + 1 )2 4

Where (x) is a unit point mass at 0 and [z] + = max(0, z). A remarkable result is that the eigenvalues of such matrices have a compact support while the entries may take any value (subject to a zero mean and variance one distribution). We have plotted the theoretical and practical eigenvalue distribution for various size matrices. Only one realization of matrix WN,K WH + 2 I has been studied. As the dimensions increase, the practical results t extremely well N,K the theoretical results as shown in g.A.1 and g.A.2.

112

B. U SEFUL

0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0

distributiob of the eigenvalues for alpha=0.25 and no noise simulations with matrix of size 256*64 theory

0.5

1.5

2 eigenvalues

2.5

3.5

0.09

distributiob of the eigenvalues for alpha=0.25 and SNR=10db simulations with matrix of size 4096*1024 theory

0.08

0.07

0.06

0.05

0.04

0.03

0.02

0.01

0.5

1.5

2 eigenvalues

2.5

3.5

113

Appendix C

Free probability is a non-commutative probability theory, in which the concept of independence of classical probability is replaced by that of freeness. The main concern for introducing this theory in this thesis is:

1 to show that NK tr((I UN UH )BN (I UN UH ) converges to a deterministic constant (see short N N proof of theorem 1 in section 4)

to compute of the spectrum of the product of two hermitian matrices (matrix W N,K WH HH HN in N,K N the SINR formula of section 4). Knowing the respective spectra of W N,K WH and HH HN , the set of possible spectra of WN,K WH HH HN N,K N N,K N depends in a complicated way on the eigenvalues and eigenvectors of W N,K WH and HH HN . HowN,K N ever, when N become large, a remarkable phenomenon occurs, and it turns out that, roughly speaking, for almost all choices of the matrices W N,K WH and HH HN with given spectra, the spectrum of N,K N WN,K WH HH HN is essentially the same and can be computed explicitly, without knowing the eigenvecN,K N tors of WN,K WH and HH HN ! In other words, under certain asymptotic and freeness assumptions, we N,K N will be able to derive the SINR based only on the knowledge of the spectra of W N,K WH and HH HN . N,K N This section aims only at introducing some useful notions relative to free probability theory. For a more thorough development of this theory , the reader should consult [117] and the monograph [58].

C.1

Algebraic Context

Denition 7 A noncommutative probability space is a couple (A, ) where A is a non commutative unit algebra (i.e., an algebra having a unit denoted by 1) over C and : A C is a linear functional such that (1) = 1. When satises (ab) = (ba) it is called a trace. As it will appear below, the role of can be compared to that of expectation in classical probability theory.

114

C. U SEFUL

Denition 8 Let (A, ) be a noncommutative probability space. In the context of free probability, a random variable is an element a of A. The distribution of a is the linear functional a on C[X], the algebra of complex polynomials in one variable, dened by a (P) = (P(a)). In particular, the distribution of a non commutative random variable a is characterized by its moments, i.e. by the sequence ((ak ))kN . We note that in certain practical cases, the distribution of a non commutative random variable is associated to a real probability measure a (see e.g. the example below) in the sense that (ak ) = R t k da (t) for each k N. In this case, the moments of all orders of a are of course nite.

C.2

We shall consider random matrices of size N N whose entries are dened on some common probability space (meant in the classical sense) and have all their moments nite. The noncommutative probability space is obtained by associating to the algebra of these matrices the functional N (X) = 1 1 N E (tr(X)) = E(xii ) N N i=1 (C.1)

which is obviously a trace. This space will be denoted by (A N , N ). Suppose X is a random matrix with real (random) eigenvalues 1 , . . . , N . The real random measure = 1 N (i ) N i=1 (C.2)

1 is called empirical eigenvalue distribution of X. The k th moment of this probability measure is N tr(Xk ) = 1 N k k N i=1 i . The distribution X of X is dened by the fact that its action on each monomial X of C[X] is given by 1 N X (X k ) = N (Xk ) = ( E(k )) i N i=1

This distribution is of course associated to the probability measure X dened by f (t)dX (t) = for each bounded continuous function. 1 N

i=1

E( f (i )

C.3

The notion of distribution introduced in denition 8 can be generalized to the case of multiple random variables. Let a1 and a2 be two random variables in a noncommutative probability space (A, ). Consider noncommutative monomials in two indeterminates, of the form X ik1 Xik2 . . . Xikn , where for all j, i j {1, 2} 1 2 n , k j 1 and i j = i j+1 . The algebra C X1 , X2 of noncommutative polynomials with two indeterminates

C.4. F REENESS

115

k k will be the linear span of 1 and these noncommutative monomials (X 1 and X2 ). The joint distribution of a1 and a2 is the linear functional on C X1 , X2 satisfying

More generally, denote by C Xi | i {1, . . . , I} the algebra of non commutative polynomials in I variables, which is the linear span of 1 and noncommutative monomials of the form X ik1 Xik2 . . . Xikn where 1 2 n k j 1 and i1 = i2 , i2 = i3 , . . . , in1 = in are less than or equal to I. The joint distribution of the random variables {ai }i{1,...,I} in (A, ) is the linear functional: : C Xi | i {1, . . . , I} Xik1 Xik2 . . . Xikn 1 2 n C kn k1 k2 (Xi1 Xi2 . . . Xin ) = (ak11 ak22 . . . aknn ) . i i i (C.3)

In short, the joint distribution of the noncommutative random variables {a i }i{1,...,I} is completely specied by their joint moments.

C.4

Freeness

Denition 9 Let (A, ) be a noncommutative probability space. A family {A i }i{1,...,I} of unital subalgebras of A is called free if (a1 a2 . . . an ) = 0 for all n-uples (a1 , . . . , an ) verifying: (i) a j Ai j for some i j I and i1 = i2 , i2 = i3 , . . . , in1 = in . (ii) (a j ) = 0 for all j = 1, . . . , n. A family of subsets of A is free if the family of unital subalgebras generated by each one of them is free. Random variables a1 , . . . , an are free if the family of subsets {{a 1 }, . . . , {an }} is free. Notice that in the statement of condition (i), only two successive random variables in the argument of (a1 a2 . . . an ) belong to two different subalgebras. This condition does not forbid the fact that, for instance, i1 = i3 . Note in particular that if a1 and a2 belong to two different free algebras, then (a1 a2 a1 a2 ) = 0 whenever (a1 ) = (a2 ) = 0. This relation cannot of course hold if a 1 and a2 are two real valued independent random variables (in the classical sense) and if coincides with the classical mathematical expectation operator. Therefore, freeness cannot be considered as a noncommutative generalization of independence because algebras generated by independent random variables in the classical sense are not free. Let us make a simple computation involving freeness. A 1 and A2 being two free subalgebras in A, any two elements a1 and a2 of A1 and A2 respectively can be written as ai = (ai )1 + ai , so (ai ) = 0: (a1 a2 ) = (a1 )1 + a1 (a2 )1 + a2 = (a1 )(a2 )

116

C. U SEFUL

In other words, the expectations of two free random variables factorize. By decomposing a random variable ai into (ai )1 + ai , the principle of this computation can be generalized to the case of more than two random variables and/or to the case of higher order moments, and one can check that noncommutativity plays a central role there. Proposition 5 ([9]) Let {Ai }i{1,...,I} be free subalgebras in (A, ) and let {a 1 , . . . , an } A be such that for all j = 1, . . . , n, one has a j Ai j , i j I. Let be the partition of {1, . . . , n} associated to the equivalence relation j k i j = ik (i.e. the random variables a j are gathered together according to the free algebras to where they belong to) . For each partition of {1, . . . , n}, let = { j1 ,..., jr } (a j1 . . . a jr ). There exists universal coefcients c(, ) such that

j1 <...< jr

(a1 . . . an ) = where " " stands for " is ner than ".

c(, )

One consequence of this proposition is that given a family of free algebras {A i }i{1,...,I} in A, only restrictions of to the algebras Ai are needed to compute (a1 , . . . , an ) for any a1 , . . . , an A such that for all j = 1, . . . , n, one has a j Ai j , i j I. The problem of computing explicitly the universal coefcients c(, ) has been solved using a combinatorial approach.

C.5

Free Multiplication

Let and be two compactly supported probability measures (A function has compact support if it is zero outside some nite interval). Then from [58], it always exists two free random variables a 1 and a2 in some noncommutative probability space (A, ) having their distributions associated to and respectively. One can see that the distribution of the random variable a 1 a2 depends only on and on . The reason for this is the following: denition 8 says that distribution of a 1 a2 is fully characterized by the moments ((a1 a2 )n ). To compute these moments, we would just need the restriction of to the algebras generated by {a1 } and {a2 }, in other words, ((a1 a2 )n ) depends on the moments of a1 and a2 only. It can be shown that the distribution of a 1 a2 is associated to a probability measure called free multiplicative convolution of the distributions and of these variables. It is denoted by . Multiplicative free convolution is commutative, and when and are compactly supported measures on [0; [, then so is ([117, p. 30]). Moments of are related in a universal manner (independent of the particular noncommutative probability space) to the moments of and to those of . It happens that direct computation of the moments of a multiplicative free convolution is hardly practicable. It is feasible on the other hand using tools of analytic function theory. By the use of the so-called S-transform introduced in [115], multiplicative free convolution is converted into a mere multiplication of power series: Proposition 6 Given a probability measure on R with compact support, let (z) be the formal power series dened by zt d(t) . (C.4) (z) = zk t k d(t) = 1 zt k1

AND

R ANDOM M ATRICES

117

t d(t) is non zero, it exists a unique function analytic in a neighborhood of 0 ( (z)) = z S (z) = (z) (C.5) (C.6) is given by

= S S .

There is also a result in the same vein for additive free convolution. It will not be needed in this thesis but the interested reader is referred to [117].

C.6

Voiculescu [115, 116, 117] discovered very important relations between free probability theory and random matrix theory. Random matrices are typical noncommutative random variables as can be seen in the example in paragraph C.2. In [116], it is shown that certain independent matrix models exhibit asymptotic free relations. Denition 10 Let {XN,i }i{1,...,I} be a family of random N N matrices that belong to the noncommutative probability space (AN , N ) dened in paragraph C.2. Their joint distribution is said to have a limit distribution (see secA for the denition) on C Xi | i {1, . . . , I} as N if (Xik1 . . . Xikn ) = lim N (Xk1 1 . . . Xkn n ) N,i N,i n 1

N

for any noncommutative monomial in C Xi | i {1, . . . , I} . Consider the particular case where I = 1 (we denote X N,1 by XN to simplify the notations), and assume XN has real eigenvalues and that the distribution of X N has a limit distribution . Then, for each k 0, (X k ) = lim

N+

t k dXN (t)

(C.7)

where XN is the measure associated to the distribution of X N . Remark 2 If is associated to a compactly supported probability measure , the convergence of the moments of XN to the moments of expressed by (C.7) implies that the sequence ( XN )NN converges weakly to , i.e. f (t)d(t) = lim

N+

f (t)dXN (t)

1

(C.8)

1 Convergence

of moments implies weak convergence if the function z eitz d(t) is analytic in a neighborhood of 0, a condition which is clearly met is is compactly supported.

118

C. U SEFUL

Denition 11 The family {XN,i }i{1,...,I} of random matrices in (AN , N ) is said to be asymptotically free if the two following conditions are satised: (i) This family has a limit distribution . (ii) For every family of integers {i 1 , . . . , in } in {1, . . . , I} verifying i1 = i2 , . . . , in1 = in , and for every family of polynomials {P1 , . . . , Pn } in one indeterminate verifying

N

n

(C.9)

we have

N

lim N

Pj (XN,i )

j=1

j

=0.

(C.10)

Stated concisely, the second condition requires that the family of algebras {C[X i ]}i{1,...,I} be free in the noncommutative probability space (C Xi | i {1, . . . , I} equipped with the linear functional . The kind of asymptotic freeness introduced by Hiai and Petz is more useful for our purpose because it deals with almost sure convergence under the normalized classical matrix traces instead of convergence under the functionals N . Following [58], the family {XN,i }i{1,...,I} in (AN , N ) is said to be have a (non random) limit almost everywhere if (Xik1 . . . Xikn ) = lim n 1 1 tr(Xk1 1 . . . Xkn n ) a.s. N,i N,i N N

for any noncommutative monomial in C Xi | i {1, . . . , I} . In the case where N = 1 and X N has real eigenvalues, if the almost sure limit distribution of X N is associated to a compactly supported probability measure , this condition means that f (t)d(t) = lim 1 N f (i,N ) a.s. N+ N i=1

for each continuous bounded function f (t). In other words, the empirical eigenvalue distribution of X N converges almost surely in distribution to the measure . The family {XN,i }i{1,...,I} in (AN , N ) is said to be asymptotically free almost everywhere (A property of X is said to hold almost everywhere if the set of points in X where this property fails has measure 0) if it has a nonrandom limit distribution on C Xi | i {1, . . . , I} almost surely, and if the family {C[Xi ]}i{1,...,I} is free in the noncommutative probability space (C Xi | i {1, . . . , I} , ). This last condition can be expanded in the same manner as in the denition above: just replace N () by tr()/N in equations (C.9) and (C.10). The rst concrete random matrix models exhibiting asymptotic freeness were given in [116]. We now give [58, proposition (4.3.9)] a useful asymptotic freeness theorem. Theorem 13 Let DN and EN be N N Hermitian random matrices, and let N be a Haar distributed unitary random matrix independent from D N and EN . Assume that the empirical eigenvalue distributions of DN and of EN converge almost surely towards compactly supported probability distributions. Then, the family {DN , N EN H } is asymptotically free almost everywhere as N . N

AND

R ANDOM M ATRICES

119

An Application Example We give hereafter a way of computing the limiting eigenvalue distribution of matrices of the form WN,K WH HH HN where HN and WN,K satisfy the assumptions formulated in section 4.3.2. In the erN,K N godic case, put DN = HH HN , and denote by d(t) = p(t)dt the compactly supported probability disN tribution of the diagonal entries of this matrix. Therefore, the empirical eigenvalue distribution of D N converges weakly almost everywhere to the probability distribution . WN,K is obtained by extracting K columns from a N N Haar distributed random unitary matrix N . Without loss of generality, this matrix can be written as W N,K = N EN , where the N N matrix EN has the structure EN = diag([1, . . . , 1, 0, . . . , 0]) with tr(E N ) = K. Assume that K/N 1 when N . Then, the empirical eigenvalue distribution of E N converges to = (1) + (1 )(0). Conditions of theorem 13 are satised and then the family {H H HN , WN,K WH = N EN H } is asymptotically free. N N N,K This allows to derive the almost sure limit eigenvalue distribution of W N,K WH HH HN . To see this, N,K N denote by the a.s. limit distribution of {W N,K WH , HH HN } in C X1 , X2 . In particular, for each k N, N,K N ((X1 X2 )k ) = lim 1 tr((WN,K WH HH HN )k ) . N,K N N+ N (C.11)

Distributions of the monomials X1 and X2 are associated to the compactly supported measures and respectively. As X1 and X2 are free, the distribution of X1 X2 is associated to the measure . As is also compactly supported, (C.11) and remark 2 imply that the empirical eigenvalue distribution of WN,K WH HH HN converges weakly to . The easiest way to evaluate consists in using the N,K N fact that the S-transform of is the product of the S-transform of by the S-transform of . Let us precise this using the same notations as in proposition 6. It is obvious that the function (z) associated to there is given by z (z) = 1z After a simple calculation, we derive its inverse (z) through (z) = From S

z . +z

(z)

1+z (z) . +z

(z)

(z)

(z)

= z

(C.12)

Therefore, the function (z), which in principle, characterizes the measure , is obtained by solving equation (C.12). This result gives straightforwardly equation4.18 of theorem 1 of section 4.

120

C. U SEFUL

121

Appendix D

Proof of Theorem 1

In this section, we rst establish theorem 1 in the ergodic case. The non-ergodic case is briey considered at the end. We assume that HN and WN,K satisfy the assumptions formulated in section 4.3.2. In the following, we will denote by AN the N N matrix dened by AN,K = (I UN,K UH )HH HN UN,K UH HH + 2 IN N,K N N,K N

1

tr(A )

(D.1)

N,K In order to prove theorem 1, we rst establish in proposition 7 that wk NK converges to zero almost surely. The result of proposition 7 can be seen as a generalization to the context of isometric random tr(AN,K ) matrices of corollary 1 in [44]. Next, we use asymptotic freeness results to show that NK converges almost surely to the quantity solution of formula (1.7).

We rst recall the following useful result of [6] and [44] : Lemma 1 Let zN be a N 1 random vector and BN a N N random matrix independent of z N . Assume that the elements of zN are i.i.d. and have a unit variance and a nite eighth order moment and that supNN BN < + where denotes the spectral norm. Denote by N the random variable dened by N = Then, where C is independent of N. We now show the following result. Proposition 7 Assume that K/N when N +. Then,

N+

lim

wk

tr(AN,K ) N K

= 0 a.s.

(D.3)

tr(AN,K ) NK .

D. P ROOF

OF

T HEOREM 1

In order to establish (D.3),it is sufcient to show that if K/N , E(|eN |4 ) = O(N 2 ) E |eN |4 4 (D.4)

Indeed, if (D.4) holds, Markovs inequality implies that > 0, P (|eN | > ) Therefore,

= O(N 2 ) .

N=1

N,K We now establish (D.4). Assume that w i is the column i of WNK . eN = wi NK is a function fi (WN,K , HN ) of random matrices WN,K and HN . Let P be the permutation matrix exchanging column i with column K, and denote by WN,K the matrix WN,K = WN,K P. It is obvious that f i (WN,K , HN ) = N,K , HN ). Moreover, WN,K and WN,K have the same distribution. Random variables f K (WN,K , HN ) f K (W and fi (WN,K , HN ) are thus identically distributed, which implies that

tr(A

E| fi (WN,K , HN )|4 = E| fK (WN,K , HN )|4 Therefore, there is no restriction to assume in the proof of identity (D.4) that w K is the column K of WN,K , a condition which is supposed to hold all along the proof. Finally, E| f K (WN,K , HN )|4 does not of course depend of the particular way W N,K is generated, provided it is Haar distributed. We can therefore assume that WN,K consists of the rst K columns of the Haar distributed random matrix V(X N ) obtained from a complex Gaussian i.i.d. random matrix X N through formula (4.5). As V(XN ) is obtained by a Gram-Schmidt orthogonalization of X N = [x1,N , . . . , xK,N , . . . , xN,N ] and vector wK is assumed to be the column K of V(XN ), wK can be written as wK = (I UN,K UH )xK,N N,K (I UN,K UH )xK,N N,K (D.5)

It is important to note that UN depends only on vectors x1,N , . . . , xK1,N . Therefore, vector xK,N is independent from UN,K . The expression (4.4) of the SINR wK becomes wK = xH AN,K xK,N K,N (I UN,K UH )xK,N N,K

2

(D.6)

where AN,K denotes the matrix dened by (D.1). Put eN = e1,N + e2,N where e1,N and e2,N are dened by e1,N = and e2,N = xH AN,K xK,N K,N (I UN,K UH )xK,N N,K 2 xH AN,K xK,N K,N N K

D. P ROOF

OF

T HEOREM 1

123

Then, E|eN |4 8(E|e1,N |4 + E|e2,N |4 ). We rst use lemma 1 to show that E|e 2,N |4 = O(N 2 ). As UN,K and xK,N are independent, AN,K and xK,N are also independent. Moreover, the spectral norm of (HN UN,K UH HH + 2 I)1 is uniformly bounded by 1/2 . As HN is uniformly bounded, the N,K N inequality CD C D implies that A N,K is uniformly bounded. Furthermore, x K,N meets the conditions of lemma 1 since it has Gaussian independent elements. This lemma and the fact that N/(N K) 1/(1 ) imply that E|e2,N |4 = O(N 2 ) converges to 0 almost surely (set z N = xN and BN = AN,K ). By the Borel-Cantelli Lemma, E|e 2,N |4 = O(N 2 ) implies that e2,N converges to 0 almost surely. We now study the behavior of E|e1,N |4 . e1,N can be written as e1,N = xH AN,K xK,N K,N N K N K (I UN,K UH )xK,N N,K

2

As e2,N converges to 0 almost everywhere, xH AN,K xK,N tr(AN,K ) K,N <2 2 AN,K (N K) (N K) a.s.

for N large enough. The last inequality comes from the facts that for a given matrix X, tr(X) X rank(X), and that the rank of AN,K does not exceed N K as can be seen from the expression of this matrix. As supNN AN,K < , xH AN,K xK,N / (N K) is bounded almost everywhere. K,N We now show that E N K (I UN,K UH )xK,N N,K

2 4

= O(N 2 ).

H Let UN,K be a N (N K) isometric matrix such that (I U N,K UH ) = UN,K UN,K . Then, (I N,K H H UN,K UH )xK,N 2 = UN,K xK,N 2 . As xK,N and UN,K are independent, UN,K xK,N is a (N K)-dimensional N,K H complex Gaussian random vector with covariance matrix I NK ). Therefore, UN,K xK,N 2 is 2 dis-

tributed with 2(N K) degrees of freedom. Its probability density is the function straightforward direct computation shows that E N K (I UN,K UH )xK,N N,K

4 2

t (NK1) t (NK1)! e ,

and a

= O((N K)2 )

which coincides with O(N 2 ) if K/N . As xH AN,K xK,N / (N K) is bounded almost everywhere, K,N we get that E|e1,N |4 = O(N 2 ). Putting all the pieces together, this implies that E|e N |4 = O(N 2 ), and that eN converges almost surely to 0. We now use the asymptotic freeness results developed in subsection C.6 to study the behavior of (tr(AN,K )) /(N K). As K/N , it is sufcient to study (tr(A N,K )) /N. Proposition 8 With the hypotheses formulated on H, N = tr(AN,K ) N

124

D. P ROOF

OF

T HEOREM 1

Proof: Assumption A4 implies that H N HH c. The spectrum of HN UN,K UH HH is thus included N N,K N in [0; c] while the spectrum of YN = HN UN,K UH HH + 2 IN is included in [2 ; 2 + c]. By StoneN,K N Weierstrass theorem, the function f (x) = 1/x can be uniformly approximated by a polynomial on the compact interval [2 ; 2 + c]. More precisely, > 0, P R[X] such that 1 |P(x) | < . x x[;+c] sup (D.7)

where R[X] is the algebra of real polynomials. In the sequel we shall approximate the matrix Y 1 which N appears in the expression of AN by a polynomial to be able to use directly the asymptotic results of free probability theory given in section C. For > 0, let P be a polynomial that satises (D.7). Using the eigenvalue decomposition YN = VN N VH , it comes from N Y1 P(YN ) = VN (1 P(N ))VH N N N that Y1 P(YN ) . Notice that the choice of a polynomial P leading to this inequality does not N depend on N. Using the inequalities CD C D and tr(C) N C and the fact that tr(CD) = tr(DC), we have 1 |N tr (IN UN,K UH )HH P(YN )HN (IN UN,K UH ) | N,K N N,K N 1 = | tr (Y1 P(YN ))HN (IN UN,K UH )HH | N,K N N N 1 H YN P(YN ) HN HN c . Denote the quantity N and M, we have

1 N tr

(IN UN,K UH )HH P(YN )HN (IN UN,K UH ) by N . For any two integers N,K N N,K

(P) (P) (P) (P) (P) (P)

(P)

|N M | |N N | + |M M | + |N M | 2c + |N M | . Thanks to this inequality and by the use of Cauchy convergence criterion, almost sure convergence of N (P) is equivalent to a.s. convergence of N . To prove this convergence, notice that for any integer n, Y n expresses as N Yn = HN UN,K UH HH + 2 IN N N,K N

(P) n

k=0

By expanding N , it is thus easy to check that it is the normalized trace of the value taken by a noncommutative polynomial in the indeterminates X1 and X2 where X1 and X2 are replaced by HH HN and N UN,K UH respectively. Thanks to theorem 13 and subsection C.6, the family {U N,K UH , HH HN } is N,K N,K N asymptotically free almost everywhere (just replace W N,K WH by UN,K UH ), and has therefore an alN,K N,K most sure limit. By the very denition of the concept of limit distribution a.s., N converges almost surely. We can now assemble the pieces and give the asymptotic value of the SINR. For this, we need to specify more carefully than above which column of W N,K is selected to evaluate the SINR. We thus modify slightly our notations as follows. The parameter K of course depends on N. As we now need to specify this dependence, we denote K by K(N). For each N, we choose k(N) as an integer such that 1 k(N) K(N), and denote wk(N) the k(N)-th column of WN,K and wk(N) the corresponding SINR.

(P)

D. P ROOF

OF

T HEOREM 1

125

The matrix AN,K dened by (D.1) is now denoted Ak(N),N . It results from propositions 7 and 8 that for each sequence (k(N))N1 , the SINR wk(N) converges a.s. to a nite value as N . We rst notice that this value does not depend on the choice of sequence (k(N)) N1 . To see this, we remark that the almost sure convergence of wk(N) toward a certain limit implies the convergence of E( wk(N) ) toward the same limit (wk(N) is less than c/ and is thus bounded almost everywhere : just use the Lebesgue dominated convergence theorem). But the statistical properties of wk(N) do not depend of the particular choice of the index k(N) because the Haar distribution is invariant by permutation. Therefore, lim N+ E(wk(N) ), and therefore limN+ wk(N) does not depend on the choice of the sequence (k(N)) N1 . We denote this limit by in the following. Using the same arguments, it follows that E| wk(N) |2 does not depend on k(N), and that lim sup E|wk |2 = 0 (D.8)

N 1kK(N)

It is immediate that wk(N) given by (4.3) after the same notation change as for wk(N) , and related to this quantity by wk(N) = wk(N) /(1 + wk(N) ) as in (4.6), converges also a.s. to a value which does not depend on the choice of the sequence k(N) neither. For a given N, there are K(N) quantities wk(N) each corresponding to the choice of a particular column code in W N,K . Their arithmetic mean N over all the columns of this matrix is N = We now prove that = =

N+

(D.9)

lim N (D.10)

exists (the limit being understood in the almost sure sense), and that coincides with . We rst prove that the limit dened in formula (D.9) exists, and give the expression of . For this, we use the results and keep the notations of subsection C.6. It is shown there that the empirical eigenvalue distribution, denoted here d N , of WN,K WH HH HN converges weakly almost surely to the N,K N compactly supported measure . But, the eigenvalues of H N WN,K WH HH of course coincide with N,K N the eigenvalues of WN,K WH HH HN . Therefore, N,K N 1 tr (HN WN WH HH + 2 IN )1 HN WN,K WH HH N N N,K N N = t dN (t) . t + 2 (D.11)

The function (t) = t/(t + 2 ) is continuous and bounded for t 0. Therefore, the limit dened in (D.10) exists almost surely and is given by = 1 t d (t) . t + 2

It remains to establish that = . We rst remark that 0 wk(N) 1. Hence, 0 N 1. The Lebesgue 2 converges to 0 when N , i.e. N dominated convergence theorem thus implies that E| N | in the mean square sense. But, converges to (E|N |2 )1/2 1 K(N) (E|wk |2 )1/2 K(N) k=1

x 1+x

D. P ROOF

OF

T HEOREM 1

|wk(N) | |wk(N) |

1 This and relation (D.8) thus implies that K(N) k=1 (E|wk |2 )1/2 converges to 0 when N . It turns out that N converges to in the mean square sense. Therefore, K(N)

== Equation (C.12) gives an expression for from this last equation that

t d (t) . t + 2

( 2 )

What is needed is the SINR , which is related to by = /(1 ). Replacing in (C.12), we have after some simple manipulations . = 2 1 + (1 ) 1+ The result is obtained after developing according to (C.4) and replacing d(t) by p(t) dt. Non-ergodic case. In that case, for a given value of the channel impulse response, the same reasoning as in the ergodic case implies the almost sure convergence of wk to the solution of (5.10). However, since the channel is a random variable, we need some more work to show the almost sure convergence of wk to . If PW denotes the probability distribution of matrices (W N )NN and Ph that of the channel impulse response, the probability measure we need to consider is P = P W Ph (the precoding matrices WN are chosen independently from the channel). Using Fubinis theorem, the almost sure convergence can be proved: P (wk (h) (h)) = = = = 0 1 { w

k

(h)

(h)}

k

1 { w 0 dPh

(h)

(h)}

127

Appendix E

Proof of Theorem 2

In this appendix, we briey justify theorem 2. We denote by F N the N N matrix dened by FN = HH (HN UN,K UH HH + 2 I)1 HN N N,K N FN = (UN,K UH + 2 (HN HH )1 )1 N,K N wk = w H F N w k k which can also be written as The SINR wk is given by

Using assumption A4, it is easily seen that sup N ||FN || < . As wk and FN are independent, corollary 1 of [44] implies that wk tr(FN ) converges to 0 almost surely when N . N In order to evaluate the behavior of tr(FN ) , we use the results of [93]. We rst recall that if is a N probability measure, its Stieltjes transform is the function G (z) dened by G (z) = d(t) . t z

immediately imply that the empirical eigenvalue distribution of U N,K UH + 2 (HN HH )1 converges N,K N weakly almost surely to a probability measure whose Stieltjes transform G (z) is dened by the equation G (z) = G z 1 + G (z)

Assumption A2 implies that the empirical eigenvalue distribution of (H N HH )1 converges weakly to the N 1 distribution of random variable |h|2 , i.e. the measure dened by d(t) = p(1/t) dt. The results of [93] t2

(E.1)

As moreover the eigenvalues of UN,K UH + 2 (HN HH )1 are greater than d = > 0 (recall that c is N,K N c the upperbound of the |hi |2 ), the measure is carried by [d, ]. The function t 1/t is bounded and continuous on [d, ]. The weak convergence of the empirical eigenvalue distribution of U N,K UH + N,K 2 (HN HH )1 thus implies that N 1 tr(FN ) lim = d(t) N N d t almost surely. Therefore, the SINR converges almost surely to d 1 d(t), which coincides with m (0). t Equation (??) follows directly from relation (E.1).

128

E. P ROOF

OF

T HEOREM 2

129

Appendix F

Proof of theorem 3

In this appendix, we prove theorem 3. As we assume in the ergodic case that the probability distribution of the channel fading is compactly supported, it is possible to show that the SINRs for each user at each iteration are uniformly bounded, for all N and K, and for all realization of the precoding matrix. In the non-ergodic case, the channel gains are also bounded for a given channel response. Consequently, is it sufcient to show the result for a function f that is continuous and bounded by a value M. First, we note N := K(N) , and we dene the random piecewise constant function dened on [0, max( N , )] N for an event : f (SINR( xN + 1 , N)) if x N gN (x) = 0 if x > N We can also dene g = f . Thus we have 1 K f (SINR(i, N)) = K i=1

N x=0

gN (x)dx

Because of the fact that N and f is bounded by a real value M, we have N gN (x)dx 0. In order to prove equation Fit is sufcient to show that a.s., x=0 gN (x)dx x=0 g(x)dx

For a xed x [0, ], P [ f (SINR( xN + 1 , N)) f ((x))] = 0 (consequence of the almost sure convergence of SINR( xN + 1 , N) shown in [39], and of the fact that f is continuous). This result can be written as:

g(x)}

=0

It is then possible to consider x as the realization a random variable X, independent from , with a uniform density over [0, ] ; thus we can calculate Ex E 1{gN (x)

g(x)}

dx = 0

130 The function that is integrated is positive, so Fubinis theorem can apply:

F. P ROOF

OF THEOREM

E

x=0

1{gN (x)

g(x)} dx

=0

Consequently, as the integrated term is positive, P -a.s., EX 1{gN (x,) g(x)} dx = 0. Let us consider one such event . We then have gN (x, ) g(x) a.s, and as g and gN are uniformly bounded by a positive value M, EX [gN (x)] EX [g(x)]. Indeed, we have |EX [gN (x)] EX [g(x)]| EX [|gN (x) g(x)|] 1 1 |gN (x) g(x)|dx 0 {gN (x)g(x)} + 1

0 0(dominated convergence)

1{gN (x)

i.e.

x=0

gN (x)dx

x=0

g(x)dx

131

Appendix G

Proof of theorem 4

In this appendix, we prove theorem 4. Let SIC denote the spectral efciency of the SIC MMSE algorithm and opt the spectral efciency when optimal detection is performed. SIC (2 ) = and opt (2 ) = = HN WN,K WH HH 1 N,K N log2 det IN + 2 N N lim log2 (1 + t )d( 2 )(t)

0

= (1 )() + (1) is the limiting eigenvalue distribution of W N,K WH and is the limiting N,K eigenvalue distribution of HN HH . N The following proof will show that: dSIC (2 ) dopt (2 ) = d2 d2 Since SIC () = opt () = 0, the equality between the two spectral efciencies is then immediate. First, notice that: dopt (2 ) d2 = = = = But noticing that we have

1 1 ( 2 )

t

)(t)

dopt (2 ) (, 2 ) = d2 2 log2

(z) 0 d(x,2 ) d2

G. P ROOF

OF THEOREM

1 + (x, 2 )

dx

= (x, z) and S(x, z) = Sx (z). Using free probability results of Appendix C, we have: S x

(z)

(z)

z = (x

(z))Sx (x (z))

We can rewrite the previous equation in the following form: (x and Finally, (x, z) = ( Differentiating with respect to x: x = z S(x, (x, z)) z S2 (x, (x, z)) Sx (x, ) Sz (x, ) x

(z))

z Sx (x z

(z))

(x, z) = (

Sx (x

(z))

Differentiating with respect to z: z = z S(x, (x, z) 1 z 2 S (x, ) z S(x, (x, z)) S (x, (x, z)) z

Combining the two previous equations yields: x z 1 Sz (x, ) z S(x, (x, z)) S2 (x, ( x, z)) = z z Sx (x, ) Sz (x, ) x S2 (x, (x, z))

which gives: x + zz

Sx (x,) Sx (x,)

= 0.

(G.1)

1 Denote (x, 2 ) = x(x, 2 ) = (x, 2 ). Let us derive the expression of S(x, z).

1 x+z 1+z x+z

G. P ROOF Therefore,

OF THEOREM

4

Sx (x,) S(x,)

133 =

1 x+

Sx (x,z) S(x,z)

1 x+z

and

x = x

1 2 = z . 4 1 If we replace z = 2 in equation G.1, we obtain:

x so that

1 4 2 =0 2 x 2 = x x 2 Therefore, 1 2. 2

2 1

xx 2 2

1 ,

2 we obtain: 1+ =

Therefore,

2 x = 2 1+ Finally, 1 log2

0

2 dx = 1+ =

1 x dx 2 log2 0 (, 2 ) 2 log2

134

G. P ROOF

OF THEOREM

135

Appendix H

Proof of theorem 6

We rst show that for l = 1, 2, ..., the correlation matrix l is of the form l = l IK for a scalar l asymptotically to the same value 1 . Since i 1 = 1 2Q

that depends only of l . At step one, 0 = 0 IK . It is shown in [39] that all the SINR converge obtain: 1 = 1 IK .

Assume that for a given l, l = l IK . The SINR l+1 for the ith symbol at the (l + 1)th iteration i veries il+1 (H.1) l+1 = i 1 (1 |l |2 )l+1 i with il+1 = wH HH 2 I + (1 |l |2 )HN WN,K WH HH i N N,K N =

1

HN wi

1

1 2 wH HH I + HN WN,K WH HH N,K N 1 |l |2 i N 1 |l |2

HN wi

We can then use the result shown in [39], to show that the (1 |l |2 )l+1 i (1 |l |2 )l+1

0 2

i=1,...,K

converge almost

l 2 l+1

2

p(t)dt

1/2 1/2

2

2

|h( f )|2

l 2

l+1

df

136

H. P ROOF

OF THEOREM

From equation (H.1), we deduce that the SINR for each user l+1 converge almost surely to the same i value

l+1

l+1 l+1

l+1

t t + 2 (1 )

l+1 + 1|l |2

2

p(t)dt

1/2

l+1

1/2

|h( f )|2 + 2 (1 )

|h( f )|2

l+1

+ 1|l |2

2

df

Thus we nd equations (5.26) and (5.27), and the correlation matrix at the (l + 1) th iteration is of the form l+1 = l+1 IK . By induction, the results is shown for all l 0.

137

Appendix I

Proof of proposition 3

In this appendix, we prove proposition 3. The slicer input for the i th symbol at the l th iteration can be expressed as H sl = cli HN wi si + vli

H H H (i) where vli = cli HN WK s dli sl1 + cli n is assumed to be a complex-valued, asymptotically Gaussian (as N ), zero-mean, and uncorrelated with s i , whose variance is a function of cli and dli , and where we denote (i) HN WK = HN [w1 | . . . |wi1 |0|wi+1 | . . . |wK ]

H

var(vli ) = cli HN WK (HN WK )H cli + dli dli + 2 cli cli cli HN WK l1 dli dli l1 (HN WK )H cli

(i) (i) (i) (i)

Using this model, dene the SINR during the l th pass as li (cli , dli ) = |cli HN wi |2 /var(vli ) for i = 1, 2, . . . , K. i (cli , dli )

l

Dl

The SINR i l (cli , dli ) is maximized with respect to cli and dli in a two step procedure. Since matrix does not appear in the numerator, the expression involving D l can be minimized independently and yields:

and

Dl

= l1

138

I. P ROOF

OF PROPOSITION

At each iteration, Dl expresses the fact that only non-diagonal elements (which correspond to interference) are retrieved. The interfering signals to be detected are not alterered. In this case, rewriting v li yields vli and

H H (i) (i) var(vli ) = cli HN WK var(sl1 l1 s)(HN WK )H cli + 2 cli cli H H (i) (i) = cli HN WK (I l1 l1 )(HN WK )H cli + 2 cli cli H H H H H (i) (i) = cli HN WK s cli HN WK l1 sl1 + cli n

= cli = cli

2 I + HN WK (I l1 l1 )(HN WK )H cli

(i) (i)

The SINR for the ith user is : li = = |cli HN wi |2 var(vli ) cli HN wi wH HN H cli i cli

H H H

H

H

139

Appendix J

Proof of proposition 4

We focus here on the case where = 1. If we denote a l = 1 |l |2 , then the limit SINR at each iteration is the unique solution of an equation of the form

l l l

al1 + 1

=E

t al1 t + 2

l

(J.1)

both in the ergodic and non-ergodic cases. As a l is a strictly decreasing function of , it is sufcient to show that (al )lN is a decreasing sequence, which we are going to show by induction. initialization: As 0 = 0 and 1 = MMSE > 0, we have a1 a0 recursion: Assume that al al1 for a xed l N. In order to show that a l+1 al , we are going l+1 l to prove that . From (J.1) we get E =

t al t+2 al t+2 t

l+1

t xt+2 1xE t 2 xt+

1 al E

(J.2)

. If F is a decreasing function on [0, 1], then the result follows from the

t xt+2

recursion assumption.

2 2

t xt+2

t t | x h(x,t)| = (xt+2 )2 2 = g(t). As is compactly supported, g is integrable (the existence of a second moment would be sufcient). Therefore U is derivable and

d U (x) = dx

dF dt

J. P ROOF

OF PROPOSITION

t 2 (xt+2 )2

1 xE

2

t xt+2 t xt+2

E

2

t xt+2

xE

t 2 (xt+2 )2

t xt+2

t xt+2 t xt+2

= sign var

The boundness of is easily shown from (J.2), reminding that a l [0, 1].

141

Appendix K

COFDM

Let 2 be the variance of the channel coefcients and 2 the variance of the noise. Hereafter, we will take 2 = 1. The mean probability of error has the following expression: Pp = we set: y =

u 2 0

Q(

u 1 ) u p1 eu du 2 (p 1)!

and =

1 2

2 1 p1 (y2 ) 2 e y Q( y)dy (p 1)! 0 y 1 y p1 e Q( y)dy p (p 1)! 0 y y y 1 1 )dy) ([e y p1 Q( y)] + e ((p 1)y p2 Q( y) y p1 e 2 0 p (p 1)! 2 2y 0

y [e y p1 Q( y)] = 0 0 y y 1 e ((p 1)y p2 Q( y) y p1 e 2 )dy 2 2y 0 +2 1 1 y p1 ey( 2 ) = Pp1 p1 dy (p 1)! 0 2 2y

Pp =

1 p (p 1)!

We set

x2 2

= y( +2 ) 2

Pp = Pp1 = Pp1

1 1 p1 (p 1)! 2 1 1 p1 (p 1)! 2

x2 p1 x2 ) e 2 +2

0

dx +2

p1 ( ) +2 +2

x2p2 e 2 dx

x2

K. COFDM

0

x2

x2

x2p4 e 2 dx

x2

x2p4 e

2 x2

Therefore, the mean probability of error can be expressed: Pp = Pp1 = Pp1 = Pp1 Hence, Pp = = 1 [1 2 1 [1 2

p l1 ] C213 + 2 l=2 p1 l ] C2l1 +2 l=1

(2p 3)! 1 1 p2 (p 2)! (p 1)! ( + 2) p1 2 +2 2 1 (2p 3)! p1 (p 2)!(p 1)! + 2 (2( + 2)) 1 C p1 + 2 (2( + 2)) p1 2p3

With 2 = 1, we have: 1 Pp = [1 2

p1 1 l ] C2l1 1 + 22 l=1

1 1 ( )l 1 2 1 + 2 (2( 2 + 2))

143

144

K. COFDM

145

Bibliography

[1] Araftek website. http://www.araftek.com. [2] A. Akansu, P. Duhamel, X. Lin, and M. de Courville. Orthogonal Transmultiplexers in Communication : A Review. IEEE Trans. on Signal Processing, 463(4):979995, April 1998. [3] N. Al-Dahir and J.M. Ciof. MMSE decision-feedback equalizers:nite-length results. IEEE Trans. on Information Theory, pages 961975, jul 1995. [4] S.M. Alamouti. A Simple Transmit Diversity Technique for Wireless Communications. IEEE Journal on Selected Areas in Communications, 16(8):14511458, October 1998. [5] V. Aue, G.P. Fettweis, and R. Valenzuela. A comparison of the performance of linearly equalized single carrier and coded OFDM over frequency selective fading channels using the random coding technique. In Proceedings of the Int. Conf. on Communications, May 1998. [6] Z.D. Bai and J.W. Silverstein. No Eigenvalues Outside the Support of the Limiting Spectral Distribution of Large Dimensional Sample Covariance Matrices. 26(1):316345, 1998. [7] E. R. Berlekamp. Algebraic coding theory. Aegean Park Press, revised edition, 1984. [8] C. Berrou and A. Glavieux. Near optimum error correcting coding and decoding: turbo codes. IEEE Trans. on Communications, 44(10):12611271, October 1996. [9] Ph. Biane. Free probability for probabilists. http://www.dma.ens.fr/~biane/, 2000. download by choosing "Des notes dintroduction aux probabilits libres". [10] E. Biglieri, G. Caire, G. Taricco, and E. Viterbo. How fading affects CDMA: an asymptotic analysis. JSAC wireless series, 2001. [11] E. Biglieri, G. Caire, and G. Tarico. CDMA system design through asymptotic analysis. IEEE Trans. on Communications, 48:18821896, November 2000. [12] E. Biglieri, J. Proakis, and S. Shamai(Shitz). Fading channels: Information-theoretic and communications aspects. IEEE Trans. on Information Theory, pages 26192692, October 1998. [13] E. Biglieri, G. Taricco, and A. Tulino. Performance of space-time codes for a large number of antennas. IEEE Trans. on Information Theory, submitted 2002. [14] P. Billingsley. Probability and measure. Wiley Series in Probability and Mathematical Statistics, 3nd edition, 1995.

146

BIBLIOGRAPHY

[15] J.A.C. Bingham. Multicarrier Modulation for Data Transmission: An Idea Whose Time Has Come. IEEE Communications Magazine, 28(5):514, May 1990. [16] M. Bonaccorso, K. Mhirsi, K. Maouche, and V. Buzenac. Reducing the Peak-to-Average Power Ratio in OFDM Systems. In Gretsi, pages 781784, 1999. [17] M. Bossert, A. Donder, and V. Zyablov. Improved channel estimation with decision feedback for OFDM systems. Electronics Letters, 34(11):10641065, May 1998. [18] J. Boutros. Rotated Multidimensional QAM constellations for Rayleigh fading channels. IEEE Information Theory Workshop, pages 2327, June 1996. [19] J. Boutros. Les Turbo Codes Paralleles et Series. Polycopi du cours de Telecom Paris, Tlcom Paris, 1998. [20] J. Boutros and G. Caire. Iterative Multiuser Joint Decoding: Unied Framework and Asymptotic Analysis. IEEE Trans. on Information Theory, July 2002. [21] J. Boutros and E. Viterbo. Signal space diversity: a power and bandwidth efcient diversity technique for the Rayleigh fading channel. IEEE Trans. on Information Theory, pages 1453 1467, July 1998. [22] J. Boutros, E. Viterbo, C. Rastello, and J.C. Belore. Good lattice constellations for both Rayleigh fading and Gaussian channel. IEEE Trans. on Information Theory, pages 502518, March 1996. [23] Joseph Boutros. Lattice codes for Rayleigh Fading Channels. PhD thesis, cole Nationale Suprieure desTlcommunications, June 1996. [24] E. O. Brigham. The Fast Fourier Transform. Prentice-Hall, Englewood Cliffs, New Jersey, USA, 1974. [25] C. S. Burrus. Efcient Fourier Transform and Convolution Algorithms. In J. S. Lim and A. V. Oppenheim, editors, Advanced Topics in Digital Signal Processing. Prentice-Hall, Englewood Cliffs, New Jersey, USA, 1988. [26] C. S. Burrus and T. W. Parks. DFT/FFT and Convolution Algorithms. John Wiley, New York, 1985. [27] A.R Calderbank. The art of signaling: fty years of coding theory. IEEE Trans. on Information Theory, pages 25612595, October 1998. [28] Albert M. Chan and Gregory W. Wornell. A Class of Asymptotically Optimum Iterated-decision Multiuser Detectors. In IEEE International Conference on Acoustics, Speech, and Signal Processing, Salt Lake City, Utah, May 2001. [29] R. W. Chang and R. A. Gibby. A Theorical Study of Performance of an Orthogonal Multiplexing Data Transmission Scheme. IEEE Trans. on Communications, 16:529540, August 1968. [30] J.M. Chaufray, W. Hachem, and Ph. Loubaton. Asymptotical Analysis of Optimum and SubOptimum CDMA Downlink MMSE Receivers. In IEEE International Symposium on Information Theory, Lausanne, Switzerland, July 2002. can be downloaded at http://www-syscom.univmlv.fr/~loubaton/index.html.

BIBLIOGRAPHY

147

[31] J.M. Ciof, G.P. Dudevoir, M.V. Eyuboglu, and G.D. Forney. MMSE decision feedback equalizers and coding- Part I: equalization results. IEEE Trans. on Communications, pages 25822594, October 1995. [32] J.M. Ciof, G.P. Dudevoir, M.V. Eyuboglu, and G.D. Forney. MMSE decision feedback equalizers and coding- Part II: coding results. IEEE Trans. on Communications, pages 25952604, October 1995. [33] ETSI Normalization Commitee. Channel Models for HIPERLAN/2 in different indoor scenarios. Norme ETSI, document 3ERI085B, European Telecommunications Standards Institute, SophiaAntipolis, Valbonne, France, 1998. [34] ETSI Normalization Committee. Radio Broadcasting Systems, Digital Audio Broadcasting (DAB) to mobile, portable and xed receivers. Norme ETSI, document ETS 300 401, European Telecommunications Standards Institute, Sophia-Antipolis, Valbonne, France, 19951997. [35] J. H. Conway and N. J. Sloane. Sphere packings, lattices and groups. Springer-Verlag, New York, USA, 3rd ed., 1995. [36] T. Cover and J. Thomas. Elements of Information Theory. Wiley, 1991. [37] Marc de Courville. Utilisation de bases orthogonales pour lalgorithmique adaptative et lgalisation des systmes multiporteuses. PhD thesis, cole Nationale Suprieure des Tlcommunications, October 1996. [38] M. Debbah, W. Hachem, P. Loubaton, and M. de Courville. Asymptotic Performance Analysis for Redundant Block Precoded OFDM Systems with MMSE Equalization. In 11th IEEE Statistical Signal Processing Workshop, Singapore, August 2001. [39] M. Debbah, W. Hachem, P. Loubaton, and M. de Courville. MMSE Analysis of Certain Large Isometric Random Precoded Systems. In 2001 Information Theory Workshop, Cairns, Australia, September 2001. [40] M. Debbah, P. Loubaton, and M. de Courville. Spread OFDM Performance with MMSE Equalization. In IEEE International Conference on Acoustics, Speech, and Signal Processing, Salt Lake City, USA, May 2001. [41] M. Debbah, B. Muquet, M. de Courville, M. Muck, S. Simoens, and P. Loubaton. A New MMSE Successive Interference Cancellation Scheme for Spread OFDM Systems. In Proceedings of the IEEE Vehicular Technology Conference, volume 2, pages 745749, Tokyo, Japan, May 2000. [42] Pierre Duhamel and Martin Vetterli. Fast Fourier Transforms: A Tutorial Review and a State of the Art. Signal Processing, (19):259299, 1990. [43] R. Van Nee et al. New high-rate wireless LAN standards. IEEE Communications Magazine, 40:140147, January 2002. [44] J. Evans and D.N.C Tse. Large System Performance of Linear Multiuser Receivers in Multipath Fading Channels. IEEE Trans. on Information Theory, pages 20592078, September 2000. [45] K. Fazel. Performance of CDMA/OFDM for mobile communication system. In International Conference on Universal Personal Communications, volume 2, pages 975979, 1993.

148

BIBLIOGRAPHY

[46] K. Fazel and L. Papke. On the Performance of Convolutionally-Coded CDMA/OFDM for Mobile Communication System. In proceedings of IEEE PIMRC, pages 468472, Yokohama, Japan, September 1993. [47] B. Le Floch, M. Alard, and C. Berrou. Coded orthogonal frequency division multiplex (TV broadcasting). Proceedings of the IEEE, 83(6):982996, June 1995. [48] G.D. Forney and G. Ungerboeck. Modulation and coding for linear Gaussian channels. IEEE Trans. on Information Theory, pages 23842415, October 1998. [49] G.J. Foschini and M.J. Gans. On Limits of Wireless Communications in a Fading Environment when Using Multiple Antennas. Wireless Personal Communications, 6:311335, 1998. [50] R.G. Gallager. Information Theory and Reliable Communication. Wiley, 1968. [51] X. Giraud and J.C Belore. Constellations Matched to the Rayleigh Fading Channel . IEEE Trans. on Information Theory, pages 106115, January 1996. [52] G. Golden, C. Foschini, R. Valenzuela, and P. Wolniansky. Detection algorithm and initial laboratory results using V-BLAST space-time communication architecture. Electronics Letters, 35(1):1416, January 1999. [53] G. H. Golub and C. F. Van Loan. Matrix Computations. John Hopkins University Press, 1996. [54] A. Graham. In Kronecker Products and Matrix Calculus with Applications. Ellis Horwood Limited, 1981. [55] J. Hagenauer, E. Offer, and L. Papke. Iterative decoding of binary block and convolutional codes. IEEE Trans. on Information Theory, 42(2):429445, March 1996. [56] S. Hara and R. Prasad. Overview of multicarrier CDMA. IEEE Communications Magazine, pages 126133, December 1997. [57] A. Hettich and M. Schrother. IEEE 802.11 or ETSI BRAN HIPERLAN/2: who will win the race for a high speed wireless LAN standard. In European Wireless Conference, pages 169174, October 1999. [58] F. Hiai and D. Petz. The Semicircle Law, Free Random Variables and Entropy, volume 77 of Mathematical Surveys and Monographs. AMS, 2000. [59] Roger A. Horn and Charles R. Johnson. Matrix Analysis. Cambridge University Press, 1990. [60] A.L.C. Hui and K.B. Letaief. Successive interference cancellation for multiuser asynchronous DS/CDMA detectors in multipath fading links. IEEE Trans. on Communications, 46(3):384391, March 1998. [61] I. Hwang, H. Lee, and K. Kang. Frequency and Timing Period Offset Estimation Technique for OFDM Systems. Electronics Letters, 34(6):520521, March 1998. [62] E. Jaffrot and M. Siala. Turbo channel estimation for OFDM systems on highly time and frequency selective channels. In IEEE International Conference on Acoustics, Speech, and Signal Processing, Istanbul, Turkey, June 2000.

BIBLIOGRAPHY

149

[63] S. Kaiser. Trade-Off between Channel Coding and Spreading in Multi-Carrier CDMA. In IEEE Spread Spectrum Techniques And Applications Proceedings, volume 3, pages 13661370, 4th International Symposium, 1996. [64] J. Khun-Jush, G. Malmgren, P. Schramm, and J. Torsner. Hiperlan type 2 for broadband wireless communication. Technical report, Ericsson Review No. 2, 2000. [65] J. Khun-Jush, P. Schramm, U. Wachsmann, and F. Wenger. Structure and Performance of the HIPERLAN/2 Physical Layer. In Proceedings of the IEEE VTC conference, volume 5, pages 26672671, Amsterdam, Netherlands, June 1999. [66] Catherine Lamy. Communications grande fcacit spectrale sur le canal vanouissement. PhD thesis, cole Nationale Suprieure des Tlcommunications, April 2000. [67] Jrgen Lindner. MC-CDMA and its Relation to General Multiuser/Multisubchannel Transmission Systems. In International Symposium on Spread Spectrum Techniques & Applications, pages 115 121, Mainz, Germany, September 1996. [68] V.A. Marchenko and L.A. Pastur. Distribution of eigenvalues for some sets of random matrices. Math USSR Sb., 1:457483, 1967. [69] P. Melsa, R.C. Younce, and C.E. Rohrs. Impulse response shortening for discrete multitone transceivers. IEEE Trans. on Communications, 44:16621672, December 1996. [70] Paul H. Moose. A Technique for Orthogonal Frequency Division Multiplexing Frequency Offset Correction. IEEE Trans. on Communications, 42(10):29082914, October 1994. [71] Markus Muck. Peak-to-average power reduction methods for digital radio mondiale. Technical report, Motorola Labs Paris, 2000. [72] Markus Muck. Pseudo-random prex ofdm. Technical report, Motorola Labs Paris, 2002. [73] R. Muller. A Random Matrix Model of Communication via Antenna Arrays. IEEE Trans. on Information Theory, submitted 2000. [74] R. Mller. Multiuser Receivers for Randomly Spread Signals: Fundamental Limits With and Without Decision-Feedback. IEEE Trans. on Information Theory, pages 268283, January 2001. [75] R. Muller and S. Verdu. Spectral Efciency of Low-Complexity Multiuser Detection. In IEEE International Symposium on Information Theory, page 439, Sorrento, Italy, 2000. [76] Bertrand Muquet. Novel receiver and decoding schemes for wireless OFDM systems with cyclic prex or zero-padding. PhD thesis, cole Nationale Suprieure desTlcommunications, June 2001. [77] R. Negi and J. Ciof. Pilot tone selection for channel estimation in a mobile OFDM system. IEEE Trans. on Consumer Electronics, 1998. [78] A. V. Oppenheim and R. W. Schafer. Discrete-Time Signal Processing. Prentice-Hall, Englewood Cliffs, New Jersey, USA, 1989. [79] T. Pollet, M. van Bladel, and M. Moeneclaey. BER sensitivity of OFDM systems to carrier frequency offset and Wiener phase noise. IEEE Trans. on Communications, 43:191193, February 1995.

150

BIBLIOGRAPHY

[80] H. Poor and S. Verd. Probability of Error in MMSE Multiuser Detection. IEEE Trans. on Information Theory, 43(3):858871, May 1997. [81] J. G. Proakis. Digital Communications. Mc Graw Hill, New York, USA, 3rd ed., 1995. [82] H. Rohling, K. Brueninghaus, and T. Mueller. Performances of Coherent OFDM-CDMA for Broadband Mobile Communications. In RACE Mobile Telecommunications Summit, pages 263 269, Cascais, Portugal, November 1995. [83] H. Rohling, K. Brninghaus, and R. Grnheid. Comparison of Multiple Access Schemes for an OFDM Downlink System. In K. Fazel and G. P. Fettweis, editors, Multi-Carrier Spread Spectrum, pages 2330. Kluwer Academix Publishers, Norwell, MA 02061, USA, 1997. [84] W. Rudin. Fourier Analysis on Groups. Wiley Classics Library. Wiley, 1990. [85] B. R. Saltzberg. Performance of an Efcient Parallel Data Transmission System. IEEE Trans. on Communications, 15(6):805811, December 1967. [86] H. Sari, G. Karam, and I. Jeanclaude. Transmission Techniques for Digital Terrestrial TV Broadcasting. IEEE Communications Magazine, pages 100109, February 1995. [87] A. Scaglione, G.B. Giannakis, and S. Barbarossa. Redundant lterbank precoders and equalizers - Part I: unication and optimal designs. IEEE Trans. on Signal Processing, 47:19882006, July 1999. [88] A. Scaglione, G.B. Giannakis, and S. Barbarossa. Redundant lterbank precoders and equalizers - Part II: blind channel estimation, synchronization and direct equalization. IEEE Trans. on Signal Processing, 47:20072022, July 1999. [89] P. Schramm and R. Muller. Spectral Efciency of CDMA Systems with Linear MMSE interference. IEEE Trans. on Communications, 47:722731, May 1999. [90] S. Shamai and S. Verdu. The Impact of Frequency-Flat Fading on the Spectral Efciency of CDMA. IEEE Trans. on Information Theory, pages 13021326, May 2001. [91] C. E. Shannon. A mathematical theory of communication. The Bell Labs Technical Journal, pages 379457,623,656, July-October, vol. 27 1948. [92] C. E. Shannon. Probability of error for optimal codes in a Gaussian channel. The Bell Labs Technical Journal, May, vol. 38 1959. [93] J.W. Silverstein and Z.D. Bai. On the empirical distribution of eigenvalues of a class of large dimensional random matrices. J. Multivariate Anal., 54(2):175192, 1995. [94] Y Sun. Bandwidth-efcient wireless OFDM. In Selected Areas in Communications, IEEE Journal on, volume 11, pages 22672278, Sorrento, Italy, 2001. [95] T. Tanaka. Density evolution for multistage CDMA multiuser detector. In IEEE International Symposium on Information Theory, Lausanne, Switzerland, July 2002. [96] V. Tarokh, H. Jafarkhani, and A.R. Calderbank. Space-Time Block Codes from Orthogonal Designs. IEEE Trans. on Information Theory, 45(5):14561467, July 1999.

BIBLIOGRAPHY

151

[97] V. Tarokh, H. Jafarkhani, and A.R. Calderbank. Space-Time Block Coding for Wireless Communications: Performance Results. IEEE Journal on Selected Areas in Communications, 17(3):451 460, March 1999. [98] V. Tarokh, N. Seshadri, and A.R. Calderbank. Space-Time Codes for High Data Rate Wireless Communication: Performance Criterion and Code Construction. IEEE Trans. on Information Theory, 44(2):744765, March 1998. [99] W.G. Teich, A. Bury, J. Egle, M. Nold, and J. Lindner. Iterative detection algorithms for extended MC-CDMA. In IEEE Spread Spectrum Techniques And Applications Proceedings, volume 1, pages 184188, 5th International Symposium, 1998. [100] I.E. Telatar. Capacity of Multi-antenna Gaussian Channels. Technical report, AT & T Bell Labs, 1995. [101] J. Tellado-Mourelo. Peak to average power reduction for multicarrier modulation. PhD thesis, Standford University, September 1999. [102] D.N.C Tse and S. Hanly. Linear Multi-user Receiver: Effective Interference, Effective bandwidth and User Capacity. IEEE Trans. on Information Theory, pages 641657, March 1999. [103] U. Tureli and H. Liu. Blind carrier synchronization and channel identication for OFDM communications. In IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 6, pages 35093512, Seattle, USA, May 1998. [104] G. Ungerboeck. Treillis Coded Modulation with Redundant Signals Sets, Part. I: Introduction. Part. II: State of the Art. IEEE Trans. on Communications, 25:521, February 1987. [105] J. van de Beek, O. Edfors, M. Sandell, S. Wilson, and P. Borjesson. On Channel Estimation in OFDM Systems. In Proceedings of the IEEE Vehicular Technology Conference, volume 2, pages 815819, Chicago, USA, July 1995. [106] R. van Nee. A new OFDM standard for high rate wireless LAN in the 5 GHz band. Proceedings of the IEEE Vehicular Technology Conference, pages 258262, 1999. [107] R. van Nee and A. de Wilde. Reducing the Peak to Average Power Ratio of OFDM. In Proceedings of the IEEE Vehicular Technology Conference, pages 20722076, 1998. [108] R. van Nee and R. Prasad. OFDM for wireless multilmedia communications. Artech House Publishers, Boston, USA, 2000. [109] L. Vandendorpe. MIMO DFE Equalization for Multitone DS/SS Systems over Multipath Channels. IEEE Journal on Selected Areas in Communications, 14(3):110, April 1996. [110] M. Varanasi and T. Guess. Achieving vertices of the capacity region of the synchronous correlatedwaveform multiple-access channel with decision-feedback receivers. In IEEE International Symposium on Information Theory, page 270, 1997. [111] S. Verdu. Fifty years of Shannon theory. IEEE Trans. on Information Theory, pages 20572078, October 1998. [112] S. Verd. Multiuser Detection. Cambridge University Press, 1998.

152

BIBLIOGRAPHY

[113] S. Verdu and S. Shamai. Spectral Efciency of CDMA with Random Spreading . IEEE Trans. on Information Theory, pages 622640, March 1999. [114] A. Viterbi. Error Bounds for Convolutional Coders and Asymptoticaly Optimum Decoding Algorithm. IEEE Trans. on Information Theory, 13:5567, April 1967. [115] D.V. Voiculescu. Multiplication of certain non-commuting random variables. J. Operator Theory, 18:223235, 1987. [116] D.V. Voiculescu. Limit laws for random matrices and free products. Inventiones mathematicae, 104:201220, 1991. [117] D.V. Voiculescu, K.J. Dykema, and A. Nica. Free Random Variables. American Mathematical Society, CRM Monograph Series, Volume 1, Providence, Rhode Island, USA, 1992. [118] Z. Wang and G.B. Giannakis. Wireless multicarrier communications: where Fourier meets Shannon. IEEE Signal Processing Magazine, 17:2948, May 2000. [119] Z. Wang and G.B. Giannakis. Linearly Precoded or Coded OFDM against Wireless Channel Fades ? In Proceedings of the 3rd Workshop on Signal Processing Advances in Wireless Communications, pages 267270, March 2001. [120] S. B. Weinstein and P. M. Ebert. Data Transmission by Frequency-Division Multiplexing using the Discrete Fourier Transform. IEEE Trans. on Communications, 19(5):628634, October 1971. [121] P.W. Wolniansky, G.J. Foschini, G.D. Golden, and R.A. Valenzuela. V-BLAST: An Architecture for Realizing Very High Data Rates Over the Rich-Scattering Wireless Channel. In International Symposium on Signals, Systems, and Electronics, volume 4, pages 295300, 1998. [122] G.W Wornell. Spread-Response Precoding for Communications over Fading Channels. IEEE Trans. on Information Theory, 42:488501, March 1996. [123] Y.Algayon and T.Brunck. Principes du COFDM. Technical report, ENST Site de Toulouse, February 1998. [124] N. Yee, J.P. Linnartz, and G. Fettweis. MultiCarrier CDMA in Indoor Wireless Radio Networks. In proceedings of IEEE PIMRC, pages 109113, Yokohama, Japan, September 1993. [125] N. Yee and J.P.M.G. Linnartz. Wiener ltering for multi-carrier cdma. In proceedings of IEEE PIMRC, volume 4, pages 13441347, The Hague, September 1994. [126] E. Zehavi. 8-PSK Trellis Codes for a Rayleigh Channel. IEEE Trans. on Communications, 40:873884, May 1992. [127] J. Zhang, E. Chong, and D. Tse. Output MAI Distributions of Linear MMSE Multiuser Receivers in CDMA Systems. IEEE Trans. on Information Theory, pages 11281144, March 2001. [128] W. Zou and W. Yiyan. COFDM: an overview. IEEE Trans. on Broadcasting, 41:18, March 1995.