Myths On Massive Mimo
Myths On Massive Mimo
Myths On Massive Mimo
Massive MIMO:
10 Myths and One Grand Question
AbstractWireless communications is one of the most successful technologies in modern years, given that an exponential
growth rate in wireless traffic has been sustained for over a
century (known as Coopers law). This trend will certainly
continue driven by new innovative applications; for example,
augmented reality and internet-of-things.
Massive MIMO (multiple-input multiple-output) has been
identified as a key technology to handle orders of magnitude
more data traffic. Despite the attention it is receiving from the
communication community, we have personally witnessed that
Massive MIMO is subject to several widespread misunderstandings, as epitomized by following (fictional) abstract:
The Massive MIMO technology uses a nearly infinite number
of high-quality antennas at the base stations. By having at least an
order of magnitude more antennas than active terminals, one can
exploit asymptotic behaviors that some special kinds of wireless
channels have. This technology looks great at first sight, but
unfortunately the signal processing complexity is off the charts
and the antenna arrays would be so huge that it can only be
implemented in millimeter wave bands.
The properties above are, in fact, completely false. In this
article, we identify 10 myths and explain why they are not true.
We also ask a grand question that will require intense future
research activities to answer properly.
I. I NTRODUCTION
Massive MIMO is a multi-user MIMO technology where
each base station (BS) is equipped with an array of M active
antenna elements and utilizes these to communicate with K
single-antenna terminalsover the same time and frequency
band. The general multi-user MIMO concept has been around
for decades, but the vision of actually deploying BSs with
more than a handful of service antennas is relatively new [1].
By coherent processing of the signals over the array, transmit
precoding can be used in the downlink to focus each signal
at its desired terminal and receive combining can be used in
the uplink to discriminate between signals sent from different
terminals. The more antennas that are used, the finer the spatial
focusing can be. An illustration of these concepts is given in
Figure 1a.
The canonical Massive MIMO system operates in timedivision duplex (TDD) mode, where the uplink and downlink
transmissions take place in the same frequency resource but
are separated in time. The physical propagation channels are
reciprocalmeaning that the channel responses are the same
in both directionswhich can be utilized in TDD operation. In
particular, Massive MIMO systems exploit the reciprocity to
estimate the channel responses on the uplink and then use the
acquired channel state information (CSI) for both uplink receive combining and downlink transmit precoding of payload
data. There are several reasons behind this protocol design.
Frequency
Uplink
Time
Frame structure
Uplink
data
Downlink
Uplink
pilots
Downlink
data
Bc
Tc
(a)
60 cm
120 cm
(b)
Fig. 1: Example of a Massive MIMO system. (a) Illustration of the uplink and downlink in line-of-sight propagation, where
each BS is equipped with M antennas and serves K terminals. The TDD transmission frame consists of = Bc Tc symbols.
By capitalizing on channel reciprocity, there is payload data transmission in both the uplink and downlink, but only pilot
transmission in the uplink. (b) Photo of the antenna array of the LuMaMi testbed at Lund University in Sweden [2]. The array
consists of 160 dual-polarized patch antennas. It is designed for a carrier frequency of 3.7 GHz and the element-spacing is 4
cm (half a wavelength).
K SNRu/d + 1
(1)
where K is the number of terminals, (1 K
) is the loss
from pilot signaling, and SNRu/d equals the uplink signal-tonoise ratio (SNR), SNRu , when Eq. (1) is used to compute
the uplink performance. Similarly, we let SNRu/d be the
downlink SNR, SNRd , when Eq. (1) is used to measure the
1
)1
downlink performance. In both cases, cCSI = (1 + KSNR
u
is the quality of the estimated CSI, proportional to the meansquared power of the MMSE channel estimate (where cCSI = 1
4
0
10
Cumulative probability
10 best terminals
All 12 terminals
10
LoSULA
Isotropic (Rayleigh)
FP channel
10
42
44
46
48
50
52
54
56
Sum capacity [bit/symbol]
58
60
62
(a)
80
Massive MIMO: Measured channels
Openloop: LoSULA
Openloop: Isotropic (Rayleigh)
70
60
50
40
30
Maximal codebook
size is reached
20
10
0
20
40
60
Number of service antennas (M)
80
100
(b)
The seminal work [1] on Massive MIMO studied the asymptotic regime where the number of service antennas M .
Numerous later works, including [4][6], have derived closedform achievable spectral efficiency expressions (unit: bit/s/Hz)
which are valid for any number of antennas and terminals, any
SNR, and any choice of pilot signaling. These formulas do not
rely on idealized assumptions such as perfect CSI, but rather
on worst-case assumptions regarding the channel acquisition
and signal processing. Although the total spectral efficiency
per cell is greatly improved with Massive MIMO technology,
the anticipated performance per user lies in the conventional
range of 1-4 bit/s/Hz [4]this is a regime where off-the-shelf
channel codes perform closely to the Shannon limits.
To show this, Figure 3 compares the empirical link performance of a Massive MIMO system with the uplink spectral
efficiency expression in Eq. (1). We consider M = 100 service
antennas, K = 30 terminals, and estimated channels using
one pilot per terminal. Each terminal transmits with QPSK
modulation followed by LDPC coding with rate 1/2, leading
to a net spectral efficiency of 1 bit/s/Hz/terminal; that is, 30
bit/s/Hz in total for the cell. By equating Eq. (1) to the same
target of 30 bit/s/Hz, we obtain the uplink SNR threshold
SNRu = 13.94 dB, as the value needed to achieve this rate.
Figure 3 shows the bit error rate (BER) performance for
different lengths of the codewords, and the BER curves
drop quickly as the length of the codewords increases. The
vertical line indicates SNRu = 13.94 dB, where zero
BER is achievable as the codeword length goes to infinity.
Performance close to this bound is achieved even at moderate
codeword lengths, and part of the gap is also explained by the
shaping loss of QPSK modulation and the fact that the LDPC
code is optimized for AWGN channels (which is actually a
good approximation in Massive MIMO due to the channel
hardening). Hence, expressions such as Eq. (1) are well-suited
to predict the performance of practical systems and useful for
resource allocation tasks such as power control (see Myth 9).
Myth 5: Too much performance is lost by linear processing
0.1
0.01
0.001
0.0001
SNR
threshold
from theory
6
100
60
40
FP (no interference)
Sum capacity
ZF
MR
20
0
20
30
40
50
60
70
80
Number of service antennas (M )
90
100
(a)
45
40
M
K
35
ZF
MR
4.3
M
K
30
1.9
25
20
15
10
5
0
20
40
60
Number of terminals (K)
80
100
(b)
A main feature of Massive MIMO is the coherent processing over the M service antennas, using measured channel
responses. Each desired signal is amplified by adding the M
signal components coherently, while uncorrelated undesired
signals are not amplified since their components add up
noncoherently.
Receiver noise and data signals associated with other terminals are two prime examples of undesired additive quantities
that are mitigated by the coherent processing. There is also
a third important category: distortions caused by impairments
in the transceiver hardware. There are numerous impairments
in practical transceivers; for example, non-linearities in amplifiers, phase noise in local oscillators, quantization errors
in analog-to-digital converters, I/Q imbalances in mixers,
and finite-order analog filters. The combined effect of these
impairments can be described either stochastically [12] or by
hardware-specific deterministic models [13]. In any case, most
hardware impairments result in additive distortions that are
substantially uncorrelated with the desired signal, plus a power
loss and phase-rotation of the desired signals. The additive
distortion noise caused at the BS has been shown to vanish
with the number of antennas [12], just as conventional noise
and interference, while the phase-rotations from phase noise
remain but are not more harmful to Massive MIMO than to
contemporary systems. We refer to [12] and [13] for numerical
examples of these facts.
In summary, the Massive MIMO gains are not requiring
high-precision hardware; in fact, lower hardware precision can
be handled than in contemporary systems since additive distortions are suppressed in the processing. Another reason for the
robustness is that Massive MIMO can achieve extraordinary
spectral efficiencies by transmitting low-order modulations to
a multitude of terminals, while contemporary systems require
high-precision hardware to support high-order modulations to
a few terminals.
Myth 9: With so many antennas, resource allocation and power
control is hugely complicated
k K
(2)
maximize
1 ,...,K [0,K]
P
k k K, R0
subject to
R
cCSI,k M SNRd,k k 2R 1
!
K
X
SNRd,k
i + 1
for k = 1, . . . , K.
i=1
This resource allocation problem is known as max-min fairness, and since we maximize the worst-terminal performance
the solution gives the same performance to all terminals.
The reformulation in Eq. (2) is the epigraph form of the
original formulation. From this reformulation it is clear that
all the constraints are linear functions of the power-control
coefficients 1 , . . . , K , thus Eq. (2) is a linear optimization
problem for every fixed worst-terminal performance R. The
whole problem is solved by line search over R, to find the
largest R for which the constraints are feasible. In other
words, the power control optimization is a so-called quasilinear problem and can be solved by standard techniques
(e.g., interior point methods) with low complexity. We stress
that the power control in Eq. (2) only depends on the largescale fadingthe same power control can be applied on all
subcarriers and over a relatively long time scale.
To summarize, the resource allocation can be greatly simplified in Massive MIMO systems. It basically reduces to
admission control (which terminals should be active) and longterm power control (in many cases it is a quasi-linear problem).
The admitted terminals may use the full bandwidththere is
no need for frequency-selective allocation when there is no
frequency-selective fading. The complexity of power control
problems such as Eq. (2) scales with the number of terminals,
but is independent of the number of antennas and subcarriers.
Myth 10: With so many antennas, the signal processing complexity
will be overwhelming
8
[W]
1000
1000
10
100
10
Not
feasible
100
Number of service antennas (M)
1000
10000
Coherence Block Length (symbols)
100
[Gflops]
1000
100
11
ZF/MMSE
10000
Coherence block length (symbols)
MR
10000
1000
100
10
Not
feasible
100
Number of service antennas (M )
1000
Fig. 5: Computational complexity (in flops) of the main baseband signal processing operations in an OFDM Massive MIMO
setup: FFTs, channel estimation, precoding/combining of payload data, and computation of precoding/combining matrices. The
complexity is also converted into an equivalent power consumption using a typical computational efficiency of 12.8 Gflops/Watt
[15].
symbol basis.
T HE G RAND Q UESTION
Can Massive MIMO work in FDD operation?
Number of terminals (K )
Overhead
50 %
(a) FDD
40 %
(100,25)
for = 200
Number of terminals (K )
30 %
20 %
(b) TDD
10 %
0%
Number of service antennas (M)
Fig. 6: Illustration of the typical overhead signaling in Massive MIMO based on (a) FDD and (b) TDD operation. The main
difference is that FDD limits the number of antennas, while TDD can have any number of antennas. For a coherence block
with = 200 even a modest Massive MIMO setup with M = 100 and K = 25 is only supported in TDD operation.
at Supelec, France, and KTH Royal Institute of Technology,
Sweden. Since 2014 he is a Research Fellow at Linkoping
University, Sweden. He received the 2014 Outstanding Young
Researcher Award from IEEE ComSoc EMEA, is the first
author of the textbook Optimal Resource Allocation in Coordinated Multi-Cell Systems, and has received best conference
paper awards in 2009, 2011, and 2014.
Erik G. Larsson (erik.g.larsson@liu.se) is Professor and
Head of the Division for Communication Systems in the
Department of Electrical Engineering (ISY) at Linkoping
University (LiU) in Linkoping, Sweden. He has published
some 100 journal papers, he is co-author of the textbook
Space-Time Block Coding for Wireless Communications and
he holds many patents. He has been Associate Editor for
several IEEE journals. He serves as chair of the IEEE Signal
Processing Society SPCOM technical committee in 2015.
He also serves as chair of the steering committee for the
IEEE Wireless Communications Letters in 20142015. He
is active in conference organization, most recently as the
General Chair of the Asilomar Conference on Signals, Systems
and Computers 2015 (he was Technical Chair in 2012). He
received the IEEE Signal Processing Magazine Best Column
Award twice, in 2012 and 2014.
Thomas L. Marzetta (tom.marzetta@alcatel-lucent.com) is
the originator of Massive MIMO, the most promising technology available to address the ever increasing demand for wireless throughput. He is a Bell Labs Fellow and Group Leader
of Large Scale Antenna Systems at Bell Labs, Alcatel-Lucent
and Co-Head of their FutureX Massive MIMO project. Dr.
I. C. Wong, V. Owall,
O. Edfors, and F. Tufvesson, A flexible 100antenna testbed for massive MIMO, in Proc. IEEE Globecom Workshop
- Massive MIMO: From Theory to Practice, 2014.
[3] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, A coordinated approach
to channel estimation in large-scale multiple-antenna systems, IEEE J.
Sel. Areas Commun., vol. 31, no. 2, pp. 264273, 2013.
[4] E. Bjornson, E. G. Larsson, and M. Debbah, Massive MIMO for
maximal spectral efficiency: How many users and pilots should be
allocated?, IEEE Trans. Wireless Commun., Submitted, Available:
http://arxiv.org/pdf/1412.7102.
[5] H. Ngo, E. Larsson, and T. Marzetta, Energy and spectral efficiency of
very large multiuser MIMO systems, IEEE Trans. Commun., vol. 61,
no. 4, pp. 14361449, 2013.
10
[6] J. Hoydis, S. ten Brink, and M. Debbah, Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?, IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160171, 2013.
[7] A. Alkhateeb, O. El Ayach, G. Leus, and R.W. Heath, Channel
estimation and hybrid precoding for millimeter wave cellular systems,
IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831846, 2014.
[8] A. Adhikary, E. Al Safadi, M. K. Samimi, Rui Wang, G. Caire, T. S.
Rappaport, and A. F. Molisch, Joint spatial division and multiplexing
for mm-wave channels, IEEE J. Sel. Areas Commun., vol. 32, no. 6,
pp. 12391255, 2014.
[9] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, Aspects of favorable
propagation in massive MIMO, in Proc. EUSIPCO, 2014.
[10] X. Gao, O. Edfors, F. Rusek, and F. Tufvesson, Massive MIMO in real
propagation environments, IEEE Trans. Wireless Commun., 2015, to
appear, http://arxiv.org/abs/1403.3376.
[11] M. Karlsson and E. G. Larsson, On the operation of massive MIMO
with and without transmitter CSI, in Proc. IEEE SPAWC, 2014.
[12] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, Massive
MIMO systems with non-ideal hardware: Energy efficiency, estimation,
and capacity limits, IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112
7139, 2014.
[13] U. Gustavsson, C. Sanchez-Perez, T. Eriksson, F. Athley, G. Durisi,
P. Landin, K. Hausmair, C. Fager, and L. Svensson, On the impact of
hardware impairments on massive MIMO, in Proc. IEEE GLOBECOM,
2014.
[14] H. Yang and T.L. Marzetta, Total energy efficiency of cellular large
scale antenna system multiple access mobile networks, in Proc.
OnlineGreenComm, 2013.
[15] D. Schneider,
Could supercomputing turn to signal processors
(again)?, IEEE Spectrum, pp. 1314, Oct. 2012.