IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO.

10, OCTOBER 2017 9005

Large System Analysis of Power Normalization

Techniques in Massive MIMO
Meysam Sadeghi , Student Member, IEEE, Luca Sanguinetti , Senior Member, IEEE,
Romain Couillet, Senior Member, IEEE, and Chau Yuen, Senior Member, IEEE

Abstract—Linear precoding has been widely studied in the con- of antennas enables each BS to focus the radiated energy into a
text of massive multiple-input multiple-output (MIMO) together specific location in space or to intercept the power of transmit-
with two common power normalization techniques, namely, ma- ted electromagnetic waves more efficiently. Therefore, Massive
trix normalization (MN) and vector normalization (VN). Despite
this, their effect on the performance of massive MIMO systems MIMO has higher spectral efficiency and energy efficiency com-
has not been thoroughly studied yet. The aim of this paper is to pared to classical multiuser MIMO systems [3], [5]–[7]. Due to
fulfill this gap by using large system analysis. Considering a system the quasi-orthogonal nature of channels in Massive MIMO, lin-
model that accounts for channel estimation, pilot contamination, ear precoding and detection schemes perform close-to-optimal
arbitrary pathloss, and per-user channel correlation, we compute
[5], [6], [8]. If the channel reciprocity is exploited, the overhead
tight approximations for the signal-to-interference-plus-noise ratio
and the rate of each user equipment in the system while employing of the channel state information (CSI) acquisition is indepen-
maximum ratio transmission (MRT), zero forcing (ZF), and reg- dent of the number of BS antennas [9]. Moreover, it has been
ularized ZF precoding under both MN and VN techniques. Such recently shown that the capacity of Massive MIMO increases
approximations are used to analytically reveal how the choice of without bound as the number of antennas increases, even under
power normalization affects the performance of MRT and ZF un-
pilot contamination [10], [11]. These remarkable features candi-
der uncorrelated fading channels. It turns out that ZF with VN
resembles a sum rate maximizer while it provides a notion of fair- date Massive MIMO as one of the most promising technologies
ness under MN. Numerical results are used to validate the accuracy for next generation of cellular networks [8], [12], [13].
of the asymptotic analysis and to show that in massive MIMO, non- Linear precoding has a central role in Massive MIMO and
coherent interference and noise, rather than pilot contamination, has been extensively studied in the past few years [5], [14]–
are often the major limiting factors of the considered precoding
[22]. The spectral efficiency and energy efficiency of maximum
schemes.
ratio transmission (MRT) and zero forcing (ZF) precoding in
Index Terms—Linear precoding, large system analysis, mas- single-cell Massive MIMO systems are investigated in [14].
sive multiple-input multiple-output (MIMO), power normalization In [15], a multicell linear precoding is proposed to mitigate
techniques, pilot contamination.
the effect of pilot contamination. Multicell processing is also
considered in [10], [11], [21]. The performance of MRT, ZF,
I. INTRODUCTION and regularized ZF (RZF) precoding in single-cell large-scale
MIMO systems is studied in [16], considering a per-user chan-
ASSIVE multiple-input multiple-output(MIMO) is a
M multiuser MIMO system that employs a large number of
antennas at the base stations (BSs) to serve a relatively smaller
nel correlation model. A seminal treatment of MRT and RZF
precoding schemes in multicell Massive MIMO systems is pre-
sented in [5], followed by [17] where downlink training and
number of user equipments (UEs) [1]–[4]. This large number
linear pilot contamination precoding are also considered. In
[18], closed-form approximations for the achievable downlink
Manuscript received November 11, 2016; revised March 7, 2017 and April rates of MRT and ZF precoding schemes are presented for mul-
27, 2017; accepted April 28, 2017. Date of publication May 15, 2017; date of
current version October 13, 2017. This work was supported by A*Star SERC ticell Massive MIMO systems. A linear truncated polynomial
under Grant 142-02-00043. The work of L. Sanguinetti and R. Couillet have expansion based precoding is proposed in [19], which reduces
been supported by the ERC Starting Grant 305123 MORE. The review of this the complexity of RZF precoding. The effect of phase noise on
paper was coordinated by Ngoc-Dũng Ðào. (Corresponding author: Meysam
Sadeghi.) the signal-to-interference-plus-noise (SINR) of MRT, ZF, and
M. Sadeghi and C. Yuen are with the Singapore University of Technol- RZF precoding schemes is studied in [20].
ogy and Design, Singapore 487372 (e-mail: [email protected]; In order to utilize linear precoding, the power should be ad-
[email protected]).
L. Sanguinetti is with the Dipartimento di Ingegneria dell’Informazione, justed to meet the power constraint at the BS. This can be done
University of Pisa, Pisa 56100, Italy, and also with the Large Systems and Net- either by optimized power allocation among the downlink data
works Group, CentraleSupélec, Université Paris-Saclay, Gif-sur-Yvette 91192, streams [21], [23]–[25], or simply by uniform power alloca-
France (e-mail: [email protected]).
R. Couillet is with the Signals and Statistics Group, CentraleSupélec, tion among downlink data streams jointly with precoder power
Université Paris-Saclay, Paris 4103, France (e-mail: romain.couillet@ normalization [5], [14], [17]–[19]. Although the latter approach
centralesupelec.fr). may provide a weaker performance compared to the former, it
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. is the most used in the Massive MIMO literature [5], [9], [14],
Digital Object Identifier 10.1109/TVT.2017.2704112 [17]–[19]. The reason for this is that power allocation presents

0018-9545 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
9006 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 10, OCTOBER 2017

the following major issues: (i) finding a global solution is a The remainder of this paper is organized as follows. Section II
challenging task [24], [26], [27]; (ii) a certain level of coordina- introduces the network model, the channel estimation scheme,
tion or cooperation among cells is required; and (iii) it should be the precoding and power normalization methods, as well as the
performed very frequently, even for static users, as scheduling downlink achievable rates. The large system analysis is provided
may change rapidly in practice. in Section III. The effect of power normalization techniques is
The two commonly used power normalization techniques in elaborated in Section IV for uncorrelated fading channels. The
Massive MIMO are matrix normalization (MN) and vector nor- provided asymptotic approximations are verified by means of
malization (VN) [26], [27]. In MN, the precoding matrix of numerical results in Section V. Some conclusions are drawn in
each BS is adjusted by multiplying it with a scalar such that Section VI.
the power constraint at the BS is met[5], [9], [14], [17], [19]. Notations: The following notation is used throughout the pa-
On the other hand, with VN the precoding matrix is normalized per. Scalars are denoted by lower case letters whereas boldface
such that equal amount of power is allocated to each UE while lower (upper) case letters are used for vectors (matrices). We
satisfying the power constraint [18], [26], [27]. Note that these denote by IN the identity matrix of size N and represent the
two methods yield the same performance with optimal power element on the ith row and kth column of A as [A]i,k . The sym-
allocation, but not with practical suboptimal power allocation bol CN (., .) denotes the circularly symmetric complex Gaussian
[27], [28]. distribution. The trace, transpose, conjugate transpose, real part,
Although linear precoding has been largely studied in Mas- and expectation operators are denoted by tr(·), (·)T , (·)H , Re(·),
sive MIMO, a detailed treatment of the impact of power nor- and E[·], respectively. The notation −−−→ represent almost sure
N →∞
malization does not exist in the literature. The first attempt convergence.
in this direction was carried out in [26] and extended in [27]
wherein the authors study the impact of MN and VN on MRT II. COMMUNICATION SCHEME
and ZF precoding schemes. However, both [26], [27] do not
Next, we introduce the system model, the channel estimation
grasp the essence of a practical Massive MIMO system since:
method, the precoding and power normalization techniques, and
(i) a single-cell network composed of three radio units is con-
compute the downlink achievable rates.
sidered; (ii) perfect CSI is assumed and thus CSI acquisition or
pilot contamination are not accounted for; and (iii) large-scale
A. System Model
attenuation is neglected, though it has a fundamental impact on
power normalization, as detailed later. We consider the downlink of a Massive MIMO system com-
The goal of this paper is to study the effect of MN and VN posed of L cells, where the set of all cells is denoted by L. The
on the performance of MRT, ZF, and RZF in Massive MIMO, BS of each cell has N antennas and serves K single-antenna
in the simple and practical case of uniform power allocation. UEs in the same time-frequency resource. The set of UEs be-
Particularly, the following contributions are provided. longing to cell l is denoted by Kl . We assume transmissions
1) We extend the analysis in [26], [27] to a multicell Massive over flat-fading channels. We employ a double index notation to
MIMO system, which accounts for channel estimation, refer to each UE as e.g., “user k in cell l”. Under this convention,
pilot contamination, an arbitrary pathloss model, and per- let hjlk ∈ C N be the channel between BS j and UE k in cell l
user channel correlation. Asymptotically tight approxima- within a block and assume that
tions of the signal-to-interference-plus-noise ratio (SINR) 1/2
hjlk = Θjlk zjlk (1)
and rate of each UE are provided and validated by numer-
ical results for MRT, ZF, and RZF with VN and MN. where zjlk ∼ CN (0, IN ) and Θjlk ∈ C N ×N accounts for the
2) Explicit asymptotic approximations for the SINR and rate corresponding channel correlation matrix. Note that (1) enables
of each UE are given for a Rayleigh fading channel model. us to assign a unique correlation matrix between each user-BS
These results are used: (i) to elaborate on how the two dif- pair and it includes many channel models in the literature as
ferent normalization techniques affect the signal, noise, special cases [16].
and interference powers as well as the pilot contamina-
tion experienced by each UE in the system; (ii) to prove B. Channel Estimation
that large-scale fading has a fundamental role on the per-
We assume that BSs and UEs are perfectly synchronized and
formance provided by the two normalization techniques
operate according to a time-division duplex (TDD) protocol.
while both perform the same if neglected; (iii) to show
Then, the channels can be found by an uplink training phase
that ZF conveys a notion of sum rate maximization with
and used in the downlink by exploiting channel reciprocity.
VN and of fairness with MN.
Using orthogonal pilots in each cell while reusing them in all
3) The asymptotic approximations of SINRs are used to-
other cells, after correlating the received training signal with the
gether with numerical results to study the main limiting
pilot sequence of UE k, the observed channel of user k in cell
factors of the investigated schemes. Particularly, we re-
j is
veal that in Massive MIMO, non-coherent interference
and noise, rather than pilot contamination, are often the 
L
1
major limiting factors for all schemes. yjk
tr
= hjjk + hjlk + √ njk (2)
ρtr
l=1,l= j
SADEGHI et al.: LARGE SYSTEM ANALYSIS OF POWER NORMALIZATION TECHNIQUES IN MASSIVE MIMO 9007

where njk ∼ CN (0, σ 2 IN ) with σ 2 being the noise spectral On the other hand, if MN is employed, then Dj = ηj IK with
density and ρtr is proportional to the training SNR. Applying
the MMSE estimation, the estimated channel can be computed K
ηj = . (11)
as follows [5] E[trFj FH
j ]

ĥjjk = Θjjk Qjk yjk

tr
(3)
D. Downlink Achievable Rate
where ĥjjk ∼ CN (0, Φjjk ). Also, Qjk and Φjlk are given by
 L −1 The received signal of user k in cell j can be written as
 σ2
Qjk = Θjlk + IN ∀j, k (4)
ρtr 
K
l=1 yjk = hH
jjk gjk sjk + hH
jjk gji sji
Φjlk = Θjjk Qjk Θjlk ∀j, l, k. (5) i=1,i= k

Note that due to the orthogonality principle of MMSE, the esti- 

L 
K

 jjk = hjjk − ĥjjk is independent of ĥjjk and such

mation error h + hH
ljk gli sli + njk (12)

that hjjk ∼ CN (0, Θjjk − Φjjk ). For notational simplicity, we
l=1,l= j i=1

denote Ĥjj = [ĥjj1 , . . . , ĥjjK ] as the matrix collecting the esti- with sli ∈ C being the signal intended to UE i in cell l, as-
mated channels of cell j. sumed independent across (l, i) pairs, of zero mean and unit
variance, and njk ∼ CN (0, σ 2 /ρdl ) where ρdl is proportional to
C. Precoding and Power Normalization Techniques the downlink signal power.
As mentioned earlier, we consider MRT, ZF, and RZF with As in [1], [5], [6], [15] (among many others), we assume
VN and MN [26], [27]. Denoting by Gj = [gj 1 , . . . , gjK ] ∈ that there are no downlink pilots such that the UEs do not have
C N ×K the precoding matrix of BS j, where gjk ∈ C N is the knowledge of the current channels but can only learn the average
precoding vector of UE k in cell j, we have channel gain E{hH jjk gjk } and the total interference power. Note
1/2 this is the common approach in Massive MIMO due to the
Gj = Fj Dj (6) channel hardening [29]. Using the same technique as in [30], an
where Fj = [fj 1 , . . . , fjK ] ∈ C N ×K determines the precoding ergodic achievable information rate for UE k in cell j is obtained
scheme and Dj ∈ C K ×K characterizes the power allocation as rjk = log2 (1 + γjk ) where γjk is given by
strategy. Therefore, Fj takes one of the following forms:
⎧ jjk gjk ]|
|E[hH 2
⎪
⎪ Ĥjj MRT (7) γjk = (13)
⎪
⎨
−1 
L 
K
σ2
+ E[|hH
ljk gli | ]
2 − |E[hH
jjk gjk ]|
2
Fj = Ĥjj Ĥjj Ĥjj
H
ZF (8) ρd l
l=1 i=1
⎪
⎪
−1
⎪
⎩ Ĥjj ĤHjj + Zj + N αj IN Ĥjj RZF (9)
where the expectation is taken with respect to the channel real-
where αj > 0 is the regularization parameter and Zj ∈ C N ×N izations. The above result holds true for any precoding scheme
is an arbitrary Hermitian nonnegative definite matrix that can be and is obtained by treating the interference (from the same
used to leverage the system performance [5]. and other cells) and channel uncertainty as worst-case Gaussian
As mentioned in the introduction, finding the optimal values noise. By using VN and MN, i.e., (10) and (11), the SINR takes
for the elements of Dj is challenging in practice [24]. This is respectively the form in Eqn. (14) and (15), given on this bottom
why VN or MN are usually employed [27]. In this case, Dj of the page.
MN VN
is diagonal with entries chosen so as to satisfy the following As for all precoding schemes, γjk and γjk depend on
average power constraint E[trGj GH j ] = K ∀j. If VN is used,
the statistical distribution of {hjlk } and {ĥjlk }. This makes
then the kth diagonal element of Dj is computed as hard to compute both in closed-form. To overcome this issue, a
large system analysis is provided next to find tight asymptotic
1
[Dj ]k ,k = djk = . (10) approximations (hereafter called deterministic equivalents) for
E[fjk
Hf ]
jk MN
γjk VN
and γjk and their associated achievable rates.

VN
djk |E[hH jjk fjk ]|
2
γjk = (14)

L 
K L
jjk fjk ) + dli E[|hH ljk fli | ] + l=1,l= j dlk E[|hljk flk | ]
σ2
ρd l + djk var(hH 2 H 2
l=1 i=1,i= k

MN
ηj |E[hHjjk fjk ]|
2
γjk = . (15)

L 
K L
jjk fjk ) ηl E[|hljk fli | ] + l=1,l= j ηl E[|hljk flk | ]
σ2
ρd l + ηj var(hH + H 2 H 2
l=1 i=1,i= k
9008 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 10, OCTOBER 2017

III. LARGE SYSTEM ANALYSIS 1

uljk = trΦljk Tl (22)
We consider a regime in which N and K grow large N
with a non-trivial ratio N/K, where 1 < lim inf N/K ≤ 1 |uljk |2 1
lim sup N/K < ∞. We will represent it as N → ∞. Under li,jk = trΘljk T l,Φ l l i + 2
× trΦllk T l,Φ l l i
N (1 + ulk ) N
this assumption, we provide asymptotic approximations, also  
called deterministic equivalents (DEs), for γjk with MRT, ZF, 2 1 ∗
− Re trΦljk Tl,Φ l l i × uljk (23)
and RZF and either MN or VN. The DE is represented by γ jk , 1 + ulk N
and it is such that γjk − γ jk −−−→ 0. By applying the contin-
N →∞ and Sl = ZNl . Also, Tl , T l,IN , and T l,Φ l l i are given in
uous mapping theorem [31], the almost sure convergence of Theorems 7 and 8 in Appendix E.
the results illustrated below implies that rjk − rjk −−−→ 0 with Proof. The proof is provided in Appendix B. 
N →∞
rjk = log2 (1 + γ jk ), where γ jk denotes one of the asymptotic Theorem 3: Let A1, A2 and A4 hold true. If ZF with VN is
(ZF−VN)
approximations computed below. VN
employed, then γjk − γ jk −−−→ 0 almost surely with
N →∞
As limiting cases are considered, the following conditions
(widely used in the literature [5], [16], [32], [33]) are needed. (ZF−VN)
ujk
γ jk = (24)
1 σ2 1
L  K 
li, jk 
L u 2l j k
A1
1/2
: lim sup||Θjlk || <∞ and lim inf tr (Θjlk ) > 0 N ρd l + N uli + ulk
N l=1 i=1 l=1,l= j

1 
A2 :∃ > 0 : λm in HH where
ll Hll > 
N 1 
1 uli = tr Φlli Tl (25)
A3 : lim sup|| Z l || < ∞ N
N N  −1
1  Φlli
K
A4 :rank(Ĥll ) ≥ K. Tl = + IN (26)
N uli
i=1

A. Large System Results for Vector Normalization 1 

uljk = tr Φljk Tl (27)
VN
In this subsection, we derive DEs for γjk , when any of MRT, N
ZF, and RZF precoding schemes is used. 1 |uljk |2 1
Theorem 1: Let A1 hold true. If MRT with VN is used, then li,jk = trΘljk T l,Φ l l i + trΦllk T l,Φ l l i
(M RT−VN) N ulk 2 N
VN
γjk − γ jk −−−→ 0 almost surely with  
N →∞
2 1
2 − Re u∗ljk trΦljk T l,Φ l l i (28)
(M RT−VN) d†jk N1 trΦjjk ulk N
γ jk = (16)  

L  
1  uli,Φ l l k Φlli
K L K
σ2
N ρd l + N1 d†li zli,jk + d†lk | N1 trΦljk |2 T l,Φ l l k = Tl + Φllk Tl (29)
l=1 i=1 l=1,l= j N uli 2
i=1
where

1 −1 where u l,Φ l l k = [u l1,Φ l l k , . . . , u lK,Φ l l k ]T ∈ C K is computed as
= d†li
trΦlli (17)
N u l,Φ l l k = (IK − Jl )−1 vl,Φ l l k (30)
1
zli,jk = trΘljk Φlli . (18)
N with the entries of Jl ∈ C K ×K and vl,Φ l l k ∈ C K are given by:
Proof. The proof is provided in Appendix A.  
Theorem 2: Let A1 and A3 hold true. If RZF with VN is 1 tr Φlln Tl Φlli Tl
(RZF−VN)
[Jl ]n ,i = 2 (31)
VN
used, then γjk − γ jk −−−→ 0 almost surely while N uli 2
N →∞
1 
u2 [vl,Φ l l k ]i = tr Φlli Tl Φllk Tl . (32)
(RZF−VN)
d◦jk (1+uj kj k ) 2 N
γ jk =

L 
K 
L
|u |2 Proof: The proof is provided in the Appendix C. 
d◦li (1+u d◦lk (1+u
σ2 1 
li, jk ljk
+ 2 + 2
N ρd l N
l=1 i=1
li )
l=1,l= j
lk ) Notice that the computation of the DEs with ZF precoding
(19) (either VN or MN) for the considered multicell Massive MIMO
system is more involved than with MRT or RZF precoding
with schemes. This is mainly due to the fact that it is not straight-
(1 + uli )2 forward to start with ZF precoder and then compute the DEs
d◦li = (20) by applying common techniques, e.g., matrix inversion lemma.
1
N trΦlli T l,IN
Therefore, in proving Theorem 3 (and also Theorem 6) we start
1 with the DE of RZF and then use a bounding and limiting tech-
ulk = trΦllk Tl (21)
N nique to compute the DE for ZF.
SADEGHI et al.: LARGE SYSTEM ANALYSIS OF POWER NORMALIZATION TECHNIQUES IN MASSIVE MIMO 9009

B. Large System Results for Matrix Normalization IV. EFFECT OF POWER NORMALIZATION TECHNIQUES
MN
Next, the DEs of γjk are given for MRT, ZF, and RZF. Note In this section, we use the asymptotic approximations pro-
MN
that the DEs of γjk for MRT and RZF are obtained from [5]. vided above to gain novel insights into the interplay between the
Theorem 4: [5, Th. 4] Let A1 hold true. If MRT with MN is different system parameters and the power normalization tech-
(M RT−M N)
MN
used, then γjk − γ jk −−−→ 0 almost surely with niques in Massive MIMO. To this end, we consider a special
N →∞
case of the general channel model of (1) in which Θjlk = djlk IN
1 2 such that
(M RT−M N) λj N trΦjjk
γ jk = (33) 

L 
K 
L
σ2
+ 1
λl zli,jk + λl | N trΦljk |
1 2 hjlk = djlk zjlk (38)
N ρd l N
l=1 i=1 l=1,l= j

where zjlk ∼ CN (0, IN ) and djlk accounts for an arbitrary

where zli,jk is given in (18) and
large-scale fading coefficient including pathloss and shadowing.

1 
K
1 −1 Note this corresponds to a uncorrelated fading channel model,
λj = trΦjjk . (34) which is a quite popular model in Massive MIMO that allows
K N us to capture the essence of the technology [1], [6]. Under the
k =1
above circumstances, we have that:
K α j i −1
Theorem 5: [5, Th. 6] Let A1 and A3 hold true. If RZF with Corollary 1: Let λj = ū K1 and ū = 1 − K
(RZF−M N) i=1 d 2j j i N.
MN
MN is used, then γjk − γ jk −−−→ 0 almost surely
N →∞ If the channel is modelled as in (38), then
with
u2 d 2j j k
λj (1+uj kj k ) 2 (ZF−VN) αjk ū
(RZF−M N) γ jk = (39)
γ jk = 
L

L 
K
li, j k 
L
|u l j k |2 d2ljk
σ2
N ρd l + 1
N λl (1+u li )
2 + λl (1+u lk )
2 νjk + ū
l=1 i=1 l=1,l= j αlk
l=1,l= j
(35)   
Pilot Contam ination
with (ZF−M N) λj
γ jk = (40)
   −1 
L
d2ljk
K 1 1 Zl λl
λl = trTl − tr + αl IN T l,IN (36) νjk +
d2llk
N N N N l=1,l= j
  
where Sl = ZNl and Tl and T l,IN are given by Theorem 7 and Pilot Contam ination

Theorem 8. Also uli , uljk , and li,jk are defined in Theorem 2.

where
Theorem 6: Let A1, A2 and A4 hold true. If ZF with MN is
(ZF−M N)
MN
used, then γjk − γ jk −−−→ 0 almost surely with  
K
N →∞ L
σ2 dljk
νjk = + dljk 1 − . (41)
(ZF−M N) λj Nρ N
 dl  l=1
αlk
γ jk = (37)  

L 
K li, j k 
L uljk 2 Noise
σ2
N ρd l + 1
N λl u l i 2 + λl u l k 2 Interference
l=1 i=1 l=1,l= j
 2
K with αlk = Ln=1 dlnk + ρσt r .
with λj = ( K1 i=1 u1j i )−1 where uli , uljk , and li,jk are given Proof: See Appendix D. 
K d 2l l i −1
in Theorem 3. 1
Corollary 2: Let θl = ( K i=1 α l i ) . If the channel is
Proof: The proof follows the same procedure as the proof modelled as in (38) and MRT is used, then
of Theorem 3 presented in Appendix C. Start with the triangle
(ZF−M N) (ZF−M N)
equality and bound |γlk − γ lk |. Then find the DE d 2j j k
(ZF−M N) (RZF−M N) (M RT−VN) αjk
of γ jk by letting α → 0 in γ jk .  γ jk = (42)
 d2ljk
The asymptotic expressions provided in Theorems 1, 2, 3, and ϑjk +
6 will be shown to be very tight, even for systems with finite αlk
l=1,l= j
dimensions, by means of numerical results in Section V. This   
Pilot Contam ination
allows us to use them for evaluating the performance of practical
Massive MIMO systems without the need for time-consuming d2
(M RT−M N)
θj ( αjjjkk )2
Monte Carlo simulations. Moreover, they lay the foundation for γ jk =  2 (43)
further analysis of different configurations of Massive MIMO L
dllk dljk
ϑjk + θl
systems (e.g., distributed massive MIMO systems [34], [35]). αlk
l=1,l= j
Next, they are used to get further insights into the system under   
investigation for uncorrelated fading channels. Pilot Contam ination
9010 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 10, OCTOBER 2017

with Notice that Lemma 1 extends the results of [26] and [27] to a
system that accounts for CSI acquisition and arbitrary pathloss
σ2 K 
L
ϑjk = + dljk . (44) and UEs’ distribution. Also, observe that (39) and (40) simplify
N ρdl N as:
    l=1
 
Noise
Interference (ZF−VN) (N − K)ρdl
γ jk = dk (46)
σ2
Proof: The proof follows a similar procedure as that of
 −1
Corollary 1.  (ZF−M N) (N − K)ρdl 1  1
K

The results of Corollaries 1 and 2 are instrumental in obtaining γ jk = (47)

σ2 K di
the following insights into MRT and ZF with either MN or VN. i=1
Remark 1 (Effect of VN and MN): The terms νjk and ϑjk in from which it follows that VN provides higher SINR to the
(41) and (44) are the same for both VN and MN. This means that UEs that are closer to the BS and lower SINR for those that are
both normalization techniques have exactly the same effect on far away from the BS (which resembles opportunistic resource
the resulting noise and interference terms experienced by each allocation). On the other hand, MN provides a uniform quality
UE in the system. On the other hand, they affect differently the of experience to all UEs. This proves evidence of the fact that
signal and pilot contamination powers. The expressions (39)– ZF with VN resembles a sum rate maximizer. On the other hand,
(44) explicitly state the relation between the SINR contributions it provides a notion of fairness under MN. Notice that fairness
(signal, interference, noise, and pilot contamination), the prop- means similar SINR (quality of experience) and it should not
agation environment, and the two normalization techniques for be confused with equal power allocation. The above results and
ZF and MRT precoding schemes. observations will be validated below in Section V by means of
Remark 2 (On the mutual effect of UEs): If VN is em- numerical results. Also, the DEs provided in Corollaries 1 and 2
ployed, then the signal power and the pilot contamination of will be used to investigate the main limiting factors of Massive
UE k in cell j, for both MRT and ZF precoding, depends only MIMO.
on the coefficients dlnk ∀l, n ∈ L through αlk . This means that
they are both affected only by the large-scale gains of the UEs V. NUMERICAL RESULTS
in the network using the same pilot. On the other hand, under
MN both terms depend on the coefficients λl ∀l ∈ L (or θl for Monte-Carlo simulations are now used to validate the asymp-
MRT) and thus are influenced by all the UEs in the network, totic analysis for different values of N and K. We consider a
even though they make use of different pilot sequences. multicell network composed of L = 7 cells, one in the center
Remark 3 (Large-scale fading and power normalization): and six around. Each cell radius is 1000 meters. A 20 MHz
Assume that the large-scale fading is neglected such that it is channel is considered and the thermal noise power is assumed
the same for every UE in the network, i.e., dljk = d ∀l, j, k. to be −174 dBm/Hz. The UEs are randomly and uniformly dis-
Then, the expressions in (39) and (40) for ZF and those in tributed within each cell excluding a circle of radius 100 meters.
(42) and (43) for MRT become equal. This means that the The channel is modeled as in [36]. In particular, we assume that
1/2
large-scale fading has a fundamental impact on VN and MN the matrices Θljk are given by
and cannot be ignored.
1/2 
Consider now, for further simplicity, a single-cell setup, Θljk = dljk A (48)
i.e., L = 1. Dropping the cell index, αlk reduces to αk =
dk + σ 2 /ρtr . Also assume that the UEs operate in the high train- where A = [a(θ1 ), . . . , a(θN )] ∈ C N with a(θi ) given by
ing SNR regime such that ρtr  1. Under these conditions, we 1
have that: a(θi ) = √ [1, e−i2π ω sin(θ i ) , . . . , e−i2π ω (N −1) sin(θ i ) ]T (49)
N
Lemma 1: If L = 1 and ρtr  1, then for ZF precoding, VN
outperforms MN in terms of sum rate and the sum rate gap where ω = 0.3 is the antenna spacing and θi = −π/2 + (i −
Δr ≥ 0 is given by 1)π/N . Also, dljk is the large-scale attenuation, which is mod-
  eled as dljk = x−β
ljk where xljk denotes the distance of UE k in
K
1 cell j from BS l and β = 3.7 is the path-loss exponent. We let
Δr = log 1 + σ 2 1
k =1 N ρ d l ū d k
ρtr = 6 dB and ρdl = 10 dB, which corresponds to a practical
⎛ ⎞ setting [5]. The results are obtained for 100 different channel
and UE distributions realizations.
⎜ ⎟
⎜ 1 ⎟ Figs. 1 and 2 validate the accuracy of the DEs provided in
− K log ⎜1 + ⎟. (45)
⎝ σ2 1

K
1 ⎠ Theorems 1, 2, 3, and 6. In particular, both figures report the er-
N ρ d l ū K di godic achievable sum rate of the center cell versus N for K = 8
i=1
and 16, respectively. The solid lines correspond to the asymp-
Proof: From Corollary 1, setting L = 1 and assuming ρtr  totic sum rate whereas the markers are achieved through Monte
2
1 we obtain that αk  dk and νk  Nσρ d l . Then, the result fol- Carlo simulation. As it is depicted, the asymptotic approxima-
lows by applying the Jensen’s inequality (by the convexity of tion match perfectly with numerical results. Notice that Figs. 1
log (1 + 1/x)).  and 2 (and also Table I) extend the results in [26] and [27] in
SADEGHI et al.: LARGE SYSTEM ANALYSIS OF POWER NORMALIZATION TECHNIQUES IN MASSIVE MIMO 9011

TABLE I
SINR OF EACH UE UNDER ZF WITH VN AND MN

No. Ant. UE MN DE MN MC Er. % VN DE VN MC Er. %

N = 40 1 2.25 2.19 2.6 1.85 1.84 0.5

2 4.89 4.84 1.0 147 149 1.3
3 3.34 3.29 1.5 3.61 3.53 2.2
4 5.14 5.12 0.3 37.5 37.6 0.2
5 4.09 4.02 1.7 1.97 1.96 0.5
6 4.26 4.41 3.5 85 87 2.3
7 3.30 3.33 0.9 2.14 2.2 2.8
8 3.52 3.50 0.5 2.52 2.49 1.2
N = 80 1 3.20 3.15 1.5 3.1 2.98 3.8
2 10.57 10.40 1.6 316 310 1.9
3 5.94 6.03 1.5 6.40 6.36 0.6
4 9.33 9.47 1.5 72.2 72.2 0
5 8.41 8.62 2.5 3.87 3.90 0.7
6 9.05 9.12 0.7 182 185 1.6
7 5.02 4.90 2 3.50 3.43 2
8 5.89 5.74 2.5 4.28 4.21 1.6

Fig. 1. Ergodic achievable sum rate of center cell for MRT, ZF, and RZF with
VN and ZF with MN for K = 8.
provides a more uniform experience for all UEs, while ZF with
VN provides very high SINRs to specific UEs (UEs 2, 4, and
6) and much lower SINRs to others. More precisely, the SINR
variance with MN is equal to 0.8 (5.79) for N = 40 (N = 80),
while for VN it is equal to 2627 (11550 for N = 80). Notice also
that the percentage of error is always less than 4%, which proves
the high accuracy of the DEs. Therefore, one can simply use the
DEs to achieve insight into the network performance, instead of
using time-consuming Monte Carlo simulations. Moreover, the
DEs do not contain any randomness and are purely based on the
large-scale statistics of the network. Hence, they can be used for
network optimization purposes.
The DEs given in Corollaries 1 and 2 and Theorems 2 and 5 are
now used to investigate a common belief in the Massive MIMO
literature, that is: under uncorrelated fading when N → ∞ the
noise and interference contributions vanish asymptotically and
pilot contamination becomes the unique bottleneck of the system
performance. This follows also from the results in Corollaries 1
and 2 by letting N grow large with K kept fixed. However, in
[29] it is shown that it is desirable for Massive MIMO systems
Fig. 2. Ergodic achievable sum rate of center cell for MRT, ZF, and RZF with to work in a regime where K N
≤ 10. Therefore, it is interesting
VN and ZF with MN for K = 16. to see what is the major impairment for Massive MIMO under
this practical regime: (i) is it pilot contamination (or coherent
interference)?; (ii) is it the noise and interference (or more
the sense that account for CSI acquisition, pilot contamination, exactly the non-coherent interference)?; (iii) how is the answer
arbitrary pathloss and UEs’ distribution. related to the choice of the power normalization technique and
In Lemma 1, it is shown that ZF under VN conveys a no- precoding scheme?
tion of sum rate maximization, while ZF with MN resembles a To answer these questions, we employ the so-called pilot
fairness provisioning precoder. Now, we use Table I to validate contamination-to-interference-plus-noise ratio (PCINR) metric,
this observation and also to verify the accuracy of the computed which is computed by using the DEs provided in Corollaries 1
DEs for the simplified channel model in (38). The first column and 2. Fig. 3 plots the PCINR as a function of N/K, i.e., the
of Table I reports the number of antennas, the second one is number of degrees of freedom per-user in the system. Although,
the UE index. The third and fourth columns are the asymptotic the optimal operating regime for maximal spectral efficiency is
(as given in (40)) and simulated SINRs of each UE under MN. for N/K < 10 [29], we consider N/K up to 20 to cover a
The corresponding results for VN are reported in the sixth and wider range of Massive MIMO configurations. Moreover, as
seventh columns. The fifth and eighth columns report the per- the interference increases by having more UEs in the system,
centage of the error while estimating a specific UE SINR with we consider three different scenarios with K = 5, K = 10, and
the computed DEs. As predicted by Lemma 1, ZF with MN K = 15.
9012 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 10, OCTOBER 2017

Fig. 3. PCINR versus the degree of freedom per-user, for different values of K . (a) K = 5. (b) K = 10. (c) K = 15.

Fig. 3 is divided into 3 regions based on the significance of the the other schemes (other than MRT-MN), Massive MIMO of-
PCINR term such that, as we move away from region 1 towards ten operates within this region. This shows that, although pilot
region 3, the importance of pilot contamination increases while contamination is a major challenge in Massive MIMO, the in-
that of the interference plus noise reduces. Region 1 is where terference and noise have still the leading role in limiting the
the noise and interference are the dominant limiting factors and system performance.
pilot contamination has a negligible effect—less than 10% of Finally, region 3 presents the superiority of pilot contam-
the noise and interference. As it is depicted, MRT with MN ination effect. If K = 10, then Fig. 3(b) shows the superi-
operates within this regime, therefore pilot contamination is ority of interference and noise over pilot contamination for
never a bottleneck for this scheme, which is mainly limited by ZF-MN and RZF-MN (ZF-VN and RZF-VN) up to N = 130
noise and interference. Notice that by adding more UEs in the (N = 233) antennas at the BS. With MRT-VN, the system re-
system, the PCINR reduces and pilot contamination becomes quires more than N = 510 to experience the superiority of pilot
even less important. Hence, when MRT with MN is studied in contamination over interference and noise. This increases to
Massive MIMO the effect of pilot contamination can be safely N = 2650 with MRT-MN. From Fig. 3, we see also that, for
neglected. a given value of N/K, the value of PCINR for the considered
Region 2 represents the regime where the noise and inter- schemes can be ordered as: ZF-MN = RZF-MN ≥ ZF-VN =
ference are the main limiting factors, but pilot contamination RZF-VN ≥ MRT-VN ≥ MRT-MN. Based on the above dis-
is not negligible any more. It is interesting to observe that for cussion, it is clear that the choice of precoding scheme and
SADEGHI et al.: LARGE SYSTEM ANALYSIS OF POWER NORMALIZATION TECHNIQUES IN MASSIVE MIMO 9013

normalization technique change the importance of pilot con- and Cl = N Cl and Cl[k ] = N Cl[k ] with
tamination, interference, and noise dramatically and it should
−1
be considered carefully when designing Massive MIMO Cl = Ĥll ĤH ll + Zl + αl N I N (52)
systems.

−1
Cl[k ] = Ĥll[k] ĤH
ll[k] + Zl + αl N I N . (53)
VI. CONCLUSION
(RZF−VN)
Linear precoding schemes, such as MRT and ZF, have a fun- Plugging (9) and (10) into (14) yields γjk . Then, we
damental role in Massive MIMO. Although these precoding divide the numerator and denominator by N and replace each
(RZF−VN)
schemes can be employed with optimized power control poli- term in γjk with its DE. Notice that DEs of signal power
cies, they are usually implemented by simple matrix or vector component, variance component, and interference components
power normalization techniques. This is due to the complexity are given in [5]. Therefore, we only need to compute a DE of
of attaining optimal power control policies [24], as it requires the VN coefficient. From (10), we have that
coordination and cooperation among cells and computationally 1 1
demanding algorithms. On the other hand, the two above pre- d◦li = dli = (E[ĥH −1
lli Cl ĥlli ]) .
2
(54)
N N
coding power normalization techniques are simple and efficient
[2], [3], [5], [6]. An asymptotic approximation for ĥH
lli Cl ĥlli can be computed
2

This work made use of large system analysis to compute tight as follows
2
asymptotic approximations of the SINR experienced by each a lli Cl[i] ĥlli
ĥH 2
b N 2 trΦlli Cl
1
UE in the system when using MN or VN. These results can be ĥH
lli Cl ĥlli =
2

(1 + ĥH
lli Cl[i] ĥlli )
2 (1 + N1 trΦlli Cl )2
used to evaluate the performance of practical Massive MIMO

N 2 trΦlli Tl,I N
systems without the need for time-consuming Monte Carlo sim- 1
c
ulations. Under uncorrellated fading channels, we analytically  (55)
(1 + N1 trΦlli T−1l )
2
showed that MN and VN treat the noise and interference in
the same manner, but have different effects on pilot contam- where (a) follows from Lemma 2 in Appendix E, (b)1 is obtained
ination and received signal power. We also revealed the key by applying Lemmas 3 and 4, and (c) follows from Theorems 7
role played by large-scale fading, positions of UEs, and pilot and 8 with Sl = ZNl . By applying the continuous mapping the-
assignment into power normalization. We explained how a sim- orem and the dominated convergence theorem yields
ple change in power normalization can resemble two totally (1 + uli )2
different behaviors, namely, sum-rate maximization or fairness d◦li  . (56)
1
N trΦlli T l,IN
provisioning. Moreover, we showed numerically how the choice
of the normalization technique can change the main bottleneck
APPENDIX C
of Massive MIMO systems.
The main idea is to first compute a DE for RZF with αl = α
APPENDIX A ∀l ∈ L and then to obtain a DE for ZF by letting α → 0. By using
(ZF−VN) (ZF−VN)
VN
the triangle inequality |γlk − γ lk | can be bounded
We begin by plugging (7) and (10) into γjk given in (14) to as follows
(M RT−VN)
obtain γjk . Then, we divide the numerator and denom- (ZF−VN) (ZF−VN) (ZF−VN) (RZF−VN)
|γlk − γ lk | ≤ |γlk − γlk
|+
d†li
(M RT−VN)
inator of γjk by N and define = N dli . By apply-
(RZF−VN) (RZF−VN) (RZF−VN) (ZF−VN)
ing the continuous mapping theorem [31] and replacing each + |γlk − γ lk | + |γ lk − γ lk |. (57)
(M RT−VN) (M RT−VN)
component of γjk by its DE, the DE of γjk Next, we show that each term in the right hand side of (57) can be
is obtained. Notice that DEs of signal power component, vari- made arbitrarily small (i.e. smaller than any given ε > 0) when
ance component, and interference components are given in [5]. (ZF−VN) (RZF−VN)
α → 0. Let us start with |γlk − γlk |. Notice that
Therefore, we only need to compute a DE of the VN coefficient. (ZF−VN) (RZF−VN)
The latter is given by γlk and γlk are different because of the different
form of Fl in (8) and (9). As α → 0 and for Zl = 0, we have
1 1 (a) 1
−1
dli = = = (50) lim f RZF
= lim Ĥ ĤH
+ N αI Ĥll ek
E[ĥ ĥ ]
H
lli trE[ĥ ĥ ]
lli
H
llitrΦllilli lk
Z l =0,α →0
ll ll
α →0
N

−1
where in (a) we have applied Lemma 3 in Appendix E and used = Ĥll ĤH Ĥ ek = flkZF . (58)
ll ll
the fact that ĥlli ∼ CN (0, Φlli ).
(ZF−VN) (RZF−VN)
Therefore, the term |γlk − γlk | can be made ar-
APPENDIX B bitrarily small α → 0. By applying Theorem 2, we have that the
(RZF−VN) (RZF−VN)
Let us define for convenience second term is such that |γlk − γ lk |  0 for any

ll[k] = Ĥll Ĥll − ĥllk ĥllk

Ĥll[k] ĤH H H
(51) 1a
N  bN is equivalent to a N − bN −−−−→ 0.
N →∞
9014 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 10, OCTOBER 2017

α > 0. Consider now the third term |γ lk − γ lk |.

(RZF−VN) (ZF−VN) d◦li = uli . On the other hand, for li,jk we have
(ZF−VN) (RZF−VN)
Let us define γ lk = limα →0 γ lk . Observe that li,jk = lim α2 li,jk
α →0
lim γ jk
(RZF−VN)  
α →0 1 2 ∗ 1
= trΘljk Tl,Φ l l i − Re uljk trΦljk Tl,Φ l l i
u2 N ulk N
d◦jk (1+uj kj k ) 2
= lim |uljk |2 1
α →0 
L 
K 
L
|u l j k |2 + trΦllk T l,Φ l l i (65)
d◦li (1+u d◦lk (1+u
σ2 1 li, j k
N ρd l + N li )
2 + lk )
2 ulk 2 N
l=1 i=1 l=1,l= j

α 2
u 2j k where T l,Φ l l i  limα →0 α2 T l,Φ l l i is
d◦jk (α +α u j k ) 2
= lim .  

L   1  u lt,Φ l l i Φllt
α →0 K L K
α 2 |α u |2
σ2
N ρd l + 1
N d◦li (α +αl iu, jlki ) 2 + d◦lk (α +αl jukl k ) 2 T l,Φ l l i = lim α Tl 2
+ Φlli Tl
l=1 i=1 l=1,l= j α →0 N (1 + ult )2
t=1
(59)  
1  ult,Φ l l i Φllt
K
Define ulk := limα →0 αulk for every l and k. Based on [16] and = Tl + Φlli Tl (66)
N ult 2
by replacing ulk from Theorem 2 we have t=1

 −1 where u lt,Φ l l i = limα →0 α2 u lt,Φ l l i . From Theorem 8, we have

1  Φlli
K
1
ulk = lim αulk = lim trΦllk + IN
α →0 α →0 N N αuli
i=1
u l,Φ l l i = lim (IK − Jl )−1 α2 vl,Φ l l i = (IK − Jl )−1 vl,Φ l l i
α →0
1 
= tr Φllk Tl (60) where Jl and vl,Φ l l i are given by (31) and (32), respectively.
N
Therefore, lk,jn = limα →0 α2 lk,jn follows (28). Using all the
with above results in (59) completes the proof.
 −1
1  Φlli
K
Tl = + IN . (61) APPENDIX D
N uli
i=1 For brevity we only consider ZF with VN. The same steps
can be used for ZF with MN. If the channel is modelled as in
Also defining uljk  limα →0 αuljk for every l, j, and k we have (38), then Θljk = dljk IN and
1 1 dllk dljk
uljk = lim α trΦljk Tl = trΦljk Tl . (62) Φljk = IN (67)
α →0 N N αlk
For the term limα →0 d◦li , we obtain L σ2
with αlk = n =1 dlnk + ρ t r . Plugging (67) into (25) and (26)
d 2l l k 1
(1 + uli )2 yields ulk = α l k N tr (Tl ) with
lim d◦li = lim
α →0 α →0 1
N trΦlli T l,IN  −1
1 
K
1
uli 2
uli 2
Tl =  + 1 IN . (68)
N tr Tl
N 1
= lim
= lim 2 (63)
α →0 α2 N1 trΦlli Tl,IN α →0 α uli,IN i=1

1
Call ū = tr (Tl ). Therefore, we have that
where T l,IN and u li,IN are given in Theorem 8 in Appendix E. N

Notice that  −1

1 K1
ū = tr (Tl ) = +1 . (69)
1 N N ū
lim α2 u li,IN = α2 trΦlli T l,IN
α →0 N
  Solving with respect to ū yields ū = 1 − K
N . Then, we eventu-
1  u lt,IN Φllt
K
2 1 ally have that
= lim α trΦlli Tl + IM Tl
α →0 N N (1 + ult )2
t=1 d2llk
  ulk =
αlk
ū (70)
1  (limα →0 α2 u lt,IN )Φllt
K
1
= trΦlli Tl + IN Tl dllk dljk
N N ult 2 and also uljk = α l k ū. Therefore, the pilot contamination
t=1
(64) (ZF−VN)
term in γ jk reduces to
from which, by replacing limα →0 α2 u lt,IN with ult , we have 
L u2ljk 
L
d2ljk
that (64) reduces to uli = N1 trΦlli Tl . Therefore, we have that ulk
=
αlk
ū. (71)
limα →0 α2 u lk ,IN = ulk . From (63), we can thus conclude that l=1,l= j l=1,l= j
SADEGHI et al.: LARGE SYSTEM ANALYSIS OF POWER NORMALIZATION TECHNIQUES IN MASSIVE MIMO 9015

Let’s now compute [Jl ]n ,i defined as in (31). Using the above N

Then, for any αl > 0, as N and K grow large with β = K such
results yields that 0 < lim inf β ≤ lim sup β < ∞ we have that
1 d2lln d2lli 1  1 d2lln αli mB l ,Q l (−αl ) − m◦B l ,Q l (−αl )  0 (81)
[Jl ]n ,i = tr T2 = . (72)
2
N αln αli uli 2 N αln d2lli
where m◦B l ,Q l (−αl ) is given by
Similarly, we have that
1
m◦B l ,Ql (−αl ) = trQl Tl (82)
d2 d2 N
[vl,k ]i = lli llk ū2 . (73)
αli αlk with Tl is given by
In compact form, we may write Jl and vl,k as  −1
1 
K
Φlli
Tl = + Sl + αl IN (83)
1 d2llk 2 N 1 + uli (−αl )
Jl = al bTl vl,k = ū al (74) i=1
N αlk
where the elements of ul (−αl ) = [ul1 (−αl ), . . . , ulK (−αl )]T
with [al ]i = d2lli /αli and [bl ]i = 1/[al ]i . Then, we have that (by (t)
are defined as uli (−αl ) = limt→∞ uli (−αl ), where for t ∈
applying Lemma 2)
{1, 2, . . .}
 −1
d2 1 d2
u l,k = llk ū2 IK − al bTl al = llk ūal = ulk al . (75)
(t)
ulk (−αl ) =
αlk N αlk
 −1
1 
K
Plugging the above result into (29) produces 1 Φlli
  trΦllk (t−1)
+ Sl + αl IN (84)
N N
d2lli K1 d2 i=1 1 + uli (−αl )
Tl,Φ l l i = Tl + 1 Tl = lli ūIN = uli IN . (76)
αli N ū αli (0)
with initial values ulk (−α) = α1 for all k.
We are thus left with evaluating (28). Using the above results Theorem 8: [16] Let Ωl ∈ C N ×N be Hermitian nonnegative
yields definite with uniformly bounded spectral norm (with respect to
  N). Under the conditions of Theorem 1
dljk dljk 1
li,jk = tr T l,Φ l l i − 2 tr Φljk T l,Φ l l i +  −1
N dllk N 1 1
trQl Ĥll ĤH
ll + S l + α I
l N Ωl (85)
d2ljk 1  N N
+ tr Φllk T l,Φ l l i (77)  −1
d2llk N 1 1
Ĥll ĤH
ll + S l + α I
l N − trQl T l,Ω l  0
from which, using (67) and (76), we obtain N N

d2ljk where T l,Ω l ∈ C N ×N is defined as

li,jk = dljk uli − uli . ⎛ ⎞
αlk
1 K
u
(−α )Φ
T l,Ω l = Tl × ⎝ + Ωl ⎠ × Tl (86)
lj,Ω l l llj
Therefore, we have that
  N (1 + ulj (−αl ))2
j =1
1 li,jk 1 dljk
= dljk 1− . (78)
N uli N αlk where Tl and ul (−α) are given by theorem 1, and u l,Ω l (−α) =
[u l1,,Ω l (−α), . . . , u lK ,,Ω l (−α)]T is computed from
Plugging (70), (71) and (78) into (24) produces
  u l,Ω l (−α) = (IK − Jl )−1 vl,Ω l (87)
1   li,jk K
L K L
dljk
= dljk 1 − . (79)
N uli N αlk where Jl ∈ C K ×K and vl ∈ C K are:
l=1 i=1 l=1
trΦllm Tl Φlln Tl
Collecting all the above results together completes the proof. [Jl ]m n = 1 ≤ m, n ≤ K (88)
N 2 (1 + uln (−α))2
APPENDIX E 1
[vl,Ω l ]t1 = trΦllt Tl Ωl Tl 1 ≤ t ≤ K (89)
USEFUL RESULTS N
Theorem 7: [16, Th. 1] Let Bl = N1 Ĥll ĤH Lemma 2 (Matrix Inversion Lemma): Let U be an N × N
ll + Sl with
Ĥll ∈ C N ×K
be random with independent column vectors invertible matrix and x ∈ C N , c ∈ C for which U + c xxH is
ĥllk ∼ CN (0, Φllk ) for k ∈ {1, . . . , K}, Sl ∈ C N ×N and Ql ∈ invertible. Then
C N ×N be Hermitian nonnegative definite. Assume that Ql and −1 xH U−1
xH U + c xxH = . (90)
the matrices Φllk for k ∈ {1, . . . , K} have uniformly bounded 1 + c xH U−1 x
spectral norms (with respect to N ). Define
Lemma 3 (Trace Lemma): Let A ∈ C N ×N and x, y ∼
1 CN (0, N1 IN ). Assume that A has uniformly bounded spec-
mB l ,Q l (−αl )  trQl (Bl + αl IN )−1 (80)
N tral norm (with respect to N) and that x and y are mutually
9016 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 66, NO. 10, OCTOBER 2017

independent and independent of A. Then, for all p ≥ 1, [18] A. Khansefid and H. Minn, “Achievable downlink rates of MRC and ZF
precoders in massive MIMO with uplink and downlink pilot contamina-
1 tion,” IEEE Trans. Commun., vol. 63, no. 12, pp. 4849–4864, Dec. 2015.
xH Ax − trA  0 and xH Ay  0. (91) [19] A. Kammoun, A. Müller, E. Björnson, and M. Debbah, “Linear precoding
N
based on polynomial expansion: Large-scale multi-cell MIMO systems,”
Lemma 4 (Rank-1 perturbation lemma): Let A1 , A2 , ..., IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 861–875, Oct. 2014.
with AN ∈ C N ×N , be deterministic with uniformly bounded [20] R. Krishnan et al., “Linear massive MIMO precoders in the presence of
phase noise—A large-scale analysis,” IEEE Trans. Veh. Technol., vol. 65,
spectral norm and B1 , B2 , ..., with BN ∈ C N ×N , be random no. 5, pp. 3057–3071, May 2016.
Hermitian, with eigenvalues λB 1
N
≤ λB2
N
≤ . . . ≤ λB
N
N
such [21] X. Li, E. Bjornson, E. G. Larsson, S. Zhou, and J. Wang, “A multi-
BN cell MMSE precoder for massive MIMO systems and new large system
that, with probability 1, there exist  > 0 for which λ1 >  analysis,” in Proc. IEEE Int. Conf. Global Commun., San Diego, CA,
for all large N . Then for v ∈ C N USA, Dec. 2015, pp. 1–6.
[22] M. Sadeghi and C. Yuen, “Multi-cell multi-group massive MIMO multi-
1 1
trAN B−1
N − trAN (B−1 H −1
N + vv )  0. (92) casting: An asymptotic analysis,” in Proc. IEEE Global Commun. Conf.,
San Diego, CA, USA, Dec. 2015, pp. 1–6.
N N
[23] T. V. Chien, E. Bjornson, and E. G. Larsson, “Joint power allocation and
where B−1 −1 H −1
N and (BN + vv ) exist with probability 1. user association optimization for massive MIMO systems,” IEEE Trans.
Wireless Commun., vol. 15, no. 9, pp. 6384–6399, Sep. 2016.
[24] Q. Zhang, S. Jin, M. McKay, D. Morales-Jimenez, and H. Zhu, “Power
REFERENCES allocation schemes for multicell massive MIMO systems,” IEEE Trans.
Wireless Commun., vol. 14, no. 11, pp. 5941–5955, Nov. 2015.
[1] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers [25] M. Sadeghi, L. Sanguinetti, R. Couillet, and C. Yuen, “Reducing the
of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 11, computational complexity of multicasting in large-scale antenna systems,”
pp. 3590–3600, Nov. 2010. IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 2963–2975, May 2017.
[2] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with [26] C. Lee, C. B. Chae, T. Kim, S. Choi, and J. Lee, “Network massive MIMO
very large arrays,” IEEE Signal Proc. Mag., vol. 30, no. 1, pp. 40–60, for cell-boundary users: From a precoding normalization perspective,”
Jan. 2013. in Proc. IEEE Globecom Workshops, Anaheim, CA, USA, Dec. 2012,
[3] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive pp. 233–237.
MIMO for next generation wireless systems,” IEEE Commun. Mag., [27] Y.-G. Lim, C.-B. Chae, and G. Caire, “Performance analysis of massive
vol. 52, no. 2, pp. 186–195, Feb. 2014. MIMO for cell-boundary users,” IEEE Trans. Wireless Commun., vol. 14,
[4] L. Liu, C. Yuen, Y. L. Guan, Y. Li, and Y. Su, “Convergence analysis no. 12, pp. 6827–6842, Dec. 2015.
and assurance for Gaussian message passing iterative detector in massive [28] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser MIMO
MU-MIMO systems,” IEEE Trans. Wireless Commun., vol. 15, no. 9, achievable rates with downlink training and channel state feedback,” IEEE
pp. 6487–6501, Sep. 2016. Trans. Inf. Theory, vol. 56, no. 6, pp. 2845–2866, Jun. 2010.
[5] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the UL/DL [29] E. Bjornson et al., “Massive MIMO: Ten myths and one critical question,”
of cellular networks: How many antennas do we need?” IEEE J. Sel. Areas IEEE Commun. Mag., vol. 54, no. 2, pp. 114–123, Feb. 2016.
Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013. [30] M. Medard, “The effect upon channel capacity in wireless communications
[6] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral of perfect and imperfect knowledge of the channel,” IEEE Trans. Inf.
efficiency of very large multiuser MIMO systems,” IEEE Trans. Commun., Theory, vol. 46, no. 3, pp. 933–946, May 2000.
vol. 61, no. 4, pp. 1436–1449, Apr. 2013. [31] P. Billingsley, Probability and Measure. New York, NY, USA: Wiley,
[7] J. Zuo, J. Zhang, C. Yuen, W. Jiang, and W. Luo, “Energy-efficient down- 2008.
link transmission for multicell massive DAS with pilot contamination,” [32] L. Sanguinetti, A. L. Moustakas, and M. Debbah, “Interference manage-
IEEE Trans. Veh. Technol., vol. 66, no. 2, pp. 1209–1221, Feb. 2017. ment in 5G reverse TDD hetnets with wireless backhaul: A large system
[8] J. G. Andrews et al., “What will 5G be?” IEEE J. Sel. Areas Commun., analysis,” IEEE J. Sel. Areas Commun., vol. 33, no. 6, pp. 1187–1200,
vol. 32, no. 6, pp. 1065–1082, Jun. 2014. Jun. 2015.
[9] T. L. Marzetta, “How much training is required for multiuser MIMO?” [33] L. Sanguinetti, A. L. Moustakas, E. Björnson, and M. Debbah, “Large sys-
in Proc. IEEE Asilomar Conf. Signals, Syst. Comput., Pacific Grove, CA, tem analysis of the energy consumption distribution in multi-user MIMO
Nov. 2006, pp. 359–363. systems with mobility,” IEEE Trans. Wireless Commun., vol. 14, no. 3,
[10] E. Björnson, J. Hoydis, and L. Sanguinetti, “Pilot contamination is pp. 1730–1745, Mar. 2015.
not a fundamental asymptotic limitation in massive MIMO,” in Proc. [34] K. T. Truong and R. W. Heath, “The viability of distributed antennas for
IEEE Int. Conf. Commun., Paris, France, May 2017. [Online]. Available: massive MIMO systems,” in Proc. 2013 Asilomar Conf. Signals, Syst.
https://arxiv.org/abs/1611.09152 Comput., Pacific Grove, CA, USA, Nov. 2013, pp. 1318–1323.
[11] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO has unlimited [35] M. Sadeghi, C. Yuen, and Y. H. Chew, “Sum rate maximization for uplink
capacity,” IEEE Trans. Wireless Commun., 2017, submitted for publica- distributed massive MIMO systems with limited backhaul capacity,” in
tion. [Online]. Available: https://arxiv.org/abs/1705.00538 Proc. 2014 IEEE Globecom Workshops, Austin, TX, USA, Dec. 2014,
[12] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO pp. 308–313.
for next generation wireless systems,” IEEE Commun. Mag., vol. 52, [36] H. Q. Ngo, T. L. Marzetta, and E. G. Larsson, “Analysis of the pilot
no. 2, pp. 186–195, Feb. 2014. contamination effect in very large multicell multiuser MIMO systems for
[13] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, physical channel models,” in Proc. 2011 IEEE Int. Conf. Acoust., Speech
“Five disruptive technology directions for 5G,” IEEE Commun. Mag., Signal Process., Prague, Czech, May 2011, pp. 3464–3467.
vol. 52, no. 2, pp. 74–80, Feb. 2014.
[14] H. Yang and T. L. Marzetta, “Performance of conjugate and zero-forcing
beamforming in large-scale antenna systems,” IEEE J. Sel. Areas Com-
Meysam Sadeghi received the B.S. and M.S. degrees
mun., vol. 31, no. 2, pp. 172–179, Feb. 2013.
in electrical engineering from Shiraz University of
[15] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot contam-
Technology, Shiraz, Iran, and K. N. Toosi University
ination and precoding in multi-cell TDD systems,” IEEE Trans. Wireless
of Technology, Tehran, Iran, respectively. He is cur-
Commun., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
rently working toward the Ph.D. degree at Singapore
[16] S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large system analysis
University of Technology and Design, Singapore. He
of linear precoding in correlated MISO broadcast channels under lim-
has been a visiting Research Scholar at Linköping
ited feedback,” IEEE Trans. Inf. Theory, vol. 58, no. 7, pp. 4509–4537,
university, CentraleSupélec, and university of Pisa.
Jul. 2012.
His research interest includes wireless communica-
[17] J. Zuo, J. Zhang, C. Yuen, W. Jiang, and W. Luo, “Multi-cell multi-user
tion and signal processing, and in particular Massive
massive MIMO transmission with downlink training and pilot contamina-
MIMO systems. He received the Merlion, SPGF, and
tion precoding,” IEEE Trans. Veh. Technol., vol. 65, no. 8, pp. 6301–6314,
SINGA awards.
Aug. 2015.
SADEGHI et al.: LARGE SYSTEM ANALYSIS OF POWER NORMALIZATION TECHNIQUES IN MASSIVE MIMO 9017

Luca Sanguinetti (SM’15) received the Laurea Chau Yuen (S’02–M’08–SM’12) received the B.Eng
Telecommunications Engineer degree (cum laude) and Ph.D. degrees from Nanyang Technological
and the Ph.D. degree in information engineering from University, Singapore, in 2000 and 2004, respec-
the University of Pisa, Pisa, Italy, in 2002 and 2005, tively. He was a Postdoctoral Fellow in Lucent
respectively. Since 2005, he has been with the Dipar- Technologies Bell Laboratory, Murray Hill, New
timento di Ingegneria dell’Informazione, University York, during 2005. He was a Visiting Assistant Pro-
of Pisa. During the period June 2007–June 2008, he fessor at Hong Kong Polytechnic University in 2008.
was a Postdoctoral Associate in the Department of During the period of 2006–2010, he was at the
Electrical Engineering, Princeton University. During Institute for Infocomm Research, Singapore, as a Se-
the period June 2010–September 2010, he was se- nior Research Engineer. He joined Singapore Univer-
lected for a research assistantship at the Technische sity of Technology and Design from June 2010. He
Universität Munchen. From July 2013 to July 2015, he was with the Alcatel- was an Associate Editor of the IEEE TRANSACTIONS ON VEHICULAR TECHNOL-
Lucent Chair on Flexible Radio, Supélec, Gif-sur-Yvette, France. He is an OGY. On 2012, he received the IEEE Asia-Pacific Outstanding Young Researcher
Assistant Professor at the Dipartimento di Ingegneria dell’Informazione, Uni- Award.
versity of Pisa. He is currently an Associate Editor for the IEEE TRANSACTIONS
ON WIRELESS COMMUNICATIONS, the IEEE SIGNAL PROCESSING LETTERS. He
was a Lead Guest Editor of the IEEE JOURNAL ON SELECTED AREAS OF COM-
MUNICATIONS Special Issue on “Game Theory for Networks” and an Associate
Editor for the IEEE JOURNAL ON SELECTED AREAS OF COMMUNICATIONS (se-
ries on Green Communications and Networking). His research interests include
communications and signal processing, game theory, and random matrix theory
for wireless communications. He received two best paper awards: the IEEE
Wireless Commun. and Networking Conference 2013 and the IEEE Wireless
Commun. and Networking Conference 2014. He also received the FP7 Marie
Curie IEF 2013 “Dense deployments for green cellular networks.”

Romain Couillet (SM’15) received the M.Sc.

degree in mobile communications from the Eure-
com Institute, Biot, France, the M.Sc. degree in
communication systems from Telecom ParisTech,
Paris, France, in 2007, and the Ph.D. degree from
Supelec, Gif-sur-Yvette, France, in November 2010.
From 2007 to 2010, he was with ST-Ericsson as an
Algorithm Development Engineer on the Long Term
Evolution Advanced project. He is currently a Full
Professor in the LSS laboratory, CentraleSupélec,
France. His research interest include random matrix
theory applied to statistics, machine learning, signal processing, and wireless
communications. In 2015, he received the HDR title from University ParisSud.
He received the 2013 CNRS Bronze Medal in the section “science of infor-
mation and its interactions,” of the 2013 IEEE ComSoc Outstanding Young
Researcher Award (EMEA Region), of the 2011 EEA/GdR ISIS/GRETSI best
Ph.D. thesis award, and of the Valuetools 2008 best student paper award.