Anda di halaman 1dari 174

Technische Universit t Dresden a

On Spectrally-Bounded Codes for Multicarrier Communications

Kai-Uwe Schmidt

der Fakult t Elektrotechnik und Informationstechnik a der Technischen Universit t Dresden a zur Erlangung des akademischen Grades eines

Doktoringenieurs (Dr.-Ing.)
genehmigte Dissertation

Vorsitzender: Prof. Dr.-Ing. Gerhard Fettweis Gutachter: Prof. Dr.-Ing. habil. Adolf Finger Prof. Dr.-Ing. Martin Bossert Prof. Kenneth G. Paterson

Tag der Einreichung: 29. 09. 2006 Tag der Verteidigung: 12. 04. 2007

Acknowledgements
First, I wish to express my thanks to Prof. Adolf Finger for his support of my research and for providing excellent working conditions during the past years at Dresden University of Technology. I wish to thank him as well for writing one of the three reviews of this thesis. I would also like to thank Prof. Martin Bossert from the University of Ulm and Prof. Kenny Paterson from Royal Holloway/University of London for writing the other reviews. I am especially grateful to Prof. Kenny Paterson for helpful discussions during the preparation of this thesis. I wish to thank all the former and present colleagues and friends in the Communications Theory Group for providing a pleasant working atmosphere. I always enjoyed working here. Last but not least, I am thankful to my family, particularly to my mother, who paved the way for me to get where I am today. My special thanks go to Anke for her patience and constant support. I owe her much of the time that I spent for the preparation of this thesis.

Kai-Uwe Schmidt Dresden, April 2007

ii

To Leonard, who was born just before the completion of this thesis.

Kurzfassung
Das Sendesignal in einem Mehrtr gersystem entsteht durch eine orthogonale a Transformation eines Datenworts. Es folgt daraus, dass die maximale Momentanleistung des Sendesignals um ein Vielfaches h her sein kann als die mittlere Seno deleistung. Dies stellt einen hohen Anspruch an die Implementierung des Systems, was als entscheidender Nachteil von Mehrtr gersystemen betrachtet wird. a Das Verh ltnis von maximaler und mittlerer Sendeleistung (der Spitzenwert) l sst a a sich allerdings durch gezielte Ausnutzung der Redundanz eines Fehlerschutzcodes reduzieren. Die vorliegende Dissertation besch ftigt sich mit der Konstruktion sola cher Codes f r zwei praxisrelevante Mehrtr gersysteme: Orthogonales Frequenzu a multiplex (OFDM) und Multicode Codemultiplex (MC-CDMA). Der erste Teil der Arbeit beinhaltet eine Analyse der maximalen Spitzenwertreduktion in einem OFDM System, die durch Drehung der Koordinaten eines bekannten bin ren Codes erreicht werden kann. Anhand der erzielten Ergebnisse a wird gezeigt, dass aus der Literatur bekannte suboptimale Konzepte zur Bestimmung der Drehwinkel den Spitzenwert oftmals bis nahe an die theoretische Grenze reduzieren. In einem weiteren Teil werden unter Benutzung der spektralen Eigenschaften von komplement ren Paaren und deren Verallgemeinerungen, wie fasta komplement re Paare und komplement re Mengen, obere und untere Schranken a a f r den maximalen Spitzenwert von bestimmten Nebenklassen eines verallgeu meinerten ReedMuller Codes erster Ordnung hergeleitet. Durch Vereinigungen solcher Nebenklassen lassen sich Codes f r OFDM und MC-CDMA mit Fehleru korrektureigenschaften und strikt begrenztem Spitzenwert konstruieren. Oftmals erh lt man dabei Codes, deren Parameter besser sind als die bisher bekannter Coa des. Ferner liefern die erzielten Ergebnisse theoretische Erkl rungen f r einige a u Vermutungen und experimentelle Beobachtungen in der Literatur. Schlielich werden so genannte verallgemeinerte Bent-Funktionen in Verbindung mit algebraischen Codes uber Z4 (verallgemeinerte ReedMuller, Ker dock und DelsarteGoethals Codes) genutzt, um nichtbin re Fehlerschutzcodes a f r MC-CDMA zu konstruieren. Diese Codes reduzieren den Spitzenwert auf den u kleinstm glichen Wert und erg nzen vergleichbare bekannte bin re Codes. o a a iv

Abstract
In multicarrier communications the transmitted signal is obtained by applying an orthogonal transform to a block of data symbols. As a consequence, the peak power of the transmitted signal can be much larger than its mean power, which makes the practical implementation of the system a challenging task. The redundancy of carefully designed error-correcting codes can be exploited to control the ratio of the peak and the mean power of the transmitted signal. In this thesis such codes are studied for two types of multicarrier communications systems: orthogonal frequency-division multiplexing (OFDM) and multicode code-division multiple access (MC-CDMA). In a rst part of this thesis a well-known approach for the construction of such codes for OFDM is examined. This concept involves a rotation of the coordinates of a known binary code by phase shifts that are independent of the individual codewords. Bounds for the maximum reduction of the peak-to-mean power ratio are proved, and it is shown that most phase-shift designs in the literature, obtained using suboptimal optimization methods, produce peak-to-mean power ratio reductions that are close to the theoretical limit. The spectral properties of complementary pairs and generalizations, namely near-complementary pairs and complementary sets, are then exploited to prove upper and lower bounds on the peak-to-mean power ratio of certain cosets of a rst-order generalized ReedMuller code. By taking unions of such cosets inside higher-order generalized ReedMuller codes, error-correcting codes for OFDM and MC-CDMA with strictly bounded peak-to-mean power ratios are constructed. These codes complement and, in many situations, improve existing coding schemes. Moreover these results provide theoretical support for conjectures and empirical observations in the literature. Finally generalized bent functions in connection with algebraic codes over Z 4 , namely generalized ReedMuller, Kerdock, and DelsarteGoethals codes, are exploited to construct families of so-called constant-amplitude error-correcting codes for MC-CDMA, which are codes that reduce the peak-to-mean power ratio to the least possible value. These codes complement previously proposed binary constant-amplitude codes. v

vi

Contents
Acknowledgements Kurzfassung . . . . Abstract . . . . . . Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i iv v vii

Chapter 1. Introduction 1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Contribution and Outline . . . . . . . . . . . . . . . . . . . . .

1 1 3 5

Chapter 2. Background 2.1 General Notation and Basic Denitions . . . . . . 2.2 Galois Rings and Fields . . . . . . . . . . . . . . 2.3 Generalized Boolean Functions . . . . . . . . . . 2.4 Error-Correcting Codes . . . . . . . . . . . . . . 2.5 Aperiodic Correlation Functions . . . . . . . . . 2.6 Multicarrier Communications . . . . . . . . . . . 2.7 The Power-Control Problem in OFDM . . . . . . 2.8 The Power-Control Problem in Multicode CDMA

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

9 9 10 13 17 21 22 24 25

Chapter 3. Phase-Shifted Codes 3.1 Introduction and Chapter Overview . . 3.2 Denitions and Problem Specication 3.3 A Lower Bound on the PMEPR . . . . 3.4 Applications . . . . . . . . . . . . . . 3.5 Conclusions and Open Problems . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

29 29 30 31 37 45

vii

Chapter 4. Near-Complementary Pairs and ReedMuller Codes 4.1 Introduction and Chapter Overview . . . . . . . . . . . . 4.2 Generalized ReedMuller Codes . . . . . . . . . . . . . 4.3 Complementarity of Sequence Pairs . . . . . . . . . . . . 4.4 Constructions of Near-Complementary Pairs . . . . . . . 4.5 Near-Complementary Pairs and Cosets of RMq (1, m) . . 4.6 Lower Bounds on the PMEPR . . . . . . . . . . . . . . . 4.7 Applications and Construction Examples . . . . . . . . . 4.8 Decoding of Unions of Cosets of RMq (1, m) . . . . . . . 4.9 Conclusions, Notes, and Open Problems . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

47 47 48 50 53 59 66 74 80 87

Chapter 5. Complementary Sets and ReedMuller Codes 5.1 Introduction and Chapter Overview . . . . . . . . 5.2 Further Notation and Denitions . . . . . . . . . 5.3 Construction of Complementary Sets . . . . . . . 5.4 Lower Bounds on the PMEPR . . . . . . . . . . . 5.5 The Effective-Degree ReedMuller Code . . . . . 5.6 OFDM Codes with Low PMEPR . . . . . . . . . 5.7 Conclusions and Open Problems . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

91 91 92 95 102 108 110 116

Chapter 6. Codes for Multicode CDMA 6.1 Introduction and Chapter Overview . . . 6.2 Fourier Transforms and Bent Functions . 6.3 Cosets of RMq (1, m) . . . . . . . . . . 6.4 Some Families of Trace Codes . . . . . 6.5 Z4 -valued Bent Functions . . . . . . . . 6.6 Quaternary Constant-Amplitude Codes . 6.7 Conclusions, Notes, and Open Problems

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

119 119 120 122 125 131 139 145

Chapter 7. Summary

149

Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

viii

Chapter 1

Introduction

1.1 Motivation Orthogonal frequency-division multiplexing (OFDM) represents a key concept in the development of wired and wireless communications systems in the past decade [2]. It provides excellent ability to cope with multipath propagation and fast-moving environment. OFDM has been proposed and standardized for many wireless applications such as wireless local area networks (WLAN), digital audio and video broadcasting (DAB/DVB), as well as for wired applications such as in digital subscriber line (DSL) systems. Code-division multiple access (CDMA) has been standardized for third generation (3G) mobile telephone systems such as the universal mobile telecommunication system (UMTS) and CDMA2000. Multicode CDMA (MC-CDMA) is a simple concept, which is backwards-compatible to conventional CDMA. It was proposed in order to satisfy the temporary demand for higher data rates of individual users in a CDMA system [45]. Both concepts, OFDM as well as MC-CDMA, are special cases of multicarrier communications systems. The latter is a technique for data transmission where the transmitted data is divided into blocks (codewords), which are used to modulate a set of pairwise orthogonal waveforms (carriers). The sum of these modulated carriers is the transmitted signal. Mathematically, this signal is obtained by applying an orthogonal transform to the transmitted codeword, so the transmitted signal may be viewed as the spectrum of the codeword. Therefore, at the receiving side, the data can be recovered by applying the corresponding inverse transform. In OFDM the transform is the Fourier transform. MC-CDMA essentially employs a Hadamard transform. 1

1 Introduction
The principal disadvantage of all kinds of multicarrier systems is that some codewords exhibit a high peak-to-mean power ratio (we shall use the terms peakto-mean envelope power ratio (PMEPR) and peak-to-average power ratio (PAPR) in OFDM and in MC-CDMA, respectively). That is, the peak transmit power can be many times the average transmit power, which happens when the modulated carriers line up at some time instant. In other words, high peaks occur in the spectrum (with respect to the transform employed in the system) of certain codewords. Thus, in order to ensure a distortionless transmission, all components in the transmission chain must be linear across a wide range of signal levels. This makes the transmitter considerably more expensive than that in a single-carrier system. Moreover, most of the time, the components in the transmitter are operated at levels much below their maximum input level, which results in power inefciency. The latter issue is particularly acute in mobile applications, where battery lifetime is a major concern. On the other hand, nonlinearities in the transmission chain may lead to a loss of orthogonality among the carriers and to out-of-band radiation. The former has the effect of degrading the total system performance and the latter is subject to strong regulations. Clipping and ltering of the signal before digital-to-analog conversion has been shown to alleviate the problem [55], however, performance degradation and out-of-band radiation (or at least one of them) remains an issue. This provides clear motivation to reduce the peak-to-mean power ratio in multicarrier systems systematically. This power-control problem can be solved by introducing redundancy such that the transmission of codewords with high peakto-mean power ratios is avoided. Thus the problem is shifted to identifying those unwanted codewords. Probabilistic methods have received a great amount of attention. Such an approach involves a multiple representation of the transmitted information. An algorithm then selects a favorable representation, which is used as the transmitted codeword. Popular variants of this technique are selected mapping and partial transmit sequences [66]. While in most situations the peak-to-mean power ratio is reduced, the main drawback of this method is the requirement of complex on-line calculations in the transmitter. Moreover the transmission of side information is usually necessary. A more sophisticated technique was introduced in [50] and further developed in [117]. The idea here is to draw the codewords from a special block code, which is designed off-line such that the peak-to-mean power ratio of all codewords is 2

1.2 Related Work


not exceeding a certain threshold. This technique becomes particularly elegant if the inherent redundancy can be exploited for error-correction [48]. Thus this approach combines error-correction coding and the reduction of the peak-to-mean power ratio in multicarrier systems. The goal of this thesis is the analysis and construction of such codes for two popular multicarrier systems: OFDM and MCCDMA. 1.2 Related Work The problem statements of constructing error-correcting codes with low peak-tomean power ratio in [50], [117], and [48] has initiated a great deal of research. Most of the research was focused on codes for OFDM, some on codes for MCCDMA. There exists a number of approaches to design such codes. In what follows we review them briey. A simple approach, proposed by Jones and Wilkinson in [48], for designing OFDM codes with low PMEPR is to offset a well-understood code. That means the coordinates of a given code are rotated by xed phase shifts, which are selected such that the maximum PMEPR, taken over all codewords, is minimized. Based on learning techniques, an algorithm to nd such phase shifts was presented in [48]. This algorithm works, however, only for very short codes. The approach became much more attractive when Tarokh and Jafarkhani [106] proposed a computationally feasible algorithm to nd good phase shifts, provided that an efcient maximum-likelihood decoding algorithm for the original code is known. Some complexity reductions of this algorithm have been reported by Wunder and Boche [119]. However there is no guarantee that the algorithm nds the optimum phase shifts, neither does this approach imply anything about the achievable PMEPR reduction. Paterson and Tarokh [83] proposed to use so-called trace codes in OFDM. These are linear codes and include the duals of BCH codes, the quaternary versions of Kerdock and DelsarteGoethals codes, as well as the weighted-degree trace code (which is the most general trace code). Using bounds on hybrid character sums, it was shown that, when the zero codeword is excluded, the PMEPR of trace codes is of the order (log n)2 , where n is the length of the code. In [102] Sol and Zinoviev e provided a slightly more general treatment of this issue and corrected some minor mistakes in [83]. Although the PMEPR of these codes is rather high for practical n, asymptotically the PMEPR is much lower than n, which is the worst-case PMEPR 3

1 Introduction
for uncoded transmission. A number of contributions [117], [111], [72] proposed to use Golay sequences [30] in OFDM. A Golay sequence is a member of a Golay complementary pair, which has the property that the aperiodic autocorrelation sidelobes of the two sequences in the pair sum to zero. This property implies that the PMEPR of any Golay sequence is at most 2 [84]. The work on Golay sequences for OFDM culminated in the paper by Davis and Jedwab [22], which revealed a striking connection between Golay sequences and ReedMuller codes. More specically, it was shown that a family of binary Golay sequences of length 2m organizes in m!/2 cosets of RM2 (1, m) inside RM2 (2, m), where RM2 (r, m) is the ReedMuller code of order r and length 2m [60, Chapter 13]. Similarly for h > 1 [22] identies m!/2 cosets of RM2h (1, m) comprised of polyphase Golay sequences inside ZRM2h (2, m). Here RMq (r, m) and ZRMq (r, m) are generalizations of the classical ReedMuller code over q-ary alphabets. Consequently [22] suggests OFDM coding schemes for 2h -ary phase-shift keying constellations with large minimum distance and PMEPR bounded by 2. The major drawback of this result is that the rate of the codes rapidly tends to zero as the number of carriers increases. Therefore Davis and Jedwab [22] proposed to include further second-order cosets of RM2h (1, m). This was motivated by the observation that second-order cosets of RM2h (1, m) seem to organize naturally in classes with the same PMEPR upper bound. For 16 carriers [22] contains a computationally obtained ranking of second-order cosets according to their PMEPR, which suggests trade-offs between the code rate and the PMEPR without compromising the minimum distance. However the complexity of such a coset ranking becomes unacceptably high for larger number of carriers. Further theoretical advance was made by Paterson [79] by showing that any coset of RMq (1, m) inside RMq (2, m) can be decomposed into complementary sets [110] of the same size. By generalizing the argument in [84], Paterson [79] established that any sequence lying in a complementary set of size N has PMEPR at most N , which yields an upper bound on the PMEPR of any coset of RMq (1, m) inside RMq (2, m). This explains much, though not all, of the empirical observations in [22]. The results in [79] led to OFDM coding schemes with PMEPR bounded by 2k , where k is a positive integer. The connection of cosets of RMq (1, m) and complementary sets inspired further studies in [105], [78], and [16], however, without explicitly deriving OFDM coding schemes. Up to now the construction of codes for MC-CDMA with low PAPR was limited 4

1.3 Contribution and Outline


to binary codes. Research was particularly focused on the construction of constantamplitude codes, which are codes that reduce the PAPR to least possible value 1. Wada et al. [114] constructed the largest possible binary constant-amplitude codes of lengths 4 and 16. The minimum distance of these codes was then analyzed by the same authors in [115]. Ottoson [74] proposed a simple algorithm to identify codewords with high PAPR, which are removed from the set of binary n-tuples to obtain a code of length n with low PAPR. A second algorithm then removes further codewords from this code such that the additional redundancy can be exploited for error correction. However this scheme is only useful for small n. A coding-theoretic framework for binary codes in MC-CDMA has been established by Paterson in [81]. It was shown that codewords with low PAPR are exactly those words that are far from RM2 (1, m), the rst-order ReedMuller code. Based on this fact, fundamental trade-offs between PAPR, minimum distance, and code rate of binary codes have been proved in [81]. Moreover it was shown that codewords that are farthest away from RM2 (1, m) (words corresponding to bent functions [87], [60, Chapter 14]) have PAPR equal to 1. Though rst recognized by Wada [113], this connection has been exploited in [81] to construct several families of binary constant-amplitude codes of length 22m , where m is a positive integer. These codes are unions of cosets of RM2 (1, 2m) inside higher-order ReedMuller, Kerdock, or DelsarteGoethals codes [60, Chapter 15]. Therefore they simultaneously enjoy the lowest possible PAPR and high minimum distance. Reference [81] also proposes several code families with PAPR greater than 1, particularly for lengths 22m+1 (m is a positive integer) for which binary constant-amplitude codes cannot exist. Recently, Sol and Zinoviev [103] constructed binary codes with PAPR much e greater than 1. However in many situations their parameters beat the Gilbert Varshamov-style lower bound derived in [81]. 1.3 Contribution and Outline In the next chapter we will introduce some useful notation, present preliminary results, and state the power-control problems in OFDM and MC-CDMA systems explicitly. Chapter 3 contains an analysis of the PMEPR of phase-shifted codes. Motivated by previous designs of phase shifts in [48], [106], and [119] using suboptimal methods, we ask: given a binary code, how much PMEPR reduction can be 5

1 Introduction
obtained when the phase shifts are taken from a 2h -ary phase-shift keying constellation? We will establish a lower bound on the PMEPR, which is related to the covering radius of the binary code. The general case, where the phase shifts can be arbitrary, is then recovered by analyzing this bound for h . Informally, our bound states that the smaller the covering radius of the binary code, the less PMEPR reduction is possible. We will then apply the bound to some well understood codes, including BCH codes and their duals, ReedMuller codes, and convolutional codes, and compare the results with some phase-shift designs from the literature. The results of Chapter 3 are accepted for publication in [93]. The main contribution of Chapter 4 is a construction of cosets of RM q (1, m) comprising sequence pairs that exhibit a near-complementary property. We prove an upper bound on the PMEPR of these cosets and establish the tightness of this bound in certain cases. As a corollary we recover the cosets from [22] that contain Golay complementary pairs. More importantly, we identify many new classes of cosets of RMq (1, m) with low PMEPR, showing that several previously unexplained phenomena, observed, e.g., in [22], can now be understood as part of a general framework. These results lead to a number of OFDM coding schemes whose PMEPRs are bounded by values between 2 and 4. We also study maximumlikelihood decoding algorithms for RMq (1, m), which can be used as well to perform maximum-likelihood decoding for any union of cosets of RM q (1, m). The results on the construction of cosets of RMq (1, m) are published in [90], special cases are also contained in [95]. The material about maximum-likelihood decoding algorithms is published in [94] and [89]. In Chapter 5 we continue the study of cosets of RMq (1, m) by relating such cosets with complementary sets. This chapter essentially contains generalizations of the results obtained by Paterson in [79]. We show that any coset of RM q (1, m) comprises sequences lying in complementary sets of size 2k , where k is a positive integer that can be easily determined from the generalized Boolean function associated with a word in the coset. Consequently we obtain an upper bound on the PMEPR of any coset of RMq (1, m). Although the bound tends to be loose for large k, the bound is shown to be tight in many situations, especially when k {1, 2}. We then combine these results with nonbinary generalizations of the ReedMuller code to construct OFDM coding schemes with low PMEPR. When directly possible, we compare our codes with existing schemes in [79], and show that in most situations our codes are improvements. The key results of this chapter are published in [92]. 6

1.3 Contribution and Outline


In Chapter 6 we construct codes for MC-CDMA, where we focus especially on practically important nonbinary codes. We prove that the PAPR of any coset of RMq (1, m) is at most its PMEPR, which implies that the OFDM coding schemes developed in Chapters 4 and 5 are also suited to bound the PAPR in MC-CDMA systems. We also show that it is impossible to use this approach to construct constant-amplitude codes when q 0 (mod 4), which includes the practical cases of 2h -ary phase-shift keying modulation when h > 1. We will then prove an upper bound for the PAPR of appropriately expurgated trace codes, in particular of the weighted-degree trace code, which has been previously proposed to limit the PMEPR in OFDM systems. Another principal goal of Chapter 6 is the construction of quaternary constant-amplitude codes. It is shown that such a code must be comprised of words corresponding to a kind of generalized bent functions. We provide a number of constructions of such functions and connect them with algebraic codes over Z4 , in particular with the quaternary ReedMuller codes RM4 (r, m) and ZRM4 (r, m) as well as with the quaternary Kerdock and DelsarteGoethals codes, proposed in [36]. These coding schemes may be viewed as quaternary analogs of the binary codes developed in [81]. We also present mappings from binary to quaternary constant-amplitude codes. The results on quaternary constantamplitude codes are submitted for publication in [91]. Chapter 7 contains a summary of the results in this thesis.

1 Introduction

Chapter 2

Background

2.1 General Notation and Basic Denitions A sequence or word of length n over a set S (the alphabet) is an n-tuple over S, i.e., an element in S n . By (a | b) we denote the concatenation of the sequences a and b. Let Zq be the set of representatives of the integers modulo q, e.g., Z q := {0, 1, . . . , q 1}. Then Zq together with addition and multiplication modulo q is a ring and, if q is prime, Zq is a prime eld denoted by Fq . By q we denote a primitive qth root of unity in C, e.g., q = exp(j2/q), where j is an imaginary unit such that j 2 = 1.1 With each word c = (c0 , c1 , . . . , cn1 ) Zn we assoq c ciate another word c = (0 , c1 , . . . , cn1 ), where ci = q i . Such a word is called c a polyphase word or polyphase sequence. We dene the weight of a word c Zn to be q
n1

wt(c) :=
i=0

min{ci , q ci }.

The weight induces a distance on Zn : the distance or metric between a, b Zn is q q dened as d(a, b) := wt(a b). The measures wt(.) and d(., .) are also called Lee weight and Lee distance, respectively. In the special case when q = 2 the respective quantities are equal to the Hamming weight and the Hamming distance.
1 Sometimes

we shall use j as an index variable; the context will make clear when j is distinct from an imaginary unit.

2 Background
The space Zn together with d(., .) is a metric space. From the space Znh to q 2 2h1 n the space Z2 there exists an isometry, i.e., a distance and weight-preserving injection, which we dene below.
2 Denition 2.1. [13] For any integer h > 1 we dene the map : Z2h Z2 as follows. If 0 a 2h1 , the image (a) is the binary word (000 111) with weight equal to a. If 2h1 < a < 2h , the image (a) is the binary word (111 000) with weight equal to 2h a. We also dene the maps , : Z2h h2 Z2 such that for each a Z2h 2
h1

(a) = ((a), (a)). Finally the maps , , and are extended in the obvious way to act on words in Znh . We will call the generalized Gray map (the classical Gray map is recovered 2 when h = 2 [36]). The denition of implies the following. Theorem 2.2. The map is distance and weight preserving, i.e., for any a, b Znh we have 2 wt(a) = wt((a)) d(a, b) = d((a), (b)). 2.2 Galois Rings and Fields We briey provide some facts about Galois rings and associated elds, for details we refer to the books by Wan [116] and McDonald [59] and the papers by Nechaev [68], Boztas et al. [5], and Hammons et al. [36]. For the purpose of this thesis it is sufcient to consider Galois rings of characteristic equal to a power of two. We dene : Z2h F2 to be the modulo-2 reduction map, i.e., if a Z2h h1 is written as a = i=0 ai 2i , where ai {0, 1}, then (a) = a0 . This map is extended to act on polynomials such that : Z2h [x] F2 [x] is given by a0 + a1 x + a2 x2 + + an xn (a0 ) + (a1 )x + (a2 )x2 + + (an )xn . A polynomial p(x) Z2h [x] is called a monic basic irreducible if p(x) is monic and its projection (p(x)) is irreducible over the prime eld F2 . 10

2.2 Galois Rings and Fields


Let be a root of a monic basic irreducible p(x) Z2h [x] of degree m. By denition, the Galois ring Rh,m is obtained by adjoining to Z2h , i.e., Rh,m = Z2h []. Alternatively, Rh,m may be viewed as a ring that is isomorphic to the factor ring Z2h [x]/p(x). It follows that the Galois ring Rh,m has 2hm elements. The characteristic of Rh,m is equal to 2h . Up to isomorphisms Rh,m is uniquely determined by h and m, and it contains the subring Rh,n if and only if n is a divisor of m. Simple examples of Galois rings are Rh,1 = Z2h and R1,m = F2m , the nite eld with 2m elements. Suppose f (x) F2 [x] is a primitive polynomial of degree m. Then there exists a unique monic polynomial p(x) Z2h [x] such that (p(x)) = f (x) and p(x) m divides x2 1 1 over Z2h . This polynomial is the Hensel lift of f (x) and called a monic basic primitive. Of course p(x) is also a monic basic irreducible. The polynomial p(x) can be constructed using the fact that the roots of p(x) are the 2h1 th powers of the roots of f (x) [10]. When h = 2, Graeffes method [116, Theorem 13.12] is a convenient way to construct p(x) and works as follows. Suppose f (x) = e(x) d(x), where e(x) contains only terms of even degree and d(x) contains only terms of odd degree. Then p(x) is given by p(x2 ) = (1)m [e2 (x) d2 (x)]. A root Rh,m of a monic basic primitive in Z2h [x] of degree m generates a cyclic group Th,m of order 2m 1. The set
Th,m := {0} Th,m

is called the Teichmuller set of Rh,m . Since () is a root of a primitive polynomial of degree m in F2 [x], () is a primitive element in F2m . Thus we have (Th,m ) = F2m . Every element z Rh,m can be uniquely written as
m1

z=
i=0

zi i ,

z i Z 2h ,

(2.1)

which is known as the additive representation of z. Alternatively, z can be uniquely expressed in the form
h1

z=
i=0

zi 2i ,

zi Th,m ,

(2.2)

called the 2-adic representation of z. 11

2 Background
Example 2.3. The Hensel lift in Z4 [x] of the primitive polynomial x3 + x + 1 (in F2 [x]) is given by p(x) = x3 + 2x2 + x + 3. Let be a root of p(x), i.e., satises 3 = 2 2 + 3 + 1. Then T2,3 = {0, 1, , 2, 3 , 4 , 5 , 6 }, and the zi s in (2.1) of the elements in T2,3 are given by: element 0 1 2 3 4 5 6 z0 0 1 0 0 1 2 3 1 z1 0 0 1 0 3 3 3 2 z2 0 0 0 1 2 3 1 1

Let z Rh,m be written in 2-adic representation, as shown in (2.2). We dene the map : Rh,m Th,m by (z) = z0 . It is straightforward to prove [68, Lemma 4] that m(h1) (z) = z 2 . For any x, y Th,m we dene the operation as x y := (x + y). Lemma 2.4. Th,m , is an abelian group. Proof. Observe that for any x, y Th,m we have x 0 = x (existence of the neutral), x x = 0 (existence of the inverse), and x y = y x Th,m (closure and commutativity). Hence Th,m , , is a eld that is isomorphic to F2m , +, . Using (2.1) and (2.2) it can be shown that every element x Th,m can be uniquely expressed as
m1

(2.3)

x=
i=0

xi i ,

xi {0, 1}.

12

2.3 Generalized Boolean Functions


Let z Rh,m have 2-adic expansion
h1

z=
i=0

zi 2i ,

zi Th,m .

Then the map : Rh,m Rh,m dened by


h1

(z) :=
i=0

2 zi 2i

is the Frobenius automorphism on Rh,m with respect to the ground ring Z2h . The absolute trace function Trm : Rh,m Z2h is then dened to be
m1

Trm (z) :=
i=0

i (z).

We shall write Tr instead of Trm if m is clear from the context. Some properties of the trace function are summarized in the following lemma. For proofs see, e.g., [116] (for h > 1 property (v) was not explicitly proved there, but it follows from [116, Theorem 14.37]). Lemma 2.5. For x, y Rh,m and a Z2h we have (i) Trm (x) Z2h , (ii) Trm (x + y) = Trm (x) + Trm (y), (iii) Trm (ax) = aTrm (x), (iv) Trm ((x)) = Trm (x), (v) When a is xed, Trm (x) = a has exactly 2h(m1) solutions. 2.3 Generalized Boolean Functions Generalized Boolean functions will be used in this thesis in various contexts; they are naturally associated with words over Zq and with polyphase words of length 2m , and they are convenient to dene generalized ReedMuller and related codes. 13

2 Background
A generalized Boolean function is dened as a mapping from the binary mtuples, {0, 1}m, to Zq . We can express any such function f : {0, 1}m Zq uniquely in the polynomial form
m1

f (x) = f (x0 , x1 , . . . , xm1 ) =


i{0,1}m

ci
j=0

xjj ,

ci Z q ,

(2.4)

called the algebraic normal form of f . Sometimes we write f in place of f (x). For any i {0, 1}m the generalized Boolean functions of the form
m1

f (x0 , x1 , . . . , xm1 ) =
j=0

xjj

are called monomial functions or monomials. Every generalized Boolean function can be expressed as a Zq -linear combination of monomials. Since in a polynomial form of a generalized Boolean function the term x2 can always be reduced to xi , i m the set of generalized Boolean functions {0, 1} Zq corresponds to the factor ring Zq [x0 , x1 , . . . , xm1 ]/(x2 x0 , x2 x1 , . . . , x2 m1 xm1 ). 1 0 The degree of a generalized Boolean function f , denoted by deg(f ), is dened to be the algebraic degree of its algebraic normal form, i.e., with the notation as in (2.4) deg(f ) := max wt(i).
ci =0

With each generalized Boolean function f : {0, 1}m Zq we associate a word m in Z2 by listing all the values f (x0 , x1 , . . . , xm1 ) as (x0 , x1 , . . . , xm1 ) ranges q m1 over {0, 1}m in lexicographic order. That is, if i = j=0 ij 2j (ij {0, 1}) is the binary expansion of the integer 0 i < 2m , then the ith element of this word is equal to f (i0 , i1 , . . . , im1 ). By convention, we shall denote this Zq -valued word by the same symbol as that used to identify the generalized Boolean function; the context should make clear to which object we refer. Another representation of the function f is its associated polyphase word f, whose ith element is given by f (i0 ,i1 ,...,im1 ) q . h1 Now let f : {0, 1}m Z2h with h 2, and let : Znh Z2 n be the 2 2 generalized Gray map (cf. Denition 2.1). Then, interpreting f as a Z 2h -valued word of length 2m , the word (f ) is a binary word of length 2m+h1 . This word is associated with a function from {0, 1}m+h1 to Z2 , which we also denote by 14

2.3 Generalized Boolean Functions


(f ). For h = 2 we shall give an explicit expression for the algebraic normal form of this function. In this case the Gray map is given by (0) = (00), (1) = (01), (2) = (11), (3) = (10). The algebraic normal form of f can always be written in 2-adic expansion f (x0 , . . . , xm1 ) = a(x0 , . . . , xm1 ) + 2b(x0 , . . . , xm1 ), where a, b : {0, 1}m {0, 1}. Then the algebraic normal form of (f ) is given by (f )(x0 , . . . , xm ) = a(x0 , . . . , xm1 )xm + b(x0 , . . . , xm1 ). Informally, the variable xm {0, 1} determines whether the output is the left or the right coordinate of the Gray map of f (x). Since the Gray map is bijective for h = 2, we can map any function f : {0, 1}m+1 Z2 to a function 1 (f ) : {0, 1}m Z4 . To this end observe that the algebraic normal form of f can always be written as f (x0 , . . . , xm ) = a(x0 , . . . , xm1 )xm + b(x0 , . . . , xm1 ). Then we have 1 (f )(x0 , . . . , xm1 ) = a(x0 , . . . , xm1 ) + 2b(x0 , . . . , xm1 ). Example 2.6. Let f : {0, 1}3 Z4 be given by f (x0 , x1 , x2 ) = 1 + x1 + 2x2 + x0 x2 + 3x1 x2 . We have f = (11223030) and f = (+j, +j, 1, 1, j, +1, j, +1). The Gray map of f reads (f ) = (0011101011110000), and its corresponding function has algebraic normal form (f )(x0 , x1 , x2 , x3 ) = (1 + x1 + x0 x2 + x1 x2 )x3 + x2 + x1 x2 = x 2 + x 3 + x 1 x2 + x 1 x3 + x 0 x2 x3 + x 1 x2 x3 . 15

2 Background
Now we turn our attention to functions dened on the Teichmuller set T h,m . For any 0 i < 2m 1 , the set {i 2j (mod 2m 1) | 0 j < mi },

where mi > 0 is the smallest number such that i 2mi i (mod 2m 1), is called the cyclotomic coset modulo 2m 1 with coset representative i. The smallest number in the set, say s, is called the coset leader of the coset, and the coset will be denoted by Cs . Let I be the set of coset leaders of the cyclotomic cosets modulo 2m 1. Then, arguing as in the case h = 1 (cf., e.g., [39]), any function f : Th,m Z2h can be uniquely written as f (x) =
iI

Trmi ai xi + a2m 1 x2

ai Rh,mi , a2m 1 Z2h , (2.5)

where mi = |Ci |. Each function of this form is equivalent to a generalized Boolean function g : {0, 1}m Z2h , say, whose algebraic normal form is given by
m1

g(x0 , x1 , . . . , xm1 ) = f
i=0

xi i

where is a generator for Th,m . With any function f : Th,m Z2h we associate Z2h -valued and polyphase words that are those words corresponding to the equivalent generalized Boolean function. We denote these words by f and f, respectively. Example 2.7. Consider the function f : T2,3 Z4 with f (x) = Tr3 (x3 ). Let be as in Example 2.3. Then, when (x0 , x1 , x2 ) ranges over {0, 1}3, x = x0 x1 x 2 2 = x0 + x1 + x2 2 + 2x0 x1 4 + 2x0 x2 + 2x1 x2 5

runs through the Teichmuller set T2,3 . By elementary calculations we obtain x3 = x0 +x1 3 +x2 6 +x0 x1 (3+3 2 +2 4 +2 6 )+x0 x2 (3 2 +3 4 +2+2 5 ) + x1 x2 (3 4 + 3 5 + 2 + 2 2 ) + 2x0 x1 x2 ( 3 + 5 + 3 + ). 16

2.4 Error-Correcting Codes


Using Lemma 2.5 and Tr3 (1) = 3 Tr3 () = Tr3 ( ) = Tr3 ( 4 ) = 2 Tr3 ( 3 ) = Tr3 ( 5 ) = Tr3 ( 6 ) = 1, we nd the algebraic normal form f (x0 x1 x2 2 ) = 3x0 + x1 + x2 + 2x0 x1 + 2x0 x2 + 3x1 x2 + 2x0 x1 x2 of the generalized Boolean function that is equivalent to f . 2.4 Error-Correcting Codes Let S be a set. A code C of length n over S is dened to be a nonempty subset C S n . Two codes are called permutation equivalent if one can be obtained from the other by permuting its coordinates. The elements of a code are called codewords. We dene the code rate of C to be R(C) := log2 |C| . n
2

In this thesis we will study codes over Zq . Typically we have q = 2h , which is of practical interest because of the compatibility of the alphabet with binary representation of information. If h = 1, 2, or 3, such a code is called binary, quaternary, or octary, respectively. A code over Zq is called linear if the following holds a, b C a + b C. (2.6)

Alternatively, such a code is linear if it is a submodule of Zn . If C Zn is linear, q q the set f + C := {f + c | c C} is a coset of C with f Zn being a representative of this coset. q Unlike linear codes over elds, which are vector spaces and, therefore, have a basis, a basis does not necessarily exist for a linear code C Zn 2 . However from q (2.6) it becomes clear that one can associate a generator matrix with C such that
2A

basis for C exists if and only if C is a free module over Zq .

17

2 Background
C consists of all distinct Zq -linear combinations of the rows of this matrix. It was shown by Calderbank and Sloane [10], generalizing earlier results by Hammons et al. [36], that every linear code over Z2h is permutation equivalent to a code with generator matrix Ik0 Z01 Z02 ... Z0h ... 2Z1h 0 2Ik1 2Z12 . , G= . .. . . . . . 0 ... 0 2h1 Ikh1 2h1 Zh1h
h1

where Iki are identity matrices of size ki ki and Zij are matrices over Z2hi . Then the code contains 2(hi)ki
i=0

codewords. Encoding can be performed as follows. Choose an information word u = (u0 |u1 | |uh1 ), where ui is a word of length ki with elements in Z2hi . The corresponding codeword is then given by u G (mod 2h ). For arbitrary q a generator matrix can be obtained by considering the unique factorization of q into prime powers, which denes the decomposition of the ring Zq into a chain of prime elds. See [8] for details on this issue. The minimum distance of a code C Zn that contains at least two codewords q is dened to be d(C) := min d(a, b).
a,bC,a=b

The following lemma is obvious. Lemma 2.8. Let C Zn with |C| > 1 be a linear code. Then q d(C) =
cC,c=(000)

min

wt(c).

The minimum distance of a code determines the number of errors that are guaranteed to be correctable. More precisely, given a word x = c + e, where c C Zn and e Zn , the codeword c can always be recovered from x as long q q as wt(e) < d(C)/2. 18

2.4 Error-Correcting Codes


Now recall from Section 2.1 that the Gray map is an isometry from Z nh to 2 2h1 n n Z2 . We dene the Gray image of a code C Z2h to be (C) = {(c) | c C}. It follows that (C) is a binary code of length 2h1 n and, by Theorem 2.2, d((C)) = d(C). Since the Gray map is bijective for h = 2, any binary code B Z2n can be mapped to a quaternary code of length n by the inverse Gray map 2 1 , i.e., 1 (B) = {1 (b) | b B}. Clearly we have d(1 (B)) = d(B). Now we turn our attention to q-ary PSK codes, which are naturally associated with codes over Zq . Given C Zn , the corresponding q-PSK code is the set q { | c C}, i.e., it comprises polyphase words. When q = 2 (q = 4), such a code c is called a BPSK (QPSK) code. Let the Euclidean norm of x Cn be dened as
n1

||x|| :=

i=0

|xi |2 .

The Euclidean distance between two words x, y Cn is equal to ||x y||. Given a code C Zn , an important parameter of its associated PSK code is its minimum q Euclidean distance, which is given by
a,bC,a=b

min

|| a b||.

For any word y Cn a maximum-likelihood (ML) decoder nds a word in C whose polyphase version is closest to y, i.e., it outputs argmincC ||y c||. If n y = c + e, where c C and e C , an ML decoder always outputs c if ||e|| is less than half the minimum Euclidean distance of the PSK code associated with C. Hence the latter parameter mainly determines the performance of the code when the codewords are transmitted over a noisy channel and the received words are decoded with an ML decoder. It is easy to see that the minimum Euclidean distance of a BPSK code is equal to 2 d(C) and that of a QPSK code equals 2 d(C), where C is the underlying binary or quaternary code, respectively. In general it is difcult to derive the minimum Euclidean distance of a PSK code from the minimum distance of the underlying code over Zq , but we can prove a lower bound. Theorem 2.9. Let C Zn . The minimum Euclidean distance of the q-PSK code q associated with C is at least 2 d(C) sin . q 19

2 Background
We need a technical lemma for the proof. Lemma 2.10. For any q 2 and any integer 0 w w sin2 . sin2 w q q q/2 we have (2.7)

Proof of Lemma. The lemma is trivial for q = 2 and q = 3 and for w = 0, so assume that q > 3 and w > 0. We proceed to prove the lemma for any w R with 1 w q/2. f (w) = sin2 (w/q) is concave for 1 w q/2. Hence the linear function that passes through f (1) and f (q/2) is a lower bound for f (w) in the range 1 w q/2. This linear function meets the RHS of (2.7) at w = 1. So it remains to prove that the slope of this linear function, given by 1 sin2 (/q) , q/2 1

is equal to or greater than sin2 (/q). Hence we have to show 1 sin2 (/q) sin2 q/2 1 q 0,

or equivalently,

q sin2 1, 2 q which is apparently true if q = 4. For q > 4 we have q 2 sin2 < 2 q 2q and the proof is completed. Proof of Theorem 2.9. For any a, b Zq we have 2 < 1, 10

| 2 = 4 sin2 [a b] a b| q = 4 sin2 d(a, b) q 4 d(a, b) sin2 , q q/2 .

where the last step follows from Lemma 2.10 and the fact 0 d(a, b) Hence d(a, b) min || a b|| 2 sin min a,bC,a=b q a,bC,a=b = 2 d(C) sin . q 20

2.5 Aperiodic Correlation Functions


2.5 Aperiodic Correlation Functions One of the main concerns of this thesis will be the construction of sequences with given aperiodic correlation properties. In this section we dene aperiodic correlation functions of sequences and discuss some of their properties. Denition 2.11. Let x, y Cn . The aperiodic crosscorrelation of x and y at a displacement Z is given by
n 1 i=0 n+ 1 i=0 xi+ yi xi yi

<n

C(x, y)( ) :=

n < < 0 otherwise.

The aperiodic autocorrelation of x at a displacement Z is dened to be A(x)( ) := C(x, x)( ). Some basic properties of aperiodic correlation functions are summarized in the following lemma. The proofs are straightforward and omitted. Lemma 2.12. Let x, y Cn and denote by x and y the reverse words of x and y, respectively. Suppose that a, b R and C with || = 1. Then we have (i) (ii) (iii) (iv) (v) (vi) C(x, y)( ) = C (y, x)( ) C( a x, b y)( ) = ab C(x, y)( ) C(, y)( ) = C (y, x)( ) x A(x)( ) = A (x)( ) A( a x)( ) = A(x)( ) A()( ) = A (x)( ). x

Now let a Zn . The aperiodic autocorrelation of the polyphase word a at the q zero shift satises
n1

A()(0) = a
i=0

a a q i q i = n.

(2.8)

21

2 Background
More generally, at shift 0 < n, we obtain
n 1 a a q i+ q i i=0 n 1 a a q i+ q i i=0

|A()( )| = a

=n . From Lemma 2.12 (iv) we see that |A()( )| a 2.6 Multicarrier Communications n + for n < < 0.

(2.9)

We begin with describing a general model of a multicarrier communications system. We will then focus on the peak-to-mean power-control problem, which is an issue that is inherent in each multicarrier system. This power-control problem is then stated explicitly in the subsequent two sections for OFDM and MC-CDMA. Let C Zn be a code of length n. Typical values for q are 2, 4, or 8, which q corresponds to BPSK, QPSK, or 8-PSK modulation of the individual carriers. Although in practical systems also M -ary quadrature-amplitude modulation schemes are used, in this thesis we consider only q-PSK modulation. Given a codeword c C, the complex baseband equivalent or complex envelope of the multicarrier signal can be written as
n1

S(c)() =
i=0

ci fi (),

< 1,

The corresponding real-valued transmitted signal is then given by S(c)() ej2 ,

where the signals fi () are the n carriers, which are pairwise orthogonal, i.e., 0 if i = k 1 fi ()fk () d = 1 if i = k. 0

where R is a positive constant that is much larger than n in practical systems. For example in HiPerLAN/2 we have /n > 300 [23]. 22

2.6 Multicarrier Communications


Assuming a noiseless transmission of this signal, at the receiver the received signal is multiplied by ej2 , i.e., S(c)()ej2 ej2 = 1 S(c)() ej2 + S (c)() ej2 ej2 2 1 1 = S(c)() + S (c)()ej4 , 2 2

and ltered with a low-pass such that the complex envelope S(c)() is recovered. The transmitted codeword c C can then be regained by correlating the received complex envelope with all n carriers. Thanks to the orthogonality of the carriers we have
1 0

S(c)()fi () d

1 n1

0 k=0 n1 1

ck fk ()fi () d fk ()fi () d

=
k=0

ck
0

= ci . In practice, however, we have a noisy transmission and the transmitted signal is superimposed by an additive error signal. Thus the receiver outputs c + e, where n e C is an error vector.

The key parameter that we study in this thesis is the ratio of the peak and the mean value of the envelope power |S(c)()|2 . By Parsevals identity the mean value of |S(c)()|2 is equal to
1 0

|S(c)()|2 d = ||||2 = n, c 1 sup |S(c)()|2 . n 0 <1

so this ratio becomes

For large (recall that this is the case in practical systems) this ratio is approximately two times the ratio of the peak and the mean value of the transmitted power ( {S(c)()ej2 })2 . It is, however, mathematically often more convenient (and a good approximation in practice) to work with the properties of the envelope power rather than with those of the true transmitted power. 23

2 Background
2.7 The Power-Control Problem in OFDM Orthogonal frequency-division multiplexing (OFDM) is a multicarrier scheme, where the n carriers are given by fi () = ej2i , i = 0, 1, . . . , n 1.

The number of carriers is typically in the range 16 to 128, say, for mobile applications, such as wireless local area networking (WLAN), while it can be much larger than 1000 for digital TV applications. The complex envelope corresponding to a codeword c C Z n now reads q
n1

S(c)() =
i=0

ci ej2i ,

< 1.

(2.10)

If we drop the restriction on the time , we see that S(c)() is a continuous-time periodic signal. This signal is essentially the Fourier transform of c (where c is interpreted as a series of weighted Dirac impulses). Alternatively, the elements of c can be interpreted as the coefcients of the Fourier series of the periodic signal S(c)(). We dene the peak-to-mean envelope power ratio (PMEPR) of the codeword c (or the corresponding complex envelope) to be PMEPR(c) := 1 sup |S(c)()|2 n 0 <1
n1 2

(2.11)

1 sup = n 0 <1

ci e
i=0

j2i

Observe that the PMEPR is at least 1 and it can be as large as n. This latter value occurs, for example, if c is the all-zero word. Informally, the PMEPR measures the atness of the spectrum of c with respect to the Fourier transform: the lower the PMEPR of c, the smoother is the envelope power |S(c)()|2 . Occasionally we shall indicate the PMEPR of c as 10 log10 PMEPR(c) [dB]. Example 2.13. Let n = 8. Figure 2.1 shows the envelope power normalized by the mean envelope power, i.e., |S(c)()|2 /n, corresponding to the words (01010101) and (00010010) (lying in Z8 ). The PMEPR of these words can now be read off 2 from the graph. The PMEPR of the former word is equal to 8 (which is the worstcase PMEPR for n = 8) and that of the latter one is equal to 2. 24

2.8 The Power-Control Problem in Multicode CDMA

Figure 2.1: Normalized envelope power of the OFDM signals corresponding to the words (01010101) (solid line) and (00010010) (dashed line) We will be interested in codewords with small PMEPR. In particular we aim at designing error-correcting codes C Zn for which the maximum PMEPR, given q by PMEPR(C) := max PMEPR(c),
cC

is bounded by a value that is much lower than n. More specically we state: Problem (Power-Control Problem in OFDM). Design codes C Zn , where q R(C) is high (ensures efciency), d(C) is high (ensures good error-correction capabilities), and PMEPR(C) is low (ensures low-cost implementation and power efciency). Remark. The above criteria are contradicting design goals. It is, therefore, an interesting problem in itself to study the achievable region of the tuple (R(C), d(C), PMEPR(C)) for a given code length. We will not directly consider this problem here and refer to [83], [98], and [99] for some results on this issue. 2.8 The Power-Control Problem in Multicode CDMA Let the canonical Sylvester-type Hadamard matrix (or WalshHadamard matrix) be recursively dened by H2m := H2m1 H2m1 H2m1 H2m1 with H1 := (1). (2.12)

  

25

2 Background
We denote the ith row of H2m by w(i) . The n = 2m sequences w (0) , . . . , w(n1) are commonly referred to as Walsh sequences. They are pairwise orthogonal, i.e., n1 0 if i = j (i) (j) wk wk = (2.13) n if i = j.
k=0 n1

Multicode code-division multiple access (MC-CDMA) is a multicarrier technique, where the n carriers are given by fi () =
k=0

wk gk p

(i)

k n

i = 0, 1, . . . , n 1.

Here, (g0 , . . . , gn1 ) with |gk | = 1 is called a spreading sequence or signature sequence that is assigned to an individual user, and p() is a pulse waveform that satises p() = 0 if [0, 1/n) and /
1 0

|p()|2 d =

1 . n

In MC-CDMA we always have n = 2m , where typical values for m range from 3 to 8. The orthogonality of the carriers is now easily seen by verifying
1 0 fi ()fj () d =

1 n

n1

wk wk ,
k=0

(i)

(j)

and taking the orthogonality of the Walsh sequences into account (cf. (2.13)). So in MC-CDMA the complex envelope corresponding to the codeword c C Zn reads q
n1 n1

S(c)() =
i=0 n1

ci
k=0

wk gk p
n1

(i)

k n
(i)

=
k=0

k gk p n

c i wk ,
i=0

< 1.

In contrast, in a CDMA system the user transmits the signal


n1 k=0

gk p

k n

a,

< 1,

26

2.8 The Power-Control Problem in Multicode CDMA


where the transmitted information is contained in a Zq . Thus the rate of this user in MC-CDMA is n times that in a conventional CDMA system. The peak power of the MC-CDMA signal can be written as follows
n1 0 <1 n1 2

sup |S(c)()|2 = max = max

0 k<n k/n <(k+1)/n

sup

=0

g p

n
n1

ci w
i=0 (i) c i wk

(i)

0 k<n k/n <(k+1)/n 2

sup

k gk p n
n1 2

i=0

= Therefore we have

sup
0 <1/n

|p ()| max

0 k<n

c i wk
i=0

(i)

1 sup |S(c)()|2 = sup |p()|2 PAPR(c), n 0 <1 0 <1/n where we dene 1 max PAPR(c) := n 0 k<n
n1 (i) c i wk i=0 2

1 = max |( Hn )k |2 c n 0 k<n as the peak-to-average power ratio (PAPR) of c. Hence the ratio of the peak value and the mean value of |S(c)()|2 is proportional to the PAPR of the corresponding codeword. The PAPR can be interpreted as the ratio of the peak value and the mean value of the discrete-time signal c Hn , which is equal to the Hadamard transform of c, so the PAPR measures the spectral atness of c with respect to the Hadamard transform. It is easy to see that the PAPR takes on values between 1 and n, where the upper bound is attained, for example, when c is the all-zero codeword. Example 2.14. Let n = 8. Suppose that p() is the rectangular function, i.e., it is equal to 1 if [0, 1/n) and equal to zero otherwise, and let g0 = g1 = = gn1 = 1. The normalized envelope power of the MC-CDMA signals corresponding to the words (01010101) and (00010010) (lying in Z 8 ) is shown in 2 Figure 2.2. Since sup0 <1/n |p()|2 = 1, it can be seen from the graph that the respective PAPRs are equal to 8 and 2. 27

2 Background

Figure 2.2: Normalized envelope power of the MC-CDMA signals corresponding to the words (01010101) (solid line) and (00010010) (dashed line) We dene the PAPR of a code C Zn as q PAPR(C) := max PAPR(c).
cC

Now we are in a position to state the power-control problem for MC-CDMA systems. Problem (Power-Control Problem in MC-CDMA). Design codes C Zn , q where R(C) is high (ensures efciency), d(C) is high (ensures good error-correction capabilities), and PAPR(C) is low (ensures low-cost implementation and power efciency). Remark. Similar as in the OFDM power-control problem, the above criteria are competing. For q = 2 we refer to [81] for bounds on the achievable region of the tuple (R(C), d(C), PAPR(C)).

28

  

Chapter 3

Phase-Shifted Codes

3.1 Introduction and Chapter Overview In 1996 Jones and Wilkinson [48] proposed a simple approach to design OFDM codes with low PMEPR. Given a code C Zn and a codeword c C, the basic q idea is to shift the ith coordinate of c by a phase shift 0 i < 2, which is independent of c. Hence we obtain a new code that is called a phase-shifted version of the PSK code associated with C. The resulting complex envelope of the OFDM signal is then given by
n1

S(c, )() =
i=0

ci eji ej2i ,

< 1.

The n phase shifts are known a priori to both the transmitter and the receiver, hence, they can be compensated at the receiver. It is then clear that the errorcorrecting properties of the modied scheme are the same as in the case where no phase shifts are applied. Moreover a standard decoder for the original code C can be employed in the receiver upon back-rotation of the phase shifts. However for properly selected phase shifts the maximum PMEPR over all codewords can be signicantly lower than that of the original code. Unfortunately, except perhaps for very short codes, the computation of the optimal phase shifts (which minimize the PMEPR) is considered to be extremely difcult and a feasible solution is unknown. Several suboptimal algorithms have been proposed in the literature [48], [106], and [119]. These techniques were applied to obtain phase shifts for short Hamming codes [48], convolutional codes [106], [119], and nonredundant transmission [119]. However the achieved PMEPR reductions are rather small. One reason may be the suboptimality of the algorithms: 29

3 Phase-Shifted Codes
the use of suboptimal techniques generally results in the convergence of the algorithm to a local minimum, and therefore, it is never certain that the optimal phase shifts are computed. In this chapter we study the fundamental limit of the achievable PMEPR of phase-shifted binary (BPSK) codes. We derive a lower bound for this limit and try answer the question whether suboptimal algorithms reported in the literature are in principle useful to approach this limit. The bound can be used to estimate the gap between the reduced PMEPR, obtained using suboptimal methods, and the global minimum. This allows us, in many cases, to establish the (near) optimality of some known phase-shift designs. We will also see that for several codes of practical importance a signicant PMEPR reduction is ruled out by our results. The organization of this chapter is as follows. We will formalize the problem in Section 3.2. This involves a quantization of the possible phase shifts, such that the phase-shifted code is a 2h -PSK code. For such codes we prove a lower bound on the PMEPR in Section 3.3. The general case, where no quantization of the phase shifts is assumed, is then recovered by analyzing the bound for h . The bound is then applied in Section 3.4 to some well understood codes, including nonredundant BPSK transmission, BCH codes and their duals, Reed Muller codes, and convolutional codes. It will be demonstrated that most of the previously obtained (presumably not optimal) phase-shift designs result in PMEPR reductions that attain or approach our bound. We close this chapter with some concluding remarks and a list of open problems in Section 3.5. 3.2 Denitions and Problem Specication Throughout this chapter we let be a permutation of {0, 1, . . . , n1}, and given a word b = (b0 , b1 , . . . , bn1 ), we write (b) in place of (b(0) , b(1) , . . . , b(n1) ). Denition 3.1. Suppose that B Zn is a binary code of length n. A code C Znh 2 2 is equivalent to B if there exists a permutation of {0, 1, . . . , n 1} and a word w Znh such that 2 C = 2h1 (b) + w | b B . (If h = 1 and B is linear, C is permutation equivalent to a coset of B.) With the notation as above, the 2h -PSK code corresponding to C is given by (0) ej2w0 /2h , (1) ej2w1 /2h , . . . , (n1) ej2wn1 /2h b B . b b b 30

3.3 A Lower Bound on the PMEPR


This code is a phase-shifted version of the BPSK code associated with B, where the ith phase shift is equal to 2wi /2h with wi Z2h . Notice that, when h tends to innity, any phase shift 0 i < 2 can be approximated in this way with arbitrarily high precision. Denition 3.2. Given B Zn , we dene Eh (B) to be the set of all codes C Znh 2 2 that are equivalent to B. The power-control problem for OFDM can now be tackled as follows. Given a good (in the classical coding-theoretic sense) binary error-correcting code B, nd a code in Eh (B), whose PMEPR is equal to (or at least close to)
CEh (B)

min PMEPR(C).

(3.1)

This problem is considered to be extremely difcult, and only suboptimal algorithms are known, which may not even nd a code whose PMEPR is close to the preceding expression. In this light we ask: what is the value of the expression in (3.1)? In the next section we shall prove a lower bound for this quantity. 3.3 A Lower Bound on the PMEPR In deriving our lower bound for (3.1) the covering radius of a binary code plays an essential role. Denition 3.3. The covering radius of a code B Zn is dened to be 2 (B) := max min wt(b + x). n
xZ2 bB

In other words, (B) is the least nonnegative integer, such that the balls of radius (B) around the codewords of B cover the space Zn . If B is linear, (B) is equal 2 to the maximum weight of the coset leaders of B. Determining the covering radius of a binary code is generally nontrivial [65]. However many results are known. For a good overview we refer to [17] and [6]. For the remainder of this chapter we x the following parameters depending on h: 1 if h = 1 := 2 if h > 1

31

3 Phase-Shifted Codes
and := 1 2

if h = 1
2h3

sin

2h

if h > 1.

We are now in a position to state the main result of this chapter. Theorem 3.4. Let B Zn be a binary code with covering radius (B) 2 n 1 1 . Then 2
CEh (B)

min PMEPR(C)

[n(2 ) 2 (B)] . n

(3.2)

We postpone the proof of Theorem 3.4, instead, we rst focus on its asymptotic analysis. Corollary 3.5. Let B Zn be a binary code with covering radius (B) 2 8 1 n 2 2 . Then
h CEh (B)

lim

min PMEPR(C)

16 2 1 2 n (B) 2n 8 4

Proof. The lower bound in (3.2) is a nonincreasing function in h. Therefore we have to analyze the RHS of (3.2) for h . Recall that = 2 for any h > 1, and with the Taylor series sin2 x = x2 + O(x4 ) we conclude
h

(3.3)

lim

= lim 22h3 sin2


h

2h

2 1.2337 . . . . 8

The corollary follows then immediately. Remark. The condition (B) n/2 (which is for h > 2 slightly stronger than those in Theorem 3.4 and Corollary 3.5) is always met if B has strength 1 [17], [6], that is, each coordinate of B takes the values 0 and 1 equally often. In particular every linear code without a coordinate that is always zero has strength 1. In the remainder of this section we prove Theorem 3.4. When h = 1, the proof is straightforward. 32

3.3 A Lower Bound on the PMEPR


Proof of Theorem 3.4 When h = 1. Here we have = 1 and (a0 , . . . , an1 ) Zn . From (2.11) and (2.10) we have 2 PMEPR(a) 1 |S(a)(0)|2 n 1 = n
n1 2

= 1. Let a =

(1)
i=0

ai

1 = [n 2wt(a)]2 . n It follows that


CE1 (B)

min PMEPR(C)

1 min min max [n 2wt((b) + w)]2 n wZn bB 2 1 = min max [n 2wt(b + w)]2 . n wZn bB 2

The condition (B) arrive at


CE1 (B)

n/2 implies that n 2(B) is nonnegative. Therefore we 1 [n 2 max min wt(b + w)]2 wZn bB n 2 1 = [n 2(B)]2 , n

min PMEPR(C)

and the theorem follows for h = 1. We need two lemmas in order to prove Theorem 3.4 for h > 1. Lemma 3.6. Let h > 1. Then for each a, b Znh we have 2 || 2 a b|| where equality holds for h = 2. Proof. It sufces to prove the lemma for n = 1. It is straightforward to establish || 2 = 4 sin2 (a b) h a b|| 2 = 4 sin2 wt(a b) Hence we have to show that sin2 w 2h 2h sin2 h 4 2 w (3.4) 33 2h sin2 2h wt(a b),

. 2h

3 Phase-Shifted Codes
holds for h > 1 and w = 0, 1, . . . , 2h1 . For the case h = 2 the equality in (3.4) is easily proved by hand, since the only possible values for w are 0, 1, 2. For the case h 3 the idea of the proof is as follows. For xed h consider the function f : R R with f (x) = sin2 (x /2h ), which is a continuous equivalent of the LHS of (3.4). First a tangent to the curve f (x) at some 0 < x0 2h1 is constructed that passes through the origin. Then it is shown that for 0 x 2 h1 and h 3 the curve f (x) is upper bounded by this tangent and the slope of this tangent is at most 2h sin2 h , (3.5) 4 2 which is the coefcient of w in the RHS of (3.4). The curve f (x) has a local minimum at x = 0 and a local maximum at x = h1 2 , and has exactly one inection point between them. Since x0 > 0, it is then obvious that the tangent must meet the curve f (x) in the concave part at some 0 < x0 2h1 , where x0 is uniquely determined. We conclude that f (x) is upper bounded by the tangent for 0 x 2h1 . Since the tangent passes through the origin, x0 must satisfy f (x0 ) = f (x0 )x0 , or equivalently, 2x0 x0 = tan . h 2 2h (3.6)

Now we aim to show that f (x0 ) is at most (3.5), which is equivalent to showing that 2x0 4 sin 22h sin2 h , (3.7) 2h 2 where x0 satises (3.6). Notice that (3.6) is a transcendental equation, which generally does not have a closed-form solution. However we can apply numerical methods to solve (3.6), and nd that the LHS of (3.7) is upper bounded by 9.12. The RHS of (3.7) is an increasing function in h, and by using (3.3), we conclude that when h it converges to 2 , clearly greater than 9.12. Hence there exists an integer h0 so that (3.7) is true for all h h0 . Numerical analysis yields h0 = 3. In the next lemma we prove some properties of the generalized Gray map (cf. Denition 2.1). 34

3.3 A Lower Bound on the PMEPR


Lemma 3.7. For a Znh and b Zn we have 2 2 (i) (a 2h2 ) = (a), (a 2h2 ) = (a) (ii) (a + 2h1 b) = (a) + (2h1 b), where the addition of scalars is understood to be componentwise, and (a) denotes the complement of a (a). Proof. The statements in the lemma are immediate consequences of the apparent fact that for a, Z2h the binary word (a + ) is equal to a negacyclic shift by positions to the left of the binary word (a). Now we complete the proof of Theorem 3.4. Proof of Theorem 3.4 When h > 1. Here = 2 and = 22h3 sin2 a = (a0 , . . . , an1 ) Znh . It follows from (2.11) and (2.10) that 2 PMEPR(a) 1 2 |S(a)(0)| n 1 = n Using the easily veried identities
n1 n1 2 2h

. Let

ai
i=0

|| 1|| = 2n 2 a || j||2 = 2n 2 a

ai
i=0 n1

ai
i=0

where we interpret the addition of scalars to sequences componentwise, we observe PMEPR(a) The inequality x2 + y 2 = 1 (x + y)2 + (x y)2 2 1 (x + y)2 , x, y R 2 35 1 n 1 a n || 1||2 2
2

1 + n || j||2 a 2

(3.8)

3 Phase-Shifted Codes
yields PMEPR(a) 1 1 2n || 1||2 + || j||2 a a 2n 2
2

Applying Lemma 3.6 we arrive at PMEPR(a) 1 2n h2 wt(a) + wt(a 2h2 ) 2n 2


2

(3.9)

provided that the expression inside the square is nonnegative. This will be assumed in the following and justied at the end of the proof. We now use the weight-preserving property (cf. Theorem 2.2) of the generalized Gray map (cf. Denition 2.1) and Lemma 3.7 (i) to establish wt(a) + wt(a 2h2 )

= 2wt((a)) + wt((a)) + wt((a)) = 2wt((a)) + 2h2 n, where (a) is the complement of (a). Hence PMEPR(a)

= wt((a)) + wt((a)) + wt((a 2h2 )) + wt((a 2h2 ))

= wt((a)) + wt((a 2h2 ))

1 2n h2 (2wt((a)) + 2h2 n) 2n 2 2 1 = n(2 ) h3 wt((a)) . 2n 2

Indeed
CEh (B)

min PMEPR(C)
wZ bB

= min min max PMEPR(2h1 (b) + w) n


2h

min min max n


wZ
2h

bB

1 n(2 ) h3 wt((w + 2h1 (b))) 2n 2


2

= min max n
wZ
2h

bB

1 n(2 ) h3 wt((w + 2h1 b)) 2n 2 1 n(2 ) h3 wt((w) + b ) 2n 2


2

= min max n
wZ
2h

bB

(3.10)

36

3.4 Applications
where Lemma 3.7 (ii) has been employed and b is given by b = (2h1 b) = (b|b| |b).
2h2 times

Recall that we imposed (B) n 1 1 ,which implies that the expression n(2 2 ) 2 (B) is nonnegative. Hence we can write [n(2 ) 2 (B)] = n(2 ) 2 max min wt(u + b) n
uZ2 bB 2 2

= min max [n(2 ) 2 wt(u + b)] n


uZ2 bB

2 2

min
uZ2 2

h2 n

max n(2 )
bB

wt(u + b ) h3

wZ

min max n(2 ) n


2h

bB

wt((w) + b ) h3

(3.11) Notice that the latter inequality follows because is nonsurjective for h > 2. Since the term inside the square in (3.11) is nonnegative, our assumption leading to (3.9) holds. For h > 1 Theorem 3.4 follows then from (3.10) and (3.11). Remark. The proof of Theorem 3.4 provides hints on how the offset w Z nh must 2 look like in order to meet the lower bound in Theorem 3.4 with equality. If there exists a code C E1 (B) such that PMEPR(C) meets the bound in Theorem 3.4, the offset w Zn must have weight equal to (B) or it must be the sum of a 2 word of weight (B) and a codeword of B. If there exists a code C E 2 (B) whose PMEPR attains the bound in Theorem 3.4, then, by plugging the equality in Lemma 3.6 directly into (3.8), one can show that n must be even and w Z n 4 must satisfy: w = w + 2b with wt((w )) = (B), wt((w )) = n/2, and b B. 3.4 Applications From Theorem 3.4 we conclude that a necessary condition to obtain a 2 h -ary code with low PMEPR from a binary code is that the underlying binary code must have a large covering radius. More specically, to obtain a sequence of codes with increasing length n and PMEPR growth strictly slower than linear in n from a sequence of binary codes, the covering radius of the binary code must grow when 37

3 Phase-Shifted Codes
n increases. In particular, in order to achieve a constant PMEPR, the covering radius of the binary code B must satisfy (B) n 1 1 2 O( n),

which can be replaced by the slightly stronger condition (B) n O( n). 2 (3.12)

We shall see later in this section that this accounts for the duals of binary primitive BCH codes and for rth-order ReedMuller codes. In the remainder of this section we apply Theorem 3.4 and Corollary 3.5 to several code families, and compare the lower bounds on the PMEPR with some phase-shift designs. Nonredundant BPSK Signaling In what follows we consider the code B = Zn . This situation corresponds to 2 BPSK signaling without redundancy. Since (B) = 0, we immediately conclude from Theorem 3.4 and Corollary 3.5: Corollary 3.8. Let B = Zn . For h > 1 we have 2
CEh (B)

min PMEPR(C)

(2 )2 n 2

and lim min PMEPR(C) 2n 1 2 16


2

h CEh (B)

0.2936 n.

In [119] Wunder and Boche proposed to use Newman phases [69] to reduce the PMEPR for nonredundant BPSK transmission. Newman phases are given by i = i2 , n i = 0, 1, . . . , n 1.

For n = 16 and n = 128 in [119] the PMEPR could be reduced from 12.04 dB and 21.07 dB to 9.97 dB and 17.89 dB, respectively. We have used the algorithm described in [119] to calculate the PMEPR reductions for lengths n = 2 m , where m ranges from 2 to 16. In addition we have rounded the Newman phases to the 38

3.4 Applications

log2 n

Figure 3.1: PMEPR reduction for nonredundant BPSK signaling and n carriers with Newman phases (+) and with Newman phases rounded to points in a QPSK constellation (), upper bounds (dashed lines) nearest points in a QPSK constellation, i.e., in {+1, +j, 1, j}. The results are presented in Figure 3.1. The gure also shows the corresponding bounds on the PMEPR reductions obtained with Corollary 3.8 (10 log10 0.5 3.01 dB for h = 2 and 10 log10 0.2936 5.32 dB for h ). By using rounded Newman phases the bound in Corollary 3.8 is met with equality for m 9. With the original Newman phases the PMEPR reduction seems to converge to approximately 3.9 dB. Thus for large n Newman phases approach our bound by about 1.4 dB. Linear Codes Next we present a lower bound for the PMEPR of 2h -ary codes that are equivalent to a given binary linear code as a function of the code rate. Corollary 3.9. Let B Zn be a binary linear code with rate R(B) 2 Then 2 4n 3 min PMEPR(C) 1 R(B) . CEh (B) 2 Proof. Let k be the dimension of B, such that R(B) =
k n. 3 2

The redundancy bound 39

# "!        % $ &

1.

3 Phase-Shifted Codes
[17], [6] states (B) n k. (3.13)

k 3 1 The condition n n 1 1 . Hence we can plug 2 then implies that (B) 2 (3.13) into the inequality in Theorem 3.4, which yields the desired expression.

The above corollary shows explicitly that one cannot construct a sequence of 2 -ary codes from binary linear codes with rate R > 3 1 and PMEPR growing 2 slower than linear in n. We remark that some improved bounds for linear codes can be obtained by taking bounds for the covering radius that are better than the redundancy bound, e.g., the Griesmer-like upper bound on the covering radius of linear codes [46], which takes the minimum distance of the code into account.
h

Binary Primitive BCH Codes and Their Duals In what follows we will analyze lower bounds on the PMEPR of codes that are equivalent to binary primitive t-error-correcting BCH codes, B(t, m), and their duals, B (t, m). These codes have length 2m 1. See, e.g., [60] for details on these codes. Corollary 3.10. Let n = 2m 1. Then we have
CEh (B(t,m))

min

PMEPR(C)

1 [n(2 ) (4t 2)]2 n 4, and for arbitrary t and

for t = 1, m 2 , for t = 2, m 3, for t = 3, m m m0 , where m0 is nite and depends on t.

Proof. It is known that (B(1, m)) = 1 for m 2 (B(1, m) is perfect), (B(2, m)) = 3 for m 3 [31], (B(3, m)) = 5 for m 4 [37]. More generally, it was proved in [112] that there exists an m0 depending on t such that (B(t, m)) = 2t 1 for all m m0 . Putting these values into Theorem 3.4 yields the statement in the corollary. Remark. It has been shown in [18] that m0 satises m0 40 2 log2 [(2t 1)! (2t 3)] .

3.4 Applications
The preceding corollary shows that every code in Eh (B(t, m)) has PMEPR growing linearly in n. Jones and Wilkinson [48] applied a learning algorithm to obtain phase shifts for B(1, 3) and B(1, 4), the rst nontrivial Hamming codes. As a result, [48] reports binary phases that reduce the PMEPR from 8.45 dB and 11.76 dB to 5.53 dB and 10.52 dB, respectively. These values meet the lower bound in Corollary 3.10, and we conclude that the binary phase shifts computed in [48] are the best possible. In [48] the learning algorithm was also applied to nd offsets over Z8 . The respective reduced PMEPRs are 3.42 dB and 8.47 dB. For B(1, 3) Corollary 3.10 fails to provide a nontrivial lower bound. For B(1, 4) Corollary 3.10 yields the lower bound 5.30 dB. Another example is provided below. Example 3.11. Consider the double-error-correcting BCH code B(2, 4) of length 15 and dimension 7 with generator matrix 101011011000000 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 1 0 0 1 1 0 0 0 1 1 0 0 0 0 G = 1 1 1 0 0 1 0 0 0 1 0 1 0 0 0 . 1 1 0 1 0 0 0 0 0 1 0 0 1 0 0 1 0 1 1 0 0 0 1 0 0 0 0 0 1 0 011001010000001

Since this is a linear code of length 15, its PMEPR is equal to 11.76 dB. Through a computational search we found the binary code {u G + w | u Z 7 } with 2 w = (0 0 0 1 0 0 0 0 0 0 0 1 0 0 1), having PMEPR equal to 7.32 dB. This value meets the lower bound in Corollary 3.10, and therefore, the phase shifts and the permutation of the code positions are optimal. We also found the quaternary code {2u G + w | u Z7 } with 2 w = (2 1 0 1 2 1 1 1 3 3 2 0 1 1 1), which has a PMEPR of 5.80 dB. This should be compared with the lower bound of 4.31 dB from Corollary 3.10. Notice that the PMEPR strongly depends on the ordering of the columns of the generator matrix. In our example the generator matrix differs from the standard cyclic form. Let us turn our attention to B (t, m). 41

3 Phase-Shifted Codes
Corollary 3.12. Let n = 2m 1. Then
CE1 (B (t,m))

min

PMEPR(C)

1 6 1 + 2( t t) n t 1 n

Proof. The corollary follows from Theorem 3.4 and a result from [108] (B (t, m)) n1 ( t 6 t) n t 1. 2

Observe that the covering radius of B (t, m) satises (3.12), and the lower bound on the PMEPR is asymptotically independent of the code length. However notice that the rate is strictly decreasing with increasing length. It is noteworthy that Paterson and Tarokh [83] obtained a bound on the PMEPR of the nonzero codewords of B (t, m), which has order (log n)2 , and they speculated that by taking cosets of B (t, m) the PMEPR may be signicantly further reduced. This conjecture is supported by Corollary 3.12. ReedMuller Codes Next we establish lower bounds on the PMEPR of codes that are equivalent to rthorder ReedMuller codes of length 2m , RM2 (r, m). We refer to Section 4.2 and to [60] for details on these codes. For r = 1 we have: Corollary 3.13. The PMEPR of any code in E1 (RM2 (1, m)) is at least 2 for odd m 7 and is lower-bounded by 1 in all other cases. Proof. It was proved in [38] that 2m1 2(m1)/2 (RM2 (1, m)) 2m1 2m/21 , (3.14)

where the upper bound is tight when m is even. The lower bound of 1 for the PMEPR follows then from Theorem 3.4 (for odd m the lower bound can be tightened slightly by taking into account that (RM2 (1, m)) must be an integer). In the special case where m is odd and m 7 it is known that the lower bound in (3.14) is tight [1], [67], [43]. Therefore Theorem 3.4 states that the PMEPR is lower bounded by 2 in this case. Indeed Davis and Jedwab [22] (see also Corollary 4.30 for a restatement of this result) explicitly constructed m!/2 binary offsets for RM2 (1, m) that reduce the 42

3.4 Applications
PMEPR to values not exceeding 2 for any m. Hence the bound in Corollary 3.13 is met for odd m 7. Corollary 3.13 only considers the case when h = 1. However, even if (RM2 (1, m)) attains the lower bound in (3.14), Theorem 3.4 yields the trivial lower bound 1 for the PMEPR of quaternary codes equivalent to RM 2 (1, m) for all choices of m. Let us turn our attention to ReedMuller codes of arbitrary order. For r > 1 general explicit results for the covering radius of RM 2 (r, m) are unknown, for small r and m a few values and bounds are given in [6, page 802]. However good asymptotic bounds are known, which enable us to analyze the asymptotic behavior of Theorem 3.4 for all ReedMuller codes of xed order. Carlet and Mesnager [15] recently improved previous asymptotic bounds for (RM2 (r, m)) (r > 1) to (RM2 (r, m)) 2
m1

15 2+1 2

r2

2m/2 + O(1).

Observe that the covering radius satises (3.12). Hence Theorem 3.4 yields a lower bound that is asymptotically a constant. More specically, we obtain the following. Corollary 3.14. When m tends to innity, we have min PMEPR(C) 15 2+1
2r4

CE1 (RM2 (r,m))

for r > 1. This bound is indeed independent of the length of the code. But notice that for xed r the code rate tends to zero as the length of the code increases. Convolutional Codes A binary convolutional code C is a set of binary sequences of innite length, which are the possible outputs of a time-invariant sequential linear system. For details see, e.g., [64]. Suppose that the system has memory length , k 0 binary inputs, and n0 binary outputs. Then C has rate k0 /n0 , and + 1 is called its constraint length. Let us consider subsequent outputs, assuming an arbitrary initial state of the system. Then the binary output sequences have length n = n 0 , and the set of all such sequences may be viewed as a binary linear block code C with generator 43

3 Phase-Shifted Codes
matrix of size ( + )k0 n0 G 0 0 0 0 0 G1 G 0 0 0 0 . .. .. .. .. . .. .. .. .. .. .. .. .. . .. .. .. .. .. .. .. . G0 G1 G 0 0 0 , 0 G0 G1 G 0 0 . .. .. .. .. . .. .. .. .. .. .. .. .. . .. .. .. .. .. .. .. . 0 0 0 0 0 G0

where G0 , G1 , . . . , G are matrices of size k0 n0 with elements in Z2 , and 0 is the all-zero matrix of the same size. Then C has length n = n0 and dimension k = ( + )k0 . Note that the rst k0 information bits are fed into the encoder to produce the initial state of the system. This situation corresponds to the case where a block of subsequent symbols of a continuous stream of system outputs is used as a binary codeword of length n0 to modulate one OFDM symbol. So the initial state of this block is the nal state of the previous block. The normalized covering radius of a convolutional code is dened to be [3], [9] (C) := lim (C ) . n0

Bounds for the normalized covering radius of some convolutional codes can be found in [3] and [9]. The covering radius of C satises [3] (C ) Therefore it follows from Theorem 3.4: Corollary 3.15. With the notation as above, suppose that (C) min PMEPR(D) 4n 1 (C) + 1 2
2 1 1 2 . Then

n (C).

DEh (C )

1 The preceding corollary states that whenever (C) 1 2 the PMEPR of any code in Eh (C ) grows linearly in n. We remark that for a particular n the result can be tightened slightly by taking into account that the covering radius of a block code of nite length must be an integer. As an example let us consider the scenario of binary signaling in the HiPerLAN/2 OFDM system [23], which uses 48 carriers for data transmission. The

44

3.5 Conclusions and Open Problems


employed convolutional code has rate 1/2, constraint length 7, and is completely described by (G0 , G1 , . . . , G6 ) = ((11), (01), (11), (11), (00), (10), (11)). Tarokh and Jafarkhani [106] have computed quaternary, octary, and 16-ary phase shifts for C24 of length 48. The respective reduced PMEPR is equal to 12.7 dB, 12.6 dB, and 12.4 dB. We used a simple search algorithm to nd a binary offset that reduces the PMEPR to 14.6 dB. Notice that the PMEPR of the original code C24 is equal to 16.8 dB since the code contains the all-zero word. From [3] we know that (C) 1/5, and therefore, (C24 ) 9. However direct computation yields (C24 ) = 6. Applying Theorem 3.4 directly, the PMEPR of any code in Eh (C24 ) is at least 14.3 dB, 11.3 dB, 8.4 dB, and 7.4 dB for h = 1, 2, 3, 4, respectively. We conclude that for h = 1 a signicant PMEPR reduction is precluded by our results and the reduced PMEPR approaches our bound by 0.3 dB, for h = 2 the PMEPR reduction obtained in [106] is close to the optimum, and for h = 3, 4 the gap between the reduced PMEPR and our lower bound is about 45 dB. This gap may have the following reasons: (i) Theorem 3.4 is not tight in this case, (ii) the optimization algorithm in [106] converged to a local minimum, or (iii) a lower PMEPR can be obtained by applying another permutation to the coordinates of the code. 3.5 Conclusions and Open Problems We have analyzed the reduction of the PMEPR of a code when a permutation and a xed phase shift is applied to its coordinates. A lower bound on the PMEPR for the case where the phase shifts are quantized (i.e., they are in the set { 2i | i = 0, 1, . . . , 2h 1}) was proved, and the asymptotic behavior for h 2h was examined. This lower bound is strongly related to the covering radius of the binary code, and generally speaking, a small covering radius of the binary code precludes a signicant PMEPR reduction. For h = 1 and h = 2 we exhibited examples where the lower bound in Theorem 3.4 is met with equality. Theorem 3.4 was also employed to show that most phase-shift designs from the literature are (almost) optimal. It was demonstrated as well that for several code families, including BCH codes and convolutional codes, Theorem 3.4 precludes a signicant PMEPR reduction. We close with some open problems and possible further research directions. 45

3 Phase-Shifted Codes
We have seen that quaternary phase shifts obtained by rounding Newman phases are optimal phase shifts for nonredundant binary signaling for lengths 2m , where 9 m 16. On the other hand, by applying original Newman phases to nonredundant binary signaling, for large lengths the PMEPR reduction seems to converge to approximately 3.9 dB. Find a theoretical explanation for this behavior. We have identied codes for which the lower bound on the PMEPR is asymptotically independent of the length of the code. Namely these are the duals of the binary primitive BCH codes and ReedMuller codes. Indeed Davis and Jedwab [22] constructed cosets of the rst-order ReedMuller code whose PMEPR is independent of the length of the code. Do there exist cosets of arbitrary ReedMuller codes whose PMEPR is upper bounded by a constant? If so, nd a way to construct them. For a given binary code B Zn nd a good upper bound for (3.1). Such a 2 bound could provide results on the existence of good phase shifts. Finally we wish to point out the most essential (and most difcult) open problem within this context. Given h and a binary code B Zn , nd a code 2 in Eh (B), whose PMEPR is equal to (3.1).

46

Chapter 4

Near-Complementary Pairs and ReedMuller Codes

4.1 Introduction and Chapter Overview A Golay complementary pair is a pair of sequences whose aperiodic autocorrelation values sum to zero at each nonzero shift [30] (see also Denition 4.5). Any member of a Golay complementary pair is called a Golay sequence. It is known that the PMEPR of Golay sequences is at most 2 [84]. Slightly generalized by Paterson [79], it was proved by Davis and Jedwab [22] that a family of Golay sequences organizes in m!/2 cosets of a generalized rst-order ReedMuller code RMq (1, m) (see Denition 4.1) inside a generalized second-order ReedMuller code. Through a computationally obtained ranking of cosets of RM q (1, m) according to their PMEPR, it was further observed in [22] that there exist many regularities in the PMEPR of such cosets. While the theory of complementary sets can explain much about the PMEPR behavior of cosets of RMq (1, m) (see [79], [78], [105], [16], and Chapter 5), several phenomena remain unexplained until now. The main result of this chapter is an explicit construction of cosets of RMq (1, m) whose PMEPRs are bounded by low values. By taking unions of such cosets, OFDM codes with low PMEPR can be easily derived. The result also shows how earlier experimental observations by Davis and Jedwab [22] and Parker and Tellambura [76], [77] arise as a part of a unied framework. An outline of the chapter is given next. We will recall denitions and some basic properties of generalized ReedMuller codes in the next section. In Section 4.3 we prove an upper bound on the PMEPR of so-called near-complementary 47

4 Near-Complementary Pairs and ReedMuller Codes


sequences. Informally, a near-complementary sequence is a member of a nearcomplementary pair, which is a pair of sequences whose aperiodic autocorrelation values sum to small values at a few nonzero shifts. This is a natural generalization of Golay sequences and leaves much more design freedom for the construction of sequences with low PMEPR. We will see in Section 4.4 that known techniques for the construction of complementary pairs are readily applicable to construct near-complementary pairs. A much more elaborate construction of nearcomplementary pairs is then proved in Section 4.5. These pairs organize in cosets of RMq (1, m), and we immediately obtain an upper bound for the PMEPR of these cosets. In Section 4.6 we prove a lower bound on the PMEPR of cosets of RMq (1, m). This lower bound is used to analyze the achievable PMEPR of such cosets and to establish the tightness of the upper bound in Section 4.5 in certain cases. The application of our theory is then illustrated in Section 4.7. First, it is shown that the m!/2 cosets of RMq (1, m) comprised of Golay sequences arise as a special case. Second, further families of cosets with PMEPR upper bounds between 2 and 4 are presented, showing that several previously unexplained phenomena can now be understood within a unied framework. We conclude Section 4.7 with a discussion on resulting coding schemes for OFDM with low PMEPR. In Section 4.8 we present two maximum-likelihood decoders for RM q (1, m). These algorithms can also be used to perform maximum-likelihood decoding of any union of cosets of RMq (1, m) by means of the supercode decoding principle. The chapter will be concluded in Section 4.9, which also contains a list of open problems related to this chapter. Convention. Throughout this chapter q is an even positive integer.

4.2 Generalized ReedMuller Codes In this section we recall two generalizations of the classical ReedMuller code (cf., e.g., [60]) from [79] based on earlier work in [22]. Denition 4.1. ([79], based on [22]) For q 2 and 0 r m the rth-order generalized ReedMuller code RMq (r, m) is dened as the set of words corresponding to generalized Boolean functions f : {0, 1}m Zq of degree at most r. 48

4.2 Generalized ReedMuller Codes


For q 4 and 0 r m + 1 the code ZRMq (r, m) is dened as the set of words corresponding to generalized Boolean functions f : {0, 1} m Zq of degree at most r, where the coefcients of the monomials in the algebraic normal form of f with degree equal to r are even. By convention, monomials with degree equal to m + 1 are zero. Both RMq (r, m) and ZRMq (r, m) are linear codes over Zq . The code RM2 (r, m) is identical to the classical ReedMuller code (cf. [60]), while RMq (r, m) generalizes this code to q-ary alphabets. The code ZRMq (r, m) is a generalization to q-ary alphabets of the quaternary code ZRM(r, m) from [36], which in turn generalizes the classical ReedMuller code. A generator matrix for RMq (r, m) can be obtained by taking a matrix whose rows are all words associated with monomials of degree at most r. Likewise a generator matrix for ZRMq (r, m) has rows corresponding to all words associated with monomials of degree at most r 1 together with all words associated with two times the monomials of degree equal to r. Example 4.2. The code RMq (2, 3) is a linear code over Zq with generator matrix 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 1 0 0 0 0 0 1 0 1 0 0 0 0 1 1 1 0 1 0 0 1 0 0 1 0 0 0 1 1 0 1 0 1 0 1 0 1 1 0 0 1 1 1 1 1 1 1 1 1 x0 x1 x2 x0 x1 x0 x2 x1 x2

and ZRMq (2, 3) is also linear over Zq and generated by 1 1 0 0 0 0 0 1 0 1 0 0 0 0 1 1 1 0 2 0 0 1 0 0 1 0 0 0 1 1 0 1 0 2 0 1 0 1 1 0 0 2 1 1 x0 1 1 x1 1 x2 2 2x0 x1 2 2x0 x2 2 2x1 x2 .

It is well known that the classical ReedMuller code RM 2 (r, m) can be recursively constructed using the so-called (u | u + v) construction (cf. [60, Chapter 13, 49

4 Near-Complementary Pairs and ReedMuller Codes


Theorem 2]). This construction naturally extends to the generalized versions. For 0 < r < m we have RMq (r, m) = { (u | u + v) | u RMq (r, m 1), v RMq (r 1, m 1)}, (4.1) while for 0 < r < m + 1 we have ZRMq (r, m) = { (u | u + v) | u ZRMq (r, m 1), v ZRMq (r 1, m 1)}. (4.2) The minimum distances of RMq (r, m) and ZRMq (r, m) can be obtained from the above relations. Theorem 4.3. d(RMq (r, m)) = 2mr d(ZRMq (r, m)) = 2mr+1 . Proof. By Lemma 2.8 we have to nd the minimum weights of the nonzero codewords. Using a straightforward generalization of [60, Chapter 2, Theorem 33] m we observe that the word (u | u + v) Z2 has weight min{2wt(u), wt(v)}. q With (4.1) and (4.2) the theorem can now be proved by induction on m using as the induction anchor that RMq (0, m) and ZRMq (0, m) are repetition codes with respective minimum nonzero weights 2m and 2m+1 , and RMq (m, m) = m ZRMq (m + 1, m) is the space Z2 with minimum nonzero weight 1. q In [22] an alternative proof of Theorem 4.3 was given for the case q = 2 h . 4.3 Complementarity of Sequence Pairs Denition 4.4. For two sequences x, y Cn we dene the operator as follows x y :=
Z

|A(x)( ) + A(y)( )| .

If a, b Zn , we obtain q
n1

a = 2n + 2 b
=1

A()( ) + A( ) , a b)(

(4.3)

50

4.3 Complementarity of Sequence Pairs


and from (2.8) and (2.9) we have 2n a b 2n2 .

Sequence pairs that attain the lower bound are of particular interest. Denition 4.5. Suppose that a, b Zn . The pair of polyphase sequences (, a b) q = 2n. Any sequence lying in a Golay is a Golay complementary pair if a b complementary pair is called a Golay sequence. The preceding denition implies that (, is a Golay complementary pair if a b) and only if 2n if = 0 A()( ) + A( ) = a b)( 0 if = 0,

which is, perhaps, the more common denition of Golay complementary pairs (cf. [24], [75]). Binary complementary pairs (which contain sequences over the alphabet {+1, 1}) have been studied by Golay [30] in connection with multislit spectrometry [28], [29]. Other applications of complementary pairs include pulse compression radar [44], surface wave encoding [109], system identication [73], and channel estimation in communications systems [104]. After the original work by Golay, the concept of complementary pairs has been generalized to polyphase alphabets [100], [27] and to multilevel alphabets [21], [7], [86], [54]. This allows the construction of complementary pairs of almost arbitrary lengths, while binary complementary pairs are only known to exist for lengths 2a 10b 26c with a, b, c being nonnegative integers [4]. A good overview of complementary pairs can be found in [24], [75], and [4]. In what follows we generalize the concept of complementary pairs to nearcomplementary pairs. Denition 4.6. Let a, b Zn . The pair of polyphase sequences (, is a neara b) q are near-complementary sequences if a complementary pair and a, b b 2n2 . The notion of near-complementary pairs seems to be rst introduced in [41] (where they are called broad-sense complementary pairs) in order to increase the design freedom for radar signals. They are also useful for the construction of sequences with low PMEPR, as we will show next. Theorem 4.7. The PMEPR of a, b Zn is at most ( a b)/n. q 51

4 Near-Complementary Pairs and ReedMuller Codes


We need a lemma, which was rst established in [107]. Lemma 4.8. [107] Given c Zn , the corresponding OFDM signal satises q
n1

|S(c)()| = n + 2 Proof of Lemma. From (2.10) we have

A()( )ej2 c
=1

|S(c)()|2 = S (c)()S(c)()
n1 i=0 n1 n1

n1

c ej2i i
k=0

ck ej2k

=
i=0 k=0 n1

c ck ej2(ki) i
n1 n1

=
i=0

c ci i

+2
i=0 k=i+1

c ck ej2(ki) . i

Setting = k i and rearranging summation yields


n1 n 1

|S(c)()| = A()(0) + 2 c
n1

c ci+ ej2 i
=1 i=0

=n+2
=1

A()( )ej2 c

Proof of Theorem 4.7. With Lemma 4.8 and (4.3) we have


n1

|S(a)()| + |S(b)()| = 2n + 2 2n + 2

[A()( ) + A( )]ej2 a b)(


=1 n1

A()( ) + A( ) a b)(
=1

=a b. Since |S(a)()|2 and |S(b)()|2 are nonnegative, the PMEPR of a and b is at most ( a b)/n. The preceding theorem is consistent with the results in [84] in the special case where (, is a Golay complementary pair. Then the PMEPR of a and b is at a b) 52

4.4 Constructions of Near-Complementary Pairs


most 2. Motivated by this fact, several authors constructed OFDM codes that contain exclusively Golay sequences [117], [111], [72], [22]. While these codes ensure a PMEPR not exceeding 2, they also result in rather low code rates when the number of carriers increases. Theorem 4.7 suggests to increase the size of the sequence set at the cost of slightly larger PMEPR by constructing sequence pairs (a, b) Zn Zn with ( a b)/n being close to 2. q q We close this section with an example. Example 4.9. Consider the two quaternary sequences a, b Z8 4 a = (00010210) b = (00230232). The aperiodic autocorrelations of the associated polyphase sequences are given by (A() | 0 a (A( | 0 b) < 8) = (+8, +1 2j, +1 + 3j, +2 + j, 0, +j, +1 + j, +1) < 8) = (+8, 1 + 2j, 1 3j, +2 j, 0, j, 1 j, 1).

The aperiodic autocorrelations cancel at each nonzero shift except at shift 3 (and 3), where the sum is equal to 4. We conclude that a = 24, and hence, the b PMEPR of a and b is at most 24/8 = 3. Direct computation yields PMEPR(a) 2.84 and PMEPR(b) 2.93. Figure 4.1 shows the envelope power of a and b scaled by the length 8. These curves are always equal to or below their sum, whose supremum equals 3, as the proof of Theorem 4.7 claims.

Figure 4.1: Normalized envelope power of a and b given in Example 4.9 (solid lines) and their sum (dashed line)

4.4 Constructions of Near-Complementary Pairs In this section we establish constructions of near-complementary pairs of length 2m . We begin with summarizing some transformations that, when applied to a 53

 

 

 



4 Near-Complementary Pairs and ReedMuller Codes


near-complementary pair, preserve the near-complementary property. These transformations are well known for the case where the pair is a Golay complementary pair, see, e.g., [30], [20], [26]. Lemma 4.10. Let a, b Zn and let (c, d) be a pair of sequences that is obtained q by applying one or more of the following transformations to (a, b). Then c d = a b. (i) interchanging the sequences, (ii) adding a constant Zq to all elements of either sequence, (iii) reversing both sequences, (iv) negating both sequences, (v) reversing one and negating the other sequence, (vi) adding the sequence (0, 1, 2, 3, . . . , n 1) to both sequences. Proof. (i) is trivial. From Lemma 2.12 (v) we know that a and q a have the same auto-correlation, which implies (ii). Setting c = a, we observe that c is a sequence whose elements are the complex-conjugated elements of a. Then (iii), (iv), and (v) follow from Lemma 2.12 (vi). It remains to prove (vi). Write = (0, 1, 2, 3, . . . , n 1), c = a + , and d = b + . Now observe for 0
n 1

A()( ) = c
i=0 n 1

ci+ c i
i+ ai+ q a q i i

=
i=0

n 1

= q

ai+ a i
i=0

= q A()( ), a and analogously, A(d)( ) = q A( ). b)( The above relations can be proved similarly for (4.3). 54 < 0, and (vi) follows from

4.4 Constructions of Near-Complementary Pairs


Next we prove an elementary lemma, which may be viewed as a slightly generalized version of the RudinShapiro construction [88], [97]. This result suggests a recursive construction method for near-complementary pairs. We then introduce some notation, which allows us to translate the recursive construction into an explicit construction using the language of generalized Boolean functions. Lemma 4.11. Suppose x, y Cn and write u = x + y and v = x y. Then u v = 2(x y). Proof. It is straightforward to show A(u)( ) = A(x)( ) + A(y)( ) + C(x, y)( ) + C(y, x)( ) A(v)( ) = A(x)( ) + A(y)( ) C(x, y)( ) C(y, x)( ). The lemma then follows immediately from Denition 4.4. Example 4.12. Let a, b Zn . Also dene the sequences a = ( | 0) and = a b q (0 | of length 2n. Here 0 denotes the all-zero word of length n. Now let c = b) and d = a , which are polyphase sequences of length 2n. Since a +b b A()( ) = A( )( ) and A( ) = A( )( ) for all Z, it follows a = a a a b)( b b b = 2( Therefore, if (, is a Golay complementary and, by Lemma 4.11, c d a b). a b) This recursive construction is also known as the Rudin pair, so will be (, d). c Shapiro construction. If we start with the initial pair a = (0) and b = (0) and apply this process repeatedly, we obtain the RudinShapiro sequences [24], [88], [97]. Next we establish the technique of extending polyphase sequences of length 2 m and their corresponding generalized Boolean functions. It will be used to prove our results on near-complementary pairs. Denition 4.13. Suppose m > k and let f : {0, 1}k Zq be a generalized Boolean function in the variables x0 , x1 , . . . , xk1 . Dene the sets I = {i0 , i1 , . . . , ik1 }

J = {j0 , j1 , . . . , jmk1 } such that I J = and I J = {0, 1, . . . , m1}. Write z = (xj0 , . . . , xjmk1 ) and let d = (d0 , . . . , dmk1 ) be a binary word of length m k. We dene the 55

4 Near-Complementary Pairs and ReedMuller Codes


extended sequence f[z=d] of length 2m as follows. As (u0 , u1 , . . . , uk1 ) ranges k over {0, 1} , at position
k1 mk1

u 2
=0

+
=0

d 2 j
)

k1 the sequence f[z=d] is equal to q 0 1 and equal to zero otherwise. We also dene the extended generalized Boolean function f[z] that is formally regarded as a generalized Boolean function in m variables, i.e., it is of type {0, 1} m Zq . This function is obtained from f by replacing the variables x by xi in the algebraic normal form of f for = 0, 1, . . . , k 1.

f (u ,u ,...,u

With the notation as above, f[z=d] comprises 2k nonzero elements and 2m 2k zeros. It is a consequence of the above denition that at the positions where f[z=d] is nonzero the sequence f[z=d] is equal to the polyphase sequence associated with f[z] . Example 4.14. Let f : {0, 1}2 Z2 be given by f (x0 , x1 ) = x0 x1 + x1 . Writing + for +1 and for 1, we have f = (+ + +). Now take z = (x0 , x2 ) and d = (10). Then f[z=d] = (0+0+000000+0000). By relabeling the variable indices in the algebraic normal form of f according to 0 1, 1 3, we obtain the extended function f[z] = f[z] (x0 , x1 , x2 , x3 ) = x1 x3 + x3 . Regarding this function as a generalized Boolean function in m variables, its associated polyphase sequence reads (+ + + + + + + + + + ++). This sequence is equal to f[z=d] at the positions where f[z=d] is nonzero. Using the language of generalized Boolean functions we now prove an explicit construction for near-complementary pairs. Theorem 4.15. Let m > k and write m k = s + t for nonnegative integers s, t. Dene the sets J = {0, . . . , s 1, m t, . . . , m 1} 56

4.4 Constructions of Near-Complementary Pairs


and I = {0, 1, . . . , m 1}\J. Denote the elements of J and I by 0 and 0 i0 < i1 < < ik1 < m, respectively. Let a, b : {0, 1}k Zq and dene f, g : {0, 1}m Zq by q f (x0 , . . . , xm1 ) = 2
mk2 mk1

j0 < j1 < < jmk1 < m

xj() xj(+1) +
=0 =0

w xj() + w

+ a(xi0 , . . . , xik1 )(1 xj(0) ) + b(xi0 , . . . , xik1 )xj(0) , q g(x0 , . . . , xm1 ) = f (x0 , . . . , xm1 ) + xj(mk1) , 2 where w0 , . . . , wmk1 , w Zq and is a permutation of {0, 1, . . . , m k 1}. Then f g = 2mk ( a b), and PMEPR(f ) a b . 2k

Proof. Write z = (xj(0) , . . . , xj(mk1) ) and let d = (d0 , . . . , dmk1 ) be a binary word of length m k. Consider the extended sequences a [z=d] and [z=d] . b These sequences can be constructed by successively inserting zeros at alternating positions, at the beginning, or at the end of the sequences a and It is then b. straightforward to establish that a[z=d] [z=d] = a b b. Let us dene the recurrence formulas u(+1) (d+1 , . . . , dmk1 ) =
w u() (0, d+1 , . . . , dmk1 ) + q v () (1, d+1 , . . . , dmk1 ) (4.5)

(4.4)

v (+1) (d+1 , . . . , dmk1 ) =


w u() (0, d+1 , . . . , dmk1 ) q v () (1, d+1 , . . . , dmk1 ), (4.6)

57

4 Near-Complementary Pairs and ReedMuller Codes


where = 0, 1, . . . , m k 1 and u(0) (d0 , d1 , . . . , dmk1 ) = a[z=d] v (0) (d0 , d1 , . . . , dmk1 ) = [z=d] . b Since the positions of the nonzero components in u() (0, d+1 , . . . , dmk1 ) () and v (1, d+1 , . . . , dmk1 ) are distinct for each = 0, . . . , m k 1, the sequences u(mk) and v (mk) are polyphase sequences of length 2m . By w Lemma 2.12 (v) we have A(q v () )( ) = A(() )( ) for each Z. Therefore, v by repeated application of Lemma 4.11 and using (4.4), we conclude u(mk) v (mk) = 2mk ( a b). (4.7)

We know that a[z=d] and the polyphase sequence corresponding to a[z] are equal in those positions where a[z=d] is nonzero. An analogous statement holds for [z=d] and b[z] . Similarly we can nd functions, say u() and v () , whose assob ciated polyphase sequences are equal to u() and v () at their respective nonzero components. Using (4.5) and (4.6) it can be veried that the functions u () and v () can be recursively constructed as follows u(+1) = u() (1 xj() ) + (v () + w )xj() q v (+1) = u(+1) + xj() , 2 where u(0) = a[z] = a(xi0 , . . . , xik1 ) and v (0) = b[z] = b(xi0 , . . . , xik1 ). Explicitly we obtain u() = a(xi0 , . . . , xik1 )(1 xj(0) ) + b(xi0 , . . . , xik1 )xj(0) q xj xj + w xj() + 2 =0 () (+1) =0 q = u() + xj(1) . 2
2 1

v ()

Now observe that f = u(mk) + w and g = v (mk) + w . The theorem follows then from (4.7) and Theorem 4.7. Once we have a suitable pair of functions (a, b), where a, b : {0, 1} k Zq , the above theorem can be used to explicitly construct a large family of nearcomplementary sequences and to bound the PMEPR of these sequences. If the 58

4.5 Near-Complementary Pairs and Cosets of RMq (1, m)


pair (a, b) itself cannot be obtained in this way, then it is referred to as a kernel. The transformations given in Lemma 4.10 can be used to generate more kernels from a given one. These are called equivalent kernels, and the maximal set of them is called an equivalence class of a kernel. 4.5 Near-Complementary Pairs and Cosets of RMq (1, m) In this section we establish a more elaborate construction for near-complementary sequences and prove an upper bound for their PMEPR. These sequences organize in cosets of the rst-order generalized ReedMuller code RMq (1, m) inside a higher-order generalized ReedMuller code. In order to state our results, we require some preliminaries. Denition 4.16. Let f : {0, 1}k Zq be a generalized Boolean function. We dene the sequence f of length (4k + 2)/3 as follows. As (u0 , u1 , . . . , uk1 ) ranges over {0, 1}k , at position
k1

u 22
=0 f (u ,u ,...,uk1 ) the sequence f is equal to q 0 1 and equal to zero otherwise.

We remark that the sequence f may also be obtained from the extended se quence f[z=d] by setting z = (x1 , x3 , . . . , x2k3 ), letting d be the all-zero word of length k 1, and deleting the trailing zeros. Denition 4.16 implies that for any generalized Boolean function f : {0, 1} k Zq at its 2k nonzero elements the sequence f coincides with the polyphase sequence associated with a function that is obtained by replacing x by x2 for each = 0, 1, . . . , k 1 in the algebraic normal form of f , where this polynomial is regarded as the algebraic normal form of a generalized Boolean function in 2k 1 variables. Example 4.17. Let f : {0, 1}2 Z4 be given by f (x0 , x1 ) = 2x0 x1 + 3x0 + x1 . We obtain: f = (+1, j, 0, 0, +j, 1). 59

4 Near-Complementary Pairs and ReedMuller Codes


At the nonzero positions the above sequence coincides with the polyphase sequence associated with the function g : {0, 1}3 Z4 whose algebraic normal form is given by g(x0 , x1 , x2 ) = 2x0 x2 + 3x0 + x2 . We are now in a position to state the main theorem of this section. Theorem 4.18. Let m > k and suppose a, b : {0, 1}k Zq are two generalized Boolean functions. Let f, g : {0, 1}m Zq be given by q f (x0 , . . . , xm1 ) = 2
m2 m1

x() x(+1) +
=k =0

w x() + w

+ a(x(0) , . . . , x(k1) )(1 x(k) ) + b(x(0) , . . . , x(k1) )x(k) , q g(x0 , . . . , xm1 ) = f (x0 , . . . , xm1 ) + x(m1) , 2 where w0 , . . . , wm1 , w Zq and is a permutation of {0, 1, . . . , m 1}. Then f g and 2mk ( a b)

a b . 2k We need a series of lemmas in order to prove the theorem. It is well known that each integer 0 i < 2m has a unique binary representation (i0 , i1 , . . . , im1 ), m1 such that i = =0 i 2 and i {0, 1}. If we now allow i {1, 0, +1}, we obtain a signed-digit representation (SDR) of i. Such a representation is in general not unique. An SDR is called sparse (sometimes referred to as nonadjacent form) if it contains no adjacent nonzero entries, i.e., if i i+1 = 0 for all = 0, . . . , m 2. The following lemma is a well-known result in redundant numberrepresentation theory (cf., e.g., [35], [47]). We include a proof for clarity. PMEPR(f ) Lemma 4.19. Every nonzero integer i has a unique sparse SDR (i 0 , i1 , . . . , im1 ) with im1 = 0. Proof. Without loss of generality we assume that i > 0. We rst show that for each nonzero integer there exists a sparse SDR. For i = 1 this is obvious. The remaining cases are proved by induction. There are three cases: (i) if i is even, take a sparse SDR for i/2 and prepend 0, (ii) if i 1 (mod 4), take a sparse SDR for (i 1)/4 and prepend 1, 0, (iii) if i 1 (mod 4), take a sparse SDR for (i + 1)/4 and prepend 1, 0. 60

4.5 Near-Complementary Pairs and Cosets of RMq (1, m)


Now let us prove the uniqueness of a sparse SDR for a given positive integer. We rst show that there is a unique sparse SDR for i = 1 and then proceed by induction. Suppose im1 = 1. Then m1 (m3)/2 2 2m + 1 2 if m is odd 2 = 3 =0 i (m4)/2 m1 2m + 2 2 2+1 if m is even. 2 = 3 =0

If im1 = 1, a similar argument yields i < 0. We conclude that, if i = 1, then m = 1 and i0 = 1. Hence a sparse SDR for i = 1 is unique. Now suppose that i > 1 is the smallest positive integer with two sparse SDRs, namely (u0 , . . . , um1 ) and (v0 , . . . , vn1 ). If i is even, then u0 = v0 = 0. By shifting the two SDRs one position to the left, we obtain two sparse SDRs for i/2 < i, which is a contradiction. If i 1 (mod 4) or i 1 (mod 4), then u0 = v0 = 1 or u0 = v0 = 1, respectively. Consequently, we would have two sparse SDRs for (i u0 )/4 < i. Again we arrive at a contradiction, which completes the proof. Now let f : {0, 1}k Zq be a generalized Boolean function and consider the sequence f . The nonzero elements of f are located at the positions
k1

u 22 ,
=0

where (u0 , . . . , uk1 ) ranges over {0, 1}k . Hence the positions of the nonzero elements in f have sparse SDRs. This property turns out to be crucial in the proofs of the following three lemmas. Lemma 4.20. Suppose a, b : {0, 1}k Zq are two generalized Boolean functions. Then a[z=d] [z=d] a b b holds for any appropriate list of variables z and for any d of suitable length. Proof. Let 0 and write 0 j0 < j1 < < jmk1 < m 61 i0 < i1 < < ik1 < m

4 Near-Complementary Pairs and ReedMuller Codes


for the remaining indices. As usual denote z = (xj0 , . . . , xjmk1 ) and let d = (d0 , . . . , dmk1 ) be a xed binary word of length m k. Also write a = a[z=d] = [z=d] . We claim that and b b A()( ) = a
L( )

A()( ) a

(4.8)

for some disjoint sets L( ), which we shall now prove. Let u and u have binary expansion
k1 mk1

u=
=0 k1

u 2

+
=0

d 2 j

u =
=0

u 22 ,

respectively, and use an analogous notation for v and v . Notice that u and u dene the positions of the nonzero components of a and a, respectively. By Def initions 4.13 and 4.16 we have au = au for any u and corresponding u . Now consider the nonzero product au av = au a with u = v (and therefore u = v) v occurring in the expression A()(u v ) and also in A()(u v). An SDR of a a = u v is given by (u0 v0 , 0, u1 v1 , . . . , 0, uk1 vk1 ). This SDR is sparse and, up to trailing zeros, unique. In other words, for each = 0, 1, . . . , k 1 the differences u v are uniquely determined by . Hence = u v (which is independent of d) is also uniquely determined by . This means that all nonzero products contributing to A()( ) also contribute to A()( ) a a for exactly one , which proves (4.8). Now we use (4.8) to establish
2m 1

a =2 b

k+1

+2
=1 2 1
m

|A()( ) + A( )| a b)( A()( ) + a A( ) b)(


L( )

= 2k+1 + 2
=1 2 1
m

L( )

k+1

+2
=1 L( )

A()( ) + A( ) . a b)(

62

4.5 Near-Complementary Pairs and Cosets of RMq (1, m)


Since the sets L( ) are disjoint and each product contributing to A()( ) for some a also contributes to A()( ) for some , the latter expression is equal to a a b, which completes the proof. Lemma 4.21. Let a, b : {0, 1}k Zq be two generalized Boolean functions and dene the functions c, d : {0, 1}k Zq by c(x0 , . . . , xk1 ) = a(x0 , . . . , xk1 ) + L(x0 , . . . , xk1 ) d(x0 , . . . , xk1 ) = b(x0 , . . . , xk1 ) + L(x0 , . . . , xk1 ), where L(x0 , . . . , xk1 ) =
=0 k1

w x

w0 , . . . , wk1 Zq .

Then c d=a b. Proof. Let u and v have binary representation (u0 , . . . , u2k2 ) and (v0 , . . . , v2k2 ), respectively. For u = v consider the products appearing in A()(u v) c cu c v =u q a =u a a v
Pk1
=0

w u2

a v

Pk1

=0

w v2

Pk1

=0

w (u2 v2 )

Let us focus our attention on the word (u0 v0 , . . . , u2k2 v2k2 ), which is an SDR of u v. We know that, if the product au a is nonzero, u and v have sparse v SDRs. Therefore an SDR of u v is also sparse and, up to trailing zeros, unique. In this case the differences u2 v2 are uniquely determined by the difference K(uv) , where K(u v) only depends on v u v. Hence we have cu c = au a q v u v and not explicitly on u and v themselves. Thus
K( A()( ) = q ) A()( ), c a

and similarly,
K( A(d)( ) = q ) A( ). b)(

We conclude that |A()( ) + A(d)( )| = |A()( ) + A( )| c a b)( for all Z, and the lemma follows. 63

4 Near-Complementary Pairs and ReedMuller Codes


Lemma 4.22. Let a, b : {0, 1}k Zq be two generalized Boolean functions and dene the functions c, d : {0, 1}k Zq by c(x0 , . . . , xk1 ) = a(x(0) , . . . , x(k1) ) d(x0 , . . . , xk1 ) = b(x(0) , . . . , x(k1) ), where is a permutation of {0, 1, . . . , k 1}. Then c d=a b. Proof. Let u and u have binary representation (u0 , u1 , u2 , . . . , u2k3 , u2k2 ) and (u2(0) , u1 , u2(1) , . . . , u2k3 , u2(k1) ), respectively. In an analogous manner we denote the binary representations of v and v . Let us consider the product cu c = au a . By the same reasoning as in v v the proof of Lemma 4.21 we know that, if au a is nonzero, an SDR of u v is v sparse and uniquely determined by u v . Similarly an SDR of u v is sparse v if cu c is nonzero. It is then clear that, if au a is nonzero, there is a one-to-one v correspondence between the sparse SDR of uv and the sparse SDR of u v , and hence, uv is uniquely determined by the difference u v and does not explicitly depend on u and v themselves. This means that each nonzero product appearing in A()(u v ) also appears in A()(u v) and, by setting := 1 and arguing a c analogously, the converse holds too. Hence A()(u v) = A()(u v ) and, c a similarly, A(d)(uv) = A(b)(u v ). The lemma follows then immediately. Proof of Theorem 4.18. Let a set of m k indices be given by {j0 , j1 , . . . , jmk1 }, where 0 j < m for each 0 < m k. Write z = (xj0 , . . . , xjmk1 ) and let d be a binary word of length m k. First we construct a number of initial pairs (, d), where c, d : {0, 1}k Zq are given c by c(x0 , . . . , xk1 ) = a(x(0) , . . . , x(k1) ) + L(x0 , . . . , xk1 ) d(x0 , . . . , xk1 ) = b(x(0) , . . . , x(k1) ) + L(x0 , . . . , xk1 ), 64

4.5 Near-Complementary Pairs and Cosets of RMq (1, m)


is a permutation of {0, 1, . . . , k 1}, and
k1

L(x0 , . . . , xk1 ) =
=0

w x ,

w0 , . . . , wk1 Zq .

From Lemmas 4.204.22 it follows c[z=d] d[z=d] a b. (4.9)

We can now proceed with the same constructive reasoning as in the proof of Theorem 4.15 to obtain the pair ((mk) , v (mk) ) from (, d). Using (4.9) and the u c arguments from the proof of Theorem 4.15 we conclude u(mk) v (mk) 2mk ( a b). (4.10)

Then u(mk) is a function where in its algebraic normal form the addition of all linear terms in m variables is possible. Moreover any permutation can be applied to the indices of the m variables in the algebraic normal form of u(mk) because and the indices j0 , j1 , . . . , jmk1 can be chosen arbitrarily, so that all 0 j < m are distinct. We can, therefore, replace i by () and j by (k + ) and set f = u(mk) + w and g = v (mk) + w . The theorem follows then from (4.10) and Theorem 4.7. We have a useful corollary of Theorem 4.18. Corollary 4.23. Let m > k and suppose a, b : {0, 1}k Zq are two generalized Boolean functions. Then the cosets f + RMq (1, m) with q f (x0 , . . . , xm1 ) = 2
m2

x(i) x(i+1)
i=k

+ a(x(0) , . . . , x(k1) )(1 x(k) ) + b(x(0) , . . . , x(k1) )x(k) , where is a permutation of {0, 1, . . . , m 1}, have PMEPR at most ( a and lie inside RMq (r, m) with max{deg(b a) + 1, deg(a)} if m = k + 1 r= max{deg(b a) + 1, deg(a), 2} if m > k + 1. k b)/2

In particular, if q 4 and deg(f mod 2) r 1, then the cosets are contained in ZRMq (r, m). The number of distinct polynomials f (x0 , . . . , xm1 ), and therefore 65

4 Near-Complementary Pairs and ReedMuller Codes


the number of distinct cosets, is at least (m k)!/2 and at most m! (the true value depends on how many permutations can be applied to the variable indices in the algebraic normal forms of a and b such that distinct pairs (a, b) are generated). The preceding corollary states that, in order to obtain families of cosets of RMq (1, m) with low PMEPR, we just have to nd two generalized Boolean functions a, b : {0, 1}k Zq with ( k being low. This can be accomplished by a b)/2 an exhaustive search or, if k is small enough, such functions can be found using a by hand construction technique. Several examples will be given in Section 4.7. 4.6 Lower Bounds on the PMEPR Lower bounds on the PMEPR of cosets of RMq (1, m) have been investigated in [79], [105], and [61]. However existing results apply only to second-order cosets of RMq (1, m), where the coset representative is binary (in the q-ary context this means that it contains values 0 and q/2). In the following we establish a general m lower bound on the PMEPR of the cosets f + RMq (1, m), where f Z2 is q arbitrary. First, we need some preliminaries. Denition 4.24. With each generalized Boolean function f : {0, 1} m Zq we associate another function Lf : Zm C that is given by q Lf (u) :=
f q (x)ux , x{0,1}m

where denotes the scalar product in Zm . By convention, in the special case where q f m = 0, we dene Lf := q . Remark. The values Lf (u) measure the correlation of the function f with all q m linear functions from {0, 1}m to Zq . Note that, if q = 2, Lf coincides with the WalshHadamard transform of the Boolean function f (see, e.g., [60, Chapter 14]). The following lemma will be useful in the sequel. It is well known for the binary case, i.e., when q = 2 (cf. [60, Chapter 14]). Lemma 4.25. Let f, g : {0, 1}m Zq be two generalized Boolean functions that are related by g(x) = f (Ax) + v x + with v Zm , Zq , and A being an m m permutation matrix. Then the q multisets {Lg (u) | u Zm } and {q Lf (u) | u Zm } are equal. q q 66

4.6 Lower Bounds on the PMEPR


Proof. Write Lg (u) = =
x{0,1}m g(x)ux q x{0,1}m f q (Ax)ux+vx+ .

Setting x = A1 y, we obtain Lg (u) = q = q


y{0,1}m f q (y)(uv)(A y{0,1}m f q (y)(A(uv))y
1

y)

= q Lf (A(u v)), and the lemma follows. Remark. In the case q = 2 the above lemma still holds for A being an arbitrary invertible binary matrix. Next we state a lower bound for the PMEPR of cosets of RM q (1, m). Theorem 4.26. Let f : {0, 1}m Zq be a generalized Boolean function. Then the PMEPR of the coset f + RMq (1, m) is at least 1 max |Lf (u)|2 . 2m uZm q (4.11)

Proof. The coset comprises the words fu , which are associated with f (x) + u x + , where u Zm and Zq . From (2.10) we have q
2m 1

S(fu )(0) =
i=0

fu

,i f q (x)+ux+

=
x{0,1}m

= Lf (u) q . 67

4 Near-Complementary Pairs and ReedMuller Codes


Hence the PMEPR of the coset f + RMq (1, m) can be bounded as follows PMEPR(f + RMq (1, m)) = 1 2m 1 2m 1 = m 2 1 = m 2 max sup |S(fu )()|2
u, 0 <1

max |S(fu )(0)|2


u,

max |Lf (u) q |2


u,

max |Lf (u)|2 ,


u

as it was to be shown. Remark. In the case where q = 2 and f is a quadratic function, it can be shown that the lower bounds in [61, Theorem 1] and [79, Section VI] coincide with that in the preceding theorem. We already know from Corollary 3.13 that the PMEPR of the coset f + RM2 (1, m) is at least 2 if m 7 is odd. Of course, the same is true for cosets of the form (q/2)f + RMq (1, m). If f is a quadratic function, we can say a bit more. Lemma 4.27. Let f : {0, 1}m (q/2) Z2 be a quadratic generalized Boolean function. Then we have min max |Lf (u)|2 = 2m+1 m
f uZq

if q 2 (mod 4) and m is odd or if q 0 (mod 4) and m is arbitrary. Proof. We begin with recalling some well-known facts concerning the weight distribution of second-order cosets of RM2 (1, m) (cf. [60, Chapters 14 and 15]). Suppose Q : {0, 1}m Z2 is a quadratic form, which is generally given by Q(x0 , x1 , . . . , xm1 ) =
0 i<j<m

cij xi xj ,

cij Z2 .

With such a form we associate a matrix A(Q) over F2 of size m m as follows. A(Q) contains cij at positions (i, j) and (j, i) while the elements in the main diagonal are equal to zero. The rank of A(Q) is always even (which is a consequence of Dicksons theorem [60, Chapter 15, Theorem 4]) and completely determines the weight distribution of the coset Q + RM2 (1, m) [60, Chapter 15, Theorem 5]. 68

4.6 Lower Bounds on the PMEPR


Suppose that A(Q) has rank 2t. Then the weight distribution of Q + RM 2 (1, m) is given by: weight number of words m1 mt1 2 2 22t 2m+1 22t+1 2m1 2m1 + 2mt1 22t Now write gu (x) = u x + , where u Zm and Z2 . Then the words in 2 the coset Q + RM2 (1, m) correspond to the polynomials Q(x) + gu (x), and their weights are given by wt(Q + gu ) = 2m1 = 2m1 1 2 1 2 (1)Q(x)+gu
x{0,1}m (x)

(1)Q(x)+ux+
x{0,1}m

1 = 2m1 (1) LQ (u). 2 Hence the distribution of LQ (u) reads: value +2mt 0 2mt Now we prove the inequality min max |Lf (u)|2 m
f uZq

frequency 22t1 2m 22t 22t1

2m+1

(4.12)

under the constraints in the lemma. Let f (x) = (q/2) Q(x). First suppose that m is odd. Then the rank of A(Q) is at most m 1. Hence min max |Lf (u)|2 m
f uZq

min
f

u(q/2) Zm 2

max

|Lf (u)|2

= min max |LQ (u)|2 m


Q uZ2

2m+1 , which proves (4.12) for odd m. Now suppose that m is even and q 0 (mod 4). 69

4 Near-Complementary Pairs and ReedMuller Codes


We have Lf ((q/4) u) = =
x{0,1}m f q (x)+(q/4) ux x{0,1}m 2Q(x)+ux

Suppose that u Zm has 2-adic expansion u = 2v + w, where v, w {0, 1}m. 4 Then the 2-adic expansion of the polynomial 2Q(x) + u x (over Z 4 ) reads 2Q(x) + (2v + w) x = 2Q(x) + 2 v x + w x where Q (x) = Q(x) +
0 i<j<m

= 2Q (x) + 2 v x + (w x mod 2), w i w j xi xj .

Now write {Lf ((q/4) u)} = = = and similarly, {Lf ((q/4) u)} = = = Therefore

x{0,1}m

(1)Q (x)+vx 4

wx

(mod 2)

x{0,1}m

(1)Q (x)+vx

1 [1 + (1)wx ] 2

1 [LQ (v) + LQ (v + w)] 2

x{0,1}m

(1)Q (x)+vx 4

wx

(mod 2)

x{0,1}m

(1)Q (x)+vx

1 [1 (1)wx ] 2

1 [LQ (v) LQ (v + w)] . 2


2 2

|Lf ((q/4) u)| = [ {Lf ((q/4) u)}] + [ {Lf ((q/4) u)}] 1 1 2 2 = [LQ (v)] + [LQ (v + w)] 2 2 1 2 [LQ (v)] . 2 70

4.6 Lower Bounds on the PMEPR


The rst row of A(Q ) reads (0 c01 + w0 w1 c02 + w0 w2 c0m1 + w0 wm1 ).

It is now easy to see that there exists a w {0, 1}m such that this row is zero and, since A(Q ) has even rank, the rank of A(Q ) is at most m 2. Hence for any Q there exist v, w {0, 1}m such that |LQ (v)| 2m/2+1 . Therefore min max |Lf (u)| m
f uZq 2

min max |Lf ((q/4) u)| m


f uZ4

1 min max m [LQ (v)]2 2 Q v,w{0,1} 2m+1 , which proves (4.12) for even m and q 0 (mod 4). By Theorem 4.26 and Corollary 4.30 there exist functions f for which the lower bound in (4.12) is attained, hence, equality holds in (4.12). The preceding lemma together with Theorem 4.26 implies that for any quadratic function f : {0, 1}m Z2 the PMEPR of the coset (q/2) f + RMq (1, m) is at least 2 if q 2 (mod 4) and m is odd or if q 0 (mod 4) and m is arbitrary. This bound is well known when m is odd [79, Section VI]. When m is even, the bound is new. We have proved in particular: Corollary 4.28. The PMEPR of any coset of RM2 (1, m) inside RM2 (2, m) with m being odd and the PMEPR of any coset of RM4 (1, m) inside ZRM4 (2, m) is at least 2. In what follows we apply Theorem 4.26 to obtain lower bounds on the PMEPR of the cosets of RMq (1, m) identied in Corollary 4.23. Theorem 4.29. Let a, b : {0, 1}k Zq be two generalized Boolean functions, and let f : {0, 1}m Zq be given by q f (x0 , . . . , xm1 ) = 2
m2

x(i) x(i+1)
i=k

+ a(x(0) , . . . , x(k1) )(1 x(k) ) + b(x(0) , . . . , x(k1) )x(k) , 71

4 Near-Complementary Pairs and ReedMuller Codes


where is a permutation of {0, 1, . . . , m 1}. Then the PMEPR of the coset f + RMq (1, m) is at least 1 2k+2 uZk ,v,wZq q if m k is even and at least 1 if m k is odd. Proof. We intend to nd a lower bound for 1 max |Lf (u)|2 . m uZm 2 q Using Lemma 4.25 we conclude that, in order to prove the theorem, it is sufcient to assume that is the identity permutation. Hence we are interested in the coset representative corresponding to fm (x0 , . . . , xm1 ) = q 2
m2

max

w v w La (u)(1 + q ) + Lb (u) q (1 q )

2k+1 uZk ,vZq q

max

v La (u) + Lb (u) q

i=k

xi xi+1 + a(x0 , . . . , xk1 )(1 xk ) + b(x0 , . . . , xk1 )xk .

We require the following expansion Lfm (u, v) = =


y{0,1}m1 f q m (x)(u,v)x x{0,1}m f v q m (y,0)uy + q y{0,1}m1 f q m (y,1)uy ,

(4.13)

m1 and v Zq . Let us rst consider the case m = k + 1, so m k where u Zq is odd. We have

fk+1 = a(x0 , . . . , xk1 )(1 xk ) + b(x0 , . . . , xk1 )xk , and thus fk+1 (x0 , . . . , xk1 , 0) = a(x0 , . . . , xk1 ) fk+1 (x0 , . . . , xk1 , 1) = b(x0 , . . . , xk1 ). 72

4.6 Lower Bounds on the PMEPR


With (4.13) we obtain
v Lfk+1 (u, v) = La (u) + Lb (u) q .

(4.14)

Now let m > k + 1. Then we have fm (x0 , . . . , xm2 , 0) = fm1 (x0 , . . . , xm2 ) q fm (x0 , . . . , xm2 , 1) = fm1 (x0 , . . . , xm2 ) + xm2 , 2 and with (4.13) it follows that q v Lfm (u, v) = Lfm1 (u0 , . . . , um2 ) + Lfm1 (u0 , . . . , um2 + ) q . (4.15) 2 For m k even suppose Lfm (u, v, 0, . . . , 0, w) =
w v w 2(mk2)/2 La (u)(1 + q ) + Lb (u) q (1 q ) , (4.16)

which is true for m = k + 2, by (4.14) and (4.15). We will use this expression as a hypothesis for the following induction. We employ (4.15) to obtain
vw Lfm+1 (u, v, 0, . . . , 0, w) = 2(mk)/2 La (u) + Lb (u) q

(4.17)

and Lfm+2 (u, v, 0, . . . , 0, w) =

w v w 2(mk)/2 La (u)(1 + q ) + Lb (u) q (1 q ) ,

which proves by induction that (4.16) and (4.17) hold in general. Now we can write for m k even 1 max Lfm (u) 2m uZm q
2

1 2 Lfm (u, v, 0, . . . , 0, w) max m uZk ,v,wZ 2 q q 1 w v w = k+2 max La (u)(1 + q ) + Lb (u) q (1 q ) k ,v,wZ 2 uZq q 1 2 max Lfm (u, v, 0, . . . , 0) m uZk ,vZ 2 q q 1 v 2 La (u) + Lb (u) q . = k+1 max k ,vZ 2 uZq q

and for m k odd

1 max Lfm (u) 2m uZm q

Now the statements in the theorem follow from Theorem 4.26. 73

4 Near-Complementary Pairs and ReedMuller Codes


4.7 Applications and Construction Examples In this section we will apply Corollary 4.23 in order to nd families of cosets of RMq (1, m) with low PMEPR. Trivially we may choose the functions a and b in Corollary 4.23 from the complete set of generalized Boolean functions in k variables. In such a way, we can construct large sequence families with their PMEPR bounded by 2k+1 , since the trivial bound states a b 2k+1 . 2k However it can be shown that these sequence sets are subsets of a larger family of complementary sets of size 2k+1 , which will be constructed in Chapter 5. The GolayDavisJedwab Set An immediate consequence of Corollary 4.23 is the well-known construction of Golay sequences over Zq , which has been stated by Davis and Jedwab in [22] for q being a power of 2 and generalized by Paterson [79] to arbitrary (even) q. Corollary 4.30. Each of the m!/2 cosets of RMq (1, m) with coset representatives corresponding to m2 q x(i) x(i+1) , 2 i=0 where is a permutation of {0, 1, . . . , m 1}, is entirely comprised of Golay sequences. In particular the PMEPR of these cosets is (i) exactly 2 if m is odd and q arbitrary (even) and if m is even and q 0 (mod 4), (ii) at least 1 + cos2
q

and at most 2 if m is even and q 2 (mod 4).

Proof. The corollary follows by setting k = 0, a = 0, and b = 0 in Corollary 4.23 and Theorem 4.29. Since a = 2, the upper bound on the PMEPR is b an immediate consequence of Corollary 4.23, which also implies that the cosets are comprised of Golay sequences. Note that La = Lb = 1. Then for odd m Theorem 4.29 immediately yields a lower bound of 2 for the PMEPR of the cosets. It remains to prove the lower bounds for even m. By Theorem 4.29, we need to evaluate 1 w v w 2 max (1 + q ) + q (1 q ) . 4 v,wZq 74

4.7 Applications and Construction Examples


For q 0 (mod 4) putting w = v = q in the preceding expression yields a lower 4 bound of 2. For q 2 (mod 4) it is not hard to see that the maximum is attained when w = q+2 and v = q2 . Hence 4 4 1 w v w 2 max (1 + q ) + q (1 q ) 4 v,wZq 1 = (1 + jej/q ) + jej/q (1 jej/q ) 4 2 1 2 + jej/q + jej/q = 4 2 1 = 2 + 2j ej/q 4 = 1 + cos2 , q which completes the proof. We remark that the upper bound is well known: it was rst proved by Davis and Jedwab for q being a power of 2 [22, Corollaries 6 and 9] and generalized to arbitrary q by Paterson [79, Corollary 11]. The lower bound in Corollary 4.30 (i) has been rst proved Cammarano and Walker [12], and the case when m is odd also by Paterson [79, Theorem 21]. Except for the trivial bound 1 when q = 2, the lower bound in part (ii) of Corollary 4.30 is new. Note that these upper and lower bounds now arise in a uniform way from a general framework. In the light of Corollary 3.13 we see that the PMEPR of cosets of RM2 (1, m) in the preceding corollary is the best possible at least for odd m 7. Moreover, by Corollary 4.28, the PMEPR of these cosets is the best possible among all cosets lying in RM2 (2, m) when m is odd and among all cosets lying in ZRM4 (2, m). Construction of More Golay Sequences of Length 2m Li and Chu [56] reported 1024 additional quaternary (with elements in {+1, 1, +j, j}) Golay sequences of length 16. Using search results by Holzmann and Kharaghani [42] and Theorem 4.15, these sequences can be explained in a natural way, as we will show next. It has been observed in [42] that there exists essentially one quaternary kernel of length 8 ((00020020) , (01120332)). 75

4 Near-Complementary Pairs and ReedMuller Codes


This pair corresponds to the generalized Boolean functions {0, 1} 3 Z4 with algebraic normal forms 2x0 x1 + 2x1 x2 and 2x0 x2 + 2x1 x2 + x0 + x1 , (4.18)

respectively. Using Lemma 4.10 from this kernel its equivalence class, containing 512 equivalent kernels, can be generated. These pairs can be used in conjunction with Theorem 4.15 to obtain an explicit construction for Golay sequences of length 2m , where m > 3. Notice that all of them correspond to cubic forms, so they are distinct from the Golay sequences identied in Corollary 4.30. Indeed, if m = 4, we obtain exactly the 1024 Golay sequences of length 16 reported by Li and Chu in [56]. Unfortunately, in this way, we cannot construct complete cosets of RMq (1, m) whose PMEPR is bounded by 2. Instead Corollary 4.23 reveals that the PMEPR is at most 5 if the 512 kernels are used in conjunction with Corollary 4.23 to build cosets of RMq (1, m). We remark that, recently, Fiedler and Jedwab [25] provided an alternative explanation of these 1024 Golay sequences by using some observations by Davis and Jedwab in [22]. Most notably, in [26], Fiedler, Jedwab, and Parker demonstrated how the kernel in (4.18) can be used to construct new complementary pairs, which cannot be obtained with any previously known construction. Further Cosets of RMq (1, m) with Low PMEPR Consider the functions a, b : {0, 1}2 Zq with a(x0 , x1 ) = q x0 x1 2 q q b(x0 , x1 ) = x0 x1 + + x0 + x1 , 2 2

where , (q/p)Zp and p divides q. We have a = (1, 1, 0, 0, 1, 1) = (1, , 0, 0, , + ) b


q q q

and (A()( ) | 0 a (A( ) | 0 b)( 76 < 6) = (4, 0, 0, 1, 0, 1)

+ < 6) = (4, 0, 0, q , 0, q ).

4.7 Applications and Construction Examples


Hence
+ a = 2(4 + |1 q | + |1 q |) b

16.

Therefore we conclude from Corollary 4.23: Corollary 4.31. For m > 2 the cosets of RMq (1, m) with coset representatives of the form m2 q x(i) x(i+1) + x(0) x(2) + x(1) x(2) , (4.19) 2 i=0 where , (q/p)Zp , p is a divisor of q, and is a permutation of {0, 1, . . . , m 1}, have PMEPR at most 2+
+ |1 q | |1 q | + 2 2

4.

The cosets are contained in RMq (2, m). In particular, if q > p and q/p is even, the cosets lie inside ZRMq (2, m). In the preceding corollary there are p2 choices for the pair (, ) and for each pair we can construct m!/2 distinct cosets by applying to the variable indices m!/2 permutations that are invariant under reversal (e.g., all permutations that satisfy (0) < (m 1)). However setting = = 0 or = = q/2 in (4.19) will generate the same set of coset representatives. Hence the total number of distinct cosets identied by Corollary 4.31 equals p2 (m!/2) if p is odd or (p2 1)(m!/2) if p is even. The PMEPR of all of them is bounded by 4. For p {2, 4, 8} the distribution of the PMEPR bounds, according to Corollary 4.31, is given below. If = = 0 or = = q/2, we obtain cosets with PMEPR at most 2. We remark that in this case the coset representatives are of the form given in Corollary 4.30. More generally, for a given p the cosets arrange in classes with the same PMEPR upper bound. For p = 2 we obtain just two classes: m!/2 cosets with maximum PMEPR 2 and in addition m! cosets with PMEPR at most 4. With Theorem 4.29 it can be shown that these bounds are tight if q is a multiple of 4. q q For q 4 and p = 4 we have , 0, 4 , 2 , 3q and the classes are: 4 + 0 arbitrary (0) q q 4 (mod 2 ) arbitrary 0 0 (arbitrary) q q 4 (mod 2 ) arbitrary max. PMEPR 2 3 2+ 2 4 # cosets m!/2 2m! 4m! m! 77

4 Near-Complementary Pairs and ReedMuller Codes


Notice that the number of cosets are counted by considering only those cosets that are not contained in a class with a lower PMEPR bound. For example the total number of cosets with PMEPR at most 3 is equal to 5m!/2, while there are 13m!/2 cosets with PMEPR at most 2 + 2. With Theorem 4.29 it can be shown that the PMEPR bounds given above are tight if q is a multiple of 8. If q is equal to 4, Theorem 4.29 yields a lower bound of 2 for the rst class, 2.5 for the second and third class, and 4 for the fourth class. The second class includes those 48 cosets of RM 8 (1, 4) inside ZRM8 (2, 4) for which Davis and Jedwab observed that their PMEPR is exactly 3 [22, Table V]. While Nieswand and Wagner [70] could explain this behavior partially, a complete proof arises as a special case of Corollary 4.23 and Theorem 4.29. Moreover we identied further cosets of RMq (1, m) with PMEPR at most 3. For q 8 and p = 8 there are already 9 classes with different PMEPR upper bounds. These are: max. PMEPR 2 + 1/ 2 2+ 2 2 2+ 1 + 1/ 2 2 + 2 3 + 1/ 2 2+ 2+ 2 2 2.707 2.765 3 3.307 3.414 3.707 3.848 4 # cosets m!/2 4m! 4m! 2m! 8m! 4m! 4m! 4m! m!

Theorem 4.29 can be used to show that the PMEPR bounds in the preceding table are tight if q is a multiple of 16. The number of cosets of RMq (1, m) with PMEPR at most 4 can be increased further. To this end, consider the pairs (a, b), where a, b : {0, 1} 2 Zq are given by a(x0 , x1 ) = x0 x1 b(x0 , x1 ) = x0 x1 + + q x0 + x1 , 2

, , , Zq , and and are selected such that a b 16. Here the coset representatives of the constructed cosets of RMq (1, m) generally correspond to cubic polynomials. 78

4.7 Applications and Construction Examples


It is noteworthy that a computer-based search failed to nd kernels that generate further cosets of RMq (1, m) with PMEPR upper bounds less than 4. Our search included all functions in k variables for k = 3 and q = 4, k = 4 and q = 2, and quadratic functions with k = 5 and q = 2.

OFDM Codes with Low PMEPR Corollary 4.31 suggests a variety of coding schemes for OFDM with low PMEPR. Since all of the cosets identied in Corollary 4.31 are contained inside RM q (2, m), the minimum distance of any union of them is at least 2m2 . If only cosets inside ZRMq (2, m) are considered, the minimum distance can be increased to 2 m1 . As already described in [22] and [79], encoding of such codes can be performed by storing a list of coset representatives and using the information bits partly to select one of the coset representatives and partly to encode one of the q m+1 words of RMq (1, m). Therefore, it is convenient to choose the number of coset representatives to be a power of 2. Also, in practice, q is typically a power of 2. Efcient decoding algorithms for such codes are discussed in the next section. Tables 4.1, 4.2, 4.3 contain selected coding options for BPSK, QPSK, and 8PSK constellations (i.e., q = 2, 4, 8), respectively. Here n = 2m is the length of the codes. Options B1, Q1, and O1 consist of 2 log2 m!/2 cosets of RMq (1, m) comprised of Golay sequences. They are exactly those described by Davis and Jedwab in [22] (cf. also Corollary 4.30). In [22] Davis and Jedwab performed a computational search for cosets of RMq (1, m) with low PMEPR for n = 16. Based on this search, coding schemes with improved rate but PMEPR slightly larger than 2 were derived. These coding options can now be explained in a common framework using more sophisticated and generally applicable techniques. The considered codes are Options B2, Q4, and O2 for n = 16, whose parameters appear in [22, Tables VIIIX], respectively. We remark that Options B2 and Q4 outperform Options 4a and 4c in Table 5.1 for n = 16, however, Options B2 and Q4 are inferior to them for n 128. For n = 32 and 64 Options B2 and Q4 have the same parameters as Options 4a and 4c in Table 5.1, respectively. It should be noted as well that [22, page 2405] reports that 29 out of the 210 cosets of RM2 (1, 5) inside RM2 (2, 5) have PMEPR not greater than 4, which suggests a code with rate 15/32 and minimum distance 8. This code outperforms Option B2 for n = 32. 79

4 Near-Complementary Pairs and ReedMuller Codes

Table 4.1: Binary Coding Options Option B1 B2 B1 B2 B1 B2 n 16 32 64 # info bits 8 10 11 13 15 17 code rate 0.50 0.63 0.34 0.41 0.23 0.27 min. dist. 4 4 8 8 16 16 max. PMEPR 2 4 2 4 2 4

Table 4.2: Quaternary Coding Options Option Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 n 16 # info bits 13 15 17 15 17 20 21 19 22 24 26 24 code rate 0.81 0.94 1.06 0.94 0.53 0.63 0.66 0.59 0.34 0.38 0.41 0.38 min. dist. 8 4 4 8 16 8 8 16 32 16 16 32 max. PMEPR 2 3 2+ 2 4 2 3 2+ 2 4 2 3 2+ 2 4

32

64

4.8 Decoding of Unions of Cosets of RMq (1, m) Overview of Existing Approaches The majority of existing approaches to decode a code that consists of a union of cosets of RMq (1, m) relies on the supercode scheme, which was described for binary codes in [19]. This scheme is generally applicable to an arbitrary union of, say M , cosets and works as follows. The received word is multiplied componentwise with the complex-conjugated versions of the M coset representatives, and the resulting words are fed into a decoder for RMq (1, m). Finally, among the M decoding results, one selects the word that is closest to the received word, from which the nal decoding result can be determined. Hence an efcient decoder for 80

4.8 Decoding of Unions of Cosets of RMq (1, m)

Table 4.3: Octary Coding Options Option O1 O2 O3 O4 O5 O1 O2 O3 O4 O5 O1 O2 O3 O4 O5 n 16 # info bits 18 20 22 22 24 23 26 28 27 29 29 31 33 33 34 code rate 1.13 1.25 1.38 1.38 1.50 0.72 0.81 0.88 0.84 0.91 0.45 0.48 0.52 0.52 0.53 min. dist. 8 8 4 8 4 16 16 8 16 8 32 32 16 32 16 max. PMEPR 2 3 3 2 + 2 2+ 2 2 3 3 2 + 2 2+ 2 2 3 3 2 + 2 2+ 2

32

64

RMq (1, m) is required. Moreover, if the algorithm used to decode RM q (1, m) is a maximum-likelihood (ML) decoder, the supercode method performs ML decoding too. This follows by a simple adaption of the reasoning in [19] to the q-ary case. Several suboptimal decoders for RMq (1, m) have been described in [22], [32], [71], [82]. Also a general algorithm is applicable that relies on lifting a decoder for linear codes over Zp to decode related linear codes over Zph [34]. This latter method suggests to use any decoder for the classical ReedMuller code RM2 (r, m) to obtain a (suboptimal) decoder for RM2h (1, m). An ML decoder for RM4 (1, m) has been proposed in [49], while [33] (together with a correction by the author [89]), and the authors work [94] contain ML decoders for RM q (1, m). These latter two schemes will be discussed in this section. The complexity of the supercode decoder is roughly M times the complexity of the decoder for RMq (1, m). So the total complexity becomes high if M is large. For this reason Paterson and Jones [82] developed a variety of suboptimal decoding methods that are not directly based on the supercode scheme and benet from a reduced complexity for large M . However these algorithms do not achieve ML decoding performance. Reference [82] contains also a comparison of several decoding strategies for unions of cosets of RM q (1, m). 81

4 Near-Complementary Pairs and ReedMuller Codes


Maximum-Likelihood Decoding of RMq (1, m) Let c RMq (1, m). We assume that c is transmitted and we receive y = c + e, where e C2 is an error vector. As described in Section 2.4, an ML decoder nds the codeword c = argmin ||y c||.
cRMq (1,m)
m

We shall call c the maximum-likelihood codeword of y. It is straightforward to verify that ||y c||2 = ||y||2 + ||||2 2 {y c}, c where denotes the scalar product in C2 . Notice that ||y||2 and ||||2 are indec 2 m pendent of the decoding result since |||| = 2 for each c RMq (1, m). Hence c c= argmax
cRMq (1,m)
m

{y c}.

By virtue of the denition of RMq (1, m), there exists a u = (u0 , u1 , . . . , um1 ) Zm and an Zq such that the ith entry of c RMq (1, m) q is given by
m1

ci =
i=0

uj i j + ,

where (i0 , i1 , . . . , im1 ) is the binary expansion of the integer i. Hence c can be obtained by estimating u and as follows
2m 1

u, = argmax
uZm , Zq q i=0 q

y i q
2m 1

Pm1
i=0

uj i j

= argmax
uZm , Zq q

y i q

Pm1
i=0

uj i j

i=0 q

= argmax
uZm , Zq q

Y (u) ,

where Y (u) :=

2m 1 i=0

y i q

Pm1
i=0

uj i j

u Zm . q

82

4.8 Decoding of Unions of Cosets of RMq (1, m)


Grant and van Nee [33] provided a fast algorithm to compute the multiset {Y (u) | u Zm }, which was then applied to obtain a decoder for RMq (1, m). q However a minor mistake in [33] distinguishes this algorithm from an ML decoder. The correct version is given below. Algorithm 4.32. (Decoder for RMq (1, m), corrected version of [33, Algorithm 1]) 1) Input y C2 . 2) Compute the multiset {Y (u) | u Zm } using the algorithm described in [33]. q
3) Determine u Zm and Zq such that {q Y ()} is a maximal value in u q m the set { {q Y (u)} | u Zq , Zq }.
m

4) Output the decoded codeword whose ith component is equal to The reasoning preceding Algorithm 4.32 implies: Theorem 4.33. Algorithm 4.32 is an ML decoder for RMq (1, m).

m1 i=0

uj ij +.

In what follows we propose a new ML decoder for RM q (1, m), which has slightly lower complexity than Algorithm 4.32. Algorithm 4.34. (Decoder for RMq (1, m)) 1) Input y = (y0 , y1 , . . . , y2m 1 ) C2 . 2) If m = 1, output the hard decisions c (obtained by rounding the entries in y to the nearest constellation points) and {y c}. Stop. 3) For v = 0, 1, . . . , q 1 calculate z(v) = (z0 (v), z1 (v), . . . , z2m1 1 (v)) with
v zk (v) = yk + yk+2m1 q ,
m

and use the algorithm to decode z(v) to z (v) and to calculate p(v) = {z(v) z (v)}. 4) Determine v = argmax p(v).
vZq

83

4 Near-Complementary Pairs and ReedMuller Codes


5) Output the decoded polyphase codeword c = (()|()q ) and p(). z v z v v v

Theorem 4.35. Algorithm 4.34 is an ML decoder for RM q (1, m). Proof. We will show that, if Algorithm 4.34 is an ML decoder for RMq (1, m 1), then it is also an ML decoder for RMq (1, m). The case m = 1 serves as the anchor for this induction since RMq (1, 1) is the space Z2 and the theorem holds trivially. q Now let m > 1. We have to show that the algorithm always outputs the polyphase version of the ML codeword c and {y c}, which is potentially re quired for the higher stage of the decoder. From the (u | u + v) construction of RMq (1, m) (cf. (4.1)) we know that, if c RMq (1, m), then c = (| q ), where uu v u RMq (1, m 1) and v Zq . By computing u, v = we nd v c = (u|u q ). Let us inspect the real part of the scalar product in (4.20)
2m1 1

argmax
uRMq (1,m1),vZq

{y (| q )}, uu v

(4.20)

(4.21)

{y (| q )} = uu v

k=0

v uk {(yk + yk+2m1 q ) }

= {z(v) u},

(4.22)

where z(v) is as calculated in Step 3) of Algorithm 4.34. By hypothesis, Algorithm 4.34 is an ML decoder for RMq (1, m 1), which is now used to get z (v) according to z (v) = argmax
uRMq (1,m1)

{z(v) u}.

(4.23)

Moreover the algorithm provides p(v) = {z(v) z (v)}. In order to nd u and v , it remains to compute v = argmax p(v),
vZq

which is done in Step 4) of Algorithm 4.34. Then we have u = z () and v , v can be obtained by (4.21). We also have, by (4.22) and (4.23), from which c p() = {() z ()} = {y c}. v z v v 84

4.8 Decoding of Unions of Cosets of RMq (1, m)


We have several remarks on Algorithm 4.34. The code RMq (1, m) itself is a union of q cosets of the code {(u | u) | u RM q (1, m 1)} with coset representatives {(0, . . . , 0, v, . . . , v) | v Zq }. Then Algorithm 4.34 basically performs the steps of the supercode decoding method, as described in [19] for binary codes. We also mention the relationship of Algorithm 4.34 and the list-decoding scheme, proposed in [58]. In this reference it was shown that an ML decoder for RM 2 (1, m) can be obtained when the list length is equal to 2. The basic idea behind Algorithm 4.34 can also be interpreted as a generalization of the list-decoding method in [58] to q-ary codes, where the list length equals q.

Complexity Analysis The complexity of Algorithm 4.32 is essentially given by the complexity of the algorithm that computes Y (u) from the input y, which was analyzed in [33]. In what follows we examine the complexity of Algorithm 4.34 and make comparisons with Algorithm 4.32 and the ML decoder for RM 4 (1, m) in [49]. In practice q is typically a power of 2. So we restrict our analysis to the case q = 2h . We consider the number of complex additions N + (2h , m) and the number of complex multiplications N (2h , m) as complexity measurements. We regard one real addition as half a complex addition. Multiplications with 1, j, 1, j are not counted since they are trivial and can be implemented by sign-bit changes and swapping of the real and imaginary part. We start with m = 1. Then in Step 2) we need to determine the hard decisions and the scalar product of two vectors of length 2. The hard decision for the ith c symbol is ci = 2i , where h
ci = argmax {yi 2h }. Z2h

(4.24)

For h 2 the above calculation is a trivial task and can be accomplished with simple sign logic. For h > 2 we may implement this maximum calculation as follows. First determine the quadrant in which the received value lies. This leaves 2h2 + 1 candidates, over which the maximum in (4.24) has to be calculated. For each candidate one complex multiplication is required. Since one of the possible signal points lies on the real and another on the imaginary axis (where the multiplication is trivial), we need for each of the two hard decisions 2h2 1 complex 85

4 Near-Complementary Pairs and ReedMuller Codes


multiplications. Hence N (2h , 1) = 0, h 2

Having {yi 2hci } for i = 0, 1, one real addition is required to calculate the scalar product in Step 2). Hence N + (2h , 1) = 1/2.

2h1 2, h > 2.

Now let m > 1. In Step 3) we require 2h 2m1 complex multiplications to calculate zk (v) for k = 0, . . . , 2m1 1 and v = 0, . . . , 2h 1. If h 2, these are trivial multiplications. For h > 2, four out of the 2h complex multiplications for v each k are trivial. Moreover, if we have computed yi 2h for v = 0, 1, . . . , 2h1 v 1, by sign-bit changes, the values yi 2h for v = 2h1 , . . . , 2h 1 can be easily v derived. After having the products yi 2h , there are 2h 2m1 additions required in Step 3). These are real additions if h = 1 and complex additions if h > 1. Finally we need to perform 2h decodings of RM2h (1, m 1). Hence in total we have 0, h 2 N (2h , m) = (2h1 2)2m1 + 2h N (2h , m 1), h > 2 0, h 2 = (2h1 2)(2hm 2m ) , h > 2, h
2 2

and

N + (2h , m) =

Table 4.4 contains the number of complex operations required to decode RMq (1, m) for different m and q = 2h . It also compares the complexity of Algorithm 4.34 with the complexity of Algorithm 4.32, which is given in [33]. Notice that for h = 1 only real additions are required, and its number is equal to half the indicated value. It is apparent that Algorithm 4.34 has slightly lower complexity 86

2h 2m1 + 2h N + (2h , m 1) h > 1 (2m 1)2m2 h=1 = hm1 h(m1) h+m 2 2 52 h > 1. h
2 2

2 2m2 + 2N + (2, m 1)

h=1

4.9 Conclusions, Notes, and Open Problems

Table 4.4: Number of Complex Operations Required to Decode RM q (1, m) m 4 q 2 4 8 16 2 4 8 16 2 4 8 16 Algorithm 4.34 N+ N 28 0 256 0 1600 1360 11392 28080 72 0 1088 0 12928 10912 182528 449376 176 0 4480 0 103680 87360 2920960 7190208 Algorithm 4.32 N+ N 32 0 480 0 5440 2720 74880 56160 80 0 1984 0 43648 21824 1198336 898752 192 0 8064 0 349440 174720 19173888 14380416

than Algorithm 4.32. To be precise, Algorithm 4.34 requires exactly one half of the multiplications required for Algorithm 4.32. For h > 1, the ratio of the number of complex additions required for Algorithm 4.34 and Algorithm 4.32 is approximately (2h+1 + 2h1 1)/22h and converges to this value when m approaches innity. We nally remark that the ML decoder for RM 4 (1, m) in [49] requires (m+1)22m+1 real additions. Hence it has higher complexity than Algorithm 4.34. 4.9 Conclusions, Notes, and Open Problems In this chapter we have shown how cosets of RMq (1, m) with low PMEPR can be obtained in a systematic way once suitable kernels are known. The PMEPR of these cosets is low because such cosets are comprised of sequences lying in near-complementary pairs. It was demonstrated that those cosets of RM q (1, m) comprised entirely of Golay sequences arise as a special case in a more general theory. We presented some simple kernels and, in this way, several previously unexplained phenomena arising in earlier works by Davis and Jedwab [22], by Paterson [79], and by Parker and Tellambura [76], [77] can now be understood within a general framework. As one example, we exhibited 2m! cosets of RM 8 (1, m) with PMEPR of exactly 3, and therefore, we have proved a conjecture by Davis and Jedwab [22]. Moreover we provided at least a partial answer to the question 87

4 Near-Complementary Pairs and ReedMuller Codes


stated in [79] (see also [80, Open Problem 3]): What other regularities appear in the PMEPRs of cosets as we move to higher alphabets, and how can they be explained in general? Our results led to OFDM coding schemes with low PMEPR, which allow a trade-off between the rate of the code and its PMEPR. Finally we presented two maximum-likelihood decoding algorithms for RM q (1, m), which can be used to decode any union of cosets of RMq (1, m). One of the algorithms corrects an earlier published version from [33], the other appears to be new and has lower complexity than any previously known maximum-likelihood decoder for RMq (1, m). Next we wish to point out some relations to the work by Parker and Tellambura [76], [77]. Indeed Theorem 4.15 is similar to [77, Theorem 5] and [76, Theorem 6], which also exploit the RudinShapiro construction. However these references do not identify the crucial connection between the PMEPR and the aperiodic autocorrelation, which allows us to extend the PMEPR bound to a slightly weaker bound in Theorem 4.18 and Corollary 4.23 that holds for complete cosets of RM q (1, m). In contrast, in order to obtain cosets of RM 2 (1, m) with low PMEPR, the approach in [76] and [77] involves a computational search for the maximum PMEPR over a number of kernels. Indeed in [77, Table 6] it is implicitly shown that the cosets with coset representatives given in Corollary 4.31 have PMEPR at most 4 when q = 2. Notice that their semi-computationally obtained upper bound is exactly as predicted by Corollary 4.31. Below we provide a list of open problems, which suggest possible further research directions. We presented some kernels of short length and we also sought good kernels of medium length. From the complexity point of view it is also realistic to perform an exhaustive search for good kernels of larger lengths, say 32 or 64, and dened over larger alphabets. We have not attempted such a search and leave it to further work. It would be desirable to nd kernels producing cosets of RMq (1, m) with PMEPR upper-bounds close to 2 or even exactly 2 (which implies the discovery of new complementary pairs). Moreover our theory would benet from having an efcient way to construct good kernels. The application of Corollary 4.23 together with a suitable kernel can produce cosets of RMq (1, m) whose PMEPR is upper bounded by at least 2. If the upper bound is exactly 2, the cosets are comprised of Golay sequences. The only known cosets that consist of Golay sequences are those in Corol88

4.9 Conclusions, Notes, and Open Problems


lary 4.30. Although more Golay sequences of length 2m are known, they do not organize in cosets of RMq (1, m). We ask: do there exist cosets of RMq (1, m) comprised of Golay sequences that are distinct from those in Corollary 4.30? A partial answer to this question was recently announced by Calderbank et al. [11]: the nonexistence of further cosets of RM 2 (1, m) comprised of Golay sequences inside RM2 (2, m) was proved. Although cosets comprised of Golay sequences that are distinct from those in Corollary 4.30 may not exist, recent results [56], [26] show that there are more individual quaternary Golay sequences of length 2 m . However no more individual binary Golay sequences of length 2m have been found so far. By numerical evidence we know that there are no more binary Golay sequences of length 2m for m 6 [105, page 51]. Are there binary complementary pairs besides those identied in Corollary 4.30? For several cases we have analyzed lower bounds for the achievable PMEPR of cosets of RMq (1, m) by examining the expression (4.11) in Theorem 4.26. However we failed to nd a general lower bound. What is the lowest possible PMEPR of the cosets f +RMq (1, m) depending on q, the alphabet of f , and the degree of f . Notice that nding the least value of (4.11) for q = 2 coincides with determining the covering radius of RM 2 (1, m), which is for odd m an old and, perhaps, one of the most difcult open problems in coding theory [101].

89

4 Near-Complementary Pairs and ReedMuller Codes

90

Chapter 5

Complementary Sets and ReedMuller Codes

5.1 Introduction and Chapter Overview In this section we continue the study of cosets of RMq (1, m) with low PMEPR by relating such cosets with complementary sets. A complementary set is a generalization of a Golay complementary pair and dened as follows (cf. [110]). Denition 5.1. A set of N sequences {x(0) , x(1) , . . . , x(N 1) } is called a complementary set of size N if A(x(0) )( ) + A(x(1) )( ) + + A(x(N 1) )( ) = 0 for = 0.

From Theorem 4.7 we know that the PMEPR of a, b Zn is bounded by 2 if q is a Golay complementary pair. This is because the corresponding OFDM (, b) a signals satisfy |S(a)()|2 + |S(b)()|2 = 2n (cf. proof of Theorem 4.7). This argument may be extended to complementary sets: suppose that a(0) , a(1) , . . . , a(N 1) Zn and the N polyphase sequences q {(0) , a(1) , . . . , a(N 1) } form such a set, then a |S(a(0) )()|2 + |S(a(1) )()|2 + + |S(a(N 1) )()|2 = N n. Therefore we obtain: Theorem 5.2. [79] Each word whose corresponding polyphase version lies in a complementary set of size N has PMEPR at most N . 91

5 Complementary Sets and ReedMuller Codes


In this chapter it will be shown that any coset of RM q (1, m) is comprised of sequences lying in complementary sets of size 2k+1 . Here k is a nonnegative integer, which can be easily determined from the generalized Boolean function associated with a coset representative. As a corollary we can upper bound the PMEPR of any coset of RMq (1, m) by a power of 2. This result is then combined with a new generalization of the ReedMuller code in order to derive exible OFDM coding schemes with low PMEPR. This chapter generalizes results obtained by Paterson in [79] and makes extensive use of the techniques developed there. While in [79] the focus is on cosets of RMq (1, m) inside RMq (2, m), we will consider arbitrary cosets of RMq (1, m). The remainder of this chapter is structured as follows. In Section 5.2 we recall some useful notation from [79]. We will prove our results on complementary sets in Section 5.3 and derive an upper bound on the PMEPR of arbitrary cosets of RMq (1, m). In Section 5.4 we will establish the tightness of this bound in certain cases. In Section 5.5 we introduce a new nonbinary generalization of the classical (binary) ReedMuller code, which we call the effective-degree ReedMuller code. It turns out that this code has better parameters than the previous generalizations RM2h (r, m) and ZRM2h (r, m) for h > 1 and h > 2, respectively. Section 5.6 proposes two OFDM coding schemes whose PMEPR is bounded by 2 k+1 . These will be analyzed in detail for k = 1 and compared with previously proposed coding schemes. The chapter will be concluded in Section 5.7, which also contains some open problems. Convention. Throughout this chapter q is an even positive integer. 5.2 Further Notation and Denitions The material in this section is from [79]. Expansion of Aperiodic Correlation Functions In what follows we recall the technique of restricting generalized Boolean functions and their associated polyphase sequences. This technique was introduced in [79] in order to expand aperiodic correlation functions, as we shall see in Lemma 5.4. Suppose that f : {0, 1}m Zq is a generalized Boolean function. Let a set of k indices be given by J = {j0 , j1 , . . . , jk1 } such that J {0, 1, . . . , m 1}. 92

5.2 Further Notation and Denitions


Dene z = (xj0 , xj1 , . . . , xjk1 ) and let d = (d0 , d1 , . . . , dk1 ) be a binary word of length k. Further let (i0 , i1 , . . . , im1 ) be the binary expansion of the integer 0 i < 2m . The restricted sequence f |z=d is a sequence of length 2m that coincides with f at the positions i where ij = d for each 0 < k. Otherwise f |z=d is equal to zero. For k = 0 we dene f |z=d := f. Now write z = (xj0 , xj1 , . . . , xjk2 ) and d = (d0 , d1 , . . . , dk2 ). Then the above denition implies f |z =d = f |z=d 0 + f|z=d 1 , which yields by induction f=
d{0,1}k

(5.1)

f |z=d .

A sequence that is restricted in k variables comprises 2m 2mk zero entries and 2mk nonzero entries. Those nonzero entries are determined by a function, which is denoted as f |z=d and called a restricted generalized Boolean function. This function is a generalized Boolean function in m k variables and is obtained by replacing the variables xj by d for all 0 < k in the algebraic normal form of f . The restricted sequence f |z=d is then recovered by associating a polyphase sequence of length 2mk with f |z=d and inserting 2m 2mk zeros at the corresponding positions. Similarly to a disjunctive normal form of a Boolean function [60], the original function f can be reconstructed from the functions f | z=d by
k1

f (x0 , . . . , xm1 ) =
d{0,1}k

f |z=d

=0

xd (1 xj )(1d ) . j

Example 5.3. Consider the generalized Boolean function f : {0, 1} 3 Z2 given by f (x0 , x1 , x2 ) = x0 x2 + x1 x2 + x1 + x2 . Write + for +1 and for 1, and consider the restricted sequences f|x1 =0 = ( + + 0 0 + 0 0 ) f|x1 =1 = ( 0 0 0 0 + ) . The sum of these sequences equals f = (++++), which is the polyphase sequence associated with f . Alternatively we may examine the restricted Boolean 93

5 Complementary Sets and ReedMuller Codes


functions f |x1 =0 = x0 x2 + x2

f |x1 =1 = x0 x2 + 1.

From these functions the algebraic normal form of the original function f can be reconstructed by (1 x1 ) f |x1 =0 + x1 f |x1 =1 . By evaluating f |x1 =0 and f |x1 =1 , we obtain the two polyphase sequences (+ + +) and ( +), respectively. After inserting zeros at the corresponding positions, the restricted sequences f |x1 =0 and f |x1 =1 are recovered. The following lemma is due to [79] and a special case of [105, Lemma 1.15]. Lemma 5.4. [79] Suppose that f : {0, 1}m Zq is a generalized Boolean function. Let I = {i0 , i1 , . . . , it1 } and J = {j0 , j1 , . . . , jk1 } be two sets of indices such that I J = and I J {0, 1, . . . , m1}. Write y = (xi0 , xi1 , . . . , xit1 ) and z = (xj0 , xj1 , . . . , xjk1 ). Then for any c {0, 1}t we have A(f |y=c )( ) =
d{0,1}k

A(f |yz=cd )( ) +

d1 ,d2 {0,1}k d1 =d2

C(f |yz=cd1 , f|yz=cd2 )( ).

Proof. We consider the case 0. For < 0 the above relations then follow from the symmetric structure of correlation functions (cf. Lemma 2.12). Applying induction to (5.1) we obtain f|y=c =
d{0,1}k

f |yz=cd .

Hence
n 1

A(f |y=c )( ) = =

i=0 n 1 i=0

(f|y=c )i+ (f |y=c ) i (f |yz=cd1 )i+ (f |yz=cd2 )i

d1 {0,1}k

d2 {0,1}k

n 1 i=0

=
d1 ,d2 {0,1}k

(f |yz=cd1 )i+ (f |yz=cd2 )i

=
d1 ,d2 {0,1}k

C(f |yz=cd1 , f |yz=cd2 )( ).

94

5.3 Construction of Complementary Sets


After extracting the terms in the sum where d1 = d2 we obtain the statement in the lemma. Quadratic Polynomials, Graphs, and Paths With each quadratic polynomial f (xi0 , . . . , xim1 ) Zq [xi0 , . . . , xim1 ]/(x20 xi0 , . . . , x2m1 xim1 ), i i generally given by
m1

f (xi0 , . . . , xim1 ) =
0 j<k<m

bjk xij xik +


j=0

c j xi j + ,

bjk , cj , Zq ,

we associate an undirected labeled graph, denoted by G(f ), as follows. The vertices of G(f ) are labeled by i0 , i1 , . . . , im1 , and the vertices ij and ik are joined by an edge labeled bjk if bjk = 0. By convention, the edge labels are omitted if q = 2 (since in this case every edge is labeled by 1). We call G(f ) a path if m = 1 (then the graph consists of a single vertex) or if m 2 and f (xi0 , . . . , xim1 ) is of the form f (xi0 , . . . , xim1 ) = q 2
m2 m1

xi(j) xi(j+1) +
j=0 j=0

c j xi j + ,

c j , Zq ,

where is a permutation of {0, 1, . . . , m 1}. The vertices labeled by i (0) and i(m1) are called end vertices of the path. If the path consists of a single vertex, the single vertex is called an end vertex as well. We will say that m is length of the path. Example 5.5. The graph of f (x0 , x1 , x2 , x3 , x4 ) = x0 x1 + x0 x2 + x0 x4 + x1 x2 + x1 x4 + x0 + x3 . is shown in Figure 5.1. 5.3 Construction of Complementary Sets We begin with recalling a construction of Golay complementary pairs of length 2m over the alphabet i {q | i Zq } {0} 95

5 Complementary Sets and ReedMuller Codes


0 1

Figure 5.1: The graph of f given in Example 5.5 from [79]. The following theorem [79, Theorem 9] slightly generalizes the construction of Golay complementary pairs described by Davis and Jedwab in [22] (cf. Corollary 4.30). Theorem 5.6. [79] Suppose m > k and let 0 j0 < j1 < < jk1 < m be a list of k indices. Write z = (xj0 , xj1 , . . . , xjk1 ) and let d be a binary word of length k. Suppose f : {0, 1}m Zq is a generalized Boolean function such that f |z=d is quadratic and G(f |z=d ) is a path of length m k. Dene another function g : {0, 1}m Zq by q g(x0 , . . . , xm1 ) = f (x0 , . . . , xm1 ) + xa + , 2 where a is an end vertex of the path G(f |z=d ) and Zq . Then (f |z=d , g|z=d ) is a Golay complementary pair. Proof. We have to show that A(f |z=d )( ) + A(|z=d )( ) = 0 for g = 0. (5.2)

We rst prove the theorem for the case k = m 1, and then proceed by induction. If k = m 1, z consists of all the variables x0 , . . . , xm1 except for xa . Hence each of the restricted sequences f|z=d and g |z=d contains exactly two nonzero elements, which are located at the positions
a1 m2

2 +

d 2 +
=0 =a

d 2+1 ,

{0, 1}.

Let the two nonzero elements in f|z=d be q and q . Then the respective + + two nonzero elements in g |z=d are given by q and q . Therefore the

96

5.3 Construction of Complementary Sets


only nonzero autocorrelation values of f|z=d and g |z=d for a positive shift are a a A(f |z=d )(2 ) = q and A(|z=d )(2 ) = q . Now it is straightforward to g verify that (5.2) is true for k = m 1. Now we proceed by induction on k to prove the remaining cases where k < m 1. Here we use the theorem as a hypothesis and the case k = m 1 as the induction anchor. Since G(f |z=d ) is a path of length m k 2 with a being one end vertex, the restricted sequences f |z=d and g|z=d correspond to restricted functions f |z=d q = 2
mk2 mk1

xi() xi(+1) +
=0 =0

c xi() + c ,

c , c Z q

q g|z=d = f |z=d + xa + , 2 respectively. Here, the indices 0 i0 < i1 < < imk1 < m

are distinct from the js, a {i(0) , i(mk1) }, and is a permutation of {0, 1, . . . , m k 1}. We now employ Lemma 5.4 to expand (5.2) as follows A(f |z=d )( ) + A(|z=d )( ) g = A(f1 )( ) + A(f2 )( ) + C(f1 , f2 )( ) + C(f2 , f1 )( ) + A(1 )( ) + A(2 )( ) + C(1 , g2 )( ) + C(2 , g1 )( ), g g g g where f1 = f |zxa =d0 g1 = g |zxa =d0 f2 = f |zxa =d1 (5.3)

g2 = g |zxa =d1 .

For the remainder of the proof we assume a = i(mk1) . The proof can be carried out similarly for the case a = i(0) . Then the respective restricted functions corresponding to f1 , f2 , g1 , and g2 are given by q f2 = f |zxa =d1 = p + xi(mk2) + cmk1 2 g1 = g|zxa =d0 = p + q q g2 = g|zxa =d1 = p + xi(mk2) + cmk1 + + , 2 2 97 f1 = f |zxa =d0 = p

5 Complementary Sets and ReedMuller Codes


where q p= 2

mk3

mk2

xi() xi(+1) +
i=0 =0

c xi() + c .

Observe that G(f1 ) and G(g1 ) are paths of length m k 1, where i(mk2) is an end vertex of both of them. Hence, by hypothesis, ( f1 , f2 ) and (1 , g2 ) are g Golay complementary pairs. Therefore A(f1 )( ) + A(f2 )( ) = A(1 )( ) + A(2 )( ) = 0 for g g = 0. (5.4)

Since f1 = g1 and f2 = g2 + q/2, we have f1 = q g1 and f2 = q g2 , and by Lemma 2.12 (ii)

C(f1 , f2 )( ) = C(1 , g2 )( ) g C(f2 , f1 )( ) = C(2 , g1 )( ). g

(5.5) (5.6)

By plugging (5.4), (5.5), and (5.6) into (5.3), we conclude that (5.2) is satised, as it was to be shown. We are now in a position to state a construction of sequences lying in complementary sets. Theorem 5.7. Let 0 j0 < j1 < < jk1 < m be a list of k indices, and write z = (xj0 , xj1 , . . . , xjk1 ). Let f : {0, 1}m Zq be a generalized Boolean function such that for each d {0, 1}k the restricted function f |z=d is quadratic and G(f |z=d ) is a path of length m k. Then f lies in a complementary set of k+1 size 2 . Proof. Write d = (d0 , d1 , . . . , dk1 ) and let c = (c0 , c1 , . . . , ck1 ) {0, 1}k . Dene q q fcc (x0 , . . . , xm1 ) = f (x0 , . . . , xm1 ) + c xj + c e(x0 , . . . , xm1 ), 2 =0 2 (5.7) where
k1 k1

e(x0 , . . . , xm1 ) =
d{0,1}k

xa d
=0

xd (1 xj )(1d ) j

and ad is an end vertex of the path G(f |z=d ). We claim that the set fcc | c {0, 1}k , c {0, 1} , 98

5.3 Construction of Complementary Sets


which contains f , is a complementary set of size 2k+1 . To prove this, it has to be shown that the sum c, c A(fcc )( ) is zero for each = 0. We employ Lemma 5.4 and write A(fcc )( ) = S1 + S2 ,
c, c

where S1 =
c, c d

A(fcc |z=d )( ) C(fcc |z=d1 , fcc |z=d2 )( ).

S2 =
c, c d1 =d2

We rst focus on the term S1 , which can be written as S1 =


c d

A(fc0 |z=d )( ) + A(fc1 |z=d )( ) .

Note that e|z=d = xad . Thus fc0 |z=d fc1 |z=d q c d = f |z=d + 2 =0 q q = f |z=d + c d + xa d . 2 =0 2
k1 k1

Hence, since G(f |z=d ) is a path and the term containing the sum over is a con stant occurring in both expressions, Theorem 5.6 implies that fc0 |z=d and fc1 |z=d form a Golay complementary pair. It follows that the inner term of S 1 is zero for = 0, and thus, also S1 itself is zero for = 0. Next we focus on the term S2 and rearrange the sum as follows S2 =
d1 =d2 c c

C(fcc |z=d1 , fcc |z=d2 )( ).

For xed d1 , d2 , and c denote the inner sum over c as Sc . We have fcc |z=d1 = f + q c e 2 q + h1 z=d1 2 q + h2 , z=d2 2 99

q fcc |z=d2 = f + c e 2

5 Complementary Sets and ReedMuller Codes


respectively, where
k1

h1 =
=0 k1

c d1, c d2, .
=0

h2 =

Now write g1 = (f + (q/2) c e)|z=d1 and g2 = (f + (q/2) c e)|z=d2 , and let g1 and g2 be their respective associated (restricted) sequences. Then Sc comprises terms of the form C( g1 , g2 )( ), where C(+1 , +2 )( ) and C(1 , 2 )( ) g g g g occur if h1 = h2 , and C(+1 , 2 )( ) and C(1 , +2 )( ) occur if h1 = g g g g h2 . By Lemma 2.12 (ii) it follows that C(+1 , +2 )( ) = C(1 , 2 )( ) = g g g g C(+1 , 2 )( ) = C(1 , +2 )( ). g g g g In order to prove that Sc is zero, we have to show that h1 = h2 and h1 = h2 occur equally often as c runs through all possible values. Since d 1 = d2 , the difference
k1

h2 h 1 =

=0

c (d2, d1, )

is a nonzero linear polynomial in the variables c0 , c1 , . . . , ck1 . According to the randomization lemma [60, page 372], such a polynomial produces the values 0 and 1 equally often as c takes on all possible values. Thus h1 and h2 are distinct for half of all cases. It follows that Sc , and hence, also S2 itself is zero for all . The preceding theorem together with Theorem 5.2 immediately provides bounds on the PMEPR of arbitrary cosets of RMq (1, m). Corollary 5.8. Let f : {0, 1}m Zq be a generalized Boolean function. Suppose that there exists a list of indices 0 j0 < j1 < < jk1 < m, dening the restricting variables z = (xj0 , xj1 , . . . , xjk1 ) such that for each d {0, 1}k the restricted function f |z=d is quadratic and the graph G(f |z=d ) is a path of length m k. Then the PMEPR of the coset f + RMq (1, m) is at most 2k+1 . We have a number of notes on Theorem 5.7 and its consequence, Corollary 5.8. In the special case when k = 0 the above corollary identies exactly the cosets of RMq (1, m) from Corollary 4.30. Hence, when k = 0, Corollary 5.8 does not provide any new results. For k > 0 Corollary 5.8 essentially generalizes [79, 100

5.3 Construction of Complementary Sets


Corollary 13]; if f is constrained to be a quadratic generalized Boolean function, then Corollary 5.8 essentially reduces to [79, Corollary 13]. The proof of Theorem 5.7 shows that the sequence f lies in a complementary set that can be decomposed into 2k Golay complementary pairs identied in Theo rem 5.6. By reversing this process, the sequence f can be constructed by interleaving 2k Golay sequences from Theorem 5.6. However the sole application of such an interleaving method would not directly admit the construction of sequences corresponding to generalized Boolean functions of a specic degree, which will be required to derive our coding schemes in Section 5.6. In the next section we will characterize several cases where Corollary 5.8 yields a tight bound. However it should be noted that the bound in Corollary 5.8 tends to be loose when k is large. Indeed [78], [105] (in particular [105, Theorem 3.6]), and the recent work [16] contain signicant improvements of Corollary 5.8 in certain situations. A worst-case example where Corollary 5.8 fails to provide a meaningful bound was exhibited in [16]: if m is even and f : {0, 1}m Zq is of the form f (x0 , . . . , xm1 ) = q 2
m2

xi xi+1 ,
i=0 i=m/21

then Corollary 5.8 yields an upper bound on the PMEPR of the coset f + RMq (1, m) of 2m/2+1 . With [105, Theorem 3.6] this bound can be improved to 2m/2 . An (almost) tight upper bound of 4 is provided by [16, Theorem 1]. In summary, the usefulness of Corollary 5.8 lies in the fact that it provides a relatively simple method to identify sets of sequences that correspond to generalized Boolean functions of a given (preferably low) degree and whose PMEPR is bounded above by a given power of 2. This will be crucial in Section 5.6, where we derive exible coding schemes for OFDM with low PMEPR. We close this section with an example for the application of Theorem 5.7 and Corollary 5.8. Example 5.9. Let f : {0, 1}4 Z2 be given by f (x0 , x1 , x2 , x3 ) = x0 x1 x2 + x0 x1 x3 + x2 x3 + x1 x3 + x2 . By restricting f in x0 (i.e., z = (x0 )), we obtain the two restricted functions f |x0 =0 = x1 x3 + x2 x3 + x2

f |x0 =1 = x1 x2 + x2 x3 + x2 , 101

5 Complementary Sets and ReedMuller Codes


which are quadratic and their associated graphs are paths of length 3. Hence, by Theorem 5.7, f lies in a complementary set of size 4, so its PMEPR is at most 4. By Corollary 5.8, the PMEPR of the whole coset f + RM2 (1, 4) is at most 4. Figure 5.2 contains the normalized envelope power |S(f )()| 2 /16 of f . Direct examination of this function yields PMEPR(f ) 3.32, while the maximum PMEPR taken over all words in the coset is equal to 4.

Figure 5.2: The normalized envelope power of f given in Example 5.9

5.4 Lower Bounds on the PMEPR In this section we will use Theorem 4.26 to obtain lower bounds on the PMEPR of cosets of RMq (1, m) identied in Corollary 5.8. In this way we are able to characterize several cases where the PMEPR bound in Corollary 5.8 is tight. We will constrain our study to cosets f + RMq (1, m), where f : {0, 1}m Zq is a quadratic generalized Boolean function. Consider the quadratic functions given by f (x0 , . . . , xm1 ) = +
i=0 j=mk

q 2

mk2

x(i) x(i+1)
i=0

mk1 m1

aij x(i) x(j) +


mk i<j<m

bij x(i) x(j) , (5.8)

where aij , bij Zq and is a permutation of {0, 1, . . . , m 1}. Setting z = (x(mk) , . . . , x(m1) ), we see that G(f |z=d ) is a path for each d {0, 1}k . Therefore, by Corollary 5.8, the PMEPR of the coset f + RMq (1, m) is upper bounded by 2k+1 . It is not hard to verify that every quadratic function that satises the condition in Corollary 5.8 for a particular k can be expressed in the form (5.8). 102

  

5.4 Lower Bounds on the PMEPR


Now recall from Theorem 4.26 that for any generalized Boolean function f : {0, 1}m Zq the PMEPR of the coset f + RMq (1, m) is at least 1 max |Lf (u)|2 , 2m uZm q where Lf (u) is given by Lf (u) =
f q (x)ux x{0,1}m

(cf. Denition 4.24). For q = 2 we have Lf (u) = (1)f (x)+ux


x{0,1}m

= 2m wt(f + u x). Hence in the case where q = 2 a lower bound on the PMEPR of the considered cosets can be obtained by inspecting the weights of the words in the cosets. This approach also provides lower bounds on the PMEPR of cosets of the form (q/2)f +RMq (1, m). By examining the weight distribution of second-order cosets of RM2 (1, m), Paterson [79] could identify several cases where the PMEPR of a coset with coset representative associated with the expression in (5.8) is exactly 2k+1 . This result is summarized in the following theorem. Theorem 5.10. [79] Let m k be odd and let f : {0, 1}m Zq be given by q f (x0 , . . . , xm1 ) = 2
mk2 i=0

q x(i) x(i+1) + 2

mk1 m1

cij x(i) x(j) ,


i=0 j=mk

where cij Z2 , is a permutation of {0, 1, . . . , m 1}, and


(mk1)/2

i=0

c(2i),j = 0 for each j = m k, . . . , m 1.

Then the PMEPR of the coset f + RMq (1, m) is exactly 2k+1 . The approach based on the weight distribution of cosets of RM2 (1, m) fails for the general case where q > 2. We will use the lower bound in Theorem 4.26 to prove the following theorem, which generalizes Theorem 5.10 to arbitrary (even) q in the case k = 1. 103

5 Complementary Sets and ReedMuller Codes


Theorem 5.11. Let m be even and let f : {0, 1}m Zq be given by q f (x0 , . . . , xm1 ) = 2
m3 m2

x(i) x(i+1) +
i=0 i=0

ci x(i) x(m1) ,

where ci Zq and is a permutation of {0, 1, . . . , m 1}. If


m/21

c2i = 0,
i=0

then the PMEPR of the coset f + RMq (1, m) is exactly 4. We need some preliminaries to prove the theorem. Let f : {0, 1}m Zq be a generalized Boolean function. In the remainder of this section we will frequently use the following expansion Lf (u, v) = =
x{0,1}m1 f q (x)(u,v)x x{0,1}m f v q (x,0)ux + q x{0,1}m1 f q (x,1)ux.

(5.9)

We will be interested in so-called path functions pm : {0, 1}m Zq , which are of the form q pm (x0 , . . . , xm1 ) = 2 From pm (x, 0) = pm1 (x) q pm (x, 1) = pm1 (x) + xm2 + cm1 2 and (5.9) it is seen that Lpm (u0 , . . . , um1 )
m2 m1

xi xi+1 +
i=0 i=0

c i xi ,

ci Z q .

q c = Lpm1 (u0 , . . . , um2 ) + q m1 um1 Lpm1 (u0 , . . . , um2 + ). (5.10) 2

The next two lemmas state some properties of Lpm . 104

5.4 Lower Bounds on the PMEPR


Lemma 5.12. Suppose that m is odd and let pm : {0, 1}m Zq be given by q pm (x0 , . . . , xm1 ) = 2 Then we have Lpm (0, u1 , 0, u3 , . . . , 0, um2 , 0) = 2(m+1)/2 . Proof of Lemma. For odd m we claim that
u Lpm (0, u1 , 0, u3 . . . , 0, um2 , um1 ) = 2(m1)/2 1 + q m1 . m2

xi xi+1 .
i=0

(5.11)

If the above expression is true, the lemma follows immediately. Since L p0 = 1, we have by (5.10) u Lp1 (u0 ) = 1 + q 0 , so (5.11) is true for m = 1. We now use (5.10) and induction on m to establish Lpm+1 (0, u1 , 0, u3 . . . , 0, um2 , 0, um )
q/2 u 0 = 2(m1)/2 1 + q + 2(m1)/2 q m 1 + q

= 2(m+1)/2 and
u Lpm+2 (0, u1 , 0, u3 . . . , 0, um2 , 0, um , um+1 ) = 2(m+1)/2 1 + q m+1 ,

which proves our claim in (5.11), and therefore, proves the lemma. Lemma 5.13. Suppose that m is odd and let pm : {0, 1}m Zq be given by q pm (x0 , . . . , xm1 ) = 2 where ci Zq and
m2 m1

xi xi+1 +
i=0 i=0

c i xi ,

(m1)/2

c2i = 0.
i=0

Then we have Lpm (0, c1 , 0, c3 , . . . , 0, cm2 , 0) = 2(m+1)/2 . 105

5 Complementary Sets and ReedMuller Codes


Proof of Lemma. For odd m we claim that Lpm (u0 , c1 , u2 , c2 , . . . , cm2 , um1 ) = 2(m1)/2 1 + which is true for m = 1 since by (5.10)
c Lp1 (u0 ) = 1 + q 0 u0 .

(m1)/2

i=0

c q 2i u2i ,

(5.12)

Then (5.12) and the relation in (5.10) lead to Lpm+1 (u0 , c1 , u2 , c2 , . . . , cm1 , um1 , um ) =
(m1)/2

2(m1)/2 1 +

i=0

c c q 2i u2i + q m um 1

(m1)/2

i=0

c q 2i u2i

and Lpm+2 (u0 , c1 , u2 , c2 , . . . , cm1 , um1 , cm , um+1 ) =2


(m+1)/2

(m1)/2

c = 2(m1)/2 2 + 2 q m+1 um+1

(m+1)/2

1+
i=0

c q 2i u2i ,

i=0

c q 2i u2i

which, by induction, proves (5.12). If

(m1)/2 c2i i=0

= 0, it is now seen that

Lpm (0, c1 , 0, c3 , . . . , 0, cm2 , 0) = 2(m+1)/2 , as it was to be shown. Now we complete the proof of Theorem 5.11. Proof of Theorem 5.11. According to Theorem 4.26, the PMEPR of the coset f + RMq (1, m) is at least 1 max |Lf (u)|2 . m uZm 2 q 106

5.4 Lower Bounds on the PMEPR


By Lemma 4.25, the above expression is independent of the choice of the permutation , so we assume that is the identity permutation. With the expansion in (5.9) we can write Lf (u, v) = =
x{0,1}m1 v = La (u) + q Lb (u), f q (x)(u,v)x x{0,1}m f v q (x,0)ux + q x{0,1}m1 f q (x,1)ux

where La and Lb are associated with the functions a, b : {0, 1}m1 Zq given by q a(x0 , . . . , xm2 ) = f (x0 , . . . , xm2 , 0) = 2 and q b(x0 , . . . , xm2 ) = f (x0 , . . . , xm2 , 1) = 2
m3 m2 m3

xi xi+1
i=0

xi xi+1 +
i=0 i=0

c i xi ,
m/21 c2i i=0

respectively. Note that a and b are path functions and we assumed 0 in the theorem. Hence Lemmas 5.12 and 5.13 imply

La (0, c1 , 0, c3 , . . . , 0, cm2 , 0) = Lb (0, c1 , 0, c3 , . . . , 0, cm2 , 0) = 2m/2 . Therefore Lf (0, c1 , 0, c3 , . . . , 0, cm2 , 0, 0) = 2m/2+1 . It follows that the PMEPR of the coset f + RMq (1, m) is at least 1 max |Lf (u)|2 2m uZm q 1 2m+2 = 4. 2m

But Corollary 5.8 states that the PMEPR of this coset is at most 4, so the PMEPR must be exactly 4. Example 5.14. The function f : {0, 1}4 Z4 with algebraic normal form f (x0 , x1 , x2 , x3 ) = 2x0 x1 + 2x0 x3 + x1 x2 + 3x2 x3 107

5 Complementary Sets and ReedMuller Codes


denes the coset f + RM4 (1, 4). By Corollary 5.8, the PMEPR of this coset is at most 4 (this is seen by restricting f in x2 ). Theorem 5.11 states that the PMEPR of this coset is exactly 4 since with the notation as in Theorem 5.11 we have c0 = 1 and c2 = 3, so c0 + c2 = 0 (mod 4). The normalized envelope power |S(f )()|2 /16 of f is shown in Figure 5.3. It is seen that the normalized envelope power has a peak of 4 at time = 0, showing that this word has PMEPR of exactly 4.

Figure 5.3: The normalized envelope power of f given in Example 5.14

5.5 The Effective-Degree ReedMuller Code In Section 4.2 we have introduced the codes RMq (r, m) and ZRMq (r, m), which generalize the classical (binary) ReedMuller code RM2 (r, m) to codes over qary alphabets. In this section we introduce the effective-degree ReedMuller code, which is another generalization of RM2 (r, m) to codes over Z2h . It is then shown how the codes RM2h (r, m) and ZRM2h (r, m) are related to this generalization. We begin with dening the effective degree of a generalized Boolean function. Any function f : {0, 1}m Z2h can be written in 2-adic expansion
h1

f (x) =
i=0

2i fi (x),

where fi : {0, 1}m {0, 1}. We dene the effective degree of f to be


0 i<h

max (deg(fi ) i) .

Also let GBF(r, m, h) be the set of all functions from {0, 1}m to Z2h with effective degree at most r. 108

  

5.5 The Effective-Degree ReedMuller Code


Note that the effective degree is at least 1 h and at most m, and when h = 1, the effective degree and the algebraic degree of a Boolean function coincide. Moreover the effective degree of a generalized Boolean function is always less than or equal to its algebraic degree. Example 5.15. The function f : {0, 1}3 Z8 with algebraic normal form f (x0 , x1 , x2 ) = 4x0 x1 x2 + x1 has algebraic degree equal to 3, while its effective degree is equal to 1. Denition 5.16. For 1 h r m we dene the effective-degree ReedMuller code ERM(r, m, h) to be the set of words associated with generalized Boolean functions of effective degree at most r. Since the effective degree and the algebraic degree coincide for h = 1, ERM(r, m, 1) is identical to the classical binary ReedMuller code, RM2 (1, m). A simple counting argument yields
r

log2 |ERM(r, m, h)| = log2 |GBF(r, m, h)| =

h
i=0

m m . + (h i) r+i i i=1

h1

It is apparent that ERM(r, m, h) is a linear code over Z2h . A generator matrix for ERM(r, m, h) has rows corresponding to the words associated with monomials in the variables x0 , x1 , . . . , xm1 of degree at most r together with 2i times the monomials of degree r + i, where i = 1, . . . , h 1. Example 5.17. The code ERM(0, 3, 3) is a linear code over Z8 and has a generator matrix 1 1 1 1 1 1 1 1 1 0 2 0 2 0 2 0 2 2x0 0 0 2 2 0 0 2 2 2x1 0 0 0 0 2 2 2 2 2x2 0 0 0 4 0 0 0 4 4x0 x1 0 0 0 0 0 4 0 4 4x0 x2 0 0 0 0 0 0 4 4 4x1 x2 . ERM(r, m, h) admits the (u | u + v) construction, i.e., for 1 h < r < m we have

ERM(r, m, h) = {(u | u+v) | u ERM(r, m1, h), v ERM(r1, m1, h)}. 109

5 Complementary Sets and ReedMuller Codes


Note that ERM(1 h, m, h) is a repetition code over 2h1 Z2 and ERM(m, m, h) m is the space Z2h . Hence, by the same reasoning as in the proof of Theorem 4.3, 2 we obtain: Theorem 5.18. d(ERM(r, m, h)) = 2mr . Next we comment on the relation of ERM(r, m, h) to the codes RM2h (r, m) and ZRM2h (r, m) given in Denition 4.1. We have RM2h (r, m) ERM(r, m, h). This inclusion is proper if h > 1 and r < m. Hence for h > 1 and r < m the code ERM(r, m, h) contains more codewords than RM2h (r, m), while both codes have minimum distance equal to 2mr . For h 2 ZRM2h (r + 1, m) ERM(r, m, h), which is a proper inclusion if h > 2 and r < m1. Hence for h > 2 and r < m1 the code ERM(r, m, h) contains more codewords than ZRM2h (r + 1, m), while their minimum distances are equal to 2mr . 5.6 OFDM Codes with Low PMEPR In Section 4.7 we have constructed OFDM codes whose PMEPRs are bounded by values between 2 and 4. In this section we will use Corollary 5.8 to derive OFDM coding schemes whose PMEPRs are at most a given power of 2. These codes are unions of cosets of the linear code A, which is dened below. Denition 5.19. For 0 k < m, 0 A = A(k, r, m, h) to be the set
mk1

k + 1, and h

1 we dene the code

xi gi (xmk , . . . , xm1 ) + g(xmk , . . . , xm1 )


i=0

g0 , . . . , gmk1 GBF(r 1, k, h), g GBF(r, k, h) . It follows from the preceding denition that A(k, r, m, h) is a linear subcode of ERM(r, m, h). Therefore the minimum distance of A(k, r, m, h) is at least 110

5.6 OFDM Codes with Low PMEPR


2mr . Since A(k, r, m, h) contains the word corresponding to the monomial xmr xmr+1 xm1 (which has weight equal to 2mr ), this bound is in fact tight. Moreover we have log2 A(k, r, m, h) = log2 GBF(r, k, h) + (m k) log2 GBF(r 1, k, h) k + = h i i=0
r h1 i=0

(h i)

k r+i
h1 i=0

k + (m k) + h i i=0

r1

(h i)

k r+i1

. (5.13)

Note that A(0, 1, m, h) = RM2h (1, m) and A(0, 0, m, h) = ZRM2h (1, m). More generally, for r 1 A(k, r, m, h) is a union of cosets of RM 2h (1, m), while A(k, 0, m, h) is a union of cosets of ZRM2h (1, m). Example 5.20. The code A(1, 0, 3, 3) is a linear subcode of ERM(0, 3, 3) (considered in Example 5.17) and has a generator matrix 1 1 1 1 1 1 1 1 1 0 2 0 2 0 2 0 2 2x0 0 0 2 2 0 0 2 2 2x1 0 0 0 0 2 2 2 2 2x2 0 0 0 0 0 4 0 4 4x0 x2 0 0 0 0 0 0 4 4 4x1 x2 .

Denition 5.21. We dene R(k, m, h) to be the set of words associated with generalized Boolean functions {0, 1}m Z2h having algebraic normal forms
mk2 k1

h1 d{0,1}k i=0

xd (i) xd (i+1)
j=0

j xmk+j (1 xmk+j )(1dj ) ,

(5.14)

where d (d = (d0 , d1 , . . . , dk1 ) {0, 1, . . . , m k 1}.

{0, 1}k ) are 2k permutations of

Corollary 5.22. The PMEPR of the cosets of A(k, r, m, h) having a coset representative in R(k, m, h) is at most 2k+1 . Proof. The corollary is easily proved with Corollary 5.8 and the following observations. By restricting any function in A(k, r, m, h) in the variables 111

5 Complementary Sets and ReedMuller Codes


xmk , . . . , xm1 , we obtain an afne function, and by restricting any function in R(k, m, h) in the same variables, we obtain a quadratic function, whose graph is a path of length m k. We are now in a position to construct a simple code. Construction 5.23. Take a single coset of A(k, r, m, h) that contains a word in R(k, m, h). This code has PMEPR at most 2k+1 and minimum distance 2mr . The number of encoded bits per codeword is equal to log2 |A(k, r, m, h)|, which is given in (5.13). In order to obtain a more elaborate code construction, we prove: Lemma 5.24. For m > k + 1 and r > 2 h the set R(k, m, h) contains (m k)! 2
2min{r+h3,k}

(5.15)

words corresponding to generalized Boolean functions of effective degree at most r. Proof. The set R(k, m, h) contains exactly [(m k)!/2]2 words, all having effective degree at most k + 3 h. Hence the lemma is true for r k + 3 h. It is also clear that the expression in (5.14) has algebraic degree at least 2, hence, its effective degree is at least 3 h. Now suppose that 3 h r < k + 3 h, and write = r + h 3, so 0 < k. By factoring out terms in the outer sum in (5.14), it can be veried that, if (d0 ,...,d
1 , d k

,...,dk1 )

= (d0 ,...,d

1 , 1d

,...,1dk1 ) ,

then (5.14) is independent of the variables xmk+ , . . . , xm1 and, therefore, has effective degree at most r = + 3 h. This leaves the choice of 2 = 2r+h3 permutations of m k symbols that are distinct under reversal (e.g., all permutations satisfying (0) < (m k 1)) to obtain distinct words in R(k, m, h) with effective degree at most + 3 h. This leads in total to the number given in (5.15). Construction 5.25. Let 2 r k + 2 when h = 1 and 1 r k + 1 when t h > 1. Write r = min{r, k + 1}. Let 2 be the largest integer power of 2 not exceeding (5.15). Now take the union of 2t distinct cosets of A(k, r , m, h), each 112

5.6 OFDM Codes with Low PMEPR

Table 5.1: Coding Options With PMEPR At Most 4 (k = 1) Opt. 4a 4b 4c 4d 4e 4f 4a 4b 4c 4d 4e 4f 4a 4b 4c 4d 4e 4f m 4 h 1 2 3 5 1 2 3 6 1 2 3 r 2 3 1 2 1 2 2 3 1 2 1 2 2 3 1 2 1 2 s 8 8 13 16 21 24 10 10 16 20 26 30 12 12 19 24 31 36 t 1 3 1 3 3 3 3 7 3 7 7 7 5 11 5 11 11 11 R1 0.50 0.81 1.00 1.31 1.50 0.31 0.50 0.63 0.81 0.94 0.19 0.30 0.38 0.48 0.56 R2 0.56 0.69 0.88 1.19 1.5 1.69 0.41 0.53 0.59 0.84 1.03 1.16 0.27 0.36 0.38 0.55 0.66 0.73 d 4 2 8 4 8 4 8 4 16 8 16 8 16 8 32 16 32 16 d2 E 16.00 8.00 16.00 8.00 4.69 2.34 32.00 16.00 32.00 16.00 9.37 4.69 64.00 32.00 64.00 32.00 18.75 9.37

containing a word in R(k, m, h) with effective degree at most r. The PMEPR of this code is at most 2k+1 , and we can encode s + t bits per codeword, where s = log2 |A(k, r , m, h)|. Since the code is a subcode of ERM(r, m, h), its minimum distance is at least 2mr (which is a tight bound if r = r). We remark that, if k = 0, Construction 5.25 essentially restates the construction by Davis and Jedwab [22] (cf. Options B1, Q1, O1 in Tables 4.1, 4.2, 4.3, respectively). A list of coding options having PMEPR at most 4 and at most 8 is compiled in Tables 5.1 and 5.2, respectively. The code with rate R 1 = s/2m is obtained with Construction 5.23, and the code with rate R2 = (s + t)/2m arises from Construction 5.25. The quantity d denotes the lower bound for the minimum distance of the code, and d2 is a lower bound for the squared minimum Euclidean E distance of the associated PSK code (obtained with Theorem 2.9). Next we wish to sketch how these codes can be generally encoded and decoded. We have noted that for r 1 the code A is a union of cosets of RM 2h (1, m). Hence for r 1 the codes in Constructions 5.23 and 5.25 are also unions of cosets 113

5 Complementary Sets and ReedMuller Codes

Table 5.2: Coding Options With PMEPR At Most 8 (k = 2) Opt. 8a 8b 8c 8d 8e 8f 8g 8h 8i 8a 8b 8c 8d 8e 8f 8g 8h 8i m 5 h 1 r 2 3 4 1 2 3 1 2 3 2 3 4 1 2 3 1 2 3 s 13 16 16 19 29 32 35 45 48 16 20 20 23 36 40 43 56 60 t 1 3 6 1 3 6 3 6 6 3 7 14 3 7 14 7 14 14 R1 0.41 0.50 0.59 0.91 1.00 1.09 1.41 1.50 0.25 0.31 0.36 0.56 0.63 0.67 0.88 0.94 R2 0.44 0.59 0.69 0.63 1.00 1.19 1.19 1.59 1.69 0.30 0.42 0.53 0.41 0.67 0.84 0.78 1.09 1.16 d 8 4 2 16 8 4 16 8 4 16 8 4 32 16 8 32 16 8 d2 E 32.00 16.00 8.00 32.00 16.00 8.00 9.37 4.69 2.34 64.00 32.00 16.00 64.00 32.00 16.00 18.75 9.37 4.69

of RM2h (1, m), and encoding and decoding can be performed by means of supercode encoding and decoding methods as described in Section 4.8. Alternatively, we can use the intrinsic structure of the linear code A to encode and decode a word of A in one step. While encoding of A is straightforward by using a generator matrix for A, for decoding one needs to have an efcient algorithm to decode the linear code A. Such a decoder would reduce the decoding complexity considerably compared to partitioning A into cosets of RM2h (1, m) and applying the supercode decoding method. We close this section with a discussion on Constructions 5.23 and 5.25 followed by several comments on the relation to previous constructions. It can be observed that quaternary codes are always better than binary codes, in the sense that we can always construct a quaternary code with higher code rate, while the bounds for the PMEPR and the minimum Euclidean distance are the same. By moving to larger alphabets, the code rate can be increased further, but only at the cost of a smaller minimum Euclidean distance. 114

5.6 OFDM Codes with Low PMEPR


We remark that Corollary 5.22 and the arising code constructions do not exploit Corollary 5.8 in the most general way. The generalized Boolean functions corresponding to the words in the cosets identied in Corollary 5.22 are characterized by the property that by restricting the functions in the variables xmk , . . . , xm1 , we obtain quadratic functions whose graphs are paths in the vertices 0, . . . , m k 1. In order to increase the size of the codes in Constructions 5.23 and 5.25, we can, according to Corollary 5.8, apply any permutation to the m variables in the polynomials corresponding to the codewords (instead of only to a xed set of m k variables). This, however, has the unwanted effect that some codewords are generated more than once. Such an approach, coupled with rather complicated techniques to remove multiple codewords, has been used in [79], where the functions are constrained to have quadratic degree. Our approach has the advantage that these difculties are avoided. Moreover, compared to the concept in [79], it allows us to construct our codes as unions of cosets of a relatively large linear code, which presumably simplies the decoding process. The penalty of this simplication is a loss of at most log2 m information bits (since, instead of m possible index k k sets, we choose just one set of m k indices that form the vertices of the paths in the graphs of the restricted functions). This loss is moderate for typical choices of m and k. Finally we wish to compare the codes from Construction 5.25 with those in [79]. The codes in [79] are contained in RM2h (2, m) or ZRM2h (2, m), which ensures a minimum distance of at least 2m2 or 2m1 , respectively. So we compare these codes with codes from Construction 5.25 having the same bound for the minimum distance, i.e., we let r {1, 2}. Also we let k = 1, which covers the majority of the codes in [79] having PMEPR greater than 2. For h = 1, r = 2, and m 3 the code from Construction 5.25 can be used to encode log2 (m 1)! + 2m 1 bits. This yields 13 and 17 bits for m = 5 and m = 6, respectively. These values should be compared with 11 and 17 bits in [79, Table I], which suggests that for h = 1 and for small m our construction is slightly stronger than that in [79]. However for h = 1 and m 8 it was stated in [79] that the number of encoded bits of the codes in [79] is equal to log2 m! + 2m 2. Hence for large m the binary code from [79] allows to encode either log2 m or log2 m 1 more bits than a comparable code from Construction 5.25. We arrive at a similar conclusion 115

5 Complementary Sets and ReedMuller Codes


for h = 2 and r = 1. For m 3 we can encode log2 (m 1)! + 3m bits, which is, compared to a code in [79] with the same minimum distance, slightly larger for small m and is either log2 m or log2 m 1 bits less for m 8. For h 2, r = 2, and m 3 Construction 5.25 yields a code, which can be used to encode 2 log2 (m 1)! + 2hm 2 bits. When h {2, 3} and m 4, the number of encoded bits for a comparable code in [79] is equal to log2 m! + 2hm 2. This is log2 m! 2 log2 m or log2 m! 2 log2 m + 1 bits less than for the code from Construction 5.25. Similar results can be established for h > 2 and r = 1. In summary, except for large m in the cases (h, r) = (1, 2) and (h, r) = (2, 1), codes resulting from Construction 5.25 outperform coding schemes proposed in [79].

5.7 Conclusions and Open Problems In this chapter we have proved a sufcient condition such that a polyphase word of length 2m lies in a complementary set of size 2k+1 . This result provides an upper bound for the PMEPR of the coset of RMq (1, m) that contains this word. When k = 1, we have shown that the bound is often tight for words corresponding to quadratic functions. It was then demonstrated how sets of sequences lying in complementary sets organize in cosets of a linear code over Z2h inside a new generalization of the ReedMuller code. In this way we obtained a number of exible coding schemes for OFDM whose PMEPRs are bounded by powers of 2. These codes extend and improve the codes presented in Chapter 4 and previously reported coding options in [22] and [79]. Moreover, compared to the codes in Chapter 4, [22], and [79] with PMEPR greater than 2, the codes developed in this chapter have higher intrinsic structure because they are unions of less cosets of a large linear code. This makes them more amenable to efcient encoding and decoding algorithms. We close this chapter with some open problems. 116

5.7 Conclusions and Open Problems


In the remarks following Corollary 5.8 we have mentioned that in certain situations Corollary 5.8 fails to provide a meaningful bound for the PMEPR, and [78], [105], [16] contain signicant improvements of our bound. Can these results be exploited to improve our coding schemes while at least some of the intrinsic structure of the codes is preserved? For quadratic Z2 -valued Boolean functions [79, Theorem 23] contains a complete characterization of the cases where the bound in Corollary 5.8 is tight. For k = 1 we have generalized this result to arbitrary quadratic functions from {0, 1}m to Zq . On the other hand, counterexamples show that for k > 1 [79, Theorem 23] cannot hold for arbitrary quadratic functions from {0, 1}m to Zq . Find more general characterizations of the cases where Corollary 5.8 yields a tight bound. We have mentioned that, in order to decode the codes that arise from Constructions 5.23 and 5.25, we need an efcient decoder for the linear code A. Find such a decoder.

117

5 Complementary Sets and ReedMuller Codes

118

Chapter 6

Codes for Multicode CDMA

6.1 Introduction and Chapter Overview In this chapter we turn our attention to codes for MC-CDMA with low PAPR. We will particularly focus on constant-amplitude codes, which are optimal codes with respect to the PAPR. Denition 6.1. A constant-amplitude code is a code with PAPR equal to 1. Binary codes with low PAPR have been constructed by Wada et al. [114], [115], Ottoson [74], Paterson [80], [81], and Sol and Zinoviev [103]. Most notably, by e exploiting the connection between so-called bent functions and binary words with PAPR equal to 1, in [80] and [81] several binary constant-amplitude codes have been proposed. These codes are unions of cosets of RM 2 (1, m) inside higherorder ReedMuller, Kerdock, or DelsarteGoethals codes [60, Chapter 15]. However, so far, nonbinary codes for MC-CDMA with low PAPR are unknown. There are several motivations to study nonbinary codes for use in MC-CDMA. First, quaternary or octary signaling rather than binary signaling is typically employed in MC-CDMA systems. Second, binary constant-amplitude codes cannot exist for lengths 2m when m is odd (we will see that nonbinary codes, more specifically quaternary codes, indeed do). In this chapter we will, therefore, concentrate on nonbinary codes. In the next section we describe a connection between the Fourier transform of a function from {0, 1}m (alternatively from the Teichmuller set Th,m ) to Zq and the PAPR of the associated word. In particular it is shown that such a word has PAPR equal to 1 if and only if it is associated with a bent function. 119

6 Codes for Multicode CDMA


In the light of the work in [80] and [81] it seems natural to study cosets of RMq (1, m) in order to construct nonbinary codes with low PAPR. Indeed we will show in Section 6.3 that the PAPR of any coset of RMq (1, m) is at most its PMEPR. As a consequence, all codes with low PMEPR developed in Chapters 4 and 5 are directly applicable in MC-CDMA with low PAPR. However we will also see that with this approach it is impossible to construct constant-amplitude codes whenever q is a multiple of 4. In Section 6.4 we turn our attention to the family of trace codes, which are codes derived from the trace of polynomials over Galois rings. A theorem about estimates on character sums is utilized to bound the PAPR of any appropriately expurgated trace code. In Sections 6.5 and 6.6 we derive a number of quaternary constant-amplitude codes. As a prelude to our code constructions in Section 6.6 we study bent functions from {0, 1}m to Z4 in Section 6.5. Several constructions of such functions are presented, including mappings from classical (Z2 -valued) bent functions. These are then exploited in connection with algebraic codes over Z 4 (in particular quaternary ReedMuller codes RM4 (r, m) and ZRM4 (r, m), quaternary Kerdock codes, and quaternary DelsarteGoethals codes) in order to construct families of quaternary constant-amplitude codes. These codes may be viewed as quaternary analogs of the binary codes constructed in [80] and [81]. Moreover some mappings from binary to quaternary constant-amplitude codes are presented. We will conclude this chapter in Section 6.7 and close with a list of open problems.

6.2 Fourier Transforms and Bent Functions Let G be a nite abelian group of order n. A character of G is dened as a homomorphism from G to the complex numbers lying on the unit circle [57, Chapter 5], i.e., we have (a)(b) = (a b), where is the operation in G. There exist exactly n characters of G, which will be denoted as u , where u ranges over G. Our focus is on the groups Zm , + and Th,m , . The characters of Zm are 2 2 given by u (x) = (1)ux , and the characters of Th,m are u (x) = (1)Tr(ux) . Here denotes the scalar product in Zm . 2 120

6.2 Fourier Transforms and Bent Functions


The characters of G are orthogonal, i.e., 0 if u = v u (x) (x) = v n if u = v. xG

(6.1)

Thus they give rise to a transform of a function dened on G. Denition 6.2. Let G be a nite abelian group. Given a function f : G Z q , its Fourier transform Ff : G C is dened to be Ff (u) :=
f q (x) u (x). xG

We will tacitly identify the space {0, 1}m with the group Zm . Thus the Fourier 2 m transform of a generalized Boolean function f : {0, 1} Zq reads Ff (u) =
f q (x) (1)ux x{0,1}m

and that of a function f : Th,m Zq is given by Ff (u) =


f q (x) (1)Tr(ux) . xTh,m

Notice that, if q = 2 in the former case, the Fourier transform of f coincides with the WalshHadamard transform of f (cf. [60, Chapter 14]). Of particular interest in this chapter are so-called bent functions, which are generally dened as follows. Denition 6.3. Let G be a nite abelian group of order n. A function f : G Z q is called a bent function if |Ff (u)| = n for every u G. The name bent function was coined by Rothaus in [87]. His denition of bent functions accounted for functions from Zm to Z2 . Later, Kumar, Scholtz, and 2 Welch [51] generalized this denition to functions from Zm to Zq for arbitrary q q. Another generalization of bent functions appeared in [62], where the functions are dened from {0, 1}m to {0, 1/2, 1, 3/2} Z4 (as an additive group) and are = called real-valued bent functions. Let us denote the elements of G by 0 , 1 , . . . , n1 . We dene the character matrix (G) of G to be a matrix of size n n with entry i (j ) at position 121

6 Codes for Multicode CDMA


(i, j). It follows from the orthogonality of characters (cf. (6.1)) that (G) is an orthogonal matrix. It is not hard to show that (Zm ) and (Th,m ) are Sylvester2 m m type Hadamard matrices of size 2 2 , which differ from that in (2.12) only in the order of columns and rows. By convention, in the case where (G) is a Sylvester-type Hadamard matrix, we shall x the order of 0 , 1 , . . . , n1 , such that (G) coincides with the canonical form shown in (2.12) (if G = Z m , 2 then 0 , 1 , . . . , 2m 1 are the binary m-tuples in lexicographic order). It is now straightforward to prove the following simple theorem. Theorem 6.4. Let G be a nite abelian group of order 2m such that (G) is a Sylvester-type Hadamard matrix. Suppose q 2 and let f : G Z q . Dene the word f = (f (0 ), f (1 ), . . . , f (2m 1 )). Then we have PAPR(f ) = 1 max |Ff (u)|2 . 2m uG

Proof. Observe that the discrete-time MC-CDMA signal f (G) takes on the values of the Fourier transform of f . The theorem then follows from the denition of the PAPR. In particular we obtain: Corollary 6.5. With the notation as in Theorem 6.4, we have PAPR(f ) = 1 if and only if f is a bent function. Hence every constant-amplitude code must be comprised of words that correspond to bent functions. We remark that in the special case where q = 2, the preceding corollary has been similarly stated by Wada [113, Theorem 2] and by Paterson [81, Theorem 8]. 6.3 Cosets of RMq (1, m) In this section we consider codes that are unions of cosets of RMq (1, m). Theorem 6.6. Let f : {0, 1}m Zq be a generalized Boolean function that denes the coset C = f + RMq (1, m). Then PAPR(C) = 1 2 max |Lf (u)| , m uZm 2 q (6.2)

where Lf : Zm C is given in Denition 4.24. q 122

6.3 Cosets of RMq (1, m)


Proof. The q m+1 words in C correspond to the functions fu : {0, 1}m Zq given by fu (x) = f (x) + u x + , where u Zm and Zq . We have q PAPR(C) = and with Theorem 6.4, PAPR(C) = 1 2m
uZm , Zq v{0,1}m q uZm , Zq q

max

PAPR(fu ),

max

max |Ffu (v)|2


2

1 = m 2 1 2m

uZm , Zq v{0,1}m q

max

max

f q (x)+ux+ (1)vx x{0,1}m 2

uZm , Zq v{0,1}m q

max

max

f q (x)+[u+(q/2)v]x+ x{0,1}m 2

1 = m max 2 uZm q =

f q (x)ux x{0,1}m

1 max |Lf (u)|2 . 2m uZm q

We can now use Lemma 4.27 to establish general lower bounds on the PAPR of cosets of RMq (1, m). More specically, the lemma implies that for any quadratic function f : {0, 1}m Z2 the PAPR of the coset (q/2) f + RMq (1, m) is at least 2 if q 2 (mod 4) and m is odd or if q 0 (mod 4) and m is arbitrary. The next theorem shows that one cannot construct constant-amplitude codes from cosets of RMq (1, m) whenever q is a multiple of 4. Theorem 6.7. If q 0 (mod 4), any coset of RMq (1, m) has PAPR greater than 1. Proof (Sketch). Suppose f : {0, 1}m Zq , where q 0 (mod 4). To prove the theorem, it has to be shown that maxuZm |Lf (u)| is strictly greater than 2m/2 . It q is not hard to show that the multiset {Lf (u) | u (q/4) Zm } is equal to the mul4 m m tiset of values of the 2 possible {H, N } transforms of f [85]. It was shown in [85, Theorem 5 and the following remark] that there is no function from {0, 1} m to Zq for which all absolute values of the 2m {H, N }m transforms have magnitudes smaller than or equal to 2m/2 . 123

6 Codes for Multicode CDMA


Theorems 6.6 and 4.26 imply the following key result. Corollary 6.8. Let f : {0, 1}m Zq be a generalized Boolean function that denes the coset C = f + RMq (1, m). Then we have PAPR(C) PMEPR(C). The preceding corollary implies that all codes with low PMEPR that are unions of cosets of RMq (1, m) (e.g., all codes developed in Chapters 4 and 5 and in [22], [79]) also enjoy low PAPR. Now the question arises what is the difference between the PAPR and the PMEPR of cosets of RMq (1, m). This question is generally difcult to answer, but at least we can prove an asymptotic result for the case where q is a power of 2. Theorem 6.9. Let m be xed and let f : {0, 1}m Z2 be a generalized Boolean function that denes the coset C = 2h f + RM2h (1, m). Then, when h tends to innity, PAPR(C) = PMEPR(C). We need a lemma to prove the theorem. Lemma 6.10. Let a, b : {0, 1}m Z2h be two generalized Boolean functions that are related by
m1

b(x0 , . . . , xm1 ) = a(x0 , . . . , xm1 ) +


i=0

2 i xi .

Then their corresponding OFDM signals satisfy S(b)() = S(a)( + 1 2h (mod 1)).

Proof of Lemma. Observe that the ith elements of the words a and are related by b i = ai i h . b 2 Hence
2m 1

S(b)() =
i=0 2m 1

i ej2i b ai ej2i/2 ej2i


i=0 2m 1
h

=
i=0

ai ej2i(+1/2 1 2h

= S(a)( +

(mod 1)).

124

6.4 Some Families of Trace Codes


Proof of Theorem 6.9. Let the word a C be one word in the coset with maximum PMEPR, and suppose the maximum power occurs at time max . Then, according to Lemma 6.10, there exists another word in C, whose OFDM signal is equal to a cyclic shift by 1/2h to the left of that of a. Hence, since h tends to innity, there exists a word b C, whose OFDM signal is a cyclic shift of that of a, such that the maximum power occurs at = 0, i.e., S(b)(0) = S(a)(max ). Therefore PMEPR(C) = 1 |S(b)(0)|2 2m
2m 1 2

1 = m 2

i b
i=0

PAPR(C). But, according to Corollary 6.8, the PAPR of C must be equal to or less than its PMEPR. Hence equality must hold in the preceding expression. 6.4 Some Families of Trace Codes Let Rh,m be the Galois ring of characteristic 2h and cardinality 2hm . Every polynomial p(x) Rh,m [x] can be written uniquely in 2-adic expansion
h1

p(x) =
i=0

2i Pi (x),

Pi (x) Th,m [x],

(6.3)

which can be obtained from the 2-adic expansion of the coefcients of p(x) (cf. (2.2)). We say that p(x) is nondegenerate if its degree is at least 1 and if it contains no terms of even degree [52]. The weighted degree of p(x) is dened as [52] D(p) := max{2h1 d0 , 2h2 d1 , . . . , dh1 }, where di is the algebraic degree of Pi (x). The following result by Kumar, Helleseth, and Calderbank [52] provides estimates on so-called exponential sums over Galois rings. Theorem 6.11. [52] For any nondegenerate polynomial p(x) R h,m [x] we have 2 h
xTh,m Tr(p(x))

(D(p) 1)2m/2 .

125

6 Codes for Multicode CDMA


We now introduce the family of lengthened weighted-degree trace codes (cf. [83], [102], and for the case q = 4 [96]). Denition 6.12. For 1 D 2 m/2 we dene the lengthened weighted-degree trace code W(h, m, D) of length 2m to be the set of words corresponding to the functions f : Th,m Z2h with f (x) = + Tr(p(x)). Here Z2h and p(x) Rh,m [x] is of the form p(x) = p1 x + p3 x3 + + pr2 xr2 + pr xr , where r = 2 (D 1)/2 + 1 and D(p) D.

The code W(h, m, D) is linear (because of the linearity of the trace function, cf. Lemma 2.5 (ii)), and when h = 1, it coincides with the lengthened dual of the t-error-correcting primitive BCH code [60, Chapter 9, 9]. Theorem 6.13. |W(h, m, D)| = 2h+m(D
D/2h

).

Proof. The rst step is to observe that the cyclotomic cosets Cj are distinct for odd 1 j 2 m/2 and contain exactly m elements [60, Chapter 9, Theorem 7]. Since every function from Th,m to Z2h can be uniquely expressed as shown in (2.5), we conclude that choosing different polynomials p(x) in Denition 6.12 results in different codewords. Therefore the number of words in W(h, m, D) is equal to 2h times the number of distinct polynomials p(x). So it remains to count these polynomials. Suppose that p(x) is written in its unique 2-adic expansion, as shown in (6.3), and di is the algebraic degree of Pi (x). Then the condition D(p) D translates to 2hi1 di D, i = 0, 1, . . . , h 1. Since di must be odd, the preceding inequality implies di 2 D 2hi 1 + 1. 2

Because Pi (x) contains no terms of even degree, the number of distinct polynomials Pi (x) Th,m [x] satisfying the above condition is equal to 2m 126
D 2hi

+1 2

6.4 Some Families of Trace Codes


Using the easily veried identity x+ 1 = 2x x , 2 x R,

the number of polynomials p(x) Rh,m [x] of weighted degree at most D can then be written as
h1

2m
i=0

D 2hi

1 +2

= 2m = 2m

Ph1
i=0

D 2hi

+1 2

Ph1
i=0

D 2hi1

Ph1
i=0

D 2hi

= 2m(D

D 2h

).

Remark. For D 4 a (more difcult) proof of Theorem 6.13 was given by Sol e and Zinoviev in [103]. As shown above, the condition D 4 can be dropped. Theorem 6.14. For h {1, 2} the minimum distance of W(h, m, D) is at least 2h2 2m (D 1)2m/2 . Proof. By Lemma 2.8 we have to nd the minimum nonzero weight of the words in W(h, m, D). Let c W(h, m, D) and suppose that this word corresponds to the polynomial + Tr(p(x)). When p(x) = 0, we have wt(c) {0, 2 m } if h = 1 and wt(c) {0, 2m, 2m+1 } if h = 2. Otherwise p(x) is a nondegenerate polynomial of weighted degree at most D. Now simple calculation yields
2m 1

2h

wt(c) = 2 = 2m 2m

c 2 i h

i=0

2 h
xTh,m

2 h
Tr(p(x))

Tr(p(x))

2 h

xTh,m

2m (D 1)2m/2 , where the latter bound follows from Theorem 6.11. 127

6 Codes for Multicode CDMA


Remark. For general h a similar argument can be used to show that the squared minimum Euclidean distance of the PSK code associated with W(h, m, D) is at least min 2m+2 sin2 h , 2m+1 (D 1)2m/2+1 , 2 where the rst term in the minimum calculation considers the minimum distance between constant words in W(h, m, D), and the second term considers the minimum distance between a nonconstant word and any other word in W(h, m, D). Let us now focus on the subcode of W(h, m, D) whose words correspond to the functions f : Th,m Z2h given by f (x) = + Tr(2h1 ax), {0, 2h1}, a Th,m .

Let be the modulo-2 reduction map, and let be a generator for T h,m . Writing = (a) and letting = () be a primitive element in F2m , the algebraic normal form of the generalized Boolean function corresponding to f reads m1 m1

f
i=0

xi

= +2

h1

Tr
i=0 m1

xi i xi i

= + 2h1 Tr
i=0 m1

= + 2h1
i=0

Tr i xi .

It is seen that the set of words associated with the functions f is the code 2h1 RM2 (1, m). Hence W(h, m, D) is a union of cosets of RM2 (1, m) and W(1, m, 1) = RM2 (1, m). If h 2, a similar argument reveals that W(h, m, D) is a union of cosets of ZRM4 (1, m) and W(2, m, 1) = ZRM4 (1, m). Now we use Theorem 6.11 to bound the PAPR of a large subcode of W(h, m, D). Theorem 6.15. Suppose that D by removing the 2h1 cosets 2+ 1/h . The expurgated code that is obtained

+ 2h1 RM2 (1, m),

= 0, 1, . . . , 2h1 1

from W(h, m, D) has PAPR at most (D 1)2 . 128

6.4 Some Families of Trace Codes


Proof. The condition D 2 + 1/h implies that after the expurgation we have a nonempty set. As in Denition 6.12, let f (x) = + Tr(p(x)) denote the polynomials associated with the words in W(h, m, D). Their Fourier transforms read Ff (u) = 2h (1)Tr(ux)
xTh,m f (x)

= 2h

2 h
xTh,m

Tr(p(x)+2h1 ux)

If p(x) = 2h1 ax, we have |Ff (a)| = 2m . Otherwise the polynomial p(x) + 2h1 ux is nondegenerate and has weighted degree at most D. Hence, by Theorem 6.11, |Ff (u)| (D 1)2m/2 for all u Th,m . Theorem 6.4 then implies that the PAPR of the associated sequences is at most (D 1)2 . It is easy to verify that the expurgation in Theorem 6.15 has the effect that, compared to the full code W(h, m, D), one can encode one bit less per codeword. For h {1, 2} the properties of the code in Theorem 6.15 are summarized below: length: rate: minimum distance: PAPR: 2m m(D D/2h ) + h 1 /2m 2h2 2m (D 1)2m/2 (D 1)2

It is clear that any lengthened trace code of length 2m (a code whose nonconstant words correspond to the trace of nondegenerate polynomials in R h,m [x]) can be understood as a subcode of W(h, m, D) for appropriate h and D. Therefore the PAPR bound in Theorem 6.15 as well as bounds on the minimum distance can be translated into bounds that hold for any particular trace code. Next we recall the family of quaternary DelsarteGoethals codes from [36] (for the trace description of these codes see also [39]). These codes play an important role in Section 6.6. Denition 6.16. For m 3 and any integer 0 t (m 1)/2 we dene the quaternary DelsarteGoethals code DG(t, m) of length 2m to be the set of words associated with the functions f : T2,m Z4 given by
t

f (x) = + Tr ax + 2

b i x2
i=1

+1

Z4 , a R2,m , bi T2,m . (6.4) 129

6 Codes for Multicode CDMA


We also dene K(m) := DG(0, m) to be the quaternary Kerdock code. We remark that, typically, the above dened codes are called quaternary Kerdock and DelsarteGoethals only when m is odd; we shall use the name for any m 3. It can be shown [36] that when m is odd, (K(m)) and (DG(t, m)) are the classical nonlinear Kerdock and DelsarteGoethals codes [60, Chapter 15], respectively. The code DG(t, m) is linear and contains 4m+1 2tm codewords. By expanding the functions in (6.4) into their algebraic normal forms, it becomes obvious that DG(t, m) is a subcode of ZRM4 (2, m) and DG( m1 , m) = ZRM4 (2, m) if m is 2 odd. More generally, DG(t, m) is a union of cosets of ZRM4 (1, m) and generates a nested chain of codes ZRM4 (1, m) K(m) DG(1, m)

DG( (m 1)/2 , m) ZRM4 (2, m).

Since DG(t, m) is a subcode of W(4, m, 2t + 1), with Theorem 6.14 the minimum distance of the quaternary code DG(t, m) can be lower bounded by 2m 2t+m/2 . This bound can be improved for odd m [40], and the true minimum distance of DG(t, m) is equal to 2m 2t+ Moreover Theorem 6.15 implies: Corollary 6.17. The code DG(t, m) \ ZRM4 (1, m) has PAPR at most 4t . Note that with the code in the preceding corollary one can encode m(t + 2) + 1 bits per codeword, as opposed to m(t + 2) + 2 bits for the full code DG(t, m). The parameters of the code DG(t, m) \ ZRM4 (1, m) are: length: rate: minimum distance: PAPR: 2m [m(t + 2) + 1] /2m 2m 2t+ 4t
m/2 m/2

130

6.5 Z4 -valued Bent Functions


6.5 Z4 -valued Bent Functions We have seen in Corollary 6.5 that every word of a constant-amplitude code corresponds to a bent function. For this reason, we study Z4 -valued bent functions in this section. We will summarize a few implicitly known constructions and establish several new constructions of such functions. Throughout this section we let := 4 be a primitive fourth root of unity in C. Primary Constructions It is known that Z2 -valued bent functions cannot exist for lengths 2m when m is odd (because the Fourier transform of a Z2 -valued Boolean function must contain integers, so it cannot attain the values 2m/2 when m is odd). The following theorem shows that Z4 -valued bent functions exist for all lengths 2m , where m is an arbitrary positive integer. The theorem essentially coincides with [63, Theorem 2], in which the construction applies to real-valued bent functions. Theorem 6.18. The function f : {0, 1}m Z4 given by f (x0 , x1 , . . . , xm1 ) = x0 + x1 + + xm1 is a bent function. Proof. We have Ff (u) = =
x{0,1}m i=0 m1

x{0,1}m

Pm1
i=0

xi +2ui xi

m1

xi +2ui xi

=
i=0 xi {0,1} m1

xi +2ui xi

=
i=0

1 + 1+2ui
m1 m/2 i=0

=2

ej(1)

ui

/4

It follows that |Ff (u)| = 2m/2 for each u {0, 1}m, as it was to be shown. 131

6 Codes for Multicode CDMA


The next theorem describes a version of the MaioranaMcFarland construction (see, e.g., [60, Chapter 14, Problem (20)]) over Z4 . Notice that a special case of this construction has been stated in [62, Theorem 1] in terms of real-valued bent functions. This case is recovered by setting equal to the identity map. We remark that our construction is different from that in [118], where another version of the MaioranaMcFarland construction over Z4 is given. Theorem 6.19. Let be a permutation on {0, 1}k , let g : {0, 1}k Z4 be arbitrary, and let f : {0, 1}2k Z4 be given by f (x, y) = 2 (x) y + g(x). Then f is a bent function. Proof. Write Ff (u, v) = =
x{0,1}k

2(x)y+g(x) (1)ux+vy
x,y{0,1}k

g(x) (1)ux g(
1

(1)((x)+v)y
y{0,1}k
1

=
x{0,1}k

(x))

(1)u

(x)

(1)(x+v)y .
y{0,1}k

If x = v, the exponent in the inner sum is a linear form, which takes the values 0 and 1 equally often (cf. the randomization lemma [60, page 372]). Hence the inner sum is zero if x = v. On the other hand, if x = v, the sum is trivial and becomes equal to 2k . Thus Ff (u, v) = 2k g( and the proof is completed. Now we turn our attention to functions dened on T2,m . Theorem 6.20. Suppose m 1 and Z4 . Let f : T2,m Z4 be given by f (x) = + Tr(ax). Then f is a bent function if and only if a R2,m \ 2T2,m . 132
1

(v))

(1)u

(v)

6.5 Z4 -valued Bent Functions


Proof. The Fourier transform of f is given by Ff (u) = Tr([a+2u]x) .
xT2,m

The distribution of the absolute values of the exponential sum (b) =


xT2,m

Tr(bx) ,

b R2,m

is completely known (cf., e.g., [39, Lemma 8.2]). In particular we have | (b)| = 2m/2 if b 2T2,m . Hence |Ff (u)| = 2m/2 for each u T2,m if a 2T2,m . On the / / other hand, if a 2T2,m , then there exists a u T2,m , such that Ff (u) = 2m . Theorem 6.21. Suppose m 3 and let f : T2,m Z4 be given by
Z4 , a R2,m , b T2,m .

f (x) = + Tr(ax + 2bx3 ),

Then f is a bent function if (a) = 0 and the equation x3 + 1 =0 (b) (6.5)

has no solution in F2m , or if (a) = 0 and the equation x3 + x + (b)2 =0 (a)6 (6.6)

has no solution in F2m . Here, is the modulo-2 reduction map. Proof. Consider the exponential sum (a, b) =
xT2,m

Tr(ax+2bx ) .

We have 2 (a, b) =
x,yT2,m

Tr(a[x+y])+2Tr(b[x

+y 3 ])

(6.7)

Set z := x y. 133

6 Codes for Multicode CDMA


With the denition of the operation in (2.3), this expression can be expanded as follows z = (x + y)2
m

= (x + y)2

2m1 2m2

= (x2 + y 2 + 2xy)2 . . .

= x2 + y 2 + 2(xy)2 = x + y + 2 xy,

m1

which together with the relation y = x + z (mod 2) implies x + y = z + 2x + 2 xz. Using Lemma 2.5 we obtain Tr(a[x + y]) = Tr(a[z + 2x + 2 xz]) = Tr(az) + 2Tr(ax) + 2Tr(a xz) = Tr(az) + 2Tr(a2 x2 ) + 2Tr(a4 x2 z 2 ) = Tr(az) + 2Tr(x2 [a2 + a4 z 2 ]) and 2Tr(b[x3 + y 3 ]) = 2Tr(b[x3 + (x + z + 2 xz)3 ]) = 2Tr(b[z 3 + x2 z + xz 2 ]) = 2Tr(bz 3 ) + 2Tr(bx2 z) + 2Tr(bxz 2 ) = 2Tr(bz 3 ) + 2Tr(bx2 z) + 2Tr(b2 x2 z 4 ) = 2Tr(bz 3 ) + 2Tr(x2 [bz + b2 z 4 ]). Since T2,m , is an abelian group (cf. Lemma 2.4), z takes on every value in T2,m when x is xed and y ranges over T2,m . Hence (6.7) can be written as 2 (a, b) =
zT2,m
3 Tr(az+2bz )

2Tr(b
xT2,m

x2 h(z))

(6.8)

where h(z) = z 4 + 134

a4 2 1 a2 z + z+ 2. b2 b b

6.5 Z4 -valued Bent Functions


The inner sum in (6.8) becomes 2Tr(b
xT2,m
2

x2 h(z))

=
xT2,m

(1)Tr((b (1)
xF2m

x2 h(z)))

= =
xF2m

Tr b h(z) x Tr (b) h(z) x

(1)

where h(z) := (h(z)). By the equi-distribution property of the trace function (cf. Lemma 2.5 (v)), it is now seen that the above sum is zero if h(z) = 0. On the other hand, if h(z) = 0, the sum is equal to 2m . Hence the absolute value of the whole sum in (6.8) is bounded above by 2m times the number of solutions of the quartic equation h(z) = 0 over F2m . Now we aim at characterizing the cases where h(z) has exactly one root, which 2 m implies that | (a, b)| = 2 . Write = (a) and = (b). First consider the case = 0. Then we have 1 h(z) = z z 3 + .

In this case h(z) has exactly one root if x3 + 1/ = 0 has no solution in F2m . Now let = 0. An afne variable substitution (which preserves the number of roots) yields 8 2 2 [1 + z] = z z 3 + z + . h 4 6 It is seen that h(z) has exactly one root if x3 + x + 2 = 0, 6

has no solution in F2m . In order to complete the proof, observe that the number of roots of h(z) only depends on (a) and (b), and therefore, if h(z) has exactly one root, then the absolute values of the Fourier transform of f Ff (u) = Tr([a+2u]x+2bx
xT2,m
3

are equal to 2m/2 for each u T2,m , and f is a bent function. 135

6 Codes for Multicode CDMA


Remark. In the case where (a) = 0 and the equation (6.5) is not solvable, f belongs to the well-known family of Gold bent functions (see, e.g., [14]). Secondary Constructions In what follows we construct new bent functions from known (mostly Z 2 -valued) bent functions. We begin with a simple lemma, whose proof is obvious and omitted. Lemma 6.22. If g : {0, 1}m Z4 is a bent function, then the function f : {0, 1}m Z4 given by f (x) = g(x) + 2 c x + , is also a bent function. A number of constructions for Z2 -valued bent functions are known. For a good overview we refer to the survey paper by Carlet [14]. In the remainder of this section we describe several mappings from Z2 -valued to Z4 -valued bent functions. To prove our results, we will frequently use the identities { x } = 1 [1 + (1)x ] 2 1 { x } = [1 (1)x ] 2 c Zm , 2 Z4

for x {0, 1}. The following two theorems cover constructions of Z4 -valued bent functions dened on {0, 1}m for even m. Theorem 6.23. Suppose that g : {0, 1}m Z2 is a bent function. Then the function f : {0, 1}m Z4 with
m1

f (x) = 2g(x) + 2
0 i<j<m

a i a j xi xj +
i=0

a i xi ,

ai {0, 1}

is a bent function. Proof. Write a = (a0 , a1 , . . . , am1 ) and observe that


m1

a x (mod 2) =

a i xi + 2
i=0 0 i<j<m

a i a j xi xj

(mod 4).

136

6.5 Z4 -valued Bent Functions


Therefore the Fourier transform of f reads Ff (u) = We have {Ff (u)} = = and similarly, {Ff (u)} = Hence |Ff (u)|2 = 1 [Fg (u) Fg (u + a)] . 2 1 [Fg (u)]2 + [Fg (u + a)]2 . 2 1 2 (1)g(x)+ux [1 + (1)ax ]
x{0,1}m

x{0,1}m

(1)g(x)+ux ax (mod 2) .

1 [Fg (u) + Fg (u + a)] , 2

By assumption we have [Fg (u)]2 = 2m for every u {0, 1}m, which together with the above relation implies |Ff (u)| = 2m/2 for each u {0, 1}m. Now recall from Section 2.3 that for any function g : {0, 1}m+1 Z2 its inverse Gray map 1 (g) is a function from {0, 1}m to Z4 such that its associated word is equal to the inverse Gray map of the word associated with g. Theorem 6.24. Suppose that a, b : {0, 1}m Z2 are bent functions, and dene g : {0, 1}m+1 Z2 by g(x, y) = a(x)(1 + y) + b(x)y. Then f = 1 (g) is a bent function. Proof. By expressing g(x, y) as g(x, y) = a(x) + [a(x) + b(x)]y, we know from Section 2.3 that f (x) can be written as f (x) = 2a(x) + [a(x) + b(x) (mod 2)]. 137

6 Codes for Multicode CDMA


Next we inspect the Fourier transform of f {Ff (u)} = = (1)a(x)+ux
x{0,1}m

a(x)+b(x)

(mod 2)

1 2

(1)a(x)+ux 1 + (1)a(x)+b(x)
x{0,1}m

= = Similarly we obtain {Ff (u)} =

1 [Fa (u) + Fb (u)] . 2

1 2

(1)a(x)+ux +
x{0,1}m x{0,1}m

(1)b(x)+ux

1 [Fa (u) Fb (u)] . 2

Since Fa (u) and Fb (u) take on only the values 2m/2 , it follows |Ff (u)| = 2m/2 . The next two theorems describe mappings from Z2 -valued bent functions to Z4 valued bent functions dened on {0, 1}m for odd m. The theorem below has been stated similarly for real-valued bent functions in [63, Theorem 3]. Theorem 6.25. Suppose m 1 and let a, b : {0, 1}m1 Z2 be bent functions. Let f : {0, 1}m Z4 be given by f (x, y) = 2a(x)(1 + y) + 2b(x)y + y. Then f is a bent function. Proof. We have Ff (u, v) = =
x{0,1}m1

2a(x)(1+y)+2b(x)y+y (1)ux+vy
x{0,1}m1 ,y{0,1}

(1)a(x)+ux + (1)v

(1)b(x)+ux
x{0,1}m1

= Fa (u) + (1)v Fb (u). Since a and b are bent functions, we see that | {Ff (u, v)}| = | {Ff (u, v)}| = 2(m1)/2 , which implies that |Ff (u, v)| = 2m/2 . 138

6.6 Quaternary Constant-Amplitude Codes


The following theorem shows that the inverse Gray map of a Z2 -valued bent function (cf. Section 2.3) is also bent. Theorem 6.26. Suppose m 1 and g : {0, 1}m+1 Z2 is a bent function. Then f = 1 (g) is also a bent function. Proof. Recall from Section 2.3 that, when g is written as g(x, y) = a(x) y + b(x), where a, b : {0, 1}m {0, 1}, then f is given by f (x) = a(x) + 2b(x). The Fourier transform of g can be written as Fg (u, v) = =
x{0,1}m

(1)a(x) y+b(x)+ux+vy
x{0,1}m ,y{0,1}

(1)b(x)+ux 1 + (1)v+a(x) .

Now observe {Ff (u)} = = = and similarly, {Ff (u)} = 1 Fg (u, 1). 2 (1)b(x)+ux { a(x) } (1)b(x)+ux 1 + (1)a(x)
x{0,1}m

x{0,1}m

1 2

1 Fg (u, 0), 2

Since |Fg (u, v)| = 2(m+1)/2 , we have |Ff (u)| = 2m/2 . 6.6 Quaternary Constant-Amplitude Codes In this section we use our results on Z4 -valued bent functions to construct families of quaternary constant-amplitude codes. In the light of Lemma 6.22 it seems natural that all codes derived in this section are unions of cosets of ZRM 4 (1, m). For practical reasons we let the number of such cosets be a power of 2. This allows to encode a whole number of information bits. Decoding of these codes can be performed by using the supercode decoding method, which reduces the decoding 139

6 Codes for Multicode CDMA


task to several decodings of ZRM4 (1, m), which can be performed as described below. The rst step is to observe that for any a Z4 we have
a 2 {ej/4 4 } = (1)(a) a 2 {ej/4 4 } = (1)(a) ,

where (a) and (a) is the left and the right half of (a), respectively (cf. Denition 2.1). The next step is to recall the relation (ZRM 4 (1, m)) = RM2 (1, m + 1) m from [36]. Now let c ZRM4 (1, m) and e C2 be an error vector, so we wish to recover c from y = c + e. By the preceding observations, we have 2 ( {yej/4 } | {yej/4 }) = ((c) | (c)) + e = (c) + e ,

where e R2

m+1

We can now use any decoder for RM2 (1, m + 1) (e.g., Algorithm 4.34) to get an estimate for (c) and, therefore, also for c. It is straightforward to verify that, if we use an ML decoder to decode RM2 (1, m + 1), then the above described scheme performs ML decoding for ZRM4 (1, m) as well. Quaternary Codes From Binary Codes Next we describe mappings from binary constant-amplitude codes to quaternary constant-amplitude codes. These can be used in connection with the results obtained by Paterson [81, Section V], where various binary constant-amplitude codes have been constructed. Recall from Corollary 6.5 that any binary constantamplitude code must comprise bent functions. Theorems 6.246.26 suggest how we can obtain quaternary constant-amplitude codes from binary codes that are entirely comprised of bent functions. The corollary below is a consequence of Theorem 6.25. Corollary 6.27. Let B be a binary constant-amplitude code of length 2 m1 (so m is necessarily odd). Then the code Q = {(2a + 1 | 2b + ( + 1) 1) | a, b B, {0, 1}} of length 2m is a quaternary constant-amplitude code with d(Q) = 2d(B) and R(Q) = R(B) + 1/2m . Here, 1 is the all-one word of length 2m1 . 140

6.6 Quaternary Constant-Amplitude Codes


Observe that the quaternary code in the preceding corollary effectively is an offset binary code that is obtained by concatenating two suitably offset codewords of a binary constant-amplitude code. This has to be compared with [81, Lemma 19 and the remarks thereafter], where essentially the same code construction is suggested resulting in, however, codes with PAPR equal to 2. The next two corollaries follow from Theorem 6.24 and Theorem 6.26 and the distance-preserving property of the Gray map (cf. Theorem 2.2). Corollary 6.28. Let B be a binary constant-amplitude code of length 2 m (so m is necessarily even). Then the quaternary code Q = 1 ({(a | b) | a, b B}) of length 2m is a constant-amplitude code with d(Q) = d(B) and R(Q) = 2R(B). Corollary 6.29. Let B be a binary constant-amplitude code of length 2 m+1 (so m must be odd). Then Q = 1 (B) is a quaternary constant-amplitude code of length 2m with d(Q) = d(B) and R(Q) = 2R(B). It is not hard to see that, if Corollaries 6.276.29 are applied to binary codes that are unions of cosets of RM2 (1, m), then the resulting quaternary codes are unions of cosets of ZRM4 (1, m(1)). Codes From Generalized ReedMuller Codes In what follows we relate Z4 -valued bent functions constructed in Section 6.5 with the quaternary ReedMuller codes RM 4 (r, m) and ZRM4 (r, m) (cf. Denition 4.1). Our rst construction takes a single coset of ZRM4 (1, m). Construction 6.30. By Theorem 6.18 and Lemma 6.22, we can construct a constant-amplitude code as a single coset of ZRM4 (1, m) with coset represenm1 tative corresponding to the polynomial i=0 xi (over Z4 ). This code contains 2m+2 codewords, and has minimum distance equal to 2m . Our next code construction exploits Theorem 6.19 and requires a lemma. Lemma 6.31. With the notation as in Theorem 6.19 we have f RM 4 (k, 2k). Proof. Observe that the function h : {0, 1}2k Z2 given by h(x, y) = (x) y is a Z2 -valued bent function, whose degree cannot exceed k [87]. Hence, since g obviously has degree at most k, the degree of f is bounded by k. 141

6 Codes for Multicode CDMA


Construction 6.32. For even m Theorem 6.19 identies 2m/2 ! 42 bent functions, which organize in cosets of ZRM4 (1, m) inside RM4 (m/2, m). The union of them is a constant-amplitude code of length 2m with minimum distance 2m/2 and rate (2m/2+1 + log2 2m/2 ! )/2m . For m = 4, 6, 8 the rate of this code is equal to 12/16, 31/64, and 76/256, respectively. By restricting the functions g in Theorem 6.19 such that the coefcient of the monomial of degree m/2 in the algebraic normal form of g is even, we can construct a subcode of the code dened above that is contained in ZRM 4 (m/2, m). This code has half as many codewords as the code contained in RM 4 (m/2, m), but minimum distance 2m/2+1 . The rate of this code is equal to 11/16, 30/64, and 75/256 for m = 4, 6, 8, respectively. It is now natural to look for subcodes of the codes derived above that are contained inside quaternary ReedMuller codes of lower order. In this way the minimum distance can be traded against code rate. This is generally a difcult task except for the case r = 2, where in Theorem 6.19 the degree of g is at most 2 and can be written as (x) = x A, where A is a nonsingular binary matrix of size k k. However this code has weaker parameters than that discussed below. Next we wish to construct subcodes of ZRM4 (2, m). First consider the code m RM2 (2, m), which consists of 2( 2 ) cosets of RM2 (1, m). The number of such cosets that are comprised of bent functions is zero for odd m and equal to
m/21
m/2

N (m) = 2

m/2(m/21)

i=0

22i+1 1

for even m [60, Chapter 15, Theorem 2]. Let m 2 be even and Q : {0, 1} m Z2 be a representative of such a coset. Then, by Theorem 6.26, 1 (Q) is a Z4 valued bent function. Since 1 (RM2 (2, m)) = ZRM4 (2, m 1) [36], any such function denes a coset of ZRM4 (1, m 1) inside ZRM4 (2, m 1) with PAPR equal to 1. On the other hand, we deduce from Theorem 6.23 that the function Q : {0, 1}m Z4 given by
m1 i=0

is also a bent function, which denes a coset of ZRM4 (1, m) inside ZRM4 (2, m) with PAPR equal to 1. Hence we obtain the following coding option. 142

Q (x) = 2 Q(x) +

0 i<j<m

a i a j xi xj +

a i xi ,

ai {0, 1}

6.6 Quaternary Constant-Amplitude Codes


Construction 6.33. As described above, if m is odd, ZRM4 (2, m) contains N (m+1) cosets of ZRM4 (1, m) with PAPR equal to 1. If m is even, ZRM4 (2, m) contains 2m N (m) such cosets. The union of these cosets is a constant-amplitude code of length 2m with minimum distance 2m1 and rate ( log2 N (m + 1) + m + 2)/2m ( log2 N (m) + 2m + 2)/2m if m is odd if m is even.

For m = 4, 5, 6, 7, 8 the rate is equal to 14/16, 20/32, 27/64, 35/128, 44/256, respectively. A crude estimate yields N (m) 2m(m/21) . Thus for either m the rate of this code is bounded below by ( m2 /2 + m + 2)/2m. Codes From Quaternary Kerdock and DelsarteGoethals Codes The following construction identies a large constant-amplitude subcode of the quaternary Kerdock code K(m). It is a direct consequence of Corollary 6.17. Alternatively it can be obtained with Theorem 6.20 together with the denition of K(m) (cf. Denition 6.16). Construction 6.34. The code K(m) consists of 2m cosets of ZRM4 (1, m), one of them is ZRM4 (1, m) itself. The union of 2m1 remaining cosets is a quaternary constant-amplitude code of length 2m with rate (2m + 1)/2m and minimum distance 2m 2 m/2 . In what follows we wish to construct a constant-amplitude subcode of the quaternary DelsarteGoethals code DG(1, m). This code is a union of 2 2m cosets of ZRM4 (1, m) and contains 23m+2 codewords. By removing the codewords of ZRM4 (1, m) from DG(1, m), we obtain a code with PAPR at most 4 (cf. Corollary 6.17), which can be used to encode 3m + 1 bits. As we shall see below, by removing further cosets of ZRM4 (1, m) from DG(1, m), we can construct a constant-amplitude code that can be used to encode 3m bits, which is only 2 bits less than for the original code DG(1, m). Our result follows from Theorems 6.20 and 6.21 together with counting the number of cases where the cubic equations (6.5) and (6.6) are not solvable. The treatment of the cubic in (6.5) is straightforward. Lemma 6.35. The number of elements a F2m for which the equation x3 + a = 0 143

6 Codes for Multicode CDMA


over F2m has no solution is equal to m is odd.
2 3

(2m 1) if m is even and equal to zero if

Proof. The equation is solvable if and only if a S = {x3 | x F2m }. So there are 2m |S| cases where the equation has no solution. Note that x x3 is a bijection on F2m if m is odd, thus, S = F2m . When m is even, 3 divides 2m 1, m which implies that S = {0, 1, 3 , 6 , . . . , 2 4 }, where is a primitive element in F2m . Hence |S| = (2m 1)/3 + 1. The following lemma is due to Kumar et al. [53, Appendix] and provides the number of cases where the cubic equation in (6.6) has no solution. Lemma 6.36. [53] The number of elements a F2m for which the equation x3 + x + a = 0 over F2m has no solution is equal to (2m (1)m )/3. Now we derive a lower bound for the number of words inside DG(1, m) that correspond to bent functions. Theorem 6.20 identies 4m+1 2m+2 bent functions inside DG(1, m). In addition, by applying Lemmas 6.35 and 6.36, Theorem 6.21 identies (4m+1 2m+2 )(2m +1)/3 bent functions in DG(1, m) for either m 3. Hence C(3, m) contains at least 4m+1 2m+2 = 8m > 8m 1+ 2m + 1 3 1 3 4m2

4 1 + m2 3 2

bent functions. It is not hard to see that these organize in cosets of ZRM 4 (1, m). The preceding analysis gives rise to the following code construction. Construction 6.37. As shown above, DG(1, m) contains 22m2 cosets of ZRM4 (1, m) that are comprised of words associated with bent functions. The union of these cosets is a constant-amplitude code of length 2 m , rate 3m/2m , and minimum distance 2m 21+ m/2 . For m = 4, 5, 6, 7 the parameters of several constant-amplitude codes constructed in this section are presented in Tables 6.16.4, respectively. These tables 144

6.7 Conclusions, Notes, and Open Problems


illustrate the trade-off between code rate and minimum distance. Notice that we omitted those codes that are outperformed by one of the codes shown in the tables. In particular the codes owing from Construction 6.32 are notoriously surpassed by other codes; they appear to be rather weak at least for small m. 6.7 Conclusions, Notes, and Open Problems In this chapter we pursued three approaches to solve the PAPR problem in MCCDMA. In our rst approach we studied the PAPR of cosets of RM q (1, m), and concluded that the PAPR of any union of them is bounded above by its PMEPR. Therefore the codes developed in Chapters 4 and 5 and those in [22], [79] are also suited for the application in MC-CDMA. It was also shown that this concept cannot be used to construct constant-amplitude codes whenever q is a multiple of 4. Our second approach was to bound the PAPR of appropriately expurgated trace codes using a theorem on exponential sums. The PAPR, however, grows rapidly when the rate of the codes increases. In our third approach we exploited the connection between bent functions and words with PAPR equal to 1 in order to construct a variety of quaternary constantamplitude codes. These codes allow a rich trade-off between minimum distance and code rate. The constructions also provide answers to [80, Open Problem 9], which is restated in [81]. Next we comment on some relations to the work by Riera and Parker [85]. In [85] the bent4 property of Boolean functions was introduced. With the notation used in this chapter, a function f : {0, 1}m Z2 is bent4 if there exists a v {0, 1}m such that the function given by 2f (x) + v x is a Z4 -valued bent function. It was proved in [85] that every Boolean function with degree smaller than or equal to 2 is bent4 . This result implies that there exists an injective map that maps any word in RM2 (2, m) to a word in ZRM4 (2, m) with PAPR equal to 1. So we can construct a quaternary constant-amplitude code with rate m + m + 1 /2m 2 and minimum distance 2m1 . However with the relation m = m2 m /2 it 2 becomes clear that this code has weaker parameters than the constant-amplitude subcodes of ZRM4 (2, m) from Construction 6.33. Finally we wish to compare code rates and minimum Euclidean distances of quaternary codes constructed in this chapter with those of previously constructed binary codes. It can be observed that all binary constant-amplitude codes proposed 145

6 Codes for Multicode CDMA

Table 6.1: Constant-Amplitude Codes for m = 4 code rate 6/16 9/16 14/16 18/16 min. dist. 16 12 8 4 reference Construction 6.30, single coset of ZRM4 (1, 4) Construction 6.34, subcode of K(4) Construction 6.33, subcode of ZRM4 (2, 4) [81, Construction 12] + Corollary 6.28

Table 6.2: Constant-Amplitude Codes for m = 5 code rate 7/32 11/32 15/32 20/32 23/32 min. dist. 32 28 24 16 8 reference Construction 6.30, single coset of ZRM4 (1, 5) Construction 6.34, subcode of K(5) Construction 6.37, subcode of DG(1, 5) Construction 6.33, subcode of ZRM4 (2, 5) [81, Construction 13] + Corollary 6.29

Table 6.3: Constant-Amplitude Codes for m = 6 code rate 8/64 13/64 18/64 27/64 30/64 40/64 46/64 min. dist. 64 56 48 32 24 16 8 reference Construction 6.30, single coset of ZRM4 (1, 6) Construction 6.34, subcode of K(6) Construction 6.37, subcode of DG(1, 6) Construction 6.33, subcode of ZRM4 (2, 6) [81, Construction 18] + Corollary 6.28 [81, Construction 12] + Corollary 6.28 [81, Construction 13] + Corollary 6.28

Table 6.4: Constant-Amplitude Codes for m = 7 code rate 9/128 15/128 21/128 35/128 41/128 60/128 min. dist. 128 120 112 64 32 16 reference Construction 6.30, single coset of ZRM4 (1, 7) Construction 6.34, subcode of K(7) Construction 6.37, subcode of DG(1, 7) Construction 6.33, subcode of ZRM4 (2, 7) [81, Construction 12] + Corollary 6.27 [81, Construction 13] + Corollary 6.29

146

6.7 Conclusions, Notes, and Open Problems


m4 in [81, Section V] except for [81, Construction 18 when h 2 ] (in which case we have no comparable quaternary code) are outperformed by a quaternary constant-amplitude code constructed in Section 6.6. For example, [81] reports a code of length 64 with minimum Euclidean distance 32 and rate 23/64, while Table 6.3 contains a quaternary code with the same length and minimum Euclidean distance but rate 40/64. Sol and Zinoviev [103] constructed binary codes derived from the most signife icant bit of weighted-degree trace codes over Z2l , which are similarly dened as W(h, m, D). A comparison of these codes with the quaternary codes identied in Theorem 6.15 shows that for D > 3 the codes in [103] have slightly higher rate, but lower minimum Euclidean distance and higher PAPR. For instance, setting l = 3 in [103] yields a code with 2 7D/8 m codewords, PAPR at most 11.66 (D 1)2 , and minimum Euclidean distance at least 2m+1 11.66(D 1)2m/2+1, while the quaternary code in Theorem 6.15 contains 2 3D/4 m+1 codewords, has PAPR at most (D 1)2 and minimum Euclidean distance at least 2m+1 (D 1)2m/2+1 . We close with some comments and open problems related to constant-amplitude codes.

We have shown that 22m+1 out of the 22m+2 words in K(m) = DG(0, m) correspond to bent functions. Similarly, at least 23m out of the 23m+2 words in DG(1, m) have PAPR equal to 1. Moreover the number of words associated with bent functions inside the classical DelsarteGoethals code (DG(t, m)) (when m is even) is completely known [60, Chapter 21, 8], [81]. Hence from the results in [81] and Theorem 6.26 we deduce that, when m is odd, at least 2m(t+2) out of the 2m(t+2)+2 words in DG(t, m) correspond to bent functions. So the problem that remains to be solved is: how many words associated with bent functions exist in DG(t, m) for t > 1 and even m? Note that the rst interesting cases are the codes DG(2, 8), DG(2, 10), and DG(3, 10) since, when m is even, DG( m 1, m) has the 2 same minimum distance as ZRM4 (2, m) while it contains less codewords than ZRM4 (2, m). Now we comment on possible ways to obtain general q-ary MC-CDMA constant-amplitude codes for q > 4. In order to address this issue, one has to investigate bent functions from {0, 1}m to Zq . We are aware of only one construction of such functions. This is a straightforward generalization of Theorem 6.19. 147

6 Codes for Multicode CDMA


Theorem 6.38. Suppose that is a permutation on {0, 1}k and q 2 is k even. Let g : {0, 1} Zq be an arbitrary generalized Boolean function, and let f : {0, 1}2k Zq be given by f (x, y) = Then f is a bent function. The functions identied in the preceding theorem organize in cosets of (q/2)RM2 (1, m) inside the RMq (m/2, m). It is now possible to formulate a generalization of Construction 6.32 for general q-ary (q even) alphabets. It would be interesting to nd further constructions of bent functions from {0, 1}m to Zq and to place these constructions in algebraic codes over Zq . In Section 6.6 we provided three mappings from binary to quaternary constant-amplitude codes. Are there mappings from binary (or quaternary) to octary or, more generally, to q-ary constant-amplitude codes? q (x) y + g(x). 2

148

Chapter 7

Summary

In this thesis we addressed the issue of reducing the peak-to-mean power ratio in multicarrier systems by exploiting the redundancy of block codes. In Chapters 36 we have provided at least partial answers to the power-control problems in OFDM and MC-CDMA stated in Sections 2.7 and 2.8, respectively. The principal results are summarized below. In Corollary 3.5 we provided a lower bound on the PMEPR of phase-shifted binary codes. The bound can be sharpened (cf. Theorem 3.4) when the alphabet of the phase shifts is taken into account. We have demonstrated by means of several examples that suboptimal algorithms from the literature often produce phase shifts that are close to the optimum ones. We also provided examples of optimum phase shifts. These include binary phase shifts for short BCH codes and quaternary phase shifts for nonredundant binary transmission and large numbers of carriers. Corollary 4.23 identies an innite family of cosets of RM q (1, m) with low PMEPR, once a suitable kernel is known. By presenting some simple kernels, we could state the relation of certain Golay sequences and cosets of RMq (1, m) in the context of a more general framework (Corollary 4.30), and more importantly, we found many new coset families with PMEPR bounded by values between 2 and 4 (Corollary 4.31). We note that, in contrast, all previously known bounds on the PMEPR of such cosets are powers of 2. We deduced several OFDM coding schemes with low PMEPR and provided theoretical support for some coding schemes reported in [22] based on empirical observations. 149

7 Summary
In Theorem 5.7 we have related generalized Boolean functions with complementary sets. As a consequence, Corollary 5.8 provides an upper bound on the PMEPR of any coset of RMq (1, m). This result generalizes [79, Corollary 13] from cosets lying in RMq (2, m) to arbitrary cosets. Based on Corollary 5.8, Constructions 5.23 and 5.25 propose two families of OFDM codes with PMEPR bounded by a power of 2. The rst one is a single coset and the second one is a union of cosets of a large linear code such that the number of cosets is kept low at least for small to medium lengths. Compared to the approaches in [22] and [79], this makes the codes more amenable to efcient decoding algorithms. Theorem 4.26 states a lower bound on the PMEPR of cosets of RM q (1, m). This lower bound was used to prove general lower bounds on the achievable PMEPR of such cosets in Corollary 4.28. This shows that among all cosets in RM2 (2, m) for odd m and among all cosets in ZRM4 (2, m) the cosets in Corollary 4.30 are optimal with respect to the PMEPR. Theorem 4.26 was also used to prove, in certain cases, the tightness of the PMEPR upper bounds in Corollaries 4.23 and 5.8. In Corollary 6.8 we have shown that the PAPR of any coset of RM q (1, m) is at most its PMEPR. As a consequence, all OFDM codes with low PMEPR proposed in Chapters 4 and 5 and in [22] and [79] also have low PAPR in MC-CDMA systems. However, in view of Theorem 6.7, it is impossible to construct constant-amplitude codes from unions of cosets of RM q (1, m) whenever q 0 (mod 4). Theorem 6.15 shows that a large subcode of the lengthened weighted-degree trace code W(h, m, D) has PAPR at most (D 1)2 . In Section 6.6 we presented families of quaternary constant-amplitude codes for MC-CDMA. We proposed constant-amplitude subcodes of quaternary ReedMuller, Kerdock, and DelsarteGoethals codes. Corollaries 6.27 6.29 provide mappings from binary to quaternary constant-amplitude codes, which can be combined with known results (e.g., from [81]) to produce additional quaternary constant-amplitude codes. These codes allow a rich trade-off between minimum distance and code rate. 150

In Algorithms 4.32 and 4.34 we proposed two maximum-likelihood decoders for RMq (1, m). It has been shown that Algorithm 4.34 (which works with slight modications also for ZRMq (1, m)) has lower complexity than any previously known maximum-likelihood decoder for RM q (1, m). In principle these algorithms in combination with the supercode decoding method can be used to decode all codes proposed in Chapters 46. Each of the Chapters 36 concludes with a list of (in total 14) open problems, which suggest possible further research directions.

151

7 Summary

152

Notation
Abbreviations BCH BPSK (QPSK) DAB (DVB) DSL HiPerLAN LHS (RHS) (MC-)CDMA ML OFDM PAPR PMEPR PSK RM SDR UMTS WLAN BoseChaudhuriHocquenghem (code) binary (quaternary) PSK digital audio (video) broadcasting digital subscriber line high performance radio local area network left (right) hand side (multicode) code-division multiple access maximum likelihood orthogonal frequency-division multiplexing peak-to-average power ratio peak-to-mean envelope power ratio phase-shift keying ReedMuller (code) signed-digit representation universal mobile telecommunication system wireless local area network

153

7 Summary
Frequently Used Symbols C R Z Zq Fq Rh,m Th,m {z} ( {z}) z |z| j q ||z|| x ( x ) f Trm (also Tr) () wt(a) d(a, b) deg A(b)( ) C(a, b)( ) S(a)() R(C) B(t, m) ERM(r, m, h) RMq (r, m) ZRMq (r, m) K(m) DG(t, m) W(h, m, D) Ff Lf 154 complex numbers real numbers integers integers modulo q nite eld with q elements Galois ring with 2hm elements and characteristic 2h Teichmuller set of Rh,m real (imaginary) part of z C complex conjugate of z C absolute value of z C or cardinality of the set z imaginary unit (also index variable) primitive qth root of unity in C Euclidean norm of z Cn x R rounded to the nearest integer towards + () polyphase version of the word f over Zq trace function on Rh,m generalized Gray map left (right) half of the generalized Gray map Lee weight of the word a Lee distance between a and b degree aperiodic autocorrelation of the word b at shift aperiodic crosscorrelation between a and b at shift multicarrier signal corresponding to the word a at time rate of a code C t-error-correcting binary BCH code of length 2m 1 2h -ary effective-degree ReedMuller code of order r a q-ary ReedMuller code of order r another q-ary ReedMuller code of order r quaternary Kerdock code quaternary DelsarteGoethals code weighted-degree trace code over Z2h Fourier transform of a function f correlation of a function f with linear functions

References
[1] E. R. Berlekamp and L. R. Welch, Weight distributions of the cosets of the (32,6) ReedMuller code, IEEE Trans. Inf. Theory, vol. 18, no. 1, pp. 203207, Jan. 1972. [2] J. A. Bingham, Multicarrier modulation for data transmission: An idea whose time has come, IEEE Commun. Mag., vol. 28, no. 5, pp. 514, May 1990. [3] I. E. Bocharova and B. D. Kudryashov, On the covering radius of convolutional codes, in Proc. First French-Israeli Workshop (Lecture Notes in Computer Science), vol. 781. Berlin, Germany: Springer-Verlag, 1993, pp. 5662. [4] P. B. Borwein and R. A. Ferguson, A complete description of Golay pairs for lengths up to 100, Mathematics of Computation, vol. 73, no. 246, pp. 967985, Jul. 2003. [5] S. Boztas, A. R. Hammons, and P. V. Kumar, 4-phase sequences with near optimum correlation properties, IEEE Trans. Inf. Theory, vol. 38, no. 3, pp. 11011113, May 1992. [6] R. Brualdi, S. Litsyn, and V. Pless, Covering radius, in Handbook of Coding Theory, V. S. Pless and W. C. Huffman, Eds. Amsterdam, The Netherlands: Elsevier, 1998. [7] S. Z. Budisin, New multilevel complementary pairs of sequences, IEE Electron. Lett., vol. 26, pp. 18611863, Oct. 1990. [8] G. Caire and E. Biglieri, Linear block codes over cyclic groups, IEEE Trans. Inf. Theory, vol. 41, no. 5, pp. 12461256, Sep. 1995. [9] A. R. Calderbank, P. C. Fishburn, and A. Rabinovich, Covering properties of convolutional codes and associated lattices, IEEE Trans. Inf. Theory, vol. 41, no. 3, pp. 732746, May 1995. [10] A. R. Calderbank and N. J. A. Sloane, Modular and p-adic cyclic codes, Designs, Codes and Cryptography, vol. 6, no. 1, pp. 2135, Jul. 1995. 155

References
[11] A. R. Calderbank, Heisenberg-weyl groups and their application to sequence design, Talk at Sequences and Codes, Vancouver, Canada, Jul. 1721 2006. [12] M. W. Cammarano and M. L. Walker, Integer maxima in power envelopes of Golay codewords, 1999. [Online]. Available: http: //www.mathcs.richmond.edu/jad/summer.html [13] C. Carlet, Z2k -linear codes, IEEE Trans. Inf. Theory, vol. 44, no. 4, pp. 15431547, Jul. 1998. [14] , On bent and highly nonlinear balanced/resilent functions and their algebraic immunities, in Proc. of 8th AAECC Symp. (Lecture Notes in Computer Science), vol. 3857. Berlin, Germany: Springer-Verlag, Feb. 2006, pp. 128. [15] C. Carlet and S. Mesnager, Improving the upper bounds on the covering radii of binary ReedMuller codes, IEEE Trans. Inf. Theory, vol. 53, no. 1, pp. 162173, Jan. 2007. [16] C.-Y. Chen, C.-H. Wang, and C.-C. Chao, Complementary sets and Reed Muller codes for peak-to-average power ratio reduction in OFDM, in Proc. of 16th AAECC Symp. (Lecture Notes in Computer Science), vol. 3857. Berlin, Germany: Springer Verlag, 2006, pp. 317327. [17] G. Cohen, I. Honkala, S. Litsyn, and A. Lobstein, Covering Codes. Amsterdam, The Netherlands: Elsevier, 1997. [18] S. D. Cohen, The length of primitive BCH codes with minimal covering radius, Designs, Codes and Cryptography, vol. 10, no. 1, pp. 516, Jan. 1997. [19] J. H. Conway and N. J. A. Sloane, Soft decoding techniques for codes and lattices, including the Golay code and the Leech lattice, IEEE Trans. Inf. Theory, vol. 32, no. 1, pp. 4150, Jan. 1986. [20] R. Craigen, W. Holzmann, and H. Kharaghani, Complex Golay sequences: Structure and applications, Discrete Math., Elsevier, vol. 252, pp. 7389, 2002. [21] M. Darnell and A. H. Kemp, Synthesis of multilevel complementary sequences, IEE Electron. Lett., vol. 24, no. 19, pp. 12511252, Sep. 1988. [22] J. A. Davis and J. Jedwab, Peak-to-mean power control in OFDM, Golay complementary sequences, and ReedMuller codes, IEEE Trans. Inf. Theory, vol. 45, no. 7, pp. 23972417, Nov. 1999. 156

References
[23] Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Physical (PHY) Layer, document RTS0023003-R2, European Telecommunications Standards Institute (ETSI), Sophia-Antipolis, Valbonne, France, Feb. 2001. [24] P. Fan and M. Darnell, Sequence Design for Communication Systems. John Wiley & Sons Inc., 1996. [25] F. Fiedler and J. Jedwab, How do more Golay sequences arise? IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 4261 4266, Sep. 2006. [26] F. Fiedler, J. Jedwab, and M. G. Parker, A framework for the construction of Golay sequences, submitted to Trans. Inf. Theory, 2006. [27] R. L. Frank, Polyphase complementary codes, IEEE Trans. Inf. Theory, vol. 26, no. 6, pp. 641647, Nov. 1980. [28] M. J. E. Golay, Multislit spectroscopy, J. Opt. Soc. Amer., vol. 39, pp. 437444, 1949. [29] , Static multislit spectrometry and its applications to the panoramic display of infrared spectra, J. Opt. Soc. Amer., vol. 41, pp. 468472, 1951. [30] , Complementary series, IRE Trans. Inf. Theory, vol. 7, no. 2, pp. 8287, Apr. 1961. [31] D. Gorenstein, W. Peterson, and N. Zierler, Two-error correcting BoseChaudhuri codes are quasi-perfect, Information and Control, vol. 3, pp. 291294, 1960. [32] A. J. Grant and R. D. van Nee, Efcient maximum-likelihood decoding of peak power limiting codes for OFDM, Proc. of Vehicular Technology Conference (VTC), Ottawa, Canada, pp. 20812084, May 1998. [33] , Efcient maximum-likelihood decoding of Q-ary modulated Reed Muller codes, IEEE Commun. Lett., vol. 2, no. 5, pp. 134136, May 1998. [34] M. Greferath and U. Vellbinger, Efcient decoding of Zpk -linear codes, IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 12881291, May 1998. [35] U. G ntzer and M. Paul, Jump interpolation search trees and symmetric u binary numbers, Inf. Process. Lett., vol. 26, no. 4, pp. 193204, Dec. 1987. [36] A. R. Hammons, Jr., P. V. Kumar, A. R. Calderbank, N. J. A. Sloane, and P. Sol , The Z4 -linearity of Kerdock, Preparata, Goethals and related e codes, IEEE Trans. Inf. Theory, vol. 40, no. 2, pp. 301319, Mar. 1994. 157

References
[37] T. Helleseth, All binary 3-error-correcting BCH codes of length 2 m 1 have covering radius 5, IEEE Trans. Inf. Theory, vol. 24, no. 2, pp. 257 258, Mar. 1978. [38] T. Helleseth, T. Klve, and J. Mykkeltveit, On the covering radius of binary codes, IEEE Trans. Inf. Theory, vol. 24, no. 5, pp. 627628, Sep. 1978. [39] T. Helleseth and P. V. Kumar, Sequences with low correlation, in Handbook of Coding Theory, V. S. Pless and W. C. Huffman, Eds. Amsterdam, The Netherlands: Elsevier, 1998. [40] T. Helleseth, P. V. Kumar, O. Moreno, and A. G. Shanbhag, Improved estimates via exponential sums for the minimum distance of Z4 -linear trace codes, IEEE Trans. Inf. Theory, vol. 42, no. 4, pp. 12121216, Jul. 1996. [41] E. E. Hollis, Constructing broad sense complementary sequences of length 4n , IEEE Trans. Aerosp. Electron. Syst., vol. AES-3, pp. 974976, Nov. 1967. [42] W. H. Holzmann and H. Kharaghani, A computer search for complex Golay sequences, Australasian J. Comb., vol. 10, pp. 251258, 1994. [43] X.-D. Hou, Covering radius of the ReedMuller code R(1, 7) a simpler proof, J. Comb. Theory (A), vol. 74, pp. 337341, 1996. [44] M. G. M. Hussain, Principles of high-resolution radar based on nonsinusoidal waves Part I: Signal representation and pulse compression, IEEE Trans. Electromagnetic Compatibility, vol. 31, no. 4, pp. 359368, Nov. 1989. [45] C.-L. I and R. D. Gitlin, Multi-code CDMA wireless personal communications networks, Proc. of Int. Conf. Commun. (ICC), Seattle, WA, USA, vol. 2, pp. 10601064, Jun. 1995. [46] H. Janwa, Some new upper bounds on the covering radius of binary linear codes, IEEE Trans. Inf. Theory, vol. 35, no. 1, pp. 110122, Jan. 1989. [47] J. Jedwab and C. J. Mitchell, Minimum weight modied signed-digit representations and fast exponentation, IEE Electron. Lett., vol. 25, no. 17, pp. 11711172, Aug. 1989. [48] A. E. Jones and T. A. Wilkinson, Combined coding for error control and increased robustness to system nonlinearities in OFDM, Proc. of IEEE 46th Vehicular Technology Conf. (VTC), Atlanta, GA, pp. 904908, Apr. 1996. 158

References
[49] , Performance of ReedMuller codes and a maximum-likelihood decoding algorithm for OFDM, IEEE Trans. Commun., vol. 47, no. 7, pp. 949952, Jul. 1999. [50] A. E. Jones, T. A. Wilkinson, and S. K. Barton, Block coding scheme for reduction of peak to mean envelope power ratio of multicarrier transmission schemes, IEE Electron. Lett., vol. 30, no. 25, pp. 20982099, Dec. 1994. [51] P. V. Kumar, R. A. Scholtz, and L. R. Welch, Generalized bent functions and their properties, J. Comb. Theory (A), vol. 40, pp. 90107, 1985. [52] P. V. Kumar, T. Helleseth, and A. R. Calderbank, An upper bound for Weil exponential sums over Galois rings and applications, IEEE Trans. Inf. Theory, vol. 41, no. 2, pp. 456468, Mar. 1995. [53] P. V. Kumar, T. Helleseth, A. R. Calderbank, and A. R. Hammons, Jr., Large families of quaternary sequences with low correlation, IEEE Trans. Inf. Theory, vol. 42, no. 2, pp. 579592, Mar. 1996. [54] H. Lee and S. Golomb, A new construction of 64-QAM Golay complementary sequences, IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 16631670, Apr. 2006. [55] X. Li and L. Cimini Jr, Effects of clipping and ltering on the performance of OFDM, IEEE Commun. Lett., vol. 2, no. 5, pp. 131133, May 1998. [56] Y. Li and W. B. Chu, More Golay sequences, IEEE Trans. Inf. Theory, vol. 51, no. 3, pp. 11411145, Mar. 2005. [57] R. Lidl and H. Niederreiter, Finite elds, 2nd ed., ser. Encyclopedia of Mathematics and its Applications. Cambridge University Press, 1997, vol. 20. [58] R. Lucas, M. Bossert, and A. Dammann, Improved soft-decision decoding of ReedMuller codes as general multiple concatenated codes, Proc. of ITG Conf. Source and Channel Coding, Aachen, Germany, 1998. [59] B. R. MacDonald, Finite Rings with Identity. 1974. New York: Marcel Dekker,

[60] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Correcting Codes. Amsterdam, The Netherlands: North Holland, 1977. [61] K. Manji and B. S. Rajan, On the PAPR of binary ReedMuller OFDM codes, Proc. of IEEE Int. Symp. Inf. Theory, Chicago, IL, USA, p. 423, Jul. 2004. 159

References
[62] S. Matsufuji and K. Imamura, Real-valued bent function and its application to the design of balanced quadriphase sequences with optimal correlation properties, in Proc. of 8th AAECC Symp. (Lecture Notes in Computer Science), vol. 508. Berlin, Germany: Springer Verlag, Aug. 1990, pp. 106112. [63] , Balanced quadriphase sequences with optimal periodic correlation properties constructed by real-valued bent functions, IEEE Trans. Inf. Theory, vol. 39, no. 1, pp. 305310, Jan. 1993. [64] R. J. McEliece, The algebraic theory of convolutional codes, in Handbook of Coding Theory, V. S. Pless and W. C. Huffman, Eds. Amsterdam, The Netherlands: Elsevier, 1998. [65] A. M. McLoughlin, The complexity of computing the covering radius of a code, IEEE Trans. Inf. Theory, vol. 30, no. 6, pp. 800804, Nov. 1984. [66] S. H. M ller, R. W. B uml, R. F. Fischer, and J. B. Huber, OFDM with u a reduced peak-to-average power ratio by multiple signal representation, Annals of Telecommunications, vol. 52, no. 1-2, pp. 5867, Feb. 1997. [67] J. Mykkeltveit, The covering radius of the (128,8) ReedMuller code is 56, IEEE Trans. Inf. Theory, vol. 26, no. 3, pp. 359362, May 1980. [68] A. A. Nechaev, Kerdock code in a cyclic form, Discrete Math. Appl., vol. 1, no. 4, pp. 365384, 1991, originally published in Russian in Diskretnaya Matematika, vol. 1, no. 4, pp. 123139, 1989. [69] D. J. Newman, An L1 extremal problem for polynomials, Proc. Amer. Math. Soc., vol. 16, pp. 12871290, Dec. 1965. [70] K. M. Nieswand and K. N. Wagner, Octary codewords with power envelopes of 3 2m , 1998. [Online]. Available: http://www.mathcs. richmond.edu/jad/summer.html [71] H. Ochiai, M. P. C. Fossorier, and H. Imai, On decoding of block codes with peak-power reduction in OFDM systems, IEEE Commun. Lett., vol. 4, no. 7, pp. 226228, Jul. 2000. [72] H. Ochiai and H. Imai, Block coding scheme based on complementary sequences for multicarrier signals, IEICE Transactions on Fundamentals, vol. E80-A, pp. 21362146, Nov. 1997. [73] M. K. Oolun, Electrical systems identication using Golay complementary series, IEE Proceedings on Science, Measurement and Technology, vol. 144, no. 6, pp. 267272, Nov. 1997. 160

References
[74] T. Ottoson, Precoding for minimization of envelope variations in multicode DS-CDMA systems, Wireless Personal Commun., vol. 13, pp. 5778, May 2000. [75] M. G. Parker, K. G. Paterson, and C. Tellambura, Golay complementary sequences, in Encyclopedia of Telecommunications, J. G. Proakis, Ed. New York: Wiley, 2002. [76] M. G. Parker and C. Tellambura, Generalised RudinShapiro constructions, Proc. of Workshop on Coding and Cryptography (WCC), Paris, France, 2001. [77] , GolayDavisJedwab complementary sequences and Rudin Shapiro constructions, 2001. [Online]. Available: http://www.ii.uib.no/ matthew/ConstaBent2.pdf [78] , A construction for binary sequence sets with low peak-toaverage power ratio, Report No. 242, Department of Informatics, University of Bergen, Norway, Feb. 2003. [Online]. Available: http: //www.ii.uib.no/matthew/ConstructReport.pdf [79] K. G. Paterson, Generalized ReedMuller codes and power control in OFDM modulation, IEEE Trans. Inf. Theory, vol. 46, no. 1, pp. 104120, Jan. 2000. [80] , Sequences for OFDM and multi-code CDMA: Two problems in algebraic coding theory, in Proc. of SETA 2001 (Discrete Mathematics and Theoretical Computer Science Series). Berlin, Germany: Springer Verlag, 2002, pp. 4671. [81] , On codes with low peak-to-average power ratio for multicode CDMA, IEEE Trans. Inf. Theory, vol. 50, no. 3, pp. 550559, Mar. 2004. [82] K. G. Paterson and A. E. Jones, Efcient decoding algorithms for generalized ReedMuller codes, IEEE Trans. Commun., vol. 48, no. 8, pp. 12721285, Aug. 2000. [83] K. G. Paterson and V. Tarokh, On the existence and construction of good codes with low peak-to-average power ratios, IEEE Trans. Inf. Theory, vol. 46, no. 6, pp. 19741987, Sep. 2000. [84] B. M. Popovi , Synthesis of power efcient multitone signals with at c amplitude spectrum, IEEE Trans. Commun., vol. 39, no. 7, pp. 10311033, Jul. 1991. 161

References
[85] C. Riera and M. G. Parker, Generalized bent criteria for Boolean functions (I), IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 4142 4159, Sep. 2006. [86] C. R ing and V. Tarokh, A construction of OFDM 16-QAM sequences o having low peak powers, IEEE Trans. Inf. Theory, vol. 47, no. 5, pp. 2091 2093, Jul. 2001. [87] O. S. Rothaus, On bent functions, J. Comb. Theory (A), vol. 20, pp. 300305, 1976. [88] W. Rudin, Some theorems on Fourier coefcients, Proc. Amer. Math. Soc., vol. 10, pp. 855859, 1959. [89] K.-U. Schmidt, Comment on Efcient maximum-likelihood decoding of Q-ary modulated ReedMuller codes, IEEE Commun. Lett., vol. 10, no. 1, p. 52, Jan. 2006. [90] , On cosets of the generalized rst-order ReedMuller code with low PMEPR, IEEE Trans. Inf. Theory, vol. 52, no. 7, pp. 32203232, Jul. 2006. [91] , Quaternary constant-amplitude codes for multicode CDMA, submitted to IEEE Trans. Inf. Theory, 2006. [92] , Complementary sets, generalized ReedMuller codes, and power control for OFDM, IEEE Trans. Inf. Theory, vol. 53, no. 2, pp. 808814, Feb. 2007. [93] , On the peak-to-mean envelope power ratio of phase-shifted binary codes, accepted for publication in IEEE Trans. Commun., 2007. [94] K.-U. Schmidt and A. Finger, Simple maximum-likelihood decoding of generalized rst-order ReedMuller codes, IEEE Commun. Lett., vol. 9, no. 10, pp. 912914, Oct. 2005. [95] , Constructions of complementary sequences for power-controlled OFDM transmission, in Proc. of WCC 2005 (Lecture Notes in Computer Science), vol. 3969. Berlin, Germany: Springer Verlag, Jun. 2006, pp. 330345. [96] A. G. Shanbhag, P. V. Kumar, and T. Helleseth, Improved binary codes and sequence families from Z4 -linear codes, IEEE Trans. Inf. Theory, vol. 42, no. 5, pp. 15821587, Sep. 1996. [97] H. S. Shapiro, Extremal problems for polynomials and power series, Master Thesis, MIT, 1951. 162

References
[98] M. Sharif and B. Hassibi, Existence of codes with constant PMEPR and related design, IEEE Trans. Signal Proces., vol. 52, no. 10, pp. 28362846, Oct. 2004. [99] , On multicarrier signals where the PMEPR of a random codeword is asymptotically log n, IEEE Trans. Inf. Theory, vol. 50, no. 5, pp. 895903, May 2004. [100] R. Sivaswamy, Multiphase complementary codes, IEEE Trans. Inf. Theory, vol. 24, no. 5, pp. 546552, Sep. 1978. [101] N. J. A. Sloane, Unsolved problems related to the covering radius of codes, in Open Problems in Communication and Computation, T. M. Cover and B. Gopinath, Eds. Springer-Verlag, NY, 1987, pp. 5156. [102] P. Sol and D. Zinoviev, Weighted degree trace codes for PAPR reduction, e in Proc. of SETA 2004 (Lecture Notes in Computer Science), vol. 3486. Berlin, Germany: Springer Verlag, 2005, pp. 406413. [103] , Low-correlation, high-nonlinearity sequences for multiple-code CDMA, IEEE Trans. Inf. Theory, vol. 52, no. 11, pp. 51585163, Nov. 2006. [104] P. Spasojevic and C. N. Georghiades, Complementary sequences for ISI channel estimation, IEEE Trans. Inf. Theory, vol. 47, no. 3, pp. 1145 1152, Mar. 2001. [105] T. E. Stinchcombe, Aperiodic autocorrelations of length 2 m sequences, complementarity, and power control for OFDM, Ph.D. dissertation, University of London, Apr. 2000. [Online]. Available: http://www.isg.rhul. ac.uk/alumni/thesis/stinchcombe t.pdf [106] V. Tarokh and H. Jafarkhani, On the computation and reduction of the peak-to-average power ratio in multicarrier communications, IEEE Trans. Commun., vol. 48, no. 1, pp. 3744, Jan. 2000. [107] C. Tellambura, Upper bound on the peak factor of n-multiple carriers, IEE Electron. Lett., vol. 33, no. 19, pp. 16081609, Sep. 1997. [108] A. A. Tiet v inen, An upper bound on the covering radius as a function of a a the dual distance, IEEE Trans. Inf. Theory, vol. 36, no. 6, pp. 14721474, Nov. 1990. [109] C.-C. Tseng, Signal multiplexing in surface-wave delay line using orthogonal pairs of Golays complementary sequences, IEEE Trans. Sonics and Ultrasonics, vol. SU-18, pp. 103107, 1971. 163

References
[110] C.-C. Tseng and C. L. Liu, Complementary sets of sequences, IEEE Trans. Inf. Theory, vol. 18, no. 5, pp. 644652, Sep. 1972. [111] R. D. J. van Nee, OFDM codes for peak-to-average power reduction and error correction, Proc. of IEEE Global Telecommunications Conference (GLOBECOM), London, U.K., pp. 740744, Nov. 1996. [112] S. G. Vladuts and A. N. Skorobogatov, Covering radius for long BCH codes, Probl. Pered. Inf., vol. 25, pp. 3845, 1989, translated in Probl. Inf. Transm., vol. 25, no. 1, pp. 28-34, 1989. [113] T. Wada, Characteristic bit sequences applicable to constant amplitude orthogonal multicode systems, IEICE Trans. Fundamentals, vol. E83-A, no. 11, pp. 21602164, Nov. 2000. [114] T. Wada, T. Yamazato, M. Katayama, and A. Ogawa, A constant amplitude coding for orthogonal multi-code CDMA systems, IEICE Trans. Fundamentals, vol. E80-A, no. 12, pp. 24772483, Dec. 1997. [115] , Error correcting capability of constant amplitude coding for orthogonal multi-code CDMA systems, IEICE Trans. Fundamentals, vol. E81-A, no. 10, pp. 21662169, Oct. 1998. [116] Z.-X. Wan, Lectures on Finite Fields and Galois Rings. Singapore: World Scientic Publishing Co. Pte. Ltd., 2003. [117] T. A. Wilkinson and A. E. Jones, Minimisation of the peak to mean envelope power ratio of multicarrier transmission schemes by block coding, Proc. of IEEE Vehicular Technology Conference (VTC), Chicago, IL, pp. 825829, Jul. 1995. [118] J. Wolfmann, Difference sets in Zm and F2m , Designs, Codes and Cryp4 2 tography, vol. 20, no. 1, pp. 7388, Apr. 2000. [119] G. Wunder and H. Boche, A baseband model for computing the PAPR in OFDM systems, Proc. of 4th Int. Conf. Source and Channel Coding, Berlin, Germany, pp. 273280, Jan. 2002.

164