MASSIVE MIMO:

AN INTRODUCTION
D
emand for wireless throughput, both mobile and fixed, will always increase.
One can anticipate that, in five or ten years, millions of augmented reality
users in a large city will want to transmit and receive holographic video
more or less continuously, say 100 megabits per second per user in each di-
rection. Massive MIMO—also called Large-Scale Antenna Systems—is a promising
candidate technology for meeting this demand. Fifty-fold or greater spectral efficiency
improvements over fourth generation (4G) technology are frequently mentioned. A
multiplicity of physically small, individually controlled antennas performs aggressive
multiplexing/demultiplexing for all active users, utilizing directly measured channel
characteristics. By leveraging time-division duplexing (TDD), Massive MIMO is scal-
able to any desired degree with respect to the number of service antennas. Adding
more antennas is always beneficial for increased throughput, reduced radiated power,
uniformly great service everywhere in the cell, and greater simplicity in signal pro-
cessing. Massive MIMO is a brand new technology that has yet to be reduced to prac-
tice. Notwithstanding, its principles of operation are well understood, and surprisingly
simple to elucidate.
.........................................
Thomas L. Marzetta Introduction
Two timeless truths are evident: first, demand for wireless throughput will always
grow; second, the quantity of available electromagnetic spectrum will never increase.
Wireless communications is fundamentally different from optical fiber communica-
tions in that more fiber can always be manufactured and laid down, so irrespective of
the cleverness of optical researchers there is no doubt that any future optical demand
will always be met. In contrast, there is no easy solution for wireless throughput.
The fundamental and perennial wireless problem is a physical layer problem:
how to provide ever-increasing total wireless throughput reliably and uniformly
throughout a designated area [1, 2]. All proposed solutions seem to fall into one of
three categories: 1) exploitation of spectrum that is currently unused or underutilized;
2) deployment of ever more access points, each covering a commensurately smaller
area; and 3) use of access points and/or terminals with multiple antennas. The first
two activities as epitomized by millimeter wave technology and small cells are more
than adequately treated elsewhere and will not be discussed in this paper. The third
activity is known as MIMO (multiple input multiple output), and it further divides
into its original form, Point-to-Point MIMO [3–6], and the later (in terms of its theo-
retical development) Multi-User MIMO [7–10] of which Massive MIMO is emerging
as its ultimate and most useful form [11–14].
Figure 1 illustrates some of the features of Massive MIMO. This could constitute
one cell of a network of cells, or it could be an isolated (single-cell) site. An array of
physically small, non-directive antennas serves a multiplicity of autonomous users
(terminals). The terminals typically would have only a single antenna each. Downlink
operation, shown in the figure, entails transmitting a different data stream to each
user. The central task is to ensure that each user receives only the data stream in-
tended for him, with minimal interference from the other data streams. Contemporary
systems typically accomplish this multiplexing by some combination of sending the
various data streams at different times (time-division multiplexing) and over different
frequencies (frequency-division multiplexing). In contrast, Massive MIMO uses

1538-7305/14©2015ALCATEL-LUCENT V O L U M E 2 0 · Bell Labs Technical Journal

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 12

FIGURE 1. Downlink operation of a Massive MIMO link. The antenna array selectively transmits a multiplicity of data streams, all occupying the
same time/frequency resources, so that each user receives only the data stream that it intended for him.

spatial-division multiplexing such that the different data Portions of the preceding description of Massive
streams occupy the same frequencies and times. A key MIMO also apply to a generic Multi-User MIMO system.
element to performing wireless spatial multiplexing is an What distinguishes Massive MIMO is that certain activi-
array of independently-controlled antennas. Under line-of- ties make it a scalable technology. The multiplexing and
sight propagation conditions, the data streams are carried demultiplexing operations are accomplished using directly
on focused beams of data. In a cluttered propagation envi- measuredVrather than assumedVchannel characteristics.
ronment, the data streams can arrive from many directions (In contrast, sectorization as well as open-loop angle-
simultaneously. The streams tend to reinforce each other selective beam-forming assume line-of-sight propagation
constructively where desired, and interfere destructively conditions, and as explained later, are ultimately not scal-
where they are unwanted. To carry out the multiplexing, able.) If measured channel responses are utilized, increas-
the array needs to know the frequency response of the ing the number of antennas always improves the
propagation channels between each of its elements and performance of the system irrespective of the noisiness of
each of the users. This channel state information (CSI) is the measurements. The benefits of growing the number
utilized in the precoding block (shown in the figure) and of service antennas relative to the number of active
it is here that the data streams are mapped into the signals users include greater selectivity in transmitting and re-
that drive each of the antennas. By increasing the number ceiving the data streams, which leads to greater through-
of antennas, the beams can be focused more selectively to put, a reduction in the required radiated power, effective
the users. power control that provides uniformly good service
throughout the cell, and greater simplicity in signal
............................................... processing.
The paper is organized as follows. We begin with an
Demand for wireless throughput will overview of Point-to-Point MIMO, and explain in particu-
always grow, but the quantity of lar why it is not a scalable technology. Next, we consider
the theory of Multi-User MIMO, which has fundamental
available electromagnetic spectrum will advantages over Point-to-Point MIMO, but which in its
original conception also is not scalable. We then provide a
never increase. detailed description of Massive MIMO: why it is scalable,
how CSI is acquired, the simplest effective types of multi-
The uplink operation of the Massive MIMO system, plexing and demultiplexing, power control, and the ulti-
shown in Figure 2, is substantially the reverse of down- mate limitations of Massive MIMO. We follow with a
link operation. The users transmit data streams at the case study of a dense urban macrocellular Massive MIMO
same time and over the same frequencies. The antenna ar- system in which each cell has a 64-antenna base station
ray receives the sum of the data streams as modified by that serves 18 users. Despite the modest size of this sys-
their respective propagation channels, and the decoding tem, it has amazingly good performance, particularly in
operation, again utilizing CSI, untangles the received sig- that it can provide uniformly good service throughout the
nals to produce the individual data streams. cell, even at the cell edges. We conclude with a brief

Bell Labs Technical Journal · V O L U M E 2 0

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 13

FIGURE 2. Uplink operation of a Massive MIMO link. The users transmit data streams that occupy the same time/frequency resources, and the
signals received by the elements of the antenna array are processed to recover the individual data streams.

summary of current research issues and describe potential where d is the expected signal-to-noise ratio (SNR) at
non-cellular applications for Massive MIMO. any receiver if full power were fed into one of the transmit
antennas. Thus, without increasing either spectral band-
Point-to-Point MIMO width or radiated power, throughput can be increased by
Figure 3 depicts a Point-to-Point MIMO link in which a adding antennas to each end of the link.
base station with a concentrated array of M antennas trans- Shannon theory yields a celebrated formula for sys-
mits data to a user who has a concentrated array of K tem capacity, expressed in bits/s/Hz:
antennas.
d 
Different users are accommodated in disjoint time/ C ¼ log2 det IK þ GH G d

M d 
frequency blocks via a combination of time division and d
frequency division multiplexing. Every use of the channel ¼ log2 det IM þ Gd GH d ; (2)
M
comprises transmitting a signal vector and receiving a
where Gd is the M  K frequency response of the matrix-
signal vector, where every received signal is a linear com-
valued channel that connects the base station antennas and
bination of transmitted signals, and the combining coeffi-
the user antennas, IK denotes the K  K identity matrix,
cients are determined by the propagation between the two
and the superscript “H” denotes “conjugate transpose.”
ends of the link. Subject to a number of assumptions, at
The validity of equation 2 depends on the receiver additive
sufficiently high signal-to-noise ratios, the spectral effi-
noise being complex Gaussian, and more importantly, on
ciency of the link expressed in bits/second/Hz, is approxi-
the receiver knowing the downlink channel matrix. The
mately as follows:
transmitter does not have to know the channel, although
C / minðM; KÞ log2 ðd Þ; d  1; (1) operation would be both simplified and improved

FIGURE 3. Point to Point MIMO link: A base station equipped with an antenna array serves a user equipped with an antenna array. Different users
are served over different time/frequency blocks.

V O L U M E 2 0 · Bell Labs Technical Journal

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 14

somewhat if both ends of the link knew the channel. If the TABLE I. Shannon Capacity (bits/s/Hz) vs. Number of Base
elements of Gd are independent, identically distributed, Station Antennas for a Four-Antenna User Having an SNR
zero-mean complex Gaussian, unit-variance random vari- of 3.0 dB
.......................................................................
ables (i.i.d. Rayleigh fading), then asymptotic random ma-
trix theory yields equation 1, a formula which is valid for
sufficiently high SNRs.
Equation 2 applies to downlink operation, with M
transmit antennas and K receive antennas. For reverse op-
eration on the same link we have K transmit antennas and an SNR of 3.0 dB, a typical value for cell-edge users.
M receive antennas, and the same capacity formula applies Table I shows the theoretical Shannon capacity (equation 2)
with the term d =M replaced by u =K, and with Gd for M ¼ 1; 2; 4; 8 base station antennas (note that here we
replaced by Gu. For FDD systems the uplink channels are take the mean of equation 2 as consistent with ergodic
different than the downlink channels, while for TDD coding over many independent realizations of the random
systems the uplink channels are theoretically equal to the channel). It is clear that barely two data streams are sup-
downlink channels. ported, and that most of the efficacy is due to the four re-
The intuitive idea behind Point-to-Point MIMO is ceive antennas which collect four-fold signal power.
that the base station transmits a vector-valued signal As we show in the next section, Multi-User MIMO
whose components contain distinct pieces of information. mitigates some of the drawbacks of Point-to-Point MIMO,
This vector is multiplied by the channel matrix to produce but as originally conceived, it still is not scalable.
the vector of received signals. If the channel matrix is
well-conditioned then the receiver can reliably recover
either the distinct transmitted signals, or at least some Multi-User MIMO
nontrivial linear combinations of transmitted signals. A comparison of Figures 1 and 2 with Figure 3 implies
In order for the receiver to learn the matrix-valued that Multi-User MIMO is equivalent to starting with a
channel, the transmitter has to send known training signals Point-to-Point MIMO link, and breaking the single K an-
(pilots) through the channel. Subject to peak power restric- tenna user into K autonomous single-antenna users. Since
tions, the optimum training signals are mutually orthogo- the users are not capable of communicating with each other,
nal [15]. Consequently the sample duration of the pilot it is evident that the throughput achievable for individual
sequences should be at least as great as the number of users on poor quality channels can be severely impacted by
transmit antennas, d  M. Uplink data transmission re- the break-up. The remarkable and non-intuitive fact is that,
quires the base station to learn the uplink channel matrix subject to certain assumptions about CSI, the sum through-
which requires uplink pilots such that u  K. Hence the put (the sum of the individual capacities to and from the
total training burden (e.g., the amount of time required for users) is not reduced by the break-up! It is interesting to
training) for a complete system, either TDD or FDD, take note of a paper from 1919 in which directional beam-
grows as forming from a multiple-antenna array is mentioned as a
possible solution to the perceived problem of inadequate
d þ u  M þ K: (3)
spectrum [16]. (At that time it was believed that transoce-
Point-to-Point MIMO is an option under current wireless anic wireless communication was only possible at wave-
standards. The 802.11ac standard, for example, permits up lengths between 10 and 20 km and that the spectrum could
to ðM; KÞ ¼ ð8; 8Þ. For a variety of reasons, Point-to- only support about 200 worldwide stations.)
Point MIMO is not readily scalable beyond 8  8. First, The Shannon sum-capacity for uplink Multi-User
the propagation environment may not support eight data MIMO is identical to that of uplink Point-to-Point MIMO,
streams. Line-of-sight conditions present a particular chal-
u 
lenge because, for compact arrays, the channel matrix has Csum up ¼ log2 det IK þ GH u Gu ; (4)
K
the minimum rank of one which permits only one data
and in both cases, only the base station has to know the
stream. Second, the activity of scaling up the number of
uplink channel matrix.
antennas requires that proportional amounts of time have
The formula for downlink Shannon sum-capacity re-
to be spent on training, as indicated by equation 3. Third,
quires the solution of a convex optimization problem,
near the edge of the cell, SINRs are typically low, and   
multiplexing gains fall short of the promised minðM; KÞ. Csum down ¼ supa log2 det IM þ d Gd Da GH d ;
Fourth, the user equipment is complicated, and indepen-
a  0; 1T a ¼ 1; (5)
dent electronics chains are required for each antenna.
Fifth, achieving performance close to the Shannon limit where Da is a diagonal matrix whose diagonal elements
requires rather involved signal processing on the part of comprise the M  1 vector, a, and 1 denotes the M  1
both the base station and the user. To illustrate the third vector of ones. Crucially, this capacity is predicated on
point, consider a user having K ¼ 4 antennas, operating at both ends of the link knowing the downlink channels.

Bell Labs Technical Journal · V O L U M E 2 0

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 15

(More precisely, the transmitter has to know the full down- with both the number of service antennas and the num-
link channel, while each user has to know only its own ber of users.
downlink channel.) A comparison of equation 5 with
equation 2 reveals that the downlink multi-user sum ca- Massive MIMO: A Scalable Technology
pacity actually exceeds that of the point-to-point link Massive MIMO breaks the scalability barrier by not at-
which can be attributed to the additional channel state in- tempting to achieve the full Shannon limit, and, paradoxi-
formation assumed in the case of the multi-user system. cally, by increasing the size of the system. It departs from
To achieve near-capacity performance requires so-called Shannon-theoretic practice in three ways. First, only the
dirty paper coding/decoding whose computational burden base station learns the downlink channel. In a TDD system
grows exponentially with the size of the system. More- the time required to acquire CSI is independent of the
over, the success of dirty paper coding depends on having number of base station antennas. Second, the number of
very accurate channel estimates. base station antennas is typically increased to several times
The base station can acquire uplink CSI through up- the number of users. Third, a simple linear precoding mul-
link pilots of duration u  K. If TDD is employed on tiplexing is employed on the downlink, coupled with lin-
the uplink, pilots will also provide the base station with ear decoding demultiplexing on the uplink. As the number
the downlink CSI, because reciprocity implies that the up- of base station antennas increases, linear precoding and
link and downlink channels are mathematically identical. decoding performance can approach the Shannon limit.
Downlink CSI for the users (required for dirty paper de-
coding) is acquired from orthogonal downlink pilots of ...............................................
duration d  M. Hence complete CSI acquisition oc-
cupies time (assuming the more favorable TDD) propor-
Massive MIMO breaks the scalability
tional to d þ u  M þ K. barrier by not attempting to achieve the
On the positive side, Multi-User MIMO has two ad-
vantages over Point-to-Point MIMO. First, it is less vulne- full Shannon limit, and, paradoxically,
rable to the propagation environment. It can function well
even under line-of-sight conditions provided the typical
by increasing the size of the system.
angular separation between users is greater than the an-
gular resolution of the base station array. Second, only
single-antenna terminals are required. 1. Linear Precoding and Decoding
What prevents the Shannon-theoretic version of Multi- Figure 4 and Figure 5 illustrate matched filter decoding
User MIMO from being scalable is first, the exponentially for uplink data transmission, and conjugate beamforming
growing complexity of dirty paper coding/decoding, for downlink data transmission, the simplest type of linear
and second, the time spent acquiring CSI which grows decoding and precoding.

FIGURE 4. Matched filter decoding for uplink Massive MIMO.

V O L U M E 2 0 · Bell Labs Technical Journal

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 16

FIGURE 5. Conjugate beamforming for downlink Massive MIMO.

On the uplink, the users transmit their data bearing A natural question follows: How do imperfect chan-
(QAM) symbols simultaneously, with no MIMO-specific nel estimates affect the performance of linear precoding
signal processing, except for optional power control weight- and decoding? In particular, as we add more antennas,
ing (which is not shown in the figure). The k-th user trans- does the quality of the CSI have to improve in order to
mits a QAM symbol denoted by qk. The frequency maintain coherent gains proportional to M? The definitive
response between the k-th user and the m-th base station an- answer is favorable: irrespective of the noisiness of the
tenna is denoted by gmk where the dependence on fre- channel estimates, the coherent gain grows in proportion
quency is not explicitly shown. To recover the k-th QAM to M [11], albeit with a proportionality factor that de-
symbol, the base station weights its m-th received signal by creases with increasing estimation error.
the complex conjugate of its estimate for the mk-th channel It is also possible to use other types of linear pre-
coefficient, ^
gmk , and sums this weighted signal over the M coding and decoding. In general, linear decoding con-
antennas to produce ^qk . With only one user, this processing sists of multiplying the received vector by a K  M
would be optimal for enhancing SNR, but multiple users matrix, and linear precoding multiplies the vector of
will generate interference (crosstalk). Under favorable prop- QAM symbols by an M  K matrix, where the matrices
agation conditions, which experimentally have been shown are functions of the channel estimates. For matched fil-
to prevail [12, 17], the vector-valued channels from one tering and conjugate beamforming, the matrices are
user to another are asymptotically orthogonal as M in- complex conjugates of the estimated channel matrix.
creases, and the expected power of the desired signal grows Zero-forcing, instead, uses the pseudo-inverse of the esti-
M times faster than the power of the interference. mated channel matrix. Under high SNR conditions, zero-
On the downlink, the k-th QAM symbol, qk (with op- forcing may perform significantly better than matched
tional power weighting), is multiplied by the complex con- filtering and conjugate beamforming [18]. An advantage
jugate of the estimate for the mk-th channel coefficient, of matched filtering and conjugate beamforming is that
^
gmk , and the weighted signals are summed over the K the Massive MIMO signal processing can be performed
users to produce the signal that is fed into the m-th an- locally at each antenna, as is apparent from Figures 4
tenna. The weighting ensures that the components of the and 5. This in turn permits a decentralized architecture
M transmitted signals that are associated with the k-th for the antenna array, which lends great resilience to the
QAM symbol arrive in phase at the k-th user. For a single system. For example, if half the antennas are lost from a
user, this weighting maximizes the received power for a lightning strike, the remaining antennas do exactly what
total given transmitted power. Again, the asymptotic or- they did before. Likewise, during periods of slack de-
thogonality of channels under favorable propagation im- mand, some antennas can be put into sleep mode, for
plies that power for desirable signals grows M times faster improved energy efficiency, without affecting the opera-
than the power of the crosstalk interference. tions of the others.

Bell Labs Technical Journal · V O L U M E 2 0

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 17

FIGURE 6. Slot structure for TDD Massive MIMO.

2. Channel Estimation
Massive MIMO relies on measuring the frequency re-
sponses of the actual propagation channels. To that end,
either the users or the base station transmit known training
signals and the receiver opposite then estimates the fre-
quency response. Once the channels have been estimated,
the CSI has to be utilized in a timely manner before the
motion of the users significantly changes the channels.
Hence there is only a limited amount of time available for
training. In a Massive MIMO system both training and
data transmission take place in a slot whose duration is
chosen so that nobody moves more than a fraction of a
wavelength within the slot. FIGURE 8. TDD enables Massive MIMO scalability for mobile users,
while FDD is strictly limited to small systems.
In a TDD system, the users transmit orthogonal pilot
sequences, of sample duration u  K. It is convenient to
think of the frequency response as piece-wise constant over Figure 8 shows the overwhelming advantage of TDD
intervals defined by the Nyquist sampling intervalVequal over FDD for mobile users. The vertical axis is the number
in Hertz to the reciprocal of the channel delay-spread. of base station antennas, and the horizontal axis is the num-
Hence, over each Nyquist interval the terminals transmit or- ber of users. The light-to-medium blue region shows the
thogonal pilot sequences. For typical OFDM parameters system dimensions obtainable with TDD versus the much
(symbol interval Ts ¼ 1=14 ms, usable interval Tu ¼ smaller red region for FDD. The slot duration is 4 ms
1=15 ms, guard interval Tg ¼ 1=ð14  15Þ ms), if the de- which permits 35 km/h mobility at 1.9 GHz, and to miti-
lay-spread is assumed to be equal to the guard interval, the gate pilot contamination (explained later) the orthogonal pi-
Nyquist interval is equivalent to TTug ¼ 14 tones. In a slot of lot sequences are made seven times longer than necessary.
duration (in seconds) Tslot , there are TTslots OFDM symbols.
We define the sample duration of the slot as the number of 3. The Importance of Using Measured Rather Than
OFDM symbols times the tone duration of the Nyquist in- Assumed Channel Characteristics
terval, T ¼ TTslots TTgu ; this is the number of channel uses in Massive MIMO utilizes channel information that is di-
each Nyquist interval of each slot. Shorter slots permit rectly measured rather than assumed. This, in turn, makes
higher mobility: at 1.9 GHz a 280 km/h user would move Massive MIMO a scalable technology: any number of
1/4 wavelength during a 500 s slot. Thus a 500 s slot base station antennas can be usefully employed with no
has an equivalent sample duration of T ¼ 7  14 ¼ 98, so tightening of array tolerances. Extra antennas always help.
as many as 98 terminals could be trained. However this up- In contrast, if an assumed channel response is used,
per limit would leave no time for transmitting data, and it the technology is ultimately not scalable. One technique
can be shown that under general conditions no more than for downlink beamforming is to form open-loop beams,
half of the slot should be expended for training. Figure 6 transmit a downlink pilot through each beam, and, on the
illustrates the TDD slot structure. uplink, have each user report back which beam is the
An FDD system requires considerably more time for strongest. The base station then transmits downlink data to
training than a TDD system [19]. First, the M base station each user through its preferred beam. The open loop
antennas transmit orthogonal downlink pilots with d  M. beams are directional beams subject to the assumption of
Next, the K users have to transmit their M received pilot line of sight propagation. While this scheme may work up
signals on the uplink, which expends an additional d sam- to a point, eventually the addition of more antennas yields
ples of resources. Finally, the K users have to transmit ordi- little further improvement. This happens for two reasons.
nary uplink pilot sequences with u  K. Figure 7 shows First, any real propagation environment has some non-
the structures of the FDD downlink and uplink slots. The zero angle spread: one can transmit ever-finer beams, but
entire process requires a minimum of 2d þ u  2M þ K the propagation medium itself will broaden the beams. To
channel uses, in contrast to u  K channel uses for TDD. put it another way, there is a mismatch between any of the
fine open loop beams and the actual propagation to the
user. The second reason for lack of scalability is that even
if the propagation were perfect line-of-sight, the activity of
forming ever-finer open loop beams would require increas-
FIGURE 7. Slot structure for FDD Massive MIMO. ingly tight array tolerances. Figure 9 illustrates the

V O L U M E 2 0 · Bell Labs Technical Journal

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 18

FIGURE 10. Total spectral efficiency versus number of base station

antennas for K equal to 16, 32, 64, 128 users operating at a minus 6 dB
FIGURE 9. The array gain of Massive MIMO always grows linearly SINR. Asterisk indicates performance of 8  4 Point-to-Point MIMO.
with the number of antennas. Open-loop beamforming eventually Crosses indicate M ¼ 4K operating points.
yields only logarithmic improvements in array gain.

Other strategies are possible. For example one could spec-

divergence of array gain achievable by Massive MIMO ify desired rates for a subset of users, and subject to these
based on the measured channels and the array gain constraints use linear programming to find the power con-
achieved by open-loop beamforming. Massive MIMO gain trol coefficients that yield max-min throughput for the re-
increases linearly with the number of antennas, but under maining users.
open-loop beamforming, when the beamwidth becomes
less than the angle spread of the medium, the array gain in- 5. Performance of Massive MIMO
creases only logarithmically with the number of antennas. Massive MIMO can operate in a regime unavailable to
Point-to-Point MIMO, as illustrated in Figure 10, com-
4. Power Control prising plots of sum spectral efficiency for K ¼
As indicated in an earlier section, power control, in both ½16 32 64 128 users as functions of M for an
uplink and downlink, entails multiplying the original set SINR of 6.0 dB. The red X's correspond to the di-
of K QAM symbols by power control coefficients, mensions, M ¼ 4K. The point ðM; KÞ ¼ ð64; 16Þ yields
a total spectral efficiency of 13.6 bits/s/Hz which is dou-
pffiffiffiffiffi X
K
downlink : qk ! k qk ; k ¼ 1;    ; K; k  1 bled for every simultaneous doubling of ðM; KÞ. In con-
k¼1 trast, 8  4 Point-to-Point MIMO has a spectral efficiency
pffiffiffiffiffi
uplink : qk ! k qk ; k  1; k ¼ 1;    ; K: (6)

One of the nice properties of Massive MIMO is that the

large number of antennas makes the beamforming gains
virtually constant over frequency, and moreover, depen-
dent only on large-scale (slow) fading coefficients, which
themselves are independent of frequency and of antenna
index. Hence the power control coefficients can be made
independent of frequency and their effect on the data rate
attained by an individual user may be computed without
regard to the short term channel estimates obtained from
the pilots. It happens that, in the mathematical expressions
for the effective SINRs of the precoded/decoded signals,
the power control coefficients enter the expressions in a
way such that inequality constraints on SINRs are equiva-
lent to linear inequality constraints on power control coef-
ficients. So max-min power control, which gives equal
FIGURE 11. Total spectral efficiency versus number of base station
throughput to everyone in the cell, is obtained by solving antennas for K ¼ 16 users and 0.0 dB SINR. Blue curve: Shannon limit
a set of linear equations for the power control coefficients. (dirty-paper coding); Red curve; Massive MIMO (linear pre-coding).

Bell Labs Technical Journal · V O L U M E 2 0

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 19

of only 1.3 bits/s/Hz. For these results perfect CSI is as- Case Study: Dense Urban Macro-Cellular
sumed. The Point-to-Point MIMO performance is ergodic Massive MIMO
Shannon capacity according to equation 2. Massive We next examine a multi-cellular scenario in which each
MIMO performance is computed as a capacity lower- cell, containing on average 18 randomly located users, is
bound for conjugate beamforming according to a formula served by a base station with a 64-element array. The cell
derived in [18]: radii (center to vertex) are 500 meters, the carrier fre-
quency is 1.9 GHz, and the spectral bandwidth is 20 MHz.
Md
Csum cb > K log2 1 þ : (7) The 64-element array, comprising 1/2 wavelength patch
Kð1 þ d Þ
antennas arranged in a cylindrical configuration, is only
Figure 11 shows that by employing additional base sta- 40 cm in diameter and 55 cm high, as shown in Figure 12.
tion antennas, the linear precoding used in Massive A persistent mythVeasily refuted by Figure 12Vis that
MIMO is highly competitive with the dirty-paper coding Massive MIMO is not practical at anything but millimeter
mandated by Shannon theory. For K ¼ 16 users and an wavelengths because of the allegedly huge physical size
SINR of 0.0 dB, the blue curve represents the Shannon of the array. The maximum total downlink radiated power
total spectral efficiency, and the red curve is the total of the base station is one Watt (16 mW per antenna),
spectral efficiency for Massive MIMO with linear precod- and the maximum uplink radiated power of each terminal
ing, as a function of the number of base station antennas. is 200 mW.
For M ¼ 64 antennas, Massive MIMO gives the same
performance as dirty-paper coding for M ¼ 55. Here the 1. Training, Pilot Contamination, and Slot Structure
Shannon limit is computed according to equation 5, while The TDD slot duration is 2 ms, permitting 71 km/h mobil-
the linear precoding lower bound is computed as the ity, and according to the OFDM parameters discussed in
greater of the conjugate beamforming lower bound (equa- the section on Massive MIMO channel estimation, the
tion 7) or the zero-forcing lower bound [18]: sample duration of the slot is T ¼ 392. Pilot sequences of
duration 18 would be sufficient to ensure orthogonality
ðM  KÞd within each cell, but re-use of the same 18 pilots from cell
Csum zf > K log2 1 þ : (8)
K to cell gives rise to pilot contamination [14]. Pilot re-use
implies that the pilot-derived estimate for the channel be-
6. Ultimate Limitation of Massive MIMO tween the home base station and one of its users is con-
When serving mobile users, there is a finite limit to the taminated by channels between the base station and users
number of users that can be served simultaneously be- in other cells which share the same pilot sequence. On
cause of the overhead required for CSI acquisition. Of downlink, the base station inadvertently transmits coherent
course one can add as many base station antennas as interference to the users in other cells. This impairment
desired which will increase proportionally the SINR does not decrease with the addition of more antennas. A
experienced by the users, but this ultimately yields only
logarithmic improvements in throughput. Short of a break-
through in channel estimation, extremely large arrays (say
conformal arrays on the sides of skyscrapers) should be re-
served primarily for fixed terminals, where theoretically,
there is an unlimited amount of time for training.

FIGURE 12. Antenna array for 1.9 GHz: Each of 64 antennas is 1/2
wavelength by 1/2 wavelength in size (8 cm  8 cm), spaced 1/2 wave-
length apart circumferentially, and spaced one-wavelength apart FIGURE 13. Pilot re-use seven. Cells having different colors have
vertically. mutually orthogonal pilot sequences.

V O L U M E 2 0 · Bell Labs Technical Journal

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 20

similar effect occurs on the uplink. A simple measure to

mitigate pilot contamination is to use pilot sequences of
duration 7  18 ¼ 126 which enables a cluster of seven
cells to assign mutually orthogonal pilot sequences to all
users. As shown in Figure 13, the home cell is surrounded
by two concentric rings of non-contaminating cells. Pilot
contamination does arise from six cells in the third ring,
but the coherent interference is reduced by 40 dB or so
[20, 21]. The fraction of time spent on training is
392 ¼ :32. The remaining fraction of the slot is divided
126

evenly into uplink and downlink data transmission intervals.

2. Propagation Modeling
The propagation model comprises slow fading (long scale)
and fast fading (short scale). Slow fading accounts for
range-dependent attenuation according to the COST231
dense urban model in combination with log-normal
shadow fading. The fast fading is independent Rayleigh. FIGURE 15. Cumulative distribution of uplink per-user net through-
put (Mb/s). Red solid curve: equal power for all users; black dashed
curve: max-min power control.
3. Multiplexing Precoding/Decoding and
Power Control
Precoding and decoding consist of conjugate beamform- Each simulation involves 49 cells as shown in Figure 13
ing and matched filtering. Max-min power control gives with periodic wrap-around and a total of 18  49 ran-
equal rate service to the users in each cell. domly located users with random shadow fading. Max-
min power control parameters are derived from the slow
4. Semi-Analytical Simulations fading coefficients. As is common in cellular practice, the
Our performance calculations are based on extensions of 5% most troublesome users are dropped from service.
the capacity bound (equation 7) that account for receiver
noise, random channel instantiations, channel estimation 5. Numerical Results
error, the overhead associated with pilot transmissions, the Figures 14 and 15 show the cumulative distributions of
imperfections of conjugate beamforming and matched fil- per-user net throughput (megabits/s) for the downlink and
tering, power control, non-coherent inter-cell interference, the uplink respectively (excluding the 5% of the users
and coherent inter-cell interference (due to pilot contami- who were dropped from service).
nation) [21]. These capacity bounds only involve slow fad- Consider first the downlink results captured in
ing coefficients. One hundred simulations are performed. Figure 14. Note that max-min power control reduces the
median net throughput per user, but advantageously flattens
the cumulative distribution. The median net throughput is
about 7 Mb/s per user, and the 95% likely throughput is
4.8 Mb/s. The extraordinary thing is that each of the users
experiences the same high throughput irrespective of his
location in the cell.
Power control is even more important on the uplink
(modeled in Figure 15) than on the downlink. Users close
to the base station are in a position to disrupt the trans-
missions of users far away. With max-min power control,
median net throughput per user is about 6.5 Mb/s, and
the 95% likely throughput is 3.3 Mb/s. Note that uplink
performance is not significantly inferior to downlink per-
formance. The total median net system spectral efficiency
is 11.5 bits/s/Hz.
The system could be scaled-up in various ways. Extra
antennas would increase the SINR of every user, resulting
in higher throughput. Simultaneously doubling the num-
ber of service antennas, the number of users, and the slot
FIGURE 14. Cumulative distribution of downlink per-user net
throughput (Mb/s). Red solid curve: equal power for all users; black duration (while reducing the mobility by a factor of two)
dashed curve: max-min power control. would double the system throughput.

Bell Labs Technical Journal · V O L U M E 2 0

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 21

6. Downlink Energy Efficiency environments, or simply accommodating more users with

Total energy efficiency for the Massive MIMO base sta- TDD [37–40].
tion is estimated at 7.8 Mb/Joule, using GreenTouch
parameters for internal power consumption [22], which
comprises RF power generation, Massive MIMO comput- 3. Non-Cellular Applications
ing, and the power consumed by internal electronics. The principles of Massive MIMO are not confined to
concentrated antenna arrays, but could also be applied to
distributed (cell free) deployments. Billboard-sized arrays
Research Issues and Non-Cellular Applications
in suburban and rural locations could provide high-speed
Despite its great promise, Massive MIMO has not yet
fixed wireless access to hundreds or thousands of homes.
been reduced to practice. A number of issues still need to
Uplink Massive MIMO could transport data continuously
be tackled, and there is much ongoing research. Further,
from tens of thousands of sensors to an access point.
the potential for leveraging Massive MIMO extends be-
yond cellular deployments.
Acknowledgements
1. Issues The author thanks Hien Ngo and Hong Yang for help with
Massive MIMO replaces costly instrument-grade 40 Watt the numerical results and figures. The quality of the paper
transceivers with a large number of low-power and possi- was greatly improved due to suggestions from anonymous
bly low-precision units. Ideally each antenna would be reviewers, as well as Alexei Ashikhmin, Oliver Blume,
contained in an inexpensive module containing all elec- Thierry Klein, Erik Larsson, Iraj Saniee, and Chris White.
tronics, signal processing, and a small power amplifier Special thanks to the Editor of BLTJ, Jane DeHaven.
and as many of these modules as desired could be assem-
bled like Lego bricks. This represents an entirely new de-
sign philosophy for which a low cost solution is needed. References
[1] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski,
Massive MIMO will likely require new standards.
“Five Disruptive Technology Directions for 5G,” IEEE Commun. Mag.,
Most Massive MIMO research has focused on the vol. 52, no. 2, pp. 74–80, Feb. 2014.
physical layer. Higher layers are a possible challenge. For [2] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano,
example, when a new user joins the system, the base A. C. K. Soong, and J. C. Zhang, “What Will 5G Be?” IEEE J. Sel. Areas.
station cannot utilize the full selectivity of the array until Commun., vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
[3] G. J. Foschini and M. J. Gans, “On Limits of Wireless Communica-
after the preliminary handshaking tasks have been com-
tions in a Fading Environment When Using Multiple Antennas,” Wireless
pleted, and the user has been assigned a pilot sequence. Pers. Commun., vol. 6, no. 3, pp. 311–335, Mar. 1998.
Massive MIMO automatically provides great gains in [4] E. Telatar, “Capacity of Multi-Antenna Gaussian Channels,” Eur.
radiated energy efficiency. To achieve total energy effi- Trans. Telecommun., vol. 10, no. 6, pp. 585–595, Nov./Dec. 1999.
ciency there has to be a commensurate reduction in inter- [5] G. G. Raleigh and J. M. Cioffi, “Spatio-Temporal Coding for Wire-
less Communication,” IEEE Trans. Commun., vol. 46, no. 3, pp. 357–366,
nal power consumption, for which there has been little
Mar. 1998.
incentive in the past. [6] A. Paulraj and T. Kailath, “Increasing Capacity in Wireless Broadcast
Massive MIMO relies on favorable propagationV Systems Using Distributed Transmission/Directional Reception (DTDR),”
vector-valued channels to different users grow asymptoti- U.S. Patent 5 345 599, Sep. 6, 1994.
cally orthogonal with the increasing numbers of antennas. [7] D. Gesbert, M. Kountouris, R. W. HeathJr., C. Chae, and T. Chae,
“Shifting the MIMO Paradigm,” IEEE Signal Process. Mag., vol. 24, no. 5,
Experiments so far support this hypothesis but many
pp. 36–46, Sep. 2007.
more experiments are required. [8] G. Caire and S. Shamai, “On the Achievable Throughput of a Multi-
Although there is a great deal of theoretical and simu- Antenna Gaussian Broadcast Channel,” IEEE Trans. Inf. Theory, vol. 49,
lation evidence for the vast superiority of Massive MIMO no. 7, pp. 1691–1706, Jul. 2003.
over 4G technology, public demonstrations of Massive [9] P. Viswanath and D. N. C. Tse, “Sum Capacity of a Vector Gaussian
Broadcast Channel and Uplink-Downlink Duality,” IEEE Trans. Inf.
MIMO on a sufficiently large scale will be required to
Theory, vol. 49, no. 8, pp. 1912–1921, Aug. 2003.
convince many people of its potential. [10] S. Vishwanath, N. Jindal, and A. Goldsmith, “Duality, Achievable
Rates, Sum-Rate Capacity of Gaussian MIMO Broadcast Channels,” IEEE
Trans. Inf. Theory, vol. 49, no. 10, pp. 658–2668, Oct. 2003.
2. Massive MIMO Research Trends
[11] T. L. Marzetta, “How Much Training is Required for Multiuser
Massive MIMO is dependent upon the ability to acquire MIMO,” in Proc. 40th Asilomar Conf. Signals, Syst., Comput., Nov. 2006,
channel information, and user mobility appears to impose pp. 359–363.
a limit on the number of active users who can be served. [12] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
Considerable research is underway to mitigate these limi- O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and Chal-
lenges With Very Large Arrays,” IEEE Signal Process. Mag., vol. 30, no. 1,
tations [23–30]. Signal processing algorithms and modu-
pp. 40–60, Jan. 2013.
lation formats are attracting much attention [31–36]. [13] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, “Massive
Additional research is aimed at CSI acquisition either MIMO for Next Generation Wireless Systems,” IEEE Commun. Mag.,
with the object of making FDD work in mobile vol. 52, no. 2, pp. 186–195, Feb. 2014.

V O L U M E 2 0 · Bell Labs Technical Journal

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.
 22

[14] T. L. Marzetta, “Noncooperative Cellular Wireless With Unlimited [34] S. K. Mohammed and E. G. Larsson, “Per-Antenna Constant
Numbers of Base Station Antennas,” IEEE Trans. Wireless Commun., Envelope Precoding for Large Multi-User MIMO Systems,” IEEE Trans.
vol. 9, no. 1, pp. 3590–3600, Nov. 2010. Commun., vol. 61, no. 3, pp. 1059–1071, Mar. 2013.
[15] T. L. Marzetta, “BLAST Training: Estimating Channel [35] S. K. Mohammed and E. G. Larsson, “Constant-Envelope Multi-
Characteristics for High-Capacity Space-Time Wireless,” in Proc. 37th User Precoding for Frequency-Selective Massive MIMO Systems,” IEEE
Annu. Allerton Conf. Commun., Control, Comput., Monticello, IL, USA, Wireless Commun. Lett., vol. 2, no. 5, pp. 547–550, Oct. 2013.
1999, pp. 958–966. [36] C. Studer and E. G. Larsson, “PAR-Aware Large-Scale Multi-User
[16] E. F. W. Alexanderson, “Transoceanic Radio Communication,” Trans. MIMO-OFDM Downlink,” IEEE J. Sel. Areas Commun., vol. 31, no. 2,
Amer. Inst. Elect. Eng., vol. 38, no. 2, pp. 1269–1285, Jul.–Dec. 1919, pp. 303–313, Feb. 2013.
reprinted Proc. IEEE, vol. 72, no. 5, May 1984. [37] J. Hoydis, K. Hosseini, S. ten Brink, and M. Debbah, “Making
[17] X. Gao, O. Edfors, F. Rusek, and F. Tufvesson, “Massive MIMO Smart Use of Excess Antennas: Massive MIMO, Small Cells, TDD,” Bell
in Real Propagation Environments,” IEEE Trans. Wireless Commun., Labs Tech. J., vol. 18, no. 2, pp. 5–21, Sep. 2013.
Mar. 2014. [38] K. T. Truong, and R. W. Heath, Jr., “The Viability of Distributed
[18] H. Yang and T. L. Marzetta, “Performance of Conjugate and Zero- Antennas for Massive MIMO Systems,” in Proc. Asilomar Conf. Signals,
forcing Beamforming in Large-Scale Antenna Systems,” IEEE J. Sel. Syst., Comput., Pacific Grove, CA, USA, Nov. 3–6, 2013, pp. 1318–1323.
Areas Commun., vol. 31, no. 2, pp. 172–179, Feb. 2013. [39] E. Bjornson, M. Kountouris, and M. Debbah, “Massive MIMO and
[19] T. L. Marzetta and B. M. Hochwald, “Fast Transfer of Channel Small Cells: Improving Energy Efficiency by Optimal Soft-Cell Coordina-
State Information in Wireless Systems,” IEEE Trans. Signal Process., vol. 54, tion,” in Proc. ICT, May 2013, pp. 1–5.
no. 4, pp. 1268–1278, Apr. 2006. [40] Z. Jiang, A. F. Molisch, G. Caire, and Z. Niu, “Achieveable Rates
[20] H. Yang and T. L. Marzetta, “A Macro Cellular Wireless Network of FDD Massive MIMO Systems with Spatial Channel Correlation,” IEEE
With Uniformly High User Throughputs,” in Proc. IEEE Veh. Technol. Trans. Wireless Commun., to be published.
Conf., Sep. 2014, pp. 1–5.
[21] H. Yang and T. L. Marzetta, “Capacity Performance of Multicell
(Manuscript approved January 2015)
Large Scale Antenna Systems,” in Proc. 51st Allerton Conf. Commun.,
Control, Comput., Oct. 2013, pp. 668–675.
[22] H. Yang and T. L. Marzetta, “Total Energy Efficiency of Cellular
Large Scale Antenna System Multiple Access Mobile Networks,” in Proc. Author
IEEE Online Conf. Green Commun., Oct. 2013, pp. 27–32. Thomas L. Marzetta heads the Large-
[23] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot Scale Antenna Systems Group in the
Contamination and Precoding in Multi-Cell TDD Systems,” IEEE Trans.
Network Energy Program at Bell
Wireless Commun., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
[24] J. Choi, D.J. Love, and P. Bidigare, “Downlink Training Techniques
Labs in Murray Hill, New Jersey. He
for FDD Massive MIMO Systems: Open-Loop and Closed-Loop Training received a Ph.D. and an S.B. in elec-
With Memory,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. trical engineering from Massachu-
802–814, Oct. 2014. setts Institute of Technology in 1978
[25] K. T. Truong and R. W. Heath, Jr., “Effects of Channel Aging in
and 1972, and an M.S. in systems
Massive MIMO Systems,” J. Commun. Netw.––Special Issue Massive
MIMO, vol. 15, no. 4, pp. 338–351, Aug. 2013.
engineering from the University of Pennsylvania in 1973.
[26] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A Coordinated Ap- He worked for Schlumberger-Doll Research in petroleum
proach to Channel Estimation in Large-Scale Multiple-Antenna Systems,” exploration and for Nichols Research Corporation in
IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, Feb. 2013. defense research before joining Bell Labs in 1995. He
[27] J. Nam, J.-Y. Ahn, A. Adhikary, and G. Caire, “Joint Spatial Division
served as director of the Communications and Statistical
and Multiplexing: Realizing Massive MIMO Gains With Limited Channel
State Information,” in Proc. CISS, Princeton, NJ, USA, Mar. 2012, pp. 1–6.
Sciences Department within the former Mathematical
[28] S. Noh, M. D. Zoltowski, Y. Sung, and D. J. Love, “Pilot Beam Sciences Research Center. He is the originator of Massive
Pattern Design for Channel Estimation in Massive MIMO Systems,” IEEE MIMO, a promising technology for addressing the ever
J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 787–801, Oct. 2014. increasing demand for wireless throughput. Currently
[29] N. Shariati, E. Bjornson, M. Bengtsson, and M. Debbah, “Low
Dr. Marzetta serves as coordinator of the GreenTouch
Complexity Polynomial Channel Estimation in Large Scale MIMO With
Arbitrary Statistics,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5,
Consortium's Large-Scale Antenna Systems Project, and
pp. 815–830, Oct. 2014. as member of the advisory board of MAMMOET (Mas-
[30] K. T. Truong, A. Lozano, and R. W. Heath, Jr., “Optimal Training sive MIMO for Efficient Transmission), an EU-sponsored
in Continuous Flat-Fading Massive MIMO Systems,” in Proc. European FP7 project. Dr. Marzetta was the recipient of the 1981
Wireless Conf., Barcelona, Spain, May 2014, pp. 1–6.
ASSP Paper Award from the IEEE Signal Processing So-
[31] A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson, “On the Opti-
mality of Single-Carrier Transmission in Large-Scale Antenna Systems,”
ciety. He was elected a Fellow of the IEEE in 2003. He
IEEE Wireless Commun. Lett., vol. 1, no. 4, pp. 276–279, Aug. 2012. became a Bell Labs Fellow in 2014. For his contributions
[32] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the to Massive MIMO he received the 2013 IEEE Guglielmo
UL/DL of Cellular Networks: How Many Antennas Do We Need?” IEEE Marconi Best Paper Award, the 2013 IEEE OnLineGreen-
J. Sel. Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013.
Comm Conference Best Paper Award, the 2014 Green-
[33] H. Huh, G. Caire, H. C. Papadopoulos, and S. A. Ramprashad,
“Achieving ‘Massive MIMO’ Spectral Efficiency With a Not-so-Large
Touch Consortium 1000x Award, the 2014 Thomas Alva
Number of Antennas,” IEEE Trans. Wireless Commun., vol. 11, no. 9, Edison patent award in telecommunications, and the 2015
pp. 3226–3239, Sept. 2012. IEEE W. R. G. Baker Award. t

Bell Labs Technical Journal · V O L U M E 2 0

Authorized licensed use limited to: Athens University of Economics and Business. Downloaded on February 26,2020 at 17:10:25 UTC from IEEE Xplore. Restrictions apply.