You are on page 1of 10

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO.

2, FEBRUARY 2013 1

Inter-cell Interference in Noncooperative TDD


Large Scale Antenna Systems
Fabio Fernandes, Student Member, IEEE, Alexei Ashikhmin, Senior Member, IEEE,
and Thomas L. Marzetta, Fellow, IEEE

Abstract—In this paper we study the performance of cellular In this paper, we extend the results of the conference
networks when its base stations have an unlimited number of paper [1] by the authors of this paper. In [1], we investigate
antennas. In previous work, the asymptotic behavior of the signal the performance of the downlink of a cellular system with
to interference plus nose ratio (SINR) was obtained. We revisit
these results by deriving the rigorous expression for the SINR a large number of antennas, analyze the impact of power
of both downlink and uplink in the scenario of infinite number allocation in this setting and devise a method for mitigating
of antennas. interference. Here, we extend these results to the uplink of
We show that the contamination of the channel estimates such systems. The main aspect of this extension is the proposal
happens whenever a pilot sequence is received at a base sta- and analysis of the performance of a scheme that decreases
tion simultaneously with non-orthogonal signals coming from
other users. We propose a method to avoid such simultaneous the residual interference, including both uplink and downlink
transmissions from adjacent cells, thus significantly decreasing communications in this analysis. The techniques proposed in
interference. We also investigate the effects of power allocation in this paper are shown to increase rates overall, with the focus on
this interference-limited scenario, and show that it results in gains guaranteeing higher quality of service to even poorly located
of over 15dB in the signal to interference ratio for the scenario users.
simulated here. The combination of these two techniques results
in rate gains of about 18 times in our simulations. This paper is organized as follows. First, in Section II, we
describe our system model and assumptions. Then, in Section
Index Terms—Antenna arrays, cellular networks, mobile com- III, we derive a rigorous expression for the asymptotic SINR
munication, multiple access interference
for both the downlink and uplink of cellular networks as
the number of BTS antennas tends to infinity for the case
I. I NTRODUCTION when all users transmit their pilots simultaneously. Later in
this section, we propose a scheme where the time overlap
HE EXPONENTIAL increase in demand for high data
T rates, as well as the higher user density in cellular
networks require new ways of mitigating interference allowing
of pilots is decreased, resulting in less contamination of the
channel estimates and consequently a substantial reduction in
interference. Noting that the asymptotic expressions for the
a larger number of users to share bandwidth. This, along with
SINR are not affected by additive noise, we investigate the
the Green Touch initiative to decrease the power consumption
impact of distributed power allocation algorithms in Section
in communications networks, motivates the analysis of cellular
IV. In Section V, we present the results of numerical sim-
systems with a very large number of base station (BTS)
ulations and show that the proposed interference reduction
antennas. Such systems have been studied extensively (see
algorithms and power allocation increase data transmission
[2]- [5] and references therein).
rates approximately by the factor of 20.
In [6], the author derived estimates for SINR values in
a non-cooperative cellular network in which the number of
BTS antennas tends to infinity. It is shown in [6] that not II. S YSTEM M ODEL
all interference vanishes, and therefore, SINR does not grow We consider a cellular network composed of L hexagonal
indefinitely. The reason is that the channel estimates made cells, each consisting of a central M -antenna base station
at the BTS contain not only the desired channel vector and (BTS) and Nu single-antenna users that share the same
additive white noise, but also components directed towards bandwidth.
users in other cells who are assigned non-orthogonal pilot We assume that Orthogonal Frequency-Division Multiplex-
sequences. The numerical results obtained in [6] show that ing (OFDM) is used. Consequently, we consider a flat-fading
this type of non cooperative cellular network may provide channel model for each OFDM √ subcarrier. For a given sub-
breakthrough data transmission rates. carrier we denote by gikl = βikl hikl the channel vector
between the i-th BTS and the k-th user of the l-th cell. We
Manuscript received 1 February 2012; revised 10 June 2012. This paper
extends the results of the conference paper presented in [1]. F. Fernandes assume that the small-scale fading vectors, hikl ∼ CN (0, I),
acknowledges the financial support of CNPq, Brazil. As this paper was co- are statistically independent across users. By βikl we denote
authored by a guest editor of this issue, the review of this manuscript was the large-scale fading coefficients (log normal distribution and
coordinated by Senior Editor Wayne Stark.
F. Fernandes is the University of Campinas (UNICAMP), Campinas, Brazil geometric decay). These coefficients are constant with respect
(e-mail: fabiogf@gmail.com). to frequency and BTS antenna index.
A. Ashikhmin and T. L. Marzetta are with Bell Laboratories, We further assume frequency block fading model, in which
Alcatel-Lucent, Murray Hill, NJ 07974 USA (e-mail: aea@lucent.com,
tlm@research.bell-labs.com). the channel vector gikl is constant across Nsmooth subcarriers.
Digital Object Identifier 10.1109/JSAC.2013.1302xx. Thus each user would need to send a pilot only in one
0733-8716/13/$31.00 
c 2013 IEEE
2 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 2, FEBRUARY 2013

one would like to have several-times as many service-antennas


as terminals. In turn the number of terminals that can be
serviced simultaneously is limited by their mobility: if half of
Fig. 1. TDD transmission protocol for U = 2, K = 5, N = 1 and D = 3.
the coherence interval is used for reverse-link pilots, then the
maximum number of terminals is equal to the coherence-time
subcarrier for every Nsmooth subcarriers. So the maximum divided by twice the channel delay-spread. If high radiated
number of users per cell is Nu = K Nsmooth , where K is energy-efficiency (bits/Joule) is a priority then even greater
the number of available pilot sequences at each cell. By ρkl number of service-antennas may be desirable, since every
and Pkl we denote the pilot power and BTS transmit power doubling of M permits a reduction in total radiated power
respectively. by a factor of two.
The block fading model in the frequency domain is a rough In this asymptotic scenario, we first analyze the case where
approximation to the actual behavior of the channel gains. all users send the pilots simultaneously. Next, we propose and
However, sending only one pilot every coherence bandwidth analyze a scheme with time-shifted pilots and show that this
can be used in real situations. It provides the base stations scheme causes a significant reduction of the interference in
with evenly spaced samples allowing it to interpolate to find the asymptotic regime.
the gains for the other subcarriers. In both cases, we assume that in all cells the same
We also assume a time block fading model. Thus channel set of K orthogonal pilots of length K is used. The k -
vectors hikl stay constant during coherence blocks of T th users in all cells use the same pilot sequence ψk =
OFDM symbols. The channel vectors in different coherence (ψk1 , . . . , ψkK ), |ψkj | = 1. Since the pilots are orthogonal
blocks are assumed to be independent. The large-scale fading we have |ψk†  ψk | = Kδk,k .
coefficients are assumed to remain constant, as they change
more slowly by some orders of magnitude.
We assume reciprocity between uplink and downlink chan- A. Aligned Pilots
nels, i.e., βikl and hikl are equal for both directions. Here, we focus on the case where pilots are sent simulta-
Let us consider here a Time-Division Duplexing (TDD) neously by all users in the system. At the pilot-transmission
scheme. For every user, each coherence interval is organized stage, the i-th base station receives the signal
in four phases:
L 
 K 
• first, each user sends uplink data to its BTS for U symbol
periods. yBi = ρkl βikl hikl ψk + zi , (1)
l=1 k=1
• then, the user sends a pilot sequence of length K to its
BTS; where zi ∈ CM×K is the additive noise. Without loss of
generality, we assume that the entries of zi are i.i.d. CN (0, 1)
• the BTS then uses this pilot to estimate the corresponding random variables and that all gains are scaled accordingly.
channel vector, with which it processes the data received The i-th base station estimates the vectors gik i for users
at the uplink phase ; yBi ψk† 
located in the same cell as ĝik i = , which results in
K
• the BTS then transmits downlink data for D symbol
periods to its mobile units using the channel estimates  L
 
as beamforming vectors; ĝik i = ρk i βik i hik i + ρk l βik l hik l + zi , (2)
l=1,l=i

We assume that the estimation process takes N OFDM z ψ† 1


symbols. Therefore, each coherence interval has length where zi = iK k ∼ CN (0, K IM ).
T = U + K + N + D, as depicted in Figure 1. The base station computes the beamforming vector to its
k  -th user as the normalized version of (2):
III. A SYMPTOTIC B EHAVIOR OF SINR ĝik i ĝik i
wk  i = = √ . (3)
In this section, we analyze the asymptotic behavior of the ĝik i 
 αk i M
SINR as the number of BTS antennas M tends to infinity
while the number of users per cell Nu remains finite, constant The scalar αk i = ĝ√ikM i  is a normalization factor. To
and equal to the length of the pilot sequence K. Here we compute its value as M → ∞, we use the following well
revisit the results obtained in [6], deriving more general known lemma.
expressions for the SINR of downlink and uplink taking into
Lemma 1. Let x, y ∈ CM×1 be two independent vectors with
account all pertaining variables, such as pilot powers, base
distribution CN (0, c I). Then
station transmit powers and normalization constants.
The analysis for M = ∞ facilitates the presentation of x† y a.s. x† x a.s.
the concept of time-shifted pilots, and it constitutes a useful lim = 0 and lim = c. (4)
M→∞ M M→∞ M
limiting case. Certainly a finite-M analysis is needed to
provide first-order performance estimates of practical systems, Using the fact that the channel vectors of different users are
but that is beyond the scope of this paper. Practically speaking independent, and applying the above lemma, we can derive the
FERNANDES et al.: INTER-CELL INTERFERENCE IN NONCOOPERATIVE TDD LARGE SCALE ANTENNA SYSTEMS 3

asymptotic behavior of α2k i : Theorem 1. The downlink SINR of the k  -th user in the i-th
 L
cell is
1  2 2
lim α2k i = lim ρk l βik l hik l 2 + D
ςik  =
Pk i βik  i /αk i
, (7)
M→∞ M→∞ M L 2 2
l=1 l=1,l=i Pk l βlk i /αk l
L
 L
 
+ ρk l1 ρk l2 βik l1 βik l2 h†ik l1 hik l2 + L 1
with αk l = j=1 ρk j βlk j + K.
l1 =1 l2 =1,l2 =l1
L 
 L
  1
Note that additive noise impacts only the normalization
zi 2 + ρk l βik l h†ik l zi
a.s.
+ = ρk l βik l + . constants αk l . Therefore base station transmit powers Pkl
K
l=1 l=1 are scalable. This allows forthe use of lower power levels,
Once the pilot sequences are received and the channel resulting in a more power-efficient system.
vectors are estimated, each BTS transmit the downlink data One may think possible SINR gains can be obtained by
to its respective users. The k  -th user of the i-th cell receives choosing optimal pilot powers ρk l , since they appear in
expression (7). The following theorem shows that in fact this
L 
 K 
is not the case.
yUk i = Pkl βlk i h†lk i wkl skl + vk i , (5)
l=1 k=1 Theorem 2. If a set of SINRs can be achieved for a particular
choice of pilot powers ρkl , it can also be achieved for any
where skl is the signal intended to the k-th user in the l-th cell
other choice of ρkl .
and vk i is the unit variance additive white Gaussian noise.

Denote each term in the double sum of (5) by Qkl and let Proof: Defining the variables Pkl = Pkl /α2kl we get
Skl = E |Qkl |2 be the variance of the received signal skl .
Then D Pk  l βik
2
i
 ςik  = . (8)
  L  2
Skl = E |Qkl |2 = E Pkl βlk i |h†lk i wkl |2 |skl |2 l=1 k l βlk i
P


2
Pkl βlk i

 

L This implies that any set of downlink SINR values obtained


† † 

=
ρ kl β lkl h 
1 lk i h lkl + h  z
lk i i

for one set of normalization factors can be obtained if they are


α2kl M
1 1
changed by appropriately scaling the BTS transmit powers.
l1 =1
Uplink communication suffers similar interference gener-
Consider first the case k = k  . Applying Lemma 1 to the terms
ated by the contamination of the channel estimates. In this
inside the absolute value, we get
phase, users send data to their respective BTSs, which then
 L 
1  multiply the received signal by the channel estimates to obtain
lim ρk l1 βlk l1 h†lk i hlk l1 + h†lk i zi the desired signal as well as interference.
M→∞ M
l1 =1 The i-th BTS receives the signal
a.s. 
= ρk i βlk i . (6)
K 
L 

yBi = U β

Note that
 Sk  l is a continuous function of Pkl ikl hikl qkl + νi
L † † 
ρ k  l βlk l h  hlk l + h  z . Thus, from l=1 k=1
l1 =1 1 1 lk i 1 lk i i
the Continuous Mapping Theorem [7, p.27], we have U
where Pkl is the uplink transmit power and qkl is the uplink
2
Sk l a.s. Pk l ρk i βlk i signal sent by the k-th user of the l-th cell and νi is additive
lim =
M→∞ M 2
αk l white Gaussian noise. To decode the signal sent by its k  -
th user, the BTS applies Maximal Ratio Combining (MRC)
Now let k = k  . According to Lemma 1, Skl vanishes as by multiplying the received signal by the channel estimate
M → ∞ since hlk i , hlkl1 are independent for any l1 . obtained in (2):
In (5), the received power of the desired signal is Sk i
and all other values Skl contribute to interference. One can †
q̂k i = ĝik  i yBi
see that only the users in the neighboring cells, the k  -th L L K 
users, who use the same pilot sequence create interference  †
= U β
ρk l1 βik l1 Pkl ikl2 hik l1 hikl2 qkl2
that does not vanish as M → ∞. The reason for this is that l1 l2 =1 k=1
2

the beamforming vectors of these users contain components L  L 


 
directed towards the k  -th user in the i-th cell, generating +

U β z †h ρk l βik l h†ik l νi
Pkl ikl ikl +
directed interference that does not vanish as M → ∞. The l=1 l=1
variance of the additive noise, however, is unitary regardless 
+ zi† νi (9)
of the number of BTS antennas, thus rendering the effect of
the noise null in the asymptotic region.
Since multiplication on the receiver does not modify transmit
We obtained the asymptotic behavior of the numerator and
power, normalization is not required in the uplink.
denominator of the SINR, which are formed of independent
variables. Therefore, we proved the following theorem. The asymptotic behavior of the uplink SINR is similar to
that of the downlink. In fact, by using Lemma 1 we can rewrite
4 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 2, FEBRUARY 2013

(9) as with only two cells with K users each. Let us assume that, in
 the first phase, all users in cell 1 are sending pilot sequences
q̂k i = ρk i PkU i βik i hik i 2 qk i while all users in cell 2 are sending uplink data to their BTS.
  
signal
As a result, BTS 1 receives
L
  K 
 
+ ρk l PkU l βik l hik l 2 qk l (10) yP1 = ρ β1k1 h1k1 ψk + Pk2 β1k2 h1k2 qk2 + z1 (13)
l=1,l=i k=1
  
interference and the estimate of the k  -th user channel is
+ o(M ) yB1 ψk  z1 ψk
ĝ1k 1 = = ρ β1k 1 h1k 1 +
Consequently, the uplink SINR for the k  -th user of the i-th K K
cell is 1 K 
U ρk i PkU i βik
2
i + Pk2 β1k2 h1k2 qk2 ψk†  . (14)
ςik  = (11) K
L U 2 k=1
l=i ρk l Pk l βik l
Note that the dependence of the SINR on the pilot powers Note that each term qk2 ψk†  is a random scalar, and thus
is different for the downlink and uplink. In the case of the the channel estimate is a linear combination of the desired
downlink SINR, as seen in expression (7), the pilot powers channel vector and the channel vectors between the BTS and
appear in the normalization constants, whereas in the uplink the users in cell 2.
SINR (expression (11)), the pilot powers appear multiplying Let us assume now that, in the second phase, users in cell
the transmit powers. As a result, any attempt to assign different 1 send uplink data and users in cell 2 send pilot sequences.
pilot powers to users would impact uplink and downlink Now, BTS 1 receives
differently, thus making it impossible to optimize pilot powers K 
 
for both directions simultaneously. Moreover, in both cases, yU1 = Pk1 β1k1 h1k1 qk1 + ρ β1k2 h1k2 ψk + ν1 (15)
any effect of pilot power allocation can be achieved through k=1
transmit power allocation. Therefore, no loss is incurred in
and performs MRC to obtain
assigning a constant pilot power ρ to all users. Consequently, 
we henceforth assume that ρkl = ρ for all k, l, resulting in an †
ĝ1k  1 yU1
= ρ Pk 1 β1k 1 h1k 1 2 qk 1 (16)
uplink SINR of K

P U β 2  + ρ Pk2 β1k2 h1k2 2 qk2 ψk†  ψk + o(M ).
U
ςik = L k i ik i (12) k=1
U 2
l=i Pk l βik l
We can see that the pilot sequences sent by the users in
We also observe that the interference terms are also different cell 2 do not vanish as M → ∞, since the channels between
for the two directions of communication. In the downlink, each them and BTS 1 are present in the estimates, as shown in (14).
user receives interference from neighboring BTSs because The example above shows that the simultaneous transmis-
of the contaminated channel estimate used as beamforming. sion of pilot sequences in one cell and uplink signals in
Thus, the long-term fading coefficients that form the SINR neighboring cells does not solve the problem of interference
expression depend on the distance between each mobile user due to the contamination of channel estimates.
and the fixed neighboring BTSs. In the uplink, interference is In the remainder of this section, we propose a scheme in
due to users in other cells who use the same pilot sequence, which pilot sequences are sent simultaneously with downlink
and hence the coefficients depend on the distances between data from other cells, and we show that this superposition does
each BTS and the users from other cells. As a result, the not create interference as M → ∞.
statistical properties of the interference are different and so This scheme is based on partitioning the L cells into
are the SINR values. Nonetheless, as shown in the numerical groups of cells A1 , A2 , . . . , AΓ . Communication during each
results of Section V, these dissimilarities amount to little and coherence interval is divided in two stages. In the first stage,
the distribution of SINR values is very similar for uplink and users from cells in the group Aγ transmit their pilot sequences
downlink, a fact was previously noted in [6]. simultaneously, while users from al other groups receive
downlink data. Once users in Aγ finish their pilot sequence,
B. Time-Shifted Pilots they start receivnig downlink data while a different group
In Section III-A, we show that, as the number of BTS starts sending pilots. This protocol is repeated until users in
antennas M tends to infinity, the only cause of interference all groups transmit their pilots.Then, at this point, all users
is the superposition of non-orthogonal pilot sequences coming switch to transmitting uplink data to their BTSs.
from different users, which contaminates the channel estimates For both the downlink and uplink, the BTSs use the channel
used for spatial filtering. As a result, we devise a scheme that estimates computed in the first stage. Figure 2 depicts this
avoids such superposition. scheme for when cells are divided into three groups. Note
In fact, when a mobile unit sends a pilot sequence to its that, for any given coherence interval, the allocation of time
BTS, any non-orthogonal signal sent simultaneously from a slots to pilot, uplink and downlink data must now follow a
mobile station will cause the channel estimate to be contami- certain structure. Note that this scheme does not require a
nated. In order to illustrate this point, let us consider a scenario longer coherence interval than the scheme presented in the last
FERNANDES et al.: INTER-CELL INTERFERENCE IN NONCOOPERATIVE TDD LARGE SCALE ANTENNA SYSTEMS 5

this case is obtained by applying Lemma 1



1 
lim α2k i = lim ⎝ ρβik j hik j 2
M→∞ M→∞ M
j∈Aγ

Fig. 2. Time- shifted pilot scheme with K = 3, N = 1, D = 5 and U = 2. K
1  2


2
+ Pkl cil hil wkl  skl ψkH
+ zi 2 ⎠(20)
K
/ γ k=1
l∈A
section, but rather it establishes a different tradeoff between
The above expression is obtained by considering that the
rates, number of users in the system, and coherence interval. cross-terms vanish since they are inner products of indepen-
Here we show that, under the proposed scheme, all the dent random vectors. In order to compute the limit
in (20),
2
we focus on the terms of the form Pkl cil hil wkl 
skl ψkH
.
interference coming from cells in different groups vanishes. 2
We also discuss important issues generated by the fact that We see that αki in this case fluctuates depending on the
pilots are now transmitted simultaneously with downlink data. signals, the beamforming vectors and the transmit powers of
Let the i-th cell be in the group Aγ . Then, in the first phase, neighboring BTSs. However, its exact value is known to the
this BTS receives BTS once it computes ĝiki . Below we derive lower and upper
K bounds on αki . These bounds are important for the design of
  
yBi = ρβikj hikj ψk a power allocation scheme, discussed in section IV.
j∈Aγ k=1 First, note that the inner product between data skl and pilot
K  sequence ψk is a random variable, independent of channel

+ Pkl cil hil wkl skl + zi , (17) coefficients. We assume that the signal is composed

of

2 PSK
/ γ k=1
l∈A
symbols, and therefore bound it from above by
skl ψkH
≤ K
(similar bounds can be obtained for other types of modula-
where skl is a K × 1 vector of the intended signals to the tion).
k-th user in the l-th cell, cil ∈ C is the fading coefficient The product between channel matrix and beamforming can
2
(constant with respect to frequency and antenna indices), and also be bounded as hil wkl  ≤ |λmax (hil )|2 , where
hil ∈ CM×M is the channel matrix between antennas of the i- λmax (hil ) is the largest singular value of hil . For the case
th and the l-th base stations. The quantities cil and hil almost where there is no line of sight between the BTSs, we assume
do not change in time, since base stations do not move. So that the channel matrices have independent entries with a
we assume that cil and hil are constant in time. Gaussian distribution. As a result, the study of the distribution
The results presented below are based on the assumption of eigenvalues in random matrices in [8] show that
that for randomly chosen base stations i and l the quantities |λmax (hil )|2
cil and hil are generated as random variable with log-normal lim = 4. (21)
M→∞ M
distribution and i.i.d. Gaussian distribution respectively. We Therefore we have the following result.
obtained similar results for a Ricing fading model that ac-
counts for a line-of-sight components, which we discuss later Theorem 3. As M → ∞, we can bound α2ki by
in the section.  1
The i-th BTS uses yBi to get the estimate ρβik j + ≤ lim α2ki (22)
K M→∞
j∈Aγ

    1
yBi ψk†  ≤ ρβik j + 4Pkl cil + .
ĝik i = = ρβik j hik j K
K j∈Aγ l∈A
/ γ
j∈Aγ
K The upper bound in expression (21) is required due to the
1  
+ Pkl cil hil wkl skl ψk†  + zi (18) intricate and dynamic dependence between channel matrix and
K beamforming vectors. The bound assumes that the beamform-
/ γ k=1
l∈A
ing vector wkl coincides with the eigenvector associated with
Note that the signal received by the base station receives the largest eigenvalue of h†il hil , and is only valid when the
not only pilot signals but also downlink signals transmitted entries of the channel matrix between BTSs are assumed to
by BTS from different groups. These downlink signals are be i.i.d. Gaussian, which corresponds to a Rayleigh fading
typically more powerful than pilots. Thus it is a priori unclear model. A Rician-like fading model that accounts for a line-
whether the BTS can obtain an accurate estimate of the of-sight component, which may be realistic for the propagation
channel vector in the presence of these strong downlink between BTSs, generates channel matrices dominated by one
signals. or a few eigenvectors corresponding to the line of sight. In
The beamforming vector is defined as it was in (3): such cases, the largest eigenvalue of h†il hil would grow as
M 2 , not M . The worst case for such scenario, with the
ĝik i ĝik i ĝik i  beamforming vector aligned with the dominating eigenvector,
wk  i = = √ , i.e. αk i = √ (19)
ĝik i  αk i M M would result in a very poor channel estimate, since the
interference coming from neighboring BTSs would be stronger
The asymptotic behavior of the normalization factor αk i in than the pilot sequences by a factor of M .
6 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 2, FEBRUARY 2013

For scenarios with line of sight between neighboring BTSs, Below we find an expanded expression of the product
a small modification in the beamforming vectors avoids the h†ik i ĝik i and analyze the asymptotic behavior of each term:
potential issue described above while keeping the asymptotic  
behavior of the system unchanged. To describe this modifi- h†ik i ĝik i = ρβik j h†ik i hik j + (24)
  
cation, let us assume the extreme case where the matrix hil j∈A
γ

a
has rank 1, assuming it is formed exclusively by the line-of- K 
1  
sight component. In order to eliminate interference, the BTS + Pkl cil h†ik i hikl wkl skl ψk†  + h†ik i zi .
projects all of its beamforming vectors onto the null space of K      
/ γ k=1
l∈A

c
hil , thus forcing hil wkl = 0. The effect of this projection
is minimal: beamforming vectors would no longer have a From Lemma 1 it follows that

component perfectly aligned with the channel vector in all M

a 1 if i = j
dimensions; it would instead be aligned in M − 1 dimensions. lim =
For very large M , this effect is negligible.
M→∞ M 0 if i =
 j.
In the general case, let us assume that the channel matrix
For the analysis of the second term, we first note that skl ψk† 
between the i-th and the l-th BTSs is of the form
is a scalar whose value does not grow with M . In fact, since
we assume PSK signals, we have that skl ψk†  ≤ K.
hil = hLOS
il + hNLOS
il ,
To establish the behavior of the interaction between chan-
hil wkl
where hLOS ∈ CM×M is a deterministic matrix whose rank nel and beamforming, consider the vector qikl = h il wkl 
.
il
m does not grow with M and hNLOS is a complex Gaussian From the model adopted here, the channel vectors gik i are
il
full-rank matrix. Then, we compute the beamforming vectors independent of qikl , for i = l. Then, we have
as h†ik i hil wkl = hil wkl  h†ik i qikl , and
il ĝlkl
h̄LOS
wkl = , hil wkl |h†ik i qikl | ≤ |λmax (hil )||h†ik i qikl |.
il ĝlkl 
h̄LOS

‡ LOS ‡
According to (21) λmax (hil ) grows proportionally to M .
where h̄LOS
il = I− (hLOS
il ) hil .Here, (hLOS
il )
denotes its Denote by f the inner product between hik i ∼ CN (0, I)
Moore-Penrose pseudo-inverse. The resulting beamforming and an independent unit-norm vector qikl . It is not difficult to
vector is the projection of ĝlkl onto the null space of hLOS
il . show that f ∼ CN (0, 1), irrespective of the vector dimensions
As a consequence, the line of sight component is cancelled, and the distribution of qikl . Therefore we have
removing its interference on the channel estimates. Moreover,
this procedure does not change the asymptotic performance |
|
b K |λmax (hil )| |h†ik i qikl | a.s.
lim ≤ lim = 0. (25)
of the technique, since the only effect is that the power at M→∞ M M→∞ M
the receiver grows with (M − m)2 instead of M 2 for a The term

c allows immediate application of Lemma 1:


constant m. Therefore, this procedure allows us to consider
only the non-line-of-sight component of the channel matrices h†ik i zi a.s.
as interference during the channel estimation process, thus lim = 0.
M→∞ M
validating the uuper bound in (22).
Collecting the above results, we have that the received
After estimating corresponding channel vectors the BTSs power of the desired signal behaves as
in group Aγ transmit data to their users. Simultaneously, all
but one of the remaining groups of BTSs (Aγ  ) are also Sk  i 2 2
lim = Pk i ρ βik  i /αk i .
transmitting data to their respective users. At this stage, the M→∞ M
k  -th user of the i-th BTS receives Similarly, signals coming from other BTSs in the group Aγ
 K
 shown in expression (23) cause directed interference, whose

yUk i = Pkj βjk i h†jk i wkj skj power behaves asymptotically as
j∈Aγ k=1

2

K 
Pkj βjk i hjk i wjk sjk
a.s.
  lim 2
= Pk j ρ βjk 2
 i /αk j
+ ρ cklk i hklk i ψk + vk i M→∞ M
l∈Aγ  k=1
The pilot signals coming from users in cells of group Aγ 
 K 
 propagate through a single-input single-output (SISO) channel,
+ Pkl βlki h†lki wlk slk , (23) with finite power and not transmitted by very large antenna
l∈A
/ γ ∪Aγ  k=1
arrays. Therefore, the received power of any such user does
not grow with M , that is
where cklk i and hklk i are slow and fast fading coefficients
 
between the k-th user of the l-th cell and the k  -th user of the  ρ c  g   ψk 2 a.s.
klk i klk i
i-th cell, and vk i is the additive noise term. lim = 0. (26)
M→∞ M
The received power of the desired signal sk i is
The BTSs from groups other than Aγ and Aγ  transmit


2 Pk i βik i


2



signals using beamforming vectors uncorrelated to the channel
Sk i =
Pk i βik i h†ik i wk i
=
h ik  i ĝ ik  i
.
α2ik M vectors between them and users in the group Aγ . Therefore,
FERNANDES et al.: INTER-CELL INTERFERENCE IN NONCOOPERATIVE TDD LARGE SCALE ANTENNA SYSTEMS 7

/ Aγ ∪ Aγ  , irrespective of k, k  , we
for any i ∈ Aγ and l ∈ C. Cell Grouping
have

2 Since grouping cells and shifting the pilot sequences in time


hlk i wlk
a.s. has the result of completely avoiding interference, it can be
lim = 0. thought of as being analogous to frequency reuse, with the
M→∞ M
advantage that the entire band can be used by all cells.
Combining these results we obtain the following expression We suggest partitioning cells into groups A1 , . . . , AΓ in ex-
for the SINR at the downlink: actly the same way as they would be partitioned in frequency
2
Pk i βik 2 reuse r = Γ systems. (Usually frequency reuse systems are
 i /αk i
ςkD i = 2 2 . (27) considered only for certain integers r; we skip those details
l∈Aγ ,l=i Pk l βlk i /αk l here.) Then, in a particular OFDM subcarrier, a network with
Expression (27) is different from (7) in that only base frequency reuse r = Γ and aligned pilots will have the same
stations in the same group cause interference. This gives SINR value as a network with time-shifted pilots with Γ
significant gains in data transmission rates. groups and frequency reuse r = 1.
The same can be accomplished for the uplink by using Using this cell grouping scheme, we can compare the rates
the channel estimates obtained in (18) at the receivers. In obtained by the aligned pilot and the shifted pilot approaches.
this scheme, users from all cells transmit to their BTSs Let us first assume that no uplink data is being transmitted.
simultaneously. The signal received by the i-th BTS can be Denote by B the available bandwidth and remind ourselves
written as that Nu = KNsmooth . The time-shifted pilot method requires
that K = T /Γ. The aligned pilot method has the optimal value
K 
L 
 K = T −N 2 . Hence the data transmission rates for aligned
yUi = U β h q +ν .
Pkl pilots and shifted pilots are
ikl ikl kl i
l=1 k=1 ap BT −N T −N D
Rki =
Nsmooth (T − N − ) log2 (1 + ςik )

In order to obtain the signal sent by its k -th user, the BTS Γ 2 2
performs MRC, resulting in sp T T D
Rki =
B Nsmooth (T − N − ) log2 (1 + ςik ).
Γ Γ
  Here the factor
accounts for the effect of cyclic prefix, guard

ĝik  i yUi = ρPkU1 l βik j1 βik1 l h†ik j1 hik1 l qkl intervals, and particular modulation constellation, equal for
l∈{1,L}
  

d both schemes. The ratio of these two rates is


k1 ∈{1,K}
sp
j1 ∈Aγ Rki T (T − N − TΓ )
1   ap = 2 . (29)
+ PkD2 j2 PkU1 l cij2 s†k2 j2 ψk wk† 2 j2 h†ij2 hik1 l qk1 l Rki T −N )
K
l∈{1,L}
   2

k1 ,k2 ∈{1,K}

e
The expression above shows that the rate gain increases
j2 ∈A
/ γ
with Γ, as expected. However, the gain saturates and
+νi sp
Rik T
lim ap = 4 .
where νi denotes the terms related to additive noise that Γ→∞ Rik T −N
vanish as M → ∞. The remaining terms have the same
As an example, consider a network with T = 9 and N = 1.
stochastic properties as those in the downlink. From Lemma Rsp
Choosing Γ = 3, we obtain Rap ik
= 2.8125. This example
1, we have that ik

 shows that the time-shifted pilots allow achieving a very



d 1 if j1 = land k1 = k  significant gain in data transmission rate even for small Γ
lim = in the case where no uplink data is sent.
M→∞ M 0 if otherwise.
The analysis above is no longer valid for cases including
The terms of the form
e have exactly the same distribution as uplink data. However, a fair comparison can be made by

b in expression (24), and thus its asymptotic behavior is as considering scenarios with the same coherence interval and
obtained in expression (25). Hence, we can rewrite the above same number of users, via time-sharing. This is done in
expression as Section V, where it is explained in detail.


 
ĝik  i yUi = ρPkU j βik j hik j 2 qk j + o(M ) IV. P OWER A LLOCATION
j∈Aγ The power allocation problem for systems with SINR
expressions of the form of (7) and (27) has been extensively
The resulting uplink SINR is thus
studied [9]–[13]. Although the primary scenario of such stud-
PkU i βik
2 ies is the uplink of CDMA systems, the algorithms can be
ςkU i =
i
U β2
, (28) applied to the scenario studied in this paper as well.
P
j∈Aγ ,j=i k j ik j
We focus on setting SINR targets ςˆkl for the users and
effectively showing that, like in the downlink, the proposed allocating powers Pkl with the objective of maximizing the
scheme guarantees that only cells in the same group interfere number of users who achieve their targets under constraints
with each other. Pkl ≤ P kl . The reason for this is that in real-world systems,
8 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 2, FEBRUARY 2013

there is a possibility that not all users are able to achieve norm of the channel vector when only one group is active:
their SINR targets. In such cases, optimal algorithms for when ĝkl 
all targets are achievable become unstable and have very ξkl =
. (31)
M
poor performance. We focus here on distributed algorithms
when there are no communication between BTSs. The only Thus, for each group, one coherence interval is enough to
assumption is that each user sends to its BTS the current compute ξkl . Hence the estimation of ξkl for all BTSs in all
SINR value once every coherence interval. This is a standard groups only requires Γ coherence intervals, which is negligible
assumption since this information is also useful at the BTS when compared to the time during which coefficients βlkj stay
for other reasons, such as rate adaptation. almost constant. Thus, according to Theorem 3, we can upper
In this section Pkl is used for both uplink and downlink bound α2kl by
in order to simplify notation. Note, however, that powers are  1
allocated independently for downlink and uplink, as different ᾱ2kl = ξl + 4P̄kj clj + . (32)
K
j ∈A
/ γ
sets of powers are optimal for the two directions of commu-

nication. In practice, downlink power allocation runs at the Define the variables = Pkl /α2kl and P̄kl
Pkl 
= P̄kl /ᾱ2kl . At
BTSs while uplink power allocation is run by users, both the i-th iteration the l-th BTS computes

distributively. Pkl (i−1)
1) Rkl (i) = ςkl (i) ;

A. Aligned Pilots 2)  
P̄kl
 ςˆkl Rkl (i) if Rkl (i) ≤ ςˆkl
It is clear in expressions (7) and (12), corresponding to Pkl (i) = 
P̄kl
the SINR at downlink and uplink respectively, that only users ςˆkl Rkl (i) otherwise,
assigned the same pilot sequence interfere with each other.
3) αkl (i) = ||ĝlkl ||2 and Pkl (i) = Pkl

(i)αkl (i)2 .
As a consequence, we have one independent power allocation
problem for each different pilot sequence. Theorem 4. The above algorithm converges and at any
We adopt here a distributed algorithm proposed in [10]. iteration Pkl (i) ≤ P̄kl .
The algorithm attempts to minimize outage by using a soft  
Proof: Step 2 guarantees that Pkl (i) ≤ P̄kl at every iter-
removal criterion. The principle of soft removal is that users
ation. This corresponds by definition to Pkl (i) αkl (i)2 ≤ P̄kl .

who cannot achieve their target SINRs will be allocated
Once the power is updated at step 3, we see that Pkl (i) ≤ P̄kl .
progressively lower power. At the 0-th iteration we assume
To show that it converges, we first note that this algorithm
equal powers Pkl = P . At the i-th iteration the algorithm
applies the power allocation method in [10] to the variables
updates powers according to 
Pkl . We can rewrite the downlink SINR expression in (27) in

ςˆkl Rkl (i) if Rkl (i) ≤ P̄ςˆkl
kl terms of such variables as
Pkl (i) = P̄kl (30) Pk  i βik
2
ςkD i
otherwise, i
ςˆkl Rkl (i) =  2 . (33)
l∈Aγ ,l=i k l βlk i
P
where Rkl (i) = Pkl (i − 1)/ςkl (i).
In [10] it is shown that the above algorithm converges. We In [10], the convergence of algorithm is proven precisely
present simulation results for this algorithm in the next section. for SINR expressions with this structure.

B. Time-Shifted Pilots V. N UMERICAL R ESULTS


In the case of time-shifted pilots, the independence of the The scenarios simulated here consist of a cellular system
power allocation problem can only be claimed for the uplink, organized in hexagonal cells with radius 1.6 km, with users
as expression (28) only contains terms pertaining to users uniformly distributed in each cell, with the exception of a
assigned the k  -th pilot within group Aγ . Thus, for the uplink, circle of 100m around each BTS. The system uses a frequency
we have one independent power allocation problem for each band of B = 20 MHz and a carrier frequency of 1.9 GHz. We
pilot sequence in each group. model slow fading coefficients βikl assuming an average decay
In the downlink, however, such independence cannot be of 38dB/decade and log normal shadowing with a standard
claimed directly. The normalization factors αki , i ∈ Aγ , deviation of 8dB. We assume T = 11 and N = 1. For aligned
depend on the transmit powers of BTSs from other groups pilots we assume K = 5, D, = 3 and U = 2, and K = 3,
Aγ  , and therefore αki can vary significantly from one iteration D = 5 and U = 2 for time-shifted pilots .
to another. This may result in that transmit powers will First, we compare the SINR obtained by assuming equal
be increasing at each iteration and significantly exceed the power allocation and the distributed soft removal algorithm
constraints P kl . To avoid this undesirable behavior we propose in (30) for both uplink and downlink. We assign equal target
the following approach. SINR to all users, and vary its value. We plot in Figure 3
As a first step, the BTSs compute the components of the fraction of users who reach the target once the algorithm
the normalization that depend only on the large-scale fading converges, for frequency reuse r = 3. Due to the ”soft”
coefficients, ξkl = j∈Aγ ρβlkj , l ∈ Aγ . Note that, asymp- characteristic of the algorithm, we allow a 0.2dB tolerance,
totically, this slowly varying component is proportional to the as the SINR of the majority of the users is around (and very
FERNANDES et al.: INTER-CELL INTERFERENCE IN NONCOOPERATIVE TDD LARGE SCALE ANTENNA SYSTEMS 9

1 1
0.95
Fraction of users above SINR

0.9 Time−Shifted Pilots, r=1


0,8 Aligned Pilots, r =3
Aligned Pilots, r =1

Fraction of users above rate


0.8 Aligned Pilots, Equal Power, r=3
Aligned Pilots, Equal Power, r=1
0,6 0.7 95 % Reference
Power Allocation
Equal Power
0.6
95% Reference
0,4
0.5
−10 −5 0 5 10 15 20 25 30 35
SINR (dB)
0.4

Fig. 3. Fraction of users above SINR for r = 3. 0.3


0 10 20 30 40 50 60 70 80 90 100
Rates (Mbps)
close) to the target. We also plot the fraction of users with
Fig. 4. Fraction of users above rate for aligned and time-shifted pilots for
SINRs larger than a given value for the case of equal power. In the downlink.
this scenario, we see that, for instance, 95% of the users have
SINR above 10.5dB after power allocation, whereas the same 1

fraction of users has SINR above −5dB with equal power. A


0.9
gain of over 10dB due to power allocation is seen for other Time−Shifted Pilots, r=1
Aligned Pilots, r = 3
percentages of users. Figure 3 also shows clearly that the SINR
Fraction of users above rate
Aligned Pilots, r = 1
0.8
for both uplink and downlink are similar, a fact also noted in Aligned Pilots, Equal Power, r=3
Aligned Pilots, Equal Power, r=1
[6]. 0.7 95 % Reference
In Figures 4 and 5, we show the rate gains obtained by using
the time-shifted pilot approach compared with the aligned 0.6

pilot approach, both with power allocation, respectively for


0.5
the downlink and uplink. In both figures, we plot the fraction
of users that can achieve at least a certain rate.
0.4
A comparison between approaches with different lengths of
pilot sequences (K = 5 and K = 3 for aligned and shifted 0.3
0 5 10 15 20 25 30 35 40 45
approaches in our cases) must account for the fact that longest Rates (Mbps)
pilot sequences allow a larger number of users. In order to do
so, we consider all systems with the same number of users, Fig. 5. Fraction of users above rate for aligned and time-shifted pilots for
the uplink.
corresponding to K = 5, and assume that the time-shifted
approach accommodates the users through time sharing. As
gain of 1.8 times.
a consequence, the rates for the time-shifted approach are
Comparing the time-shifted approach to the aligned ap-
decreased by a factor of 35 .
proach for r = 1, we also see substantial gains for the worst
First, note in Figure 4 that the downlink rates corresponding
60% of the users. In fact, at the 95% level, the gain in this case
to the time-shifted approach and r = 1 is roughly 3 times
is of over 18 times( 7.39 Mbps for the time-shifted approach
larger than the aligned pilot approach with r = 3. Also
versus 0.39 Mbps for the aligned pilot approach).
consider the curve for the aligned pilot approach with r = 1,
It is worth noting that although we see substantial gains for
which shows that time-shifted pilots have provided a rate
the 95% level, the best 40% of the users perform better in
increase for all users in this scenario, even though it required
this specific scenario without pilot shifting. This behavior can
time-sharing to serve the same number of users.
be easily explained. The time-shifted approach guarantees a
We also see in Figure 4 that the rate gains of the time-shifted
distance of at least one cell to the nearest interferer, which
approach for the worst 1% − 30% and of the users is even
greatly increases the SINR of less favorably located users.
more substantial (we see a rate gain of 32 times for the 95%
Well-positioned users, however, do not get a proportional
level, from 548 Kbps to 17.6 Mbps). The gain drops when we
increase in rates even if their SINRs also increases, since
consider higher SINR users, but is still very significant (1.8
their relationship is logarithmic. Thus, the decrease in rates
times for the 50% level).
due to time sharing (by a factor of 35 in this scenario) is more
Finally, we also note that in the case of aligned pilots and
significant to those users. Nonetheless, guaranteeing a much
equal powers, for the 95% level, we have rates 0.016 Mbps
larger rate to the vast majority of the users is a substantial
and 0.68 Kbps for r = 1 and r = 3 respectively. Thus the
advantage of the proposed method.
time-shifted approach together with power allocation gives us
a 26-fold increase in data transmission rates at this level.
Figure 5 also shows significant gains for the uplink. In this VI. C ONCLUSION
case, the comparison between the shifted pilot approach with In this paper, we derive expressions for the asymptotic
r = 1 and the aligned pilot approach with r = 3 results in a behavior of the SINR in both the downlink and the uplink of a
10 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 31, NO. 2, FEBRUARY 2013

cellular network as the number of base station antennas tends [13] K.-K. Leung and C. W. Sung, “An opportunistic power control algorithm
to infinity. We show that the fundamental limitation of such for cellular network,” IEEE/ACM Trans. on Networking, vol. 14, pp. 470
–478, 2006.
networks is the interference present in the channel estimates
computed by the base stations, due to the overlapping of non-
orthogonal pilot sequences from neighboring cells. Fabio G. Fernandes received the B.Sc. and M.S.
degrees in Electrical Engineering from the Uni-
We analyze two different cases based on timing of pilot versity of Campinas (UNICAMP), Brazil, in 2006
sequences: first, when all users transmit pilots to their base and 2009 respectively. He was in the Wireless Net-
stations simultaneously; then, when the transmission of pilots working and Communications Group (WNCG) at
the University of Texas at Austin as a part of his
is shifted in time from one cell to the next, avoiding overlap. PhD program, where he held teaching and research
We see, in the latter case, that it is possible to completely assistantship positions from 2009 to 2011. He was in
cancel interference from adjacent cells, as long as the pilots do a summer internship at Alcatel-Lucent Bell Labs in
Murray Hil, NJ in 2010 and 2011. He is currently
not overlap in time. It is possible to harness SINR gains similar pursuing his PhD at the School of Electrical and
to those of frequency reuse while still sharing the same band, Computer Engineering at UNICAMP. His research interests are in the areas
therefore resulting in a substantial increase in rates. In the of Communications Theory, Information Theory and Signal Processing.
simulations, we obtained a gain of 32 times for the downlink
rate and 18 times for the uplink rate for the chosen scenario. Alexei Ashikhmin (M’00–SM’08) received the
Ph.D. degree in electrical engineering from the Rus-
We also discuss the use of power allocation algorithms in sian Academy of Science, Moscow, Russia, in 1994.
this scenario, and show that such systems greatly benefit from In 1995 and 1996, he was with the Department of
it, as communication is limited not by additive noise, but Mathematics and Computer Science, Delft Univer-
sity of Technology, Delft, The Netherlands. From
by interference from adjacent terminals. Numerical simula- 1997 to 1999, he was a Postdoctoral Fellow with
tions show that applying power allocation algorithms in this the Modeling, Algorithms, and Informatics Group,
interference-limited setting provides significant gains, on the Los Alamos National Laboratory, Los Alamos, NM.
Since 1999, he has been with Bell Laboratories,
order of 15 dB for the scenario in question. Murray Hill, NJ, where he is currently a Member of
Finally, we show that the proposed techniques are espe- Technical Staff with the Communications and Statistical Sciences Department.
cially beneficial to users in unfavorable locations that would In 2005–2007, he was an Adjunct Professor with Columbia University, New
otherwise suffer with low SINR. As a consequence, these York, NY, where he taught error correcting codes and communications theory.
His research interests include communications theory, the theory of error
techniques would result in significantly more users with high correcting codes, and classical and quantum information theory.
quality of service. Dr. Ashikhmin is a Senior Member of the IEEE Information Theory Society.
He is currently an Associate Editor for the IEEE TRANSACTIONS ON
R EFERENCES INFORMATION THEORY, which he also served from 2003 to 2006. He
is the recipient of the President’s Gold Award from Bell Laboratories in
[1] F. Fernandes, A. Ashikhmin, T. Marzetta, ”Interference Reduction on 2002 for breakthrough research that has resulted in the ability to deliver
Cellular Networks with Large Antenna Arrays” in IEEE International unprecedented wireless bit rates, the S. O. Rice Prize Paper Award from
Conference on Communications (ICC 2012), Ottawa, Canada, June the IEEE Communications Society in 2005 for work on low-density parity
2012. check codes for information transmission with multiple antennas, and the Bell
[2] T. Marzetta, ”How much training is required for multiuser MIMO?” in Laboratories Team Award for crosstalk cancellation in digital subscriber line
Proc. Asilomar Conference on Signals, Systems and Computers, Pacific systems in 2010.
Grove, CA, USA, Oct./Nov. 2006, pp.359–363.
[3] K. S. Gomadam, H. C. Papadopoulos, and C.-E. W. Sundberg, ”Tech-
niques for Multi-User MIMO with Two-Way Training,” in Proc. IEEE Thomas L. Marzetta was born in Washington,
International Conference on Communications (ICC’08), Beijing, China, D.C. He received the PhD in electrical engineer-
May 2008, pp. 4100-4105. ing from the Massachusetts Institute of Technology
[4] J. Jose, A. Ashikhmin, T. Marzetta, S. Vishwanath, ”Pilot Contamination in 1978. His dissertation extended the three-way
and Precoding in Multi-Cell TDD Systems,” IEEE Trans. on Wireless equivalence of autocorrelation sequences, minimum-
Communications, vol. 10, 2011 , pp. 2640 – 2651. phase prediction error filters, and reflection coeffi-
[5] K. Appaiah, A. Ashikhmin, T. Marzetta, ”Pilot Contamination Reduction cient sequences to two dimensions. He worked for
in Multi-User TDD Systems,” Proc. IEEE International Conference on Schlumberger-Doll Research (1978–1987) to mod-
Communications (ICC’00), South Africa, May 2010, pp.1–5. ernize geophysical signal processing for petroleum
[6] T. Marzetta, “Noncooperative cellular wireless with unlimited numbers exploration. He headed a group at Nichols Research
of base station antennas,” , IEEE Trans. on Wireless Communications, Corporation (1987–1995) which improved automatic
vol. 9, pp. 3590 –3600, 2010. target recognition, radar signal processing, and video motion detection. He
[7] P. Billingsley, Convergence of probability measures, ser. Wiley series in joined Bell Laboratories in 1995 (formerly part of AT&T, then Lucent
probability and statistics: Probability and statistics. Wiley, 1999. Technologies, now Alcatel-Lucent). He has had research supervisory re-
[8] A. Tulino and S. Verdú, ”Random matrix theory and wireless com- sponsibilities in communication theory, statistics, and signal processing. He
munications”, series Foundations and trends in communications and specializes in multiple-antenna wireless, with a particular emphasis on the
information theory. Now Publishers, 2004. acquisition and exploitation of channel-state information.
[9] S. Stanczak, M. Wiczanowski, and H. Boche, Fundamentals of Resource Dr. Marzetta was a member of the IEEE Signal Processing Society
Allocation in Wireless Networks: Theory and Algorithms, 2nd ed. Technical Committee on Multidimensional Signal Processing, a member of
Springer Publishing Company, Incorporated, 2009. the Sensor Array and Multichannel Technical Committee, an associate editor
[10] M. Rasti and A. Sharafat, “Distributed uplink power control with soft for the IEEE TRANSACTIONS ON SIGNAL PROCESSING, an associate
removal for wireless networks,” IEEE Trans. on Communications, vol. editor for the IEEE TRANSACTIONS ON IMAGE PROCESSING, and a
59, pp. 833 –843, 2011. guest associate editor for the IEEE TRANSACTIONS ON INFORMATION
[11] J. Zander, “Distributed cochannel interference control in cellular radio THEORY Special Issue on Signal Processing Techniques for Space-Time
systems,” IEEE Trans. on Vehicular Technology, vol. 41, pp. 305–311, Coded Transmissions (Oct. 2002) and for the IEEE TRANSACTIONS ON
1992. INFORMATION THEORY Special Issue on Space-Time Transmission, Re-
[12] F. Berggren, R. Jantti, and S.-L. Kim, “A generalized algorithm for ception, Coding, and Signal Design (Oct. 2003). He was the recipient of the
constrained power control with capability of temporary removal,” IEEE 1981 ASSP Paper Award from the IEEE Signal Processing Society. He was
Trans. on Vehicular Technology, vol. 50, pp. 1604 –1612, 2001. elected a Fellow of the IEEE in Jan. 2003.

You might also like