Massive MIMO With Non-Ideal Arbitrary Arrays: Hardware Scaling Laws and Circuit-Aware Design

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TWC.2015.2420095, IEEE Transactions on Wireless Communications
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS 1
Massive MIMO with Non-Ideal Arbitrary Arrays:

Hardware Scaling Laws and Circuit-Aware Design
Emil Bjornson, Member, IEEE, Michail Matthaiou, Senior Member, IEEE, and Merouane Debbah, Fellow, IEEE
AbstractMassive multiple-input multiple-output (MIMO) Massive densification, in terms of more service antennas per
systems are cellular networks where the base stations (BSs) are unit area, has been identified as a key to higher area throughput
equipped with unconventionally many antennas, deployed on co- in future wireless networks [5][7]. The downside of densifi-
located or distributed arrays. Huge spatial degrees-of-freedom
are achieved by coherent processing over these massive arrays, cation is that even stricter requirements on the interference co-
which provide strong signal gains, resilience to imperfect channel ordination need to be imposed. Densification can be achieved
knowledge, and low interference. This comes at the price of more by adding more antennas to the macro BSs and/or distributing
infrastructure; the hardware cost and circuit power consumption the antennas by ultra-dense operator-deployment of small
scale linearly/affinely with the number of BS antennas N . Hence, BSs. These two approaches are non-conflicting and represent
the key to cost-efficient deployment of large arrays is low-cost
antenna branches with low circuit power, in contrast to todays the two extremes of the massive MIMO paradigm [7]: a
conventional expensive and power-hungry BS antenna branches. large co-located antenna array or a geographically distributed
Such low-cost transceivers are prone to hardware imperfections, array (e.g., using a cloud RAN approach [8]). The massive
but it has been conjectured that the huge degrees-of-freedom MIMO topology originates from [9] and has been given many
would bring robustness to such imperfections. We prove this alternative names; for example, large-scale antenna systems
claim for a generalized uplink system with multiplicative phase-
drifts, additive distortion noise, and noise amplification. Specifi- (LSAS), very large MIMO, and large-scale multi-user MIMO.
cally, we derive closed-form expressions for the user rates and a The main characteristics of massive MIMO are that each cell
scaling law that shows how fast the hardware imperfections can performs coherent processing on an array of hundreds (or even
increase with N while maintaining high rates. The connection thousands) of active antennas, while simultaneously serving
between this scaling law and the power consumption of different tens (or even hundreds) of users in the uplink and downlink.
transceiver circuits is rigorously exemplified.
This reveals that one
can make the circuit power increase as N , instead of linearly, In other words, the number of antennas, N , and number of
by careful circuit-aware system design. users per BS, K, are unconventionally large, but differ by
a factor two, four, or even an order of magnitude. For this
Index TermsAchievable user rates, channel estimation, mas-
sive MIMO, scaling laws, transceiver hardware imperfections. reason, massive MIMO brings unprecedented spatial degrees-
of-freedom, which enable strong signal gains from coherent
reception/transmit beamforming, give nearly orthogonal user
channels, and resilience to imperfect channel knowledge [10].
I. I NTRODUCTION Apart from achieving high area throughput, recent works
Interference coordination is the major limiting factor in have investigated additional ways to capitalize on the huge
cellular networks, but modern multi-antenna base stations degrees-of-freedom offered by massive MIMO. Towards this
(BSs) can control the interference in the spatial domain end, [5] showed that massive MIMO enables fully distributed
by coordinated multipoint (CoMP) techniques [1][3]. The coordination between systems that operate in the same band.
cellular networks are continuously evolving to keep up with Moreover, it was shown in [11] and [12] that the transmit
the rapidly increasing demand for wireless connectivity [4]. uplink/downlink powers can be reduced as 1N with only a
minor loss in throughput. This allows for major reductions
E. Bjornson was with the Alcatel-Lucent Chair on Flexible Radio, Supelec, in the emitted power, but is actually bad from an overall
Gif-sur-Yvette, France, and with the Department of Signal Processing, KTH
Royal Institute of Technology, Stockholm, Sweden. He is currently with the energy efficiency (EE) perspectivethe EE is maximized by
Department of Electrical Engineering (ISY), Linkoping University, Linkoping, increasing the emitted power with N to compensate for the
Sweden (email: emil.bjornson@liu.se). increasing circuit power consumption [13].
M. Matthaiou is with the School of Electronics, Electrical Engineering
and Computer Science, Queens University Belfast, Belfast, U.K. and with This paper explores whether the huge degrees-of-freedom
the Department of Signals and Systems, Chalmers University of Technology, offered by massive MIMO provide robustness to transceiver
Gothenburg, Sweden (email: m.matthaiou@qub.ac.uk). hardware imperfections/impairments; for example, phase
M. Debbah is with the Alcatel-Lucent Chair on Flexible Radio, Supelec,
Gif-sur-Yvette, France (email: merouane.debbah@supelec.fr). noise, non-linearities, quantization errors, noise amplification,
Parts of this work were published at the IEEE Conference on Acoustics, and inter-carrier interference. Robustness to hardware imper-
Speech, and Signal Processing (ICASSP), Florence, Italy, May 2014 and at fections has been conjectured in overview articles, such as [7].
the IEEE International Symposium on Communications, Control, and Signal
Processing (ISCCSP), Athens, Greece, May 2014. Such a characteristic is notably important since the deployment
This research has received funding from the EU 7th Framework Programme cost and circuit power consumption of massive MIMO scales
under GA no ICT-619086 (MAMMOET). This research has been supported by linearly with N , unless the hardware accuracy constraints can
ELLIIT, the International Postdoc Grant 2012-228 from the Swedish Research
Council and the ERC Starting Grant 305123 MORE (Advanced Mathematical be relaxed such that low-power, low-cost hardware is deployed
Tools for Complex Network Engineering). which is more prone to imperfections. Constant envelope
1536-1276 (c) 2015 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
precoding was analyzed in [14] to facilitate the use of power- Receive

efficient amplifiers in the downlink, while the impact of phase- antenna 1
drifts was analyzed and simulated for single-carrier systems
in [15] and for orthogonal frequency-division multiplexing Filter LNA Mixer ADC
(OFDM) in [16]. A preliminary proof of the conjecture was
given in [17], but the authors therein considered only additive LO
distortions and, thus, ignored other important characteristics DSP
Receive LO
of hardware imperfections. That paper showed that one can antenna N
tolerate distortion variances that increase as N with only
minor throughput losses, but did not investigate what this Filter LNA Mixer ADC
implies for the design of different transceiver circuits.
In this paper, we consider a generalized uplink massive
Fig. 1. Block diagram of a typical N -antenna receiver. The main circuits are
MIMO system with arbitrary array configurations (e.g., co- shown, but these can be complemented with additional intermediate filters
located or distributed antennas). Based on the extensive liter- and amplifiers depending on the implementation. Most of the circuits affect
ature on modeling of transceiver hardware imperfections (see only one antenna, whilst the LO can be either common for all antennas or
different.
[3], [4], [15], [18][24] and references therein), we propose
a tractable system model that jointly describes the impact
of multiplicative phase-drifts, additive distortion noise, noise also rigorously supported by the analytic scaling law.
amplification, and inter-carrier interference. This stands in This paper extends substantially our conference papers [25]
contrast to the previous works [15][17], which each inves- and [26], by generalizing the propagation model, generalizing
tigating only one of these effects. The following are the main the analysis according to the new model, and providing more
contributions of this paper: comprehensive simulations. The paper is organized as follows:
We derive a new linear minimum mean square error In Section II, the massive MIMO system model under consid-
(LMMSE) channel estimator that accounts for hardware eration is presented. In Section III, a detailed performance
imperfections and allows the prediction of the detrimental analysis of the achievable uplink user rates is pursued and the
impact of phase-drifts. impact of hardware imperfections is characterized, while in
We present a simple and general expression for the Section IV we provide guidelines for circuit-aware design in
achievable uplink user rates and compute it in closed- order to minimize the power dissipation of receiver circuits.
form, when the receiver applies maximum ratio combin- Our theoretical analysis is corroborated with simulations in
ing (MRC) filters. We prove that the additive distortion Section V, while Section VI concludes the paper.
noise and noise amplification vanish asymptotically as Notation: The following notation is used throughout the
N , while the phase-drifts remain but are not paper: Boldface (lower case) is used for column vectors, x,
exacerbated. and (upper case) for matrices, X. Let XT , X , and XH
We obtain an intuitive scaling law that shows how fast denote the transpose, conjugate, and conjugate transpose of X,
we can tolerate the levels of hardware imperfections respectively. A diagonal matrix with a1 , . . . , aN on the main
to increase with N , while maintaining high user rates. diagonal is denoted as diag(a1 , . . . , aN ), while IN is an N N
This is an analytic proof of the conjecture that massive identity matrix. The set of complex-valued N K matrices
MIMO systems can be deployed with inexpensive low- is denoted by CN K . The expectation operator is denoted
power hardware without sacrificing the expected major E{} and , denotes definitions. The matrix trace function is
performance gains. The scaling law provides sufficient tr() and is the Kronecker product. A Gaussian random
conditions that hold for any judicious receive filters. variable x is denoted x N (x, q), where x is the mean and
The practical implications of the scaling law are exem- q is the variance. A circularly symmetric complex Gaussian
plified for the main circuits at the receiver, namely, the random vector x is denoted x CN (x, Q), where x is the
analog-to-digital converter (ADC), low noise amplifier mean and Q is the covariance matrix. The big O notation
f (x)
(LNA), and local oscillator (LO). The main components f (x) = O(g(x)) means that g(x) is bounded as x .
of a typical receiver are illustrated in Fig. 1. The scaling
law reveals the tradeoff between hardware cost, level of II. S YSTEM M ODEL WITH H ARDWARE I MPERFECTIONS
imperfections, and circuit power consumption. In partic-
ular, it shows how a circuit-aware design We consider the uplink of a cellular network with L 1
can make the
circuit power consumption increase as N instead of N . cells. Each cell consists of K single-antenna user equipments
The analytic results are validated numerically in a realis- (UEs) that communicate simultaneously with an array of N
tic simulation setup, where we consider different antenna antennas, which can be either co-located at a macro BS or
deployment scenarios, common and separate LOs, dif- distributed over multiple fully coordinated small BSs. The
ferent pilot sequence designs, and two types of receive analysis of our paper holds for any N and K, but we
filters. A key observation is that separate LOs can provide are primarily interested in massive MIMO topologies, where
better performance than a common LO, since the phase- N K 1. The frequency-flat hchannel from UE iT k in cell
(1) (N )
drifts average out and the interference is reduced. This is l to BS j is denoted as hjlk , hjlk . . . hjlk CN 1
and is modeled as Rayleigh block fading. This means that Nevertheless, the distortions can be classified into three dis-
it has a static realization for a coherence block of T chan- tinct categories: 1) received signals are shifted in phase; 2)
nel uses and independent realizations between blocks.1 The distortion noise is added with a power proportional to the total
UEs channels are independent. Each realization is complex received signal power; and 3) thermal noise is amplified and
Gaussian distributed with zero mean and covariance matrix channel-independent interference is added. To draw general
jlk CN N : conclusions on how these distortion categories affect massive
hjlk CN (0, jlk ). (1) MIMO systems, we consider a generic system model with

(1) (N )
hardware imperfections. The received signal in cell j at a given
The covariance matrix jlk , diag jlk , . . . , jlk is channel use t {1, . . . , T } is modeled as
assumed to be diagonal, which holds if the inter-antenna
L
distances are sufficiently large and the multi-path scattering X
yj (t) = Dj (t) Hjl xl (t) + j (t) + j (t) (3)
environment is rich [27].2 The average channel attenuation
(n) l=1
jlk is different for each combination of cells, UE index, and
receive antenna index n. It depends, for example, on the array where the channel matrices Hjl and transmitted signals xl (t)
geometry and the UE location. Even for co-located antennas are exactly as in (2). The hardware imperfections are defined
(n)
one might have different values of jlk over the array, because as follows:
of the large aperture that may create variations in the shadow 1) The matrix Dj (t) , diag ej1 (t) , . . . , ejN (t) de-

fading. scribes multiplicative phase-drifts, where is the imag-
The received signal yj (t) CN 1 in cell j at a given inary unit. The variable jn (t) is the phase-drift at the
channel use t {1, . . . , T } in the coherence block is conven- nth receive antenna in cell j at time t. Motivated by the
tionally modeled as [9][12] standard phase-noise models in LOs [21], jn (t) follows
L
X a Wiener process
yj (t) = Hjl xl (t) + nj (t) (2)
l=1
jn (t) N (jn (t 1), ) (4)
where the transmit signal in cell l is xl (t) = which equals the previous realization jn (t 1) plus
[xl1 (t) . . . xlK (t)]T CK1 and we use the notation an independent Gaussian innovation of variance . The
Hjl = [hjl1 . . . hjlK ] CN K for brevity. The scalar signal phase-drifts can be either independent or correlated
xlk (t) sent by UE k in cell l at channel use t is either a between the antennas; for example, co-located arrays
deterministic pilot symbol (used for channel estimation) or an might have a common LO (CLO) for all antennas
information symbol from a Gaussian codebook; in any case, which makes the phase-drifts jn (t) identical for all
we assume that the expectation of the transmit energy per n = 1, . . . , N . In contrast, distributed arrays might have
symbol is bounded as E{|xlk (t)|2 } plk . The thermal noise separate LOs (SLOs) at each antenna, which make the
vector nj (t) CN (0, 2 IN ) is spatially and temporally drifts independent, though we let the variance be equal
independent and has variance 2 . for simplicity. Both cases are considered herein.
The conventional model in (2) is well-accepted for small- 2) The distortion noise j (t) CN (0, j (t)), where
scale MIMO systems, but has an important drawback when
K
L X
applied to massive MIMO topologies: it assumes that the large

(1) (N )
X
antenna array consists of N high-quality antenna branches j (t) , 2 E{|xlk (t)|2 }diag |hjlk |2 , . . . , |hjlk |2
l=1 k=1
which are all perfectly synchronized. Consequently, the de- (5)
ployment cost and total power consumption of the circuits for given channel realizations, where the double-sum
attached to each antenna would at least grow linearly with gives the received power at each antenna. Thus, the
N , thereby making the deployment of massive MIMO rather distortion noise is independent between antennas and
questionable, if not prohibitive, from an overall cost and channel uses, and the variance at a given antenna is
efficiency perspective. proportional to the current received signal power at this
In this paper, we analyze the far more realistic scenario antenna. This model can describe the quantization noise
of having inexpensive hardware-constrained massive MIMO in ADCs with gain control [19], approximate generic
arrays. More precisely, each receive array experiences hard- non-linearities [4, Chapter 14], and approximate the
ware imperfections that distort the communication. The exact leakage between subcarriers due to calibration errors.
distortion characteristics depend generally on which modula- The parameter 0 describes how much weaker the
tion scheme is used; for example, OFDM [18], filter bank distortion noise magnitude is compared to the signal
multicarrier (FBMC) [28], or single-carrier transmission [15]. magnitude.
1 The size of the time/frequency block where the channels are static depends 3) The receiver noise j (t) CN (0, IN ) is independent
on UE mobility and propagation environment: T is the product of the of the UE channels, in contrast to the distortion noise.
coherence time c and coherence bandwidth Wc , thus c = 5 ms and This term includes thermal noise, which typically is
Wc = 100 kHz gives T = 500. amplified by LNAs and mixers in the receiver hardware,
2 The analysis and main results of this paper can be easily extended to
arbitrary non-diagonal covariance matrices as in [11] and [17], but at the cost and interference leakage from other frequency bands
of complicating the notation and expressions. and/or other networks. The receiver noise variance must
(a) Pilot sequence Data symbols A. Channel Estimation under Hardware Imperfections
Based on the transmission protocol, the pilot sequence of
(b) Data symbols Pilot sequence Data symbols
UE k in cell j is xjk , [xjk (1 ) . . . xjk (B )]T CB1 . The
(c) pilot sequences are predefined and can be selected arbitrarily
under the power constraints. Our analysis supports any choice,
(d) but it is reasonable to make xj1 , . . . , xjK in cell j mutually
orthogonal to avoid intra-cell interference (this is the reason
Coherence block to have B K).
Example 1: Let X e j , [xj1 . . . xjK ] denote the pilot
Fig. 2. Examples of different ways to distribute the B pilot symbols over the sequences in cell j. The simplest example of linearly inde-
coherence block of length T : (a) beginning of block; (b) middle of block; (c)
uniform pilot distribution; (d) preamble and a few distributed pilot symbols.
pendent pilot sequences (with B = K) is
e temporal , diag(pj1 , . . . , pjK )

X (6)
j
satisfy 2 . If there is no interference leakage,
F = 2 is called the noise amplification factor. where the different sequences are temporally orthogonal since
This tractable generic model of hardware imperfections at only UE k transmits at time k . Alternatively, the pilot
the BSs is inspired by a plethora of prior works [3], [4], sequences can be made spatially orthogonal so that all UEs
[15], [18][24] and characterizes the joint behavior of all transmit at every pilot transmission time, which effectively
hardware imperfections at the BSsthese can be uncalibrated increases the total pilot energy by a factor K. The canonical
imperfections or residual errors after calibration. The model example is to use a scaled discrete Fourier transform (DFT)
in (3) is characterized by three parameters: , , and . The matrix [31]:
model is compatible with the conventional model in (2), which
1 1 ... 1

is obtained by setting = 2 and = = 0. The analysis in K1
1 WK ... WK
this paper holds for arbitrary parameter values. Section IV

e spatial
X , ..

.. .. ..
e temporal
Xj (7)
exemplifies the connection between imperfections in the main j
. . . .
transceiver circuits of the BSs and the three parameters. These 1 B1
WK . . . WK
(B1)(K1)
connections allow for circuit-aware design of massive MIMO
systems.
where WK , e2/K .
In the next section, we derive a channel estimator and
The pilot sequences can also be jointly designed across
achievable UE rates for the system model in (3). By analyzing
cells, to reduce inter-cell interference during pilot transmis-
the performance as N , we bring new insights into the
sion. Since network-wide pilot orthogonality requires B
fundamental impact of hardware imperfections (in particular,
LK, which typically is much larger than the coherence block
in terms of , , and ).
length T , practical networks need to balance between pilot
orthogonality and inter-cell interference. A key design goal is
III. P ERFORMANCE A NALYSIS to allocate non-orthogonal pilot sequences to UEs that have
In this section, we derive achievable UE rates for the uplink nearly orthogonal channel covariance matrices; for example,
multi-cell system in (3) and analyze how these depend on by making tr(jjk jlm ) small for any combination of a UE
the number of antennas and hardware imperfections. We first k in cell j and a UE m in cell l, as suggested in [32].
need to specify the transmission protocol.3 The T channel For any given set of pilot sequences, we now derive esti-
uses of each coherence block are split between transmission of mators of the effective channels
uplink pilot symbols and uplink data symbols. It is necessary
to dedicate B K channel uses for pilot transmission if hjlk (t) , Dj (t) hjlk (8)
the receiving array should be able to spatially separate the
different UEs in the cell. The remaining T B channel at any channel use t {1, . . . , T } and for all j, l, k. The
uses are allocated for data transmission. The pilot symbols conventional multi-antenna channel estimators from [33][35]
can be distributed in different ways: for example, placed in cannot be applied in this paper since the generalized system
the beginning of the block [17], in the middle of the block model in (3) has two non-standard properties: the pilot trans-
[29], uniformly distributed as in the LTE standard [30], or a mission is corrupted by random phase-drifts and the distortion
combination of these approaches [22]. These different cases noise is statistically dependent on the channels. Therefore, we
are illustrated in Fig. 2. The time indices used for pilot derive a new LMMSE estimator for the system model at hand.
T
transmission are denoted by 1 , . . . , B {1, . . . , T }, while CBN

Theorem 1: Let j , yjT (1 ) . . . yjT (B )
D , {1, . . . , T } \ {1 , . . . , B } are the time indices for data denote the combined received signal in cell j from the pilot
transmission. transmission. The LMMSE estimate of hjlk (t) at any channel
3 We assume that the same protocol is used in all cells, for analytic
use t {1, . . . , T } for any l and k is
simplicity. It was shown in [12, Remark 5] that nothing substantially different
hjlk (t) = xHlk D(t) jlk 1

will happen if this assumption is relaxed. j j (9)
pjk |E{vjk
H
(t)hjjk (t)}|2
SINRjk (t) = L P
K
(20)
P
plm E{|vjk (t)hjlm
H
(t)|2 } pjk |E{vjk (t)hjjk
H
(t)}|2 + E{|vjk (t) j
H
(t)|2 } + E{kvjk (t)k2 }
l=1m=1
where Next, we use these channel estimates to design receive filters

and derive achievable UE rates.
2 |t1 | 2 |tB |
D(t) , diag e ,...,e , (10)
L X
X K
j , X`m j`m + IBN , (11)
B. Achievable UE Rates under Hardware Imperfections
`=1 m=1
X`m , X`m + 2 D|x`m |2 , (12) It is difficult to compute the maximum achievable UE rates

when the receiver has imperfect channel knowledge [36], and
D|x`m |2 , diag |x`m (1 )|2 , . . . , |x`m (B )|2 , (13)
hardware imperfections are not simplifying this task. Upper
bounds on the achievable rates were obtained in [17] and
while the element (b1 , b2 ) of X`m CBB is [37]. In this paper, we want to guarantee certain performance
(
|x`m (b1 )|2 , b1 = b2 , and thus seek simple achievable (but suboptimal) rates. The
[X`m ]b1 ,b2 = following lemma provides such rate expressions and builds
x`m (b1 )x`m (b2 )e 2 |b1 b2 | , b1 6= b2 .
upon well-known techniques from [9], [15], [36], [38], [39]
(14)
for computing lower bounds on the mutual information.
The corresponding error covariance matrix is
n H o Lemma 1: Suppose the receiver in cell j has complete
Cjlk (t) = E hjlk (t) hjlk (t) hjlk (t) hjlk (t) statistical channel knowledge and applies the linear receive
filters vjkH
(t) C1N , for t D, to detect the signal from its
= jlk xHlk D(t) jlk 1
H
j (D(t) xlk jlk ) kth UE. An ergodic achievable rate for this UE is
(15)
and the mean-squared error (MSE) is MSEjlk (t) = 1 X
Rjk = log2 1 + SINRjk (t) [bit/channel use] (19)
tr(Cjlk (t)). T
tD
Proof: The proof is given in Appendix B.
It is important to note that although the channels are block
fading, the phase-drifts caused by hardware imperfections where SINRjk (t) is given in (20) at the top of this page and
make the effective channels hjlk (t) change between every all UEs use full power (i.e., E{|xlk (t)|2 } = plk for all l, k).
channel use. The new LMMSE estimator in Theorem 1 pro- Proof: The proof is given in Appendix C.
vides different estimates for each time index t D used for The achievable UE rates in Lemma 1 can be computed for
data transmissionthis is a prediction, interpolation, or retro- any choice of receive filters, using numerical methods; the
spection depending on how the pilot symbols are distributed in MMSE filter is simulated in Section V. Note that the sum in
the coherence block (recall Fig. 2). The LMMSE estimator is (19) has |D| = T B terms, while the pre-log factor T1 also
the same for systems with independent and correlated phase- accounts for the B channel uses of pilot transmissions. The
drifts which brings robustness to modeling errors, but also next theorem gives new closed-form expressions for all the
means that there exist better non-linear estimators that can expectations in (20) when using MRC receive filters.
exploit phase-drift correlations, though we do not pursue this Theorem 2: The expectations in the SINR expression (20)
issue further in this paper. are given in closed form by (21)(24), at the top of the next
The estimator expression is simplified in the special case of MRC
page, when the MRC receive filter vjk (t) = hjjk (t) is used
co-located arrays, as shown by the following corollary. in cell j. The nth column of IN is denoted by en CN 1 in
Corollary 1: If jlk = jlk IN for all j, l, and k, the this paper.
LMMSE estimate in (9) simplifies to Proof: The proof is given in Appendix D.

By substituting the expressions from Theorem 2 into (20),
hjlk (t) = jlk xHlk D(t) 1
j I N j (16)
we obtain closed-form UE rates that are achievable using
and the error covariance matrix in (15) becomes MRC filters. Although the expressions in (21)(24) are easy
to compute, their interpretation is non-trivial. The size of each
H 1 H
Cjlk (t) = jlk 1 jlk xlk D(t) j D(t) xlk IN (17) term depends on the setup and scales differently with N ; note
that each trace-expression and/or sum over the antennas give
where j is the Hermitian matrix a scaling factor of N . This property is easily observed in the
special case of co-located antennas:
L X
K
j ,
X
j`m X`m + IB . (18) Corollary 2: If jlk = jlk IN for all j, l, and k, the MRC
receive filter yields:
`=1 m=1

E{kvjk (t)k2 } = tr xHjk D(t) jjk j1 DH(t) xjk jjk

(21)
E{vjkH
(t)hjjk (t)} = E{kvjk (t)k2 } (22)
1 H
H 2 H
E{|vjk (t)hjlm (t)| } = tr jlm xjk D(t) jjk j D(t) xjk jjk (23)
N N
P P (n1 ) (n1 ) (n2 ) (n2 ) H
jjk jlm jjk jlm xjk D(t) eHn1 1 Xlm en1 eHn2 1

j j DH(t) xjk en2 if a CLO
+ n1 =1n2 =1
2
1
jk (t) jjk j DH(t) xlm jlm
tr xH D if SLOs
N 2
(n) (n)
xHjk D(t) eHn 1 2 D|xlm |2 en eHn 1
P
jjk jlm DH(t) xjk en if a CLO

j j
n=1
+ N 2
(n) (n)
xHjk D(t) eHn 1 (Xlm DH(t) xlm xHlm D(t) ) en eHn 1
P
jjk jlm DH(t) xjk en if SLOs

j j
n=1
L X
X K
(t) j (t)|2 } = 2 plm tr jlm xHjk D(t) jjk 1
H

E{|vjk j DH(t) xjk jjk (24)
l=1 m=1
L X K XN 2
(n) (n)
X
+ 2 xHjk D(t) eHn 1 1
H
plm jjk jlm j (Xlm en en )j DH(t) xjk en
l=1 m=1 n=1
Example 2: The interference power in (20) consists of

multiple terms of the form E{|vjkH
(t)Dj (t) hjlm |2 }. Suppose
E{kvjk (t)k } = N 2jjk xHjk D(t) 1
2 H
j D(t) xjk
that the receive filter is set to some constant vjk (t) =
E{vjkH
(t)hjjk (t)} = E{kvjk (t)k2 } vjk . If a CLO is used, we have E{|vjk H
Dj (t) hjlm |2 } =
H
E{|vjk (t)hjlm (t)|2 } = jlm E{kvjk (t)k2 } 2
E{|vjk hjlm | }, which is independent of the phase-drifts since
H
+ N 2jjk 2jlm xHjk D(t) 1 1 H all elements of vjk are rotated in the same way. In con-
j Xlm j D(t) xjk + N (N 1)
( trast, each component of vjk is rotated in an independent
2jjk 2jlm xHjk D(t) 1 1 H
j Xlm j D(t) xjk if a CLO random manner with SLOs, which reduces the average in-

2jjk 2jlm |xHjk D(t) 1
j D(t) xlm |
H 2
if SLOs terference power since the components of the inner product
H
L X
K vjk Dj (t) hjlm add up incoherently. Consequently, the re-
X
H
E{|vjk (t) j (t)|2 } = 2 E{kvjk (t)k2 } plm jlm ceived interference power is reduced by SLOs while it remains
l=1 m=1 the same with a CLO.
L X
X K To summarize, we expect SLOs to provide larger UE rates
+2 plm N 2jjk 2jlm xHjk D(t) 1 1 H
j Xlm j D(t) xjk .
than a CLO, because the interference reduces with t when
l=1 m=1 the phase-drifts are independent, at the expense of increasing
As seen from this corollary, most terms scale linearly with the deployment cost by having N LOs. This observation is
N but there are a few terms that scale as N 2 . The latter terms validated by simulations in Section V.
dominate in the asymptotic analysis below.
The difference between having a CLO and SLOs C. Asymptotic Analysis and Hardware Scaling Laws
only manifests itself in the second-order moments The closed-form expressions in Theorem 2 and Corollary
E{|vjk H
(t)hjlm (t)|2 }. Hence, the desired signal quality 2 can be applied to cellular networks of arbitrary (finite)
is the same in both cases, while the interference terms are dimensions. In massive MIMO, the asymptotic behavior of
different;
PL PKthe case with the smallest interference variance large antenna arrays is of particular interest. In this section, we
2
jlm (t)| } gives the largest rate
H
l=1 p
m=1 lm E{|v jk (t)h assume that the N receive antennas in each cell are distributed
for UE k in cell j. These second-order moments depend over A 1 spatially separated subarrays, where each subarray
on the pilot sequences, channel covariance matrices, and contains N A antennas. This assumption is made for analytic
phase-drifts. By looking at (23) in Theorem 2 (or the tractability, but also makes sense in many practical scenarios.
corresponding expression in Corollary 2), we see that the Each subarray is assumed to have an inter-antenna distance
only difference is that two occurrences of X`m in the case much smaller than the propagation distances to the UEs, such
(a)
of a CLO are replaced by DH(t) x`m xH`m D(t) in the case of that jlk is the average channel attenuation to all antennas in
SLOs. These terms are equal when there are no phase-drifts subarray a in cell j from UE k in cell l. Hence, the channel
(i.e., = 0), while the difference grows larger with . In covariance matrix jlk CN N can be factorized as
particular, the term X`m is unaffected by the time index t,
(1) (A)
while the corresponding terms for SLOs decay as et (from jlk = diag jlk , . . . , jlk I N .
A
(25)
D(t) ). The following example provides the intuition behind | {z }
this result. (A)
,jlk CAA
By letting the number of antennas in each subarray grow large, of second-order channel statistics such as different path-losses
we obtain the following property. and spatial correlation [32]).
Corollary 3: If the MRC receive filter is used and the Apparently, the detrimental impact of hardware imperfec-
channel covariance matrices can be factorized as in (25), then tions vanishes almost completely as N grows large. This result
pjk Sigjk holds for any fixed values of the parameters , , and . In
SINRjk (t) = L P
K
(26) fact, the hardware imperfections may even vanish when the
1

hardware quality is gradually decreased with N . The next
P
plm Intjklm pjk Sigjk +O N
l=1m=1 corollary formulates analytically such an important hardware
where the signal part is scaling law.
2 Corollary 4: Suppose the hardware imperfection parameters
(A) e 1

(A)
Sigjk = tr xHjk D(t) jjk j D(t) xjk jjk are replaced as 2 7 20 N z1 , 7 0 N z2 , and 7 0 (1 +
(27) loge (N z3 )), for some given scaling exponents z1 , z2 , z3 0
the interference terms with a CLO are and some initial values 0 , 0 , 0 0. Moreover, let all pilot
A X
A symbols be non-zero: xjk (b ) > 0 for all j, k, and b. Then, all
(a ) (a ) (a ) (a )
X
IntCLO the SINRs, SINRjk (t), under MRC receive filtering converge

jklm = jjk1 jlm
1
jjk2 jlm
2
xHjk D(t) eHa1
a1 =1a2 =1 to non-zero limits as N if

e 1 Xlm ea eH
1 H
max(z1 , z2 ) 1 and z3 = 0
j 1 a2
e
j D (t) x jk ea 2
(28) 2 for a CLO
0 |t |
max(z1 , z2 ) + z3min 2 12 for SLOs.
and the interference terms with SLOs are {1 ,...,B }
2 (30)
(A) e 1

(A)
IntSLOs
jklm = tr x H
jk D (t) jjk j D H
(t) xlm jlm . Proof: The proof is given in Appendix F.
(29) This corollary proves that we can tolerate stronger hardware
imperfections as the number of antennas increases. This is
In these expresssions e j , PL PK X`m (A) +IAB , a very important result for practical deployments, because
`=1 m=1 j`m
ea is the ath column of IA , and the big O notation O( N1 ) we can relax the design constraints on the hardware quality
denotes terms that go to zero as N1 or faster when N . as N increases. In particular, we can achieve better energy
Proof: The proof is given in Appendix E. efficiency in the circuits and/or lower hardware costs by
This corollary shows that the distortion noise and receiver accepting larger distortions than conventionally. This property
noise vanish as N . The phase-drifts remain, but has been conjectured in overview articles, such as [7], and
have no dramatic impact since these affect the numerator was proved in [17] using a simplified system model with only
and denominator of the asymptotic SINR in (26) in similar additive distortion noise. Corollary 4 shows explicitly that the
ways. The simulations in Section V show that the phase-drift conjecture is also true for multiplicative phase-drifts, receiver
degradations are not exacerbated in massive MIMO systems noise, and inter-carrier interference. Going a step further,
with SLOs, while the performance with a CLO improves with Section IV exemplifies how the scaling law may impact the
N but at a slower pace due to the phase-drifts. circuit design in practical deployments.
The asymptotic SINRs are finite because both the signal Since Corollary 4 is derived for MRC filtering, (30) provides
power and parts of the inter-cell and intra-cell interference a sufficient scaling condition for any receive filter that performs
grow quadratically with N . This interference scaling behavior better than MRC. The scaling law for SLOs consists of
0 |t |
is due to so-called pilot contamination (PC) [9], [40], which two terms: max(z1 , z2 ) and z3 min 2 . The first
{1 ,...,B }
represents the fact that a BS cannot fully separate signals from
term max(z1 , z2 ) shows that the additive distortion noise and
UEs that interfered to each other during pilot transmission.4
receiver noise can be increased simultaneously and indepen-
Intra-cell PC is, conventionally, avoided by making the pilot
dently (as fast as N ), while the sum of the two terms
sequences orthogonal in space; for example, by using the
DFT pilot matrix X e spatial in Example 1. Unfortunately, the manifests a tradeoff between allowing hardware imperfections
j that cause additive and multiplicative distortions. The scaling
phase-drifts break any spatial pilot orthogonality. Hence, it is
law for a CLO allows only for increasing the additive distor-
reasonable to remove intra-cell PC by assigning temporally
orthogonal sequences, such as X e temporal in Example 1. Note tion noise and receiver noise, while the phase-drift variance
j should not be increased because only the signal gain (and not
that with temporal orthogonality the total pilot energy per UE,
1 the interference) is reduced by phase-drifts in this case; see
kxjk k2 , is reduced by K since the energy per pilot symbol is
Example 2. Clearly, the system is particularly vulnerable to
constrained. Consequently, the simulations in Section V reveal
phase-drifts due to their accumulation and since they affect
that temporally orthogonal pilot sequences are only beneficial
the signal itself; even in the case of SLOs, the second term
for extremely large arrays. Inter-cell PC cannot generally be
of (30) increases with T and the variance can scale only
removed, because there are only B T orthogonal sequences
logarithmically with N . Note that we can accept larger phase-
in the whole network, but it can be mitigated by allocating
drift variances if the coherence block T is small and the pilot
the same pilot to UEs that are well separated (e.g., in terms
symbols are distributed over the coherence block, which is in
4 Pilot contamination can be mitigated through semi-blind channel estima- line with the results in [15].
tion as proposed in [41], but the UE rates will still be limited by hardware
imperfections [17]. IV. U TILIZING THE S CALING L AW:
C IRCUIT-AWARE D ESIGN characterized by the figure-of-merit (FoM) expression

The generic system model with hardware imperfections in G
FoMLNA = (31)
(3) describes a flat-fading multi-cell channel. This channel can (F 1)PLNA
describe either single-carrier transmission over the full avail- where F 1 is the noise amplification factor, G is the
able flat-fading bandwidth as in [22] or one of the subcarriers amplifier gain, and PLNA is the power dissipation in the
in a system based on multi-carrier modulation; for example, LNA. Using this notation, the LNA contributes to the receiver
OFDM or FBMC as in [18], [28]. To some extent, it can also noise variance with F 2 . For optimized LNAs, FoMLNA is
describe single-carrier transmission over frequency-selective a constant determined by the circuit architecture [43]; thus,
channels as in [15]. The mapping between the imperfections in FoMLNA basically scales with the hardware cost. The scaling
a certain circuit in the receiving array to the three categories of law in Corollary 4 allows us to increase as N z2 for z2 12 .
distortions (defined in Section II) depends on the modulation The noise figure, defined as 10 log10 (F ), can thus be increased
scheme. For example, the multiplicative distortions caused by by z2 10 log10 (N ) dB. For example, at z2 = 21 we can allow
phase-noise leads also to inter-carrier interference in OFDM an increase by 10 dB if we deploy 100 antennas instead of
which is an additive noise-like distortion. one.
In this section, we exemplify what the scaling law in For a given circuit architecture, the invariance of the
Corollary 4 means for the circuits depicted in Fig. 1. In FoMLNA in (31) implies that we can decrease the power
particular, we show that the scaling law can be utilized dissipation (roughly) proportional to 1/N z2 . Hence, we can
for circuit-aware system design, where the cost and power make the total power dissipation of the N LNAs, N PLNA ,
dissipation per circuit will be gradually decreased to achieve increase as N 1z2 instead of N by tolerating higher noise
a sub-linear cost/power scaling with the number of antennas.
amplification. The scaling can thus be made as small as N .
For clarity of presentation, we concentrate on single-carrier
transmission over flat-fading channels, but mention briefly if
the interpretation might change for multi-carrier modulation. C. Local Oscillator (LO)
Phase noise in the LOs is the main source of multiplicative
phase-drifts and changes the phases gradually at each channel
A. Analog-to-Digital Converter (ADC) use. The average amount of phase-drifts that occurs under a
The ADC quantizes the received signal to a b bit resolution. coherence block is T and depends on the phase-drift variance
Suppose the received signal power is Psignal and that auto- and the block length T . If the LOs are free-running, the phase
matic gain control is used to achieve maximum quantization noise is commonly modeled by the Wiener process (random
accuracy irrespective of the received signal power. In terms of walk) defined in Section II [15], [21][23], [44] and the phase
the originally received signal power Psignal , the quantization noise variance is given by
in single-carrier transmission can be modeled as reducing the
= 4 2 fc2 Ts (32)
signal power to (1 22b )Psignal and adding uncorrelated
quantization noise with power 22b Psignal [19, Eq. (17)]. This where fc is the carrier frequency, Ts is the symbol time, and
model is particularly accurate for high ADC resolutions. We is a constant that characterizes the quality of the LO [21]. If
can include the quantization noise in the channel model (3) and/or T are small, such that T 0, the channel variations
by normalizing the useful signal. The quantization noise is dominate over the phase noise. However, phase noise can
included in the additive distortion noise j (t) and contributes play an important role when modeling channels with large
22b
to 2 with 12 2b , while the receiver noise variance is scaled coherence time (e.g., fixed indoor users, line-of-sight, etc.) and
by a factor 1212b due to the normalization. The scaling law as the carrier frequency increases (since = O(fc2 ) while the
in Corollary 4 allows us to increase the variance 2 as N z1 for Doppler spread reduces T as O(fc1 ) [22]. Relevant examples
z1 21 . This corresponds to reducing the ADC resolution by are mobile broadband access to homes and WiFi at millimeter
around z21 log2 (N ) bits, which reduces cost and complexity. frequencies.
For example, we can reduce the ADC resolution per antenna The power dissipation PLO of the LO is coupled to ,
by 2 bits if we deploy 256 antennas instead of one. For very such that PLO FoMLO where the FoM value FoMLO
large arrays, it is even sufficient to use 1-bit ADCs (cf. [42]). depends on the circuit architecture [21], [45] and naturally
The power dissipation of an ADC, PADC , is proportional to on the hardware cost. For a given architecture, we can allow
22b [19, Eq. (14)] and can, thus, be decreased approximately larger and, thereby, decrease the power PLO . The scaling
as 1/N z1 . If each antenna has a separate ADC, the total law in Corollary 4 allows us to increase as (1 + loge (N z3 ))
power N PADC increases with N but proportionally to N 1z1 , when using SLOs. The power dissipation per LO can then be
1
for z1 12 , instead of N , due to the gradually lower ADC
reduced as 1+z3 log
e (N )
. This reduction is only logarithmic

resolution. The scaling can thus be made as small as N . in N , which stands in contrast to the 1/ N scalings for
ADCs and LNAs (achieved by z1 = z2 = 12 ). Since linear
increase is much faster than logarithmic decay, the total power
B. Low Noise Amplifier (LNA) N PLO with SLOs increases almost linearly with N ; thus, the
The LNA is an analog circuit that amplifies the received benefit is mostly cost and design related. In contrast, the phase
signal. It is shown in [43] that the behavior of an LNA is noise variance cannot be scaled when having a CLO, because
massive MIMO only relaxes the design of circuits that are minimum distance of 25 m from any array location). We thus
placed independently at each antenna branch. have K = 8 and use B = 8 as pilot length in this section.
Imperfections in the LOs also cause inter-carrier interfer- Each sector is allocated an orthogonal pilot sequence, while
ence in OFDM systems, since the subcarrier orthogonality is the same pilot is reused in the same sector of all other cells.
broken [18]. When inter-carrier interference is created at the The channel attenuations are modeled as [47]
receiver side it depends on the channels of other subcarriers. (n)
(n) 10sjlk 1.53
It is thus uncorrelated with the useful channel in (3) and jlk = (33)
(n)
can be included in the receiver noise term. Irrespective of (djlk )3.76
the type of LOs, the severity of inter-carrier interference is (n)
where djlk is the distance in meters between receive antenna n
suppressed by z2 10 log10 (N ) dB according to Corollary 4. (n)
Hence, massive MIMO is less vulnerable to in-band distortions in cell j and UE k in cell l and sjlk N (0, 3.16) is shadow-
than conventional systems. fading (it is the same for co-located antennas but independent
The phase-noise variance formula in (32) gives other possi- between the 4 distributed arrays). We consider statistical power
bilities than decreasing the circuit power. In particular, one can control with pjk = 1 PN (n) to achieve an average received
N n=1 jjk
increase the carrier frequency fc with N by using Corollary signal power of over the receive antennas. The thermal noise
4. This is an interesting observation since massive MIMO variance is 2 = 174 dBm/Hz. We consider average SNRs,
has been identified as a key enabler for millimeter-wave / 2 , of 5 and 15 dB, leading to reasonable transmit powers
communications [6], in which the phase noise is more severe (below 200 mW over a 10 MHz bandwidth) for UEs at cell
since the variance in (32) increases quadratically with the edges.
carrier frequency fc . Fortunately, massive MIMO with SLOs
has an inherent resilience to phase noise. A. Comparison of Deployment Scenarios
We first compare the co-located and distributed deployments
D. Non-Linearities in Fig. 3. We consider the MRC filter, set the coherence block
to T = 500 channel uses (e.g., 5 ms coherence time and 100
Although the physical propagation channel is linear, prac- kHz coherence bandwidth), use the DFT-based pilot sequences
tical systems can exhibit non-linear behavior due to a variety of length B = 8, and send these in the beginning of the
of reasons; for examples, non-linearities in filters, converters, coherence block. The results are averaged over different UE
mixers, and amplifiers [18] as well as passive intermodulation locations.
caused by various electro-thermal phenomena [46]. Such non- The average achievable rates per UE are shown in Fig. 4
linearities are often modeled by power series or Volterra series for / 2 = 5 dB, using either ideal hardware or imperfect
[46], but since we consider a system with Gaussian transmit hardware with = 0.0156, = 1.58 2 , and = 1.58 104 .
signals the Bussgang theorem can be applied to simplify the These parameter values were not chosen arbitrarily, but based
characterization [4], [24]. For a Gaussian variable X and any on the circuit examples
non-linear function g(), the Bussgang theorem implies that in Section IV. More specifically, we
obtained = 2b / 1 22b = 0.0156 by using b = 6 bit
g(X) = cX + V , where c is a scaling factor and V is ADCs and = 12 F 2 2
2b = 1.58 for a noise amplification
a distortion uncorrelated with X; see [24, Eq. (15)]. If we
factor of F = 2 dB. The phase noise variance = 1.58 104
let g(X) describe a nonlinear component and let X be the
was obtained from (32) by setting fc = 2 GHz, Ts = 107 s,
useful signal, the impact of non-linearities can be modeled
and = 1017 . Note that the curves in Fig. 4 are based on
by a scaling of the useful signal and an additional distortion
the analytic results in Theorem 2, while the marker symbols
term. Depending on the nature of each non-linearity, the
correspond to Monte Carlo simulations of the expectations in
corresponding distortion is either included in the distortion
(20). The perfect match validates the analytic results.
noise or the receiver noise.5 The scaling factor c of the useful
Looking at Fig. 4, we see that the tractable ergodic rate
signal is removed by scaling 2 and by |c|12 .
from Lemma 1 approaches well the slightly higher achievable
rate from [12, Eq. (39)]. Moreover, we see that the hardware
V. N UMERICAL I LLUSTRATIONS imperfections cause small rate losses when the number of
Our analytic results are corroborated in this section by antennas, N , is small. However, the large-N behavior depends
studying the uplink in a cell surrounded by 24 interfering cells, strongly on the oscillators: the rate loss is small for SLOs at
as shown in Fig. 3. Each cell is a square of 250 m 250 m any N , while it can be very large if a CLO is used when N is
and we compare two topologies: (a) co-located deployment large (e.g., 25% rate loss at N = 400). This important property
of N antennas in the middle of the cell; and (b) distributed was explained in Example 2 and the simple explanation is that
deployment of 4 subarrays of N4 antennas at distances of the effect of phase noise averages out with SLOs, but at the
62.5 m from the cell center. To mimic a simple user scheduling cost of adding more hardware.
algorithm, each cell is divided into 8 virtual sectors and one Fig. 4 also shows that the distributed massive MIMO
UE is picked with a uniform distribution in each sector (with a deployment achieves roughly twice the rates of co-located
massive MIMO. This is because distributed arrays can ex-
5 The distortion from non-linearities are generally non-Gaussian, but this
ploit both the proximity gains (normally achieved by small
has no impact on our analysis because the achievable rates in Lemma 1 were
obtained by making the worst-case assumption of all additive distortions being cells) and the array gains and spatial resolution of coherent
Gaussian distributed. processing over many antennas.
Pilot 2 Pilot 3 Pilot 4 Pilot 2 Pilot 3

Pilot 1 Pilot 1 Pilot 4
Cell
under
study
250
meters
Pilot 8 Pilot 7 Pilot 6 Pilot 5 Pilot 8 Pilot 7 Pilot 6 Pilot 5
(a) (b)
Fig. 3. The simulations consider the uplink of a cell surrounded by two tiers of interfering cells. Each cell contains K = 8 UEs that are uniformly distributed
in different parts of the cell. Two site deployments are considered: (a) N co-located antennas in the middle of the cell; and (b) N/4 antennas at 4 distributed
arrays.
5 consider the distributed deployment in Fig. 3, while a similar

Average Rate per UE [bit/channel use]
figure for the co-located deployment is available in [25]. Fig. 5

4 shows the UE rates as a function of the number of antennas,
for ideal hardware and the same hardware imperfections as in
Distributed
3 deployment the previous figure. The simulation validates the convergence
to the limits derived in Corollary 3, but also shows that the
2
convergence is very slowwe used logarithmic scale on the
horizontal axis because N = 106 antennas are required for
Ideal Hardware, (39) in [12] convergence for ideal hardware and for hardware imperfec-
1 Ideal Hardware, Lemma 1
Co-located NonIdeal Hardware: SLOs tions with SLOs, while N = 104 antennas are required for
deployment NonIdeal Hardware: CLO hardware imperfections with a CLO. The performance loss for
0
0 50 100 150 200 250 300 350 400 hardware imperfections with SLOs is almost negligible, while
Number of Receive Antennas per Cell the loss when having a CLO grows with N and approaches
Fig. 4. Achievable rates with MRC filter and either ideal hardware or
50 %.
imperfections given by (, , ) = (0.0156, 1.58 2 , 1.58 104 ). Co- Two types of pilot sequences are also compared in Fig. 5:
located and distributed antenna deployments are compared, as well as, a CLO the temporally orthogonal pilots in (6) and the spatially
and SLOs.
orthogonal DFT-based pilots in (7). As discussed in relation
10
to Corollary 3, temporal orthogonality provides slightly higher
rates in the asymptotic regime (since the phase noise cannot

break the temporal pilot orthogonality). However, this gain is
8
barely visible in Fig. 5 and only kicks in at impractically large
Asymptotic limits Temporal is slightly better N . Since temporally orthogonal pilots use K times less pilot
6 than spatial orthogonality energy, they are the best choice in this simulation. However, if
the average SNR is decreased then spatially orthogonal pilots
4 can be used to improve the estimation accuracy.
2 Ideal Hardware C. Impact of Coherence Block Length

NonIdeal Hardware: SLOs
NonIdeal Hardware: CLO Next, we illustrate how the length of the coherence block, T ,
0 affects the UE rates with the MRC filter. We consider a practi-
1 2 3 4 5 6
10 10 10 10 10 10
Number of Receive Antennas per Cell cal number of antennas, N = 240, while having / 2 = 5 dB
and imperfections with {, , } = {0.0156, 1.58 2 , 1.58
Fig. 5. Average UE rate with MRC filter for different numbers of antennas,
different hardware imperfections, and spatially or temporally orthogonal
104 }, as before. The UE rates are shown in Fig. 6 as a
pilots. Note the logarithmic horizontal scale which is used to demonstrate function of T . We compare two ways of distributing the pilot
the asymptotic behavior. sequences over the coherence block: in the beginning or in the
middle (see (a) and (b) in Fig. 2).
With ideal hardware, the pilot distribution has no impact.
B. Validation of Asymptotic Behavior We observe in Fig. 6 that the average UE rates are slightly
Next, we illustrate the asymptotic behavior of the UE rates increasing with T . This is because the pre-log penalty of using
(with MRC filter) as N . For the sake of space, we only only T B out of T channel uses for data transmission is
4 4.5

4 z1 = z2 = z3 = 0
3.5 Pilot sequences (Fixed imperfections)
in the middle 3.5
3
3
2.5
Pilot sequences in the beginning 2.5
z1 = z2 (= z3)= 0.48 z1 = z2 (= z3)= 0.96
2
2 (Satisfies scaling laws) (Faster than scaling laws)
1.5
1.5
1
Ideal Hardware 1 Curves bend Ideal Hardware
0.5 NonIdeal Hardware: SLOs toward zero NonIdeal Hardware: SLOs
0.5
NonIdeal Hardware: CLO NonIdeal Hardware: CLO
0 0
0 500 1000 1500 2000 0 50 100 150 200 250 300 350 400
Coherence Block [channel uses] Number of Receive Antennas per Cell
Fig. 6. Average UE rate with MRC filter as a function of the coherence block Fig. 7. Average UE rate with MRC filter for different numbers of antennas,
length, for different pilot sequence distributions. The maximum is marked at N , and with hardware imperfections that are either fixed or increase with N
each curve and is the preferable operating point for the transmission protocol. as in Corollary 4.

smaller when T is large (and B is fixed). In the case with
hardware imperfections, Fig. 6 shows that the rates increase 6 z1 = z2 = z3 = 0
(Fixed imperfections)
with T for small T (for the same reason as above) and then 5
decrease with T since phase-drifts accumulate over time.
Interestingly, slightly higher rates are achieved and larger 4
coherence blocks can be handled if the pilot sequences are sent Ideal Hardware
3 z1 = z2 (= z3)= 0.48 NonIdeal Hardware: SLOs
in the middle of the coherence block (instead of the beginning)
(Satisfies scaling laws) NonIdeal Hardware: CLO
since the phase drifts only accumulate half as much. From an 2
implementation perspective it is, however, better to put pilot z1 = z2 (= z3)= 0.96
1 Curves bend
sequences in the beginning, since then there is no need to toward zero (Faster than scaling laws)
buffer the incoming signals while waiting for the pilots that 0
0 50 100 150 200 250 300 350 400
enable computation of receive filters. Number of Receive Antennas per Cell
Fig. 6 shows, once again, that systems with SLOs have
higher robustness to phase-drifts than systems with a CLO. Fig. 8. Average UE rate with MMSE filter (in (34)) for different numbers of
antennas, N , and with hardware imperfections that are either fixed or increase
To make a fair comparison, we need to consider that the with N as in Corollary 4.
coherence block is a modeling conceptwe can always choose
a transmission protocol with a smaller T than prescribed
by the coherence block length, at the cost of increasing As expected, the z-combinations that satisfy the scaling laws
the pilot overhead B/T . Hence, it is the maximum at each give small performance losses, while the bottom curves go
curve, indicated by markers, which is the operating point to asymptotically to zero since the law is not fulfilled (the points
compare. The difference between SLOs and a CLO is much where the curves bend downwards are marked).
smaller when comparing the maxima, but these are achieved at We have considered the MRC filter since its low computa-
very different T -values; the transmission protocol should send tional complexity is attractive for massive MIMO topologies.
pilots much more often when having a CLO. The true optimum MRC provides a performance baseline for other receive filters
is achieved by maximizing over T and B, and probably by which typically have higher complexity. In Fig. 8 we consider
spreading the pilots to reduce the accumulation of phase drifts. the (approximate) MMSE receive filter
MMSE
vjk (t)
D. Hardware Scaling Laws with Different Receive Filters !1
L X
K
Next, we illustrate the scaling laws for hardware imper- ,
X
2
plm Gjlm (t)+ DGjlm (t) +IM

hjjk (t)
fections established in Corollary 4 and set / 2 = 15 dB to l=1 m=1
emphasize this effect. We focus on the distributed scenario (34)
for T = 500, since the co-located scenario behaves similarly
and can be found in [26]. Using the notation from the where Gjlm (t) , hjlm (t)hHjlm (t) + Cjlm (t) and DGjlm (t)
scaling law, we set the baseline hardware imperfections to is a diagonal matrix where the diagonal elements are the
(0 , 0 , 0 ) = (0.05, 3 2 , 7 105 ) and increase these with N same as in Gjlm (t). By comparing Figs. 7 and 8 (notice the
using different values on the exponents z1 , z2 , and z3 . different scales), we observe that the MMSE filter provides
The UE rates with MRC filters are given in Fig. 7 for ideal higher rates than the MRC filter. Interestingly, the losses due
hardware, fixed hardware imperfections, and imperfections to hardware imperfections behave similarly but are larger for
that either increase according to the scaling law or faster than MMSE. This is because the MMSE filter exploits spatial
the law (observe that we always have z3 = 0 for a CLO). interference suppression which is sensitive to imperfections.
VI. C ONCLUSION and interpretations are, however, outside the scope of this
paper.
Massive MIMO technology can theoretically improve the
spectral and energy efficiencies by orders of magnitude, but to
A PPENDIX A: A U SEFUL L EMMA
make it a commercially viable solution it is important that the
N antenna branches can be manufactured using low-cost and Lemma 2: Let u CN (0, ) and consider some determin-
low-power components. As exemplified in Section IV, such istic matrix M. It holds that
components are prone to hardware imperfections that distort
E |uH Mu|2 = |tr(M)|2 + tr(MMH ).

(35)
the communication and limit the achievable performance.
In this paper, we have analyzed the impact of such hardware Proof: This lemma follows from straightforward com-
imperfections at the BSs by studying an uplink communica- putation, by exploiting that uH Mu = uH 1/2 M1/2 u =
1/2
M1/2 ]i,j uj where u CN (0, I).
P
tion model with multiplicative phase-drifts, additive distortion i,j ui [
noise, noise amplifications, and inter-carrier interference. The
system model can be applied to both co-located and distributed A PPENDIX B: P ROOF OF T HEOREM 1
antenna arrays. We derived a new LMMSE channel estima-
tor/predictor and the corresponding achievable UE rates with We exploit the fact that hjlk (t) =
H H
1
MRC. Based on these closed-form results, we prove that only E{hjlk (t) j } E{ j j } j is the general expression
the phase-drifts limit the achievable rates as N . This of an LMMSE estimator [33, Ch. 12]. Since the additive
showcases that massive MIMO systems are robust to hardware distortion and receiver noises are uncorrelated with hjlk (t)
imperfections, which is a property that has been conjectured and the UEs channels are independent, we have that
in prior works (but only proved for simple models with one
E{hjlk (t) Hj }
type of imperfection). This phenomenon can be attributed to
the fact that distortions are uncorrelated with the useful signals = E{Dj (t) hjlk hHjlk [DHj (1 ) xlk (1 ) . . . DHj (B ) xlk (B )]}
and, thus, add non-coherently during the receive processing. = jlk E{[Dj (t) DHj (1 ) xlk (1 ) . . . Dj (t) DHj (B ) xlk (B )]}
Particularly, we established a scaling law showing that the
variance of the distortion noise and receiver noise can increase = jlk [xlk (1 )e 2 |t1 | IN . . . xlk (B )e 2 |tB | IN ]

simultaneously as N . If the phase-drifts are independent = xHlk D(t) jlk (36)
between the antennas, we can also tolerate an increase of the 2 |t1 t2 |
phase-drift variance with N , but only logarithmically. If the since E{en,t1 en,t2 } = e and by exploiting the
phase-drifts are the same over the antennas (e.g., if a CLO fact that diagonal matrices commute. Furthermore, we have
is used), then the phase-drift variance cannot increase. The that
numerical results show that there are substantial performance E{ j Hj }
benefits of using separate oscillators at each antenna branch K
L X n
instead of a common oscillator. The difference in performance
X
= E [DHj (1 ) x`m (1 ) . . . DHj (B ) x`m (B )]H hj`m
might be smaller if the LMMSE estimator is replaced by a `=1 m=1
Kalman filter that exploits the exact distribution of the phase- o
drifts [16], [17]. Interestingly, the benefit of having SLOs hHj`m [DHj (1 ) x`m (1 ) . . . DHj (B ) x`m (B )]
remains also under idealized uplink conditions (e.g., perfect + E{[ Hj (1 ) . . . Hj (B )]H [ Hj (1 ) . . . Hj (B )]}
CSI, no interference, and high SNR [48]). In any case, the + E{[ Hj (1 ) . . . Hj (B )]H [ Hj (1 ) . . . Hj (B )]}
transmission protocol must be adapted to how fast the phase-
L X K
drifts deteriorate performance. The scaling law was derived X
= X`m j`m + 2 D|x`m |2 j`m + IBN .
for MRC but provides a sufficient condition for other judicious
`=1 m=1
receive filters, like the MMSE filter. We also exemplified what | {z }
the scaling law means for different circuits in the receiver ,j
(e.g., ADCs, LNAs, and LOs). This quantifies how fast the (37)
requirements on the number of quantization bits and the The LMMSE estimator in (9) now follows from (36) and (37).
noise amplification can be relaxed with N . It also shows The error covariance matrix in (15) is computed as jlk
that a circuit-aware design can make the total circuit power
1
E{hjlk (t) Hj } E{ j Hj } (E{hjlk (t) Hj })H [33, Ch. 12].
consumption of the N ADCs and LNAs increase as N ,
instead of N which it would conventionally be the case.
A natural extension to this paper would be to consider A PPENDIX C: P ROOF OF L EMMA 1
also the downlink with hardware imperfections at the BSs. Since the effective channels vary with t, we follow the
If maximum ratio transmission (MRT) is used for precoding, approach in [15] and compute one ergodic achievable rate
then more-or-less the same expectations as in the uplink for each t D. We obtain (19) by taking the average of
SINRs in (20) will show up in the downlink SINRs but at these rates. The SINR in (20) is obtained by treating the
different places [49]. Hence, we believe that similar closed- uncorrelated inter-user interference and distortion noise as
form rate expressions and scaling laws for the levels of independent Gaussian noise, which is a worst-case assumption
hardware imperfections can be derived. The analytic details when computing the mutual information [39]. In addition, we
follow an approach from [38] and only exploit the knowl- computed by expanding the expression as
H
edge of the average effective channel E{vjk (t)hjjk (t)} in n
the detection, while the deviation from the average effective E{|tr(jlm Mjklm (t))|2 } = E tr Bjklm (t)

channel is treated as worst-case Gaussian noise with variance 2 o
H
(t)hjjk (t)|2 } |E{vjk
H
(t)hjjk (t)}|2 . [DTj (1 ) xlm (1 ) . . . DTj (B ) xlm (B )]T DHj (t)

E{|vjk
N X
nX B

A PPENDIX D: P ROOF OF T HEOREM 2 =E [Bjklm (t)Eb1 ]n1 n1 xlm (b1 )e n1 ,b1 en1 ,t
n1 =1 b1 =1
The expressions in Theorem 2 are derived one at the time. N B o
n2 ,b
X X
For brevity, we use the following notations in the derivations: [EHb2 BHjklm (t)]n2 n2 xlm (b2 )e 2 en2 ,t
n2 =1 b2 =1
Ajlk (t) = xHlk D(t) jlk 1

j (38) X
= [Bjklm (t)Eb1 ]n1 n1 [EHb2 BHjklm (t)]n2 n2
(1) (N )
D|hjlk |2 = diag |hjlk |2 , . . . , |hjlk |2 (39) n1 ,n2 ,b1 ,b2
n o
( + )
xlm (b1 )xlm (b2 )E e n1 ,b1 n1 ,t n2 ,b2 n2 ,t
Bjklm (t) = jlm Ajjk (t) (40)
H
(46)
Mjklm (t) = Dj (t) Ajjk (t)
[DTj (1 ) xlm (1 ) . . . DTj (B ) xlm (B )]T . (41) where Eb = eb IN and eb CB1 is the bth column of IB .
The phase-drift expectation depends on the use of a CLO or
We begin with (21) and exploit that vjk (t) = hjlk (t) is an SLOs:
LMMSE estimate to see that n
( + )
o
E e n1 ,b1 n1 ,t n2 ,b2 n2 ,t
E{kvjk (t)k2 } = tr(jjk Cjjk (t))
|b b |
(42) e 2 1 2 ,
if a CLO, (47)
= tr xHjk D(t) jjk 1
H
j D x
(t) jk jjk
| |
= e 2 b 1 b 2 , if SLOs and n1 = n2 ,
2 |tb1 | 2 |tb2 |

e e , if SLOs and n1 6= n2 .
which proves (21). Next, we exploit that hjjk (t) = Ajjk (t) j
and note that
Since xlm (b1 )xlm (b2 )e 2 |b1 b2 | = [Xlm ]b1 ,b2 =
H
E{vjk (t)hjjk (t)} = tr E{hjjk (t) Hj }AHjjk (t) eHb1 Xlm eb2 in the case of a CLO, (46) becomes

= tr xHjk D(t) jjk AHjjk (t)
(43)
X
[Bjklm (t)Eb1 ]n1 n1 [EHb2 BHjklm (t)]n2 n2 eHb1 Xlm eb2
= tr xHjk D(t) jjk 1
H
j D x
(t) jk jjk n1 ,n2 ,b1 ,b2
X
where the second equality follows from (36) and the third = eHn1 Bjklm (t)(Xlm en1 eHn2 )BHjklm (t)en2
equality follows from the full expression of Ajjk (t) in n1 ,n2
(38). Observe that the expression (43) is the same as for (48)
E{kvjk (t)k2 } in (42).
where en CN 1 is the nth column of of IN (recall also the
Next, the second-order moment in (23) can be expanded as
definitions of Eb and eb above).

E |vjk (t)hjlm (t)|2 Next, we note that e 2 |tb | xlm (b ) = [DH(t) xlm ]b . In the
H
= E{tr(AHjjk (t)hjlm (t)hHjlm (t)Ajjk (t) j Hj )} case of SLOs, (46) then becomes
X
= tr AHjjk (t)jlm Ajjk (t) j Xlm jlm [Bjklm (t)Eb1 ]n1 n1 [EHb2 BHjklm (t)]n2 n2
n1 ,n2 ,b1 ,b2
+ E tr(MHjklm (t)hjlm hHjlm Mjklm (t)hjlm hHjlm ) n1 6=n2
+ 2 E tr AHjjk (t)hjlm hHjlm Ajjk (t)(D|xlm |2 D|hjlm |2 )

[DH(t) xlm ]b1 [xHlm D(t) ]b2
X
(44) + [Bjklm (t)Eb1 ]nn [EHb2 BHjklm (t)]nn [Xlm ]b1 ,b2
n,b1 ,b2
where the first term follows from computing separate ex- X 2
pectations for the parts of j Hj that are independent of =

[Bjklm (t)Eb ]nn [DH(t) xlm ]b

hjlm (t)hHjlm (t). The remaining two terms take care of the n,b
statistically dependent terms. The middle term is simplified as +
X
[Bjklm (t)Eb1 ]nn [EHb2 BHjklm (t)]nn
n,b1 ,b2

E tr(MHjklm (t)hjlm hHjlm Mjklm (t)hjlm hHjlm )
[Xlm DH(t) xlm xHlm D(t) ]b1 ,b2
= E |tr(jlm Mjklm (t))|2

(45) 2 X H
= tr(Ajjk (t)(DH xlm jlm )) + en Bjklm (t)
H

+ E tr(jlm Mjklm (t)jlm Mjklm (t)) (t)
n H
by computing the expectation with respect to hjlm using H H H
(Xlm D(t) xlm xlm D(t) ) en en Bjklm (t)en .
Lemma 2 in Appendix A. The first expectation in (45) is now (49)
The second expectation in (45) is computed along the same The expectation in the middle term of (52) is computed as
lines as in (37) and becomes
E{tr(MHjklm (t)D|hjlm |2 Mjklm (t)hjlm hHjlm )}
X
E |hHjlm MHjklm (t)en eHn hjlm |2

=
E{tr(jlm Mjklm (t)jlm MHjklm (t))} n
(50)
X
E |eHn jlm MHjklm (t)en |2

= tr jlm Ajjk (t)(Xlm jlm )AHjjk (t) . =
n
X (53)
+ E eHn jlm en eHn Mjklm (t)jlm MHjklm (t)en
n
It remains to compute the last term in (44). We exploit = tr jlm Ajjk (t)(Xlm jlm )AHjjk (t)

the following expansion of diagonal matrices: D|xlm |2 = X
PB 2
PN 2 + eHn jlm Ajjk (t)(Xlm en eHn )AHjjk (t)jlm en
b=1 |xlm (b )| eb eb and D|hjlm |2 = n=1 |en hjlm | en en ,
H H H
n
where eb is the bth column of IB and en is the nth column
of IN . Plugging this into the last term in (44) yields where the first equality follows from the same diagonal matrix
expansion as in (51), the second equality is due to Lemma 2
(and that diagonal matrices commute), and the third equality
X follows from computing the expectation with respect to phase-
|xlm (b )|2 E |hHjlm Ajjk (t)(eb en eHn )hjlm |2

drifts as in (37) and then reverting the matrix expansions
b,n
wherever possible.
xlm (b )|2 |tr(jlm Ajjk (t)(eb en eHn ))2
X
= Similarly, we have
b,n
X E{tr AHjjk (t)D|hjlm |2 Ajjk (t)(D|xlm |2 D|hjlm |2 ) }
+ |xlm (b )|2 tr jlm Ajjk (t)(eb en eHn ) X
|xlm (b )|2 E |hHjlm en1 eHn1 Ajjk (t)(eb en2 eHn2 )hjlm |2

b,n =
n1 ,n2 ,b

jlm (eHb en eHn )AHjjk (t)
X 2
= |xlm (b )|2 tr jlm en1 eHn1 Ajjk (t)(eb en2 eHn2 )
X
= eHn jlm Ajjk (t)(D|xlm |2 en eHn )AHjjk (t)jlm en
n1 ,n2 ,b
n
H

+ tr jlm Ajjk (t)(D|xlm |2 jlm )Ajjk (t) + tr jlm en1 eHn1 Ajjk (t)(eb eHb en2 eHn2 jlm )AHjjk (t)
(51) X
= eHn1 jlm Ajjk (t)(D|xlm |2 en1 eHn1 )AHjjk (t)jlm en1
n1

where the first equality follows from Lemma 2 and the + tr jlm Ajjk (t)(D|xlm |2 jlm )AHjjk (t)
second equality from reverting the matrix expansions wherever (54)
possible. Plugging (45)(51) into (44) and utilizing Xlm + where the first equality follows from the same diagonal matrix
2 D|xlm |2 = Xlm , we obtain (23) by removing the special expansions as above, the second equality follows from Lemma
notation that was introduced in the beginning of this appendix. 2 (and that diagonal matrices commute), and the third equality
Finally, we compute the expectation in (24) by noting that from reverting the matrix expansions wherever possible.
By plugging (53) and (54) into (52) and utilizing Xlm +
2 D|xlm |2 = Xlm , we finally obtain (24).
H
E{|vjk (t) j (t)|2 } = E{tr(AHjjk (t)j (t)Ajjk (t) j Hj )}
L X
K A PPENDIX E: P ROOF OF C OROLLARY 3
X
2
= plm This corollary is obtained by dividing all the terms in
2
l=1 m=1 SINRjk (t) by N A2 and inspecting the scaling behavior as

N . Using the expressions in Theorem 2 and utilizing
tr AHjjk (t)jlm Ajjk (t) j Xlm jlm
that 1 e 1 A2 2
j = j I N , we observe that N 2 E{kvjk (t)k } =
A
+ E{tr(MHjklm (t)D|hjlm |2 Mjklm (t)hjlm hHjlm )} A2
N 2 tr(F I N ) = A 1
N tr(F) = O( N ), where F = (xjk D(t)
H
A
(A) e 1 (A)
+ 2 E{tr AHjjk (t)D|hjlm |2 Ajjk (t)(D|xlm |2 D|hjlm |2 ) } . jjk ) j (D(t) xjk jjk ). Similarly, it is straightforward

A2 2 1
2 E{|vjk (t) j (t)| } = O( N ).
H
but lengthy to prove that N
(52)
The only terms in the SINR that remain as N are
A2 2 2 A2 2
N 2 (E{kvjk (t)k }) = Sigjk and N 2 E{|vjk (t)hjlm (t)| } =
H
1
Intjklm + O( N ).
The first equality follows by taking the expectation with
respect to j (t) for fixed channel realizations. The second
equality follows by taking A PPENDIX F: P ROOF OF C OROLLARY 4
PLseparate
PK expectations with respect
to the terms of j = 2 l=1 m=1 plm D|hjlm |2 and j Hj The first step of the proof is to substitute the new pa-
that are independent. These give the first term in (52) while the rameters into the SINR expression in (20) and scale all
last two terms take care of the statistically dependent terms. terms by 1/N 1+z3 0 min |t | . Since the distortion noise
and receiver noise terms normally behave as O(N ), it is [17] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, Massive
straightforward (but lengthy) to verify that the (scaled) dis- MIMO systems with non-ideal hardware: Energy efficiency, estimation,
and capacity limits, IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112
tortion noise and receiver noise terms go to zero when 7139, Nov. 2014.
N . Similarly, the signal term in the numerator which [18] T. Schenk, RF Imperfections in High-Rate Wireless Systems: Impact
normally behave as O(N 2 ), will after the scaling behave and Digital Compensation, Springer, 2008.
[19] A. Mezghani, N. Damak, and J. A. Nossek, Circuit aware design of
as O(N 12 max(z1 ,z2 )z3 0 min |t | ). In the case of SLOs, power-efficient short range communication systems, in Proc. IEEE
the second-order interference moments E{|vjk H
(t)hjlm (t)|2 } ISWCS, Sep. 2010, pp. 869873.
in the denominator exhibit the same scaling as the signal [20] M. Wenk, MIMO-OFDM Testbed: Challenges, Implementations, and
Measurement Results, Series in microelectronics. Hartung-Gorre, 2010.
term. The scaling law in (30) then follows from that we want [21] D. Petrovic, W. Rave, and G. Fettweis, Effects of phase noise on OFDM
the signal and interference terms to be non-vanishing in the systems with and without PLL: Characterization and compensation,
asymptotic limit; that is, 1 2 max(z1 , z2 ) z3 0 min |t IEEE Trans. Commun., vol. 55, no. 8, pp. 16071616, Aug. 2007.
[22] H. Mehrpouyan, A. A. Nasir, S. D. Blostein, T. Eriksson, G. K.
| > 1. In the case of a CLO, the second-order interference Karagiannidis, and T. Svensson, Joint estimation of channel and
moments behave as O(N 12 max(z1 ,z2 ) ) and do not depend on oscillator phase noise in MIMO systems, IEEE Trans. Signal Process.,
z3 . To make the signal and interference terms have the same vol. 60, no. 9, pp. 47904807, Sep. 2012.
scaling and be non-vanishing, we thus need to set z3 = 0 and [23] G. Durisi, A. Tarable, and T. Koch, On the multiplexing gain of MIMO
microwave backhaul links affected by phase noise, in Proc. IEEE ICC,
max(z1 , z2 ) 21 . June 2013, pp. 32093214.
[24] W. Zhang, A general framework for transmission with transceiver
distortion and some applications, IEEE Trans. Commun., vol. 60, no.
R EFERENCES 2, pp. 384399, Feb. 2012.
[25] E. Bjornson, M. Matthaiou, and M. Debbah, Massive MIMO systems
[1] M. K. Karakayali, G. J. Foschini, and R. A. Valenzuela, Network with hardware-constrained base stations, in Proc. IEEE ICASSP, May
coordination for spectrally efficient communications in cellular systems, 2014, pp. 31423146.
IEEE Wireless Commun. Mag., vol. 13, no. 4, pp. 5661, Aug. 2006. [26] E. Bjornson, M. Matthaiou, and M. Debbah, Circuit-aware design of
[2] D. Gesbert, S. Hanly, H. Huang, S. Shamai (Shitz), O. Simeone, and energy-efficient massive MIMO systems, in Proc. IEEE ISCCSP, May
W. Yu, Multi-cell MIMO cooperative networks: A new look at 2014, pp. 101104.
interference, IEEE J. Sel. Areas Commun., vol. 28, no. 9, pp. 1380 [27] X. Gao, O. Edfors, F. Rusek, and F. Tufvesson, Massive MIMO in real
1408, Dec. 2010. propagation environments, IEEE Trans. Wireless Commun., 2014, To
[3] E. Bjornson and E. Jorswieck, Optimal resource allocation in coordi- appear, Available: http://arxiv.org/abs/1403.3376.
nated multi-cell systems, Foundations and Trends in Communications [28] A. Farhang, N. Marchetti, L. E. Doyle, and B. Farhang-Boroujeny,
and Information Theory, vol. 9, no. 2-3, pp. 113381, 2013. Filter bank multicarrier for massive MIMO, in Proc. IEEE VTC Fall,
[4] H. Holma and A. Toskala, LTE for UMTS: Evolution to LTE-Advanced, 2014.
Wiley, 2nd edition, 2011. [29] H. Yang and T. L. Marzetta, Total energy efficiency of cellular large
[5] J. Hoydis, K. Hosseini, S. ten Brink, and M. Debbah, Making smart scale antenna system multiple access mobile networks, in Proc. IEEE
use of excess antennas: Massive MIMO, small cells, and TDD, Bell OnlineGreenComm, Oct. 2013, pp. 2732.
Labs Technical Journal, vol. 18, no. 2, pp. 521, Sep. 2013. [30] E. Dahlman, S. Parkvall, J. Skold, and P. Beming, 3G Evolution: HSPA
[6] R. Baldemair, E. Dahlman, G. Fodor, G. Mildh, S. Parkvall, Y. Selen, and LTE for Mobile Broadband, Academic Press, 2nd edition, 2008.
H. Tullberg, and K. Balachandran, Evolving wireless communications: [31] M. Biguesh and A. B. Gershman, Downlink channel estimation in
Addressing the challenges and expectations of the future, IEEE Veh. cellular systems with antenna arrays at base stations using channel
Technol. Mag., vol. 8, no. 1, pp. 2430, Mar. 2013. probing with feedback, EURASIP J. Appl. Signal Process., vol. 2004,
[7] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, Massive no. 9, pp. 13301339, 2004.
MIMO for next generation wireless systems, IEEE Commun. Mag., [32] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, A coordinated approach
vol. 52, no. 2, pp. 186195, Feb. 2014. to channel estimation in large-scale multiple-antenna systems, IEEE J.
[8] China Mobile Research Institute, C-RAN: The road towards green Sel. Areas Commun., vol. 31, no. 2, pp. 264273, Feb. 2013.
RAN, Tech. Rep., White Paper, Oct. 2011. [33] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
[9] T. L. Marzetta, Noncooperative cellular wireless with unlimited Theory, Prentice Hall, 1993.
numbers of base station antennas, IEEE Trans. Wireless Commun., [34] J. H. Kotecha and A. M. Sayeed, Transmit signal design for optimal
vol. 9, no. 11, pp. 35903600, Nov. 2010. estimation of correlated MIMO channels, IEEE Trans. Signal Process.,
[10] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, vol. 52, no. 2, pp. 546557, Feb. 2004.
O. Edfors, and F. Tufvesson, Scaling up MIMO: Opportunities and [35] E. Bjornson and B. Ottersten, A framework for training-based es-
challenges with very large arrays, IEEE Signal Process. Mag., vol. 30, timation in arbitrarily correlated Rician MIMO channels with Rician
no. 1, pp. 4060, Jan. 2013. disturbance, IEEE Trans. Signal Process., vol. 58, no. 3, pp. 1807
[11] J. Hoydis, S. ten Brink, and M. Debbah, Massive MIMO in the UL/DL 1820, Mar. 2010.
of cellular networks: How many antennas do we need?, IEEE J. Sel. [36] T. Yoo and A. Goldsmith, Capacity and power allocation for fading
Areas Commun., vol. 31, no. 2, pp. 160171, Feb. 2013. MIMO channels with channel estimation error, IEEE Trans. Inf. Theory,
[12] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, Energy and spectral vol. 52, no. 5, pp. 22032214, May 2006.
efficiency of very large multiuser MIMO systems, IEEE Trans. [37] G. Durisi, A. Tarable, C. Camarda, R. Devassy, and G. Montorsi,
Commun., vol. 61, no. 4, pp. 14361449, Apr. 2013. Capacity bounds for MIMO microwave backhaul links affected by
[13] E. Bjornson, L. Sanguinetti, J. Hoydis, and M. Debbah, Designing phase noise, IEEE Trans. Commun., vol. 62, no. 3, pp. 920929, Mar.
multi-user MIMO for energy efficiency: When is massive MIMO the 2014.
answer?, in Proc. IEEE WCNC, Apr. 2014, pp. 242247. [38] M. Medard, The effect upon channel capacity in wireless communica-
[14] S. K. Mohammed and E. G. Larsson, Per-antenna constant envelope tions of perfect and imperfect knowledge of the channel, IEEE Trans.
precoding for large multi-user MIMO systems, IEEE Trans. Commun., Inf. Theory, vol. 46, no. 3, pp. 933946, May 2000.
vol. 61, no. 3, pp. 10591071, Mar. 2013. [39] B. Hassibi and B. M. Hochwald, How much training is needed in
[15] A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson, Uplink perfor- multiple-antenna wireless links?, IEEE Trans. Inf. Theory, vol. 49, no.
mance of time-reversal MRC in massive MIMO systems subject to phase 4, pp. 951963, Apr. 2003.
noise, IEEE Trans. Wireless Commun., vol. 14, no. 2, pp. 711723, Feb. [40] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, Pilot
2015. contamination and precoding in multi-cell TDD systems, IEEE Trans.
[16] R. Krishnan, M. R. Khanzadi, N. Krishnan, A. Graell i Amat, T. Eriks- Commun., vol. 10, no. 8, pp. 26402651, Aug. 2011.
son, N. Mazzali, and G. Colavolpe, On the impact of oscillator phase [41] R. Muller, M. Vehkapera, and L. Cottatellucci, Blind pilot decontami-
noise on the uplink performance in a massive MIMO-OFDM system, nation, IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 773786,
Available online, http://arxiv.org/abs/1405.0669. Oct. 2014.
[42] C. Risi, D. Persson, and E. G. Larsson, Massive MIMO with 1-bit Michail Matthaiou (S05M08SM13) was born
ADC, Available online, http://arxiv.org/abs/1404.7736. in Thessaloniki, Greece in 1981. He obtained the
[43] I. Song, M. Koo, H. Jung, H.-S. Jhon, and H. Shinz, Optimization of Diploma degree (5 years) in Electrical and Com-
cascode configuration in CMOS low-noise amplifier, Micr. Opt. Techn. puter Engineering from the Aristotle University of
Lett., vol. 50, no. 3, pp. 646649, Mar. 2008. PLACE Thessaloniki, Greece in 2004. He then received the
[44] D. Petrovic, W. Rave, and G. Fettweis, Common phase error due PHOTO M.Sc. (with distinction) in Communication Systems
to phase noise in OFDM-estimation and suppression, in Proc. IEEE HERE and Signal Processing from the University of Bristol,
PIMRC, Sept. 2004, pp. 19011905. U.K. and Ph.D. degrees from the University of
[45] K.-G. Park, C.-Y. Jeong, J.-W. Park, J.-W. Lee, J.-G. Jo, and C. Yoo, Edinburgh, U.K. in 2005 and 2008, respectively.
Current reusing VCO and divide-by-two frequency divider for quadra- From September 2008 through May 2010, he was
ture LO generation, IEEE Microw. Wireless Compon. Lett., vol. 18, no. with the Institute for Circuit Theory and Signal
6, pp. 413415, June 2008. Processing, Munich University of Technology (TUM), Germany working as a
[46] J. R. Wilkerson, Passive Intermodulation Distortion in Radio Frequency Postdoctoral Research Associate. He is currently a Senior Lecturer at Queens
Communication Systems, Ph.D. thesis, North Carolina State University, University Belfast, U.K. and also holds an adjunct Assistant Professor position
2010. at Chalmers University of Technology, Sweden. His research interests span
[47] Further advancements for E-UTRA physical layer aspects (Release 9), signal processing for wireless communications, massive MIMO, hardware-
3GPP TS 36.814, Mar. 2010. constrained communications, and performance analysis of fading channels.
[48] M. R. Khanzadi, G. Durisi, and T. Eriksson, Capacity of Dr. Matthaiou was the recipient of the 2011 IEEE ComSoc Best Young
multiple-antenna phase-noise channels with common/separate oscil- Researcher Award for the Europe, Middle East and Africa Region and a co-
lators, IEEE Trans. Commun., 2014, To appear, Available: recipient of the 2006 IEEE Communications Chapter Project Prize for the best
http://arxiv.org/abs/1409.0561. M.Sc. dissertation in the area of communications. He was co-recipient of the
[49] E. Bjornson, M. Bengtsson, and B. Ottersten, Optimal multiuser Best Paper Award at the 2014 IEEE International Conference on Communi-
transmit beamforming: A difficult problem with a simple solution cations (ICC) and was an Exemplary Reviewer for IEEE C OMMUNICATIONS
structure, IEEE Signal Process. Mag., vol. 31, no. 4, pp. 142148, L ETTERS for 2010. In 2014, he received the Research Fund for International
July 2014. Young Scientists from the National Natural Science Foundation of China.
He has been a member of Technical Program Committees for several IEEE
conferences such as ICC, GLOBECOM, VTC etc. He currently serves as
an Associate Editor for the IEEE T RANSACTIONS ON C OMMUNICATIONS,
IEEE C OMMUNICATIONS L ETTERS and was the Lead Guest Editor of the
special issue on Large-scale multiple antenna wireless systems of the IEEE
J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS. He is an associate
member of the IEEE Signal Processing Society SPCOM and SAM technical
committees.
Merouane Debbah (SM08, F15) entered the Ecole

Normale Suprieure 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 Supelec (Gif-
sur-Yvette, France). From 2007 to 2014, he was
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
Emil Bjornson (S07, M12) received the M.S. de- and Algorithmic Sciences Lab.
gree in Engineering Mathematics from Lund Univer- His research interests are in information theory, signal processing and
sity, Sweden, in 2007. He received the Ph.D. degree wireless communications. He is an Associate Editor in Chief of the journal
in Telecommunications from the KTH Royal Insti- Random Matrix: Theory and Applications and was an associate and senior
PLACE tute of Technology, Stockholm, Sweden, in 2011. area editor for IEEE Transactions on Signal Processing respectively in 2011-
PHOTO From 2012 to July 2014, he was a joint postdoc at 2013 and 2013-2014. Mrouane Debbah is a recipient of the ERC grant MORE
HERE Suplec, Gif-sur-Yvette, France, and at KTH Royal (Advanced Mathematical Tools for Complex Network Engineering). He is a
Institute of Technology. He is currently an Assistant IEEE Fellow, a WWRF Fellow and a member of the academic senate of
Professor and Docent at the Department of Elec- Paris-Saclay. He is the recipient of the Mario Boella award in 2005, the 2007
trical Engineering (ISY) at Linkoping University, IEEE GLOBECOM best paper award, the Wi-Opt 2009 best paper award,
Sweden. the 2010 Newcom++ best paper award, the WUN CogCom Best Paper 2012
His research interests include multi-antenna cellular communications, radio and 2013 Award, the 2014 WCNC best paper award as well as the Valuetools
resource allocation, energy efficiency, massive MIMO, and network topology 2007, Valuetools 2008, CrownCom2009 , Valuetools 2012 and SAM 2014 best
design. He is the first author of the textbook Optimal Resource Allocation student paper awards. In 2011, he received the IEEE Glavieux Prize Award
in Coordinated Multi-Cell System published in Foundations and Trends in and in 2012, the Qualcomm Innovation Prize Award.
Communications and Information Theory, 2013. He is also dedicated to re-
producible research and has made a large amount of simulation code publicly
available. Dr. Bjornson received the 2014 Outstanding Young Researcher
Award from IEEE ComSoc EMEA, the 2015 Ingvar Carlsson Award, and has
received 4 best paper awards for novel research on multi-cell multi-antenna
communications: WCNC 2014, SAM 2014, CAMSAP 2011, and WCSP 2009.

Massive MIMO With Non-Ideal Arbitrary Arrays: Hardware Scaling Laws and Circuit-Aware Design

Uploaded by

Copyright:

Available Formats

Massive MIMO With Non-Ideal Arbitrary Arrays: Hardware Scaling Laws and Circuit-Aware Design

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Massive MIMO With Non-Ideal Arbitrary Arrays: Hardware Scaling Laws and Circuit-Aware Design

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been