IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO.

9, SEPTEMBER 2004 2123

Now Analysis of Multiple-Antenna Wireless Links at Low SNR

g (t) = log det(X + Z + tN ) 0 log det(Z + tN ) Chaitanya Rao, Student Member, IEEE, and Babak Hassibi
n
= log(1 + ti (NY
01 ))
i=1
0 log(1 + ti (NZ 0 )) 01 ) Abstract—Wireless channels with multiple transmit/receive antennas
+ log det(ZY
1
are known to provide a high spectral efficiency both when the channel is
known to the receiver, and when the channel is not known to the receiver
so that if the signal-to-noise ratio (SNR) is high. Here we analyze such systems
at low SNR, which may find application in sensor networks and other
n
d2 g low-power devices. The key point is that, since channel estimates are not
dt2
=
1
2
0 1
2 reliable, it is often not reasonable to assume that the channel is known
i=1 t +  (NZ 1
)
t +  (NY
1
)
at the receiver at low SNR. In this unknown channel case, we show that

0
for sensible input distributions, in particular all practical modulation
schemes, the capacity is asymptotically quadratic in the SNR, , and thus
where for the last inequality we have used (15) and the fact that t 2 T .
much less than the known channel case where it exhibits a linear growth
in . We show that under various signaling constraints, e.g., Gaussian
Strict inequality holds for X  0. modulation, unitary space–time modulation, and peak constraints, that
mutual information is maximized by using a single transmit antenna.
We also show that at low SNR, sending training symbols leads to a rate
REFERENCES reduction in proportion to the fraction of training duration time so that it
[1] T. W. Anderson, “The integral of a symmetric unimodal function over is best not to perform training. Furthermore, we show that the per-channel
a symmetric convex set and some probability inequalities,” Proc. Amer. use mutual information is linear in both the number of receive antennas
Math. Soc., vol. 6, pp. 170–176, 1955. and the channel coherence interval.
[2] T. Başar, “The Gaussian test channel with an intelligent jammer,” IEEE
Index Terms—Low-signal-to-noise ratio (SNR) regime, multiple-antenna
Trans. Inform. Theory, vol. IT-29, pp. 152–157, Jan. 1983.
systems, noncoherent channels, Rayleigh fading.
[3] T. Başar and G. J. Olsder, Dynamic Noncooperative Game
Theory. Philadelphia, PA: SIAM, 1999.
[4] J. M. Borden, D. M. Mason, and R. J. McEliece, “Some information
theoretic saddlepoints,” SIAM J. Control Optimiz., vol. 23, no. 1, pp. I. INTRODUCTION
129–143, 1985.
Multiple-antenna wireless systems have been shown to provide high
[5] S. N. Diggavi and T. M. Cover, “The worst additive noise under a covari-
ance constraint,” IEEE Trans. Inform. Theory, vol. 47, pp. 3072–3081, capacity, exploiting the presence of fading in such channels. However,
Nov. 2001. this is based on the premise that either the channel coefficients are
[6] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, U.K: known to the receiver, or that the signal-to-noise ratio (SNR) of the
Cambridge Univ. Press, 1985. channel is high [1]–[3].
[7] A. Kashyap, T. Başar, and R. Srikant, “Correlated jamming on MIMO
Gaussian fading channels,” in Proc. IEEE Int. Conf. Communications Wireless systems operating at low SNR (exhibiting weak signaling
(ICC), Paris, France, 2004, pp. 458–462. or in noisy environments) find increasing use in energy-efficient
[8] Y. H. Kim and A. Lapidoth, “On the log-determinant of the noncen- devices such as sensor networks. Recent work on analyzing the
tral wishart distribution,” in Proc. IEEE Int. Symp. Information Theory capacity of low-SNR multiple-antenna links, assuming that the
(ISIT), Yokohama, Japan, 2003, p. 54.
channel is known at the receiver, has appeared in [4]. However, at
[9] M. Medard, “Capacity of correlated jamming channels,” in Proc. 35th
Allerton Conf., 1997, pp. 1043–1052. low SNR, channel estimates in some circumstances are unreliable
[10] W. E. Stark and R. J. McEliece, “On the capacity of channels with block and so it is sensible to assume that the channel is unknown. In the
memory,” IEEE Trans. Inform. Theory, vol. 34, pp. 322–324, Mar. 1988. following analysis we, therefore, assume the channel is unknown to
[11] İ. E. Telatar, “Capacity of multi-antenna Gaussian channels,” Europ. both transmitter and receiver. As shown later, this leads to results
Trans. Telecommun., vol. 10, pp. 585–589, Nov./Dec. 1999.
qualitatively different from the known channel case.
We use the block-fading model of a wireless multiple-antenna
system proposed by Marzetta and Hochwald in [5], expressing the
mutual information between input and output as a function of the
model parameter  (proportional to the SNR) up to second order.
This model is described in detail in the next section. Maximizing this
expression gives us insight about desired signaling at low SNR as
well as the optimal number of antennas to be used at the transmitter
and receiver. It has been shown in [6] that the optimum signaling at
low SNR achieves the same minimum energy per bit as the known
channel cases for single transmit antenna systems. We show that
the on–off optimal signaling found in [6] also generalizes to the
multiple-antenna setting (a result that also follows from [7, Theorems

Manuscript received March 10, 2003; revised February 1, 2004. This work
was supported in part by the National Science Foundation under Grant CCR-
0133818, by the Office of Naval Research under Grant N00014-02-1-0578, and
by Caltech’s Lee Center for Advanced Networking.
The authors are with the Department of Electrical Engineering, California
Institute of Technology, Pasadena CA 91125 USA (e-mail: rao@systems.cal-
tech.edu; hassibi@systems.caltech.edu).
Communicated by G. Caire, Associate Editor for Communications.
Digital Object Identifier 10.1109/TIT.2004.833369

0018-9448/04$20.00 © 2004 IEEE

2124 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004

1 and 5]). However, this scheme requires increasingly peaky signals Computing the capacity of this multiple-antenna system for generic
(indeed, ones with unbounded fourth-order moment) and so may not  is an open problem. In [5], however, it is shown that capacity is
be acceptable from a practical point of view in some situations. We achieved when the input signal S has the form
therefore focus our attention on signaling schemes with bounded
fourth-order moment.
S = 8D; (2)
Recent work by Verdú [7] has shown that knowledge of the first and where D is a diagonal M 2 M matrix with nonnegative entries and 8
second derivatives of capacity at low SNR also tells us about band- is a T 2 M isotropically distributed unitary random matrix. This means
width and energy efficiency for signal transmission. For example, these
quantities are used to see how spectral efficiency grows with energy per • 83 8 = IM (the M 2 M identity matrix) although 883 6= IT
bit. More work on constrained signaling in the low-power regime for for T > M , and
Rayleigh-fading channels is given in [8], [9] and [10], while [11] and • the distribution of 8 is unaltered when left multiplied by a de-
[12] study the Rician case. In [13], among other things, the low-SNR terministic T 2 T unitary matrix or when right multiplied by a
mutual information of the same block-fading multiple-antenna channel deterministic M 2 M unitary matrix [5].
of [5] is also calculated. Similar results to ours have also been obtained
in [14], as a by-product of their study of the capacity of general commu- 8 and D are independently distributed.
nication channels under small-peak constraints. Our results differ from
[13] and [14] in two ways. First, we require a weaker assumption on III. MUTUAL INFORMATION TO SECOND ORDER
the input signals; essentially, conditions on the fourth- and sixth-order In this section we prove the following result.
moments, rather than an exponentially decaying input distribution as
in [13], or a peak constraint on the singular values of the transmitted Theorem 1: Consider the model (1) and let p(S ) denote the proob-
signal as in [14], both of which render all moments finite. Second, we ability density function (pdf) of S .
study the optimal signaling structure derived in [5] and further opti- 1) First-order result: If i) @p(S )=@ exists at  = 0 and ii)
mize mutual information subject to various signaling constraints such lim!0 Etr(SS 3 )2 = 0, the mutual information between the
as training. transmitted and received signals S and X for the multiple-an-
There are two main parts to this correspondence. In the first part, tenna system (1) is zero to first order in , i.e., I (X ; S ) = o().
2) Second-order result: If, in addition, i) @ 2 p(S )=@2 exists
we expand the mutual information of the wireless link to second order
in the SNR  using an approach that may be applied to other channel
at  = 0, ii) the fourth-order moment of S is finite, i.e.,
models. Secondly, we optimize this expression under both peak and
fourth-order moment signal constraints to determine what signaling
Etr(SS 3 )2 4 < 1, and iii) lim!0 Etr(SS 3 )3 = 0,
then the mutual information between S and X up to second
should be applied to the input and how many transmit antennas should
order in  is given by
be employed. We also study Gaussian modulation, unitary space–time
N tr[E(SS 3 )2 0 (ESS 3 )2 ] 2
I (X ; S ) =  + o(2 ):
modulation. and training-based schemes.
2M 2 (3)

II. MODEL
The second-order part of the theorem is essentially a result in [13]
We consider a discrete-time block-fading channel model [5] in which and [14]. However, we here require a much less stringent condition
there are M transmit and N receive antennas. The channel is described on the input distribution. Moreover, we shall optimize (4) for various
by a propagation matrix H that is assumed constant for a coherence in- signaling schemes.
terval of length T symbols. For the next T symbols, the propagation The reason for the condition on p(S ) in Theorem 1 is that the choice
matrix changes to a new independent value, and so on. Signals are rep- of distribution may depend on the SNR . Condition ii) of the first-order
resented by complex-valued matrices with T rows, where the ith row result limits the growth of the fourth-order moment, whereas conditions
is what is transmitted or received via each of the multiple antennas at ii) and iii) of the second-order result, respectively bound and limit the
the ith time slot. growth of the fourth- and sixth-order moments. The regularity condi-
For each coherence interval, the T 2 N received matrix X is related tions i) on p(S ) at  = 0 are required for reasons that will be seen
to the T 2 M transmitted matrix S by shortly (see Sections III-B and V).
 For the optimum signaling structure (2), (3) can be replaced by
X = SH + V (1)
M N (4 0 22 =T ) 2
I (X ; S ) =  + o ( 2 ):
where H is an M 2 N matrix and V is a T 2 N noise matrix, both 2M 2 (4)
comprised of zero-mean and unit-variance circularly symmetric com- Note that under any reasonable input distribution (and certainly all
plex Gaussian entries. The matrices H , V , and S are assumed to be practical modulation schemes) the mutual information has no linear
independent and the values of H and V are unknown to both trans- term in  and so the capacity is much less than the known channel case
mitter and receiver. S satisfies the power constraint where the low-SNR expansion of the well-known log det formula has
T M a nonzero first-order term. Since 4 and 2 are independent of N , (4)
EtrSS 3 = E jsij j2 2  Pmax suggests that the capacity increases linearly in the number of receive
i=1 j =1 antennas. The dependence of the mutual information on M is more
where E and tr denote the expectation and trace operators, respectively, complicated since both the denominator (M 2 ), as well as the numer-
and sij is the (i; j )th entry of S . Throughout this work, the 3 operator ator (via 2 and 4 ) depend on M . However, careful analysis will show
denotes the conjugate transpose of a matrix. When 2 = T M , the that for most practical signal constraints the optimal value is M = 1

normalization factor M in (1) makes  equal to the SNR at each transmit antenna.
receive antenna. Otherwise, the SNR at each receive antenna is given Finally, note that the mutual information is affine linear in 4 sug-
by E[trXX 3 ]=E[trV V 3 ] = 2 =(T M ). It is also known that there is gesting that it increases as the input becomes more peaky, in good
no performance gain in having the number of transmit antennas greater agreement with the results of [6] and their multiple-antenna general-
than T [5]. Hence, we will assume that T  M . izations.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004 2125

A. Conditional Entropy Approximation Also, p(X jS ) in (5) is a function whose derivatives with respect to 
We compute I (X ; S ) = h(X ) 0 h(X jS ) via the conditional pdf at  = 0 may be calculated. These two facts imply that
p(X jS ). Given S , X is zero-mean complex Gaussian with covariance p(X ) = p(S )p(X jS ) dS
E(XX 3 jS ) = N (IT + M SS 3 ) and so, as in [5]
can also be expanded to second order as
e0trX (I + SS ) X
p ( X jS ) =  SS 3 ) N : (5) p(X ) = p(X; 0) + p0 (X; 0) +
2
p00 (X; 0) + o(2 )
NT det(IT + M 2
where p0 (X; 0) and p00 (X; 0) are used to denote the first and second
Here IT denotes a T 2 T identity matrix. From p(X jS ) it is possible partial derivatives of p(X ) with respect to , respectively, evaluated at
to compute the conditioned entropy h(X jS ) directly  = 0. Also, to second order
h ( X jS ) =
0 E log p(X jS ) p0 (X; 0)
log p(X )  log p(X; 0) + 
= N T log  + E log det(IT + M SS 3 )N p(X; 0)
p00 (X; 0) p0 (X; 0)
2
2
3
+ EtrX (IT +

SS 3 )0 X 1 +
2 p(X; 0)
0 p(X; 0)
:
M
SS 3 )

= N T log  + N E log det(IT + This leads us to the following quadratic approximation:
M
SS 3 )0 XX 3

+ Etr(IT +
1
h (X ) = 0 p(X ) log p(X ) dX
M
(using tr AB = tr BA)
0 p(X; 0) log p(X; 0) dX
= N T log  + NE log det(IT + M SS 3 )+ ES trNIT
= N T log e + N E log (1 +
 2
d ) 0 p0 (X; 0) log p(X; 0) + p0 (X; 0) dX
i M i
p00 (X; 0) + (p0 (X; 0))2=p(X; 0)
2

are the eigenvalues of SS 3 )

(where di
2 0 2
 2 00
= N T log e + N E log(1 + di ) (6) + p (X; 0) log p(X; 0) dX: (10)
i M
We now claim that the integrals in (10) involving the second deriva-
2 4
tive p00 (X; 0) are equal to zero.
 2
 N T log e + N E (
M
di 0
2M 2
di ) (7)
i First, note that
 N 2
= N T log e + N 2 0 4 (8) p(X; 0) + p0 (X; 0) + p00 (X; 0) + o(2 ) dX = 1:
2
M 2M 2 2

since 2 = EtrSS 3 = E 32
i di and 4 = Etr(SS ) = E
2
i di .
4 Comparing coefficients of n on both sides we conclude

p(n) (X; 0) dX
The approximation step (7) is made assuming that the second-order
= 0; n = 1 ; 2: (11)
approximation
Also, since S , H , and V are independent, (1) implies that
 2  2 2 4
E log 1+ d
M i
E di 0 di
M 2M 2 EtrXX 3 = Etr M SHH 3 S 3 + 
M
EtrSHV 3
is valid for each i. Consider the inequality
+

M
EtrV H 3 S 3 + EtrV V 3
 2 2 4 2 4 3 6
d 0 d
M i 2M 2 i
 log(1+ M di2 )  M di2 0 2M d +
2 i 3M 3
di
=
 3 3
trEHH S S + 0 + 0 + trNIT
M
) 0  log(1 + M di ) 02( M di 0 2M di )  3M 3 di6 :
 2  2  4
=
 3
trNIM ES S + N T
M
(9) = N (2 =M + T ): (12)
Thus,
For the second-order approximation to be a valid one the limit of
the expression between the two inequalities in (9) should go to zero in trXX
3 p(X; 0) + p0 (X; 0) + 2 p00 (X; 0) + o(2 ) dX
expectation as  ! 0 for each i. The condition 2
N (2 =M + T ):
Etr(SS 3 )3 = E
=
(di )
2 3
!0
Comparing coefficients of n of on both sides
in the second-order statement of Theorem 1 ensures that this occurs.
The first-order condition Etr(SS 3 )2 similarly ensures that d2i ! 0, trX
3 Xp(X; 0) dX = N T
making log(1 + M  d2 )   d2 a valid first-order approximation.
i M i
trX
3 Xp0 (X; 0) dX = N 2 =M
B. Entropy Approximation
The pdf p(X ) depends on the input distribution p(S ). Our regularity trX
3 Xp00 (X; 0) dX = 0:
conditions i) on p(S ) in Theorem 1 guarantee that the distribution can
be expanded to second order around  = 0 as
Now p(X; 0) = e  and so
p(S ) = p(S; 0) + p0 (S; 0) + p00 (S; 0) + o(2 ):
2
2
log p(X; 0) = 0trX 3X 0 N T log :
2126 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004

Hence the zeroth-order term in (10) is from (14) and this together with (8) gives
( ; ) = h ( X ) 0 h ( X jS )
I X S
( 0) log p(X; 0) dX  (N T log e + N 2 =M  0 2M
N 32
2 tr (ESS ) )
p X; 2


=0 ( 0)trX 3 X dX 0 N T log 
p X; ( 0) dX
p X; 0 (N T log e + N 2=M  0 2NM42 2 )
32 32
= 0N T 0 N T log  = N tr E(SS ) 0 (ESS ) 2
2M 2 (16)
= 0N T log e: (13)
as stated in Theorem 1.
Similarly We remark that to show the first-order result I (X ; S ) = o() we
only require first-order expansions of p(X ) and h(X ), and so the con-
p X;
0( 0) log p(X; 0) dX = 0 N 2 =M ditions stated in the first-order result of Theorem 1 suffice.
In the special capacity-optimizing case of S = 8D , we have
and
p
00 (X; 0) log p(X; 0) dX = 0: E(SS 3 )ij = E(8D2 83 )ij
= E ik dk2 jk (dk2 are both the diagonal entries of
The preceding calculations combined with (10) lead to k
2
2
D and the eigenvalues of SS
3)
( )  N T log e+N 2 =M 0 2 (p0 (X; 0))2 =p(X; 0)dX:
h X (14) = E [d k ]E
2
ik jk (since 8 and D are independent):
k
This shows that to express h(X ) to second order, it suffices to cal-
culate only the first derivative of p(X ) at  = 0. We use the following The expectation E ik jk is evaluated by noticing that the expec-
result, proved in the Appendix. tation is unchanged by adding T 0 M orthonormal columns to 8 to
make it a T 2 T unitary, denoted say by 9 = ( ij ). Then using the
Lemma 1: For the model (1), the first derivative of the pdf of X relation 993 = IM we have
evaluated at  = 0 is given by T
ik jk = ij :
0( e0trX X k=1
p X; 0) = M  NT
(trX 3P X 0 N 2 ) Taking expectations of both sides and using the fact that each entry
of 9 has the same distribution, gives us
where P = ESS 3 .
This gives us E ik jk = Tij ; for k = 1 to T :
This implies E ik jk = = 1 to M .

(p0 (X; 0))2 dX T for k
Hence,
p(X; 0)

e0trX X
E(SS 3 )ij = Tij E[d2k ]
= M 2  NT
(trX 3 P X 0 N 2 )2 dX k
ij
= 2 :
= M1 2 [EG (trG3 P G)2 0 2N 2 EG trG3 P G + N 2 22 ] (15)
T
In other words, ESS 3 = 
T IT , and so tr(ESS 3 )2 = 22 =T . This
where G is a T 2 M random matrix having the pdf p(G) =  e in (16) gives
and EG denotes expectation over the random variable G.
( ; ) = N (420M22 =T ) 2 + o(2 ):
2
To proceed we use the following lemma proved in the Appendix. I X S

Lemma 2: If G is a T 2 N matrix with independent zero-mean

unit-variance complex Gaussian entries, then IV. EXAMPLES
• EG trG3 P G = N 2 ; We now compute the low-SNR mutual information for some cases
• EG (trG3 P G)2 = N 2 22 + N trP 2 ; of interest.
2 T deterministic matrix P satisfying trP = 2 .
for any T
A. Gaussian Modulation
For P = ESS 3 , from (15) and Lemma 2 we have
Suppose the transmitted signal S has independent zero-mean unit-
(p0 (X; 0))2 dX variance complex Gaussian entries. Then
p(X; 0)
(ESS 3 )ij = Esik sjk = M ij
= M1 2 N 2 22 + N tr (ESS 3 )2 0 2N 2 N 2 + N 2 22 k

so that ESS 3 = M IT . In the Appendix , we show that for a Gaussian

= MN2 tr (ESS 3 )2 :
matrix, 4 = E(tr(SS 3 )2 ) = M T (M + T ) and so
1 I (X ; S ) = N 2
(
MT M ( + T ) 0 (T M )2 =T ) + o(2 )
Hence,
T 2M 2T
( )  N T log e + N 2 =M  0 2M
N 32 NT 2
2 tr (ESS ) =2 + ( )
2 2
h X   o  :
M
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004 2127

This has two interesting ramifications. First, the capacity per channel Note that the above distribution does not satisfy the regularity condi-
use increases linearly in T (I (X ; S ) is actually quadratic in T ) and, tions i) of Theorem 1.
second, the optimal number of transmit antennas is M = 1. Then from (6)

B. Unitary Space–Time Modulation ( j ) =N T log e + N E log det(IT + M SS 3 )

h X S
p
In this scheme, we let S = 8 T (where 8 has an isotropic unitary = N T log e + N  log(1 + M T 0 )M
distribution), which gives 2 = T M and 4 = T 2 M . Using this in (4)
yields  N T log e + N  M log T

0 as 10 is large
1
M
1 I (X ; S ) = N (T 0 M )  2 + o (2 ) = N T log e + N M  [(1 0 ) log  + log(T =M )]
T 2M = N T log e + o(); since  > 1:
which is strictly less than the Gaussian case. Again, the optimal number
of transmit antennas is M = 1.
Also, we have
0trX (I + A) X 0trX X

C. Training-Based Schemes ( ) =  eN T det(I +  A)N + (1 0  ) e N T : (19)

p X
T M
In these schemes, we have For  small and  > 1,  is small while 10 is large. Hence, in the
S first term of (19), the determinant in the denominator is
S =
1 + M T 0  M N 0
Sd MN
(1 )

where S is a T 2 M fixed training matrix and Sd is a Td 2M

zero-mean random matrix. Furthermore which is large while the numerator is bounded above by 1. Therefore,
the second term is much larger than the first and so
trS3 S = 2; ; EtrSd Sd3 = 2;d ; 2; + 2;d = 2 ; T + Td = T :
( ) = 0 E log p(X )
h X
Under these conditions it can be readily shown that 0trX X
 0 E log(1 0  ) e N T
tr(ESS 3 )2 = tr(S S3 )2 + tr(ESd Sd3)2 = 0 log(1 0  ) + EtrX 3 X + N T log 
and
  + N T log  + EtrX 3X
trE(SS 3 )2 = tr(S S3 )2 + trE(Sd Sd3)2 : = N T log  + N T ( + 1) +  using (12)
Therefore, using (3), we obtain = N T (log e + ) + o():
32 32
( ; ) = N tr[E(Sd Sd2)M02 (ESd Sd ) ] 2 + o(2 ):
I X S (17)
Then I (X ; S )=T = N  + o(), so the first-order term corresponds
to that of the capacity when the channel is known, equal to
The latter equation shows that the mutual information is independent
of S . In fact, the right-hand side of (17) is just the mutual information
E log det I + M HH 3 = N  + o():
of a system with coherence interval Td = T 0T . Thus, in the low-SNR However, such signals cannot be used in practice and so we shall con-
regime, training actually contributes a rate reduction proportional to sider signals that are peak constrained or have a fourth-order moment
the fraction of time that one is sending the training symbols. One may constraint.
contrast this with the result of [15] which shows that training-based
schemes achieve capacity at high SNR. A. Fourth-Order Moment Constraint
Suppose we enforce the fourth-order moment constraint 4  K22
V. OPTIMAL SIGNALING for some positive constant K . This may be a practical constraint to
In this section, we shall optimize (4) to determine what kind of sig- impose but as mentioned in [7] a bounded fourth-order moment will
naling should be applied to maximize the mutual information between not lead to mutual information optimality at low SNR.
the transmitted and received signals. It is known that, under the stan- By the root mean square–arithmetic mean inequality we have
2
dard power (second-order) constraint, capacity approaches up to first
order the capacity of a channel where the channel matrix is perfectly
M
di 
4 1 M
di
2
M
known to the receiver. This is achieved by a peaky input distribution i=1 i=1
[6]. from which we conclude 4  22 =M . Also, T  M as stated in
We can show this is also the case for the multiple-antenna channel Section II and so 4  22 =T . Hence, we require that K > 1=M 
as follows. For any  > 1 and assuming  < 1, define our transmitted 1=T .
signal to satisfy Then 4 0 22 =T  (K 0 1=T )22 and as K 0 1=T > 0, from
3= A; w.p.  (4) it follows that maximizing the mutual information is equivalent to
w.p. 1 0 
SS
0T 2T ; maximizing 2 . We therefore have the following result.

where Theorem 2: Consider the model (1) and suppose that the input
signal must satisfy the constraints 2  T M and 4  K22 . Then,
A = T 0 0 IM 0M 2(T 0M ) : to second order, the mutual information is maximized by any input
(T 0M )2M 0 (T 0M )2(T 0M )
(18)
distribution that simultaneously achieves 2 = T M and 4 = K22
Then S satisfies the power constraint and is given by
1 I (X ; S ) = N (K 0 1=T )T 2 + o(2 ):
EtrSS 3 = tr(T 0 IM ) 2  = T M: T 2 (20)
2128 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004

One distribution that achieves this is given by (2) where optimal number of transmit antennas is M = 1 and the optimal mutual
information is
2 2 2
(d1 ; d2 ; . . . ; dM ) NK (T 01) 2
1 2T
 + o (2 ); if T > 1
(T KM; T KM; . . . ; T KM ); w.p. 1=(KM ) I (X ; S ) =
T NK 2 + o(2 );
w.p. 1 0 1=(KM ).
= if T = 1.
(0; 0; . . . ; 0); 8

One distribution that achieves this is given in (2), where the diagonal
Note here that the optimal mutual information (per channel use) is entries of D2 are given by (23) or (24) depending on whether Pmax 
independent of the number of transmit antennas and is proportional to T K=2 or Pmax < T K=2, respectively.
both N and T .
VI. CONCLUSION
B. Peak Constraint
For the block-fading multiple-antenna channel model in which
Due to the isotropic unitary matrix in (2) it is not possible to di- the channel is unknown to the transmitter and receiver, the low-SNR
rectly enforce a peak constraint on the transmitted signals. However, asymptotic mutual information has a quadratic leading term. This
it is possible to force a maximum constraint on the diagonal entries mutual information may be maximized by using one transmit antenna,
of D (the singular values of S ). Thus, assume that di2  K for some many receive antennas, and with on–off signaling. When there is a
positive constant K and for all i. For any fixed M , we wish to max- maximum constraint on the singular values of the transmit signal, it
imize 4 0 22 =T subject to the constraint 2  Pmax . We also have is possible to obtain a higher capacity by lowering the signal power
2 = E M i=1 di  M K . Now
2
from its maximum allowed level.
M
4 0 22 =T E 0 APPENDIX
= di4 2
2 =T
i=1 Here we prove the lemmas used in the main sections.
M
 KE di2 0 2
2 =T Lemma 1: For the model (1), the first derivative of the pdf of X
i=1 evaluated at  = 0 is given by
=  2 (K 0  =T ) (21)
e0trX X
p0 (X; 0) = 3
2
(trX P X 0 N 2 )
with equality iff all the di ’s are equal to 0 or K . This quantity is maxi- M  NT
mized at either 2 = T K=2, 2 = Pmax , or 2 = M K , depending on where P = ESS 3 .
which of the three quantities T K=2, Pmax , or M K is smallest. This Proof: We first approximate p(X jS ) to first order in . Ex-
leads to panding the numerator to first order

T K 2 =4; if L = T K=2 e0trX (I + SS ) X

 e0 trX (I 0 SS )X

4 0  =T
2
2  Pmax (K 0 Pmax =T ); if L = Pmax (22) =e
0trX X etr( X SS X )
M K 2 (T 0 M )=T; if L = M K  e0trX X 1 +  trX 3 SS 3 X :
M
where L = minfT K=2; Pmax ; M K g. Equality holds in (22) when 2 To expand the denominator we use
is set to minfT K=2; Pmax ; M K g. Corresponding distributions that
log det(I+A)
achieve equality are det(I + A) = e
tr log(I +A)
=e
(K; K; . . . ; K ); w.p. minf1; T =2M g
(d
2
1
2
2
2
; d ; . . . ; dM ) =
(0; 0; . . . ; 0); w.p. 1 0 minf1; T =2M g  e A
tr( )

(23)  1 + tr(A)
so that
and
0N 0N
SS 3  1 + M trSS 3

2 2 2 (K; K; . . . ; K ); w.p. minf1; Pmax =M K g det IT +
M
(d1 ; d2 ; . . . ; dM ) =
(0; 0; . . . ; 0); w.p. 1 0 minf1; Pmax =M K g 3
(24)
 1 0 N
M
trSS :

Putting everything together

depending on whether Pmax  T K=2 or Pmax < T K=2, respectively.
e0trX X  3 3 3
The mutual information bounds are p ( X jS ) 
 NT
1+
M
trX SS X 1 0 N
M
trSS
2 2 2
N T K  =8M ; if L = T K=2
I (X ; S )  N Pmax (K 0 Pmax =T )2 =2M 2 ; if L = Pmax and by taking the coefficient of  of both sides it follows that
N K 2 (T 0 M )2 =2T M; e0trX X
p0 (X jS;  = 0) = 3 3 3
if L = M K .
(trX SS X 0 N trSS ): (25)
M  NT
Note that all the above bounds are decreasing functions of M . There-
fore, it is clear that the optimal choice of transmit antennas is M = 1. To find p0 (X; 0) we take the expectation of (25) over S , leading to
Since it is most likely that K < Pmax (that is, L = M K = K unless the required result.
T = 1 in which case L = T K=2 = K=2), we have the following Lemma 2: If G is a T 2 N matrix with independent zero-mean
theorem. unit-variance complex Gaussian entries, then
Theorem 3: In the model (1) with optimal signaling as in (2) sup- • EG trG3 P G = N 2
pose the diagonal entries di of D satisfy di2 < K for all i, where K is • EG (trG3 P G)2 = N 2 22 + N trP 2
some constant less than Pmax . Then, for asymptotically low SNR, the for any T 2 T matrix P satisfying trP = 2 .
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004 2129

Proof: Proof: Denote the (i; j )th entry of S by sij . Then, as the pdf of
1) Denoting the (i; j )th entries of G and P by gij and pij , respec- the Gaussian matrix S is
tively, we have e0trS S
p(S ) =
 MT
EG trG3 P G = EG gji pjk gki we have
0trS S
i;j;k
Etr(S 3 S )2 = e MT tr(SS 3 )2 dS
= E
pjk G [gji gki ]
i;j;k e0trS S
= sij skj skl sil dS
pjk jk  MT
= i;j;k;l
e0trS S
i;j;k
(since gki and gji are independent unless = sij skj skl sil dS:
 MT
i;j;k;l
j = k; in which case the expectation is unity)
=N pjk jk In this summation, the indices i and k each range from 1 to T while
j;k the indices j and l each range from 1 to N . If both i 6= k and j 6= l,
the integral
=N pjj
j e0trS S
= N trP sij skj skl sil dX = 0
 MT
as required. as the integrand is an odd function of the variable sij .
2) We have Using the elementary integrals e  = ds e  jsj2 ds = 1,
 jsj ds = 2 where the integrals are over s 2 , we have
2 e 4
EG (trG3 P G)2 = EG gji pjk gki
e0trSS e0trS S
i;j;k 3 2
tr(S S ) dS = jsij j2 jsil j2 dS
 MT  MT
= EG gji pjk gki gml pmn gnl i;j;l
j 6=l
i;j;k;l;m;n
e0trS S
= pjk pmn EG [gji gki gml gnl ] : +
 MT
jsij j2 jskj j2 dS
j;k;m;n i;l i;j;k
i6=k
E[gji gki gml gnl ] will be nonzero only when terms pair up as jgj4 e0trS S
or jgi j2 jgj j2 . This will occur in the following instances: +
 MT
jsij j4 dS
• j = k = m = n, i = l i;j
• j = k = m = n, i 6= l = 1+ 1+ 2
• j = k 6= m = n i;j;l i;j;k i;j
• j = n 6= k = m, i = l. j 6=l i6=k
Using the Kronecker delta to indicate nonzero terms when all its = T M (M 0 1)+ T M (T 0 1)+2T M
subscripts are equal we have = M T (M + T );

EX (trG3 P G)2 = pjk pmn EX [gji gki gml gnl ] as required.

j;k;m;n i;l
REFERENCES
= pjk pmn (2jkmn il + jkmn (1 0 il ) [1] G. J. Foschini, “Layered space-time architecture for wireless commu-
j;k;m;n i;l
0 0
nication in a fading environment when using multi-element antennas,”
+ jk mn (1 km )+ jn km (1 kn )il ) Bell Labs. Tech. J., vol. 1, no. 2, pp. 41–59, 1996.
= pjk pmn 2jkmn N + jkmn (N
2
0N ) [2] İ. E. Telatar, “Capacity of multi-antenna Gaussian channels,” Europ.
Trans. Telecommun., vol. 10, pp. 585–595, Nov. 1999.
j;k;m;n [3] L. Zheng and D. N. C. Tse, “Communication on the Grassman manifold:
+jk mn (1 0 km)N 2 + jn km (1 0 kn)N A geometric approach to the noncoherent multiple-antenna channel,”

pjj +(N 0 N )
2 2 2 2 IEEE Trans. Inform. Theory, vol. 48, pp. 359–383, Feb. 2002.
= 2N pjj + N pjj pmm [4] A. Lozano, A. Tulino, and S. Verdú, “Multiple-antenna capacity in
j j j;m the low-power regime,” IEEE Trans. Inform. Theory, vol. 49, pp.
0N 2 2
pjj + N pjk pkj 0N 2
pjj
2527–2544, Oct. 2003.
[5] T. L. Marzetta and B. M. Hochwald, “Capacity of a mobile multiple-
j j;k j antenna communication link in flat fading,” IEEE Trans. Inform. Theory,
2 vol. 45, pp. 139–157, Jan. 1999.
=N pjj pmm + N pjk pkj
[6] I. C. Abou-Faycal, M. D. Trott, and S. Shamai (Shitz), “The capacity
j;m j;k of discrete-time memoryless Rayleigh-fading channels,” IEEE Trans.
2 2 2
= N (trP ) + N trP ; Inform. Theory, vol. 47, pp. 1290–1301, May 2001.
[7] S. Verdú, “Spectral efficiency in the wideband regime,” IEEE Trans.
completing the proof. Inform. Theory, vol. 48, pp. 1319–1343, June 2002.
[8] İ. E. Telatar and D. N. C. Tse, “Capacity and mutual information of
The following lemma was used when considering a Gaussian mod- wideband multipath fading channels,” IEEE Trans. Inform. Theory, vol.
ulation scheme in Section IV-A. 46, pp. 1384–1400, July 2000.
[9] M. Médard and R. Gallager, “Bandwidth scaling for fading multipath
Lemma 3: Let S be a T 2 M matrix with independent zero-mean channels,” IEEE Trans. Inform. Theory, vol. 48, pp. 840–852, Apr, 2002.
unit-variance complex Gaussian entries. Then [10] V. Subramaniam and B. Hajek, “Broad-band fading channels: Signal

Etr(SS 3 )2 = M T (M + T ):
burstiness and capacity,” IEEE Trans. Inform. Theory, vol. 48, pp.
809–827, Apr. 2002.
2130 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 9, SEPTEMBER 2004

[11] M. Gursoy, H. V. Poor, and S. Verdú, “Efficient signaling for low-power which are essentially a generalization of the Alamouti scheme. How-
Rician fading channels,” in Proc. 40th Annu. Allerton Conf. Communi- ever, among the implications of employing multiple antennas is the as-
cation, Control and Computing, Oct. 2–4, 2002. sociated increase in the cost of the additional hardware required in the
[12] M. Godavarti and A. O. Hero, III, “Multiple-antenna capacity in a deter-
ministic Rician fading channel,” IEEE Trans. Inform. Theory, submitted form of radio frequency (RF) chains, in addition to the constraints im-
for publication. posed by the physical limitation of wireless devices.
[13] V. Prelov and S. Verdú, “Second-order asymptotics of mutual informa- In an effort to overcome these problems, while utilizing the advan-
tion,” IEEE Trans. Inform. Theory, vol. 50, pp. 1567–1580, Aug. 2004. tages of using multiple antennas, several papers have appeared recently
[14] B. Hajek and V. Subramaniam, “Capacity and reliability function for
small peak signal constraints,” IEEE Trans. Inform. Theory, vol. 48, pp.
in the literature in which the notion of antenna selection was intro-
828–839, Apr. 2002. duced for both STTCs and STBCs [4]–[18]. The idea behind antenna
[15] B. Hassibi and B. Hochwald, “How much training is needed in mul- selection is to use only a subset of the transmit and/or receive an-
tiple-antenna wireless links?,” IEEE Trans. Inform. Theory, vol. 49, pp. tennas in multiple-input multiple-output (MIMO) systems. In [4], the
951–963, Apr. 2003. authors consider the joint transmit and receive antenna selection based
on the second-order channel statistics, which is assumed to be avail-
able to the transmitter. The authors in [5] consider antenna selection
for low-rank matrix channels where selection is done only at the trans-
mitter. In [6], antenna selection is considered only at the transmitter
with the assumption that the channel statistics are available to the trans-
Performance Bounds for Space–Time Block Codes With mitter. In [7], Ghrayeb and Duman show that, for full-rank STTCs over
Receive Antenna Selection quasi-static fading channels, the diversity order is maintained as that of
the full-complexity system. In [8], Molisch et al. studied the effect of
Xiang Nian Zeng, Student Member, IEEE, and
antenna selection from a channel capacity perspective. It was shown
Ali Ghrayeb, Member, IEEE
that only a small loss in capacity is suffered when the receiver uses a
good subset of the available receive antennas. Other work related to
Abstract—In this correspondence, we present a comprehensive perfor- antenna selection for STTCs can be found in [9]–[15].
mance analysis of orthogonal space–time block codes (STBCs) with receive In [16], the authors consider antenna selection for STBCs at the
antenna selection. For a given number of receive antennas , we assume transmitter. They demonstrate through computer simulations that the
that the receiver uses the best of the available antennas, where, typ- performance is improved by increasing the number of transmit an-
ically, . The selected antennas are those that maximize the in-
stantaneous received signal-to-noise ratio (SNR). We derive explicit upper
tennas while keeping the number of selected antennas fixed. [17] con-
bounds on the bit-error rate (BER) performance of the above system for siders antenna selection for MIMO systems employing three transmit
any and , and for any number of transmit antennas. We show that and one receive antennas. They introduce an algorithm for selecting
the diversity order, with antenna selection, is maintained as that of the full the best two transmit antennas based on the channel statistics provided
complexity system, whereas the deterioration in SNR is upper-bounded by by the receiver. In [18], the authors consider antenna selection at the
10 log ( ) decibels. Furthermore, we derive a tighter upper bound
for the BER performance for any and when = 1, and derive an transmitter (with full knowledge of the channel statistics) or at the re-
expression for the exact BER performance for the Alamouti scheme when ceiver for orthogonal STBCs with particular emphasis on the Alamouti
= 1. We also present simulation results that validate our analysis. scheme [2]. In their analysis, they adopt a selection criterion that max-
Index Terms—Antenna selection, multiple-input multiple-output imizes the channel Frobenius norm and, accordingly, derive expres-
(MIMO) systems, performance bounds, space–time block codes (STBCs). sions for the improvement in the average signal-to-noise ratio (SNR)
and outage capacity. They use the outage probability analysis to argue
that the spatial diversity, when the underlying space–time code is or-
I. INTRODUCTION thogonal, is maintained with antenna selection. They do not, however,
In a wireless environment, unlike other applications, achieving reli- investigate the direct impact of antenna selection on the bit-error rate
able communication is much more challenging due to the possibility (BER) performance, which is the focus of this work.
that received signals from multipaths may add destructively, which, In [19], the authors propose a new scheme that involves using hybrid
consequently, results in a serious performance degradation. It has been selection/maximal-ratio transmission where the transmitter uses a good
shown that a key issue to achieving reliable wireless communication subset of the available antennas and the receiver uses maximum-ratio
is to employ spatially separated antennas at the transmitter and/or at combining. They investigate this scheme in terms of signal-to-noise
the receiver. To this end, several space–time coding schemes have been ratio (SNR), BER, and capacity. They demonstrate the effectiveness of
developed in recent years, including space–time trellis codes (STTCs) their scheme relative to already existing schemes. The same scheme
[1], a simple transmit diversity scheme developed by Alamouti [2], was also treated in [20] but when the transmitter selects the best single
and space–time block codes (STBCs) developed by Tarokh et al. [3], antenna. Other schemes that use hybrid selection/maximal-ratio com-
bining were also considered in [21]–[24]. A nice overview of antenna
selection for MIMO systems can be found in [25].
Manuscript received December 15, 2003; revised April 27, 2004. This work In this correspondence, we present a comprehensive performance
was supported in part by the Natural Sciences and Engineering Research analysis of STBCs with receive antenna selection. We limit our analysis
Council of Canada under Grant N00858, and by the National Science Foun-
dation under Grant INT-0217549. The material in this correspondence was to orthogonal STBCs simply because they are easy to design and they
presented in part at the IEEE International Symposium on Information Theory, achieve the maximum diversity order possible for a given number of
Chicago, IL, June/July 2004. transmit/receive antennas. In our analysis, we assume that, for a given
The authors are with the Department of Electrical and Computer Engi- number of receive antennas M , the receiver uses L out of the avail-
able M antennas where the selected antennas are those whose instan-
neering, Concordia University, Montreal, QC H3G 1MB, Canada (e-mail:
x_zeng@ece.concordia.ca; aghrayeb@ece.concordia.ca).
Communicated by R. R. Müller, Associate Editor for Communications. taneous SNRs are largest. This is achieved by comparing the sums of
Digital Object Identifier 10.1109/TIT.2004.833363 the magnitude squares of the fading coefficients at each receive antenna

0018-9448/04$20.00 © 2004 IEEE