Professional Documents
Culture Documents
Dongning Guo
Dept. of Electrical Engineering
Princeton University
Princeton, NJ 08544, U.S.A.
Sergio Verdu
Dept. of Electrical Engineering
Princeton University
Princeton, NJ 08544, U.S.A.
1. Introduction
Code-division multiple access (CDMA) is becoming the ubiquitous
air-interface in future generation communication systems. In a CDMA
V. K. Bhargava et al. (eds.), Communications, Information and Network Security
© Springer Science+Business Media New York 2003
230 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
system, all frequency and time resources are allocated to all users simul-
taneously. To distinguish between users, each user is assigned a user-
specific spreading code on which the user's information symbol is mod-
ulated before transmission. By selecting mutually orthogonal spreading
codes for all users, each user can be separated completely by matched
filtering to one's spreading code. It is not very realistic to maintain
orthogonality in a mobile environment and hence multiple access inter-
ference (MAl) arises. The problem of demodulating in the presence of
the MAl becomes vital for a CDMA system.
A variety of multiuser detectors [1] have been proposed to mitigate the
MAL The simplest one is the single-user matched filter, which totally
ignores the existence of the MAL Its performance is not very satisfac-
tory and is particularly limited by the near-far problem. In the other
extreme, the individually optimal (10) and the jointly optimal (JO) de-
tectors achieve the minimum probability of error but entail prohibitive
complexity which is exponential in the number of users. A wide spec-
trum of multiuser detectors offer performance in between the matched
filter and the optimal detectors with substantially reduced complexity.
The most popular ones include the MMSE detector and the decorrelat-
ing detector. The performance of multiuser detectors have been stud-
ied extensively in the literature. A collection of results is found in [1].
In general, the performance is dependent on the spreading factor, the
number of users, the transmit signal-to-noise ratios (SNR), and the in-
stantaneous spreading codes. The dependence on the large number of
parameters results in very complex expressions for all but the simplest
cases. Not only are these expressions hard to evaluate, but the com-
plication allows little useful insight into the problem. To eliminate the
dependency by averaging over all spreading codes (e.g. [2]) is plausible
but usually a prohibitive task.
Recently, it is found that performance analysis can be greatly sim-
plified for randomly spread systems the size of which tend to infin-
ity [1, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]. Of special interest is the case
where the number of users and the spreading factor both tend to infin-
ity with their ratio fixed. This is referred to as the large-system limit
in the literature. As far as linear multiuser detector is concerned, an
immediate advantage of the large-system setting is that the multiple-
access interference, as a sum of contributions from all interfering users,
becomes Gaussian-like in distribution under mild conditions as the num-
ber of users goes to infinity [13]. This allows the signal-to-interference
ratio (SIR) and the uncoded bit-error-rate (BER) to be easily character-
ized for linear detectors such as the matched filter, the MMSE detector
and the decorrelator. The large-system treatment also finds its success
Multiuser Detection and Statistical Mechanics 231
(13.1)
Multiuser Detection and Statistical Mechanics 233
(13.2)
(13.3)
where the N x K matrix S = [SI, ... ,SK]. Let r = [rI, ... ,rN]T, A =
diag( y'F{ , ... , VPK) and v = [VI, ... , VN ]T, we have a more compact
form,
and
(13.6)
each symbol interval and tries to recover the transmitted symbols us-
ing knowledge of the instantaneous spreading codes S. In general, the
detector outputs a soft decision statistic for each user of interest, which
is a function of (r, S),
ih = A(r, S), k E {I, ... , K}. (13.7)
Whenever the soft output can be separated as a useful signal component
and an interference, their energy ratio gives the SIR. Usually, a hard
decision is made according to the sign of the soft output,
(13.8)
Assuming binary symmetric priors, the bit-error-rate for user k is
(13.9)
(13.11)
s~r (13.12)
yIP; dk + 2)S~Si)JP: di + aOWk (13.13)
i#
Multiuser Detection and Statistical Mechanics 235
(13.14)
It can be shown using the central limit theorem that the MAl converges
to a Gaussian random variable in the large-system limit. Thus the BER3
IS
(13.15)
'1l(mf) _ _1_
./ - {3 • (13.16)
1+ -(72o
The performance of the matched filter is solely determined by the mul-
tiuser efficiency, since the SIR can be obtained as
and the BER by (13.15). Indeed, from every user's point of view, the
MAl has the same statistical property in the large-system limit. The
only difference in the users is their own energies. Due to normalization
with respect to one's own energy, the multiuser efficiency is the same for
every user.
(13.18)
~(m=') = 1 - ~ ( J( + /i3) + U6 - J
1 2 (1 _ /i3) 2 + U6 ) 2
(13.20)
f3TJ
"1+ 2=1. (13.21)
"I + 0'0
Equation (13.21) is generalized in [4] to an arbitrary energy distribution
using random matrix theory and the so-called Tse-Hanly equation is ob-
tained. In the large-system limit, with probability 1, the output decision
statistic given by (13.19) converges in distribution to a Gaussian random
variable [9, 13]. Hence the BER is determined by the SIR by a simple
expression similar to (13.15). The Tse-Hanly equation is distilled to the
following fixed-point equation for the multiuser efficiency in [10]
(13.23)
where
(dl S) ~ po(rld, S) (13.26)
Po r, - 2:dE{-I,I}K po(rld, S)Po(d)
The individually optimal detector maximizes the marginal posterior prob-
ability
(13.27)
where
po(dlr, S) (13.28)
dk E { -1,1}J(-1
where d k denotes the vector d with the kth element struck out. The
IO detector achieves the minimum possible BER among all multiuser
detectors.
The asymptotic (low-noise) multiuser efficiency of the optimal detec-
tors is shown to be 1 in [6]. For finite SNRs, the large-system perfor-
mance of the optimal detectors has been solved for the case where all
users' energies are the same [27, 29]. In parallel with the format of the
result in (13.22), we show that the multiuser efficiency of the individually
optimal detector is also the solution to a fixed-point equation
1
4)
ry- (13.29)
- 1+ 4- [1 -
----~--------------~--------~--~
(To V
k27r J e- z: tanh ( V14
(Til
z+
(To
dZ] .
The efficiency of the jointly optimal detector is also found in [27] but
omitted here. We solve the multiuser efficiency for an arbitrary energy
distribution for both the IO and the JO detectors in section 5.
p(d) p(rJd)
Figure 13.1. The Bayes retrochannel and the conditional mean estimator.
(13.30)
where SIR~I) is the user's output SIR. It can be easily justified by notic-
ing that these linear detectors output asymptotically Gaussian decision
statistics.
Without any constraint on the type of detector, the spectral efficiency
of the CDMA channel is [10]
f3 { log ( 1 +
C = -E
2
'f/
(mmse)
0"5
p) }+ 1
-log
2
1
'f/(mmse)
+ -21 ('f/(mmse) - 1)
(13.31)
where the expectation is taken over the received energy distribution due
to fading.
where the operator (-) gives the expectation taken over the distribution
p(dlr). Indeed, this estimator outputs the mean value of the random
estimate generated by the Bayes retrochannel. Clearly, this estimator
also depends on the same postulated prior and conditional distribution
as that of the Bayes retrochannel.
Interestingly, by tuning the postulated prior and conditional distribu-
tions, the conditional mean estimator arrives at many popular detectors.
For example, suppose K = N = 1 and the scalar prior is symmetric bi-
nary, Po(d = 1) = Po(d = -1) = !.
Let the postulated prior be the
same as the true prior, po(d) = p(d), V d, and that the postulated condi-
tional distribution be the true channel law, p(rld) = po(rld), Vr, d. The
conditional mean estimate is then
which may be different from the true binary symmetric prior distribution
Po(d). Specific choices are made in the following subsection. Let the
conditional distribution be
(13.36)
which differs from the true channel law po(rld, S) by a positive control
parameter a, where in case a = 0'0, po(rld, S) = p(rld, S). The posterior
probability distribution is then
(13.37)
where
The conditional mean estimator outputs the mean value of the pos-
terior distribution,
(13.39)
where the expectation is taken over p(dlr, S). We identify a few choices
for the prior distribution p( d) and the control parameter a for the condi-
tional distribution under which the conditional mean estimator becomes
equivalent to each of the detectors discussed in section 2.2.
(13.40)
where
8p(I)(dlr, S)
8d H = 0, (13.43)
which yields
(13.44)
where we use the subscript to denote an average over the posterior prob-
ability distribution with control parameter (j.
By choosing different values for (j, we arrive at different types of linear
detectors. If (j -+ 00, we get
(13.45)
Hence the conditional mean estimate is consistent in sign with the matched
filter output. If a = (jo, it is exactly the soft output of the MMSE re-
ceiver as given by (13.19). If a -+ 0, we get the soft output of the
decorrelator as given by (13.23). We have seen that the control param-
eter can be used to tune a parameterized conditional mean estimator to
the desired one in a set of detectors.
(13.46)
where
(13.48)
242 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
Clearly, this soft output is consistent in sign with the hard decision of
the 10 detector as given by (13.27).
Alternatively, if a --+ 0, most of the probability mass of the distribu-
tion p(o)(dlr, S) is concentrated on a vector d that achieves the minimum
of IIr-SAdll, which also maximizes the posterior distribution po(dlr, S).
The conditional mean estimator output
(13.49)
will single out the kth component of this d at the minimum. Therefore
by letting a --+ 0 the conditional mean estimator is equivalent to the JO
detector as given by (13.25).
Worth mentioning here is that if a --+ 00, the conditional mean es-
timator reduces to the matched filter. This can be easily verified by
noticing that
(13.51)
Multiuser Detection and Statistical Mechanics 243
(13.53)
where
1
F = E + <: 2:p(d) logp(d). (13.57)
d
244 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
(13.60)
(13.62)
(13.63)
and
(13.64)
Multiuser Detection and Statistical Mechanics 245
3.5 Overlap
Here we aKe interested in one of the most important performance
measures, the bit-error-rate. Due to symmetry of the spreading codes,
as far as the BER is concerned, we can assume that all transmitted
symbols are +1, i.e., dk = 1 for all k. Note that the spreading codes are
randomly chosen and not dependent on the received energies.
For the time being consider equal-energy users, i.e., Pk = 1 for all
k. Conditioned on (r, S) and the transmitted symbols being all equal
to 1, the percentage of erroneously detected bits in the current symbol
interval is ~[1 - M(r, S)] where
1 K 1 K
M(r,S) = KLdk = KLsgn(dk») (13.65)
k=l k=l
is called the overlap of the detector output. We can regard the overlap
as a macroscopic quantity of the spin glass system. Essentially it is the
magnetism of the system.
A fundamental principle in statistical mechanics states that the over-
lap, as a macroscopic quantity, is self-averaging. That is, with probabil-
ity 1, M(r, S) converges to its average over the probability distribution
of (r, S), denoted as «M(r, S)), as the system size K tends to infinity.
Here, «.» is defined explicitly as
where d is the all 1 vector and Es {.} denotes the average over the
subscript variable 5, which is the spreading codes in this case. Note
that the overlap
(13.67)
(13.70)
where
(13.71)
F=- lim
K-+oo \
//~IOgZ(r,s))).
\I< (13.73)
The second step is to evaluate the average in (13.74) only for integer
u. Finally, we take the formal limit as u -7 0 of the resulting expres-
sion. There are intensive ongoing efforts in the mathematics and physics
community to find a rigorous proof for the replica method.
248 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
Consider u independent replicas of the same spin glass with the same
quenched randomness (r, S). Statistically there is no difference among
these replicas. We can view this as u identical and independent Bayes
retrochannels in parallel. Let d a = [dab ... , daK]T denote the channel
outputs (or the spin glass configuration) of the ath replica. The posterior
probability distribution for the replicated system is
g{(211")-~
p({da}~=llr, S)
(13.77)
! Dzf(z) ~ Joo
-00 y
~e-z22 f(z), Vf
211"
(13.79)
Note that in order to average with respect to r, we need to use the true
data prior which in this case puts unit mass on the vector d = [1, ... ,1 r
Multiuser Detection and Statistical Mechanics 249
= J (IT IT Dd {Jdr
a=1 k=1
ak ) k ES { exp [- 2~2 (r - ~N
0 y
t
k=1
g
21T0'0
where the expectation in (13.82) is now taken over Sk'S which are inde-
pendent random chips. For convenience, we introduce dOk = dk = 1 for
k = 1, ... , J(. For a = 0, 1, ... , u, define a set of random variables
(13.83)
(13.84)
as I< --+ 00. Because of a generalization of the central limit theorem, the
variables {va}~=o converge to a set of jointly Gaussian random variables,
with zero mean and covariance matrix Q, where
(13.86)
Note that although not made explicit in the notation, Qab is a function
of {dak' dbdf=l' Trivially, Qoo = l.
Therefore the expectation over va's is dependent on the symbols {dad
only through Q in the large-system limit. We can separate the integral
in (13.85) into the symbol dependent part Go and the rest G l for further
calculations:
exp[I<Go(u)] = J (
uK
l11]Dd ak
)U'(L
~ 8 ~Pl ~l dakdbk - I<Qab
)
(13.88)
where 8(·) is the Dirac function and
where the minimum is over the set of possible Q's that can be obtained
by varying dak'S in (13.86). To find the extremum with respect to Qab'S
Multiuser Detection and Statistical Mechanics 251
E {VOva} = m (13.91a)
E {VaVb} = q (13.91b)
E {vn = p. (13.91c)
Vo (13.92a)
(13.92b)
Then
[kVo
2a02
ylvW +vlxY (13.94a)
Hence
exp[G 1 (u)]
= vk Jdr
211"a5
Ev {ex p [- ~
2ao
(r - ...jfivO)2]
g exp [ - 2!' (r - .n v.) ,] } (13.95)
where a = 1/{1+2x) and b = ua5J[(1 + 2y)a 2]. The integral in the right
hand side of (13.98) is
= (1 + 2x)
_1
2 (1 + 2y)
_~
2
[1+
1 2x
ua5 2ua5
+ (1 + 2y)a2 + (1 + 2x)(1 + 2y)a2
= (1+ 2yr "~' [1+ 2y + u(1+ 2x) :~ + 2u (:. v'v - Vw) T' 1
(13.100)
= ( 1 + af32 (p - q) ) - u-1
2
[ f3
1 + a2 (p - q) + :2 (a5 + f3 (1 - 2m + q))
] -!.
2.
(13.101)
Therefore,
-~log[1+
2
f3 (p-q)+
a2
U
a2
(a 5+ f3 (1-2m+ q))].
(13.102)
1 jjoo+t dQexp (
8(x)=-. )
Q.x, 'lit E IR, (13.103)
211") -joo+t
we have
(13.104)
(13.105)
254 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
(13.106)
where the integral over each Qab is from -joo +tab to joo + tab for some
real number tab, and
(13.108)
E (13.109a)
F (13.109b)
G
(13.109c)
2
Multiuser Detection and Statistical Mechanics 255
In this case,
XI(U)
= J l1 (
u )
Dda exp
[U U
PIE~da + PzF oEb 1 U]
dadb + 2PzG~d;
(13.111)
(13.112)
= Ez {
(1 + PI(F -
- U
G))-z exp
[u (PzE+ v!PzF zr] }
2(1 + Pz(F _ G)) (13.116)
= (1 + Pz(F - -
G))--2 (1 + (1 - u)PIF - PzG)-z
u-1 - - 1
exp [ UP?I E2 ]
(13.117)
2(1 + (1 - u)PIF - PIG) .
256 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
Meanwhile,
U I U U u
L QabQab =L QOaQoa + L QabQab +L QaaQaa (13.118)
a<b a=1 O<a<b a=1
u(u-1) u
= uEm+ 2 Fq+ "2 Gp . (13.119)
Therefore,
GO(u)
L U i
t,
Z=1 a~b
(13.122)
(13.123)
1
E (13.124a)
a2 + (3(p - q)
F
a5 + (3(1 - 2m + q) (13.124b)
[a 2 + (3(p - q)]2
G
(3(1 - 2m + 2q - p) + (a5 - ( 2 ) (13.124c)
[a 2 + (3(p - q))2
L P2 E
m
tt
~a
I
I
1 + PI (F - G)
(13.124d)
L P3E2 + PI2F
tt
q ~a I (13.124e)
l[l+PI(F-G)F
p (13.124f)
Immediately,
G F-E (13.125a)
p 1- m+q. (13.125b)
1
E (13.126a)
a2 + (3(1 - m)
L P2E
m
tt
~ al---,-I-=-=
l+PIE
(13.126b)
F
a5 + (3(1 - 2m + q)
(13.126c)
[a 2 + (3(1 - m)]2
(13.126d)
(13.127)
(13.128)
(13.129)
(13.130)
Multiuser Detection and Statistical Mechanics 259
Consider
(13.131)
(13.132)
(13.133)
lim ZU(r, S)
u-+o
= 1, (13.134)
we have
Similar to (13.85),
We can write
1
1(1 log ((ZU(r, Sj h)))
where
exp[KGo(u; h)]
and G 1 (u) is the same as defined by (13.89). For the purpose of solving
the overlap by (13.135), only Go(u; h) is relevant. We have
Multiuser Detection and Statistical Mechanics 261
= J(il'
a<b
~~a.b) J
exp [K (allogXl (u; h) + t
/=2
a/log X/(u)
- t'QabQab)] (13.142)
a9
Xl (U; h)
exp t,
[(P,E + Jp,F z) d. - ~l"Et, d;] } (13.144)
262 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
Therefore,
:h logXI(U;h)lh=o
- ~PlEt,d:] } (13.145)
= Ez {[Jo(z)](u-i)[h(z)Ji} (13.146)
Ez {[Jo(z)]u}
where
(13.147)
(13.148)
and
= (PIE + y'P1F z) (1
- -
+ P1E)-'2 exp
- 3 [ (PI E + n;F z
2(1 + PIE)
r1 .
(13.150)
Therefore,
:h logX1(u; h)lh=O
- - i - [i(PIE+n;Fzf]
(PIE + y'PIF z) (1 + PIE)-T exp 2(1 + PIE)
3i
= Ez {
(13.151)
(13.152)
(13.153)
From (13.137),
T.'_I
L\.I
81og((ZU(r,S;h))) I
8h
= ~l X ( 'h)1
>lh og I U, . (13.154)
h=O U h=O
The above is true for all groups of users. Also noting that E > 0, we
have for every group l = 1, ... , L,
Mz Ez{sgn(PIE+JPIFz)} (13.157)
(13.159)
264 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
By (13.70), we can easily conclude that all users in group I have the
same BER which is
(13.160)
and
(L) 1
al = L' 1= 1, ... , L. (13.162)
F
CT5 + ,8(1 - 2m + q)
(13.167c)
[CT 2 + ,8(1 - m)]2
3 2
= { P E +P2F}
(1 + PE)2
q (13.167d)
E .
The BER of user k with energy Pk converges to
(13.168)
P 2 (F - E)}
q - m = E { (1 + P E) 2 •
(13.171)
266 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
p2E2 }
F - E = (F - E) . /3 E{ (1 + PE)2 . (13.172)
Unless
(13.173)
(13.174)
and as a consequence,
q=m. (13.175)
(13.176)
0 2
q= 1+ _0_. (13.178)
1-/3
Therefore, by (13.169),
f3 E { PE } = 1. (13.180)
I+PE
m is obtained immediately from (13.167b). q can then be solved from
(13.167c)-(13.167d). The multiuser efficiency is found as a positive num-
ber by (13.169). In the special case of equal-energy users, E = 1/f3 - 1
and some algebra gives
(13.181)
((ZU(r, S)))
= J(ITn=l dr n ) L
{dad
2- uK (21l"(J5)-~
Es {ITn=l
exp [- 2:
0
2 (rn - ~ tk=l Msnk) 2]
YlV
268 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
(13.184)
where
(13.186)
XI(U)
Hence we have the same expression for the overlap as (13.156). The
multiuser efficiency is given by the same (13.169).
The conditional mean estimator reduces to the matched filter if 0" --+
00. In this case m, q --+ 0 and we get (13.16) by (13.169).
In case of 0" = 0"0, we get the 10 detector. Notice that for all x,
Ez {tanh(x + v'x z) - tanh 2 (x + v'x z)}
~Je-(y;:)2 [tanh(y) - tanh 2 (y)] dy (13.193)
v27rx
-- 1 e - 2X
y'27rX
1 J e- Y!...
2 :r
eY - e- Y dy
(e Y + e- y)2
(13.194)
o (13.195)
since the integrand in (13.194) is an odd function. It can be shown that
the solution to the saddle-point equations satisfies F = E and q = m.
Therefore the multiuser efficiency, 0"5E, can be found as the solution to
the following fixed-point equation
1
1] -
- 1+ !I [1 - Ep,z { P tanh ( fJ z + ~ ) }] . (13.196)
270 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
6. Discussions
The parameters m and q have interesting physical meanings. Take
the linear detector for example. Consider a simple energy distribution.
By letting i = 1 in (13.155), we get
(13.197)
and by letting i = 2,
(13.198)
(13.199)
Then
« (d) )) (13.200)
«(d))) = m. (13.203)
Similarly,
(13.204)
Multiuser Detection and Statistical Mechanics 271
The same expressions are resulted if we start with (13.192) for the op-
timal detectors. Indeed, (13.203)-(13.204) are true for all multiuser de-
tectors of interest.
Let
E = 1- 2m+ q. (13.205)
Recall that the transmitted symbol dk = 1 for each user k. The right
hand side of (13.207) is the mean square error of the conditional mean
estimator output. Interestingly, the multiuser efficiency is then expressed
in the mean square error as
1
ry=---::- (13.208)
1 + EP...
(}"~
'I1(mf) _ _1_
./ - {3 , (13.209)
1 + (}"2o
K
.;p; d1 + Lsfsk.j"P; (dk - (dk)) + O"OWI (13.211)
k=2
272 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
Conditional
Mean Estimator
7. Spectral Efficiency
The maximum sum of the rates of a CDMA channel is equal to the
maximum mutual information between transmitted symbols and the re-
ceived signal
D(PrldllPrIPd) (13.212)
E {log P(rld)} + E {-log P(r)}. (13.213)
Assume the spreading codes are known and fixed. By (13.6), P r1d is a
Gaussian density, and hence
Then Pr is given by
(13.216)
= (27ra6)-~ Z(l)(r, S) (13.217)
= ~E
2
{log(l + PE)} + ,8E (m -
2
1) + ~2 log (1 + a,82 (1 -
o
m)) .
(13.222)
By (13.167a)-(13.167b) and that 1](mmse) = aJE, the spectral efficiency
is exactly (13.31), which is repeated here,
,8 { log ( 1 + 1] (mmse))
C = -E } + -log
1 1 + _(1](mmse)
1
2 aJ P 2 1](mmse) 2
- 1).
(13.223)
Not surprisingly, independent fading on an equal transmit energy system
is a perfect cause of the energy distribution. Therefore we have obtained
the spectral efficiency as a function of the multiuser efficiency. In fact,
we can identify the first term on the right hand side of (13.223) to be
the spectral efficiency of the linear MMSE receiver.
The spectral efficiency gives the maximum total number of bits per
second per chip that can be transmitted reliably from all users to the
receiver over the multiuser channel. Unfortunately it does not easily
break down to a rate combination of individual users that achieves it
as (13.223) may seem to suggest.
274 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY
8. Conclusions
This paper exploits the connection between large-system multiuser
detection and statistical mechanics, and presents a new interpretation
of multiuser detection in general.
We first introduced the concept of Bayes retrochannel, which takes
the received signal as the input and outputs a stochastic estimate of
the transmitted symbols. By assuming an appropriate prior distribu-
tion and a channel characteristic that may be different to the true ones,
a multiuser detector can be expressed as a conditional mean estima-
tor that outputs the mean value of the stochastic output of the Bayes
retrochannel.
The Bayes retrochannel is equivalent to a spin glass in the sense that
the distribution of its stochastic output conditioned on the received sig-
nal is exactly the distribution of the spin glass at thermal equilibrium.
The performance of the multiuser detector is then found as a certain
macroscopic property of the spin glass. In particular, the BER can
be obtained through calculating the overlap of the spin glass. In the
large-system limit, the macroscopic properties as such can be solved by
powerful tools developed in statistical mechanics.
In this paper we have solved through a unified analysis the large-
system uncoded BER of the matched filter, the MMSE detector, the
decorrelator, the individually and jointly optimal detectors. We show
that under arbitrary received energy distribution, the large-system BER
is uniquely determined by the multiuser efficiency, which has a very
simple relationship with the output mean square error of the conditional
mean estimator. The relationship also implies that a multiuser detector
is in general equivalent in performance to interference subtraction using
a conditional mean estimator obtained for certain prior and conditional
distribution (depending on the detector), and the remaining interference
is always Gaussian distributed in the large-system limit.
Using the techniques developed in this paper, we have also solved the
spectral efficiency of the multiuser channel.
Notes
1. Imagine a system consisting molecules with random magnetic spins that evolve over
time, while the random positions of the molecules are fixed for each concrete instance as in
a piece of glass.
2. The SIR is defined as the energy ratio of the useful signal to the noise in the output.
In contrast, the SNR of user k is usually defined as Pk/(2a~).
3. Precisely, we refer to the large-system BER. Unless otherwise stated, all performance
indexes such as BER, SIR, multiuser efficiency and spectral efficiency refer to large-system
performance hereafter.
REFERENCES 275
References
[1] S. Verdu, Multiuser Detection. Cambridge University Press, 1998.
[2] D. Guo, L. K. Rasmussen, and T. J. Lim, "Linear parallel interfer-
ence cancellation in long-code CDMA multiuser detection," IEEE
J. Selected Areas Commun., vol. 17, pp. 2074-2081, Dec. 1999.
[3] A. J. Grant and P. D. Alexander, "Random sequence multisets for
synchronous code-division multiple-access channels," IEEE Trans.
Inform. Theory, vol. 44, pp. 2832-2836, Nov. 1998.
[4] D. N. C. Tse and S. V. Hanly, "Linear multiuser receivers: effective
interference, effective bandwidth and user capacity," IEEE Trans.
Inform. Theory, vol. 45, pp. 622-640, March 1999.
[5] S. Verdu and S. Shamai, "Spectral efficiency of CDMA with ran-
dom spreading," IEEE Trans. Inform. Theory, vol. 45, pp. 622-640,
March 1999.
[6] D. N. C. Tse and S. Verdu, "Optimum asymptotic multiuser effi-
ciency for randomly spread CDMA," IEEE Trans. Inform. Theory,
vol. 46, pp. 2718-2722, Nov. 2000.
[7] R. R. Miiller, "Multiuser receivers for randomly spread signals: fun-
damentallimits with and without decision-feedback," IEEE Trans.
Inform. Theory, vol. 47, pp. 268-283, Jan. 2001.
[8] P. Viswanath, D. N. C. Tse, and V. Anantharam, "Asymptoti-
cally optimal water-filling in vector multiple-access channels," IEEE
Trans. Inform. Theory, vol. 47, pp. 241-267, Jan. 2001.
[9] J. Zhang, E. K. P. Chong, and D. N. C. Tse, "Output MAl distri-
bution of linear MMSE multiuser receivers in DS-CDMA systems,"
IEEE Trans. Inform. Theory, vol. 47, pp. 1128-1144, March 2001.
[10] S. Shamai and S. Verdu, "The impact of frequency-flat fading on
the spectral efficiency of CDMA," IEEE Trans. Inform. Theory,
vol. 47, pp. 1302-1327, May 2001.
[11] D. Guo, S. Verdu, and L. K. Rasmussen, "Asymptotic normality
of linear CDMA multiuser detection outputs," in Proceedings 2001
IEEE International Symposium on Information Theory, p. 307,
Washington, D.C., June 2001.
[12] L. Li, A. Tulino, and S. Verdu, "Asymptotic eigenvalue moments
for linear multiuser detection," Communications in Information and
Systems, vol. 1, pp. 273-304, Sept. 2001.
[13] D. Guo, S. Verdu, and L. K. Rasmussen, "Asymptotic normality
of linear multiuser receiver outputs," accepted for IEEE Trans. on
Inform. Theory, 2001.
276 COMMUNICATIONS, INFORMATION AND NETWORK SECURITY