1

Massive MIMO for Maximal Spectral Efficiency:

How Many Users and Pilots Should Be Allocated?
Emil Björnson, Member, IEEE, Erik G. Larsson, Senior Member, IEEE, and Mérouane Debbah, Fellow, IEEE

Abstract—Massive MIMO is a promising technique to increase per cell, while maintaining robustness to inter-user interfer-
the spectral efficiency (SE) of cellular networks, by deploying ence. The research on massive MIMO has so far focused on
antenna arrays with hundreds or thousands of active elements at establishing the fundamental physical (PHY) layer properties;
the base stations and performing coherent transceiver processing.
A common rule-of-thumb is that these systems should have an in particular, that the acquisition of channel state information
order of magnitude more antennas, M , than scheduled users, (CSI) is limited by the channel coherence block (i.e., the fact
K, because the users’ channels are likely to be near-orthogonal that channel responses are only static in limited time/frequency
arXiv:1412.7102v3 [cs.IT] 20 Oct 2015

when M/K > 10. However, it has not been proved that this rule- blocks) and how this impacts the SEs and the ability to
of-thumb actually maximizes the SE. In this paper, we analyze mitigate inter-cell interference [2], [6], [7]. In addition, the
how the optimal number of scheduled users, K ? , depends on M
and other system parameters. To this end, new SE expressions aggressive multiplexing in massive MIMO has been shown to
are derived to enable efficient system-level analysis with power provide major improvements in the overall energy efficiency
control, arbitrary pilot reuse, and random user locations. The [8]–[11], while [12]–[14] have shown that the hardware im-
value of K ? in the large-M regime is derived in closed form, pairments of practical transceivers have smaller impact on
while simulations are used to show what happens at finite M , in massive MIMO than contemporary systems. In contrast, the
different interference scenarios, with different pilot reuse factors,
and for different processing schemes. Up to half the coherence research community has only briefly touched on the resource
block should be dedicated to pilots and the optimal M/K is less allocation problems in the media access control (MAC) layer
than 10 in many cases of practical relevance. Interestingly, K ? (e.g., user scheduling)—although the truly achievable SEs can
depends strongly on the processing scheme and hence it is unfair only be understood if the PHY and MAC layers are jointly
to compare different schemes using the same K. optimized.
Index Terms—Coordinated multipoint, massive MIMO, multi- The importance of resource allocation for massive MIMO
cell, pilot contamination, spectral efficiency, user scheduling. was described in [15], where initial guidelines were given. A
main insight is that the limited number of orthogonal pilot
I. I NTRODUCTION sequences needs to be allocated intelligently among the UEs
to reduce interference, which can be done by capitalizing on
Cellular communication networks are continuously evolving pathloss differences [16], [17] and spatial correlation [15],
to keep up with the rapidly increasing demand for wireless [18], [19].
data services. Higher area throughput (in bit/s per km2 ) has In this paper, we consider a related resource allocation
traditionally been achieved by a combination of three multi- question: how many UEs should be scheduled per cell to
plicative factors [1]: more frequency spectrum (Hz), higher cell maximize the spectral efficiency? This question has, to the best
density (more cells per km2 ), and higher spectral efficiency of our knowledge, not been answered for multi-cell systems.1
(bit/s/Hz/cell). This paper considers the latter and especially We show how the coherence block length, number of antennas,
the massive multiple-input multiple-output (MIMO) concept, pilot allocation, hardware impairments, and other system pa-
proposed in [2], which has been identified as the key to rameters determine the answer. To this end, we derive new SE
increase the spectral efficiency (SE) by orders of magnitude expressions which are valid for both uplink (UL) and downlink
over contemporary systems [3]–[5]. (DL) transmission, with random user locations and power
The massive MIMO concept is based on equipping base control that yields uniform UE performance. We consider
stations (BSs) with hundreds or thousands of antenna elements both conventional linear processing schemes such as max-
which, unlike conventional cellular technology, are operated imum ratio (MR) combining/transmission and zero-forcing
in a coherent fashion. This can provide unprecedented array (ZF), and a new full-pilot zero-forcing (P-ZF) scheme that
gains and a spatial resolution that allows for multi-user MIMO actively suppresses inter-cell interference in a fully distributed
communication to tens or hundreds of user equipments (UEs) coordinated beamforming fashion. The following are the main
E. Björnson and E. G. Larsson are with Department of Elec-
contributions of each section:
trical Engineering (ISY), Linköping University, Linköping, Sweden • Section II presents the UL/DL massive MIMO system
({emil.bjornson,erik.larsson}@liu.se). M. Debbah is with CentraleSupélec,
Gif-sur-Yvette, France (merouane.debbah@supelec.fr) and with the Mathe-
model, where the unique features are the power control
matical and Algorithmic Sciences Lab, Huawei, Paris, France. and random UE locations.
This research has received funding from the EU FP7 under ICT-619086 • Section III provides new analytic results for channel
(MAMMOET), from ELLIIT, the Swedish Research Council (VR), and the
ERC Grant 305123 MORE.
estimation with arbitrary pilot signals and new tractable
Part of the material in this paper was presented at the IEEE Global Con-
ference on Signal and Information Processing (GlobalSIP), Atlanta, Georgia, 1 A few results for single-cell systems are available in the literature; for
December 3-5, 2014. example, in [8].
 2

SE expressions for the UL and DL with random UE Frequency

locations and power control. MR and ZF processing are
considered, as well as the new P-ZF scheme.
• Section IV provides extensive simulation results on the
maximal SE, where the impact of all system parameters
are explained. The expected massive MIMO gains are
illustrated. Time
• Section V extends the previous results to systems with
hardware impairments.
• Finally, Section VI summarizes the main results and
Frame structure
insights obtained in the paper.
UL pilots: UL data: DL data: Wc
II. S YSTEM M ODEL B symb ζ (ul)(S−B) symb ζ (dl)(S−B) symb

We consider a cellular network where payload data is Tc

transmitted with universal time and frequency reuse. Each cell Fig. 1: The transmission is divided into frames of S = Tc Wc
is assigned an index in the set L, where the cardinality |L| is symbols, whereof B symbols are dedicated to pilot transmis-
the number of cells. The BS in each cell is equipped with sion. The remaining S − B symbols are used for payload data,
an array of M antennas and communicates with K single- where ζ (ul) and ζ (dl) are respectively the fractions of UL and
antenna UEs at the time, out of a set of Kmax UEs. We are DL transmission.
interested in massive MIMO topologies where M and Kmax
are large and fixed, while K is a design parameter and all
UEs have unlimited demand for data. The subset of active reserved for UL pilot signaling. There is no DL pilot signaling
UEs changes over time, thus the name UE k ∈ {1, . . . , K} and no feedback of CSI, because the BSs can process both UL
in cell l ∈ L is given to different UEs at different times. The and DL signals using the UL channel measurements due to the
geographical position zlk ∈ R2 of UE k in cell l is therefore channel reciprocity in TDD systems. The remaining S − B
an ergodic random variable with a cell-specific distribution. symbols are allocated for payload data and are split between
This model is used to study the average performance for a UL and DL transmission. We let ζ (ul) and ζ (dl) denote the
random rather than fixed set of interfering UEs. The time- fixed fractions allocated for UL and DL, respectively. These
frequency resources are divided into frames consisting of Tc fractions can be selected arbitrarily, subject to the constraint
seconds and Wc Hz, as illustrated in Fig. 1.2 This leaves room ζ (ul) + ζ (dl) = 1 and that ζ (ul) (S − B) and ζ (dl) (S − B) are
for S = Tc Wc transmission symbols per frame. We assume positive integers. Below, we define the system models for the
that the frame dimensions are such that Tc is smaller or equal UL and DL.
to the coherence time of all UEs, while Wc is smaller or equal The BSs are not exchanging any short-term information
to the coherence bandwidth of all UEs. Hence, all the channels in this work, but we will see how the pilot allocation and
are static within the frame; hjlk ∈ CN denotes the channel transmission processing can be coordinated in a distributed
response between BS j and UE k in cell l in a given frame. fashion.
These channel responses are drawn as realizations from zero-
mean circularly symmetric complex Gaussian distributions: A. Uplink
 
hjlk ∼ CN 0, dj (zlk )IM , (1) The received UL signal yj ∈ CM at BS j in a frame is
modeled, similar to [7] and [8], as
where IM is the M × M identity matrix. This is a theoretical
K
model for non-line-of-sight propagation that is known to give XX √
yj = plk hjlk xlk + nj (2)
representative results with both few and many BS antennas
l∈L k=1
(see recent channel measurements reported in [21]). The
deterministic function dj (z) gives the variance of the channel where xlk ∈ C is the symbol transmitted by UE k in cell
attenuation from BS j to any UE position z. The value of l. This signal is normalized as E{|xlk |2 } = 1, while the
dj (zlk ) varies slowly over time and frequency, thus we assume corresponding UL transmit power is defined by plk ≥ 0. The
that the value is known at BS j for all l and k and that each additive noise nj ∈ CM is modeled as nj ∼ CN (0, σ 2 IM ),
UE knows its value to its serving BS. The exact UE positions where σ 2 is the noise variance.
zlk are unknown. Contrary to most previous works on massive MIMO, which
We consider the time-division duplex (TDD) protocol shown assume fixed UL power, we consider statistics-aware power
in Fig. 1, where B ≥ 1 out of the S symbols in each frame are control3 ; the symbols from UE k in cell l have the transmit
2 This paper concentrates on frames that carry user-specific signals, in 3 Channel-aware power control was considered in [22] and [9], but it
particular, payload data and pilots. From time to time, the network also needs requires a rapid feedback mechanism where UEs are provided with instan-
special frames to transmit cell-specific control and system information and taneous CSI. Since the small-scale fading average out in massive MIMO
to enable random access. The design of these control frames is outside the systems [2], statistical power control policies are expected to be almost equal
scope of this paper, but some initial results are found in [20]. to channel-aware policies [23], but are considerably easier to implement.
 3

power plk = dl (zρlk ) , where ρ > 0 is a design parameter.4 This interfering signals. This requires, however, some knowledge
√
power-control policy inverts the average channel attenuation of the UEs’ channels; for example, plk hjlk in the UL, for
dl (zlk ) and has the merit of making the average effective all l and k. Such CSI is typically acquired by pilot signaling,
channel gain the same for all UEs: E{plk khllk k2 } = M ρ. where the UEs send known signals in a predefined manner.
Hence, this policy guarantees a uniform user experience, saves Accurate CSI acquisition is a challenging task in multi-cell
valuable energy at UEs, and avoids near-far blockage where systems, where the transmission resources are reused across
weak signals drown in stronger signals due to the finite cells, because the pilot signals are inevitably affected by inter-
dynamic range of analog-to-digital converters (ADCs). cell interference. This so-called pilot contamination limits the
B. Downlink quality of the acquired CSI and the ability to reject inter-cell
interference (unless intricate subspace methods can be used
Building on the UL/DL channel reciprocity in calibrated for decontamination, as suggested in [17]).
TDD systems, the received DL signal zjk ∈ C at UE k in cell
The impact of pilot contamination is usually studied under
j in a frame is modeled as
the assumption that exactly the same pilot signals are used in
K
XX all cells. In contrast, this section derives the main properties
zjk = hTljk wlm slm + ηjk (3) of massive MIMO systems (with power control) for arbitrary
l∈L m=1 pilot reuse, where each cell might only use a subset of the
where (·)T denotes transpose, slm is the symbol intended for pilots. As shown in Fig. 1, the pilot signals are assumed to
UE m in cell l, wlm ∈ CM is the corresponding precoding span B symbols of each frame, where 1 ≤ B ≤ S.6 Each pilot
vector, and kwlm k2 is the allocated DL transmit power. Any signal can be represented by a deterministic vector v ∈ CB
power control can be considered in the DL since the BS has and the fixed per-symbol power implies that all entries have
access to the estimated CSI. We show later how to select the unit magnitude: |[v]s | = 1, where [·]s denotes the sth element
transmit power to achieve the same SEs in the DL as in the for s ∈ {1, . . . , B}. We assume that all pilot signals originate
UL. The additive noise at UE k in cell j is modeled as ηjk ∼ from a fixed pilot book V, defined as
CN (0, σ 2 ), with the same variance as in the UL.5 (
B, b1 = b2 ,
Remark 1 (Synchronization Issues). The UL/DL system mod- V = {v1 , . . . , vB } where vb1 vb2 =
H
(4)
els in (2) and (3) assume perfect synchronization across all 0, b1 6= b2 ,
cells, as commonly done in the massive MIMO literature;
cf. [2], [6]–[8], [15]. Local synchronization is achievable, where (·)H denotes the conjugate transpose. Hence, the B pilot
for example, using the cyclic prefix in OFDM-based systems, signals form an orthogonal basis and can, for example, be the
but network-wide synchronization is probably infeasible over columns of a discrete Fourier transform (DFT) matrix [24].
large coverage areas. The processing techniques analyzed The pilot signal transmitted by UE k in cell l is denoted by
in this paper can thus be used to suppress the strong in- vilk , where ilk ∈ {1, . . . , B} is the index in the pilot book.
terference between the closest tiers of neighboring cells, By transmitting these pilot signals over B symbols in the UL
while the interference from distant cells is asynchronously system model of (5), the collective received UL signal at BS
received and practically insuppressible. We expect that the j is denoted as Yj ∈ CM ×B and given by
simplified synchronization modeling used here and elsewhere K
has negligible impact on the system performance, since the XX √
Yj = plk hjlk viTlk + Nj , (5)
insuppressible distant interferers are weak as compared to l∈L k=1
(partially suppressed) interference from neighboring cells.
where Nj ∈ CM ×B contains the additive noise at the receiver
III. AVERAGE P ER -C ELL S PECTRAL E FFICIENCY during the pilot signaling.
In this section, we derive and analyze the SE for multi-cell The following lemma derives the minimum mean-squared
systems with random UE positions. error (MMSE) estimator of the effective power-controlled UL
√
channels, which are defined as heff
jlk = plk hjlk .
A. Pilot-Based Channel Estimation Lemma 1. The MMSE estimate at BS j of the effective power-
BS j can use its multitude of antennas for coherent receive controlled UL channel heff
jlk , for any UE k ∈ {1, . . . , K} in
combining in the UL and transmit precoding in the DL, any cell l ∈ L, is
which can adaptively amplify desired signals and suppress
dj (zlk )
4 The parameter ρ needs to be selected such that UEs at the cell edge do
ĥeff
jlk = Yj (ΨTj )−1 vi∗lk (6)
dl (zlk )
not use more transmit power than their amplifiers can handle or the spectrum
regulations allow. This is not a critical limitation in massive MIMO since high
SEs are provided also at low SNRs (see Fig. 12), but it might be necessary 6 The pilot signals need not be synchronized across the cells as assumed
to occasionally drop severely shadowed UEs from service. herein, but there is little to gain from shifting the pilot signals and UL payload
5 The noise variance is conventionally lower in the UL, due to better data signals between cells; this leads to a mix of deterministic pilots and
hardware characteristics at the BS, but since massive MIMO has an inherent stochastic data signals at each symbol transmission, but the average pilot
robustness to noise amplification [13] it is possible to use handset-like contamination will not change in any substantial way [8, Remark 5]. The
hardware at the BSs. In any case, any disparity in noise power between the UL new full-pilot interference suppression concepts proposed in this paper are
and DL can be absorbed into the transmit powers without loss of generality. also harder to implement in such cases.
 4

(ul) pjk |E{h} {gjk

H
hjjk }|2
SINRjk = K
. (12)
P P
plm E{h} {|gjk
H
hjlm |2 } − pjk |E{h} {gjk
H
hjjk }|2 + σ 2 E{h} {kgjk k2 }
l∈Lm=1

where (·)∗ denotes the complex conjugate and the normalized Remark 2 (Mobility and Pilot Sharing). Each UE might have
covariance matrix Ψj ∈ CB×B of the received signal is a different dimension of its coherence block, defined by some
K coherence time T̃c and coherence bandwidth W̃c , depending on
XX dj (z`m ) σ2 the propagation environment and the UE’s mobility. Suppose
Ψj = vi`m viH`m + IB . (7)
d (z )
m=1 ` `m
ρ that T̃c = aTc and W̃c = bWc for a certain UE, where a ≥ 1
`∈L
and b ≥ 1 since the frame structure was defined to fit into
The estimation error covariance matrix Cjlk ∈ CM ×M is the coherence block of all UEs. Then, τ = bacbbc is the total
given by number of frames that fits into the coherence block of this
n o
Cjlk = E (heff eff eff eff H particular UE, where b·c stands for truncation. If τ > 1, there
jlk − ĥjlk )(hjlk − ĥjlk )
  is no need to send pilots in every frame; it is sufficient with
dj (zlk )
dj (zlk )  dl (zlk ) B 1/τ of the frames. Hence, multiple UEs with τ > 1 can share
=ρ 1− P PK dj (z`m ) H  IM a pilot signal, without disturbing one another, by using it in
dl (zlk ) σ2
`∈L m=1 d` (z`m ) vilk vi`m + ρ different frames.
(8)
and the mean-squared error (MSE) is MSEjlk = tr(Cjlk ).
B. Achievable UL Spectral Efficiencies
Proof: The proof is given in the appendix. The channel estimates in Lemma 1 enable each BS to (semi-
There are two important differences between Lemma 1 )coherently detect the data signals from its UEs. In particular,
and the channel estimators that are conventionally used in we assume that BS j applies a linear receive combining vector
the massive MIMO literature: 1) we estimate the effective gjk ∈ CM to the received signal, as gjk H
yj , to amplify the
channels including the UL power control; and 2) the MMSE signal from its kth UE and reject interference from other UEs
estimator supports arbitrary pilot allocation. in the spatial domain. We want to derive the ergodic achievable
The covariance matrix in (8) reveals the causes of estimation SE for any UE, where codewords span over both the Rayleigh
errors; it depends on the inverse signal-to-noise ratio (SNR), fading and random locations of the interfering UEs—specific
σ 2 /ρ, and on which UEs that use the same pilot signal UE distributions are considered in Section IV. For notational
(i.e., which of the products viHlk vi`m that are non-zero). The convenience, we assume that β = K B
is an integer that we refer
ratio dj (z`m )/d` (z`m ) describes the relative strength of the to as the pilot reuse factor. The cells in L are divided into
interference received at BS j from UE m in cell `; it is almost β ≥ 1 disjoint subsets such that the same K pilot sequences
one for cell-edge UEs of neighboring cells, while it is almost are used within a set, while different pilots are used in different
zero when cell ` is very distant from BS j. sets. We refer to this as non-universal pilot reuse. An explicit
Although Lemma 1 allows for estimation of all channel
example is provided in Section IV for hexagonal cells, while
vectors in the whole cellular network, each BS can only
the result in this section holds for any network topology. The
resolve B different spatial dimensions since there are only
following lemma shows how the SE depends on the receive
B orthogonal pilot signals. To show this explicitly, we define
combining, for Gaussian codebooks where xjk ∼ CN (0, 1).
the M × B matrix
b V,j = Yj (ΨT )−1 [v∗ . . . , v∗ ] Lemma 2. In the UL, an ergodic achievable SE of an arbitrary
H j B (9)
1 UE k in cell j is
using each of the B pilot signals from V. The channel estimate 
B
 n o
(ul) (ul)
in (6) for UE k in cell l, which uses the pilot vilk , is parallel ζ 1− E{z} log2 (1 + SINRjk ) [bit/s/Hz]
S
to the ilk th column of H
b V,j ; more precisely, we have
(11)
dj (zlk ) b where the effective signal-to-interference-and-noise ratio
ĥeff
jlk = HV,j eilk (10) (ul)
(SINR), SINRjk , is given in (12) at the top of the page.
dl (zlk )
The expectations E{z} {·} and E{h} {·} are with respect to UE
where ei denotes the ith column of the identity matrix IB .
positions and channel realizations, respectively.
This is the essence of pilot contamination; BSs cannot tell
apart UEs that use the same pilot signal and cannot reject the Proof: By coding over variations in the channel re-
corresponding interference since the estimated channels are alizations {h} and positions {z} of the interfering UEs,
parallel. In some cases (e.g., for slow changes in the user an achievable SE is given by E{z,h} {I(xlk , yj )}, where
scheduling and high spatial channel correlation), statistical I(xlk , yj ) is the mutual information between the transmitted
prior knowledge can be used to partially separate the UEs [18], and received signal in (2) for fixed channel realizations and
but this possibility is not considered herein since we want to UE positions. The lemma follows from computing a lower
develop methods to suppress pilot contamination that can be bound on I(xlk , yj ), similar to [6], [7], [13], [25], [26], by
utilized in any propagation environment. making three limiting assumptions: 1) a Gaussian codebook
 5

MR
is used; 2) the signal component received over the effective Zjl = K, while ZF combining is obtained by GZF = M −K
channel mean E{h} {gjk H
hjjk } is the only desired signal, while and
the interference and the signal component over the remaining   
uncorrelated channel gjkH
hjjk − E{h} {gjkH
hjjk } are treated as 
K 1 − P jl µ
(1)
 if l ∈ Lj (β),

noise (i.e., not exploited in the decoding); and 3) the noise ZF
Zjl =
(1) σ2
µj` + Bρ (17)
is taken as worst-case Gaussian distributed in the decoding, 
 `∈Lj (β)

K if l 6∈ Lj (β).

leading to a further lower bound on the mutual information.

The ergodic achievable SE in Lemma 2, for any UE in The following notation was used:
cell j, is a lower bound on the ergodic capacity, which is  ω 
unknown for general multi-cell networks. Similar bounds are (ω) dj (zlm )
µjl = Ezlm for ω = 1, 2. (18)
found in [6]–[8] and the bounding technique interacts with dl (zlm )
the Rayleigh fading, which is why its expectations end up
inside the logarithm while the user positions are averaged at Proof: The proof is given in the appendix.
the outside. To compute these expectations we need to specify The closed-form SE expressions in Theorem 1 are lower
the receive combining. The combining schemes for massive bounds on the ergodic capacity and slightly more conservative
MIMO can have either passive or active interference rejection. than the non-closed-form bound in Lemma 2; see Section
The canonical example of passive rejection is maximum ratio IV-B for a numerical comparison. We stress that the closed-
(MR) combining, defined as form SEs are only functions of the pilot allocation and the
(1) (2)
MR b V,j ei = ĥeff , propagation parameters µjl and µjl defined in (18). The
gjk =H jjk (13)
jk
latter two are the average ratio between the channel variance to
which maximizes the gain of the desired signal and relies on BS j and the channel variance to BS l, for an arbitrary UE in
that interfering signals are rejected automatically since the co- cell l, and the second-order moment of this ratio, respectively.
user channels are quasi-orthogonal to ĥeff
jjk when M is large.
7 These parameters are equal to 1 for j = l and otherwise go to
In contrast, active rejection is achieved by making the zero as the distance between BS j and cell l increases. The SE
receive combining as orthogonal to the interfering channels as expression manifests the importance of pilot allocation, since
possible. This is conventionally achieved by zero-forcing (ZF) the interference term in (16) contains summations that only
combining, where the combining is selected to orthogonalize consider the cells that use the same pilots as cell j.
the K intra-cell channels: The first term in (16) describes the pilot contamination,
ZF b V,j Ej EH HbH H

−1 while the second term is the inter-user interference. The
gjk =H j V,j V,j Ej
b eijk , (14)
difference between MR and ZF is that the latter scheme
scheme
where Ej = [eij1 . . . eijK ] ∈ CB×K and all the UEs in cell j cancels some interference through Zjl , at the price of
scheme
are required to use different pilots. reducing the array gain G from M to M − K.
The next theorem provides closed-form expressions for the ZF combining only actively suppresses intra-cell interfer-
per-cell SEs with MR and ZF. ence, while the inter-cell interference is passively suppressed
just as in MR combining. Further interference rejection can
Theorem 1. Let Lj (β) ⊂ L be the subset of cells that uses
be achieved by coordinating the combining across cells, such
the same pilots as cell j. In the UL, an achievable SE in cell
that both intra-cell and inter-cell interference are actively
j is
! suppressed by the receive combining. We propose a new full-
 
(ul) (ul) B 1 pilot zero-forcing (P-ZF) combining, defined as
SEj = Kζ 1− log2 1 + scheme [bit/s/Hz/cell]
S Ij
−1
P-ZF bH H
(15) gjk =H
b V,j H
V,j V,j
b eijk . (19)
where the interference term
  2  In contrast to the conventional ZF in (14), which only orthog-
(2) (1)
X  (2) µjl − µjl onalize the K intra-cell channels in H b V,j Ej , P-ZF exploits
Ijscheme = µjl +

G scheme  that all the B estimated channel directions in H b V,j are known
l∈Lj (β)\{j} at BS j and orthogonalizes all these directions to also mitigate
  ! parts of the inter-cell interference; a similar downlink concept
(1) scheme 2 (1) 2
σ σ
was proposed in [15]. The cost is a loss in array gain of B,
P P
µjl Zjl + ρ µj` + Bρ
l∈L `∈Lj (β) instead of K as with conventional ZF. There is no signaling
+ (16)
Gscheme between BSs in this coordinated multipoint (CoMP) scheme—
depends on the receive combining scheme through Gscheme BS j estimates H b V,j from the UL pilot signaling—and thus
scheme
and Zjl . MR combining is obtained by GMR = M and the P-ZF scheme is fully distributed and scalable. Achievable
SEs with P-ZF are given by the following theorem.
7 With quasi-orthogonality we mean that two vectors a, b ∈ CM satisfy
aH b
→ 0 as M → ∞, although aH b will
Theorem 2. Let Ll (β) ⊂ L be the subset of cells that uses
M √ not converge to zero and the same pilots as cell l. In the UL, an achievable SE in cell
might even go to infinity, e.g., proportionally to M as with Rayleigh fading
channel vectors. j with P-ZF combining is given by (15) for GP-ZF = M − B
 6

K K
and   policy {qjk }, with
P P
qjk =
P P
pjk , for which
(1) j∈Lk=1 j∈Lk=1
P-ZF
 µjl 
Zjl 1 −
=K . (20) (dl) (ul)
P (1) σ2  SINRjk = SINRjk (24)
µj` + Bρ
`∈Ll (β) scheme
by using ǧjk = gjk for all j and k. Consequently, an
Proof: The proof is given in the appendix. achievable SE in the DL of cell j is
The SE expressions were derived assuming that M and K   !
(dl) (dl) B 1
are the same in all cells, for notational brevity. However, the SEj = Kζ 1− log2 1 + scheme [bit/s/Hz/cell]
S Ij
results in this section are straightforward to extend to cell-
(25)
specific M and K values.
where the interference term Ijscheme is the same as in the UL
(for MR, ZF, or P-ZF).
C. Achievable DL Spectral Efficiencies Proof: The proof is given in the appendix.
The channel estimates from Lemma 1 are also used for This theorem shows that the SINRs that are achieved in
linear precoding in the DL, where the M channel inputs are the UL are also achievable in the DL, by selecting the
utilized to make each data signal add up (semi-)coherently at power control coefficients {qjk } properly. The total transmit
its desired UE and to suppress the interference caused to other power is the same, but is allocated differently over the UEs.
UEs. Recall from (3) that wjk ∈ CM is the precoding vector This is a consequence of the uplink-downlink duality [28],
associated with UE k in cell j. We express these precoding conventionally considered for single-cell systems with perfect
vectors as CSI, which is applicable also in our general multi-cell massive
qjk
r
wjk = ǧ∗ (21) MIMO setup with estimated CSI. The exact expression for
E{h} {kǧjk k2 } jk the power control coefficients is only given in the proof, since
the main purpose of Theorem 3 is the fact that equal UL/DL
where the average transmit power qjk ≥ 0 is a function of the
performance is possible, which allows for joint analysis in
UE positions, but not the instantaneous channel realizations.
what follows.
The vector ǧjk ∈ CM defines the spatial directivity of the
Motivated by Theorem 3, this paper considers three types
transmission and is based on the acquired CSI; the normaliza-
of linear precoding vectors: MR precoding which amplifies
tion with the average squared norm E{h} {kǧjk k2 } gives the MR
the desired signal by setting ǧjk = gjk ; ZF precoding that
analytic tractability that enables the following results.8 ZF
actively rejects intra-cell interference by setting ǧjk = gjk ;
Lemma 3. In the DL, an ergodic achievable SE of an arbitrary and P-ZF precoding that actively rejects both intra- and inter-
P-ZF
UE k in cell j is cell interference by setting ǧjk = gjk . We stress that P-
  ZF precoding is a fully distributed coordinated beamforming
B n
(dl)
o
ζ (dl) 1 − E{z} log2 (1 + SINRjk ) [bit/s/Hz] scheme tailored to massive MIMO systems, since each BS
S only uses locally estimated CSI.
(22)
(dl)
with the effective SINR, SINRjk , given by
D. Finite and Asymptotic Analysis
H
|E {ǧjk hjjk }|2
qjk E{h} 2
{h} {kǧjk k }
Based on Theorems 1–3, the sum of the per-cell achievable
K
. SEs in the UL and DL are given by the following corollary.
H h
|E{h} {ǧjk 2
jjk }|
H h
E{h} {|ǧlm 2
ljk | }
P P
qlm E{h} {kǧlm k2 } − qjk E{h} {kǧjk k2 } + σ2 Corollary 1. Looking jointly at the UL and DL, an achievable
l∈Lm=1
(23) SE in cell j is
(ul) (dl)
Proof: This follows from the same procedures as the SEj = SEj + SEj
proof of Lemma 2.   !
(26)
B 1
Note that Lemma 3 takes into account the fact that each UE =K 1− log2 1+ [bit/s/Hz/cell]
only knows the expectations in (23) and not the instantaneous
S Ijscheme
channels (see [6, Theorem 1] for more details). where the interference term Ijscheme for UE k is given Theorem
The precoding can be designed in a variety of ways. The 1 for MR and ZF and in Theorem 2 for P-ZF. This SE can be
next theorem shows that there is a strong connection between divided between the UL and DL arbitrarily using any positive
transmit precoding in the DL and receive combining in the fractions ζ (ul) and ζ (dl) , with ζ (ul) + ζ (dl) = 1.
UL.
This is a convenient result that allows us to analyze and
scheme
Theorem 3. Let {gjk } be the set of receive combining optimize the SE of the network as a whole, without having
vectors used in the UL. Then, there exist a DL power control to separate the UL and DL. Since it is hard to gain further
8 Conventionally, the power is normalized by kǧ k2 instead of
insights from the structure of the SE expression in (26), we
jk analyze it for a particular network topology in Section IV. In
E{h} {kǧjk k2 } in multi-user MIMO systems [27], but the difference is small
in massive MIMO since |E{h} {kǧjk k2 } − kǧjk k2 |/M → 0 as M → ∞, the remainder of this section, we consider the limit of a large
for most precoding schemes. number of antennas.
 7

Corollary 2. Let Lj (β) ⊂ L be the subset of cells that uses the

same pilots as cell j. When M → ∞ (with K, B ≤ S < ∞),
the effective SINRs with MR, ZF, and P-ZF converge to the
same limit:
1 1 1 1
MR
, ZF , P-ZF → (2)
. (27)
Ij Ij Ij
P
µ jl
l∈Lj (β)\{j}

The ultimate effect of pilot contamination is very clear in β=1 β=3

(27), since only the cells that interfered with cell j during pilot
transmission (i.e., cells with indices in the set Lj (β) \ {j})
affect the asymptotic limit. To maximize the asymptotic SINR
(2)
in (27), one should place the cells with large µjl in different
subsets (i.e., Lj (β) ∩ Ll (β) = ∅) so that these cells use
different pilots. The asymptotic limit can be used as follows
to find the optimal K.
Corollary 3. Let Lj (β) ⊂ L be the subset of cells that uses β=4 β=7
the same pilots as cell j. The SE in cell j approaches
    Fig. 2: Part of a hexagonal network, colored for different pilot
∞ Kβ 1
SEj = K 1 − log2 1 + P (28) reuse factors β.
S (2)
µ
l∈Lj (β)\{j} jl

when M → ∞. This SE is maximized jointly for jall cells k

S
when the number of scheduled UEs is either K ? = 2β or
l m
S S Axis 2
K ? = 2β (i.e., one of the closest integers to 2β ).
(2)
αj
Proof: The logarithmic part of (28)
 is independent of K,

r
while the concave pre-log factor K 1 − Kβ is maximized
S Axis 1
S
by K = 2β . The concavity implies that the optimal integer αj
(1)
∗ S
K is one of the closest integers to 2β .
Corollary 3 is a main contribution of this paper and proves Fig. 3: The coordinate system for a hexagonal grid.
that the number of scheduled UEs should be proportional to At first sight, these results bear some similarity with the
the frame length S (when M is large enough); for example, we results in [26] and [29] for block-fading noncoherent point-
get K ? = S2 for β = 1 and K ? = S6 for β = 3. Since both S = to-point (P2P) MIMO channels, where the maximal degrees
200 and S = 10000 are reasonable coherence block lengths of freedom (DoF) are S4 and are achieved by having S2
in practice, depending on the UE mobility and propagation transmit/receive antennas and using pilot signals of the same
environment, this means that we should schedule between tens length. The fundamental difference is that the DoF concept
and several thousands of UEs for simultaneous transmission in for P2P MIMO channels, where unbounded SE is achieved
order to be optimal. This is only possible if the UE selection at high SNRs, does not apply to cellular networks [30].
S
policy is scalable and there is a high load of UEs. If K ? = 2β S
Instead, the pre-log factor 4β in (29) may be interpreted as the
is an integer, the asymptotically optimal SE is relative improvement in SE that can be achieved by aggressive
S

1
 scheduling of UEs in massive MIMO systems.
SE∞j = log 2 1 + (2)
(29) We have now established the asymptotically optimal number
4β
P
l∈Lj (β)\{j} µjl of scheduled UEs, as M → ∞. Next, we investigate the
and increases linearly with the frame length S (in the large-M impact on practical systems with finite M for a certain network
regime). topology.
Interestingly, the asymptotically optimal scheduling gives
B = S2 for any β, which means that half the frame is IV. O PTIMIZING N UMBER OF UE S IN H EXAGONAL
allocated to pilot transmission. This extraordinary fact was N ETWORKS
initially conjectured in [2] for β = 1. The rationale is that the The concept of cellular communications has been around for
SE gain from adding an extra UE outweighs the pre-log loss decades [31]. Although practical deployments have irregular
at the existing UEs if at least half the frame is used for data cells, it is common practice to establish general properties
(a criterion independent of β). The asymptotically optimal β by analyzing symmetric networks where the cells are regular
cannot be computed in closed-form, but we notice that a larger polygons [32]; in particular, hexagons.
β leads to fewer interferers in Lj (β) and also reduces the pre- In this section, we consider the symmetric network topology
log factor; hence, a larger β brings SINR improvements until depicted in Fig. 2 with hexagonal cells. All the time/frequency
a certain point where the pre-log loss starts to dominate. resources allocated to payload data transmission are used in all
 8

400
the cells. However, inspired by [15], we consider pilot books

Spectral Efficiency (SE) [bit/s/Hz/cell]

of size B = βK to allow for non-universal pilot reuse that 350
Asymptotic limit
mitigates the pilot contamination from neighboring cells. 300 P−ZF
The hexagonal grid is infinitely large, to avoid edge effects ZF
250 MR
and to give all cells the same properties. The cell radius is
200
denoted by r > 0 and is the distance from the cell center to
the corners. Each cell can be uniquely indexed by a pair of 150 Switching
(1) (2) points
integers αj , αj ∈ Z, where Z is the set of integers. This 100
integer pair specifies the location of BS j [31]: 50 Switching
points
√
√   
3r/2 (1) 0 (2) 0
bj = 3 αj + √ α ∈ R2 . (30) 10
1 2
10
3
10 10
4 5
10
r/2 3r j Number of BS Antennas (M)
(1) (2) (a) Optimized SE per cell.
The coordinate system imposed by αj and αj is illustrated
in Fig. 3. Every cell on the hexagonal grid has 6 interfering
200

Optimal Number of Scheduled UEs (K)

cells in the first surrounding tier, 12 in the second tier, etc. As P−ZF
shown in the early works on hexagonal networks [31], [32], ZF
MR
this limits which pilot reuse factors that give symmetric reuse 150
patterns: β ∈ {1, 3, 4, 7, 9, 12, 13, . . .}. Reference
curve
Our simulations consider a classic pathloss model where the M =K Pilot reuse:
C 100 β=1
variance of the channel attenuation in (1) is dj (z) = kz−b jk
κ,

where k · k is the Euclidean norm, C > 0 is a reference value,

and κ ≥ 2 is the pathloss exponent. These assumptions allow 50
(ω)
us to compute µjl in (18) as Pilot reuse: β = 3
ω  κω  0
kzlm −bl k
 
(ω) dj (zlm ) 1 2 3 4 5
µjl = Ezlm = Ezlm 10 10 10 10 10
dl (zlm ) kzlm −bj k Number of BS Antennas (M )
(31)
(b) Corresponding optimal number of UEs: K ? .
for any UE distributions in the cells. We note that C and r
cancel out in (31), if the UE distributions in each cell are Fig. 4: Simulation of optimized SE, as a function of M , with
independent of C and r. Since the power control makes the average inter-cell interference.
SEs in Theorems 1–3 independent of the UEs’ positions, we
only need to define the parameter ratio ρ/σ 2 ; that is, the
average SNR (over fading) between any UE and any antenna bandwidth), set the SNR to ρ/σ 2 = 5 dB, and pick κ = 3.7
at its serving BS. as pathloss exponent.9 The impact of changing the different
system parameters is considered in Section IV-B.
A. Optimizing SE for Different Interference Levels We consider three propagation environments with different
severity of inter-cell interference:
We simulate the SE in an arbitrary cell on the hexagonal grid
1) Average case: Averaging over uniform UE locations in
in Fig. 2 and take all non-negligible interference into account.
all cells.
The UEs can be anywhere in the cells, but at least 0.14r
2) Best case: All UEs in other cells are at the cell edge
from the serving BS (this makes the analysis independent of
furthest from BS j (for each j).
r). Since the SE expressions in Section III are the same for
3) Worst case: All UEs in other cells are at the cell edge
the UL and DL, except for the fractions ζ (ul) and ζ (dl) , we
closest to BS j (for each j).
simulate the sum of these SEs and note that it can be divided (1) (2)
arbitrarily between the UL and DL. The same linear processing The corresponding values of the parameters µjl and µjl were
6
schemes are used in both directions. The simulations consider computed by Monte-Carlo simulations with 10 UE locations
MR, ZF, and P-ZF precoding/combining, and all results are in each cell.
obtained by computing the closed-form expressions from The best case is overly optimistic since the desirable UE
Section III for different parameter combinations. The simu- positions in the interfering cells are different with respect to
lations were performed using Matlab and the code is available different cells. However, it gives an upper bound on what is
for download at https://github.com/emilbjornson/maximal-SE, achievable by coordinated scheduling across cells. The worst
which enables reproducibility as well as simple testing of other case is overly pessimistic since the UEs cannot all be at the
parameter values. worst locations, with respect to all other cells, at the same time.
For each number of antennas, M , we optimize the SE with The average case is probably the most applicable in practice,
respect to the number of UEs K and the pilot reuse factor where the averaging comes from UE mobility, scheduling,
β (which determine B = βK) by searching the range of all and random switching of pilot sequences between the UEs
reasonable integer values. We set the coherence block length 9 A higher pathloss exponent reduces the inter-cell interference, but requires
to S = 400 (e.g., 2 ms coherence time and 200 kHz coherence more signal power to maintain a certain SNR.
 9

200
2500

Spectral Efficiency (SE) [bit/s/Hz/cell]

Spectral Efficiency (SE) [bit/s/Hz/cell]

Asymptotic limit Asymptotic limit

P−ZF P−ZF
2000 ZF 150 ZF Switching
MR MR points
1500
100

1000
Switching
50 points
500

0 0
1 2 3 4 5 1 2 3 4 5
10 10 10 10 10 10 10 10 10 10
Number of BS Antennas (M) Number of BS Antennas (M )

(a) Optimized SE per cell. (a) Optimized SE per cell.

200 60
Optimal Number of Scheduled UEs (K)

Optimal Number of Scheduled UEs (K)

P−ZF Reference Pilot reuse:
ZF curve β=3
50 M = K
MR
150
Reference 40
curve Pilot reuse:
M =K β=4 Pilot reuse:
100 30 β=4
Pilot reuse:
β=1
20
50
P−ZF
10 ZF
Pilot reuse: β = 7 MR
0 0
1 2 3 4 5 1 2 3 4 5
10 10 10 10 10 10 10 10 10 10
Number of BS Antennas (M) Number of BS Antennas (M )

(b) Corresponding optimal number of UEs: K ? . (b) Corresponding optimal number of UEs: K ? .

Fig. 5: Simulation of optimized SE, as a function of M , with Fig. 6: Simulation of optimized SE, as a function of M , with
best-case inter-cell interference. worst-case inter-cell interference.

in each cell. Results for the average case are shown in Fig. 4, they are achieved; that is, which number of UEs K ? and which
the best case in Fig. 5, and the worst case in Fig. 6. The pilot reuse factor β that are used. The general behavior is that
optimized SE and the corresponding K ? are shown in (a) and larger M implies a higher K ? and a smaller β, because the
(b), respectively. channels become more orthogonal with M . Since the reuse
factor is an integer, K ? changes non-continuously when β
The achievable SEs (per cell) are very different between the
is changed; smaller β allows for larger K ? , and vice versa.
best case interference and the two other cases—this confirms
MR schedules the largest number of UEs and switches to a
the fact that results from single-cell analysis of massive MIMO
smaller reuse factor at fewer antennas than the other schemes.
is often not applicable to multi-cell cases (and vice versa). ZF
In contrast, P-ZF schedules the smallest number of UEs and
brings much higher SEs than MR under the best case inter-
has the highest preference of large reuse factors, since this it
cell interference, since then the potential gain from mitigating
can suppress more inter-cell interference in these cases. Simply
intra-cell interference is very high. P-ZF is equivalent to ZF in
speaking, MR gives low per-user SEs to many UEs (sometimes
the best case, but excels under worst case inter-cell interference
more than M ), while ZF and P-ZF give higher per-user SEs
since it can actively suppress also inter-cell interference. In
to fewer UEs.
the realistic average case, the optimized SEs are rather similar S
Recall from Corollary 3 that K = 2β becomes the optimal
for MR, ZF, and P-ZF; particularly in the practical range of
number of UEs as M → ∞. This property is confirmed by
10 ≤ M ≤ 200 antennas. In all cases, the largest differences
Figs. 4–6, since K ? → 67 in the average case (where β = 3),
appear when the number of antennas is very large (notice the
K ? → 200 in the best case (where β = 1), and K ? → 50 in
logarithmic M -scales). At least M = 105 is needed to come
the worst case (where β = 4).
close to the asymptotic limit in (29), which was proved by
Corollary 3, and many more antennas are required under best
case interference. Clearly, the asymptotic limits should not B. Impact of System Parameters
be used as performance indicators since unrealistically many We now focus on the average case of inter-cell interference,
antennas are needed for convergence. due to its practical relevance, and investigate how each system
As seen from Figs. 4–6, the main difference between MR, parameter affects the simulation results. We focus on the range
ZF, and P-ZF is not the values of the optimized SE but how 10 ≤ M ≤ 1000 antennas, and when other system parameters
 10

80 180
P−ZF P−ZF
Spectral Efficiency (SE) [bit/s/Hz/cell]

Spectral Efficiency (SE) [bit/s/Hz/cell]

70 ZF 160 ZF
MR 140 MR
60
120 Pilot reuse:
50 β=3
100
40
80
30 Pilot reuse:
60
β=1
20 40
10 20

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Number of BS Antennas (M) Number of BS Antennas (M )

Fig. 7: Per-cell SE for K = 10. The lines are based on Fig. 8: Impact of changing the pilot reuse factor β, for a system
Theorems 1 and 2 while the markers are computed numerically optimized for high per-cell SE.
from Lemma 2.
3

Per−UE Spectral Efficiency [bit/s/Hz/user]

P−ZF
2.5 ZF
than M are varied we only consider M = 100 (medium MR
massive MIMO setup) and M = 500 (large massive MIMO 2
setup).
Pilot reuse:
We begin by verifying the accuracy of the closed-form 1.5 β=3
expressions in Theorems 1 and 2, by comparing the formulas
1
to Monte-Carlo simulations based on Lemma 2. The formulas
are exact in the best and worst interference cases, but Fig. 7 0.5 Pilot reuse:
shows that the interference variations in the average case result β=1
in some loss in SE. The figure considers K = 10 UEs and 0
0 200 400 600 800 1000
Monte-Carlo simulations are represented by markers. The MR Number of BS Antennas (M )
and ZF formulas in Theorem 1 are very tight. However, there Fig. 9: Achievable SE per UE, for a system optimized for high
is a few percent of deviation for P-ZF in Theorem 2, since per-cell SE.
a lower bound on the ability of cancel inter-cell interference
is used to get a tractable formula. Hence, P-ZF will actually
perform slightly better than reported in the simulations in this thumb which says that massive MIMO systems should have an
paper. order of magnitude more BS antennas than UEs. The operating
Next, we study the impact of the pilot reuse factor β points that satisfy this guideline are above the horizontal dotted
using the formulas from Theorems 1 and 2. Fig. 8 shows line. This simulation indicates that an optimized system might
the per-cell SE for β = 1 and β = 3, which provide the not follow this guideline; in fact, there is a few occasions
highest SEs for M ≤ 1000. The curves are smooth and there where MR even prefers to have M/K ? < 1. Generally
are wide regions around the β-switching points where both speaking, it seems that having 2–8 times more BS antennas
values provide almost equal SEs. This robustness simplifies than UEs is the range to aim at for practical deployments.
cell planning and scheduling based on user load. Since the cells might not be fully loaded at every time
Changes in the pilot reuse factor have major impact on instant, Fig. 11 shows the per-cell SE as a function of the
the optimal number of UEs and their achievable performance. number of scheduled UEs. As noted before, the peak numbers
The SE per UE is shown in Fig. 9 for the operating points (which are star marked) are at different K for each scheme. If
that maximize the SE in the cell; this is basically the ratio MR, ZF, and P-ZF are compared for a given K, the differences
SE/K ? , where SE was given in Fig. 4(a) and K ? was given between the schemes can either be larger or smaller than at the
in Fig. 4(b). We notice that MR gives the lowest SE per peak numbers. Although ZF and P-ZF often provide better SE
scheduled UE, while P-ZF gives the highest SE per scheduled than MR, it is interesting to note that MR is competitive when
UE. The numbers are around 1 bit/s/Hz for MR, in the range K is large—both in terms of SE and since its computational
1–2.5 bit/s/Hz for ZF, and in the range 1–3 bit/s/Hz for P-ZF. complexity scales as O(M K), while the complexity of ZF
Since the pilot signaling consumes between 2 and 40 percent and P-ZF scales as O(M K 2 ) [9].
of the frame in this simulation, the payload data need to be Next, Fig. 12 investigates how the average SNR ρ/σ 2 affects
encoded with up to 4.5 bit/symbol, which can be achieved the results. The SE saturates already at an SNR of 5 dB due
by conventional 64-QAM with a 3/4 coding rate. Hence, all to the array gain from coherent processing—this is why 5
the per-user SEs in Fig. 9 are straightforward to implement in dB was used in the previous figures. Massive MIMO can
practice. operate also at lower SNRs, but with a performance loss.
Fig. 10 shows the ratio M/K ? for the same scenario as ZF and P-ZF are particularly sensitive to the SNR level,
in the previous figures. This ratio can be interpreted as the since the active interference suppression requires a higher CSI
number of BS antennas per UE [7]. There is a common rule of estimation quality than simple MR processing.
 11

14 300
P−ZF P−ZF

Spectral Efficiency (SE) [bit/s/Hz/cell]

12 ZF ZF
250
BS Antennas per UE (M/K)

MR MR
10x more antennas than UEs
10
200
8
150
6 Pilot reuse:
β=3 100 M = 500
4
Pilot reuse:
2
β=1 50
M = 100
0 0
0 200 400 600 800 1000 0 500 1000 1500 2000
Number of BS Antennas (M) Coherence Block Length (S)

Fig. 10: Number of BS antennas per UE with different Fig. 13: Per-cell SE as a function of the coherence block length
processing schemes, for a system optimized for high per-cell S.
SE.

140
V. S PECTRAL E FFICIENCIES WITH H ARDWARE
I MPAIRMENTS
Spectral Efficiency (SE) [bit/s/Hz/cell]

120
The analytic and numeric analysis in the previous sections
100 have focused on cellular networks where the BSs and UEs are
M = 500 equipped with ideal transceiver hardware, which can radiate
80
P−ZF any waveform without distortions and which can receive
60 ZF any waveform with infinite resolution. However, practical
MR
M = 100 transceivers inevitably operate with non-linearities in ampli-
40
fiers, clock drifts in local oscillators, finite-precision ADCs,
20 I/Q imbalance in mixers, and non-ideal analog filters [33]–
0
[36]. In this section, we provide a prediction of how these
0 50 100 150 200 hardware impairments affect the achievable SEs in multi-cell
Number of UEs (K)
massive MIMO systems. We notice that it was recently shown
Fig. 11: Achievable per-cell SE as a function of the number in [12], using impairments models developed and evaluated
of scheduled UEs. in [33]–[35], that the hardware impairments caused by the
BS array are negligible in massive MIMO systems, since the
150
desired signals are amplified by the array gain from coherent
Spectral Efficiency (SE) [bit/s/Hz/cell]

processing while the distortions add non-coherently. Hence,

the hardware impairments in the UE hardware are expected to
100 be the main hardware limitation [12] and henceforth we only
M = 500 P−ZF
ZF
consider those impairments in this section.
M = 100 MR Similar to [33]–[35], we model the hardware impairments√
50 as a reduction of the original signals by a factor 1 − 2
and replacing it with Gaussian distortion noise that carries
the removed power. More precisely, the UL system model in
(2) is generalized as
0
−10 −5 0 5 10 15 20 K p 
Signal−to−Noise Ratio (SNR), ρ/σ2
XX
yj = hjlk (1 − 2 )plk xlk + εlk + nj , (32)
Fig. 12: Impact of SNR variations on the SE. l∈L k=1

where εlk ∼ CN (0, 2 plk ) is the UL distortion noise caused at

UE k in cell l, and the DL system model in (3) is generalized
Finally, Fig. 13 investigates how the length of the coherence
as
block, S, affects the per-cell SE. In the case of M = 100
K
!
antennas, the gain of increasing S above 500 is relatively p XX
zjk = 1 − 2 hTljk wlm slm + ηjk + ejk , (33)
small—the system cannot schedule more UEs since the ratio
l∈L m=1
M/K would then be too small, so the gain mainly comes from PK
reducing the prelog factor (1 − B where ejk ∼ CN (0, 2 ( l∈L m=1 khTljk wlm k2 + σ 2 )) is
P
S ). However, in the case of
M = 500, the system can utilize an increasing S to schedule the DL
P distortion
PK noise caused at UE k in cell j. Notice
more UEs and achieve major improvements in SE. As the that l∈L m=1 khTljk wlm k2 + σ 2 is the power of the term
number of UEs increases, the part of the intra-cell interference in parenthesis in (33). The parameter  determines the level
that cannot be rejected due to imperfect CSI becomes the main of impairments and can be interpreted as the error vector
limiting factor. The benefit of P-ZF then diminishes. magnitude (EVM) [34]; typical values in LTE are in the range
 12

(1 − 2 )pjk |E{h} {gjk

H
hjjk }|2
SINR
^ jk =
K
. (35)
P P
plm E{h} {|gjk
H
hjlm |2 } − (1 − 2 )pjk |E{h} {gjk
H
hjjk }|2 + σ 2 E{h} {kgjk k2 }
l∈Lm=1

400
0 ≤  ≤ 0.17 [37]. Based on these generalized system models,

Spectral Efficiency (SE) [bit/s/Hz/cell]

the following counterpart of Lemmas 2 and 3 is obtained. 350 Ideal Hardware
=0
300
Lemma 4. Under hardware impairments, a jointly achievable
SE in the UL and DL of an arbitrary UE k in cell j is 250
Hardware
  200 Impairments
B n o
1− E{z} log2 (1 + SINR
^ jk ) [bit/s/Hz] (34) 150
 = 0.1
S
Asymptotic limits
100 P−ZF
where the effective SINR is given in (35) at the top of the page.
ZF
50
Proof: The proof is given in the appendix. MR

The SE expression in Lemma 4 resembles our previous 0

1 2 3 4 5
10 10 10 10 10
results in Section III, with the only differences that there is a Number of BS Antennas (M)
loss in desired signal power by a factor (1 − 2 ) and that this Fig. 14: Optimized per-cell SE with or without hardware
power is turned into self-interference in the denominator of the impairments.
SINR. Under the assumption of MR, ZF, or P-ZF processing
in the UL and DL, we have the following closed-form SE
expression. [37]. Interestingly, there is only a tiny difference in SE for
M < 5000, mainly because the SE per UE is relatively small
Theorem 4. Let Lj (β) ⊂ L be the subset of cells that uses at the optimized operating points and thus the distortion noise
the same pilots as cell j. Looking jointly at the UL and DL, is only a minor limiting factor. For higher number of antennas,
an achievable SE in cell j under hardware impairments is the difference is substantial because of the asymptotic limits
!
1 − 2
 
B for ideal hardware in (27) and for hardware impairments in
SEj = K 1 − log2 1 + scheme [bit/s/Hz/cell]
S Ij + 2 (37) are different. We conclude that hardware impairments
(36) seem to have a small impact on practical massive MIMO
where the interference term Ijscheme is defined in (16) and systems, which have been optimized for high SE.
depends on Gscheme and Zjl scheme
. The parameter values with
MR, ZF, and P-ZF are as follows:
VI. C ONCLUSION
Scheme Gscheme scheme
Zjl
MR M (1 − 2 ) K This paper investigated how many UEs, K, that should
   be scheduled in massive MIMO systems to maximize the
(1)
(1−2 )µjl

 SE per cell for a fixed M . Conventional SE expressions are
K 1 −

  
(1) σ2
 P
µj` + Bρ strongly dependent on the UE positions, which makes it hard
ZF (M − K)(1 − 2 ) `∈Lj (β)
to optimize K. In contrast, we derived new SE expressions

 if l ∈ Lj (β)


K if l 6∈ L (β)
 that are independent of the instantaneous UE positions, due
j to power control and averaging over random UE locations. In
 
2
(1− )µjl
(1) fact, the new expressions are the same for the UL and DL,
P-ZF (M − B)(1 − 2 ) K 1 − P (1) σ2
 which allows for joint network optimization. When applied
µj` + Bρ
`∈Ll (β) to symmetric network topologies, where each cell is repre-
If M → ∞ (with K, B ≤ S < ∞), the effective SINRs with sentable for any cell, these expressions can directly provide the
these processing schemes approach the upper limit network-wide performance—which otherwise would require
extensive Monte-Carlo simulations.
1 − 2 The analytic results treat MR and ZF processing and a new
P (2)
. (37)
µjl + 2 distributed cooperation scheme, coined P-ZF, that suppresses
l∈Lj (β)\{j} inter-cell interference by listening to the pilot transmissions
Proof: This result follows straightforwardly from Theo- from neighboring cells. The asymptotic analysis shows that
S
rems 1–3, since the SINR expressions in (35) only differ from the SE-optimal K ∗ approaches 2β as M → ∞, irrespective
those in Section III by the (1 − 2 )-factors. of the processing scheme. Hence, B = βK ∗ → S2 which
Using the tractable SE expression in Theorem 4 for simu- means that half the frame should be spent on pilot signaling
lation, Fig. 14 shows the per-cell SE in the average inter-cell when M is large enough. The corresponding asymptotic SE
interference. This figure shows results for ideal hardware with limit is not reached for practical M , but an unconventionally
 = 0 (as in Fig. 4(a)) and for hardware impairments with large fraction of the frame should still be allocated to pilots:
 = 0.1, which is a large EVM number in these contexts 5% to 40% were observed in simulations for M ≤ 1000.
 13

(ul) viHjk Ψ−1

j vijk
SINRjk = K
  2  (42)
dj (zlm ) 1 d (zlm ) −1
− viHjk Ψ−1 σ2
P P
dl (zlm ) M + djl (zlm ) v H
Ψ
ijk j v ilm j vijk + Mρ
l∈Lm=1

Generally speaking, high per-cell SEs are achieved by matrix Cjlk is given by
scheduling many UEs for simultaneous transmission, while
dj (zlk )
the SE per UE might only be 1–4 bit/s/Hz. P-ZF gives the ρ IM − E{h} {heff
jlk vec(Yj ) }
H

highest performance per UE, while MR gives the lowest SE dl (zlk )

−1
per UE. In contrast, MR schedules the largest number of UEs × E{h} {vec(Yj )vec(Yj )H } E{h} {heff vec(Yj )H }H
 jlk
and P-ZF the smallest number. ZF processing is often the best

dj (zlk ) dj (zlk ) H −1
choice in terms of per-cell SE, thus showing that the inter-cell =ρ 1− v Ψ vilk IM
dl (zlk ) dl (zlk ) ilk j
interference suppression offered by P-ZF is only needed in 
dj (zlk )

special cases with strong inter-cell interference. The extensive dj (zlk )  dl (zlk ) B
=ρ 1− P PK dj (z`m ) H  IM ,
simulations show that massive MIMO prefers an SNR of 0– dl (zlk ) v vi`m + σ
2
`∈L m=1 d` (z`m ) ilk ρ
5 dB, that a non-universal pilot reuse of β = 3 is often (41)
a decent choice, and that the technology is very robust to
distortion noise from hardware impairments. Based on the where the last equality follows from the fact that the pilot
simulations, we notice that massive MIMO with M = 100 signals form an orthogonal basis.
can easily achieve a 10× gain in SE over the IMT-Advanced
Proof of Theorem 1: The first step for MR combining
requirement of 3 bit/s/Hz/cell. For large arrays with M = 500
is to compute the expectations in (12) with respect to the
antennas, massive MIMO can even provide a 40× gain over
channel realizations. These are obtained from [13, Corollary
IMT-Advanced. The results in this paper are for uncorrelated d (zlm )
2] by setting κ = δ = 0, ξ = σ 2 , and λjlm = djl (zlm ).
fading, while spatially correlated fading is expected to reduce
Plugging these expressions into (12) yields, for MR, the
the inter-user interference [18] thus leading to higher SEs and
expression in (42) at the top of the page, by multiplying each
allowing for smaller β.
term by M 2 ρ2 vH 1Ψ−1 v . The expression in (16) for MR
ijk j ijk

is now obtained by considering an achievable lower bound

1
A PPENDIX : C OLLECTION OF P ROOFS E{z} {log2 (1 + f ({z}) )} ≥ log2 (1 + E{z} {f1({z})} ) where the
expectation with respect to user positions is moved to the
Proof of Lemma 1: As shown in [38, Theorem 11.1], the
denominator of the SINRs using Jensen’s inequality. This leads
expression for an MMSE estimator ĥeff eff
jlk of hjlk with jointly
to expectations of the following types:
Gaussian channels and (colored) noise is ( )

−1 1
E{h} {heff
jlk vec(Yj ) } E{h} {vec(Yj )vec(Yj ) }
H H
vec(Yj ) E{z}
viHjk Ψ−1
j vijk
(38) P 
σ2 
PK dj (z`m̃ ) H
where vec(·) denotes vectorization. Direct algebraic com-  v i vi +
`∈L m̃=1 d` (z`m̃ ) jk ` m̃ ρ
putation using the vectorization rule (CT ⊗ A)vec(B) = = E{z}
B
vec(ABC), where ⊗ is the Kronecker product, shows that
 
(1) σ2
P
E{h} {heff `∈Lj (β) µj` B + ρ
 eff eff H H

jlk vec(Yj ) } = E{h} hjlk (hjlk ) vilk ⊗ IM
H
= (43)
  B
dj (zlk )
= viHlk ⊗ ρ IM (39)
dl (zlk ) ( K  2 H −1 )
XX dj (zlm ) vijk Ψj vilm X (2)
E{z} −1 = µjl
since the channels are independent. Similarly, the mutual dl (zlm ) viHjk Ψj vijk
l∈L m=1 l∈L (β) j
independence of the UE channels implies that
(44)
E{h} {vec(Yj )vec(Yj )H } = σ 2 IM B  
K K
X X dj (zlm ) 
XX dl (zlm )
E{hj`m } vec(heff eff
 T T H

+ j`m vi`m )vec(hj`m vi`m )
E{z}
 v H
Ψ−1 vijk
l∈L m=1 ijk j

`∈L m=1
K σ2
! P (1)
XX dj (z`m ) `∈Lj (β) µj` B +
v i v H + σ 2 IB (1) ρ (2) (1)
X X
= ρ ⊗ IM . (40) = Kµjl + µjl − (µjl )2
d` (z`m ) `m i`m B
`∈L m=1 l∈L l∈Lj (β)
(45)
The expression (6) is obtained by substituting (39) and (40)
into (38), normalizing by ρ and using the vectorization rule where we have utilized the definition in (18) and the non-
above. According to the definition in (8), the error covariance universal pilot reuse assumption to identify the expectations.
 14

(ul) 1
SINRjk = d (z )
 dj (zlm ) H
 (49)
K 2 viH vilm ρ dj (z lm) 1−AZF −1
jl dl (zlm ) vilm Ψj vilm

P P dj (zlm ) σ2
dl (zlm )
jk
B + l lm
H −1
(M −K)ρvi Ψj vijk
−1+ (M −K)ρviH Ψ−1
l∈Lm=1 jk j vijk
jk

(ul) 1
SINRjk = d (z )
 d (z )
 (54)
K 2 viH vilm ρ dj (z lm) 1− dj (z lm) viH Ψ−1 j vilm

P P dj (zlm ) σ2
dl (zlm )
jk
B + l lm l lm
(M −B)ρviH Ψ−1
lm
vijk
−1+ (M −B)ρviH Ψ−1
l∈Lm=1 jk j j vijk
jk

2 H
vijk vilm

The expectations in (12) with respect to the channel real- P-ZF H dj (zlm )
plm E{h} {|(gjk ) hjlm |2 } =
izations for ZF combining are dl (zlm ) B
 
dj (zlm ) dj (zlm ) H −1
1 ρ dl (zlm ) 1 − dl (zlm ) vilm Ψj vilm
ZF 2
E{h} {kgjk k }= (46) + (53)
(M − K)ρviHjk Ψ−1
j vijk (M − B)ρviHjk Ψ−1 j vijk
ZF H
pjk |E{h} {(gjk ) hjjk }|2 = 1 (47) by following the procedures used for ZF. Using (51)–(53),
we obtain the expression (54) at the top of the page for P-ZF.
The final expression is obtained by considering an achievable
2 H 1 1
vijk vilm lower bound E{z} {log2 (1 + f (z) )} ≥ log2 (1 + E{z} {f (z)} )

ZF H dj (zlm )
plm E{h} {|(gjk ) hjlm |2 } =
dl (zlm ) B using Jensen’s inequality, similar to the ZF case in Theorem1.
 
dj (zlm ) ZF dj (zlm ) H −1
ρ dl (zlm ) 1 − Ajl dl (zlm ) vilm Ψj vilm Proof of Theorem 3: Suppose that γjk = SINRjk is the
(ul)
+ (48) UL SINR value achieved by UE k in cell j for a given receive
(M − K)ρviHjk Ψ−1 j vijk
combining scheme. The goal of the proof is to show that we
(dl)
with AZF = 1 if l ∈ Lj (β) and zero otherwise, where can also achieve γjk = SINRjk for the DL SINR in (23).
jl
(46) follows from the definition of ZF and by utilizing well- This condition can also be expressed as
known properties of Wishart matrices (see e.g., [8, Proof γjk E{h} {kǧjk k2 }
of Proposition 2]) and (47) is a consequence of the ZF |E{h} {ǧjk
H
hjjk }|2
principle. The first term in (48) follows from (47) whenever qjk
viHjk vilm 6= 0 (i.e., when the same pilot signal is used). = K
.
H h
|E{h} {ǧjk 2
jjk }|
H h
{|ǧlm 2
E ljk | }
qlm {h}
P P
The second term is the product between E{h} {kgjk ZF 2
k } and E{h} {kǧlm k2 } − qjk E{h} {kǧ 2
jk k }
+ σ2
l∈Lm=1
the variance of the estimation error of the effective channel (55)
√
plm hjlm if AZF
jl 6= 0 (i.e., if the UE is in a cell l ∈ Lj (β))
√ We define the K|L|×K|L| block matrix Ψ, where each block
or the original variance of plm hjlm if AZF jl = 0. Using
(46)–(48), we obtain the expression (49) at the of the page for is K × K and the (j, l)th block is denoted Ψjl . Its (k, m)th
ZF. Finally, the achievable SE in the theorem is obtained by element is given by
using Jensen’s inequality in the same way as for MR, where [Ψjl ]k,m
the expectation in (43)–(45) reappear along with
 H
H 2
 E{h} {|ǧlm hljk2| } − |E{h} {ǧjk hjjk }|2
E{h} {kǧlm k } E{h} {kǧjk k2 } if k = m, j = l,
= E {|ǧH h |2 }
vilm Ψ−1
(  2 H )
 {h} lm ljk2
dj (zlm ) j vilm E{h} {kǧlm k } otherwise.
E{z} −
dl (zlm ) viHjk Ψ−1j vijk
(56)
P (1)
µj` + Bρ σ 2
(50) Moreover, we define the K|L| × K|L| block diagonal matrix
(1) `∈Lj (β) D, where the jth K × K block is Dj and its kth diagonal
≤ −(µjl )2 P (1) , element is
σ2
µj` + Bρ γjk E{h} {kǧjk k2 }
`∈Ll (β) [Dj ]k,k = . (57)
|E{h} {ǧjk
H
hjjk }|2
where the inequality is once again from Jensen’s inequality. Using this notation, (55) can be expressed as
Proof of Theorem 2: Similar to the proof of Theorem 1, qjk
[Dj ]k,k = K
for P-ZF we obtain P P
qlm [Ψjl ]k,m + σ 2
l∈Lm=1
P-ZF 2 1 (58)
E{h} {kgjk k }= (51) K
(M − B)ρviHjk Ψ−1
XX
j vijk
2
⇔ [Dj ]k,k σ = qjk − qlm [Dj ]k,k [Ψjl ]k,m .
l∈L m=1

P-ZF H
In matrix form, the DL SINR conditions for all UEs in
pjk |E{h} {(gjk ) hjjk }|2 = 1 (52) all cells can be expressed as Dσ 2 = q − DΨq, where
 15

(1 − 2 )|E{s}|2 (1 − 2 )|E{s}|2
= (61)
(1 − 2 ) (E{|n|2 } + E{|s| } − |E{s}| ) +  (E{|n| } + E{|s| })
2 2 2 2 2 E{|n| } + E{|s|2 } − (1 − 2 )|E{s}|2
2

q = [qT1 . . . qT|L| ]T and qj = [qj1 . . . qjK ]T contain the DL [6] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot
transmit powers in the jth cell. This expression can now be contamination and precoding in multi-cell TDD systems,” IEEE Trans.
Commun., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
solved for q. The matrix D depends only on the precoding [7] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
vectors, thus for any choice of precoding scheme the sought of cellular networks: How many antennas do we need?” IEEE J. Sel.
SINRs are achieved by the power control policy Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
[8] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral effi-
q? = σ 2 (IK|L| − DΨ)−1 D1 (59) ciency of very large multiuser MIMO systems,” IEEE Trans. Commun.,
vol. 61, no. 4, pp. 1436–1449, Apr. 2013.
? [9] E. Björnson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal design
where 1 is the vector with only ones. q is a feasible
of energy-efficient multi-user MIMO systems: Is massive MIMO the
power control (i.e., has positive values) if all eigenvalues of answer?” IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3059–3075,
(IK|L| − DΨ) are larger than zero. We need to show that this Jun. 2015.
always holds. We notice that the UL SINR condition, which is [10] D. Ha, K. Lee, and J. Kang, “Energy efficiency analysis with cir-
cuit power consumption in massive MIMO systems,” in Proc. IEEE
satisfied by assumption, can be expressed in a similar matrix Int. Symp. Personal, Indoor and Mobile Radio Commun. (PIMRC), 2013.
form where Ψ is replaced by ΨT : [11] H. Yang and T. Marzetta, “Total energy efficiency of cellular large scale
antenna system multiple access mobile networks,” in Proc. IEEE Online
Dσ 2 = p − DΨT p ⇔ p = σ 2 (IK|L| − DΨT )−1 D1, Conference on Green Communications (OnlineGreenComm), 2013.
(60) [12] E. Björnson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive
MIMO systems with non-ideal hardware: Energy efficiency, estimation,
where p = [pT1 . . . pT|L| ]T and pj = [pj1 . . . pjK ]T , if ǧjk = and capacity limits,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112–
scheme
gjk for all j and k. Since the eigenvalues of (IK|L| −DΨ) 7139, Nov. 2014.
and (IK|L| −DΨT ) are the same, we can always select the DL [13] E. Björnson, M. Matthaiou, and M. Debbah, “Massive MIMO with
arbitrary non-ideal arrays: Hardware scaling laws and circuit-aware
powers according to (59). It is straightforward to verify that design,” IEEE Trans. Wireless Commun., vol. 14, no. 8, pp. 4353–4368,
1T q? = pT 1, thus the total transmit power is the same in the Aug. 2015.
DL and UL. Since the same SINRs as in the UL are achieved [14] A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson, “Uplink perfor-
mance of time-reversal MRC in massive MIMO systems subject to phase
in the DL for any UE positions, the SE in (25) follows directly noise,” IEEE Trans. Wireless Commun., vol. 14, no. 2, pp. 711–723, Feb.
from Theorems 1 and 2. 2015.
[15] H. Huh, G. Caire, H. Papadopoulos, and S. Ramprashad, “Achieving
Proof of Lemma 4: The derivations of Lemmas 2 and 3 are “massive MIMO” spectral efficiency with a not-so-large number of
antennas,” IEEE Trans. Wireless Commun., vol. 11, no. 9, pp. 3226–
based on the following principle: if we receive s+n, where s is 3239, Sept. 2012.
a Gaussian information signal and n is uncorrelated noise, then [16] M. Li, Y.-H. Nam, B. Ng, and J. Zhang, “A non-asymptotic throughput
|E{s}|2
for massive MIMO cellular uplink with pilot reuse,” in Proc. IEEE
an achievable SE is log2 1+ E{|n|2 }+E{|s| 2 }−|E{s}|2 [25]. For
Globecom, 2012.
the hardware impairment models in (32) and√(33), the received [17] R. Müller, M. Vehkaperä, and L. Cottatellucci, “Blind pilot decontami-
signals (after linear processing) behave as 1 − 2 (s + n) + nation,” in Proc. WSA, 2013.
η instead, where E{|η|2 } = E{|n|2 } + E{|s|2 }. Since the [18] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
to channel estimation in large-scale multiple-antenna systems,” IEEE J.
distortion η is uncorrelated with s and n by assumption, n + η Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, Feb. 2013.
is also uncorrelated with s and the corresponding SINR is [19] M. Li, S. Jin, and X. Gao, “Spatial orthogonality-based pilot reuse for
computed in (61) at the top of the page. The only impact of multi-cell massive MIMO transmission,” in Proc. WCSP, 2013.
[20] M. Karlsson and E. G. Larsson, “On the operation of massive MIMO
the distortion is thus the (1 − 2 )-factors in front of |E{s}|2 with and without transmitter CSI,” in Proc. IEEE SPAWC, 2014.
in the numerator and denominator. The UL SINRs in (35) [21] X. Gao, O. Edfors, F. Rusek, and F. Tufvesson, “Massive MIMO
follow directly from this observation, while the DL SINRs are performance evaluation based on measured propagation data,” IEEE
Trans. Wireless Commun., vol. 14, no. 7, pp. 3899–3911, July 2015.
achieved by also utilizing the power control policy from the [22] K. Guo, Y. Guo, G. Fodor, and G. Ascheid, “Uplink power control
proof of Theorem 3. with MMSE receiver in multi-cell MU-massive-MIMO systems,” in
Proc. IEEE ICC, 2014.
R EFERENCES [23] H. Yang and T. Marzetta, “A macro cellular wireless network with
uniformly high user throughputs,” in Proc. IEEE VTC-Fall, 2014.
[1] Nokia Siemens Networks, “2020: Beyond 4G radio evolution for the [24] M. Biguesh and A. B. Gershman, “Downlink channel estimation in
Gigabit experience,” White Paper, Tech. Rep., 2011. cellular systems with antenna arrays at base stations using channel
[2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- probing with feedback,” EURASIP J. Appl. Signal Process., vol. 2004,
bers of base station antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 9, pp. 1330–1339, 2004.
no. 11, pp. 3590–3600, Nov. 2010. [25] M. Medard, “The effect upon channel capacity in wireless communica-
[3] R. Baldemair, E. Dahlman, G. Fodor, G. Mildh, S. Parkvall, Y. Selén, tions of perfect and imperfect knowledge of the channel,” IEEE Trans.
H. Tullberg, and K. Balachandran, “Evolving wireless communications: Inf. Theory, vol. 46, no. 3, pp. 933–946, May 2000.
Addressing the challenges and expectations of the future,” IEEE Veh. [26] B. Hassibi and B. M. Hochwald, “How much training is needed in
Technol. Mag., vol. 8, no. 1, pp. 24–30, Mar. 2013. multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no. 4,
[4] F. Boccardi, R. Heath, A. Lozano, T. Marzetta, and P. Popovski, “Five pp. 951–963, Apr. 2003.
disruptive technology directions for 5G,” IEEE Commun. Mag., vol. 52, [27] E. Björnson and E. Jorswieck, “Optimal resource allocation in coordi-
no. 2, pp. 74–80, Feb. 2014. nated multi-cell systems,” Foundations and Trends in Communications
[5] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, “Massive and Information Theory, vol. 9, no. 2-3, pp. 113–381, 2013.
MIMO for next generation wireless systems,” IEEE Commun. Mag., [28] H. Boche and M. Schubert, “A general duality theory for uplink and
vol. 52, no. 2, pp. 186–195, Feb. 2014. downlink beamforming,” in Proc. IEEE VTC-Fall, 2002, pp. 87–91.
 16

[29] L. Zheng and D. Tse, “Communication on the Grassmann manifold: A Erik G. Larsson received his Ph.D. degree from
geometric approach to the noncoherent multiple-antenna channel,” IEEE Uppsala University, Sweden, in 2002. Since 2007,
Trans. Inf. Theory, vol. 48, no. 2, pp. 359–383, Feb. 2002. he is Professor and Head of the Division for Com-
[30] A. Lozano, R. Heath, and J. Andrews, “Fundamental limits of cooper- munication Systems in the Department of Electrical
ation,” IEEE Trans. Inf. Theory, vol. 59, no. 9, pp. 5213–5226, Sept. PLACE Engineering (ISY) at Linköping University (LiU)
2013. PHOTO in Linköping, Sweden. He has previously been As-
[31] V. M. Donald, “The cellular concept,” Bell System Technical Journal, HERE sociate Professor (Docent) at the Royal Institute
vol. 58, no. 15-41, pp. 113–381, 1979. of Technology (KTH) in Stockholm, Sweden, and
[32] D. Cox, “Cochannel interference considerations in frequency reuse Assistant Professor at the University of Florida and
small-coverage-area radio systems,” IEEE Trans. Commun., vol. 30, the George Washington University, USA. In the
no. 1, pp. 135–142, Jan. 1982. spring of 2015 he was a Visiting Fellow at Princeton
[33] T. Schenk, RF Imperfections in High-Rate Wireless Systems: Impact and University, USA, for four months.
Digital Compensation. Springer, 2008. His main professional interests are within the areas of wireless commu-
[34] M. Wenk, MIMO-OFDM Testbed: Challenges, Implementations, and nications and signal processing. He has published some 100 journal papers
Measurement Results, ser. Series in microelectronics. Hartung-Gorre, on these topics, he is co-author of the textbook Space-Time Block Coding
2010. for Wireless Communications (Cambridge Univ. Press, 2003) and he holds 15
[35] W. Zhang, “A general framework for transmission with transceiver issued and many pending patents on wireless technology.
distortion and some applications,” IEEE Trans. Commun., vol. 60, no. 2, He has served as Associate Editor for several major journals, including the
pp. 384–399, Feb. 2012. IEEE Transactions on Communications (2010-2014) and IEEE Transactions
[36] U. Gustavsson et al., “On the impact of hardware impairments on on Signal Processing (2006-2010). He serves as chair of the IEEE Signal
massive MIMO,” in Proc. IEEE GLOBECOM, 2014. Processing Society SPCOM technical committee in 2015–2016 and as chair
[37] H. Holma and A. Toskala, LTE for UMTS: Evolution to LTE-Advanced, of the steering committee for the IEEE Wireless Communications Letters in
2nd ed. Wiley, 2011. 2014–2015. He is the General Chair of the Asilomar Conference on Signals,
[38] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Systems and Computers in 2015 (he was Technical Chair in 2012). He received
Theory. Prentice Hall, 1993. the IEEE Signal Processing Magazine Best Column Award twice, in 2012
and 2014, and he is receiving the IEEE ComSoc Stephen O. Rice Prize in
Communications Theory in 2015.

Mérouane Debbah (SM’08, F’15) entered the Ecole

Normale Supérieure de Cachan (France) in 1996
where he received his M.Sc and Ph.D. degrees
respectively. He worked for Motorola Labs (Saclay,
PLACE France) from 1999-2002 and the Vienna Research
PHOTO Center for Telecommunications (Vienna, Austria)
HERE until 2003. From 2003 to 2007, he joined the Mobile
Communications department of the Institut Eurecom
(Sophia Antipolis, France) as an Assistant Professor.
Since 2007, he is a Full Professor at CentraleSupelec
(Gif-sur-Yvette, France). From 2007 to 2014, he was
the director of the Alcatel-Lucent Chair on Flexible Radio. Since 2014, he
is Vice-President of the Huawei France R&D center and director of the
Mathematical and Algorithmic Sciences Lab.
His research interests lie in fundamental mathematics, algorithms, statistics,
information & communication sciences research. He is an Associate Editor
in Chief of the journal Random Matrix: Theory and Applications and was an
associate and senior area editor for IEEE Transactions on Signal Processing
respectively in 2011-2013 and 2013-2014. Dr. Debbah is a recipient of the
Emil Björnson (S’07, M’12) received his M.S. de- ERC grant MORE (Advanced Mathematical Tools for Complex Network
gree in Engineering Mathematics from Lund Univer- Engineering). He is a IEEE Fellow, a WWRF Fellow and a member of
sity, Sweden, in 2007. He received his Ph.D. degree the academic senate of Paris-Saclay. He has managed 8 EU projects and
in Telecommunications from the KTH Royal Insti- more than 24 national and international projects. He received 14 best paper
PLACE tute of Technology, Stockholm, Sweden, in 2011.
PHOTO awards, among which the 2007 IEEE GLOBECOM best paper award, the
From 2012 to July 2014, he was a joint postdoc at Wi-Opt 2009 best paper award, the 2010 Newcom++ best paper award, the
HERE Suplec, Gif-sur-Yvette, France, and at KTH Royal WUN CogCom Best Paper 2012 and 2013 Award, the 2014 WCNC best
Institute of Technology. He is currently an Assistant paper award, the 2015 ICC best paper award, the 2015 IEEE Communications
Professor at the Department of Electrical Engineer- Society Leonard G. Abraham Prize and 2015 IEEE Communications Society
ing (ISY) at Linköping University, Sweden. Fred W. Ellersick Prize as well as the Valuetools 2007, Valuetools 2008,
His research interests include multi-antenna cel- CrownCom2009, Valuetools 2012 and SAM 2014 best student paper awards.
lular communications, radio resource allocation, energy efficiency, massive He is the recipient of the Mario Boella award in 2005, the IEEE Glavieux
MIMO, and network topology design. He is the first author of the textbook Prize Award in 2011 and the Qualcomm Innovation Prize Award in 2012. He
Optimal Resource Allocation in Coordinated Multi-Cell System (Foundations is the co-founder of the start-up Ximinds.
and Trends in Communications and Information Theory, 2013). He is also
dedicated to reproducible research and has made a large amount of simulation
code publicly available. Dr. Björnson received the 2014 Outstanding Young
Researcher Award from IEEE ComSoc EMEA and the 2015 Ingvar Carlsson
Award. He has received 5 best paper awards for novel research on optimization
and design of multi-cell multi-antenna communications: ICC 2015, WCNC
2014, SAM 2014, CAMSAP 2011, and WCSP 2009.