Massive MU-MIMO Downlink TDD Systems with
Linear Precoding and Downlink Pilots
Hien Quoc Ngo∗ , Erik G. Larsson∗, and Thomas L. Marzetta†
∗ Department of Electrical Engineering (ISY)
Linköping University, 581 83 Linköping, Sweden
† Bell Laboratories, Alcatel-Lucent
Murray Hill, NJ 07974, USA
arXiv:1310.1510v1 [cs.IT] 5 Oct 2013
Abstract—We consider a massive MU-MIMO downlink time-
division duplex system where a base station (BS) equipped with
many antennas serves several single-antenna users in the same
time-frequency resource. We assume that the BS uses linear
precoding for the transmission. To reliably decode the signals
transmitted from the BS, each user should have an estimate of its
channel. In this work, we consider an efficient channel estimation
scheme to acquire CSI at each user, called beamforming training
scheme. With the beamforming training scheme, the BS precodes
the pilot sequences and forwards to all users. Then, based on
the received pilots, each user uses minimum mean-square error
channel estimation to estimate the effective channel gains. The
channel estimation overhead of this scheme does not depend on
the number of BS antennas, and is only proportional to the
number of users. We then derive a lower bound on the capacity
for maximum-ratio transmission and zero-forcing precoding
techniques which enables us to evaluate the spectral efficiency
taking into account the spectral efficiency loss associated with
the transmission of the downlink pilots. Comparing with previous
work where each user uses only the statistical channel properties
to decode the transmitted signals, we see that the proposed
beamforming training scheme is preferable for moderate and
low-mobility environments.
I. I NTRODUCTION
Recently, massive (or very large) multiuser multiple-input
multiple-output (MU-MIMO) systems have attracted a lot of
attention from both academia and industry [1]–[4]. Massive
MU-MIMO is a system where a base station (BS) equipped
with many antennas simultaneously serves several users in the
same frequency band. Owing to the large number of degrees-
of-freedom available for each user, massive MU-MIMO can
provide a very high data rate and communication reliability
with simple linear processing such as maximum-ratio combin-
ing (MRC) or zero-forcing (ZF) on the uplink and maximum-
ratio transmission (MRT) or ZF on the downlink. At the
same time, the radiated energy efficiency can be significantly
improved [5]. Therefore, massive MU-MIMO is considered
as a promising technology for next generations of cellular
systems. In order to use the advantages that massive MU-
MIMO can offer, accurate channel state information (CSI) is
required at the BS and/or the users.
The work of H. Q. Ngo and E. G. Larsson was supported in part by
the Swedish Research Council (VR), the Swedish Foundation for Strategic
Research (SSF), and ELLIIT.
In small MU-MIMO systems where the number of BS
antennas is relatively small, typically, the BS can acquire an
estimate of CSI via feedback in frequency-division duplex
(FDD) operation [6]. More precisely, each user estimates the
channels based on the downlink training, and then it feeds
back its channel estimates to the BS through the reverse
link. However, in massive MU-MIMO systems, the number
of BS antennas is very large and channel estimation becomes
challenging in FDD since the number of downlink resources
needed for pilots will be proportional to the number of BS
antennas. Also, the required bandwidth for CSI feedback
becomes very large. By contrast, in time-division duplex
(TDD) systems, owing to the channel reciprocity, the BS can
obtain CSI in open-loop directly from the uplink training.
The pilot transmission overhead is thus proportional to the
number of users which is typically much smaller than the
number of BS antennas. Therefore, CSI acquisition at the BS
via open-loop training under TDD operation is preferable in
massive MU-MIMO systems [1]–[3], [7], [8]. With this CSI
acquisition, in the uplink, the signals transmitted from the
users can be decoded by using these channel estimates. In the
downlink, the BS can use the channel estimates to precode
the transmit signals. However, the channel estimates are only
available at the BS. The user also should have an estimate of
the channel in order to reliably decode the transmitted signals
in the downlink. To acquire CSI at the users, a simple scheme
is that the BS sends the pilots to the users. Then, each user
will estimate the channel based on the received pilots. This
is very inefficient since the channel estimation overhead will
be proportional to the number of BS antennas. Therefore, the
majority of the research on these systems has assumed that
the users do not have knowledge of the CSI. More precisely,
the signal is detected at each user by only using the statistical
properties of the channels [7]–[9]. Some work assumed that the
users have perfect CSI [10]. To the authors’ best knowledge, it
has not been previously considered how to efficiently acquire
CSI at each user in the massive MU-MIMO downlink.
In this paper, we propose a beamforming training scheme
to acquire estimates of the CSI at each user. With this scheme,
instead of forwarding a long pilot sequence (whose length is
proportional to the number of BS antennas), the BS just beam-
forms a short pilot sequence so that each user can estimate
the effective channel gain (the combination of the precoding
vector and the channel gain). The channel estimation overhead
of this scheme is only proportional to the number of users.
To evaluate the performance of the proposed beamforming
training scheme, we derive a lower bound on the capacity of
two specific linear precoding techniques, namely MRT and ZF.
Numerical results show that the beamforming training scheme
works very well in moderate and low-mobility environments.
Notation: We use upper (lower) bold letters to denote
matrices (vectors). The superscripts T , ∗, and H stand for
the transpose, conjugate, and conjugate-transpose, respectively.
A) denotes the trace of a matrix A , and I n is the n × n
tr (A
identity matrix. The expectation operator and the Euclidean
norm are denoted by E {·} and k · k, respectively. Finally, we
Fig. 1. Massive MU-MIMO downlink system model.
use z ∼ CN (0, Σ ) to denote a circularly symmetric complex
Gaussian vector z with zero mean and covariance matrix Σ .
B. Downlink Transmission
II. S YSTEM M ODEL AND B EAMFORMING T RAINING
Let sk be the symbol to be transmitted to the kth user, with
We consider the downlink transmission in a MU-MIMO E |sk |2 = 1. The BS uses the channel estimate Ĥ H to linearly
system where a BS equipped with M antennas serves K precode the symbols, and then it transmits the precoded signal
single-antenna users in the same time-frequency resource, see vector to all users. Let W ∈ CM×K be the linear precoding
Fig. 1. Here, we assume that M ≫ K. We further assume that matrix which is a function of the channel estimate ĤH . Then,
the BS uses linear precoding techniques to process the signal the M × 1 transmit signal vector is given by
before transmitting to all users. This requires knowledge of √
x = pd W s (3)
CSI at the BS. We assume TDD operation so that the channels
T
on the uplink and downlink are equal. The estimates of CSI where s , [s1 s2 ... sK ] , and pd is the average transmit
are obtained from uplink training. power at the BS. To satisfyn the o power constraint at the
BS, W is chosen such as E kx xk2 = pd , or equivalently
A. Uplink Training n  o
E tr W W H = 1.
Let τu be the number of symbols per coherence interval used The vector of samples collectively received at the K users
entirely for uplink pilots. All users simultaneously transmit is given by
pilot sequences of length τu symbols. The pilot sequences of √
K users are pairwisely orthogonal. Therefore, it is required y = H T x + n = pd H T W s + n (4)
that τu ≥ K.
where n is a vector whose kth element, nk , is the additive
Denote by H ∈ CM×K the channel matrix between the noise at the kth user. We assume that nk ∼ CN (0, 1). Define
BS and the K users. We assume that elements of H are i.i.d. aki , h Tk w i , where h i and w i are the ith columns of H and
Gaussian distributed with zero mean and unit variance. Here, W , respectively. Then, the received signal at the kth user can
for the simplicity, we neglect the effects of large-scale fading. be written as
Then, the minimum mean-square error (MMSE) estimate of
K
H is given by [11] √ √ X
yk = pd akk sk + pd aki si + nk . (5)
√
τu pu τu pu i6=k
H=
Ĥ H+ Nu (1)
τu pu + 1 τu pu + 1 Remark 1: Each user should have CSI to coherently detect
the transmitted signals. A simple way to acquire CSI is to
where N u is a Gaussian matrix with i.i.d. CN (0, 1) entries,
use downlink pilots. The channel estimate overhead will be
and pu denotes the average transmit power of each uplink pilot
proportional to M . In massive MIMO, M is large, so it is inef-
symbol. The channel matrix H can be decomposed as
ficient to estimate the full channel matrix H at each user using
H +E
H = Ĥ (2) downlink pilots. This is the reason for why most of previous
studies assumed that the users have only knowledge of the
where E is the channel estimation error. Since we use MMSE statistical properties of the channels [8], [9]. More precisely,
H and E areindependent [11]. Further-
channel estimation, Ĥ in [8], [9], the authors use E {akk } to detect the transmitted
H has i.i.d. CN 0, τuτpuup+1
more, Ĥ u
elements, and E has i.i.d. signals. With very large M , akk becomes nearly deterministic.
  In this case, using E {akk } for the signal detection is good
CN 0, τu p1u +1 elements. enough. However, for moderately large M , the users should
 n o
2
have CSI in order to reliably decode the transmitted signals. where Var (aki ) , E |aki − E {aki }| , and ỹp,ki is the ith
We observe from (5) that to detect sk , user k does not need element of ỹy p,k . Let ǫki be the channel estimation error. Then,
the knowledge of H (which has a dimension of M × K). the effective channel aki can be decomposed as
Instead, user k needs only to know akk which is a scalar value.
Therefore, to acquire akk at each user, we can spend a small aki = âki + ǫki . (11)
amount of the coherence interval on downlink training. In the Note that, since we use MMSE estimation, the estimate âki
next section, we will provide more detail about this proposed and the estimation error ǫki are uncorrelated.
downlink beamforming training scheme to estimate akk . With
this scheme, the channel estimation overhead is proportional III. ACHIEVABLE D OWNLINK R ATE
to the number of users K. In this section, we derive a lower bound on the achievable
downlink rate for MRT and ZF precoding techniques, using
C. Beamforming Training Scheme
the proposed beamforming training scheme. To obtain these
The BS beamforms the pilots. Then, the kth user will achievable rates, we use the techniques of [12].
estimate aki by using the received pilots. Let S p ∈ CK×τd User k uses the channel estimate â ak in (10) to detect the
be the pilot matrix, where τd is the duration (in symbols) of transmitted signal sk . Therefore, the achievable downlink rate
the downlink training. The pilot matrix is given by of the transmission from the BS to the kth user is the mutual
√ information between the unknown transmitted signal sk and
S p = τd pdΦ. (6)
the observed received signal yk given by (5) and the known
We assume that the rows of Φ are pairwisely orthonormal, i.e., channel estimate âak = [âk1 ... âkK ]T , i.e., I (sk ; yk , â
ak ).
ΦΦ H = I K . This requires that τd ≥ K. Following a similar methodology as in [12, Appendix A],
The BS beamforms the pilot sequence using the precoding we obtain a lower bound on the achievable rate of the
matrix W . More precisely, the transmitted pilot matrix is transmission from the BS to the kth user as:
W S p . Then, the K × τd received pilot matrix at the K users
  
 
is given by
 
  2 
 pd |âkk |  
√ Rk = E log21 + .
 
Y Tp = τd pdH T W Φ + N Tp . K K

(7) 
  P n
2
o P 2 

 pd E |ǫki | + pd |âki | + 1  

i=1 i6=k
where N p is the AWGN matrix whose elements are i.i.d.
(12)
CN (0, 1). The received pilot matrix Y Tp can be represented by
Y Tp Φ H and Y Tp Φ H H
⊥ , where Φ ⊥ is the orthogonal complement
We next simplify the capacity lower bound given by (12)
of Φ , i.e., Φ ⊥ = I τd − Φ Φ . We can see that Y Tp Φ H
H H H
⊥ only
for two specific linear precoding techniques at the BS, namely,
includes noise which is independent of Y Tp Φ H . Thus, it is MRT and ZF.
sufficient to use Y Tp Φ H for the channel estimation. Let A. Maximum-Ratio Transmission
T √ T With MRT, the precoding matrix W is given by
Y
Ỹ p , Y Tp Φ H = τd pdH T W + Ñ
Np (8)
∗
T H
W = αMRTĤ (13)
where Np
Ñ , N Tp Φ H
has i.i.d. CN (0, 1) elements. From (8),
the 1 × K received pilot vector at user k is given by where αMRT is a normalization constant chosen
n to satisfy othe
H
√ √ transmit power constraint at the BS, i.e., E tr W W =
ỹy Tp,k = τd pdhTk W + ñnTp,k = τd pdaTk + ñ
nTp,k (9)
1. Hence,
T
where a k , [ak1 ak2 ... akK ] , and ỹy p,k and ñ
np,k are the kth
v r
u 1 τu pu + 1
columns of Ỹ Y p and Ñ
N p , respectively. t n  ∗ T o =
αMRT = u . (14)
E tr Ĥ H ĤH M Kτu pu
From the received pilot ỹy Tp,k , user k estimates a k . De-
pending on the precoding matrix W , the elements of a k can
be correlated and hence, they should be jointly estimated. Proposition 1: With MRT, the lower bound on the achiev-
However, here, for the simplicity of the analysis, we estimate able rate given by (12) becomes
ak1 , ..., akK independently, i.e., we use the ith element of ỹy p,k ( !)
pd |âkk |2
to estimate aki . In Section V, we show that estimating the Rk = E log2 1 + Kp PK 2
(15)
τd pd +K + pd i6=k |âki | + 1
d
elements of a k jointly will not improve the system perfor-
mance much compared to the case where the elements of a k where
are estimated independently. The MMSE channel estimate of √ s
aki is given by [11] τd pd K τu pu M
âki = ỹy + δki (16)
√ τd pd + K p,ki τd pd + K K (τu pu + 1)
τd pd Var (aki ) √
âki = E {aki } + (ỹp,ki − τd pd E {aki })
τd pd Var (aki ) + 1 where δki = 1 when i = k and 0 otherwise.
(10) Proof: See Appendix A.
 12,0 30,0
ZF, with Beamforming Training (Proposed)
With Beamforming Training (Proposed) ZF, without Beamforming Training [9]
10,0
Without Beamforming Training [9] 25,0 MRT, with Beamforming Training (Proposed)

Spectral Efficiency (bits/s/Hz)

Spectral Efficiency (bits/s/Hz)

MRT, without Beamforming Training [9]

8,0 20,0
M = 50

6,0 15,0
M = 50

4,0 10,0

2,0 5,0
M = 10
M = 10
K = 1, pu = 0 dB K = 5, pu = 0 dB
0,0 0,0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
SNR (dB) SNR (dB)

Fig. 2. Spectral efficiency versus SNR for a single-user setup (K = 1, Fig. 3. Spectral efficiency versus SNR for a multiuser setup (K = 5, pu = 0
pu = 0 dB, and T = 200). dB, and T = 200).

B. Zero-Forcing where Rk is given by (15) for MRT, and (19) for ZF.
With ZF, the precoding matrix is For comparison, we also consider the spectral efficiency
∗
 T ∗ −1 for the case that there is no beamforming training and that
W = αZFĤ
H Ĥ H Ĥ H (17) E {akk } is used instead of akk for the detection [9]. The
spectral efficiency for this case is given by [9]
where the normalization
n  constantoαZF is chosen to satisfy the   
H  T−τu K log2 1 + M τu pu pd
power constraint E tr W W = 1, i.e., [9] T K (pd +1)(τ p
u u +1) , for MRT
S0 = T−τ   (22)
 u K log2 1 + M−K τu pu pd
K τu pu +pd +1 , for ZF
s
(M − K) τu pu T
αZF = . (18)
K (τu pu + 1) In all examples, we choose τu = τd = K and pu = 0dB. We
define SNR , pd .
Proposition 2: With ZF, the lower bound on the achievable We first consider a single-user setup (K = 1). When K = 1,
rate given by (12) becomes the performances MRT and ZF are the same. Fig. 2 shows
( !) the spectral efficiency versus SNR for different number of
pd |âkk |2
Rk = E log2 1 + Kpd PK 2
BS antennas M = 10 and M = 50, at T = 200 (e.g.
τd pd +K(τu pu +1) + pd i6=k |âki | + 1 1ms×200kHz). We can see that the beamforming training
(19) scheme outperforms the case without beamforming training.
where The performance gap increases significantly when the SNR
√ increases. The reason is that, when SNR (or the downlink
τd pd
âki = ỹy power) increases, the channel estimate at each user is more
τd pd + K (τu pu + 1) p,ki
p accurate and hence, the advantage of the beamforming training
K (M − K) τu pu (τu pu + 1) scheme grows.
+ δki . (20)
τd pd + K (τu pu + 1) Next, we consider a multiuser setup. Here, we choose the
Proof: See Appendix B. number of users to be K = 5. Fig. 3 shows the spectral
efficiency versus SNR for the MRT and ZF precoders, at
IV. N UMERICAL R ESULTS
M = 10, M = 50, and T = 200. Again, the beamforming
In this section, we illustrate the spectral efficiency per- training offers an improved performance. In addition, we can
formance of the beamforming training scheme. The spectral see that the beamforming training with MRT precoding is more
efficiency is defined as the sum-rate (in bits) per channel use. efficient than the beamforming training with ZF precoding.
Let T be the length of the coherence interval (in symbols). This is due to the fact that, with ZF, the randomness of the
During each coherence interval, we spend τu symbols for effective channel gain akk at the kth user is smaller than the
uplink training and τd symbols for beamforming training. one with MRT (with ZF, akk becomes deterministic when the
Therefore, the spectral efficiency is given by BS has perfect CSI) and hence, MRC has a higher advantage
K
T − τu − τd X of using the channel estimate for the signal detection.
STB = Rk (21) Furthermore, we consider the effect of the length of the
T
k=1 coherence interval on the system performance of the beam-
 35,0 30,0
With Beamforming Training (Proposed) ZF, with Beamforming Training (Proposed)
Without Beamforming Training [9] MRT, with Beamforming Training (Proposed)
30,0
25,0 ZF, MRT, with a Genie Receiver
Spectral Efficiency (bits/s/Hz)

Spectral Efficiency (bits/s/Hz)

25,0 ZF
20,0
M = 50
20,0
15,0
15,0
MRT
10,0
10,0

5,0 5,0
K = 5, M = 50 M = 10
pu = 0 dB, pd = 20 dB
0,0 0,0
50 100 150 200 0 2 4 6 8 10 12 14 16 18 20
Coherence Interval T (symbols) SNR (dB)

Fig. 4. Spectral efficiency versus coherence interval for MRT and ZF Fig. 5. Spectral efficiency versus SNR with a genie receiver (K = 5, pu = 0
precoding (M = 50, K = 5, pu = 0 dB, and pd = 20 dB). dB, and T = 200).

forming training scheme. Fig. 4 shows the spectral efficiency antennas. Therefore, it is suitable and efficient for massive
versus the length of the coherence interval T at M = 50, MU-MIMO systems. Furthermore, the down-link pilots will
K = 5, and pd = 20 dB. As expected, for short coherence in- add robustness to the beamforming process which otherwise
tervals (in a high-mobility environment), the training duration is dependent on the validity of the prior (Bayes) assumptions.
is relatively large compared to the length of the coherence
A PPENDIX
interval and hence, we should not use the beamforming
training to estimate CSI at each user. At moderate and large A. Proof of Proposition 1
∗
T , the training duration is relatively small compared with With MRT, we have that aki = αMRTh Tk ĥ
hi .
the coherence interval. As a result, the beamforming training • Compute E {aki }:
scheme is preferable. From (2), we have
Finally, we consider the spectral efficiency of our scheme  T  ∗
but with a genie receiver, i.e., we assume that the kth user can hk + ε Tk ĥ
aki = αMRT ĥ hi
estimate perfectly a k in the beamforming training phase. For T ∗ ∗
this case, the spectral efficiency is given by = αMRTĥ hi + αMRTε Tk ĥ
hk ĥ hi (24)
  
where ĥhk and ε k are the kth columns of Ĥ H and E ,
∗

 

K  2 
respectively. Since ε̂εk and ĥ
hi are uncorrelated with all
T −τu −τd X   pd |akk | 
SG = E log21 + . (23)
 
T K
 i, k = 1, ..., K, we obtain
  P 2 
k=1  
 pd |aki | + 1 

 n T ∗o
i6=k E {aki } = αMRT E ĥhk ĥ
hi
Figure 5 compares the spectral efficiency given by (12), where
(
0, if i 6= k
the kth user estimates the elements of a k independently, with = q
τu pu M (25)
the one obtained by (23), where we assume that there is a K(τu pu +1) , if i = k

genie receiver at the kth user. Here, we choose K = 5 and • Compute Var (aki ) for i 6= k:
T = 200. We can see that performance gap between two cases From (24) and (25), we have
is very small. This implies that estimating the elements of a k n
2
o
independently is fairly reasonable. Var (aki ) = E |aki |
T ∗ 2
     
(a) T ∗ 2
V. C ONCLUSION AND F UTURE W ORK = E αMRTĥ

hk ĥ
hi  + E αMRTε k ĥ

hi 
In this paper, we proposed and analyzed a scheme to acquire  2
τu pu τu pu M
CSI at each user in the downlink of a MU-MIMO system, = α2MRT M + α2MRT 2
called beamforming training scheme. With this scheme, the BS τu pu + 1 (τu pu + 1)
uses linear precoding techniques to process the pilot sequence = 1/K (26)
before sending it to the users for the channel estimation. T ∗
The channel estimation overhead of this beamforming training hk ĥ
where (a) is obtained by using the fact that ĥ hi and
∗
scheme is small and does not depend on the number of BS ε Tk ĥ
hi are uncorrelated.
 • Compute Var (akk ): Using [13, Lemma 2.10], we obtain
Similarly, we have 1
n o Var (aki ) = . (36)
Var (akk ) = E |akk | 2 2
− |E {akk }| . (27) K (τu pu + 1)
Substitutingn(34) and o (36) into (10), we get (20).
From (24), we have 2
     • Compute E |ǫki | :
4  T ∗ 2
n o
2 2
E |akk | = αMRT E ĥ 2
hk + αMRT E εεk ĥ
hk  . If i 6= k, from (9) and (20), we have
n o n o
(28) E |ǫki |2 = E |aki − âki |2
Using [13, Lemma 2.9], we obtain
( √ 2)
 K (τu pu + 1) aki τ d p d ñ p,ki 
2 = E  − 
τd pd + K (τu pu + 1) τd pd + K (τu pu + 1)

n
2
o
2 τu pu
E |akk | = αMRT M (M + 1) 2 n
τu pu + 1 
K (τu pu + 1) 2
o
τu pu = E |aki |
+ α2MRT M. (29) τd pd + K (τu pu + 1)
(τu pu + 1)2 τd pd
+
Substituting (25) and (29) into (27), we obtain (τd pd + K (τu pu + 1))2
1
Var (akk ) = 1/K. (30) = (37)
τd pd + K (τu pu + 1)
Substitutingn(25), (26),
o and (30) into (10), we get (16). where the last equality
2 n is obtained by using (36). Sim-
Compute E |ǫki | :
o
•
ilarly, we obtain E |ǫkk |2 = τd pd +K(τ 1
u pu +1)
. There-
If i 6= k, from (9) and (16), we have fore, we arrive at the result in Proposition 2.
n o n o
2 2
E |ǫki | = E |aki − âki | R EFERENCES
( √ 2 )
 K τ d p d  [1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
= E  aki − ñp,ki  bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9,
τd pd + K τd pd + K no. 11, pp. 3590–3600, Nov. 2010.
 2 n [2] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, “Massive
K 2
o τd pd MIMO for next generation wireless systems,” IEEE Commun. Mag.,
= E |aki | + 2 (31) 2013, to appear. [Online]. Available: http://arxiv.org/abs/1304.6690.
τd pd + K (τd pd + K) [3] Y.-H. Nam, B. L. Ng, K. Sayana, Y. Li, J. C. Zhang, Y. Kim, and
where ñp,ki is the ith element of ñnp,k . Using (26), we J. Lee, “Full-dimension MIMO (FD-MIMO) for next generation cellular
technology,” IEEE Commun. Mag., vol. 51, no. 6, pp. 172–178, 2013.
obtain [4] C. Shepard, H. Yu, N. Anand, L. E. Li, T. L. Marzetta, R. Yang, and
n
2
o 1 L. Zhong, “Argos: Practical many-antenna base stations,” in Proc. ACM
E |ǫki | = . (32) Int. Conf. Mobile Computing and Networking (MobiCom), Istanbul,
τd pd + K Turkey, Aug. 2012.
[5] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi-
n o
2
Similarly, we obtain E |ǫkk | = τd pd1+K . Therefore, ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
we arrive at the result in Proposition 1. vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
[6] M. Kobayashi, N. Jindal, and G. Caire, “Training and feedback optimiza-
B. Proof of Proposition 2 tion for multiuser MIMO downlink,” IEEE Trans. Commun., vol. 59,
no. 8, pp. 2228–2240, Aug. 2011.
With ZF, we have that aki = h Tk w i , where w i is the ith [7] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
∗
 T ∗ −1 T of cellular networks: How many antennas do we need?” IEEE J. Sel.
H Ĥ
column of αZFĤ H Ĥ
H H W = αZFI K , we
. Since Ĥ Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
have [8] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot
 T  contamination and precoding in multi-cell TDD systems,” IEEE Trans.
aki = ĥ hk + ε Tk w i = αZF δki + ε Tk w i . (33) Wireless Commun., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
[9] H. Yang and T. L. Marzetta, “Performance of conjugate and zero-
forcing beamforming in large-scale antenna systems,” IEEE J. Sel. Areas
Therefore, Commun., vol. 31, no. 2, pp. 172–179, Feb. 2013.
[10] E. Björnson, J. Hoydis, M. Kountouris, and M. Debbah, “Hardware
E {aki } = αZF δki . (34) impairments in large-scale MISO systems: Energy efficiency, estimation,
and capacity limits,” in Proc. Signal Processing and Optimization for
• Compute Var (aki ): Green Energy and Green Communications (DSP’13), Santorini, Greece,
From (33) and (34), we have 2013.
[11] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
n 2 o 1 n o
Var (aki ) = E ε Tk w i  = E kw w i k2 Theory. Englewood Cliffs, NJ: Prentice Hall, 1993.
[12] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “The multicell multiuser
τ p +1
 u u −1   MIMO uplink with very large antenna arrays and a finite-dimensional
α2ZF T ∗ channel,” IEEE Trans. Commun., vol. 61, no. 6, pp. 2350–2361, Jun.
= E H Ĥ
Ĥ H
τu pu + 1 ii
2013.
[13] A. M. Tulino and S. Verdú, “Random matrix theory and wireless commu-
α2ZF
   
1 T ∗ −1 nications,” Foundations and Trends in Communications and Information
= E tr Ĥ H Ĥ H . (35) Theory, vol. 1, no. 1, pp. 1–182, Jun. 2004.
τu pu + 1 K