1

Energy-Efficient Power Control:

A Look at 5G Wireless Technologies
Alessio Zappone, Member, IEEE, Luca Sanguinetti, Senior Member, IEEE, Giacomo Bacci, Member, IEEE,
Eduard Jorswieck, Senior Member, IEEE, and Mérouane Debbah, Fellow, IEEE

Abstract—This work develops power control algorithms for I. I NTRODUCTION

energy efficiency (EE) maximization (measured in bit/Joule) in
wireless networks. Unlike previous related works, minimum-rate Currently, the percentage of the global world CO2 emissions
constraints are imposed and the signal-to-interference-plus-noise due to the information and communications technology (ICT)
ratio takes a more general expression, which allows one to en- is estimated to be 5% [1]. While this may seem a small
compass some of the most promising 5G candidate technologies. percentage, it is rapidly increasing, and the situation will
Both network-centric and user-centric EE maximizations are escalate in the near future with the advent of 5G networks. It
considered. In the network-centric scenario, the maximization
of the global EE and the minimum EE of the network are is anticipated that the number of connected devices will reach
performed. Unlike previous contributions, we develop centralized 50 billions by 2020 [2], and that a 1000x data rate increase is
algorithms that are guaranteed to converge, with affordable required to serve so many connected devices [3]. However, it is
computational complexity, to a Karush-Kuhn-Tucker point of also clear that obtaining the required 1000x by simply scaling
the considered non-convex optimization problems. Moreover, up the transmit powers is not possible, as it would result in an
closed-form feasibility conditions are derived. In the user-centric
scenario, game theory is used to study the equilibria of the unmanageable energy demand, and in greenhouse gas emis-
network and to derive convergent power control algorithms, sions and electromagnetic pollution above safety thresholds.
which can be implemented in a fully decentralized fashion. Both Instead, the data rate must be increased by a factor 1000,
scenarios above are studied under the assumption that single at a similar power consumption as in present networks. This
or multiple resource blocks are employed for data transmission. requires a 1000× increase of the energy efficiency (EE), i.e.,
Numerical results assess the performance of the proposed so-
lutions, analyzing the impact of minimum-rate constraints, and the efficiency with which ICT systems use energy to transmit
comparing the network-centric and user-centric approaches. data [4]. This is of paramount importance for operators (e.g.,
to save on electricity bills) and end-users (e.g., to prolong the
Index Terms—Energy efficiency, resource allocation, power lifetime of batteries) and thus has motivated a great interest
control, user-centric, network-centric, Nash equilibrium, rate
in studying and designing power control strategies taking into
constraints, QoS constraints, massive MIMO, relay networks,
hardware impairments, 5G technologies. account the cost of energy.
Power control for energy efficiency can be performed in a
centralized or decentralized manner. Both approaches are of
interest in the context of 5G networks, for which network-
centric techniques like Cloud-RAN and cooperative multi-
point (CoMP) [5], as well as user-centric techniques like
Copyright (c) 2015 IEEE. Personal use of this material is permitted. device-to-device (D2D) [6] and the use of femto-cells [7],
However, permission to use this material for any other purposes must be have been proposed. In a network-centric approach, resource
obtained from the IEEE by sending a request to [email protected].
A. Zappone and E. Jorswieck are with TU Dresden, Faculty of Electrical allocation is performed centrally and all network-nodes co-
and Computer Engineering, Communications Laboratory, Dresden, Germany operate to maximize a common system-wide performance
(email: {alessio.zappone,eduard.jorswieck}@tu-dresden.de). function. Instead, a user-centric approach is implemented in
L. Sanguinetti is with the University of Pisa, Dipartimento di Ingegneria
dell’Informazione, Pisa, Italy ([email protected]) and also with a distributed way, with the different nodes behaving in a self-
Large Systems and Networks Group (LANEAS), CentraleSupélec, Université organizing, and often competitive, fashion. Network-centric
Paris-Saclay, 3 rue Joliot-Curie, 91192 Gif-sur-Yvette, France. approaches typically grant better performance, with respect
G. Bacci is with the Mediterranean Broadband Infrastructure (MBI) srl,
Pisa, Italy (email: [email protected]). to self-organizing algorithms, but on the other hand they are
M. Debbah is with Large Systems and Networks Group (LANEAS), more complex and require a larger feedback overhead to be
CentraleSupélec, Université Paris-Saclay, 3 rue Joliot-Curie, 91192 Gif-sur- implemented. An overview of recent results in these two
Yvette, France and also with the Mathematical and Algorithmic Sciences Lab,
Huawei France R&D, Paris, France (e-mail: [email protected]). directions is provided below.
The work of A. Zappone and E. Jorswieck is funded by the German
Research Foundation (DFG), within the projects CEMRIN - grant ZA 747/1-3,
and the Collaborative Research Center 912 “Highly Adaptive Energy-Efficient A. State-of-the-art
Computing”, respectively. This research is also supported by the FP7 Network
of Excellence in Wireless COMmunications NEWCOM# (Grant agreement A non-exhaustive list of recent works dealing with EE
no. 318306). L. Sanguinetti is also funded by the People Programme (Marie optimization in centralized networks includes [8]–[15] and
Curie Actions) FP7 PIEF-GA-2012-330731 Dense4Green. The research of M. references therein. Given the non-convex, fractional nature of
Debbah have been also supported by the ERC Starting Grant 305123 MORE.
Part of this work has been presented at the IEEE International Conference EE, the main mathematical tool for centralized optimization
on Communications (ICC15), London, UK, June 8 – 12, 2015. of EE-related metrics is fractional programming – a branch of
 2

optimization theory that provides algorithms with polynomial and decentralized (user-centric) EE power allocation policies
complexity to globally maximize fractional functions with a in a wireless network in which K transmitters (possibly) share
concave numerator and a convex denominator [16]. However, N mutually orthogonal resource blocks for data transmission.
even this powerful tool fails when interference-limited net- Unlike most previous related works, we aim at maximizing
works must be optimized. This is due to the fact that the different EE metrics while satisfying minimum rate constraints
presence of multi-user interference makes the numerator of (or quality-of-service (QoS) requirements). Moreover, we as-
EE non-concave. A common way out to this problem is to sume that the signal-to-interference-plus-noise-ratio (SINR)
rely on orthogonal or semi-orthogonal transmission schemes experienced by transmitter k at its intended receiver on re-
as well as on interference cancellation techniques (to fall back source block n takes the following general form:
to the noise-limited case). Contributions in this direction are αk,n pk,n
given in [9], [14], [15]. In [9], [15] multi-carrier networks γk,n = 2 P (1)
σk,n + φk,n pk,n + j6=k ωkj,n pj,n
are considered, and the global energy efficiency (GEE) of the
system (defined as the ratio between the sum achievable rate where pk,n is the k-th user’s transmit power over resource
and the total consumed power) is optimized using orthogonal block n, whereas αk,n , φk,n , ωkj,n are positive quantities
or semi-orthogonal subcarrier allocation schemes. In [14], that do not depend on the users’ transmit powers, but only
the authors consider a multiple-antenna system and aim to on system parameters and propagation channels. In particular,
maximize the GEE when non-linear interference cancellation αk,n and φk,n are assumed to depend only on user k’s
techniques are used. However, orthogonal interference sup- channels on resource block n, while the coefficients ωkj,n
pression schemes inevitably result in a poor resource reuse depend on the other users’ channels on resource block n. The
factor and are not practical in large networks. An alternative main motivation behind the adoption of (1) is that there exist
approach consists in handling the interference by means of several communication systems and technologies in which
heuristic solutions, typically based on the use of alternating the SINR takes this form.1 Interestingly, this is the case of
optimization techniques. Examples in this context are given some candidate technologies for 5G networks, e.g., practical
by [10] in which the minimum of the individual EEs is massive MIMO networks in which the massive amount of
maximized and also by [11], [12] where both the maximization deployed circuitry prevents the use of high-quality hardware
of GEE and of the sum of the individual EEs are considered. and thus gives rise to hardware impairments [24]. The form
While these approaches can operate in interference-limited in (1) arises also when imperfect channel state information
networks, they do not guarantee convergence and/or are not (CSI) is available due to channel estimation errors. This is
supported by strong optimality claims. Moreover, they are again a typical situation in practical massive MIMO sys-
typically tailored to the maximization of specific EE metrics. tems [25]. Other relevant examples are heterogeneous, relay-
A first attempt to provide a unified framework to tackle EE assisted interference networks (e.g. multi-cell and/or small-cell
optimization problems in a centralized way is given in [13], relay-assisted MIMO networks [26], multi-cell and/or small-
where polynomial-time algorithms to optimize the GEE as cell orthogonal frequency division multiple access (OFDMA)
well as the sum and product of the individual EEs are provided. networks [20]), and relay-assisted D2D networks [27]. Fi-
As for EE maximization through decentralized solutions, nally, other well-established communication technologies are
in [17] the authors study the Nash equilibrium problem for a also included, such as ultra wide-band systems [28], or,
group of players aiming at maximizing their own EE while sat- generally, transmissions affected by inter-symbol interference
isfying power constraints in single and multi-carrier systems, and frequency-selective fading [29]. In Section II-C we will
similarly to what was done in [18] for rate maximization. A describe in detail some examples of communication systems
quasi-variational inequality approach is taken in [19], where using candidate 5G wireless technologies, in which the SINRs
power control algorithms for networks with heterogenous users take the form in (1).
are developed. In [20], [21] a similar problem is considered,
with regard to relay-assisted systems, whereas single-user C. Contributions and paper outline
multiple-input multiple-output (MIMO) systems are consid-
The major contributions of this work are as follows:
ered in [22]. However, all of these previous works do not
account for rate requirements, and so the resulting users’ rates • A unified framework for EE optimization is developed

at the equilibrium could be fairly low. Incorporating target for both centralized and decentralized networks with rate
rates changes the setting drastically since any user’s admissible and power constraints in which the users’ SINRs take the
power allocation policy depends crucially on the policies of more general expression in (1). This allows encompassing
all other users. First results in this context are provided in [23] some of the emerging technologies for 5G.
wherein Nash equilibria are found to be the fixed points of a • The maximization of the GEE as well as of the mini-

water-filling best-response operator whose water level depends mum EE is considered in the network-centric case. Both
on the rate constraint and circuit power. problems are non-convex and thus hard to solve. We
first derive closed-form feasibility conditions, and then
exploit the tools of fractional programming and sequential
B. Motivation
1 Observe that (1) includes as a special case the SINR expression typically
The aim of this paper is to develop a unified framework for encountered in wireless communication systems, which can be obtained by
the analysis and design of both centralized (network-centric) simply letting φk,n = 0 for all k, n.
 3

QK
convex optimization to develop centralized power control Accordingly, we call P , k=1 Pk the feasible set of p =
algorithms that are guaranteed to converge to a Karush- [pT1 , pT2 , . . . , pTK ]T ∈ RKN
+ .
Kuhn-Tucker (KKT) point of the non-convex problems
with affordable computational complexity.
• In the decentralized setting, the users in the network are A. Network-centric formulation
modeled as rational, self-organizing agents that engage
in a non-cooperative game wherein each one aims at Based on the user-centric EE metric (2), two relevant
maximizing its individual EE while targeting its own network-centric performance metrics are investigated in this
power and rate constraint. The existence and uniqueness work. The network GEE ψ is given by the system achievable
of Nash equilibrium points are studied and a fully dis- sum-rate over the total power consumed in the system:
tributed algorithm based on best-response dynamics is PK PN
proposed to reach equilibrium. k=1 n=1B log2 (1 + γk,n )
ψ, PK (6)
• The above scenarios are studied under the assumption pc + k=1 1T pk
that one or more resource blocks are employed for data PK
transmission. with pc = k=1 pc,k being the total circuit power dissipated
The remainder of this paper is organized as follows. Sec- in the network.2 Another important energy-efficient metric is
tion II introduces the signal model and formulates the EE the minimum of the weighted EEs, defined as
maximization problems from both network- and user-centric
perspectives. Moreover, Section II-C provides a detailed de- η, min wk ηk . (7)
k=1,...,K
scription of some 5G candidate technologies for which the
SINRs are expressed as in (1). The centralized and decentral- Within the above setting, the GEE maximization problem can
ized approaches for EE maximization in single-resource block be mathematically formulated as:
transmissions are analyzed in Sections III and IV, respectively. PK PN
The counterpart cases of multiple-resource blocks are studied ? k=1 B log2 (1 + γk,n )
n=1
in Sections V and VI. In Section VII the performance of the ψ = max ψ = max PK
p∈P p∈P pc + k=1 1T pk
proposed algorithms is numerically analyzed with reference to (8)
case-studies inspired to 5G technologies. Concluding remarks
are given in Section VIII. whereas the weighted minimum-EE maximization problem
can be written as:
II. E NERGY-E FFICIENCY P ROBLEM F ORMULATIONS
Consider a wireless interference network with K transmit- η ? = max η = max min wk ηk , (9)
ters, S receivers, and N available resource blocks (that might p∈P p∈P k=1,...,K

represent time or frequency bins) of bandwidth B in which

with {wk }k non-negative weights. As discussed later, both
the SINR of user k at its intended receiver takes the general
(8) and (9) are non-convex fractional problems, which will
form in (1). The EE ηk (measured in bit/Joule) of user k is
be tackled by means of fractional programming theory3 and
defined as the ratio of the achievable rate over the N resource
sequential convex optimization.
blocks and the total consumed power [16]
PN
B log2 (1 + γk,n ) Remark 1. Observe that ψ and η represent the two extreme
ηk , n=1 (2) points in the trade-off between global performance and fair-
pc,k + 1T pk
ness. In particular, ψ can be seen as the benefit-cost ratio
with pc,k being the circuit power dissipated to operate of the system, being defined as the ratio between the sum
the k-th transmitter and its intended receiver, and pk = achievable rate and the total consumed power. However, it
[pk,1 , pk,2 , . . . , pk,N ]T ∈ RN+ being the power allocation does not directly depend on the users’ EEs, and therefore it
vector of user k over the N resource blocks. We assume that does not allow one to tune the individual EEs according to
pk must satisfy the following (local) power constraint: users’ needs, as it might be useful in heterogeneous networks.
1T pk − pk ≤ 0 (3) On the other hand, maximizing the (weighted) minimum of the
EEs allows us to achieve a fairer resource allocation policy.
where pk denotes user k’s maximum power. Unlike most
In particular, it is known that maximizing (7) yields a Pareto-
previous related works, we assume that minimum achievable
efficient point where each quantity wk ηk is the same for all k.
rates need to be satisfied. This amounts to setting:
The whole Pareto-boundary can be simply achieved by varying
N
X the weights {wk }. However, this usually comes at the price of
log2 (1 + γk,n ) − θk ≥ 0 (4) a performance loss in terms of benefit-cost ratio of the system.
n=1
where θk is the target rate of user k (in bit/s/Hz/user). The 2 A slightly more general definition of the GEE considers the weighted sum-
feasible set of pk is thus given by: rate at the numerator. This might be useful to control the users’ individual
( N
) rates. All the following results can be straightforwardly applied to this
N T
X definition of GEE.
Pk , pk ∈ R+ : 1 pk ≤ pk , log2 (1 + γk,n ) ≥ θk . (5) 3 For completeness, a brief background on fractional programming is
n=1 provided in Appendix A.
 4

Table I
L IST OF F UNCTIONS AND VARIABLES

ηk EE of user k γ k,n SINR in absence of interference and noise

ψ GEE θk Target rate of user k
η Minimum-EE γk Minimum SINR requirement for user k
γk,n SINR of user k over block n pk Maximum power of user k
µk,n Equivalent channel gain of user k over block n pc , pc,k System and per-user hardware-dissipated power

B. User-centric formulation worst-case uncorrelated noise) of the uplink SINR of UE k in

A game-theoretic approach will be taken to solve the user- cell i takes the form [30]
centric power allocation problem formulated as: 2
PN pik E{cH
ik hiik }
n=1 B log2 (1 + γk,n ) γk = (16)
max ηk = max ∀k. (10) pik var{cH
ik hiik } + iik
pk ∈Pk pk ∈Pk pc,k + 1T pk
For later convenience, we denote by with
X
α iik = σ 2 E{||cik ||2 } + plj E{|cH 2
ik hilj | }. (17)
µk,n , 2 Pk,n (11)
σk,n + ωkj,n pj,n (l,j)6=(i,k)
j6=k
Assume that a maximum ratio combining (MRC) receiver is
the equivalent channel gain when φk,n = 0, and call
employed for data recovery. This amounts to setting cik = ĥiik
αk,n where ĥiik denotes the estimate of hiik given by
γ k,n , (12)
φk,n
hiik = ĥiik + h̃iik (18)
the SINR in the absence of interference and thermal noise (i.e.,
the maximum achievable SINR). Using the above definitions, with h̃iik being the estimation error statistically independent
(1) can be rewritten as4 of ĥiik . Assume that hilj ∼ CN (0, dilj IM ) where dilj
γ k,n µk,n pk,n accounts for the corresponding large-scale channel fading and
γk,n = (13) pathloss from UE j in cell l to BS i. If a minimum mean
γ k,n + µk,n pk,n
square error (MMSE)-based channel estimation scheme is used
or, equivalently, at the BS (with full pilot reuse) [30], then we have that
γk,n

γk,n
−1 ĥiik ∼ CN (0, ρiik IM ) and h̃iik ∼ CN (0, (diik − ρiik )IM )
pk,n = 1− . (14) where
µk,n γ k,n diik
ρiik = P (19)
Note that γk,n is a strictly increasing function of pk,n as it τ + l dilk
easily follows observing that
with τ being a given parameter that depends on the pilot
∂γk,n γ 2k,n µk,n transmit power and the pilot sequence length. Under the above
=
2 ≥ 0. (15)
assumptions, we have that
∂pk,n γ k,n + µk,n pk,n
pik αk
γk = (20)
σk2
P
+ pik φk + plj ωkj
C. Applications to 5G technologies (l,j)6=(i,k)

As mentioned in the introduction, there exist several exam- with

ples of communication technologies in which the SINR may
take the form in (1). Two case-studies are briefly detailed in αk = ρ2iik , ωkj = dilj ρiik (21)
the sequel. φk = diik ρiik +
X
ρ2ilk (22)
1) Massive MIMO: Consider the uplink (similar results can l6=i
be obtained for the downlink) of a massive MIMO system and σk2 = σ 2 ρiik . Except for the cell index (which is only
composed of S cells wherein the base station (BS) of each needed to ease understanding), the above SINR turns out to
cell uses M antennas to communicate with K single-antenna be in the same form of (1), with non-zero coefficients φk .
user equipments (UEs). Each UE is associated to a specific
Similar results can be obtained when the system is affected
serving BS while interfering with all other UEs. As such, a
by hardware impairments [24], [25]; for example, unavoidable
double index notation is used to refer to each UE as e.g., “user
clock drifts in local oscillators, finite-precision digital-to-
k in cell j”. Under this convention, let us define hilj ∈ CM as
analog converters, amplifier non-linearities, finite-order analog
the channel from UE j in cell l to BS i. Denoting by cik ∈ CM
filters, and so forth. For the sake of simplicity, let us assume
the receive combining vector of UE k at its intended BS i, a
that the hardware impairments are only at the UEs. Following
lower bound (obtained using a standard bound based on the
[25], the hardware impairments result in a reduction of the
4 Note that (as expected) when there is no self-interference (SI) (i.e., φ
k,n =
uplink signals by a factor 1 − 2 with  being the error vector
0 ∀k, n) γk,n → µk,n pk,n . magnitude, and in an additive Gaussian distortion noise which
 5

carries the removed useful power. In these circumstances, a in (1) with

lower bound of the achievable SINR can be computed as
αk,n = pr,n |hk,n |2 |cH 2 2 2
k,n hik ,n | , φk,n = σik ,n |hk,n | kck,n k
2
2  
pik (1 − 2 ) E{cH
ik hiik } ωkj,n = pr,n |cH 2 2 2
k,n hik ,n | + σik ,n kck,n k |hj,n |
2

γk =
pik (1 − 2 )var{cH 2 H 2
ik hiik } + pik  E{|cik hiik | } + iik 2
= σn2 (pr,n |cH 2 2 2
and σk,n k,n hik ,n | + σik ,n kck,n k ). We should
(23) stress that the above results apply to any interference network
in which the transmitters reach the receivers via an AF relay.
with iik given in (17). Plugging cik = ĥiik into the above
For example, the above scenario applies to relay-assisted
equation and taking into account that√in the presence of
multi-cell and small-cell networks [26], as well as to relay-
hardware impairments √ ĥiik ∼ CN (0, 1 − 2 ρiik IM ) and
assisted D2D networks [27].
h̃iik ∼ CN (0, (diik − 1 − 2 ρiik )IM ), the SINR is found
to be in the same form of (20). The above analysis can also Remark 3. Observe that in both scenarios described above,
be extended (as shown in the numerical results) to the case in the expression of the receive filters cik and ck,n impacts the
which hardware imperfections are experienced at the BS. SINR expressions through the coefficients αk , φk , ωkj and
the equivalent noise power σk2 . Therefore, any choice of the
Remark 2. An alternative approach for evaluating the SINR receive filters that does not depend on the transmit powers,
in massive MIMO systems relies on exploiting the large results in a SINR expression as in (1). This means that the
dimensions of the network. Basically, results from random developed framework can be applied to both MRC and zero-
matrix theory are used to compute an asymptotic expression forcing receivers whereas it does not readily apply to MMSE-
(known as deterministic equivalent) for (16) in the limit of based receivers as they depend on the transmit powers of UEs
K
M, K → ∞ with M ∈ (0, 1). Such a deterministic equivalent through the covariance matrix of the interference.
turns out to be in the same form of (1) and close to the effective
SINR even for a finite system [24], [30].
III. C ENTRALIZED P OWER C ONTROL IN N ETWORKS WITH
2) Relay-assisted CoMP interference network: Consider A S INGLE R ESOURCE B LOCK
the uplink of a two-hop multi-point to multi-point network
with S BSs each equipped with M antennas and using We start our analysis considering the case of a single
N subcarriers to communicate with K single-antenna UEs resource block, which, for example, models single-carrier
exploiting a single-antenna amplify-and-forward (AF) relay. systems. When N = 1, deeper analytical insights can be
(r) gained compared to the case with N > 1. In particular, the
Denoting by hk,n the channel from user k to the relay on
single-resource block scenario is analytically more tractable
subcarrier n, the received signal at the relay can be written as
and allows one to obtain necessary and sufficient feasibility
(r) √ (r)
X√ (r) conditions for the centralized energy-efficient optimization
xk,n = pk,n hk,n bk,n + pj,n hj,n bj,n + ζn(r) (24)
j6=k problems. On the other hand, only sufficient feasibility con-
ditions can be obtained in the multi-resource block setting
where bk,n is the information symbol transmitted by UE k on (Section V). This also applies to the distributed scenario in
(r)
subcarrier n and ζn ∼ CN (0, σn2 ) is the relay thermal noise. terms of more compact existence and uniqueness conditions
The total power received at the relay on subcarrier n is thus of the equilibria (Section IV). Instead, the corresponding
given by condition for the multi-resource block setting will be very
K cumbersome, and thus more difficult to handle. Moreover,
(r) 2
X
(r)
P̄n = pj,n hj,n + σn2 . (25) the techniques to be presented in this section carry over to
j=1 scenarios with N > 1, and are preparative for the more
involved multiple resource block scenario.
In order to avoid amplifier saturation, the received signal needs
(r) Setting N = 1 into (1) and (2) and neglecting the block
to be normalized by P̄n before being amplified by a factor
√ index, (8) and (9) reduce to:
pr,n and forwarded to the receivers. The signal received at PK
BS ik associated to transmitter k, over subcarrier n, is ? k=1 B log2 (1 + γk )
ψ = max ψ = max PK (28)
p∈P p∈P pc + k=1 pk
pk,n pr,n
r
(r)
yk,n = (r)
hik ,n hk,n bk,n + ik,n + wik ,n (26)
P̄n and
B log2 (1 + γk )
where hik ,n ∈ CM is the channel vector from the relay to BS η ? = max η = max min wk (29)
p∈P p∈P k=1,...,K pc,k + pk
ik on subcarrier n and ik,n ∈ CM is defined as
X r pj,n pr,n wherein p = [p1 , p2 , . . . , pK ]T ∈ RK+,
pr,n
r
(r)
ik,n = (r)
h h b
ik ,n j,n j,n + h ζ (r) (27)
(r) ik ,n n
αk pk
P̄ n P̄ n
γk = 2 P (30)
j6=k σk + φk pk + j6=k ωkj pj
whereas wik ,n ∼ CN (0, σi2k ,n IM ) is the thermal noise at
QK
and P , k=1 Pk with
receiver ik on subcarrier n. After linear reception by the filter
ck,n and upon plugging (25) into (26) the SINR takes the form Pk = {pk ∈ R+ : pk ≤ pk , log2 (1 + γk ) ≥ θk }. (31)
 6

For later convenience, we define algorithm which is guaranteed to converge to a first-order

θk optimal solution of (28). The general idea of sequential convex
γk , 2 −1 (32)
programming is to find local optima of a difficult problem
the minimum SINR requirement for user k. Observe that γ k with objective f to maximize, by solving a sequence of easier
must be such that problems with objectives {fi }i . In the generic i-th step of the
sequence, we require the following three properties:
0 ≤ γk ≤ γk ∀k. (33)
1) fi (x) ≤ f (x), for all x;
The above condition follows observing that when the noise is 2) fi (x(i−1) ) = f (x(i−1) );
negligible (i.e., σk2 → 0) and only transmitter k is active then 3) ∇fi (x(i−1) ) = ∇f (x(i−1) ).
(30) reduces to γk = αk /φk = γ k and thus the rate constraint wherein x(i−1) denotes the maximizer of fi−1 . This approach
log2 (1 + γk ) ≥ θk can be met only if (33) holds true. has been used for resource allocation in wireless networks in
[34], [35] for rate maximization, and, more recently, in [13],
A. Feasibility [26], [36] for EE maximization in different settings.
The critical issue of this approach is to find suitable approxi-
The feasibility of (28) and (29) simply amounts to verifying mations {fi }i which fulfill the listed requirements, while at the
that for given values of {αk }k , {φk }k , and {wk,j }k6=j , the same time resulting in simpler optimization problems. As far
feasible set P is not empty. Closed-form necessary and as GEE maximization is concerned, this can be accomplished
sufficient conditions for P to be non-empty are provided in by leveraging the following lower-bound of the logarithmic
the following result. function [37]. Specifically, ∀γ, γ̃ ≥ 0 we have that
Lemma 1. Let F ∈ CK×K be a matrix whose (k, j)-th log2 (1 + γ) ≥ a log2 γ + b (37)
element is defined as
( with
0 j=k
[F]k,j , ωkj γ (34) γ̃ γ̃
k
j 6= k a= b = log2 (1 + γ̃) − log2 γ̃. (38)
αk −φk γ
k
1 + γ̃ 1 + γ̃
and denote by ρF its spectral radius. The solutions to (28) The right-hand side (RHS) and left-hand side (LHS) of (37)
and (29) exist if and only if are equal at γ = γ̃, and the same holds for their derivatives
−1 with respect to γ evaluated at γ = γ̃. Therefore, we may
ρF < 1 and (I − F) s≤p (35) lower-bound ψ as follows:
where p = [p1 , p2 , . . . , pK ]T ∈ RK×1
+ and s ∈ RK
+ has
K
X K
X
elements given by [s]k , σk γ k (αk − φk γ k )−1 .
2 B [ak log2 (γk ) + bk ] B [bk + ak log2 (αk pk )]
k=1 k=1
Proof: Following the same steps of the proof of Lemma 2 ψ≥ K
= K
X X
in [31] allows us to prove the sufficiency of (35). For the pc + pk pc + pk
necessity, assume that there is a vector p0 satisfying the target k=1 k=1
rates but such that p0j ≥ pj for some j. Since γk is a stricly
  
K
increasing function of pk , we have that
X X
B ak log2 σk2 +φk pk + ωkj pj 
αk pk k=1 j6=k
< γk (36) − = ψ̃ (39)
σk2 + φk pk + j6=k ωkj p0j
P
K
X
pc + pk
for any pk ≤ pk . Otherwise stated, there exists no power pk ≤
k=1
pk such that γk = γ k . Now, suppose that ρF ≥ 1. Since F
qk
is non-negative, from [32, Theorem 2.1] it follows that there from which, letting pk = 2 , one gets
does not exist a power vector p ≥ 0 such that γk = γ k . This PK
B [bk + ak log2 (αk ) + ak qk ]
proves that (35) is also necessary. ψ̃ = k=1 PK
pc + k=1 2qk
PK h  i
2 qk qj
P
B. GEE maximization k=1 B ak log2 σk +φk 2 + j6=k ωkj 2
− PK . (40)
As described in Appendix A, fractional programming pro- pc + k=1 2qk
vides efficient tools to maximize ratios in which the numerator Using the above results, the solution to (28) can be lower
is a concave function, the denominator is a convex function, bounded as
and the constraint set is convex, whereas no low-complexity
optimization method is available if any of these properties is ψ ? ≥ ψ̃ ? = max ψ̃ (41)
q∈Q
not met. Unfortunately, the objective function in (28) does not QK
have a concave numerator, and therefore finding the global so- with q = [q1 , q2 , . . . , qK ]T , Q = k=1 Qk and
lution of (28) with affordable complexity appears difficult. To
Qk = {qk ∈ R : 2qk ≤ pk , log2 (1 + γk ) ≥ θk }. (42)
overcome this difficulty, we integrate fractional programming
theory with the framework of sequential convex programming Observe now that for any given {ak }k and {bk }k , the nu-
[33]. This allows us to develop a computationally-efficient merator and denominator of (40) are both differentiable, and
 7

Algorithm 1 Network-centric EE maximization for N = 1 show how the generalized Dinkelbach’s procedure together
1: Test feasibility by Lemma 1. with sequential convex optimization can be successfully ap-
2: if Feasible then plied to solve (29).
3: Set i = 0 and choose any p(0) ∈ P;
(0) (0) (0) (0) To begin with, observe that the min(·) function is increasing
4: Set γ̃k = γk (p(0) ) and compute ak , bk as in (38);
5: repeat so that the inequality in (37) can be used to lower-bound the
6: i = i + 1; solution to (29) as
7: if GEE then
8:
(i−1)
Solve (41) with parameters ak
(i−1)
and bk and set η ? ≥ η̃ ? = max min wk η̃k (45)
(i) q∈Q k=1,...,K
(i) (i) qk
{qk }k
= arg max ψ̃i , pk
=2 ;
9: end if where
10: if Minimum EE then [bk + ak log2 (αk ) + ak qk ]
11:
(i) (i)
Solve (45) with parameters ak and bk and set η̃k = B −
(i) (i) (i)
qk
pc,k + 2qk
{qk }k = arg max η̃ i , pk = 2 ; h  i
ak log2 σk2 + φk 2qk + j6=k ωkj 2qj
P
12: end if
(i) (i) (i)
13: Set γ̃k = γk (p(i) ) and compute ak , bk as in (38); −B (46)
14: until convergence pc,k + 2qk
15: end if and qk is still given by qk = log2 pk . Since each ratio in
(46) has a concave numerator and a convex denominator, η̃ ?
can be computed by means of the Generalized Dinkelbach’s
respectively concave5 and convex in {qk }k . Finally, the set algorithm, and the maximization of η can be tackled as in
Qk can be shown to be convex for all k. Indeed, the k-th rate Algorithm 1, whose convergence is stated in the following
constraint can be equivalently rewritten as proposition and proved in Appendix C.
 
X Proposition 2. Algorithm 1 monotonically increases the value
2qk (αk − γ k φk ) ≥ γ k σk2 + ωkj 2qj  . (43) of η and converges to a point fulfilling the KKT conditions of
j6=k the epigraph-form representation of the original problem (29).
Since (33) must hold true, one gets (applying the logarithm Remark 4. Algorithm 1 can be straightforwardly specialized
function to both sides) to maximize the system sum rate and the minimum of the users’
  ! rates, since these two metrics coincide with the numerator of
2
X
qj 
αk − γ k φk the GEE and with the minimum of the numerators of the users’
qk − log2 σk +
 ωkj 2 +log2 ≥ 0 (44)
γk EEs, respectively.
j6=k

which turns out to be a convex constraint. As a consequence, IV. D ISTRIBUTED P OWER C ONTROL IN N ETWORKS WITH
(41) is a fractional problem which can be globally and A S INGLE R ESOURCE B LOCK
efficiently solved by means of fractional programming tools
A decentralized power control algorithm looks for the
[16], such as the Dinkelbach’s algorithm [39]. This leads to the
solution of the following coupled problems [17], [20], [23]:
general iterative procedure formulated in Algorithm 1 whose
convergence is proved in Appendix B. arg max ηk (pk , p−k ) ∀k (47)
pk ∈Pk (p−k )
Proposition 1. Algorithm 1 monotonically increases the GEE T
value and converges to a point fulfilling the KKT conditions where p−k = [p1 , . . . , pk−1 , pk+1 , . . . , pK ] is the inter-
of the original problem (28). ference vector containing all powers except user k’s, and
Pk (p−k ) is defined as in (31). This problem can be formulated
as the non-cooperative game in normal form:
C. Weighted Minimum EE Maximization
The key difference between (28) and (29) is that the G , {K, {Pk (p−k )}k , {ηk (pk , p−k )}k } (48)
objective function η in (29) involves K fractional functions where (in game theory parlance) K = [1, 2, . . . , K] is the
{ηk } rather than a single one. This makes (29) fall within set of players, Pk (p−k ) is player k’s strategy set, uk (p) =
the framework of generalized fractional programming, which ηk (pk , p−k ) is player k’s utility function. The K coupled
studies the maximization of functions of multiple ratios. In problems in (47) define the best-response dynamics (BRD)
this more general scenario, Dinkelbach’s algorithm fails, even of the game, while the solution of the k-th problem in (47) is
assuming that each ratio {ηk } has a concave numerator and a the k-th player’s best-response to the other players’ choices.
convex denominator. Instead, the problem can be tackled using More formally, let us define the best response Bk (p−k ) of
an extension of Dinkelbach’s algorithm known as Generalized player k to an interference vector p−k (or, equivalently, µk as
Dinkelbach’s algorithm (see Appendix A), which is guaranteed easily follows from (11)) as
to converge to the global solution of a max-min fractional
problem with limited complexity, provided each ratio has a Bk (p−k ) , arg max ηk (pk , p−k ) . (49)
pk ∈Pk (p−k )
concave numerator and a convex denominator [40]. Next, we
Any fixed point of the BRD is a Nash equilibrium of the game.
5 Recall that the log-sum-exp function is convex [38]. In general a non-cooperative game might admit zero, one, or
 8

more equilibria, and even if one or more equilibria exist, the 2) The sets Pk (p−k ) vary continuously with p−k (in the
convergence of the BRD is not guaranteed. As a consequence, sense that the graph of the set-valued correspondence
crucial issues in the analysis of a non-cooperative game are p−k 7→ Pk (p−k ) is closed).
to establish the existence and uniqueness of an equilibrium, 3) Each user’s payoff function ηk (pk , p−k ) is quasi-concave
and whether implementing the BRD eventually yields an in pk for all p−k ∈ P−k .
equilibrium. In our scenario, answering these questions is In our setting, if the sufficient condition (50) is satisfied, then
more challenging due to the fact that, unlike regular non- the sets Pk (p−k ) are nonempty, convex,6 closed and bounded
cooperative games, not only the utility functions, but also for every p−k . Moreover, each of them varies continuously
the players’ strategy sets are mutually coupled, depending with p−k since the rate constraint log2 (1 + γk ) ≥ θk in
on the other players’ actions p−k . A similar non-cooperative Pk (p−k ) is itself continuous in p−k . Finally, following [20]
game is termed a generalized non-cooperative game, and ηk (pk , p−k ) is proved to be strictly pseudo-concave since it
more restrictive conditions have to be fulfilled for a (unique) is given by the ratio between a strictly concave and an affine
generalized Nash equilibrium (GNE) to exist and for the BRD function. Since any strictly pseudo-concave function is also
to converge. To begin with, the following result is given: quasi-concave [16], the third condition is fulfilled.
Lemma 2. If The following result shows that a unique GNE exists, and
that the BRD always converges to such point.
σk2 +
P
j6=k ωkj pj
pk ≥ γ k ∀k (50) Proposition 4. The game G admits a unique GNE point,
αk − φk γ k
which can be obtained by starting from any feasible power
then Bk (p−k ) takes the form vector {pk }K
k=1 and iteratively updating the transmit powers
n n oo according to (51).
Bk (p−k ) = min pk , max p?k , pk (51)
Proof: The proof builds upon the standard function frame-
wherein work [42], which states that a non-cooperative game admits
−1
σk2 +
P
γk γk ωkj pj a unique equilibrium (reachable by iteratively computing the

j6=k
pk (p−k ) , 1− = γk (52) players’ best-responses) provided the game admits at least
µk γk αk − φk γ k
one equilibrium and the best-response function is a standard
and function.7 Since we have already shown that the game admits
p?k , arg max ηk (pk , p−k ) . (53) a GNE (see Proposition 3), we are left with proving that (51)
pk ∈R+ is a standard function for all k. Towards this end, p?k (p−k ) is
Proof: The first part of the thesis easily follows from proved to be standard in [20, Appendix A]. As for the function
rewriting the rate constraints γk ≥ γ k (using (30)) as pk (p−k ) in (52), it is non-negative because γ k ≤ γ k , and it
also fulfills the monotonicity property because it is increasing
σk2 + j6=k ωkj pj
P
in all {pj }j6=k . As for the scalability property, take any β > 1,
pk ≥ γ k . (54)
αk − φk γ k then it holds
2
σk
Since pk ≤ pk for all k ∈ K, then
P
β + j6=k ωkj pj
pk (βp−k ) = βγ k (56)
σk2 + j6=k ωkj pj σk2 + j6=k ωkj pj αk − φk γ k
P P
γk ≥ γk . (55)
σk2 + j6=k ωkj pj
P
αk − φk γ k α k − φk γ k
< βγ k = βpk (p−k ) . (57)
Hence, if ∀k ∈ K (50) holds, then there always exists a power αk − φk γ k
pk ∈ [0, pk ] such that γk ≥ γ k is fulfilled. The last part of the Finally, since both p?k (p−k ) and pk (p−k ) are standard func-
proof follows by leveraging [20], where it is shown that for any tions, and since pk does not depend on p−k , we may conclude
given p−k , ηk is unimodal and thus admits a unique maximizer that (51) is also a standard function because both max(·) and
pk ∈ R+ . Accounting for the power and rate constraints and min(·) are increasing functions.
imposing (50) eventually yields (51).

A. Analysis of the Equilibria B. Distributed implementation

The existence and uniqueness of the GNE points of G are The best response of a generic player k is characterized in
now studied under the assumption that (50) holds. the sequel to come up with an iterative algorithm that allows
each player to reach the GNE in a distributed manner. Toward
Proposition 3. The game G admits a nonempty set of GNE this end, let us first define
points.  +
x
Proof: Observe that the existence of a GNE is guaranteed νk (x) , γ k 1 + (γ k − gk (x)) (58)
2Bµk
under the following assumptions [41]:
1) The players’ feasible action sets Pk (p−k ) are nonempty, 6 Note that the constraint function log2 (1 + γk ) is concave in pk .
7 Recall
closed, convex, and contained in some compact set Ck that a vector function g(p) is standard if it fulfills the properties
Q of i) non-negativity: g(p) ≥ 0 for all p; ii) monotonicity: g(p1 ) ≥ g(p2 )
for all p−k ∈ P−k ≡ `6=k P` . for all p1  p2 ; iii) scalability: g(βp) < βg(p), for all p and β > 1.
 9

Algorithm 2 Iterative algorithm to solve (47). V. C ENTRALIZED POWER CONTROL IN NETWORKS WITH
1: initialize i = 0 and ∀k pk [0] ∈ R+ in the feasible set MULTIPLE RESOURCE BLOCKS
2: repeat
3: for k = 1 to K do In this section, we turn our attention to the case in which
4: receive γk [i] from the serving access point each transmitter can use multiple resource blocks, i.e., N > 1.
5: compute µk [i] using (64) Differently from the case in which N = 1, the rate constraints
6: use µk [i] to update pk [i] in (63) in (8) and (9) are not in a convex form. Nevertheless, the
7: use µk [i] in (61) to run the Dinkelbach’s algorithm methodology used in Section III can be successfully extended
8: set λ?k [i] equal to the Dinkelbach’s output and update the
to find sufficient feasibility conditions and to derive low
power as:
n n oo complexity algorithms that converge to KKT points.
pk [i + 1] = min pk , max πk (λ?k [i]) , pk [i]
A. GEE maximization
9: end for
10: update i = i + 1 Following the same steps of Section III-B, we leverage the
11: until convergence lower bound in (37) to obtain
K X
X N
B [bk,n + ak,n log2 (αk,n ) + ak,n qk,n ]
and k=1 n=1
r ψ ≥ ψ̃ = K X
N
−
4Bµk
X
qk,n
gk (x) , γ 2k + (1 + γ k ) (59) pc + 2
x k=1 n=1
  
with µk and γ k given by (11) and (12), respectively. K X
X N X
2
Lemma 3. For any given p−k (or, equivalently, µk ), the B ak,n log2 σk,n +φk,n 2qk,n + ωkj,n 2qj,n
k=1 n=1 j6=k
solution to (53) is found to be (65)
−1 K X
N
νk (λ?k ) νk (λ?k )
 X
qk,n
? ?
pk = πk (λk ) , 1− (60) pc + 2
µk γk k=1 n=1

where λ?k is obtained through the Dinkelbach’s algorithm as with qk,n = log2 pk,n . Although the numerator and denom-
the solution of the following equation: inator of ψ̃ in (65) are again jointly concave and convex in
{qk,n }k,n (as in the case of a single resource block), the rate
Blog2 (1 + νk (λ?k )) − λ?k (pc,k + πk (λ?k )) = 0. (61) constraints in (8) are not in a convex form yet, due to the sum
Proof: The proof is given in Appendix D. over the multiple resource blocks. To overcome this issue, we
Denote by pk [i] the transmit power of the k-th player at the resort to the same trick and lower-bound the LHS of the rate
i-th iteration step. By virtue of Proposition 4 and Lemma 3, constraint by the concave function,
it follows that an iterative algorithm operating according to N
X N
X
n n oo log2 (1 + γk,n ) ≥ [bk,n + ak,n log2 (αk,n ) + ak,n qk,n ]
pk [i + 1] = min pk , max πk (λ?k [i]) , pk [i] (62) n=1 n=1
  
N
where pk [i] is computed as (using (52)) X
2
X
− ak,n log2 σk,n +φk,n 2qk,n + ωkj,n 2qj,n = R̃k .
γk γ k −1
 
n=1 j6=k
pk [i] = 1− (63)
µk [i] γk (66)
converges to the unique GNE of G, with µk [i] being the Finally, we may write
equivalent channel gain in (11) at the i-th iteration step. The ψ ? ≥ ψ̃ ? = max ψ̃ (67)
pseudo-code is reported in Algorithm 2. q∈Q

A close inspection of (58) – (61) and (63) reveals that the QK

with Q = k=1 Qk and Qk being now given by
computation of pk [i+1] through (62) only requires knowledge ( N
)
of µk [i]. Although not available at the k-th terminal, this X
Qk = qk ∈ R N : 2qk,n ≤ pk , R̃k ≥ θk . (68)
information can be easily acquired taking into account that:
n=1
 −1
γk [i] γk [i] The solution to the above problem can be computed by means
µk [i] = 1− (64)
pk [i] γk of Dinkelbach’s algorithm, which leads to the power allocation
procedure illustrated in Algorithm 3. Using similar arguments
where γk [i] denotes the SINR of transmitter k measured at
as in Proposition 1, the following result can be proved:8
its intended receiver at iteration i. Since pk [i] and γ k are
locally available at the transmitter, the computation of µk [i] Proposition 5. Algorithm 3 monotonically increases the GEE
only requires knowledge of γk [i]. The latter can be easily value and converges to a point fulfilling the KKT conditions
estimated at the receiver and sent back to the corresponding of the original non-convex problem (8).
transmitter via a return downlink channel. Therefore, besides 8 As observed in the proof of Proposition 2, the sequential convex optimiza-
being guaranteed to converge to the unique GNE, Algorithm 2 tion tool allows us to find a KKT point of the original problem also when the
can also be implemented in a fully decentralized fashion. constraint functions are lower-bounded together with the objective.
 10

Algorithm 3 Network-centric EE maximization for N > 1 conditions as in Section III can in principle be derived if per-
1: Set i = 0 and choose a starting point p(0) ; resource-block QoS constraints are considered. This amounts
(0) (0) (0)
2: For any k and n, set γ̃k,n = γk,n (p(0) ) and compute ak,n and to enforcing log2 (1 + γk,n ) ≥ θk,n with θk,n being such that
(0) PN
bk,n as in (38); n=1 θk,n = θk , for all k, with {θk }k given target rates.
3: if (67) or (71) is unfeasible then
4: break
5: end if VI. D ISTRIBUTED P OWER C ONTROL IN N ETWORKS WITH
6: repeat M ULTIPLE R ESOURCE B LOCKS
7: i = i + 1;
8: if GEE then As done for the single resource block case, we define
(i−1) (i−1)
9: Solve (67) with parameters ak,n and bk,n and set γ k,n h x
i +
(i) (i) q
(i) νk,n (x) , 2µk,n + γ k,n − gk,n (x) (72)
{qk,n } = arg max ψ̃i , pk,n = 2 k,n ; 2µk,n B
10: end if
11: if Minimum EE then and
(i) (i)
Solve (71) with parameters ak,n and bk,n and set
r
12: 4Bµk,n
γ 2k,n +


(i) (i)
{qk,n } = arg max η̃ i , pk,n = 2 ;
(i)
qk,n gk,n (x) , 1 + γ k,n . (73)
x
13: end if
14:
(i) (i) (i)
γ̃k,n = γk,n (p(i) ) and compute ak,n and bk,n as in (38); Let us also define pk as the power vector minimizing the
15: until convergence transmit power while satisfying the rate constraints. Mathe-
matically, we have that:
pk , arg max 1T pk (74)
pk ∈RN
B. Weighted minimum EE maximization +

N
A similar approach as in Section V-A can be used to solve subject to
X
log2 (1 + γk,n ) − θk ≥ 0
(9). In particular, exploiting the fact that the min(·) function n=1
is increasing, using (37), and setting pk,n = 2qk,n allows us
from which (using the same arguments of Appendix D) one
to lower-bound η as
gets:
η≥ min wk η̃k = η̃ (69) pk,n = πk,n (λk ) ∀n (75)
k=1,...,K

where with λk being such that:

PN N
B [bk,n + ak,n log2 (αk,n ) + ak,n qk,n ]
n=1
X
η̃k = PN − log2 (1 + νk,n (λk )) − θk = 0 (76)
pc,k + n=1 2qk,n n=1
PN h  i
2 qk,n qj,n
P
n=1 B ak,n log 2 σ k,n + φ k,n 2 + j6=k ωkj,n 2 and
PN .  −1
pc,k + n=1 2 qk,n νk,n (x) νk,n (x)
πk,n (x) , 1− . (77)
(70) µk,n γ k,n
Then, we have that Lemma 4. For any given p−k , the entries of Bk (p−k ) ∈ RN
+
are found to be
η ? ≥ η̃ ? = max min wk η̃k . (71) n n oo
q∈Q k=1,...,K
[Bk (p−k )]n = min pk,n , max πk,n (λ?k ) , pk,n (78)
Observe that η̃k in (70) has a concave numerator and a
where λ?k is obtained through the Dinkelbach method as the
convex denominator, meaning that (71) can be globally solved
solution of the following equation:
by means of the Generalized Dinkelbach’s algorithm. The !
resulting power allocation procedure is given in Algorithm 3. XN XN
? ? ?
By a similar reasoning as in Proposition 2, the following result Blog2 (1 + νk,n (λk ))−λk pc,k + πk,n (λk ) = 0.
can be proved. n=1 n=1
(79)
Proposition 6. Algorithm 3 monotonically increases the η Proof: The proof relies on similar arguments (omitted for
value and converges to a point fulfilling the KKT conditions space limitations) of those in Appendix D for N = 1.
of the epigraph-form representation of the original non-convex Observe that (78) can be equivalently rewritten as (since
problem (9). πk,n (·) in (77) is a strictly decreasing function):
Remark 5 (Feasibility of (8) and (9)). Observe that Line 3 [Bk (p−k )]n = πk,n (λ0k ) (80)
in Algorithm 3 is also a sufficient feasibility test for (8) and
(9) since both are guaranteed to be feasible provided (67) with
and (71) are feasible, which can be checked by means of a
λ0k , max λk , min {λ?k , λk }

(81)
convex feasibility test. Moreover, this implies that (67) and

(71) will remain feasible for all iterations of Algorithm 3. and λk such that ∀n πk,n λk = pk,n . A sufficient condition
Observe also that similar necessary and sufficient feasibility for (80) (or, equivalently, for (78)) to be a contraction mapping
 11

Algorithm 4 Algorithm to reach the GNE of G for N > 1.

1
1: initialize i = 0 and ∀k, n pk,n [0] ∈ R+ in the feasible set
2: repeat 0.9
3: for k = 1 to K do
4: receive {γk,n [i]}n from the intended receiver 0.8
5: compute {µk,n [i]}n using (84)
6: use {µk,n [i]}n to compute λk in (76) by means of the 0.7
Dinkelbach algorithm 0.6
7: use λk to update {pk,n [i]}n in (75)
use {µk,n [i]}n in (79) to run the Dinkelbach algorithm 0.5

Pf
8:
9: set λ?k [i] equal to the Dinkelbach output and update the
power as: 0.4
n n oo
pk,n [i + 1] = min pk,n , max πk,n (λ?k [i]) , pk,n [i] 0.3
R = 15%
0.2
10: end for R = 20%
11: update i = i + 1 0.1 R = 25%
12: until ∀k, n pk,n [i] = pk,n [i − 1] R = 30%
0
-40 -35 -30 -25 -20 -15 -10
P [dBW]
is provided in the following proposition (see Appendix E for
the proof), thus ensuring the existence and uniqueness of a Figure 1. K = 5; M = 50; τ = 10−2 . Probability of feasibility Pf versus
GNE for G when N > 1, and the convergence of the BRD. P with minimum per user-rate constraints: (a) R = 15%; (b) R = 20%;
(c) R = 25%; (d) R = 30%.
Proposition 7. The EE game G admits a unique GNE and the
BRD converges to such equilibrium whenever, for all k
K XN
  hardware impairments are present at the UEs. However, we
 X h γ k,n i 2 
assume that hardware impairments are present at the BS, and
sk,n 1n=` +
X
2
ωkj,n sup ξk,` < 1 (82)
j=1 n=1 µk ∈Ωk ?
gk,n  we denote by BS the resulting error magnitude. As observed
`,n∈Sk
j6=k in Section II-C1, this scenario results again in an SINR as in
QN µ2 ∂(1/λ0k ) (1). Indeed, following [43] and by similar steps as in Section
where Ωk , n=1 (0, αk,n ] and ξk,` , αk,` k,` ∂µk,`
whereas II-C1, the k-th UE’s SINR is written as in (1) with
Sk? , {n = 1, . . . , N : µk,n > λ0k } and sk,n is defined as 2
αk = hH
k hk , φk = 2BS hH
k Dk hk (85)
gk,n (λ0k ) − (2 + γ k,n ) γ µk,n
sk,n , γ k,n
− 0 k,n . (83) ωkj = |hH
k hj |
2
+ 2BS hH
k Dj hk (86)
2αk,n 1 + γ k,n λk αk,n gk,n (λ0k )
Similarly to the single resource block case, denote by pk,n [i] with Dj = diag({|hj (m)|2 }M 2 2 H
m=1 ) and σk = σ hk hk .

the transmit power of the k-th player over block n at the i-th All channels are generated according to the Rayleigh fading
iteration step and, accordingly, compute µk,n [i] as follows: model with path-loss model as in [44]. All UEs have the
 −1 same maximum feasible power pk = P ∀k and hardware-
γk,n [i] γk,n [i] dissipated power pc,k = 10 dBm ∀k. The receive noise power
µk,n [i] = 1− (84)
pk,n [i] γk is σ 2 = F BN0 , with F = 3 dB and N0 = −174 dBm/Hz.
The minimum rate constraint θk is set as a percentage Rk %
with γk,n [i] being the SINR of transmitter k over block n
of the maximum rate that user k can achieve when pk → ∞,
measured at its intended receiver at iteration i. From the
while the other users’ powers are finite, namely:
results of Proposition 7, it follows that the iterative procedure
illustrated in Algorithm 4 converges to the unique GNE of G,
 
Rk Rk αk
and can be implemented in a fully decentralized fashion. θk = log2 (1 + γ k ) = log2 1 + (87)
100 100 φk

VII. N UMERICAL R ESULTS For simplicity, we assume Rk % = R% for all users k.

We begin by analyzing the feasibility probability Pf of the
Numerical results are now given to assess the performance
EE maximization problems as a function of P for different
of the proposed solutions. To this end, two case-studies
values of R. The results are obtained by averaging over 5 · 104
are considered, namely, a hardware-impaired massive MIMO
independent scenarios of users’ drops and channel coefficients.
system and a multi-carrier relay-assisted interference network.
As seen, Pf approaches 1 for realistic values of P up to
25% of the maximum rate, whereas for R = 30% the typical
A. Hardware-Impaired Massive MIMO System transmit power levels which are used in the uplink of present
Consider the uplink of a massive MIMO system as described cellular systems, are not enough to ensure a high Pf .
in Section II-C1, with K = 5, S = 1, and M = 50. The Next, we analyze the system performance in terms of
communication bandwidth is B = 1 MHz, and MRC is used EE and data rate. Fig. 2 shows the average GEE for the
for data recovery under the assumption of perfect channel following resource allocation policies: (a) Algorithm 1 with
estimation. We also set  = 1, thereby considering that no R = 20%; (b) Algorithm 1 without QoS constraints, i.e.
 12

Table II
K = 5; M = 50; τ = 10−2 . AVERAGE NUMBER OF REQUIRED ITERATIONS TO REACH CONVERGENCE VERSUS Pmax FOR : ( A ) A LGORITHM 1 FOR GEE
MAXIMIZATION WITH R = 20%; ( B ) A LGORITHM 1 FOR GEE MAXIMIZATION WITH R = 0%; ( C ) A LGORITHM 2 FOR DISTRIBUTED RESOURCE
ALLOCATION WITH R = 20%; ( D ) A LGORITHM 2 FOR DISTRIBUTED RESOURCE ALLOCATION WITH R = 0%.

Maximum power P [dBW]

P = −38 P = −34 P = −30 P = −26 P = −22 P = −18 P = −14 P = −10
Algorithm 1. R = 0% 2.63 3.69 4.68 6.30 6.53 6.49 6.50 6.51
Algorithm 1. R = 20% 2.63 3.67 4.87 6.68 6.70 6.76 6.76 6.77
Algorithm 2. R = 0% 1 1.01 1.07 1.42 2.54 3.66 4.12 4.50
Algorithm 2. R = 20% 1 1.01 1.07 1.42 2.54 3.67 4.19 6.71

300 250

250
200

200
150

[Mbit/J/user]
[Mbit/J]

150

100
100

(a) Algorithm 1. R = 20%

(b) Algorithm 1. R = 0% 50
50 (a) Algorithm 1. R = 20%
(c) Algorithm 1 for sum-rate maximization (b) Algorithm 1. R = 0%
(d) Maximum power allocation (c) Algorithm 1 for min-rate maximization
0 0
-40 -35 -30 -25 -20 -15 -10 -40 -35 -30 -25 -20 -15 -10
P [dBW] P [dBW]

Figure 2. K = 5; M = 50; τ = 10−2 . Average achieved GEE versus P Figure 3. K = 5; M = 50; τ = 10−2 . Average achieved minimum users’
for: (a) Algorithm 1 for GEE maximization with R = 20%; (b) Algorithm 1 energy efficiency versus P for: (a) Algorithm 1 for Min-EE maximization
for GEE maximization with R = 0%; (c) sum-rate maximization by adapted with R = 20%; (b) Algorithm 1 for Min-EE maximization with R = 0%;
Algorithm 1; (d) Maximum transmit power allocation. (c) Min-rate maximization by adapted Algorithm 1.

R = 0%; (c) Algorithm 1 specialized to maximize the sum- whereas it increases to 2.35 [bit/s/Hz/user] when scheme (a)
rate; (d) Maximum power allocation, i.e. pk = P for all k, is used.
considered as a baseline scheme. In scheme (a), if the problem Similar considerations can also be made when Algorithm 1
turns out to be unfeasible, the QoS constraint is relaxed and is used to maximize the minimum of the users’ EEs. Fig.
the solution from scheme (b) is taken. For low values of P , 3 compares the minimum EE (with wk = 1 for all k)
this circumstance is very frequent, and indeed schemes (a) and versus P , achieved by: (a) Algorithm 1 with R = 20%; (b)
(b) perform similarly. At the same time, schemes (b) and (c) Algorithm 1 with R = 0%; (c) min-rate maximization by
also perform similarly for low P , thus suggesting that in this adapted Algorithm 1. The results show a similar behavior as
range of P , GEE and sum-rate maximization are equivalent. for Fig. 2.
Instead, different performance are achieved for larger values Fig. 4 compares the GEE performance of the centralized
of P . Indeed, increasing P eventually allows attaining the Algorithm 1 and of its distributed counterpart Algorithm 2 for
peak of the GEE. At this point, the GEE achieved by scheme R = 0% and 20%. We observe that, while the centralized
(b) remains constant, as using the excess transmit power scheme suffers a little performance gap in terms of GEE
would only decrease the GEE. Indeed, the GEE achieved when QoS constraints are introduced, having minimum rate re-
by scheme (c) decreases, because this scheme makes use of quirements causes a larger GEE degradation in the distributed
the excess transmit power to maximize the sum-rate. Instead, scenario, especially for increasing P . This is expected because
scheme (a) strikes a balance between these two extremes. unlike the centralized scheme, in the distributed setting the
Some of the excess power is used to fulfill the QoS constraints, interference among the users is not jointly managed, which
which results in a slightly lower GEE. However, once the results in high multi-user interference, especially for large P .
constraints are met, the transmit power is not further increased Table II shows the average number of iterations required by
and the achieved GEE keeps constant. We remark that this Algorithms 1 and 2 to converge when R = 0% and R = 20%.
slight reduction of the GEE grants a higher minimum rate. Convergence is declared when kq(i) − q(i−1) k2 /kq(i) k2 ≤
In particular, in the saturation region of the GEE, the average 10−4 . It is seen that both algorithms converge after a small
minimum rate granted by scheme (b) is 1.6 [bit/s/Hz/user], number of iterations, which slightly increases for larger values
 13

350
220
300 200
180
250 160
140
200
120

[Mbit/J]
[Mbit/J]

100
150
80

(a) Cooperative GEE. R = 20%

100 (a) Algorithm 1. R = 20% 60 (b) Cooperative GEE. R = 0%
(b) Algorithm 1. R = 0% (c) Distributed GEE. R = 20%
(c) Algorithm 2. R = 20% (d) Distributed GEE. R = 0%
(d) Algorithm 2. R = 0% (e) Cooperative sum-rate
40
-40 -35 -30 -25 -20 -15 -10 -40 -35 -30 -25 -20 -15 -10
P [dBW] P [dBW]

Figure 4. K = 5; M = 50; τ = 10−2 . Average achieved GEE versus P Figure 5. K = 3; N = 16, M = 3. Average achieved GEE versus P for:
for: (a) Algorithm 1 for GEE maximization with R = 20%; (b) Algorithm 1 (a) Algorithm 3 for GEE maximization with R = 20%; (b) Algorithm 3 for
for GEE maximization with R = 0%; (c) Algorithm 2 for distributed resource GEE maximization with R = 0%; (c) Algorithm 4 for distributed resource
allocation with R = 20%; (d) Algorithm 2 for distributed resource allocation allocation with R = 20%; (d) Algorithm 4 for distributed resource allocation
with R = 0%. with R = 0%; (e) sum-rate maximization
. by adapted Algorithm 3

of P . This is because increasing P results in a larger feasible C. Computational complexity discussion

set. Observe that the distributed algorithm exhibits faster con-
The computational complexity of the proposed algorithms
vergence than the centralized one. This makes it particularly
depends on the number of iterations required to reach con-
suitable for self-organizing networks.
vergence as well as on the complexity of each iteration.
A theoretical analysis for the number of iterations to reach
convergence is very challenging. In fact, no theoretical results
B. Relay-assisted OFDMA interference network are available also for simpler scenarios than energy efficiency
maximization. To overcome this problem, we have resorted to
Consider a relay-assisted multi-cell network as described in a numerical analysis. From Table II, it follows that the number
Section II-C2. Assume S = 3, N = 16, K = 3, M = 3, and of outer iterations for the single-resource-block case is very
B = 180 kHz. The UEs are placed at a distance from the relay limited for both centralized and distributed algorithms (the
randomly generated in the interval [100; 300] m and the same latter require a slightly lower number of iterations). Similar
path-loss model and noise parameters as for previous figures results are obtained also in the multi-resource-block setting.
are used. The complexity of each iteration is analyzed in the following.
Fig. 5 reports the GEE versus P achieved by the following The centralized algorithms are considered first.
resource allocation algorithms: a) Algorithm 3 with R = 20%; 1) Centralized Algorithms: We begin by observing that
b) Algorithm 3 with R = 0%; c) Algorithm 4 with R = 20%; the complexity of the centralized algorithms depends on the
d) Algorithm 4 with R = 0%; e) Algorithm 3 specialized to complexity of each fractional problem to solve. To fix ideas,
maximize the sum-rate. If the problems with QoS constraints let us consider the maximization of the GEE – similar consid-
happen to be infeasible, the solution of the corresponding erations apply to the maximization of the Min-EE. Although
scheme with R = 0% is taken. Similar remarks can be made there exist several different methods to solve fractional pro-
as for the massive MIMO system. In particular, by introducing gramming problems, the most widely used is the Dinkelbach’s
QoS constraints we trade-off a slight reduction in GEE with algorithm. As shown in Algorithm 5, Dinkelbach’s algorithm
a significant increase of the users’ minimum rate, which, works by finding the solution of an auxiliary problem in each
for the simulated scenario, increases from 3.8 [bit/s/Hz/user] iteration, say x∗n , and then updating the parameter λ as λn =
when R = 0%, to 7.16 [bit/s/Hz/user] when R = 20%. As f (x∗n )/g(x∗n ). One well-known result of the Dinkelbach’s
for the comparison between the centralized and decentralized algorithm is that the sequence {λn }n converges with a super-
approaches, we observe that the gap is rather limited in this linear rate. Moreover, this convergence rate does not depend on
scenario. This might be due to both the channel diversity the complexity required to compute x∗n . In this respect, observe
granted by the use of multiple sub-carriers and to the rather also that the auxiliary problems to obtain {x∗n } are convex and
limited number of non-cooperating transmitters. Indeed, it is therefore can be solved with polynomial complexity in the
known that that a limited number of players tend to result in number of variables and constraints (which are KN and 2K,
a lower price-of-anarchy in game-theoretic settings [45]. respectively). Putting all these facts together, we can conclude
 14

that the centralized algorithms have a polynomial complexity Definition 1 (Fractional program). Let C ⊆ Rn be a convex
in each stage. This makes them of affordable complexity. set, and consider the functions f : C → R and g : C → R+ .
2) Decentralized Algorithms: As for the distributed case, A fractional program is the optimization problem
a similar analysis can be performed since the proposed al-
f (x)
gorithms consist of an outer loop, which requires to solve max . (88)
x∈C g(x)
at each stage K fractional problems with N variables and 2
constraints. Following the above arguments, we can state that Proposition 8 ([39]). An x∗ ∈ C solves (88) if and only if
the algorithms have a polynomial complexity in each stage. x∗ = arg maxx∈C {f (x) − λ∗ g(x)}, with λ∗ being the unique
Observe that the complexity of each stage does not depend zero of F (λ) = maxx∈C {f (x) − λg(x)}.
on K (as in the centralized case) but only on the number of
This result allows us to solve (88) by finding the zero of
resource blocks N and on the number of constraints (which
F (λ). An efficient algorithm to do so is the Dinkelbach’s
is two). The parameter K only scales the complexity of each
algorithm [39], reported in Algorithm 5 for the reader’s conve-
iteration in a multiplicative way.
nience. If f (x) and g(x) are concave and convex, respectively,
VIII. C ONCLUSION then the Dinkelbach’s algorithm requires to solve one convex
problem in each iteration.9 Moreover, the convergence rate of
In this paper, we have proposed a framework to develop
Dinkelbach’s algorithm is known to be super-linear [39].
centralized and decentralized power control algorithms for
EE optimization in wireless networks. Unlike most previous
Algorithm 5 Dinkelbach’s algorithm
related works, we have considered rate constraints and a more
Set ε > 0; λ = 0;
general SINR expression so as to encompass emerging 5G repeat
technologies. The resulting optimization problems have been x∗ = arg maxx∈C {f (x) − λg(x)}
tackled by an interplay of fractional programming, sequential F = f (x∗ ) − λg(x∗ );
convex optimization, and game theory. This has allowed us λ = f (x∗ )/ g(x∗ );
to derive centralized algorithms achieving first-order optimal until F ≤ ε
points of the GEE and of the minimum of the users’ EEs,
and to develop decentralized algorithms that have been shown A considerable extension of (88) is to consider the maxi-
to converge to the GNE of the associated game. The analysis mization of the minimum of a set of ratios {fi (x)/gi (x)}Ii=1 .
above has been performed in the case in which a single or This problem is usually referred to as generalized fractional
multiple resource blocks are used for transmission. Numerical programming, and has been first studied in [40], wherein an
results have been used to corroborate the theoretical results. optimization procedure is provided, based on a modification
For this purpose, two case-studies have been considered: a of Dinkelbach’s algorithm. Specifically, the auxiliary function
hardware-impaired massive MIMO network and a multi-cell to be considered is F (λ) = min1≤i≤I {fi (x) − λgi (x)}, and
multi-carrier relay-assisted network. the algorithm works as shown in Algorithm 6.
The numerical analysis indicates that the centralized algo-
rithms are quite robust to the enforcement of demanding rate Algorithm 6 Generalized Dinkelbach’s algorithm
constraints. Instead, the distributed algorithms are more sen- Set ε > 0; λ = 0;
sitive to rate constraints, especially for increasing maximum repeat
feasible powers, due to the lack of centralized interference x∗ = arg maxx∈C {min1≤i≤I [fi (x) − λgi (x)]};
management. Also, the centralized algorithms perform better F = min1≤i≤I {fi (x∗ ) − λgi (x∗ )};
λ = min1≤i≤I fi (x∗ )/ gi (x∗ );
than their distributed counterparts, both with and without until F < ε
rate constraints, at the expense of a higher computational
complexity and feedback requirements.
Then, if each ratio has a concave numerator and a convex
An important follow-up of this work is to analyze the
denominator, we can solve the generalized fractional problem
gap between the proposed centralized framework and the
by solving a sequence of convex problems,10 with a linear
global optimum of the GEE and minimum EE. To this end,
convergence rate [40].
an optimization framework is required, which allows one to
effectively determine the global solution of energy-efficient
optimization problems in interference-limited networks. This A PPENDIX B
is a challenging problem, which is still much open. An attempt In the i-th iteration of Algorithm 1, we solve (41) and
in this direction is taken in [46], wherein tools borrowed from compute the corresponding vector p(i) of transmit powers,
monotonic optimization theory, fractional programming, and which maximizes the lower-bound ψ̃i at the i-th iteration, as
sequential optimization are combined together to characterize given by (39), subject to the same constraints of the original
the global maximum of the energy efficiency in interference- problem (28). Then, the following chain of inequalities holds
limited wireless networks.
ψ(p(i) ) ≥ ψ̃i (p(i) ) ≥ ψ̃i (p(i−1) ) = ψ(p(i−1) ) (89)
A PPENDIX A
9 It is also required that λ ≥ 0 in each iteration. This can be shown to always
Major concepts from fractional programming are reviewed hold if the algorithm starts with λ = 0, and provided maxx f (x) ≥ 0.
here. For a more comprehensive overview, we refer to [16]. 10 Observe that the minimum of concave functions is also concave [38].
 15

wherein the first inequality follows because ψ̃i is a lower- From (15), using (14) one gets ∂γk /∂pk = µk (1 − γγk )2 , from
k
bound of ψ, the second inequality follows because p(i) is the which (92) reduces to:
maximizer of ψ̃i , while the final equality holds because the  2
(i) (i) Bµk γk
parameters ak and bk in ψ̃i are such that the bound is tight 1− − λk = 0. (93)
in p (i−1)
. As a consequence of (89), the value of ψ increases 1 + γk γk
after each iteration, and the algorithm must converge because In the attempt of solving (91), let us study the properties
ψ is upper-bounded. of (93) as a function of γk . This amounts to studying
Next, let us denote by ψ̃ and p̃ the lower-bound of ψ and the fk (γk ) = 0 with fk (γk ) = Bµk γk2 − 2Bµk γ k + λk γ 2k γk +


power vector at convergence. By construction, p̃ maximizes γ 2k (Bµk − λk ). The first derivative of fk (γk ) with respect to
ψ̃ subject to the same constraints of the original problem γk is
(28), and therefore fulfills the associated KKT conditions. ∂fk (γk )
= 2Bµk (γk − γ k ) − λk γ 2k < 0 (94)
Such KKT conditions are the same as the KKT conditions ∂γk
of the original problem (28), except for the different objective
where the last inequality follows since 0 ≤ γk ≤ γ k , and
function. However, upon convergence of Algorithm 1, we have
µk ≥ 0, λk > 0, and γ k > 0. As a consequence, fk (γk ) is
ψ̃(p̃) = ψ(p̃) and ∇ψ̃(p̃) = ∇ψ(p̃), and the thesis follows.
a strictly decreasing function, which admits a solution if and
A PPENDIX C only if fk (0) ≥ 0 and fk (γ k ) ≤ 0 (with the equalities not
simultaneously active). Since fk (γ k ) = −λk γ 2k (1 + γ k ) < 0,
The first part of the proof follows along the same lines of
we need to ensure fk (0) = γ 2k (Bµk − λk ) ≥ 0. This trans-
Proposition 1. As for the KKT conditions, we remark that in
lates into Bµk ≥ λk . By solving the second-order equation
this case we can not directly consider the KKT conditions of
fk (γk ) = 0, one gets γk = νk (λk ) where νk (x) is defined in
(29), because the objective of (29) is not differentiable. We
(58). Plugging γk = νk (λk ) into (14) yields (60).
can remove the non-differentiability by expressing (29) in its
equivalent epigraph form:
A PPENDIX E
max t subject to ηk ≥ t ∀k. (90) 2
P µk,n in (11) as µk,n = αk,n /(σk,n + Ik,n ) where
Rewrite
{t∈R+ ,p∈P}

Now, let us consider a modified version of Algorithm 1, which Ik,n = j6=k ωkj,n pj,n . Using [17, Theorem 4], the GNE is
operates on (90) rather than on (29). In each iteration of this unique if k∂Ik /∂p−k k supIk ∈RN +
kBk (p−k )/∂Ik k < 1, for all
modified Algorithm 1, (37) is again used to lower-bound ηk k. with Ik = [Ik,1 , . . . , Ik,N ]T . From [17, Eq. (19)], we have
with η̃k . This amounts to solving an approximation of (90), v
u K
(i) N
with η̃k in place of ηk , and updating the parameters ak u X X
2
(i) k∂Ik /∂p−k k = t ωkj,n (95)
and bk as in Algorithm 1. Following similar arguments as
j=1,j6=k n=1
in the proof of Proposition 1, it follows that this modified s
version of Algorithm 1 converges to a point fulfilling the N P
N
2
|∂πk,n (λ0k ) /∂Ik,` | . (96)
P
k ∂Bk (p−k )/ ∂Ik k =
KKT conditions of (90), the only difference being that now the `=1 n=1
lower-bound is computed for the constraint function ηk rather
than for the objective function. However, since ηk ≥ η̃k , the Observe now that πk,n (λ0k )
> 0 when µk,n > λ0k (see
solution of the approximate problem is always in the feasible Appendix D). After lengthy computations (not shown for space
set of (90). Moreover, upon convergence, ηk and η̃k are equal, limitations), one gets
and so are their gradients. Therefore, the modified version ∂p?k,n h γ k,n i
of Algorithm 1 yields a power vector fulfilling the KKT = sk,n 1n=` + 0 ξk,` 1µk,n >λ0k (97)
∂Ik,` gk,n (λk )
conditions of (90). Observe now that replacing ηk with η̃k in
(90) yields the epigraph form of (45), i.e., the problem which where we have defined (for notational compactness) sk,n as
is solved in each iteration of Algorithm 1 by the Generalized in (83), whereas ξk,` is defined in the text of Proposition 7.
Dinkelbach’s method. Hence, Algorithm 1 and its modified Plugging (97) into the RHS of (96) yields
version (considered in this proof) converge to the same power v
uX X h γ k,n i2
allocation vector p̃, and hence the thesis. sk,n 1n=` + ξk,` (98)
u
0
t
? ?
gk,n (λk )
`∈Sk n∈Sk
A PPENDIX D
Using [8, Sect. II.A], it follows that solving the EE problem where Sk? , {n = 1, . . . , N : µk,n > λ0k }. Putting all the
(47) is equivalent to finding the root of the nonlinear function above results together, Proposition 7 follows.
Φ(λk ) = max {Blog (1 + γk ) − λk (pc,k + pk )} (91)
pk ∈R+
R EFERENCES
where λ?k ∈ R+ is such that Φ(λ?k ) = 0 and can be [1] A. Fehske, J. Malmodin, G. Biczók, and G. Fettweis, “The global
obtained through the Dinkelbach’s algorithm. Setting to zero footprint of mobile communications: The ecological and economic
the derivative of (91) with respect to pk yields: perspective,” IEEE Commun. Mag., pp. 55–62, Aug. 2011.
[2] Ericsson White Paper, “More than 50 billion connected devices,” Eric-
B ∂γk sson, Tech. Rep. 284 23-3149 Uen, Feb. 2011.
− λk = 0. (92) [3] “The 1000x data challenge,” Qualcomm, Tech. Rep.
1 + γk ∂pk
 16

[4] “Reducing the net energy consumption in communications networks by [26] A. Zappone, E. A. Jorswieck, and S. Buzzi, “Energy efficiency and
up to 90% by 2020,” GreenTouch Green Meter Research Study, Tech. interference neutralization in two-hop MIMO interference channels,”
Rep., June 2013. IEEE Transactions on Signal Processing, vol. 62, no. 24, pp. 6481–
[5] D. Gesbert, S. Hanly, H. Huang, S. Shamai Shitz, O. Simeone, and 6495, December 2014.
W. Yu, “Multi-cell MIMO cooperative networks: A new look at inter- [27] Y. Pei and Y.-C. Liang, “Resource allocation for device-to-device com-
ference,” IEEE J. Sel. Areas Commun., vol. 28, no. 9, Dec. 2010. munications overlaying two-way cellular networks,” IEEE Transactions
[6] F. Boccardi, R. W. H. Jr., A. Lozano, T. L. Marzetta, and P. Popovski, on Wireless Communications, vol. 12, no. 7, pp. 3611–3621, July 2013.
“Five disruptive technology directions for 5G,” IEEE Communications [28] G. Bacci, M. Luise, H. V. Poor, and A. M. Tulino, “Energy efficient
Magazine, vol. 52, no. 2, pp. 74–80, February 2014. power control in impulse radio UWB wireless networks,” IEEE J. Sel.
[7] J. Hoydis, M. Kobayashi, and M. Debbah, “Green small-cell networks,” Topics. Signal Process., vol. 1, no. 3, pp. 508–520, Oct. 2007.
IEEE Vehicular Technology Magazine, vol. 6, no. 1, pp. 37 – 43, March [29] S. Buzzi, V. Massaro, and H. V. Poor, “Energy-efficient resource
2011. allocation in multipath CDMA channels with band-limited waveforms,”
[8] C. Isheden, Z. Chong, E. Jorswieck, and G. Fettweis, “Framework IEEE Trans. Signal Process., vol. 57, no. 4, pp. 1494–1510, April 2009.
for link-level energy efficiency optimization with informed transmitter,” [30] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
IEEE Trans. Wireless Commun., vol. 11, no. 8, pp. 2946 – 2957, 2012. of cellular networks: How many antennas do we need?” IEEE Journal
[9] D. W. K. Ng, E. S. Lo, and R. Schober, “Energy-efficient resource al- on Selected Areas in Communications, vol. 31, no. 2, pp. 160 – 171,
location in multi-cell OFDMA systems with limited backhaul capacity,” February 2013.
IEEE Trans. Wireless Commun., vol. 11, no. 10, pp. 3618–3631, Oct. [31] G. Bacci, E. Belmega, and L. Sanguinetti, “Distributed energy-efficient
2012. power optimization in cellular relay networks with minimum rate
[10] B. Du, C. Pan, W. Zhang, and M. Chen, “Distributed energy-efficient constraints,” in IEEE Intl. Conf. Acoustics, Speech and Signal Processing
power optimization for CoMP systems with max-min fairness,” IEEE (ICASSP), Florence, Italy, May 2014, pp. 7014 – 7018.
Commun. Lett., vol. 18, no. 6, pp. 999–1002, Jun. 2014. [32] E. Seneta, Non-negative Matrices and Markov Chains, 3rd ed. New
[11] S. He, Y. Huang, S. Jin, and L. Yang, “Coordinated beamforming York, NY: Springer, 2006.
for energy efficient transmission in multicell multiuser systems,” IEEE [33] B. R. Marks and G. P. Wright, “A general inner approximation algorithm
Trans. Commun., vol. 61, no. 12, pp. 4961–4971, Dec. 2013. for nonconvex mathematical programs,” Operations Research, vol. 26,
[12] S. He, Y. Huang, L. Yang, , and B. Ottersten, “Coordinated multicell no. 4, pp. 681–683, July-Aug. 1978.
multiuser precoding for maximizing weighted sum energy efficiency,” [34] M. Chiang, C. Wei, D. P. Palomar, D. O’Neill, and D. Julian, “Power
IEEE Trans. Signal Process., vol. 62, no. 3, pp. 741–751, Feb. 2014. control by geometric programming,” IEEE Trans. Wireless Commun.,
[13] L. Venturino, A. Zappone, C. Risi, and S. Buzzi, “Energy-efficient vol. 6, no. 7, pp. 2640–2651, Jul. 2007.
scheduling and power allocation in downlink OFDMA networks with [35] L. Venturino, N. Prasad, and X. Wang, “Coordinated scheduling and
base station coordination,” IEEE Trans. Wireless Commun., vol. 14, power allocation in downlink multicell OFDMA networks,” IEEE Trans.
no. 1, pp. 1–14, 2015. Veh. Technol., vol. 58, no. 6, pp. 2835–2848, Jul. 2009.
[36] D. Nguyen, L.-N. Tran, P. Pirinen, and M. Latva-aho, “Precoding for
[14] J. Xu and L. Qiu, “Energy efficiency optimization for MIMO broadcast
full duplex multiuser MIMO systems: Spectral and energy efficiency
channels,” IEEE Trans. Wireless Commun., vol. 12, no. 2, pp. 690 –
maximization,” IEEE Transactions on Signal Processing, vol. 61, no. 16,
701, Feb. 2013.
pp. 4038 – 4050, Aug 2013.
[15] Q. Xu, X. Li, H. Ji, and X. Du, “Energy-efficient resource allocation
[37] J. Papandriopoulos and J. Evans, “Low-complexity distributed algo-
for heterogeneous services in OFDMA downlink networks: Systematic
rithms for spectrum balancing in multi-user DSL networks,” in Proc.
perspective,” IEEE Trans. Veh. Technol., vol. 63, no. 5, pp. 2071–2082,
IEEE Intl. Conf. Commun. (ICC), Istanbul, Turkey, Jun. 2006, pp. 3270–
June 2014.
3275.
[16] A. Zappone and E. Jorswieck, “Energy efficiency in wireless networks [38] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, UK:
via fractional programming theory,” Foundations and Trends R in Com- Cambridge Univ. Press, 2004.
munications and Information Theory, vol. 11, no. 3-4, pp. 185–396, [39] W. Dinkelbach, “On nonlinear fractional programming,” Management
2015. Science, vol. 13, no. 7, pp. 492–498, Mar. 1967.
[17] G. Miao, N. Himayat, G. Li, and S. Talwar, “Distributed interference- [40] J.-P. Crouzeix and J. A. Ferland, “Algorithms for generalized fractional
aware energy-efficient power optimization,” IEEE Trans. Wireless Com- programming,” Mathematical Programming, vol. 52, no. 1-3, pp. 191–
mun., vol. 10, no. 4, pp. 1323–1333, Apr. 2011. 207, 1991.
[18] G. Scutari, D. Palomar, and S. Barbarossa, “Optimal linear precoding [41] G. Debreu, “A social equilibrium existence theorem,” Proceedings of
strategies for wideband noncooperative systems based on game theory the National Academy of Science, vol. 38, pp. 886—893, 1952.
part I: Nash equilibria,” IEEE Trans. Signal Process., vol. 56, no. 3, pp. [42] R. D. Yates, “A framework for uplink power control in cellular radio
1230 – 1249, March 2008. systems,” IEEE J. Sel. Areas Commun., vol. 13, no. 7, pp. 1341–1347,
[19] I. Stupia, L. Sanguinetti, G. Bacci, and L. Vandendorpe, “Power control September 1995.
in networks with heterogeneous users: A quasi-variational inequality [43] E. Björnson, M. Matthaiou, and M. Debbah, “Massive MIMO
approach,” IEEE Trans. Signal Process., vol. vol. 63, no. 21, pp. 5691– with arbitrary non-ideal arrays: Hardware scaling laws and circuit-
5705, November 2015. aware design,” CoRR, vol. abs/1409.0875, 2014. [Online]. Available:
[20] A. Zappone, Z. Chong, E. Jorswieck, and S. Buzzi, “Energy-aware com- http://arxiv.org/abs/1409.0875
petitive power control in relay-assisted interference wireless networks,” [44] G. Calcev, D. Chizhik, B. Goransson, S. Howard, H. Huanga, A. Ko-
IEEE Trans. Wireless Commun., vol. 12, no. 4, pp. 1860–1871, Apr. giantis, A. Molisch, A. Moustakas, D. Reed, and H. Xu, “A wideband
2013. spatial channel model for system-wide simulations,” IEEE Trans. Veh.
[21] F. Shams, G. Bacci, and M. Luise, “Energy-efficient power control Technol., vol. 56, no. 2, March 2007.
for multiple-relay cooperative networks using Q-learning,” IEEE Trans. [45] S. Lasaulce and H. Tembine, Game Theory and Learning for Wireless
Wireless Commun., vol. 14, no. 3, pp. 1567 – 1580, Mar. 2015. Networks: Fundamentals and Applications, ser. Academic Press. Else-
[22] E.-V. Belmega and S. Lasaulce, “Energy-efficient precoding for multiple- vier, 2011.
antenna terminals,” IEEE Trans. Signal Process., vol. 59, no. 1, pp. [46] A. Zappone, E. Björnson, L. Sanguinetti, and E. A. Jorswieck, “Globally
329–340, Jan. 2011. optimal resource allocation for energy efficiency in wireless networks,”
[23] G. Bacci, E. Belmega, P. Mertikopoulos, and L. Sanguinetti, “Energy- IEEE Transactions on Signal Processing, submitted, 2015.
aware competitive power allocation for heterogeneous networks under
QoS constraints,” IEEE Trans. Wireless Commun., vol. 14, no. 9, pp.
4728 – 4742, Sept 2015.
[24] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive
MIMO systems with non-ideal hardware: Energy efficiency, estimation,
and capacity limits,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112–
7139, Nov 2014.
[25] E. Björnson, E. G. Larsson, and M. Debbah, “Massive MIMO for
maximal spectral efficiency: How many users and pilots should be
allocated?” IEEE Trans. Wireless Commun., 2015. [Online]. Available:
http://arxiv.org/pdf/1412.7102.pdf