1

Asymptotic Analysis of Reduced-Feedback

Strategies for MIMO Gaussian Broadcast
Channels
Jordi Diaz, Osvaldo Simeone, and Yeheskel Bar-Ness
Center for Wireless Communications and Signal Processing Research (CWCSPR)
New Jersey Institute of Technology
Newark, NJ 07102, USA
Email: {jordi.diaz,osvaldo.simeone,barness}@njit.edu

Abstract

Achieving the sum-capacity of a MIMO Gaussian broadcast channel is known to require full channel state
information at the base station, which implies the need of a large amount of feedback information from the users.
Different asymptotics of the sum-capacity, such as its scaling law with respect to the number of users n or the
multiplexing gain, are conventionally used to assess the performance of suboptimal schemes with reduced feedback,
or equivalently with partial channel state information at the transmitter.
In this paper, the optimal scaling law of the sum-rate with respect to n, for fixed signal-to-noise ratio, fixed
number of transmit antennas M and any number of receiving antennas N (i.e., M log log nN ), is proved to be
achievable with a deterministic feedback of only one bit per user. The proof builds on opportunistic beamforming and
binary quantization of the signal-to-noise-plus-interference ratio. Moreover, the amount of feedback is shown to be
further reduced with no asymptotic optimality loss by applying the selective feeback principle, leading to an average
feedback rate that scales as log n. Letting the number of antennas M grow with n, the linear scaling of the sum-rate
of the scheme with respect to M is demonstrated to hold only if M grows no faster than log n. The impact of adding
feedback bits is also investigated, showing relevant multiuser diversity gains. Finally, the asymptotic performance
with respect to signal-to-noise ratio is studied, by assessing how fast the number of users needs to increase with the
signal-to-noise ratio in order to guarantee a non-interference limited behavior.

I. I NTRODUCTION

Optimal performance on a Gaussian multi-antenna broadcast channel is by now known to be achieved by

techniques that require full Channel State Information (CSI) at the base station (BS). In particular, the capacity
region of the channel has been recently derived in [1], following earlier references [2][5], by exploiting a technique
proposed in [6] termed Dirty Paper Coding (DPC). Moreover, a simpler strategy based on zero-forcing beamforming
[2], has been proved to be optimal in terms of sum-capacity in the limit of a large number of users n in [7]. Both

October 9, 2006 DRAFT

 2

DPC and zero-forcing beamforming require full CSI, which implies the need for a feedback channel of sufficient
capacity (if Frequency Division Duplex is employed). This fact has motivated intensive research on reduced feedback
techniques which can still exploit the multiuser diversity [8] of the broadcast channel. One line of work attempts
to reduce the amount of feedback by means of quantization of CSI parameters [9][17]. An alternative approach,
referred to as selective feedback, allows a user to feed back information or not depending on its current channel
conditions, and thus reduces the overall amount of feedback on the average [17][19].
Unlike single-user MIMO systems, where optimal singular-value decomposition and water-filling power allocation
based on full CSI is known to provide marginal gains over CSI-unaware transmission schemes [20], multi-user
MIMO broadcast channels are known to be very sensitive to the availability of CSI [21]. Reduced feedback schemes
are generally designed so as to minimize the impact on the sum-rate. In particular, the conventional benchmark
is the ability of limited feedback techniques to preserve optimal asymptotic properties, such as scaling law of the
sum-rate with respect to the number of users n (see, e.g. [7], [13], [22][24]) and the asymptotics of the sum-rate
with respect to the signal-to-noise ratio (SNR) [11].
For fixed SNR, fixed number of transmit antennas M and any number of receiving antennas (per user) N, the
asymptotic measure of interest is the scaling law of the sum-rate. It is shown in [7] that the optimal value of
the scaling law, namely M log log nN, can be achieved by a zero-forcing beamforming technique with appropriate
scheduling. Moreover, [22] proved that the asymptotic optimality of zero-forcing beamforming can be preserved
even if the transmitter only knows the singular vectors corresponding to the largest singular value of the channel
matrices for each user. The scaling law is also considered in [23], where an opportunistic scheme that prescribes
transmission on a random set of M orthogonal beams to M users is investigated. Each user measures the signal-
to-noise-plus-interference ratio (SINR) on the M beams and feeds back only the best SINR and the corresponding
beam index (N real and N integer numbers). In spite of the significant reduction of feedback, the scheme in [23]
is proved to achieve the optimal scaling law of M log log nN .
The growth of sum-rate with respect to the SNR, fixing the other parameters (i.e., multiplexing gain) of limited
feedback techniques for the multiantenna Gaussian channel has been addressed in [11], focusing the analysis on
zero-forcing beamforming techniques. It is shown therein that the number of feedback bits per user has to grow
linearly with the number of transmit antennas M and logarithmically with the SNR, in order to retain the optimal
multiplexing gain of M . If the condition on the feedback bits is not satisfied, the system shows to be interference-
limited and the multiplexing gain can be as low as 1, reducing the sum-capacity down by a factor of M .

A. Main contributions

In this paper, we first show that the optimal scaling law M log log n of the sum-rate of a Gaussian broadcast
multi-antenna channel (with N = 1) can be achieved with only one bit feedback per user (Sec. III, Theorem 1). The
transmission scheme used to prove this result is similar to [23] in that it applies the the opportunistic beamforming
principle [25] to a set of orthogonal beams. However, each user, instead of feeding back the SINR for the best
beam (one real plus one integer number), only transmits one bit to the BS, indicating whether or not the SINR

October 9, 2006 DRAFT

 3

on a pre-selected beam is above a given threshold. The BS then schedules for transmission on each beam a user
randomly selected among the ones that have signaled the corresponding SINR to be above the threshold. The
scheme is similar to the one proposed in [13], but the latter only employs one beam at each time, proving to be
unable to attain the optimal scaling law.
Allowing the number of transmit antenna M to grow with the number of users n, [23] showed that in order to
maintain the linear scaling law of the sum-rate with M, the number of transmit antennas is constrained to increase
no faster than log n. Here we prove that such a growth rate of M yields a linear scaling of the sum-rate even for
the one bit feedback scheme (Sec. IV, Theorem 2). Extension of the results discussed above to the scenario where
users are provided with multiple receiving antennas N > 1 is discussed in Sec. V, where it is proved (Theorem 3)
that one bit of feedback per user is enough to guarantee the optimal scaling law of M log log nN for any N .
Even though the one-bit protocol explained above presents, asymptotically in the number of users n, the same
sum-rate growth of the optimal scheme, performance benefits for finite n are expected by increasing the number
of feedback bits. The threshold-based scheme discussed above is therefore extended to accommodate 1 b
log2 (M + 1) bits per user on the uplink channel and the corresponding advantages are investigated (Sec. VI).
If the number of feedback bits is allowed to be a random variable1 , the optimal scaling law M log log n can
be guaranteed by using an average number of feedback bits that scales as log n with the number of users n (see
Sec.VII, Theorem 4). In this case, similarly to the selective feedback principle [19] [18], the transmission scheme
that is proved to attain the stated result works as described above with the only modification that no signal is fed
back by a user if there is no beam whose SINR crosses the threshold.
Finally, in Sec. VIII we let the SNR grow to infinity with the number of users n, and ask how fast n needs
to scale so as to guarantee a non-interference limited behavior. It is proved that in order for the one-bit feedback
scheme to guarantee the optimal scaling law of M of the sum-rate, the number of users has to scale proportionally
to a power of the SNR. A connection with the result in [11] concerning the multiplexing gain of zero-forcing
beamforming can be drawn to support our conclusion.

II. M ULTI - BEAM TRANSMISSION WITH ONE - BIT FEEDBACK PER USER

We consider a multi-antenna downlink system with n users as depicted in fig. 1. The BS employs a set of M
random orthonormal beams U = {u1 , . . . , uM }, generated from an isotropic distribution [25] [26], and constructs
the transmitted signal x as
M
X
x= um sm , (1)
m=1

where we assume that sm s are letters from a Gaussian codebook. An ergodic block-fading channel model is
assumed. In particular, the M 1 channel vector hi , modelling the flat-fading link between transmitting array
at the ith user, is constant within each block and presents ergodic variations across blocks. Moreover, the entries
of channel hi are independent identically distributed zero mean circularly symmetric complex Gaussian with unit

1 by, e.g., using a contention channel for the uplink feedback.

October 9, 2006 DRAFT

 4

s1 s2 s M

u1 u2 ::: uM

Transmitter x
Channel

h1 h2 ::: hn

w1 w2 w n

y1 y2 n
y

Fig. 1. Block diagram of a multi-antenna Gaussian broadcast channel.

variance CN (0, 1), and independent among different users. The signal received by the ith user in a given block is
given by

yi = hTi x + wi , (2)
2
where the transmitted vector x C M 1 has a power constraint E[kxk ] = M and the AWGN wi has unit variance
CN (0, 1). Hence, the average SNR at the receiver is M . Notice that here we are assuming that each user is
equipped with N = 1 receiving antennas; the multi-antenna case will be discussed in Sec. V. The channel vector
hi is assumed to be perfectly known at the ith receiver. On the contrary, the base station has only a partial knowledge
of the channels {hi }ni=1 . This knowledge is achieved through the reception of one bit of feedback from each user
at the beginning of each coherence block. This is further elaborated below.
Let us assume that the rth beam is intended for user i. The received signal in (2) may now be restated as
M
X T
yi = hTi ur sr + hi um sm + wi , (3)
m6=r

and the corresponding SINR for the ith user on the rth beam is given by
|hTi ur |2
Si,r = M
. (4)
|hTi um |2
P
+ 1/
m=1
m6=r

A. Scheduling

The scheduling of transmission onto the M beams is carried out at the BS aided by one bit of feedback per
user for each coherence block. Each user, say the ith, measures the SINR only on one beam, say the rth (Si,r ),

October 9, 2006 DRAFT

 5

previously assigned to it by the BS in a pseudo-random fashion. This procedure differs from the scheduling in
[13], [17], [23], where each user measures the SINR on all beams. The SINR measurement Si,r , is compared to a
given threshold , which is a network parameter known by the BS and all users. One bit of feedback from user
i informs the BS on whether or not Si,r is above the threshold (Si,r > ). After receiving feedback from all
users, the BS schedules for each beam one user picked randomly among those who have signaled the SINR on the
corresponding beam to be above the threshold. Notice that there is a small probability that a certain beam is not
requested by any user, i.e., that no user measures a strong enough SINR on the beam. In this event, we assume
that the BS communicates through the unrequested beam to a user picked randomly from the entire set of users.
This assumption ensures that equation (4) holds at all times, which is mathematically convenient for our analysis.
A codeword is assumed to span multiple coherence blocks. However, as further discussed in Sec. III, our main
results hold even in the case of block-by-block coding (i.e., each codework spans only one block).

B. Sum-rate

Let Pm be the probability that beam m is requested, i.e., that at least one user measures a SINR above the
threshold on that beam. Because of the symmetry of the setup Pm is equal for all beams and henceforth the index
is dropped. The sum-rate of the multi-beam scheme with one bit feedback described above satisfies the following.
Lemma 1: Let P be the probability of any beam to be requested (by at least one user) and be the pre-determined
threshold. Then, the achievable sum-rate of the multi-beam transmission scheme with one-bit feedback is lower
bounded by
R1 M P E[log(1 + S)|S > ], (5)

where
n

P = 1 (F ()) M , (6)

and F (x) is the cumulative distribution function of the SINR (4) ( [23], see also Appendix-A)
ex/
F (x) = 1 . (7)
(1 + x)M 1
Proof of Lemma 1 is provided in Appendix-A. For the case of M = 1, (5) reduces to the expression obtained in
[27]. As elaborated in the following sections, the key parameter in (5) is the threshold : fixing the other parameters,
increasing the threshold is expected to decrease the probability P of the beam being requested, that represents the
pre-log term in (5), and, on the other hand, increase the rate E[log(1 + S)|S > ]. Therefore, an appropriate
selection of the threshold is of crucial importance in order to obtain the optimum performance.

III. S CALING LAW OF THE SUM - RATE WITH ONE - BIT FEEDBACK AND FIXED M

In this section we show that the optimal scaling law of the sum-rate achieved by DPC with full CSI is also
attainable by the multi-beam transmission scheme with only one bit of feedback per user described in section II.

October 9, 2006 DRAFT

 6

Theorem 1: Let M and be fixed. The sum-rate R1 of the multi-beam transmission scheme with one bit feedback
per user, satisfies
R1
lim = 1. (8)
n M log log n
Proof: According to Lemma 1, the sum-rate is bounded by (5), which can be further lower bounded by
substituting S by its minimum value , i.e., exploiting the inequality E [log (1 + S) |S > ] log(1 + ):
n

R1 M 1 (F ()) M log (1 + ) = Rlb . (9)

If we choose the threshold as follows

= log n M log log n, (10)

then it is easy to verify that

n n
e/
 M  M
log n
lim 1 = lim 1 (M 1) = 0. (11)
n (1 + )M 1 n n
As a result, the pre-log term of Rlb tends to M when n goes to infinity. Therefore, the asymptotic behavior (with
respect to n) of the lower bound in (9) coincides with the upper bound set by the performance of DPC with full
CSI:
Rlb R1
lim = lim = 1, (12)
n M log log n n M log log n

thus concluding the proof.

Remark 1: The proof of Theorem 1 is based on the lower bound Rlb (9) on the achievable rate of the one-bit
scheme. It is interesting to note that this lower bound can be attained without incurring in the delay implicit in the
scheme presented in Sec. II, which requires coding across multiple blocks. In fact, the rate Rlb can be achieved
through block-by-block coding by simply transmitting to the selected users in each block with the minimum
guaranteed rate log (1 + ) . This implies that the asymptotic result (8) holds even if we limit each codeword to
span one coherence time. The same conclusion applies to the other asymptotic results presented in the following
(see Theorem 2-5). However, numerical results throughout this correspondence will refer to rate (5), which requires
multi-block coding.
Remark 2: In a one-bit feedback scheme, transmission errors on the feedback channel may be unlikely. On the
contrary, outdated feedback due to a delay between the measurement of the feedback bit and the channel access
by the base station can be a source of performance degradation. In any case, the effect of either feedback errors or
outdated feedback can be modelled as a probability of error Pe . The latter measures the event where the received
feedback bit is positive even though the current signal-to-noise ratio is below the threshold. While the effect of
such events on the rate (5) is not easy to assess, it is apparent that the lower bound Rlb (9) employed in the proof
of Theorem 1 is simply multiplied by a constant term (1 Pe ). This implies that, in the presence of a feedback
error, the scaling law of the rate is lower bounded by (1 Pe ) M log log n, thus showing the optimal trend with
increasing number of users n (unless Pe = 1). Further investigation on the effect of outdated feedback can be found
in [28] for a single antenna system (M = 1).

October 9, 2006 DRAFT

 7

IV. S CALING LAW OF M TO MAINTAIN LINEAR SCALING OF SUM - RATE

So far the number of transmit antennas M has been assumed to be fixed. Here, we consider the case where M
is allowed to grow with the number of users n. In [23] the authors show that the throughput of their scheme scales
M
linearly with M as long as M grows no faster than log n. In particular, it is proved that if limn log n =
M
then the ratio of the achievable throughput to M tends to zero. However, if limn log n is a constant, then the
sum-rate scales linearly with M. Since the proposed multi-beam scheme with one bit feedback yields an achievable
M
rate that is upper bounded by the performance of the scheme in [23], it is clear that if limn log n = , then
limn R
M
1
= 0. Moreover, in this section we show that if M
limn log n is a constant, then the sum-rate with one
bit feedback scales linearly with M as in [23]. This result is relevant not only for infinite number of users, but
also for any large n, because it suggests that given n there should exist a value for M on the order of log n that
maximizes the sum rate2 .
Theorem 2: Let be fixed. If M is given by
log n 2 log log n c/
M= + 1, (13)
log (1 + c)
where c is a positive constant, then the sum-rate of the multi-beam transmission with one bit feedback satisfies
R1
lim = 1. (14)
n M log(1 + c)
Proof: Consider the lower bound of the sum-rate in (9) and set the threshold = c. Clearly, we have
n
ec/
 M
Rlb
lim = lim 1 1 =
n M log(1 + c) n (1 + c)M 1
 c (M 1) log(1+c) 
n
log 1e
= lim 1 e M . (15)
n

Therefore, in order to show (14), we have to prove that

n  c

lim log 1 e (M 1) log(1+c) = . (16)
n M

For our choice of M in (13), the previous limit becomes

(log n)2
 
n
lim log 1 , (17)
n log n2 log log nc/ +1 n
log 1+c

and using the equality log(1 x) = x + O x2 for small x, we can simplify (17) as

n (log n)2
lim
n log n2 log log nc/+1 n
log 1+c

(log n)4
 
n
+ lim log n2 log log nc/ O , (18)
n + 1 n2
log 1+c

2 Notice that transmission schemes that perform antenna selection and/or power allocation based on available channel state information are
able to achieve a non-decreasing sum-rate with the number of transmit antennas [29]. The result in Theorem 2, similarly to [23], depends on
the assumption of very limited feedback, that hinders further system optimization.

October 9, 2006 DRAFT

 8

13

12

11

10

sumrate [b/s/hz]
8

5
=10db, opportunistic scheme [23]
4 =10db, 1 bit feedback
=0db, opportunistic scheme [23]
=0db, 1 bit feedback
3
1 2 3 4 5 6 7 8 9 10
M

Fig. 2. Sum-rate of the one-bit scheme and the opportunistic scheme in [23], as a function of the number of transmit antennas M for n = 1000
users.

but the second term in (18) is 0 and the first term is clearly . Hence, (16) and (15) are proved. Again, the
lower bound in (15) coincides with the upper bound of R1 set by the sum-capacity, which allows to conclude the
proof of (14).
It is important to remark the even though the growth of M in our scheme is asymptotically the same as in [23],
log n+3 log log nc/
there is a difference between the two techniques. It is shown in [23] that choosing M = log 1+c +1
allows the scheme proposed therein to retain linear growth in sum-rate with M . However, it is easy to verify that
for the same choice of M the ratio of R1 /M tends to 0 in our scheme. Therefore, the number of transmit antennas
effectively supported by the method in [23] is larger than with the one bit scheme.
The conclusion above leads to the conjecture that, even with a finite number of users, the optimal number of
transmit antennas M for the technique in [23] is larger than with the one-bit scheme. Fig. 2 confirms this result,
by showing the sum-rate of the two methods as a function of the number of antennas for n = 1000. Notice that
for a finite number of users n, the reduction in sum-rate due to the limited feedback as compared to the scheme
in [23] is relevant. In Sec. VI, it is demonstrated that increasing the number of feedback bits per users allows to
partially fill this gap [23].

V. E XTENSION TO N RECEIVE ANTENNAS

When user terminals are equipped with any number N of receive antennas (experiencing uncorrelated fading), the
sum-rate of DPC scales as M log log nN [30]. In this section we show that the same scaling law can be achieved
with only one-bit feedback. The scheme studied in the previous sections is modified as follows. Treating each

October 9, 2006 DRAFT

 9

antenna independently, the SINR from beam m at the jth antenna of user i is given by
|hTi,j ur |2
Si,r,j = M
. (19)
T 2
P
1/ + |hi,j um |
m=1
m6=r

where hi,j is the channel vector seen at the jth antenna of user i. A given user only measures the SINR of the rth
transmitted beam for all its receive antennas, and feeds back one bit indicating whether or not Si,r = maxj Si,r,j is
larger than the threshold . Scheduling then is carried out as explained in Sec. II-A. The scheduled user on beam
r (say the ith) processes only the signal received on the antenna at which Si,r is measured.
Theorem 3: Let M and be fixed. The sum-rate of a MIMO Gaussian broadcast channel with one-bit feedback
per user satisfies for any N
R1
lim = 1. (20)
n M log log nN
Proof: The result in Lemma 1 can be easily extended to the scenario at hand by noticing that the probability
nN
of a given beam r being requested is now P = 1 (F ()) M . That leads to a lower bound of the sum-rate given
by (see Appendix-B for further details)
 nN

R1 M 1 (F ()) M E [log (1 + S) |S > ] . (21)

Notice that (21) is exactly (5) if a change of variable n by nN is made. Therefore, the rest of the proof follows
the same steps of the proof of Theorem 1.
In interpreting the result of Theorem 3, it should be noticed that the opportunistic scheme of [23] that guarantees
the optimal scaling law requires the feedback of N real and N integer numbers per user. Moreover, from (21), it
is seen that multiple antennas at the receiver have the same effect on the sum rate as multiplying the number of
users by a factor of N , which can be interpreted as a multiuser diversity gain. Finally, following the analysis in
Sec. IV, it can be easily shown that this result implies a linear scaling of the sum-rate with respect to M for M
growing no faster than log nN .
From Theorems 1 and 3, it is clear that asymptotic optimality in terms of scaling law of the sum-rate with
respect to the number of users n can be guaranteed with a very small amount of feedback (that can be even further
reduced as explained in Sec. VII). Next section shows that even though the first bit of feedback captures most of the
multiuser diversity, additional bits of feedback are very valuable because they yield significant gains in sum-rate for
the regime of a finite number of users. In the following sections, we focus on the case N = 1 with the understanding
that extension of a MIMO scenario follows the guidelines explained in this section.

VI. I MPACT OF INCREASING FEEDBACK BITS ON SUM - RATE

As discussed in the previous sections, one bit of feedback is enough to guarantee the optimal scaling law of
the sum-rate for large n. Here we want to quantify the impact of increasing the feedback to b > 1 bits (and
b log2 (K + 1)). A way to increase the number of bits is to modify the transmission scheme described in
Sec. II-A by allowing each user to measure the SINR on more than one pre-determined beam. In particular, if

October 9, 2006 DRAFT

 10

18

16

14

sumrate (b/s/hz)
12

10

8
Opportunistic scheme [23]
b=3
b=2
b=1
6
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
number of users n

Fig. 3. Sum-rate with b = 1, 2, 3 bits of feedback per user and sum-rate of the opportunistic scheme in [23] (M = 7, N = 5, = 10dB).

b = log2 (K + 1) bits are granted for any user in the uplink channel, the SINRs on K beams can be measured by
each user. Then, the index of the beam with the maximum SINR, provided that it crosses the threshold, is fed back
to the BS (the all-zero string is sent if no beam crosses the threshold). Scheduling on each beam is carried out at
the BS by selecting randomly a user among the ones that signaled the corresponding index.
As shown in Appendix-C, if > 1, the result in Lemma 1 can be generalized to this scenario, leading to the
following lower bound on the rate of the proposed scheme with b bits of feedback (b = log2 (K + 1)):
 nK

RK M 1 (F ()) M E[log(1 + S)|S > ] > 1. (22)

Since RK with K > 1 is lower bounded by the sum-rate with one-bit feedback R1 (5), the optimal scaling law
M log log n clearly holds also for RK . In [17], this result was proved for the case of K = M , showing that the
constraint > 1 in the sum rate expression (22) does not constitute an impediment for the proof.
The impact of increasing number of feedback bits is depicted in fig. 3, where the lower bound on the sum-rate
(22) is evaluated versus the number of users n for a threshold obtained through numerical maximization of
(22). The optimal threshold is shown in fig. 4. Moreover, as a performance reference, the sum-rate of the scheme
proposed in [23] is plotted as well. It is seen that the first bit of feedback captures most of the multiuser diversity
because the growth of the sum-rate with one-bit feedback resembles that of the scheme proposed in [23]. However,
a significant gap in sum-rate exists between the two schemes. This gap is partially bridged by the second and the
third bit of feedback, with decreasing gain for increasing b.
The gap identified above attests to the limitation of asymptotic criteria as measures of system performance. As
shown in fig. 3, an optimal scaling law of the sum-rate might be attained even if a very large and impractical
number of users n is actually needed to reach optimality of the sum-rate. Reference [7] further elaborates on this
point by comparing the sum-rate of the opportunistic scheme of [23] with the performance of Dirty Paper Coding.

October 9, 2006 DRAFT

 11

4
b=3
b=2
b=1
3.5

Threshold
2.5

1.5

0.5
2 3 4
10 10 10
number of users n

Fig. 4. Optimal threshold as a function of the number of users for b = 1, 2, 3 bits of feedback (M = 7, N = 5, = 10dB).

VII. S ELECTIVE F EEDBACK

Previous sections have dealt with deterministic feedback schemes, where users communicate b bits (1 b
log2 (M +1)) at each channel coherence period. In order to further reduce the feedback, in this section we consider an
alternative approach, based on the selective feedback principle [19]. Accordingly, each users feeds back information
to the BS only when the SINR is above the threshold. As a result, the feedback load is reduced on the average in
two ways: (i) Each feedback transmission is reduced from log2 (K + 1) to max(log2 (K), 1) bits, because the new
alphabet does not contain the message no beam over the threshold3 ; (ii) users with no beam above the threshold
do not transmit at all. Apart from the feedback generation, scheduling and transmission are carried out as explained
in the previous sections. Therefore, feedback reduction due to the selective approach does not imply any loss in
performance as compared to the deterministic feedback case.
A remark on the implementation of the selective feedback scheme is in order. Requiring the feedback of a random
number of bits, the scheme cannot rely on a fixed (e.g., time or frequency) resource allocation on the feedback
channel. On the contrary, some variant of a contention channel with random access is necessary. In the result stated
in Theorem 4, we ignore the loss in spectral efficiency on the feedback channel inherent in random access schemes.
Therefore, the amount of feedback predicted in Theorem 4 has to be interpreted as a lower bound on the amount
of feedback necessary for the one-bit scheme with selective feedback.
Let us define as bn the average number of feedback bits per cell (dependence on the number of users n is made
explicit). We clearly have bn = n b, where b = max(log2 (K), 1) and n is the average number of users employing
the feedback channel. The following theorem quantifies the asymptotic behavior of the amount of feedback bn that
is necessary in order to guarantee the optimal scaling law of the sum-rate.

3 The max() operator is necessary in order to cope with the case K = 1.

October 9, 2006 DRAFT

 12

Theorem 4: Let M and be fixed and N = 1. The sum-rate scaling law M log log n is attainable with a an
average feedback per cell that satisfies
bn 1
lim = M 1 . (23)
n bK log n
Proof: The average number of users employing the feedback channel, for > 1, is given by

n = nK(1 F ()), (24)

since the probability that a user feeds back information (i.e., measures the assigned beams above the threshold )
is (1 F ()). Recalling (7), we have
Ke
n = n . (25)
(1 + )M 1
Let be chosen as in (10) so as to guarantee the optimal scaling law of the sum-rate. Then, the limit in (23) is
easily verified:

bn n e
lim = lim (26)
n bK log n n log(n) (1 + )M 1

n log n 1
= lim = M 1 (27)
n log(n) nM 1

Corollary 1: Let gn = bn /n be the average feedback per user. A direct consequence of Theorem 4 is that

lim gn = 0. (28)
n
Hence, the average number of feedback bits per user approaches zero as n grows to infinity. This result can
be intuitively explained by noticing that in order to fully capitalize on multiuser diversity, it is enough that only
the users with the best SINR feed back information to the BS. Therefore, the fraction of users using the feedback
channel vanishes for an asymptotically large number of users.

VIII. S CALING LAW OF n TO GUARANTEE NON - INTERFENCE LIMITED BEHAVIOR

Subject of the previous sections has been the asymptotic behavior of the sum-rate of limited feedback transmission
schemes for increasing number of users n and fixed SNR . Here, we focus on the asymptotic performance of such
schemes with respect to the SNR. It is known that, for fixed M and n, opportunistic transmission techniques are
interference-limited, that is the rate R satisfies R/ log2 () 0 for [21]. Notice that, fixing M and n, the
amount of feedback per cell for the opportunistic schemes proposed in both [23] and in this paper is independent
of the SNR. In the following theorem we find a sufficient condition for the growth rate of the number of users n
with the SNR that allows to guarantee the optimal scaling of the sum-rate of M .
Theorem 5: Consider the multi-beam transmission scheme described in Sec. VI, where M is fixed and K = M .
If the number of users is n = M , then
RM
lim = M. (29)
log ()
This implies that the optimal scaling of M is guaranteed and that the system is not interference limited.

October 9, 2006 DRAFT

 13

Proof: The sum-rate is lower bounded for > 1 by (22), which is further lower bounded substituting S by
its lowest possible value . Now, choosing = log M log log , which implies > 1 for large enough,
we have  

log(1 + ) log log (log )M
lim = lim 1 + =1 (30)
log () log
and for the pre-log term
 nK
e/
 M

lim 1 =
(1 + )M 1
KMM
(log )M 1
lim 1 =

M 1

 

M 1 log (log )M

  KM
log M
lim 1 M = 0, (31)

thus concluding the proof.
Corollary 2: The multi-beam scheme with selective feedback described in Sec. VII preserves the optimal scaling
(29) with an average number of feedback bits per cell that satisfies
bn
lim = M log2 (M ) (32)
log()
Proof: According to (24), the average number of feedback bits reads

bn = nM (1 F ()) log2 M (33)

From Theorem 5 a number of users n = M guarantees the optimal scaling law. Moreover, from equation (31) we
have that lim (1 F ()) = (log )/M . Substituting these results into (33) concludes the proof.
The total feedback required in (32) is proportional to M log2 M and the logarithm of the SNR. This result is
reminiscent of the one proved in [11] in the context of zero-forcing beamforming with feedback of quantized CSI.
Therein, it is shown that, fixing the SNR and the number of users n, in order to avoid interference-limited behavior,
the feedback load must approximately scale linearly with M and the SNR (in dB). Notice that, while [11] evaluates
the multiplexing gain (i.e., scaling of the sum-rate with SNR for fixed parameters), here we let the number of users
grow as well. In fig. 5, the sum-rate of the selective feedback scheme with K = M is plotted versus the SNR for
different number of antennas M and users n. It can be seen that for the same number of users, deploying more
transmit antennas is advantageous at low SNR but causes a more pronounced performance floor due to interference
at high SNR.

IX. C ONCLUSION

Asymptotics of the sum-rate have been used in order to assess analytically the performance of transmission
schemes based on limited feedback for the downlink of a MIMO Gaussian channel. In this paper, the scaling law of
the sum-rate with respect to the number of users has been considered at first, showing that optimality in this regard

October 9, 2006 DRAFT

 14

11
M=3, n=1000
M=2, n=1000
10 M=3, n=100
M=2, n=100

sumrate b/s/hz
6

1
0 5 10 15 20 25 30 35
SNR dB

Fig. 5. Sum-rate of the selective scheme with K = M versus SN R for M = 2, 3 and n = 100, 1000.

can be guaranteed with a smaller feedback rate than previously reported. In particular, a deterministic feedback of
one bit per user or an even smaller average number of bits per users if selective feedback is employed have been
proved to be enough. Scaling law of the number of required feedback bits in the latter case of selective feedback has
been derived as well. Moreover, the impact of an increase of feedback bits has been evaluated for a finite number
of users. Finally, the scaling of the sum-rate with respect to the SNR has been addressed. In accordance to recently
reported results, it has been demonstrated that a feedback rate that scales with the SNR (in dB) is sufficient to
guarantee the optimal scaling and thus to prevent an interference-limited behavior.

X. A PPENDIX

A. Proof of (7)

For completeness, here we provide a brief proof of (7), adapted from [23]. In the expression of the SINR Si,r (4),
the numerator Z = |hTi ur |2 is easily seen to be a chi-square distributed random variable with 2 degrees of freedom,
M
|hTi um |2 has the same distribution with 2(M 1) degrees of
P
while the first term at the denominator Y =
m=1,m6=r
freedom. Therefore, from the law of total probability, the probability density function (pdf) of Si,r = Z/(Y + 1/)
reads
Z
ex/
 
1
f (x) = f (x|y)fY (y)dy = (1 + x) + M 1 . (34)
(1 + x)M
0
where we used the fact that the pdf of Si,r conditioned on Y, f (x|y), and the pdf of Y are easily obtained from the
Rx
discussion above. The cumulative distribution function (7) is then obtained by using the definition F (x) = f (t)dt.
0

B. Proof of Lemma 1

The achievable sum-rate of the multi-beam scheme with one bit feedback described in Sec. II is
XM X n
R1 Ri,m (35)
m=1 i=1

October 9, 2006 DRAFT

 15

where Ri,m is the rate to user i through beam m that is achievable by coding only during the scheduling intervals
where the ith user has actually requested the beam (i.e., when Si.m > ). Expression (35) is only a lower bound
on the actual achievable rate since it neglects the time instants where user i is scheduled on a beam that no user
has requested (recall the description of the scheduling process in Sec. II-A). Notice that the probability that no
user measures a strong enough SINR on a given beam (for an appropriately selected threshold ) is expected to
be small for a large number of users n.
The rate Ri,m reads
Ri,m = m,i E[log(1 + Si,m )|Si,m > ], (36)

where m,i is the percentage of time in which the BS schedules user i to beam m (provided that the user has
requested the beam). The SINR Si,m is identically distributed for all users and beams, henceforth the indexes are
dropped. Then, (35) becomes
M X
X n
R1 E[log(1 + S)|S > ] m,i , (37)
m=1 i=1

where the inner sum term is clearly the probability that a given beam m is requested (by at least one user):
n
X
m,i = Pm . (38)
i=1

By symmetry Pm = P , thus substituting (38) in (37) leads to the sum-rate expression in (5). Moreover, the
probability that any given beam is requested reads

P = 1 (P1 )L ,


(39)

where L is the number of users measuring the SINR on the beam, and P1 is the probability that a user does not
request transmission after measuring the beam (i.e., the probability of the event S < ): P1 = F (). Assuming
that the number of users measuring each beam is the same, we have L = n/M and (6), thus completing the proof.

C. Proof of (21)

Since the scheduling processes is the same as for the N = 1 case studied above, here only the main differences
are discussed In particular, the probability that each beam is requested still follows (39) with L = n/M since we
have n/M users measuring each beam. However, the probability that a user does not request transmission is here
P1 = F ()N because the SINR measured by each users on different receive antennas are independent. The rest of
the proof follows the arguments in Appendix-A leading to (21).

D. Proof of (22)

According to the scheme described in Sec. VI, the number of users measuring a certain beam is nK/M . Therefore,
the probability P of a beam being used follows (39), whereas P1 is given by

P1 = P{S < } + P{(S > ) (S is not the maximum)}. (40)

October 9, 2006 DRAFT

 16

For > 1, P{(S > ) (S is not the maximum)} = 0, because the SINR can not be greater than one in more
than one beam [23]. Therefore, P1 = F () and
nK
P = 1 (F ()) M > 1. (41)

Therefore, following the same reasoning as in the previous sections, we finally get (22).

ACKNOWLEDGMENT

This work is partially supported by Samsung Electronics Co. LTD.

R EFERENCES

[1] H. Weingarten, Y. Steinbert, and S. Shamai (Shitz), The capacity region of the Gaussian MIMO broadcast channel, in Proc. of IEEE
Int. Symp. Inform. Theory, Chicago, IL, June 2004, p. 174.
[2] G. Caire and S. Shamai (Shitz), On the achievable throughput of a multiantenna gaussian broadcast channel, IEEE Trans. Inform. Theory,
vol. 49, no. 7, pp. 16911706, Jul 2003.
[3] S. Vishwanath, N. Jindal, and A. Goldsmith, Duality, achievable rates and sum rate capacity of Gaussian MIMO broadcast channels,
IEEE Trans. Inform. Theory, vol. 49, no. 10, pp. 26582668, October 2003.
[4] P. Viswanath and D. N. C. Tse, Sum capacity of the vector Gaussian broadcast channel and downlink-uplink duality, IEEE Trans. Inform.
Theory, vol. 49, no. 8, pp. 19121921, August 2003.
[5] W. Yu and J. Cioffi, Sum capacity of Gaussian vector broadcast channels, IEEE Trans. Inform. Theory, vol. 50, no. 9, pp. 18771892,
Sept. 2004.
[6] M. Costa, Writing on dirty paper, IEEE Trans. Inform. Theory, vol. IT-29, pp. 439441, May 1983.
[7] T. Yoo and A. Goldsmith, On the optimality of multi-antenna broadcast scheduling using zero-forcing beamforming, IEEE J. on Sel.
Areas Commun., vol. 24, no. 3, pp. 528 541, March 2006.
[8] R. Knopp and P. Humblet, Information capacity and power control in single cell multiuser communications, in Proc. of IEEE ICC,
vol. 1, Seattle, WA, June 1995, pp. 331335.
[9] W. Santipach and M. L. Honig, Asymptotic capacity of beamforming with limited feedback, in Proc. ISIT, Chicago, IL, June 2004, p.
290.
[10] , Achievable rates for MIMO fading channels with limited feedback and linear receivers, in Proc. ISSSTA, Sydney, Australia,
Aug./Sep. 2004, pp. 16.
[11] N. Jindal, MIMO broadcast channels with finite feedback, in Proc. of IEEE Global Telecomunications Conference, St. Louis, MO,
November 2005.
[12] S. Sanayei and A. Nosratinia, Exploiting multiuser diversity with only 1-bit feedback, in Proc. IEEE WCNC, vol. 2, March 2005, pp.
978 983.
[13] , Opportunistic beamforming with limited feedback, in Proc. Asilomar conference on signals, systems, and computers, November
2005.
[14] K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, On beamforming with finite rate feedback in mulitple-antenna systems, IEEE
Trans. Inform. Theory, vol. 49, no. 10, pp. 25622579, Oct. 2003.
[15] D. Love, R. Heath, W. Santipach, and M. Honig, What is the value of limited feedback for mimo channels? IEEE Commun. Mag.,
vol. 42, no. 10, pp. 5459, Oct. 2004.
[16] V. Hassel, M.-S. Alouini, D. Gesbert, and G. E. Oien, Exploiting multiuser diversity using multiple feedback thresholds, in Proc. of
IEEE VTC, 2005.
[17] J. Diaz, O. Simeone, O. Somekh, and Y. Bar-Ness, Scaling law of the sum-rate for multi-antenna broadcast channels with deterministic
or selective binary feedback, in Proc. Inform. Theory Workshop, Punta del Este, Uruguay, March 2006, pp. 298301.
[18] D. Gesbert and M. S. Alouini, Selective multi-user diversity, in Proc. ISSPIT, Dec. 2003, pp. 162165.
[19] , How much feedback is multi-user diversity really worth? in Proc. of ICC, Paris, France, June 2004, pp. 234238.

October 9, 2006 DRAFT

 17

[20] E. Telatar, Capacity of multi-antenna Gaussian channel, Europ. Trans. Telecommun., vol. 10, pp. 585595, Nov. 1999.
[21] A. Lapidoth, S. Shamai (shitz), and M. A. Wigger, On the capacity of fading mimo broadcast channels with imperfect transmitter
side-information, in Proc. Allerton Conference on Commun., Control and Computing, 2005.
[22] A. Bayesteh and A. K. Khandani, On the user selection for MIMO broadcast channels, in Proc. Int. Symp. on Inform. Theory ISIT,
Adelaide, Australia, Sep. 2005, pp. 23252329.
[23] M. Sharif and B. Hassibi, On the capacity of MIMO broadcast channels with partial side information, IEEE Trans. Inform. Theory,
vol. 51, no. 2, pp. 506522, February 2005.
[24] M. Airy, R. W. Heath, and S. Shakkottai, Multiuser diversity for the multiple antenna broadcast channel with linear receivers: asymptotic
analysis, in Conference Record of the Thirty-Eighth Asilomar Conference on Signals, Systems and Computers., vol. 1, Nov. 2004, pp. 886
890.
[25] P. Viswanath, D. N. C. Tse, and R. Laroia, Opportunistic beamforming using dumb antennas, IEEE Trans. Inform. Theory, vol. 48, no. 6,
pp. 12771294, June 2002.
[26] B. Hassibi and T. L. Marzetta, Multiple-antennas and isotropically random unitary inputs: The received signal density in closed form,
IEEE Trans. Inform. Theory, vol. 48, no. 6, pp. 14731484, June 2002.
[27] O. Somekh, A. Haimovich, and Y. Bar-Ness, Sum-rates analysis of downlink channels with 1-bit feedback, submitted to IEEE Comm.
Letters, 2005.
[28] B. Niu, O. Simeone, O. Somekh, and A. M. Haimovich, On the sum-rate of broadcast channels with outdated 1-bit feedback, in Asilomar
Conference on Signals, Systems and Computers, 2006.
[29] W. Santipach and M. L. Honig, Capacity of beamforming with limited training and feedback, in ISIT, Seattle, WA, July 2006.
[30] M. Sharif and B. Hassibi, Scaling laws of sum rate using time-sharing, DPC, and beamforming for MIMO broadcast channels, in Proc.
of Int. Symp. on Inform. Theory, Chicago, IL, Jun./Jul. 2004, p. 175.

October 9, 2006 DRAFT