1006

IEEE TRANSACTIONS

ON INFORMATION

THEORY, VOL. 37, NO. 4, JULY 1991

Optimally Near-Far Resistant Multiuser Detection in Differentially Coherent Synchronous Channels

Mahesh K. Varanasi, Member, IEEE, and Behnaam Aazhang, Member, IEEE

Abstract -The noncoherent demodulation of differentially phase-shift keyed signals transmitted simultaneously via a synchronous CDMA channel is studied under the assumption of white Gaussian background noise. A class of noncoherent linear detectors is defined with the objective of obtaining the optimal one. The performance criterion considered is near-far resistance, which denotes worst-case multiuser asymptotic efficiency over near-far environments. It is shown that the optima1 linear detector is a noncoherent decorrelating detector. This detector is analogous to the coherent decorrelating detector that was obtained from similar considerations for the coherent CDMA channel by Lupas and Verdu. The commonality between the properties of the decorrelating detectors for coherent and noncoherent channels is established. In particular, it is shown that no other DPSK multiuser detector achieves a higher near-far resistance than does the noncoherent decorrelator, i.e., the optimally near-far resistant linear detector is optimally near-far resistant.
Index Terms-Code-division multiaccess, differential modulation, noncoherent detection, multiuser detection, tiuser channels, multiaccess communications. PSK mul-

I. INTRODUCTION N a code-division multiple-access (CDMA) system, several information-bearing signals are simultaneously transmitted over a common channel. Each signal is a result of digitally modulating a sequence of information symbols using distinct preassigned code or signature waveforms. The incoming signal at the receiver is therefore a superposition of such signals. Upon observation of this composite signal and equipped with a knowledge of the signature waveforms, the receiver is required to demodulate the information from each transmission. The

Manuscript received September 25, 1989; revised January 28, 1991. This work was supported in part by the Advanced Technology Program of the Texas Higher Education Coordinating Board under Grant 003604-018 and in part by the National Science Foundation under Grant NCR-8710844. This work was presented in part at the Conference on Information Sciences and Systems, Johns Hopkins University, Baltimore, MD, 1989, and at the IEEE International Symposium on Information Theory, San Diego, CA, Jan. 14-19, 1990. M. K. Varanasi was with the Department of Electrical and Computer Engineering, Rice University, Houston, TX. He is now with the Department of Electrical and Computer Engineering, University of Colorado, Boulder, CO 80309. B. Aazhang is with the Department of Electrical and Computer Engineering, Rice University, P.O. Box 1892, Houston, TX 77251. IEEE Log Number 9144608.

study of such demodulation strategies is referred to as multiuser detection [21]. There has been considerable research on the coherent multiuser detection problem in recent years. The underlying assumption in this work is that the receiver is able to estimate and track the energy and phase of each component signal. Optimum coherent multiuser detection strategies were obtained and their performance thoroughly studied in [19], [20], where it was established that a dramatic improvement over conventional single-user detection can be achieved. However, these optimum strategies are computationally intensive. Motivated by the need for high-performance, low-complexity detection schemes, a robust decorrelating strategy [81, [91 and a suboptimum multistage detection technique based on successivemultiple-access rejection [15], [16] were obtained for synchronous and asynchronous channels, respectively. The reader is also referred to [21] for a lucid summary of recent activity on this problem, and the references therein, and also to [17] for additional results on multistage detection. Factors that dissuade or preclude the use of coherent single-user communication such as fading, oscillator phase instability at the transmitter, dynamically evolving positions of transmitter and receiver in a mobile environment etc., are stronger deterrents in multiuser environments because the estimation and tracking of the energy and phase of each component signal has to be carried out not only in the presence of additive noise, but also in the presence of the interfering signals. It is for this reason that even in less turbulent channels, which may be viable for coherent single-user communication, the decision to employ complex estimation and tracking strategies could prove to be counterproductive in multiuser channels. The primary objective in this paper is therefore the study of a CDMA systemwhere the modulation technique employed for each transmission allows noncoherent demodulation of the information, i.e., no knowledge of the energies and phases of any of the component signals is assumed at the receiver. In particular, attention is focused on CDMA channels where the modulation technique employed for each transmission is differential phase-shift keying (DPSK).
01991 IEEE

0018.9448/91/0700-1006$01.00

VARANASI AND AAZHANG: OPTIMALLY NEAR-FAR RESISTANT MULTIUSER DETECTION

1007

The conventional approach to differentially coherent and E,, the actual energy of the k th user. Asymptotic DPSK detection in CDMA channels involves demodulat- efficiency therefore characterizes performance loss when ing each transmitted signal as if it were the only one the primary degradation is due to the presence of interpresent. The conventional noncoherent DPSK detector fering users. therefore consists of a bank of K decoupled single-user A primary motivation for studying noncoherent detecDPSK detectors, one for each transmission. Although tion is to address the need for reliable communication in each single-user demodulator is optimal in error probabil- channels, such as in near-far environments in mobile ity in the corresponding single-user channel, this is no communication, where signal energies and phases vary longer true in multiuser channels. Decision statistics are too rapidly for the receiver to estimate and track them. In corrupted by multiple-access interference in addition to order to find such detection strategies that are robust to additive noise. Performance evaluation based on approxi- near-far effects, we invoke the notion of near-fur resismate bit-error rates of the conventional noncoherent de- tance, a performance measure introduced by Lupas and tector for a variety of spread-spectrum multiple-access Verdu in [8]. For noncoherent multiuser detection, we (SSMA) systems has been undertaken in several works introduce a modification of this measure to reflect the that include direct-sequence[3], frequency-hopped 121,151, lack of both energy as well as phase information at the and hybrid direct-sequence-frequency-hopped141signal- receiver. In particular, the near-far resistanceof a DPSK ing schemes. Acceptable performance for the conven- multiuser detector, the mth user asymptotic efficiency of tional detector can be expected only in systemswhere the which is vrn, is defined as signal energies are not very dissimilar and the crosscorreinf vrn, r, = (1.2) lations between the signature signals are kept low by E, P 0 employing spread-spectrum pseudonoise sequences or hopping patterns of long periodicity, i.e., in low bandwidth efficiency situations. When the received signal en- where Ei and e1 denote the energy and phase of the ith ergies are dissimilar (near-far environments), this de- transmission. A nonzero value of near-far resistancewill tector is unable to demodulate the weak signals reliably therefore guarantee an exponential decay of error probaeven in the low bandwidth efficiency mode. In order to bility with increasing signal-to-noise ratio, irrespective of remedy the near-far problem, currently operational the specific values of energies and phases of the interferspread-spectrum systemsrely on power control strategies ing signals. in which transmitters are required to adaptively adjust The rest of the paper is organized as follows. In Section power levels so that all signals arrive at the receiver with II, we model a general CDMA-DPSK systemthat encomessentially similar energies. However, such a strategy is passes as a special case the hybrid direct-sequenceself-defeating [21] because it dictates a significant reduc- frequency-hopped spread-spectrummodel (and hence dition in most transmitter powers to accommodate the rect-sequenceand frequency-hopped SSMA). This allows weakest transmitter, thereby diminishing the multiple- a unified treatment of multiuser detection problems which traditionally have been dealt with separately. In this paaccesscapability of the overall system. The performance measure of primary interest in com- per, we focus attention on the synchronous additive munication systems is probability of error. In multiuser Gaussian CDMA channel because such a study promotes systems, probability of error in the high signal-to-noise a fundamental understanding of key issues involved in a ratio region emphasizes degradation due to interfering simple setting. The goal in Section III is to motivate the need for users rather than that due to the additive noise. Multiuser asymptotic efficiency is a performance measure that not detection strategies which are robust to near-far effects. only captures this information but is analytically tractable This is done by characterizing the performance limitation as well. It was first introduced in the context of coherent of the conventional noncoherent single-user detection multiuser communication by Verdu in [19]. For differen- strategy in a multiuser environment. In particular, it is tially coherent multiuser detection, asymptotic efficiency shown that the conventional noncoherent detector is not for the mth user whose bit-error rate is P,(v) can be near-far resistant. The objective in Section IV is to alleviate the perforformally defined as mance limitation of the conventional detection scheme. rl,=sup OIr<l; To this end, a class of noncoherent linear detectors is ( introduced, and exact expressions for the bit-error rate J@ooP,,,(g)/exp[-rEm/2a2] < +a), (1.1) and asymptotic efficiency of an arbitrary member of this class are obtained. In Section IV-A, we characterize the where E, is the mth signal energy and cr2 denotes power linear detector that is optimally near-far resistant, i.e., spectral density level of the background white noise. the linear detectors that achieve the highest worst-case Since the error probability in single-user channels em- asymptotic efficiency over near-far environments. Under ploying binary DPSK is O[exp( - E, /2a2>], vm is equal mild restrictions on the signature signal constellation, it is to the limit, as (T-+ 0, of the ratio between the energy shown that noncoherent decorrelating detectors solve this minimax optimization problem. The noncoherent decorrerequired by a single user to achieve an error rate P,(a)

1008

IEEETRANSACTIONSONINFORMATIONTHEORY,VOL.37,NO.4,JULY1991

In this section we consider binary DPSK modulation and obtain the exact bit-error probability of the conventional noncoherent single-user detector in multiuser channels. The performance limitation of this detector is II. SYSTEM MODEL characterized in terms of near-far resistance. The conNoncoherent detection of multiple differentially ventional detector consists of K decoupled noncoherent phase-shift keyed transmissions of digital information made over a code-division multiple-access(CDMA) chan- single-user detectors where the mth single-user detector nel is considered. It is assumedthat a superposition of K is optimum in the absenceof all but the mth signal. In a transmissionsarrive at the receiver in symbol synchronism single-user channel, it is well-known that the decision perturbed by additive white Gaussian noise with noise variable evaluates the phase difference between the respectral density u2. In the time interval [ - T, T], the ceived signal in the current and previous symbol intervals and selects the symbol whose phase is closest to this complex envelope of the received signal is difference. In the binary DPSK multiuser environment, therefore, the conventional noncoherent detector decision on b, for each m are made as follows (cf. [12]): k=l t~[-T,T]. (2.3) where z,Ji) = ;/(i+lTl(l)f,,(t
IT

lating detector is an analog of the coherent decorrelating detector obtained from similar considerations in [8]. The commonality between the properties of the coherent and the noncoherent decorrelating detectors is established. Notable among these properties are that the bit-error rate of a noncoherent decorrelating detector (and hence its asymptotic efficiency) for each user is independent of the interfering signalsenergies and phases,thereby alleviating the near-far problem. Further, in Section IV-B, it is shown that no other DPSK multiuser detector, linear or nonlinear, has a higher near-far resistance than does the decorrelator, i.e., the optimally near-far resistant linear decorrelator is optimally near-far resistant. The paper is concluded in Section V.

As essential parameter of the system is the K X K matrix of normalized complex cross-correlations R of the complex-valued signature signals whose elements are defined by

so that the diagonal elements are all equal to unity. III. CONVENTIONALNONCOHERENT
SINGLE-USER DETECTOR

6, = en [ Re{zm(-l)zm(~)}],
- iT)dt.

(3.4)

The finite-energy modulating signal {U,(t); t E [O,T]) can be expressed as (E,)*f,(t)exp(jB,), where {fk(t); t E [O,T]} is in general a complex-valued, unit-energy signature waveform assignedfor the k th transmission.In DPSK modulation, information is encoded into phase differences between successive symbol intervals. In binary DPSK, a 1 or a 0 is transmitted by shifting the phase of the carrier by r or zero radians relative to the carrier in the previous signaling interval. Therefore, b, = + 1 (or - 1) in the previous representation denotes transmission of a 0 (or 1) in the kth packet in the time interval [O,T]. Note that implicit in the signal model, is the assumption that the energies and phases of the signals remain constant over two successivesymbol intervals. The problem can now be simply stated as the demodulation of the symbols b,, b,; * a,b,, given the received signal in the interval under consideration. In general, the kth signature signal fk(t) is assumedto be of the form a,(t>exp(jo,(t)t). This general model encompasses as special cases, direct-sequence spreadSpWtrUm Sign& (with ok(t) = 0 and u,(t) as the kth users spreading waveform), frequency-hopped spreadspectrum signals (with a,(t) = 1 and w,(t) representing the k th users hopping pattern in the zeroth time interval), and hybrid direct-sequence-frequency-hoppedsignals being represented by a combination of nontrivial a,(t) and w,(t). Results obtained from such signal models promote a unified understanding of each of these SSMA signaling techniques.

The bank of decoupled single-user detectors based on matched filter realizations is shown in Fig. 1 where, for simplicity, we have chosen to represent all operations over complex low-pass representations of real bandpass signals. Consider the following lemma, which is a standard result in communication theory and can be found in several references (e.g., see [13]). Lemma 1: Let X and Y be uncorrelated, complex-valued Gaussian random variables with means pLx and pLy, respectively, and a common variance u2. Then, the probability that the decision variable D = Re(XY) is less than zero is given as

where Q(a, b) denotes Marcums Q-function2 and the parameters a and b are defined as up @qlIpx and - pyl

The symbol denotes complex conjugation. The Marcum Q-function [lo] has the integral representation lFxexp(-(x2 + b2)/2)10(ax)dx, where lo(x) is the zeroth order modified Bessel function of the first kind.

VARANASI AND AAZHANG: OPTIMALLY NEAR-FAR RESISTANT MULTIUSER DETECTION

1009

Matched

Filter

1 $1

Matched

Filter

2 6

Fig. 1.

Conventional detection: A bank of K single-user DPSK detectors.

Next, we will use this lemma to obtain the error probability of the conventional single-user DPSK detector in multiuser Gaussian channels, the decision of which is expressed in (3.4). Since the mth user error probability conditioned on the mth information bit b, is equal for b, = 1 or - 1, we have3 Pz)( a) = Pr [ Re (Z,J - l)Z,(O)) = 21eK c
bsB,

random variables with means pL,( -1) = E E;/2exp(j0j)Rj,

j=l

and p,(O) = 5 Ej/2 exp ( jej)b,Rjm,

j=l

(3.7)

< Olb,,, + I] = <Olb], (3.5)

and variances u:(i) = E[ Iym( i)y,( i)] = 2a2, for i= -l,O.

Pr [Re(Z,( -l)Z,(O))

Further, these statistics are uncorrelated becausethey are where B, denotes the set {b E { - 1, l}K; b, = 1) and the obtained from the received signal over nonoverlapping second equality follows by averaging over equiprobable time intervals. Therefore, we can apply Lemma 1 to the interfering bit combinations. From (2.3) and the defini- probability inside the summation in (3.5) to obtain tion of Z,(i), the statistics Z,(- 1) and Z,(O) can be written as Q(a,,b,) - ;I,,(a,b,) P,$(,) =21-K c
beBm

Z,( - 1) = ; E;* eXP(.Sj)Rjm + Ym( -I>,

j=l

.exp( -q)), where a, = (2a) -lpc( - 1) - PC(0)I and

(3.8)

Z,(O) = t E;12 exp (@j)bjRj, + Y,(O) 7 (3.6)

j=l

where y,(i) are complex-valuedzero-mean Gaussianrandom variables defined as

b, = (2~) -I/-4 - 1) + ~$91, which in turn can be evaluated from the expressionsfor the means in (3.7). In the following proposition, we show that the conventional single-user detector in a multiuser environment is not near-far resistant. t)f,( t - iT) dt. Proposition I: The near-far resistance of the conventional noncoherent detector for the mth user, denoted as Conditioning on the interfering bit-combinations and the ?j$, is identically equal to zero unless the mth users energies and phases of the different transmissions, the signature signal is orthogonal to the subspacespanned by statistics Z,( - 1) and Z,(O) are complex-valuedGaussian the other K - 1 signature signals. Proof: From the definition of asymptotic efficiency in (1.1) and the expression for the bit-error probability for

3The superscript (c) in Pz) identifies the analysis of conventional noncoherent system.

1010

IEEE TRANSACTIONS

ON INFORMATION

THEORY, VOL. 37, NO. 4, JULY 1991

the conventional noncoherent detector in (3.8), we have

21pKb FB Pr [Re (.?,A - VXO))

qE)=sup i
bsB,

OIr11;

CT-0

lim

< ($1

exp(-rE,/2a2) Pr [ Re (-%-A Wm(0)) exp( - rE,/2a2) < ($1

<+m

-\

The second equality follows from the fact that, in a low totic efficiency, we have background noise region, the summation of error probabilities in P$(a> is dominated by the term corresponding to the transmission of least-favorable bits of the interfering users. The last equality defines conditional asymptotic l/2 efficiency that we denote as 4,(b), as the asymptotic efficiency conditioned on the vector of transmitted bits. Therefore, the near-far resistance of the conventional =min{l,laTr,12}, detector can be bounded above as where the second equality is obtained by evaluating bz yp) = inf min +2(b) = min &f. d?( b > m and the last equality by denoting the mth column of R as E, t 0 bEB, bs4n 1 eit[-a,al @,E[--,Tl rrn and defining the K-dimensional normalized amplitude i=#m i#m vector
exp(j(% %))Ri,

,,fo

4??(b)

>

VbEB,.

o,E[1-T,?71 i#m

a= [ (E,/E,)2exp(j(8,-t),),...,(E,/E,)12
exp(j(eK-em)].

In order to show that the near-far resistance of the conventional detector is zero, it is sufficient to show that the expression on the right-hand side of this inequality-henceforth referred to as conditional near-far resistance-is equal to zero for some appropriately chosen value of b E B, because near-far resistance is nonnegative by definition. Let us consider the probability of error given that the transmitted bits of the interfering users are all equal to + 1, i.e., when b = u = [l, 1; . *, llT. Substituting b = u in (3.8), we have Pr[Re(Z,(-l)Z,(O))<Olb=u]

Hence, corresponding to the set of admissible values of interfering energies and phases is the set of all admissible values of a denoted by A, and equal to A,,, = {a E CK; am = 1). Now, the conditional near-far resistance, conditioned on b = u can be written as

=Q(L?~,~,)

where a,, 6, are equal to a,, b, evaluated at b = u, respectively. It is easily seen that 6, is identically equal to zero. Therefore, the error probability conditioned on b = u is given as Pr (0:) < Olb = u) = Q(O,&,) - :1,(O) exp (- @/2)

where the last equality is obtained by observing that the objective function in the minimization problem is nonnegative and choosing a vector a, normalized so that a, = 1, that lies in the null spaceof r,-the existence of which is guaranteed, provided that r,,, is not equal to the mth unit vector, i.e., when the mth users signature signal is not orthogonal to the other signature signals-thereby yielding a minimum that is equal to zero. Since 7:) is nonnegative and is no greater than inf,, A, +$(u>, the result follows. 0

The near-far resistance of the conventional noncoherent detector is bounded away from zero only when the corresponding signature signal is orthogonal to the subspace spanned by the other signature signals. In practice, (3.9) however, it is more by chance than by design that the signature signals are orthogonal, because of bandwidth where equality follows from the property of the Q-func- restrictions, lack of synchronism and other design contion that Q<O, = exp(- x2/2> and also from the fact straints. A similar result for single-user detection in the X) that I,(O) = 1. From the conditional error probability ex- coherent multiuser channel was obtained by Lupas and pression in (3.9) and the definition of conditional asymp- Verdu in [8].

VARANASIANDAAZHANG:OPTIMALLYNEAR-FARRESISTANTMULTIUSERDETECTlON

1011

Matched fl(T-t)

Filter

1 t=iT

<z(i),

1) h> Rec.)

r(t)

Matched Filter f,jT-t)

2 t=iT Rec.)

Fig. 2.

Linear multiuser DPSK detector.

In the next section, we would like to remedy the inability of the conventional noncoherent detector to cope with the uncertainties associatedwith the transmissions of the interfering users. IV. NONCOHERENTLINEAR MULTIUSER DETECTORS We consider a class of noncoherent linear multiuser detectors. The computational complexity of these detectors is independent of the number of users. We obtain the exact bit-error probability for an arbitrary member of this class. The objective is to characterize the linear detector which maximizes near-far resistance, or equivalently, maximizes the worst-case asymptotic efficiency over the energies and phases of the interfering signals. Such a detector would then be near-far resistant and at the same time retain the computational simplicity of the conventional detector. A noncoherent linear multiuser detector for the mth user, denoted by a nonzero transformation h E CK, is defined by the decision 6, = sgn Re 5 h,)z,( - 1) 5 h$)~I(0) . (4.10) I=1 I i k=l The class of linear detectors will be denoted as 0 and membership to this class will be denoted by h() E 0. Note that the conventional noncoherent detector for the mth user is a degenerate member of the class fi with a representation h = urn, the mth unit vector. Furthermore, this class is the noncoherent counterpart of the class of coherent linear detectors introduced by Lupas and Verdu in [8]. A centralized matched-filter based structure of a linear detector is shown in Fig. 2. For simplicity, we show operations over equivalent low-pass representations of actual signals. Note that all matched-filter outputs now enter into the decision variable in the demodulation of each user via the corresponding transformation. The vector of complex inner product evaluations represent a

computational complexity per demodulated symbol of O(K) in complex operations. An alternative structure of the linear detector, suggested in the coherent context in [21], can be obtained by observing that the decision on the mth users symbol can be expressedas 6, = w [ Re {xm(-xm(0)}], where ~,~(i) = ilCi+)r(t)gm(t
IT

-iT)dt

and g,(t) is a function that lies in the subspacegenerated by the signature signals with a representation g,(t) = 5 hi%(t).
k=l

(4.11)

11

The resulting structure is identical to that of the conventional detector in Fig. 1 with the impulse response of each users matched filter, say f,,( T - t), replaced by g,( T - t). Since the use of more complex matched filters replaces the O(K) software complexity, the choice of structure depends solely on the tradeoff between computational complexity and ease of realization. In decentralized reception, the second realization is minimal in the sense that it requires only as many matched filters as the number of signals to be demodulated. Proposition 2: The asymptotic efficiency of user m for an arbitrary linear detector h is given as4 1 (h) Tm = 4E,hTR% .max2 O,min l,Br$; ( i -,eT(ZleT(Z+B)RhI (4.12)
the analysis of a

z3)$},

4The superscript (h) in 7:) and Pch) identifies m noncoherent linear multiuser detector.

1012

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 37, NO. 4, JULY 1991

where e denotes the vector of complex amplitudes with from straightforward calculations as kth entry as Ei/ exp( j0,) and where B, is the set of a;(i) = E[ (x,(i) - Ph(i))(Xm(i) - ph(i))] admissiblevalues of the diagonal information symbol matrix B restricted by b, = + 1. = E f hkyk(i) ; I=1 Proof: Let us obtain the mth user bit-error rate, [ k=l denoted as P$(v), of the linear multiuser detector h E R = 2cr2hTRit i= -l,O. (we drop the superscript for convenience). Since this Therefore the random iariables X,( - 1) and X,(O) have error probability, conditioned on the mth users bit, is a common variance. Further, they are uncorrelated beequal for b, = + 1 or b, = - 1, we have cause they are linear combinations of statistics obtained by processing the received signal over nonoverlapping P(h)(c) m time intervals [ - T,O] and [O,T], respectively. We can apply Lemma 1 to obtain the probability inside the summation in (4.13) so that 5 hkZk(-1) :x$,(O)
k=l I=1

= 21pK C

Pr [Re (X,( - 1)X,(O)) < oJ~], (4.13)

P,$)(a) = 2l-k

je(uh,bh) )}. (4.14)

B E 4,

1 - y( a/$,) exP(-fy

where the last equation results from averaging over the where ah and b, are given as restricted set B, of admissible values of B that are the ah = (2&?%) 7erZ#Z - eTBR%I, equiprobable interfering bit combinations. Conditioning on the bit combinations and the energies and phases of (4.15) b, = (2~di%-~1e~Rh + eBI&l. different transmissions, X,<- 1) and X,(O) are linear combinations of Gaussian random variables and hence We obtain asymptotic efficiency by substituting (4.14) into (1.2),

nE)=sup

Osrll;

u-0

lim

exp( - rE,/2a2)

= min sup B E4

exp( - rE,/2a2)

where the second equality follows from the fact that in are Gaussian. Further, the means and variances of the the limit as (T+ 0, the summation of conditional error Gaussian random variables X,(i) are denoted as ph(i) probabilities in (4.14) is dominated by the term correand a;(i) for i = - l,O. Each of these quantities is easily sponding to the least favorable interfering bit combinaobtained as tion. The penultimate equation is obtained by using the fact that Q<a, b) - 1/2Za(ab) exp (-[a* + b2]/2> K c h,Z,(-1) =eTRx, O(U exp[ -(b - aj2/2]> for b > a [l]. The result in (4.12) ph(-l)=E [ k=l is obtained by substituting the expressionsfor ah and b, from (4.15). The proof of Proposition 2 is therefore comwhere equality follows from the first equation. of (3.6). 0 plete. Similarly, We note here that in the hypothetical scenario where the received signal phases are all equal, and where the ~~(0) = E ; h,Z,(O) = eTBR%, signature signals of all the users are real-valued, as in I=1 direct-sequence spread-spectrum signaling, and with a where equality follows from the second equation of (3.6). further restriction on the linear transformation h to be The variances of X,(- 1) and X,(O) can be obtained real-valued, the expression for asymptotic efficiency in

VARANASI AND AAZHANG: OPTIMALLY NEAR-FAR RESISTANT MULTIUSER DETECTION

1013

CDMA channels. This point is further clarified. From (4.10) and the expressions in (3.61, it is seen that the desired signal plus multiuser interference component of the decision statistic of any linear detector is identically equal to Re(eTRd;tTRBZ). For the decorrelating linear detector d, and assuming that the mth user is linearly independent, we use the fact that Rd= u,, whence the desired signal plus multiuser interference component is simply E, b,, i.e., it corresponds to a detector in 1Rthat eliminates the multiuser interference from its decision statistic. Equivalently, note from (4.11) that the mth user matched filter of the decorrelator is matched to g:(t) = Cfz,dim)fk(t). Since g:(t) is orthogonal to each of the interfering signature signals f,(t), it is able to effectively tune-out the multiple-accessinterference. Therefore, owing to its striking similarity with its counterpart in coherent multiuser detection [8] and in keeping with the (h) = &maxz(O,min(l,min{~e,[R%]l nomenclature suggestedtherein, the linear detector d will 71 be referred to as the noncoherent decorrelating detector. It is important to note the distinction between the coherent and the noncoherent decorrelators. The normalized signal correlation matrix in [8] is independent of the I[J5wl~2ml2I =max2 O,min 1, signal energies but depends on the signal phases, and so m does the coherent decorrelating detector. In contrast, the Ii ) i i where a2 = \IE2/EI exp(j(0, - 0,)) and [.li denotes the noncoherent decorrelating detectors introduced here are ith element of the vector in square brackets. Consider the independent of both the signal energies and phases. Proposition 3: If the signature signal of user m is linspecial case of the two-user conventional detector. It is early independent, the bit-error rate of the mth user easily verified that noncoherent decorrelator is independent of the complex amplitudes of the interfering users and is equal to5 where R,, is an element of the normalized complex crosscorrelation matrix R defined in Section II. Therefore asymptotic efficiency of the conventional detector is identically equal to zero when the interfering user is sufficiently strong relative to the desired user, i.e., when
JE2/EI 2 1/ IR,,l.

(4.12) is identical to that of the linear detector for coherent CDMA channels employing binary phase-shift keyed modulation obtained in [S]. It is in this sense that the problem considered here is a complex arithmetic generalization of the real-valued problem considered in that paper. However, in contrast to the situation for the coherent problem, where the discrete minimization over the interfering bit-combinations was immediate, no closed form expressionis available for the minimization in (4.12) in the general K-user case. It is instructive to consider the two-user case. A closed form expressionfor asymptotic efficiency can be obtained in this case by using the fact that the magnitude of the sum of two complex numbers is not less than the difference between their magnitudes. For instance, the asymptotic efficiency of user 1 in a two-user system is given as

p.I:(V)=;exP(-

2V:;,.,)?

(4.16)

Having obtained the asymptotic efficiency of a noncoherent, linear multiuser DPSK detector, we now proceed E,[R:,l- . to the next section with the objective of finding the Proof: We evaluate the bit-error probability of d optimal detector from the class of linear detectors, the from (4.14) and (4.15) that characterizes the bit-error optimality criterion being near-far resistance. probability of any linear detector. Since the mth user is linearly independent, the decorrelating transformation A. Optimal@ of DecorrelatingDetectors satisfies Ra= u,, and we have A noncoherent decorrelating detector for user m is Pid( a) = 2-K c Q( ad, bd) defined by the decision in (4.10) with the linear transforB E 4, mation h = d where d denotes the complex conjugate of 2 1 the mth column of a generalized inverse R of R. If the ) - zZo( a,b,) exp (-a+b y mth user is linearly independent, it can be shown (cf. [8, Lemma 11) that Ra= u,, the mth unit vector. If all the signature signals are linearly independent, then R- ex=2-K C exp(-P)=iexp( -&), ists and the decorrelating transformation d is uniquely B E& characterized as the complex conjugate of the mth colwhere the penultimate equation follows from the fact that umn of the inverse of R. In the sequel, we do not need ad = 0 and the last equality from the fact that b, is the latter assumption. The definition of the class of decorrelating transforma5The superscript (d) in q!$) and PA) identifies the analysis of tions is similar to that introduced in 181for coherent noncoherent decorrelating multiuser detectors.

where R,:, is the mth diagonal element of the unique Moore-Penrose generalized inverse R+ of R. This property is of special significance since it implies that no matter how strong the interfering signalsare or what their carrier phases may be, the decorrelating detector has an error rate equal to that of a single user with energy

1014

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 37, NO. 4, JULY 1991

r-----jh

f,(t)

z
& m 2 2 2 .: g

0.9
0.8 0.70.60.5 0.4 0.30.2\\\ \ --

Decorrelating

detector

---

y, q
(a)
IO cj

fp
I

.o
5

Single-user

detector

E $ a

'L

O.l0.0
IO

\t . I I I
-8 -6 -4

I I
-2 0 E2/El(in

I I I I ,2 dB) 4 6 8 10

Fig. 4. Asymptotic efficiency comparison between the single-user and decorrelating detectors in the 2-user case when the correlation is 0.5.

Decorrelator(Z-user)

Decorrelator(3-user) 4; --

0.6

Single-user

detector(2

10

12

14

SNR(1) in dB
(b) Fig. 3. (a) Direct-sequence signature signals derived from Gold sequences of length 7 assigned to the four users of a four-user DS-SSMA system. (b) Bit-error rate of first user as a function of the first users signal-to-noise ratio. These error rates are independent of interfering signal energies and phases.

:I::
0.3 -10 -8 -6

Single-user

detector(3-user)

-4

-2 0 Ei/El(in

2 dB)

10

independent of B and is equal to a-Id-, where d, is the mth element of d and hence the mth diagonal element of any generalized inverse of R, which in turn can be shown to be equal to RL, [a]. The expressionfor error probability in (4.16) follows. 0 Fig. 3(a) shows four direct-sequence signature signals derived from Gold sequencesof length 7 assigned to a four-user direct-sequence SSMA system. Fig. 3(b) depicts the bit-error rate of the decorrelator for the first user as a function of the first users signal-to-noise ratio, as the second, third and the fourth users become active. These error rates are independent of the interferers signal energies and phases. Corollary 1: The asymptotic efficiency of any decorrelating detector for a linearly independent user is given as 1 (4 -R>o. Tm mm (4.17)

Fig. 5. Asymptotic efficiency comparison of the conventional singleuser and the decorrelating detectors for the four-user DS-SSMA system as the first two, three, and finally four users become active.

the decorrelator has a nonzero near-far resistanceand is therefore robust to interfering signal uncertainties. Proof: Substituting the bit-error rate of a decorrelating detector from (4.16) of Proposition 3 into the definition of asymptotic efficiency in (1.2), we have

The same expressioncan also be obtained by using Rd= urn in the expression for asymptotic efficiency for an arbitrary linear detector in (4.12). In order to obtain (4.17), we have to show that RL,k 1, a fact that we establish in the Appendix. This completes the proof of Corollary 1. 0 The asymptotic efficiency of the decorrelator in comparison with the conventional detector for the two-user case is depicted in Fig. 4 and for the four-user directsequenceSSMA systemof Fig. 3(a), is depicted in Fig. 5.

Since asymptotic efficiency of a decorrelating detector is independent of interfering signal energies and phases, it is also equal to its near-far resistance. This implies that

VARANASI AND AAZHANG: OPTIMALLY NEAR-FAR RESISTANT MULTIUSER DETECTION

1015

The performance measureof interest is bit-error probability. The set of design strategies is the set of linear detectors 0. The uncertainty for the detector designer arises from two quantities; the bit-combination of the interfering users and their energies and phases that are both determined by the transmitter. These two quantities determine the operating point for the receiver. The objective is to find the linear detector in R that is robust to operating point uncertainties. Therefore we will consider a minimax approach, where the design goal would be to find an optimal detector from R that optimizes the worst-case performance over all admissible operating points. Consider the next proposition. Proposition 4: For a linearly independent user, a decorrelating detector achievesthe highest near-far resistance from among the class of linear detectors, i.e.,
d = arg ~z; e54, r$),

pend only on the information bit of the desired user. The specific choice of this dependency is dictated by the dual-objective of obtaining a tight bound and at the same time retaining a mathematically tractable problem. Fortunately, these two objectives are served simultaneously by the choice b = u, i.e., when all the information bits are equal to that of the desired user. In this case, the minimax problem has a natural setting in the finite-dimensional complex semi-inner*product space CK with semi-inner product defined as (x, Y)~ = xTRj. Note that it denotes a valid semi-inner product since R is a nonnegative definite Hermitian matrix. Therefore, choosing &$i)(u> as an upper bound for q$r), we have

(4.18)

where E,,, denotes the uncertainty set for demodulation of the mth users signal which is the set of admissible values of interfering signal energies and phases and can therefore be written as 5m = {e E CK; [el, constant}.
Proof: We show that the highest near-far resistance where the second equality follows from the nonnegativity achievableby any linear detector is upper bounded by the of the objective function +(h, a) e ](a, hjR12/(h, h>R and

near-far resistanceof a decorrelating detector. Using the expression for asymptotic efficiency in (4.12) and the definition of near-far resistance in (l.l), an upper bound on the highest achievable near-far resistance is obtained as follows: i-p) = sup inf max2 111 hEi Ed%? 1, 1 24s

the last inequality follows from the minimax inequality [6]. Interestingly, the penalty function q?(h,a> is identical to the signal-to-noise ratio functional encountered in the robust matched filtering problem [ll], [18]. The inner

min le(Z+ B)ti BE%

jlr(Z-B)zy

e&y(b),

VbeB,,

(4.19) maximization in the last equation of (4.20) is easily obtained by using the Cauchy-Schwarz inequality (which can be shown by a simple extension of the Cauchy-Schwarz inequality for inner products [6]), since R can be chosen to contain A, so that l(a,h)R12
sup hECK hRh#O ch,hjR =(a,a)R,

where the normalized amplitude vector a and the set of its admissiblevalues A, were defined in Section II and 0 is chosen as the set {h E CK; hTZ8z# 0). The first inequality is obtained from the minimax inequality [6] and the second inequality holds for each interfering bit-combination b E B, and the expression on the right side of this inequality, which we have defined as &z)(b), can be interpreted as the highest near-far resistance achievable by a linear detector in a hypothetical communication environment. This environment consists of co-operative interfering transmitters sending information bits that de-

(4.21)

with the result that if the mth user is linearly indepen-

1016

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 37, NO. 4, JULY 1991

dent, (a,a)R}=min m

possible to find a vector a with a, = 1 such that it is orthogonal to the vector & when the latter is not equal to the mth unit vector. Further, since it constitutes an l,R+ upper bound for near-far resistance, which is nonnegamm(4.22) tive by definition, we have the desired result. 0

The last equality is obtained by noting that the minimum-norm optimization problem in the first of the above equations is the complex;arithmetic version of that solved in Proposition 1 of [8] and that the proof of that proposition extends to the complex case under consideration here. However, a simpler proof involves restating [9] the minimum-norm optimization problem as inf(a,a)R subject to (a,djR = 1, where d is a decorrelating transformation. The existence of d is guaranteed by the linear independence of the mth user. We invoke the Cauchy-Schwarz inequality for semi-inner product spacesagain, so that
LZEA,

This corollary together with the characterization of the near-far resistance of the decorrelating detectors in Proposition 3 could have been used to establish Proposition 4. However, the game-theoretic approach that we followed brings out the close parallels between the optimum near-far resistanceproblem formulations for coherent and noncoherent channels. Corollary 3: For a linearly dependent user, there is no linear detector that is near-far resistant. Proof If the mth users signature signal is linearly dependent, it can be shown following [81 that inf aEA (a, a)R = 0, since there is a linear combination of the colimns of R with a nonzero coefficient for the mth column that is identically equal to zero. Therefore, it is possible to find an a with a, = 1 such that Rii = 0. Together with the first equation in (4.22), we have $?)<u) = 0. Since optimum near-far resistance, Yjr)---nonnegative by definition-is bounded from above by &F)(U), we have q(o) = 0, which completes the proof. 0 m B. Optimum Near-Far Resistance

min (a,a)R=

&=bw

(4.23)

Summarizing, from (4.191, (4.201, (4.20, and (4.23), and using the fact that R L, 2 1 which was obtained in Proposition 3, we have the following:
T$fk &f(u) < [RLm] -I.

(4.24)

The discussionthus far does not exclude the possibility From Proposition 3, a decorrelating detector achieves a of the existenceof detection schemes,that are not of the near-far resistance equal to the upper bound above, whence the desired result follows. 0 linear type considered so far, which may achieve a higher near-far resistancethan that of the noncoherent decorreWe remark here that equality in the inequality of (4.20) lators. Consider the following proposition. holds becausethe functional $(h, a) has a saddle point, a Proposition5: No detector, linear or otherwise, achieves fact that we now establish. The seminorm in (4.23) is a higher near-far resistance than that of a noncoherent minimized uniquely for the least-favorableoperating point linear decorrelator. aL = [ RL,]-d. The optimal detector for the least-favorProof: The central idea here is to find an upper able operating point can be obtained from the condition bound on the highest near-far resistance achievable by for equality in the maximization problem in (4.211,which is h, = aaL = ad, and we assume LYto be unity for analyzing the performance of an optimum receiver that, convenience. The pair (d,[RL,]-d) can be shown to be in addition to observing{r(t); t E [ - T, T]}, has additional a saddle point if it can be verified that [R~,]-d is the side information. In particular, suppose that the receiver has a perfect knowledge of the signal energies and phases. worst operating point for d, i.e., if In this case, the observation of the received signal in the [RLm]-d=argarnl; $(d,a), time interval [ - T,O] provides no information about the m information symbols in [O,T]. Therefore, the minimum or equivalently, that error probability in the demodulation of the mth users bit is equal to the error probability of the coherent I(d,[R~,]-d),l~I(d,a)RI, vaEA,> multiuser minimum-error probability. detector, exponenwhich can be immediately confirmed. tially tight bounds for which have been obtained in [19]. Corollary 2: A linear detector that is not a decorrelator In particular, it was shown that for any 6 > 0, there exists is not near-far resistant. a0 > 0 such that for all (+< uo, the minimum-error probability of the mth user satisfies Proof: From (4.201, the conditional near-far resistance of a detector h that is not a decorrelator is given as 4,) = min 1, inf
a=A,

l(a,h)Ri2
(h,h)R

C,Q(A,i,(m)/u> ~f$)~ C,(l+ ~>Q(A,i,(m)/u),

(4.25)

where C, and CE are positive constants and A,Jm> is since the penalty function is nonnegative and it is always the Euclidean distance between signals corresponding to

VARANASI AND AAZHANG: OPTIMALLY NEAR-FAR RESISTANT MULTIUSER DETECTION

1017

the two closest hypotheses that differ in the mth users where the last equation was obtained from the solution to bit, i.e., the minimum-norm optimization problem of (4.22) and the fact that R+ 2 1, both of which were obtained in mm Proposition 4. Finally, since the near-far resistanceof the minimum error probability receiver with side information overbounds the near-far resistanceof any DPSK receiver = (without this side information), and since from Proposimin ETHE) EE{-l,o,+l)K tion 3 the near-far resistanceof the decorrelator achieves Em=+1 this upper bound, we have the sought result. 1 where Hkl = (E,E,)/Rk, exp(j(0, - 6,)). Using the upper and lower bounds for the complementary error funcV. CONCLUSION tion [14], we have from (4.25) that for any 6 > 0, there The main contribution of this paper is finding the exists a ut, > 0 such that for all u < (TV, error probabilthe decorrelating linear multiuser detector for noncoherent ity of the mth user satisfies demodulation of differentially phase-shift keyed transmissions in a synchronousCDMA channel. This detector was obtained as a solution to the minimax optimization problem of finding the linear detector that optimizes the worst-case asymptotic efficiency over near-far environments. It was also shown that not only is the decorrelator optimal in terms of near-far resistanceamong the classof linear detectors, but it also achieves the highest near-far resistanceachievableby any noncoherent DPSK detector. The only restriction for the linear decorrelator to be near-far resistant for a particular user is that the correTherefore, the asymptotic efficiency of the minimum er- sponding signature signal be linearly independent of the ror probability coherent multiuser detector, denoted as rest of the interfering signature signals-a mild restricT& is given as tion compared to the requirement of the conventional detector that it be orthogonal to the other signature ,t=min(l, A::) ) signals. The performance invariance of the decorrelator to interfering signal uncertainties together with its ease of realization-neither requiring elaborate energy and phase min E, e*He estimation and tracking schemesnor any computation in cE(-1,0,+1)K excessof the conventional detector-make it eminently Em=+1 suitable for practical applications. = min {1,2Az&}, While it is conceivable that transmissions emanating EE{-l,O,+l)K c,=+l from a central transmitter in multicast networks may where A = diag(a). The near-far resistance of the mini- arrive at a receiver synchronously,it is difficult to achieve mum-error probability receiver with side information is this synchronization in multipoint-to-point communicaobtained from the following equations that are reminis- tion. Of considerable interest, both from the viewpoint of cent of those that yield the near-far resistancefor coher- theory and practice, is the generalization of this work for ent channels (note that the definitions of asymptotic the noncoherent asynchronousCDMA channel. efficiency and near-far resistance for coherent and noncoherent channels are not identical) of the minimumACKNOWLEDGMENT error probability detector in [8] The authors would like to thank the reviewers for their helpful comments and one reviewer in particular for YjL=aF{ min 1, min ETARAE providing detailed and insightful suggestionsfor improvm i l E{-1,0,+1JK 1 Em=+1 ing a previous version of this paper in both style and content. =min 1, min inf E*& i EE{-l,O,+ljKaE& 1 APPENDIX Em=+1 =rnin(l,~~~+v*R~) = l/R;,,,, 0, if m th user is linearly independent, if mth user is linearly dependent, In this appendix, we show that R,&,, 2 1, a result that was required to prove Corollary 1 of the paper. The proof given here is an extension of the corresponding result in [7] for real and invertible matrices. Since R is a nonnegative definite Hermitian matrix, it is unitarily similar to the diagonal matrix A = diaglh,, . . *, AK} of its nonnegative

1018

IEEE TRANSACTIONS ON INFORMATION

THEORY, VOL. 37, NO. 4, JULY 1991

eigenvaluesso that R = UAGT, where U has the eigenvectors as its columns. Further, the Moore-Penrose inverse can be written as R+ = UA+ UT, where A = diag{h:; . ., hi}with A:=hL: when iEZ with Z={~E u,2,*. ., K}; Ai > 0) and A,?= 0 when j E Z. Therefore, R+ = l&G, = Ci EIAil~,i12 (real and positive) with uz bz:g the mth row of U. Since the diagonal elements of R are all equal to unity, we have Cit IAil~,i12 = 1, so that

where we use the fact that n + x-l 2 2 for x > 0. Since UaT= I, we have C~rl~,~l~ = 1. For some k E I, i.e., when A, = 0, we have Ru, = 0. A linear combination of the columns of R that is equal to zero cannot include the linearly independent columns, i.e., u,~ = 0 for every m E I. Therefore Ci t II~,i12 = 1 whence it follows that Rz, 2 1.
REFERENCES [l] S. Benedetto, E. Biglieri, and V. Castellani, Digital Transmission Theory. Englewood Cliffs, NJ: Prentice-Hall, 1987. [2] G. R. Cooper and R. W. Nettleton, A spread-spectrum technique for high-capacity mobile communications, IEEE Trans. Veh. Technol., vol. VT-27, pp. 264-275, Nov. 1978. [3] E. A. Geraniotis, Performance of noncoherent direct-sequence spread-spectrum multiple-access communications, Special issue on-military communications, IEEE J. Select. Areas Commun., vol. SAC-3, pp. 687-694, Sept. 1985. Noncoherent hybrid DS-SFH spread-spectrum multiple141 -1 access communications, IEEE Trans. Commun., vol. COM-34, no. 9, pp. 862-872, Sept. 1986.

[5] A. W. Lam and D. V. Sarwate, Multiple-user interference in FHMA-DPSK spread-spectrum communications, 1EEE Trans. Commun., vol. COM-34, no. 1, pp. l-13, Jan. 1986. [6] D. Luenberger, Optimization by Vector Space Methods. New York: John Wiley, 1969. [7] R. Lupas-Golaszewski and S. Verdu, Asymptotic efficiency of linear multiuser detectors, in Proc. 25th Conf. Decision Contr., Athens, Greece, Dec. 1986. [8] R. Lupas and S. Verdu, Linear multiuser detectors for synchronous code-division multiple-access channels, IEEE Trims. Inform. Theory, vol. 35, no. 1, pp. 123-136, Jan. 1989. [9] -, Near-far resistance of multiuser detectors in asynchronous channels, IEEE Trans. Commun., vol. COM-38, pp. 496-508, Apr. 1990. [lo] J. Marcum, A statistical theory of target detection by pulsed radar, IEEE Trans. Inform. Theory, vol. IT-6, no. 2, pp. 59-267, 1960. [I 11 H. V. Poor, Robust matched filters, IEEE Trans. Inform. Theory, vol. IT-29, no. 5, pp. 677-687, Sept. 1983. [12] J. G. Proakis, Digital Communicutions. New York: McGraw Hill, 1989. [13] S. Stein, Unified analysis of certain coherent and non-coherent binary communication systems, IEEE Trans. Inform. Theory, vol. IT-lo, pp. 43-51, Jan. 1964. 141 H. L. van Trees, Detection, Estimation, and Modulation Theory, Part I. New York: Wiley, 1968. 151 M. K. Varanasi and B. Aazhang, Probability of error comparison of linear and iterative multiuser detectors, in Lecture Notes in Control and Information Sciences Series, Springer-Verlag & Proc. Advances in Communication & Control Systems, ComCon 88, Baton Rouge, LA, Oct. 1988, pp. 54-65. 161 ~, Multistage detection in asynchronous code-division multiple-access systems, IEEE Trans. Commun., vol. COM-38, pp. 509-519, 1990. [17] -, Near-optimum detection in synchronous code-division multiple-access systems, to appear in IEEE Truns. Commun., vol. 39, no. 5, May 1991. [18] S. VerdG and H. V. Poor, Minimax robust discrete-time matched filters, IEEE Trans. Commun., vol. COM-31, no. 2, pp. 208-215, Feb. 1983. [19] S. Verdti, Minimum probability of error for asynchronous Gaussian multiple-access channels, IEEE Trans. Inform. Theory, vol. IT-32, no. 1, pp. 85-96, Jan. 1986. [20] ~, Optimum multiuser asymptotic efficiency, IEEE Trans. Commun., vol. COM-34, pp. 890-897, Sept. 1986. [21] -, Recent progress in multiuser detection, in Lecture Notes in Control and Information Sciences Series, Springer-Verlag & Proc. Aduances in Communication & Control Systems, ComCon 88, Baton Rouge, LA, Oct. 1988, pp. 66-77.