Sie sind auf Seite 1von 12


12, DECEMBER 2002 3117

[12] , “Space–time block codes for high data rate wireless communica- have shown that, with the simultaneous use of multiple transmit and re-
tion: Performance results,” IEEE J. Select. Areas Commun., vol. 17, pp. ceive antennas, very large capacity increases can be unleashed [1]–[4].
451–460, Mar. 1999. At the same time, it is reasonable to expect that terminals supportive of
[13] V. Tarokh, A. Naguib, N. Seshadri, and A. R. Calderbank, “Space–time
codes for high data rates wireless communications: Performance cri- progressively higher data rates will tend to be naturally larger in size
teria in the presence of channel estimation errors, mobility and multiple and, consequently, they will be able to accommodate multiple closely
paths,” IEEE Trans. Commun., vol. 47, pp. 199–207, Feb. 1999. spaced antennas. Hence, the deployment of arrays at both base stations
[14] A. Wittneben, “Base station modulation diversity for digital SIMUL- and terminals appears as an attractive scenario for the evolution of mo-
CAST,” in Proc. IEEE Vehicular Technology Conf. (VTC’91), May 1991,
pp. 848–853.
bile data access.
[15] , “A new bandwidth efficient transmit antenna modulation diversity Great progress has been made toward understanding the informa-
scheme for linear digital modulation,” in Proc. IEEE Int. Communica- tion-theoretical capacity and the performance of multiple-antenna ar-
tions Conf. (ICC’93), 1993, pp. 1630–1634. chitectures with thermal noise as the only impairment (see [5]–[16]).
[16] L. Zheng and D. N. C. Tse, “Communication on the Granssmann Within the context of a wireless system, however, the dominant impair-
manifold: A geometric approach to the noncoherent multiple-antenna
channel,” IEEE Trans. Inform. Theory, vol. 48, pp. 359–383, Feb. 2002. ment is typically not thermal noise, but rather cochannel interference.
Thus, the objective of the present work is to extend this understanding
to the realm of spatially colored interference. We invoke, as central tool,
recent results on the asymptotic distribution of the singular values of
random matrices and their application to randomly spread code-divi-
sion multiple access (CDMA) [17]–[19]. Although these distributions
Capacity of Multiple-Transmit Multiple-Receive Antenna pertain asymptotically in the number of antennas, the results we derive
Architectures therefrom become virtually universal under ergodic conditions.
Since the focus is on mobile systems, we consider only “open-loop”
Angel Lozano, Senior Member, IEEE, and architectures wherein the transmitter does not have access to the instan-
Antonia Maria Tulino, Member, IEEE taneous state of the channel. Only large-scale information—defined as
information that varies slowly with respect to the fading rate—is avail-
able to the transmitter.
Abstract—The capacity of wireless communication architectures This correspondence is organized as follows. In Section II, the met-
equipped with multiple transmit and receive antennas and impaired by rics and models are introduced. In Section III, the noise-limited ca-
both noise and cochannel interference is studied. We find a closed-form
solution for the capacity in the limit of a large number of antennas. This
pacity is reviewed using the tools of asymptotic analysis. Such anal-
asymptotic solution, which is a sole function of the relative number of ysis is generalized, in Section IV, to environments containing spatially
transmit and receive antennas and the signal-to-noise and signal-to-inter- colored interference. The main result therein is an expression of the
ference ratios (SNR and SIR), is then particularized to a number of cases asymptotic capacity in the presence of both noise and interference. Fi-
of interest. By verifying that antenna diversity can substitute for time nally, Section V concludes the correspondence.
and/or frequency diversity at providing ergodicity, we show that these
asymptotic solutions approximate the ergodic capacity very closely even
when the number of antennas is very small. II. DEFINITIONS AND MODELS
Index Terms—Adaptive antennas, antenna arrays, asymptotic analysis, A. Propagation Model
channel capacity, diversity, fading channels, multiantenna communication,
multiuser detection. With transmit and receive antennas, the channel responses
from every transmit antenna to every receive antenna can be assembled

into an 2N M random matrix G
whose underlying random process
is presumed zero-mean and ergodic. The propagation scenario and the
With the explosive growth of both the wireless industry and the In- spatial arrangement of the antennas determine the correlation among
ternet, the demand for mobile data access is expected to increase dra- G
the entries of . The scenario we consider, typical of a mobile system,
matically in the near future. As a result, the ability to support higher is based on the existence of an area of local scattering around each ter-
capacities will be paramount. Capacity can be pushed by exploiting minal. Accordingly, the power angular spread is expected to be very
the space dimension inherent to any wireless communication system. large—possibly as large as 360 —at the terminals rendering the an-
Nonetheless, due to economical and environmental aspects, it is highly tennas therein basically uncorrelated. At the base station, the angular
desirable not to increase the density of base stations. Under such con- spread tends to be small [20], [21] but the antennas can be also decor-
straint, antenna arrays are the tools that enable spatial processing on a related by spacing them sufficiently apart [22]–[25]. Consequently, we
per-base-station basis. Recognizing this potential, the use of arrays at focus on channel matrices containing only independent entries.
base-station sites is becoming universal. Array-equipped terminals, on G
Since the elements of are identically distributed, it is possible to
=p .
define a normalized channel matrix with unit-variance entries such
the other hand, had not been contemplated in the past because of size
and cost considerations. However, recent results in information theory that G
B. “Open-Loop” Capacity
Manuscript received May 3, 2001; revised April 8, 2002. Perfect channel estimation at the receiver [26]–[28] is presumed.1
A. Lozano is with the Wireless Communications Research Department, Bell The impairment comprises additive white Gaussian noise (AWGN) as
Laboratories, Lucent Technologies, Holmdel, NJ 07733 USA (e-mail: aloz@lu-
A. M. Tulino is with the Department of Electrical Engineering, Princeton
University, Princeton, NJ 08540 USA (e-mail: 1The penalty associated with channel estimation is small as long as the coher-
Communicated by D. N. C. Tse, Associate Editor for Communications. ence time of the channel—measured in symbols—is large enough with respect
Digital Object Identifier 10.1109/TIT.2002.805084 to the number of transmit antennas [27].

0018-9448/02$17.00 © 2002 IEEE


well as interference, which—conditioned on its fading—is also pre- TABLE I

sumed Gaussian.2 As a consequence, we are interested in the spatial
covariance of the impairment conditioned on the fading of the inter-

ference. Such spatial covariance, which we define as , is also pre-

sumed to be estimated perfectly at the receiver (but unknown to the

transmitter). Notice that, since depends on the random fading of the
interference, it can be viewed as a random variable itself. Furthermore,
given that is—in general—not proportional to the identity, it follows
that the interference is spatially colored.
When the entries of Gare independent and unknown to the trans-
mitter, the mutual information is maximized by transmitting a Gaussian by the spreading factor, corresponds in our case to the number of re-
signal with spatial covariance ceive antennas, while the number of users corresponds to the number
8 = MP IM (1) of transmit antennas. There are, nonetheless, significant differences.
• The spatial signatures for the different transmit antennas are not
given a total radiated power P [2]. Since the channel is time varying in
chosen by the system designer, as in CDMA, but rather imposed
nature, such mutual information fluctuates with it. With a sufficiently
by nature. Moreover, their distribution depends on the type of
long coding horizon, it is possible to code over the short-term channel
fading being experienced. Remarkably, though, the distribution
fluctuations and approach the ergodic capacity given by [11]
of the eigenvalues of y converges to the same exact function
C =E IN + MP gHHy Q01
log2 det (2) regardless of the distribution of its entries so long as those are
independent and identically distributed [17].
with expectation over the distributions of H and Q. It is also possible to
• Because of the total power constraint imposed at the transmitter,
code over the short-term channel randomness in the frequency domain,
the SNR typically used in CDMA has to be normalized by the
whereby the ergodic capacity is approached as the signal bandwidth
dimensionality ratio .
increases [30].
If it is not possible to code over the short-term channel variations, • The transmit antennas are colocated and part of a single trans-
one must resort to the idea of outage capacity, wherein the capacity ceiver whereas, in CDMA, the multiple users are geographically
itself is regarded as a random variable that fluctuates with the channel. dispersed. Hence, joint coding of the transmit signals is feasible
As we shall see, nonetheless, the outage capacity hardens around its in our problem, but not in CDMA.
average as the number of antennas increases and thus the outage and
With these considerations, the asymptotic capacity per dimension
ergodic capacities coincide asymptotically.
derived in [17] and [32] for synchronous CDMA can be modified to
express the noise-limited asymptotic capacity per receive antenna as
A. Nonasymptotic Noise-Limited Capacity C ( ; SNR) = log2 1 + SNR 0 F ;

When the impairment consists exclusively of AWGN, we have
Q = 2 IN (3)
+ log2 1 +

0F ;

where  2 is the noise power per receive antenna. Defining 2 (e) F
def P
0 logSNR ;

SNR = g (4)
2 with
as the average signal-to-noise ratio (SNR), we can write
F (x; y) def
= 4
1+y 1+
px 2 0 1+y 1 0 px 2
: (9)
C = E log2 det IN + SNR
HHy : (5)
The asymptotic capacity is solely a function of and the SNR. Fur-
thermore, it yields an extremely accurate approximation to the ergodic
B. Asymptotic Noise-Limited Capacity capacity even when the number of antennas is very small (see Table I).
As the number of antennas increases, the empirical distribution of the Next, we particularize C for a number of special cases and study
eigenvalues of the random matrix y converges to a deterministic
HH in detail its dependence on both and SNR.
function [31]. Defining the ratio of transmit and receive antennas as 1) Dependence on : C is maximized by letting ! 1.
Hence, the capacity with an optimal receiver3 increases monotonically
= (6) with , as seen in Fig. 1. Although additional antennas thus increase
N capacity regardless of whether they are added to the transmitter or to
and the capacity per receive antenna as the receiver, they add different value depending on where they are de-
C def
(7) ployed. Such difference stems from the fact that the total radiated power
is bounded whereas the total captured power increases with additional
we now look into finding C as the number of antennas is driven to receive antennas. Therefore, the capacities corresponding to and to
infinity with a constant ratio . 1= are different.4 Additionally, Fig. 1 hints that there is little incen-
Fundamentally, we are faced with a multiple-input multiple-output tive in increasing beyond unity. This result is formalized in the next
problem that is akin to that of a synchronous CDMA channel with section.
random spreading. The dimensionality of such a channel, represented
3This is not necessarily true if a suboptimal receiver is utilized [6].
2Ifthe interference is not Gaussian, the capacities we derive serve as lower 4In “closed-loop” architectures, wherein the channel realization is known at
bounds [29]. the transmitter, these capacities coincide [2].

Fig. 1. Asymptotic noise-limited capacity per receive antenna as a function of for various levels of SNR.

Of particular interest is C with = 1, which can be expressed 2) Dependence on SNR: Insightful expressions can be obtained by
[14] as particularizing C at high SNR, wherein a series expansion of (8)
p yields
C (1; SNR) = 2 log2
1+ 1 + 4 SNR log2
e 0 ( 0 1)
p 1 log2 1 0 1 + O SNR 1 ; 1
0 log 2 (e) 01
: (10) C ( ; SNR) =
log2 e 0 (1 0 )
1 + 4 SNR SNR

On the other hand, if N is kept fixed while M ! 1, the capacity per 1 log2 (1 0 ) + O SNR
; 1
receive antenna becomes
which, for = 1, particularize to

!1 C C SNR 1
lim ( ; SNR) = log2 (1 + SNR) (11) (1; SNR) = log2 +O (16)
as derived in [1]. Therefore, the high-SNR capacity is proportional to
which can be obtained by taking the limit on (8) or, alternatively, by
min(M; N ) and grows logarithmically with the SNR. As shown in
observing [2] that the channel decouples asymptotically
Fig. 2, this trend sets in very fast.
!1 M HH = IN :
1 For an in-depth study of the low-SNR capacity, the reader is referred
lim (12) to [36], [37].

Notice that, for N = 1, this configuration reverts to that of transmit C. How Many Antennas Should be Used?
diversity, which has been actively researched in recent times [33]–[35].
Conversely, if M is kept fixed while N ! 1, the capacity per
The asymptotic expressions derived thus far can be used to quantify
the benefit of pushing beyond unity. The largest gain, which occurs
receive antenna behaves as for SNR ! 1 is

C ( ; SNR) = log2
+ O ( ) (13) !1
C (1; SNR) 0 C (1; SNR) = log2 (e) (17)
or merely the equivalent of a 4.34-dB increase in SNR (Fig. 2), a van-
and the channel again decouples ishingly small improvement in capacity per receive antenna as the SNR
!1 N H H = IM :
lim (14) Given the cost associated with deploying multiple antennas with sep-
N arate radio-frequency chains, it is important to ensure that those an-
Interestingly, this decoupling also implies that the capacity would not tennas are as effective as possible. Ineffective antennas raise the cost
increase if the channel realization H
were known at the transmitter and add unnecessary complexity.
[11]. Hence, the “open-loop” and “closed-loop” capacities coincide for In “open-loop” noise-limited conditions, the following applies.
! 0. Intuitively, as the number of channel dimensions becomes • Given a number of receive antennas, at most as many transmit
much larger than the number of modes, none of these gets favored over
antennas should be used.
the others. Moreover, as can be inferred from [19] for the corresponding
multiuser problem, capacity can be approached for small with scalar • Given a number of transmit antennas, at least as many receive
coding at every transmit antenna and a simple linear receiver. antenna should be used. Additional receive antennas are always

Fig. 2. Asymptotic noise-limited capacity per receive antenna as a function of SNR for various levels of .

advantageous. However, if we denote by R the highest rate that In the presence of outside interference, the covariance of the impair-
can be supported with a realizable constellation by each transmit ment conditioned on the fading of the interference can be expressed as
antenna, there is little point in further increasing the number of K Pg
Q= k k Hk Hy + 2 IN
receive antennas once
k (19)
C k=1 Mk
> R: (18)
with K the number of outside interferers (assumed mutually indepen-

dent) and with Pk and Mk the total radiated power and number of
As we will see next, these conclusions vary in the presence of
transmit antennas for the k th interferer, respectively. In the downlink,
colored interference.
each interferer would correspond to a neighboring base station; in the
uplink, it would correspond to a terminal in a neighboring cell. Each
N 2 Mk channel matrix k contains the transfer coefficients between
every transmit antenna of interferer k and every receive antenna at the
A. Nonasymptotic Capacity
desired user. We denote by k the normalized counterpart of each
Within the context of a mature system, the dominant impairment is
usually not thermal noise, but rather cochannel interference. Further-
G p
channel matrix so that k = gk k . Defining the signal-to-inter-
ference ratio (SIR) with respect to each interferer as
more, most emerging data systems feature time-multiplexed downlink
channels, certainly those evolving from time-division multiple access SIRk = PPk ggk (20)
(TDMA) [38], but also those evolving from CDMA [39]–[41]. Hence,
same-cell users are mutually orthogonal and thus the downlink interfer- we can rewrite the capacity as
ence arises exclusively from other cells. This holds also for a code-mul-
tiplexed downlink in frequency-flat fading channels5 and, of course, for
a time-multiplexed uplink.
C = E log2 det IN + HHy
In order to simplify the analyses, it is not unusual that such outside
K 01
interference be regarded as additional AWGN [44]. In single-antenna
Hk Hky + SNR
IN (21)
k=1 Mk SIRk
systems, this approximation assumes the interference to be Gaussian. In
multiple-antenna systems, it has a second—and more profound—im-
plication: it assumes the interference to be spatially white and it thus with expectation over the distribution of the various channels and with
neglects the fact that interference does have a spatial structure, or color, the aggregate SIR and signal-to-interference-and-noise ratio (SINR)
that can be exploited by a multiple-antenna receiver. This structure being

1 = K
tends to be particularly strong in the downlink, wherein the entire in-
terference contribution of every cell emanates from the base station, a 1 (22)
single localized source. SIR
1 = 1 + 1 :
5In frequency-selective channels, same-cell interference can be suppressed The performance of receive antenna arrays in interference-limited
by preceding the receiver with a chip-level equalizer [42]. Alternatively, well- conditions has been thoroughly studied, but mostly for single-antenna
established multiuser detection principles can be applied [43]. transmitters [45]–[47]. With the addition of transmit arrays, the

problem becomes much more complex (see [11] as well as [48]–[51] how the detection (by a linear MMSE receiver) of each of the
for related work). In general, the individual channel matrices k H transmit signals is impaired by the presence of the other transmit
cannot be estimated, but only the aggregate covariance , and thus Q antennas plus the outside interference.
the receiver can only perform linear processing against the outside in-
• 2 represents the asymptotic ratio between i) the SNR at the
terferers while optimally detecting the desired signals [52]. Hence, we
output of a linear MMSE receiver detecting the signal transmitted
find it convenient to distinguish between two classes of interference.
from any outside antenna in the presence of the rest of outside an-
• The mutual interference among the transmit antennas of the de- tennas (excluding the desired user), and ii) the SNR at the output
sired user, which we shall refer to as multiantenna interference. of a matched filter detecting that same antenna without any in-
terference. Although it may seem intriguing that 2 plays a role
• The outside interference that those antennas suffer from the K
interferers within .Q in the computation of the capacity given that the actual receiver
does not make any attempt to decode the outside interference, a
justification is given in the Appendix.
B. Asymptotic Capacity
The asymptotic capacity in the presence of both AWGN and outside Clearly, both 1 and 2 are bound to lie within [0; 1].
interference (derived in the Appendix) constitutes the main result of 1) Limiting Cases: Solving for the asymptotic efficiencies be-
the correspondence. Although we present it next for a homogeneous comes trivial in some limiting cases. For growing
system, wherein the number of transmit antennas is the same for the 1
user of interest as well as each of the outside interferers, the solution in
!1 1 = 1 + SNR 1+ 1
its most general form is included in the Appendix. In a homogeneous
system, the asymptotic capacity per receive antenna is given by SINR

!1 2 = SNR (26)
+K ( ; SNR; fSIRk g) =
C log2
indicating that both efficiencies become independent of the composi-
k=1 tion of the outside interference; they become a function of only the
1 aggregate SIR and SINR. These efficiencies, in turn, yield
+ log2 1 + SNR
+K ( ; SNR; fSIRk g) = log
+ log2
+ (1 0 2 ) log2 (e)
!1 C 2 (1 + SINR) (27)

(23) which has the exact same form of the asymptotic noise-limited capacity
obtained in (11) for ! 1. Hence, as the number of interfering an-
with 1 and 2 the positive solutions to tennas grows much larger than the number of receive antennas, the pro-
K gressively fine color of the interference cannot be discerned and it thus
SNR 1 SNR 1 appears white to the receiver. The capacity depends only on the total
1 + + =1
SNR  + 1 SNR  + SIRk
k=1 impairment power, irrespective of how it breaks down into noise and
outside interference.
SNR 2 On the other hand, for diminishing and finite K
2 + = 1: (24)
1 = 1 + O( )
Because of the implicit nature of these equations, obtaining explicit
2 = 1 + O( ) (28)
expressions therefrom requires solving for 1 and 2 in equations of
order K + 2 and K + 1, respectively. Hence, the complexity of the
confirming that the penalty due to a fixed number of interfering
solution is directly determined by the number of outside interferers.
antennas vanishes as the number of receive antennas grows without
Fortunately, the equations containing 1 and 2 are decoupled and thus, +K behaves as
bound. With that, C
as we shall see, it is possible to obtain meaningful expressions for a
large number of cases. If the number of outside interferers is set to
C +K ( ; SNR; fSIRk g) = log SNR + O( )
K = 0, it is rather straightforward to solve for 1 and 2 as 2 (29)

1 = 1 0 F SNR
; and becomes determined only by the underlying AWGN, irrespective
SNR of the SIR.
2 = 1 (25) Finally, as the SNR grows

and, substituting them into (23), obtain the asymptotic noise-limited  = [1 0 (K + 1) ]+

SNR!1 1
capacity per receive antenna of (8). Therefore, the above solution is a
generalization of the one presented in Section III.  = [1 0 K ]+ :
SNR!1 2
lim (30)
The roles of 1 and 2 can be interpreted by relating them to the
multiuser efficiency, a quantity commonly used in multiuser detection
In this regime, the receiver operates in zero-forcing mode against the
problems [19], [43]. Specifically, the following holds.
interference and thus the loss in 1 is given exactly by the total number
• 1 represents the asymptotic ratio between i) the SNR at the of transmit antennas (desired plus outside interference) per receive an-
output of a linear minimum mean-square error (MMSE) receiver tenna. If that total number of transmit antennas exceeds the number of
detecting the signal transmitted from any of the desired antennas receive antennas, then 1 = 0. The loss in 2 , on the other hand, is
in the presence of multiple-antenna as well as outside interfer- exactly the total number of outside interfering antennas per receive an-
ence, and ii) the SNR at the output of a matched filter detecting tenna. Again, if the number of outside interfering antennas exceeds the
that same antenna without any interference. Hence, it quantifies number of receive antennas, then 2 = 0.

Fig. 3. Asymptotic capacity per receive antenna as a function of for K = 1 and various combinations of SIR and SNR that correspond to SINR = 10 dB.
Homogeneous system.

2) Low-SNR Behavior: The behavior of the asymptotic efficiencies At high SNR, the asymptotic efficiencies can be found to be
at low SNR is given by O 1
; > 1

1 = 1 0 SNR
+ O (SNR )
SIR 1 = 1
+O 1
; = 1

2 = 1 0
+ O (SNR ):
1 0 2 + O 1
; < 1
SIR and
Analogous to ! 1, a low SNR renders both efficiencies indepen- (
01)SNR + O
; >1
dent of the composition of the outside interference; only the aggregate
SIR becomes relevant. With the above efficiencies, C +K
becomes 2 = SIR
+O 1
; =1 (36)
C +K
( ; SNR; fSIRk g) = log2 (e) SNR2 1 0 +O 1
; <1

1 0 2 1 + + O(SNR ): (32)
1 1 1 3
with the first-order coefficient of 1 for > a convoluted function 1
of and SIR. For = 1, this coefficient admits a simple form
3) Interference-Limited Behavior: Single Outside Interferer: p
High-capacity wireless systems are typically designed to operate in in- 1 = +O : (37)
terference-limited conditions [53]. In the remainder, we concentrate, SNR SNR
therefore, on studying the asymptotic capacity at high SNR. We begin In terms of asymptotic capacity, the above efficiencies yield
by evaluating the capacity in the presence of a single outside interferer. log2 SNR 102 0 (1 0 )
With K = 1, the asymptotic capacity becomes e
1 1 2 0 +O 1
; < 1
C +1
( ; SNR; SIR)
log2 1 0 SNR 2

log2 SNR
; = 1
SIR + SNR  1 C +1
2 e 2
= log2 + log2 1 + SNR
0 ) log
SIR + SNR  + O (1); < <1
(1 2 e

2 2 log2 1+ SIR +O pSNR

; =1
+ log2
+ (1 0  ) log (e)
2 2 (33)
log2 (1 + SIR) + O 1
; ! 1.
1 + + =1
SNR  + 1 SNR  + SIR We observe the following.
• As the SNR grows, C +1
becomes independent of the SIR for
2 + =1 (34)  12 . Hence, while the combined number of desired plus out-
SNR  + SIR side transmit antennas does not exceed the number of receive an-
which requires solving cubic and quadratic equations for 1 and 2 , tennas, the receiver approaches capacity by suppressing—purely
respectively. Fig. 3 depicts C +1
as function of for various levels through linear processing—the outside interference while simul-
of SNR and SIR corresponding to SINR = 10 dB. taneously detecting the desired signals. Once that interference

Fig. 4. Asymptotic noise- and interference-limited capacity per receive antenna with = 1 and K = 1 as a function of SINR. Homogeneous system.
is suppressed, capacity becomes limited only by the underlying corresponding, to first order, to the equivalent of a 4.34-dB im-
noise. provement in SINR. Hence, as illustrated in Fig. 4, operating over
• For 12  < 1, the number of receive antennas exceeds the
colored impairment is noticeably beneficial even when the re-
ceiver has no spare degrees of freedom. Interestingly, the advan-
number of desired antennas, but not the combined number of
tage appears to hold throughout the entire range of SINR levels.
desired plus outside interfering antennas. Therefore, the receiver
must compromise between assigning its degrees of freedom to • Unlike in the noise-limited case, the interference-limited
interference suppression and to signal detection. As the SNR C +1 does not increase monotonically with . It grows with +1
increases without bound, such compromise favors sacrificing a up to = 12 and it diminishes thereafter. Hence, C
fraction of the receive antennas for interference suppression. is maximized—in these conditions—by having the number of
For = 12 , the capacity becomes exactly half that of an transmit antennas equal half the number of receive antennas just
architecture with = 1 operating, free of outside interference, so the receiver has enough spare antennas to suppress the outside
at the same SNR. That is, interference in its entirety while decoding the desired signals.
Notice that this holds only if the SNR is sufficiently large with
C +1 1
; SNR; SIR 1C (1; SNR):
SNR!1 SNR!1 2
lim = lim respect to the SIR (Fig. 3), specifically—from (16) and (40)—if
(39) SNR p 2
> 1+ SIR : (42)
• For  1, the dependence on the SNR becomes weak. Once the e
number of outside interfering antennas has exceeded the number If this condition is not met, the underlying noise influences the
of receive antennas, the receiver can no longer suppress the to- capacity sufficiently to render it monotonic in . Hence, (42) can
tality of that interference and thus even the high-SNR perfor- be used as a criterion to assess whether the system is effectively
mance is determined mostly by the SIR. Nonetheless, the col- interference-limited when the interference arises from a single
ored interference renders the capacity higher than in equivalent outside transmitter.
noise-limited conditions. Only as ! 1 does this advantage In the presence of colored interference, the capacity clearly depends
dissipate (Fig. 3). not only on the number of transmit and receive antennas and the SINR,
• Particularly revealing is C +1 at = 1 which, in interference- but also on the degrees of freedom of the interference. Specifically, a
limited high-SIR conditions, becomes given amount of interference is more benign if it occupies fewer de-
grees of freedom.
C +1 (1; SNR; SIR)
4) Interference-Limited Behavior: Multiple Outside Interferers: In
the presence of K > 1 outside interferers, the high-SNR asymptotic
= log2 SIR + O p1 +O p1 (40)
efficiencies of the previous section generalize to
1 ;
O SNR > K1+1
indicating an advantage over its high-SNR noise-limited counter-
part—see (16)—of
+1 (1; SNR; SIR) 0 C 1+ SIR
C (1; SINR) 1 = 1 +O
1 = K1+1
K +1 SNR SNR ;
= log2 e + O p1 +O p1 (41)
0 (K + 1) + O 1 < K1+1

O 1
; > K1

2 = SIR
+O 1
; = K1
1 0 K + O 1
; < K1 .
Unfortunately, the first-order coefficients for both 1 at > K1+1 and
2 at > K1 do not amend themselves to simplification and thus we
obtain a more restricted set of explicit expressions for the asymptotic

log2 SNR 10(K +1)

0(1 0 K ) log 0 0KK 2
( +1)

+O 1
; < K1+1

C + K= 1
log2 SNR
+ K1+1
1+ SIR

Fig. 5. Simplified downlink scenario with two outside interferers at SIR =

1+ SIR
5 dB and SIR = 8 dB, respectively.

1 log 2 e +O pSNR
; = K1+1
log2 (1 + SIR) + O 1
; ! 1. indicating that an increasing number of equal-strength outside inter-
ferers will progressively whiten the impairment. Hence, C can also
For other values of , obtaining tractable expressions does not appear serve as a lower bound, but it is never as tight as C +K
for finite K .
feasible. However, it is possible to show that the multiple-interferer An example corresponding to a simple downlink scenario with two
capacity can always be bounded by outside interferers is illustrated in Fig. 5. A multiantenna terminal is
C +K
C +K
C +K
illuminated by its serving base as well as two nearby interfering bases,
all of them equipped with the same number of transmit antennas. Ac-
with the bounds obtained by replacing the K outside interferer by a counting for different range and shadow fading to every base, a typical
single “equivalent” interferer generating the same exact aggregate SIR set of SIR and SNR levels are chosen. The corresponding asymptotic
with the highest and lowest possible levels of structure. capacity appears in Fig. 6 along with its upper and lower bounds. Also
• The upper bound corresponds to concentrating the aggregate in- shown is the corresponding noise-limited capacity at the same aggre-
terference contribution of all K outside interferers into a single gate SINR.
one and thus Note how neglecting the fact that the interference is colored can
lead to gross miscalculations of the actual capacity or, conversely, of
C +K
( ; SNR; SIR) = C +1
( ; SNR; SIR): (46) the SINR required to achieve a certain level of capacity. This con-
• The lower bound corresponds to forcing the K outside interferer verse look is provided in Fig. 7, which displays curves of constant
capacity on a SNR–SIR plane. Specifically, the curves shown corre-
spond to C
to be of equal strength while preserving the aggregate SIR. Since +K
= 2.5 b/s/Hz per receive antenna with = 0:5,
K equal-power interferers are equivalent to a single interferer
with K times as many antennas, C +K
can be derived from = 1, and = 2. For each configuration, the SNR–SIR combinations
the Appendix to be that achieve such capacity are depicted as a function of the number of
equal-strength outside interferers.

C +K
( ; SNR; SIR) = K log2 
SIR + SNR K C. How Many Antennas Should be Used?
1 It was shown in Section III that, in noise-limited conditions, there is
+ log2 1+ SNR little incentive in pushing beyond unity. From the results presented

throughout this section, it is clear that the presence of interference can
+ (1 0  ) log (e)
2 2 (47) only lessen that incentive, at least in homogeneous systems wherein all
transmitters are equipped with the same number of active antennas. In
with such conditions, the presence of a dominant outside interferer can ac-
SNR 1 SNR 1 tually render the capacity for > 1 smaller than for = 1 (Fig. 3).
1 + +  + SIR = 1
SNR  + 1 SNR K On the other hand, the capacity with K outside interferers increases
monotonically for < K1+1 . The best choice for will thus usually
SNR 2 lie somewhere within [ K1+1 ; 1] depending on the structure of the inter-
2 + = 1: (48)
SNR  + SIR ference and, therefore, on the geometry and propagation environment
In general, the tightness of both bounds depend on the specific set of the corresponding system. The following applies.
of fSIRk g. The upper bound becomes tight in the presence of a domi-
nating interferer whereas the lower bound, in turn, tightens in the pres- • If each transmitter is interfered by a small number of dominant
ence of comparable-strength interferers. Furthermore outside interferers, the best will lie around K1+1 with no ad-
vantage—or even possibly a loss—in pushing it any further. A
K !1
C +K
( ; SNR; SIR) = C ( ; SINR) (49) similar conclusion is reached in [54].

Fig. 6. Asymptotic capacity per receive antenna (solid line) as a function of with K 2, SIR = 5 dB, SIR = 8 dB, and SNR = 12 dB. Upper and lower
bounds (dashed lines) and the corresponding noise-limited capacity at the same aggregate SINR (circles) are also shown. Homogeneous system.

Fig. 7. Combination of SNR and SIR levels required to attain an asymptotic capacity of C = 2.5 b/s/Hz per receive antenna with = 0:5, = 1, and
= 2 as a function of the number of equal-strength outside interferers. Homogeneous system.

• With a large number of comparable-strength outside interferers, V. SUMMARY

the best will lie close to = 1.
The main result of the correspondence is the asymptotic capacity of
As in the noise-limited case, practical considerations preclude multiple-transmit multiple-receive antenna architectures impaired by
making so small that AWGN as well as spatially colored interference. From the asymptotic
capacity, we gathered insight on how the capacity behaves. Although
C +K the derivation of the asymptotic capacity requires driving the number
>R (50) of transmit and/or receive antennas to infinity, the asymptotic solution
applies to virtually any number of antennas under ergodic conditions.
While in noise-limited conditions, the asymptotic capacity depends
with R the highest realizable rate per transmit antenna. only on the number of antennas and the SNR, the capacity in the pres-

ence of interference depends also on the color thereof. At any given into our problem with a simple SNR normalization to yield the fol-
SINR, the capacity increases with such color. The lowest capacity at lowing asymptotic results:
any SINR is attained in the presence of AWGN exclusively.
C1 = + k E log2 1 + SNR

The starting point for our derivation is (21), which can be manipu- + log2
+ (1 0 1) log2 (e)
lated into
C =E log2 det IN + SNR HHy C2 = k E log2 1 + SNR
M k
Hk Hky + log2
+ (2 0 1) log2 (e) (54)
K where the expectation is with respect to the nonnegative random vari-
0 log2 det IN + SNR
Hk Hky : ables B1 and B2 whose distribution is identical to the asymptotic em-
pirical distribution of the diagonal elements of 1 and 2 , respec- B B
tively. The efficiencies 1 and 2 are the solutions to
Defining some block matrices
H2 = [ H1 H2 1 1 1 HK ] 1 + + k E
SNR1 B1 +

H1 = [ H H2 ] (52)

and 2 + k E = 1: (55)
SNR2 B2 +
SIR IM O 111 0 k
With (54) and (55), we can now calculate the asymptotic capacity per
receive antenna, in its most general form, as C +K = C1 –C2 .
B2 = 0 M
SIR IM 111 0 We note here that the above result holds even if the columns of as H
. 0 H
well as of k (k = 1; . . . ; K ) are correlated, that is, if the transmit

0 0 IM
arrays of the user of interest and the interferers exhibit antenna corre-
111 M
lation. In this case, the matrices 1 and 2 would be given by
IM O SIR 21 O 111 0
B1 = 0 B2

B2 = 0 M SIR 22 1 1 1
M 0
the capacity can be further manipulated into C = C1 –C2 given ..
. 0
C1 = E log2 det IN + SNR
H1 B1H1 y
0 0 111 M
C2 = E log2 det IN + SNR H B Hy
M 2 2 2
(53) B1 = 20 BO
with 1 and 2 diagonal matrices. Notice that—provided the desired 2
where the M 2 M matrix contains the transmit antenna correlation
user and the various outside interferers are mutually independent—the at the desired user and the Mk 2 Mk matrices k contain the transmit 2
block matrices H 1 and H 2 are composed by independent and identi- antenna correlations at each of the interferers. The expectation in (54)
cally distributed entries if only the individual and k matrices are.H and (55) would, in this case, be with respect to the asymptotic empirical
Both C1 and C2 have meaningful interpretations. B B
distribution of the eigenvalues of 1 and 2 If, on the other hand, we
wanted to take into account antenna correlation at the receiver as well,
• C1 corresponds to the capacity of the entire set of desired plus
we would have to resort to more sophisticated tools [55], but that is
outside transmit antennas and the receiver. Hence, it regards the
beyond the scope of this correspondence.
outside interfering antennas as additional desired signals that
In most instances, the number of transmit antennas is the same for the
could be decoded with proper knowledge of their corresponding
user of interest as well as for every outside interferer (that is, k =
channel matrices.
8 k), in which case C +K simplifies to
• C2 corresponds to the capacity of the outside transmit antennas
K SIRk + SNR  1
and the receiver, excluding the desired user transmit antennas. C = log2
+ log2 1+ SNR

The difference between both terms yields the actual capacity for the
desired user in the presence of outside interferers of which only the + log2
+ (1 0 2 ) log2 (e) (56)
aggregate covariance is known.
In order to obtain the asymptotic value of C1 and C2 as the dimen- given
sions of H 1 and H 2 are driven to infinity, we apply [19, Theorem IV.1]. K
With this result, derived within the context of randomly spread CDMA 1 + + =1
SNR  + 1 SNR  + SIRk
in fading channels, Shamai and Verdú obtained the asymptotic capacity k=1
of an optimum receiver using the Tse–Hanly equation [18] whose solu- K
tion is the asymptotic SNR at the output of an MMSE receiver. Defining 2 + = 1:
C1 = CN , C2 = CN , and k = MN , the above theorem can be mapped

REFERENCES [26] T. L. Marzetta, “BLAST training: Estimating channel characteristics for

high capacity space–time wireless,” in Proc. 37th Annu. Allerton Conf.
Communication, Control, Computing, Monticello, IL, Sept. 1999.
[1] G. J. Foschini and M. J. Gans, “On the limits of wireless communica- [27] B. Hassibi and B. M. Hochwald, “How much training is needed in mul-
tions in a fading environment when using multiple antennas,” Wireless tiple-antenna wireless links?,” IEEE Trans. Inform. Theory, submitted
Personal Commun., pp. 315–335, 1998.
for publication.
[2] I. E. Telatar, “Capacity of multi-antenna Gaussian channels,” Europ.
[28] Q. Sun, D. C. Cox, A. Lozano, and H. C. Huang, “Training-based
Trans. Telecommun., vol. 10, pp. 585–595, Nov. 1999.
channel estimation for continuous flat-fading BLAST,” in Proc. IEEE
[3] G. Raleigh and J. M. Cioffi, “Spatio–temporal coding for wireless com-
Int. Conf. Communications (ICC’02), New York, Apr. 2002.
munications,” IEEE Trans. Commun., vol. 46, pp. 357–366, Mar. 1998.
[4] A. Lozano, F. R. Farrokhi, and R. A. Valenzuela, “Lifting the limits in [29] T. M. Cover and J. A. Thomas, Elements of Information Theory. New
high-speed wireless data access using antenna arrays,” IEEE Commun. York: Wiley, 1990.
Mag., vol. 39, pp. 156–162, Sept. 2001. [30] H. Bolcskei, D. Gesbert, and A. J. Paulraj, “On the capacity of OFDM-
[5] G. J. Foschini, “Layered space–time architecture for wireless commu- based multi-antenna systems,” in Proc. IEEE ICASSP’2000, Istanbul,
nications in a fading environment when using multi-element antennas,” Turkey, 2000, pp. 2569–2572.
Bell Labs. Tech. J., pp. 41–59, 1996. [31] D. Jonsson, “Some limit theorems for the eigenvalues of a sample co-
[6] G. J. Foschini, G. D. Golden, R. A. Valenzuela, and P. W. Wolnianski, variance matrix,” J. Multivariate Anal., vol. 12, pp. 1–38, 1982.
“Simplified processing for high spectral efficiency wireless communica- [32] P. Rapajic and D. Popescu, “Information capacity of a random signature
tion employing multi-element arrays,” IEEE J. Select. Areas Commun., multiple-input multiple-output channel,” IEEE Trans. Commun., vol. 48,
vol. 17, pp. 1841–1852, Nov. 1999. pp. 1245–1248, Aug. 2000.
[7] A. Lozano and C. Papadias, “Layered space–time receivers for fre- [33] S. M. Alamouti, “A simple transmit diversity technique for wireless
quency-selective wireless channels,” IEEE Trans. Commun., vol. 50, communications,” IEEE J. Select. Areas Commun., vol. 16, pp.
pp. 65–73, Jan. 2002. 1451–1458, Oct. 1998.
[8] G. G. Raleigh and V. K. Jones, “Multivariate modulation and coding for [34] A. Wittneben, “A new bandwidth efficient transmit antenna modulation
wireless communications,” IEEE J. Select. Areas Commun., vol. 15, pp. diversity scheme for linear digital modulation,” in Proc. IEEE Int. Conf.
851–866, May 1999. Commun. (ICC’93), vol. 3, 1993, pp. 1630–1634.
[9] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space–time codes for [35] N. Seshadri and J. H. Winters, “Two schemes for improving the per-
high data rate wireless communications: Performance criterion and code formance of Frequency-Division Duplex (FDD) transmission systems
construction,” IEEE Trans. Inform. Theory, vol. 44, pp. 744–765, Mar. using transmitter antenna diversity,” Int. J. Wireless Inform. Networks,
1998. vol. 1, pp. 49–60, Jan. 1994.
[10] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space–time block [36] S. Verdú, “Spectral efficiency in the wideband regime,” IEEE Trans.
codes from orthogonal designs,” IEEE Trans. Inform. Theory, vol. 45, Inform. Theory, vol. 48, pp. 1319–1343, June 2002.
pp. 1456–1467, July 1999. [37] A. Lozano, A. M. Tulino, and S. Verdú, “Capacity of multi-antenna
[11] F. Rashid-Farrokhi, G. J. Foschini, A. Lozano, and R. A. Valenzuela, channels in the low-power regime,” in Proc. Information Theory Work-
“Link-optimal space–time processing with multiple transmit and receive shop (ITW’02), Bangalore, India, Oct. 2002.
antennas,” IEEE Commun. Lett., vol. 5, pp. 85–87, Mar. 2001. [38] A. Furuskar, S. Mazur, F. Muller, and H. Olofsson, “EDGE: Enhanced
[12] E. Biglieri, G. Caire, and G. Taricco, “Limiting performance of block- Data Rates for GSM and TDMA/136 evolution,” IEEE Personal
fading channels with multiple antennas,” IEEE Trans. Inform. Theory, Commun. Mag., vol. 6, pp. 56–66, June 1999.
vol. 47, pp. 1273–1289, May 2001. [39] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushayana, and A.
[13] S. L. Ariyavisitakul, “Turbo space–time processing to improve wireless Viterbi, “CDMA/HDR: A bandwidth-efficient high-speed wireless data
channel capacity,” IEEE Trans. Commun., vol. 48, pp. 1347–1358, Aug. service for nomadic users,” IEEE Commun. Mag. , vol. 38, pp. 70–77,
2000. July 2000.
[14] C. Chuah, D. Tse, J. Kahn, and R. Valenzuela, “Capacity scaling in dual- [40] A. Jalali, R. Padovani, and R. Pankaj, “Data throughput of CDMA-HDR:
antenna-array wireless systems,” IEEE Trans. Inform. Theory, vol. 48, a high efficiency-high data rate personal communication wireless
pp. 637–650, Mar. 2002. system,” in Proc. IEEE Vehicular Technology Conf. (VTC’2000), May
[15] J. B. Andersen, “Array gain and capacity for known random channels 2000.
with multiple element arrays at both ends,” IEEE J. Select. Areas [41] “3rd Generation Partnership Project; Technical Specification Group
Commun., vol. 18, pp. 2172–2178, Nov. 2000. Radio Access Network; UTRA High Speed Downlink Packet Access,”
[16] A. Lozano, “Capacity-approaching rate function for layered multi-an- Tech. Rep., 3G TR 25.950, Mar. 2001.
tenna architectures,” IEEE Trans. Wireless Commun., to be published. [42] A. Klein, “Data detection algorithms specially designed for the downlink
[17] S. Verdú and S. Shamai (Shitz), “Spectral efficiency of CDMA with of CDMA mobile radio systems,” in IEEE Vehicular Technology Conf.
random spreading,” IEEE Trans. Inform. Theory, vol. 45, pp. 622–640,
(VTC’97), Phoenix, AZ, 1997, pp. 203–207.
Mar. 1999.
[43] S. Verdú, Multiuser Detection. Cambridge, U.K.: Cambridge Univ.
[18] D. N. C. Tse and S. Hanly, “Linear multiuser receivers: Effective in-
Press, 1998.
terference, effective bandwidth and user capacity,” IEE Trans. Inform.
[44] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. J. Viterbi, L. A. Weaver,
Theory, vol. 45, pp. 641–657, Mar. 1999.
and C. E. Wheatley, “On the capacity of a cellular CDMA system,” IEEE
[19] S. Shamai (Shitz) and S. Verdú, “The effect of frequency-flat fading on
the spectral efficiency of CDMA,” IEEE Trans. Inform. Theory, vol. 47, Trans. Veh. Technol., vol. 40, pp. 303–312, May 1991.
May 2001. [45] J. H. Winters, “Optimum combining in digital mobile radio with
[20] T.-S. Chu and L. J. Greenstein, “A semiempirical representation of an- cochannel interference,” IEEE J. Select. Areas Commun., vol. 2, pp.
tenna diversity gain at cellular and PCS base stations,” IEEE Trans. 528–539, July 1984.
Commun., vol. 45, pp. 644–646, June 1997. [46] J. H. Winters, J. Salz, and R. D. Gitlin, “The impact of antenna diver-
[21] S. B. Rhee and G. I. Zysman, “Results of suburban base station diversity sity on the capacity of wireless communication systems,” IEEE Trans.
measurements in the UHF band,” IEEE Trans. Commun., vol. 22, pp. Commun., vol. 42, no. 2/3/4, pp. 1740–1751, Feb./Mar./Apr. 1994.
1630–1636, Oct. 1974. [47] A. Shah and A. M. Haimovich, “Performance analysis of optimum com-
[22] D. Chizhik, F. R. Farrokhi, J. Ling, and A. Lozano, “Effect of antenna bining in wireless communications with Rayleigh fading and cochannel
separation on the capacity of BLAST in correlated channels,” IEEE interference,” IEEE Trans. Commun., vol. 46, pp. 473–479, Apr. 1998.
Commun. Lett., vol. 4, pp. 337–339, Nov. 2000. [48] V. Tarokh, A. Naguib, N. Seshadri, and A. R. Calderbank, “Combined
[23] D. Gesbert, H. Bolcskei, D. Gore, and A. J. Paulraj, “MIMO wireless array processing and space–time coding,” IEEE Trans. Inform. Theory,
channels: Capacity and performance prediction,” in Proc. IEEE vol. 45, pp. 1121–1128, May 1999.
GLOBECOM’2000, San Francisco, CA, Dec. 2000. [49] A. Naguib, N. Seshadri, and A. R. Calderbank, “Space–time block
[24] C. C. Martin, J. H. Winters, and N. R. Sollenberger, “Multiple-input and interference suppression for high capacity wireless systems,” in
multiple-output (MIMO) radio channel measurements,” in Proc. IEEE Proc. Int. Conf. Personal, Indoor and Mobile Radio Communications
Vehicular Technology Conf. (VTC’2000), Boston, MA, Sept. 2000. (PIMRC’99), vol. A7, 1999.
[25] D. Chizhik, G. J. Foschini, and R. A. Valenzuela, “Capacities of multi- [50] B. Lu and X. Wang, “Iterative receivers for multiuser space–time coding
element transmit and receive antennas: Correlations and keyholes,” IEE systems,” IEEE J. Select. Areas Commun., vol. 18, pp. 2322–2335, Nov.
Electron. Lett., pp. 1099–1100, June 2000. 2000.

[51] R. S. Blum, J. H. Winters, and N. R. Sollenberger, “On the capacity allows for an adaptive implementation [3]. Hence, the MMSE detector
of cellular systems with MIMO,” in IEEE Vehicular Techn.ology Conf. has been a subject of considerable study.
(VTC’2001 Fall), Atlantic City, NJ, Oct. 2001.
Detailed performance analysis for the MMSE detector was first con-
[52] B. M. Zaidel, S. Shamai, and S. Verdú, “Spectral efficiency of randomly
spread DS-CDMA in a multi-cell model,” in Proc. 37th Annu. Allerton sidered in [4]. The spreading sequences were assumed to be arbitrary
Conf. Communication, Control, Computing, Monticello, IL, Sept. 1999. but fixed, and the Gaussianity of the multiaccess interference at the
[53] D. C. Cox, “Universal digital portable radio communications,” Proc. output of the detector was analyzed under various asymptotic scenarios.
IEEE, vol. 75, pp. 436–477, Apr. 1987. A more promising approach for analysis was introduced in [5], [6].
[54] R. S. Blum, “MIMO capacity with interference,” in Proc. Conf. Infor-
mation Sciences and Systems (CISS’02), Mar. 2002. Here, the spreading sequences were treated as independent random
[55] V. L. Girko, Theory of Random Determinants. Boston, MA: Kluwer vectors, and limits of the SIR and capacity were studied as the number
Academic, 1990. of users (K ) and the processing gain (N ) tend to infinity with the ratio
K=N approaching a constant. The limitation of the analysis in [5], [6]
is that it is restricted to the situation where the users are symbol-syn-
chronous. In [7], the SIR analysis of [5] was extended to the case where
the users are symbol-asynchronous but chip-synchronous, i.e., the de-
lays of all the users are aligned to the chip timing.
MMSE Detection in Asynchronous CDMA Systems: An While it allows for accurate large-system analysis, the synchronous
Equivalence Result or chip-synchronous assumption is not realistic for the received signal
on the reverse link of a cellular CDMA system, especially with user
Ashok Mantravadi, Student Member, IEEE, and mobility and the resulting variations in the delay. Thus, we would
Venugopal V. Veeravalli, Senior Member, IEEE like to allow the users to be completely asynchronous, i.e., symbol-
as well as chip-asynchronous. Analysis of the MMSE detector with
random spreading sequences and completely asynchronous users was
Abstract—The analysis of linear minimum mean-square error (MMSE)
considered in [8]. However, the performance measure was the average
detection in a band-limited code-division multiple-access (CDMA) system
that employs random spreading sequences is considered. The key features near–far resistance of the detector and bounds were obtained on this
of the analysis are that the users are allowed to be completely asynchronous, quantity for finite K and N . Furthermore, the analysis relied on the
and that the chip waveform is assumed to be the ideal Nyquist sinc function. assumption that the chip waveform was limited to a chip interval.
It is shown that the asymptotic signal-to-interference ratio (SIR) at the de- In this correspondence, we allow the users to be completely asyn-
tector output is the same as that in an equivalent chip-synchronous system.
It is hence been established that synchronous analyses of linear MMSE de- chronous and consider SIR at the detector output as the performance
tection can provide useful guidelines for the performance in asynchronous metric. We also assume that the system employs the ideal band-limited
band-limited systems. (and hence, of infinite duration) sinc chip waveform. For single-user
Index Terms—Asymptotic analysis, asynchronous systems, band-limited narrow-band systems, the sinc waveform maximizes the signaling rate
communication, code-division multiple access (CDMA), least mean squares when the symbol waveforms are constrained to have a given bandwidth
methods, matched filters (MFs), minimum mean-square error (MMSE) de- and to have no intersymbol interference [9]. In spread-spectrum sys-
tection, sinc function. tems, we have an additional degree of freedom, since the processing
gain of the system can be varied with the excess bandwidth of the
chip waveform to keep the symbol rate and occupied bandwidth fixed.
In such a framework, the sinc waveform maximizes the processing
Multiuser detection in code-division multiple-access (CDMA) sys- gain since it has zero excess bandwidth. For the matched-filter (MF)
tems has been a topic of intense research for more than a decade [1]. detector, the maximum processing gain also results in the maximum
Several criteria have been used for designing multiuser detectors, and output SIR across all waveforms [10], [11]. Hence, practical CDMA
a particularly appealing one is to minimize the mean-squared error systems (e.g., [12]) employ waveforms that have an approximately flat
(MSE) of the symbol estimates at the output of the detector. When the spectrum over the band of operation. Similar observations hold for the
detector is further constrained to be linear we obtain the linear min- MMSE detector as well, although a formal proof of the optimality of the
imum mean-squared error (LMMSE or simply, MMSE) detector [2]. sinc waveform appears to be open [13]. Based on the above remarks,
Equivalently, the MMSE detector also maximizes the output signal-to- the sinc waveform can be considered to be a benchmark for band-lim-
interference ratio (SIR) over the class of linear detectors. In addition, it ited systems. Hence, analysis of the MMSE detector when the users are
completely asynchronous and employ the sinc waveform is of much in-
terest, from a theoretical as well as a practical viewpoint.

Manuscript received August 14, 2001; revised July 22, 2002. This
work was supported by the National Science Foundation under Grant II. SYSTEM MODEL AND MF DETECTION
CCR-9980616, through a subcontract with Cornell University, and by the NSF
We consider a direct-sequence CDMA (DS/CDMA) model with K +
CAREER/PECASE Award CCR-0049089. The material in this correspondence
was presented in part at the IEEE International Symposium on Information 1 users, where the received complex baseband signal is given by
Theory, Washington, DC, June 2001.
0 k Tc )ei 2 [01; 1]
A. Mantravadi was with the School of Electrical Engineering, Cornell Uni-
versity, Ithaca, NY 14853 USA. He is now with Qualcomm, Inc., San Diego, r(t) = sk (t + w (t); t (1)
CA (e-mail: k=0
V. V. Veeravalli is with the Department of Electrical and Computer Engi-
neering and the Coordinated Science Laboratory, University of Illinois at Ur- where sk (t) is the signal transmitted by user k
bana-Champaign, 128 Computer and Systems Research Laboratory, Urbana, IL
61801 USA (e-mail: 1 p
Communicated by D. N. C. Tse, Associate Editor for Communications. sk (t) = Ek bkm ckm (t):
( ) ( )
Digital Object Identifier 10.1109/TIT.2002.805078 m=01
0018-9448/02$17.00 © 2002 IEEE