360 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO.

3, MARCH 2015

Energy Harvesting Wireless Communications:

A Review of Recent Advances
Sennur Ulukus, Member, IEEE, Aylin Yener, Fellow, IEEE, Elza Erkip, Fellow, IEEE,
Osvaldo Simeone, Senior Member, IEEE, Michele Zorzi, Fellow, IEEE, Pulkit Grover, Member, IEEE, and
Kaibin Huang, Senior Member, IEEE

Abstract—This paper summarizes recent contributions in the breaking away from conventional battery recharging, and an
broad area of energy harvesting wireless communications. In ability to deploy wireless networks at hard-to-reach places such
particular, we provide the current state of the art for wireless as remote rural areas, within concrete structures, and within
networks composed of energy harvesting nodes, starting from the
information-theoretic performance limits to transmission schedul- the human body. As such, energy harvesting wireless networks
ing policies and resource allocation, medium access, and net- will make it possible to develop new medical, environmental,
working issues. The emerging related area of energy transfer for monitoring/surveillance and safety applications which are oth-
self-sustaining energy harvesting wireless networks is considered erwise impossible with conventional battery-powered operation.
in detail covering both energy cooperation aspects and simulta- There are several different natural sources and associated
neous energy and information transfer. Various potential models
with energy harvesting nodes at different network scales are re- technologies for energy harvesting: solar, indoor lighting, vi-
viewed, as well as models for energy consumption at the nodes. brational, thermal, biological, chemical, electromagnetic, etc.
[1]–[8]. In addition, energy may be harvested from man-made
Index Terms—Energy harvesting communications, energy co-
operation, simultaneous wireless information and energy transfer. sources via wireless energy transfer, where energy is trans-
ferred from one node to another in a controlled manner. These
I. I NTRODUCTION technologies have varying degrees of harvesting capacities
and efficiencies. While the devices/circuits side of engineering

P ROVIDING energy harvesting capability to wireless de-

vices enables the nodes to continually acquire energy
from nature or man-made phenomena. This in turn provides a
has been continually working to improve energy harvesting
mechanisms and devices and their efficiency, the efforts on the
signals/systems side of engineering to develop communication
promising future for wireless networks: self-sustainability and schemes for networks composed of energy harvesting nodes
virtually perpetual operation with network lifetimes limited by have been quite recent, e.g., [9]–[56]; see also [57]. The goal
those of the hardware rather than the energy storage. Energy of this review article is to summarize recent results in energy
harvesting wireless networks are expected to introduce several harvesting wireless communications and wireless energy trans-
transformative changes in wireless networking as we know it: fer from the perspectives of communication theory, signal pro-
in addition to energy self-sufficiency and perpetual operation, cessing, information theory and wireless networking. Energy
expected benefits include reduced use of conventional energy harvesting brings new dimensions to the wireless communica-
and accompanying carbon footprint, untethered mobility by tion problem in the form of intermittency and randomness of
available energy, as well as the possibility of sharing energy
Manuscript received April 1, 2014; revised September 15, 2014; accepted among the nodes in a network via wireless energy transfer,
December 16, 2014. Date of publication January 15, 2015; date of current which necessitate a fresh look at wireless communication pro-
version March 19, 2015. tocols at the physical, medium access and networking layers, as
S. Ulukus is with the Department of Electrical and Computer Engi-
neering, University of Maryland, College Park, MD 20742 USA (e-mail: well as at the fundamental performance limits, i.e., the channel
[email protected]). capacity. In this article, we summarize such approaches taken
A. Yener is with the Department of Electrical Engineering, Pennsylvania in the past few years in this new research field.
State University, University Park, PA 16802 USA (e-mail: [email protected]).
E. Erkip is with the Department of Electrical Engineering, New York The remainder of the paper is organized as follows. In
University Polytechnic School of Engineering, Brooklyn, NY 11201 USA Section II, we summarize the efforts in identifying the
(e-mail: [email protected]).
O. Simeone is with the Department of Electrical and Computer Engineer-
information-theoretic limits of energy harvesting communi-
ing, New Jersey Institute of Technology, Newark, NJ 07102 USA (e-mail: cations, i.e., capacity, by considering energy harvests at the
[email protected]). channel use level. Next, in Section III, we consider throughput
M. Zorzi is with the Department of Information Engineering, University of
Padova, 35131 Padova, Italy (e-mail: [email protected]).
maximization, where energy harvests are at the communication
P. Grover is with the Department of Electrical and Computer Engi- slot level, and describe the efforts in identifying the through-
neering, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail: put optimal transmission power and scheduling policies. This
[email protected]).
K. Huang is with the Department of Electrical and Electronic Engi- section considers offline availability of energy harvesting times
neering, The University of Hong Kong, Pokfulam, Hong Kong (e-mail: and amounts. In Section IV, we consider online optimiza-
[email protected]). tion of general reward functions, including the throughput. In
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. Section V, we go above the physical layer and develop medium
Digital Object Identifier 10.1109/JSAC.2015.2391531 access layer control protocols for energy harvesting wireless

0733-8716 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 361

Fig. 1. AWGN channel with random energy arrivals with (a) an unlimited battery, (b) no battery.

devices. Next, Section VI considers the case where energy That is, at each channel use, the cumulative energy expended
can be shared and transferred between the nodes in a wireless cannot exceed the cumulative energy harvested. We note that
network, jointly with information. Here, we consider both while the average power constraint in the classical information
information-theoretic and network-theoretic advances in infor- theory setting in (1) imposes a single constraint for the entire
mation and energy transfer including the notions of energy codeword, the energy harvesting scenario in (2) imposes n
cooperation and interactive exchange of information and en- power constraints on the codeword. We also note that the
ergy. In Section VII, we consider energy harvesting sensor cumulative form of the constraints (2), and the unbounded
networks where a number of operations other than transmis- nature of the battery, allow for saving of the energy harvested
sion incur energy costs and impact the performance. Next, in in any channel use to be used at a later channel use. On the
Section VIII, we examine the potential of large-scale energy other hand, when there is no battery to save harvested energy
harvesting networks, including mobile ad hoc and cellular net- for future use, the constraints on the codewords become
works composed of energy harvesting wireless devices, as well
as heterogeneous systems with nodes of varying capabilities. Xi2 ≤ Ei , i = 1, . . . , n (3)
Section IX describes the advances in identifying total energy
consumption models for energy harvesting communication sys- which impose instantaneous stochastic amplitude constraints
tems. Finally, Section X provides the conclusions for the article on the code symbols, as depicted in Fig. 1(b). When we have
and lists some future directions in the broad area of energy a finite-sized battery (of maximum size Emax ) as shown in
harvesting communications. Fig. 2(a), the battery size (i.e., the amount of available energy
in the battery) at channel use i, denoted as Bi , will evolve as
II. A N I NFORMATION -T HEORETIC V IEW follows
OF E NERGY H ARVESTING  
Bi+1 = min Bi − Xi2 + Ei , Emax (4)
Consider the classical AWGN channel with input X, additive
zero-mean unit-variance Gaussian noise N, and output Y =
which denotes that first an Xi2 amount of energy exits the
X + N. The capacity of this channel is C = 12 log(1 + P). In
battery (due to the transmission of symbol Xi ), and then Ei
this classical result of Shannon [58], the codewords are average
amount of energy is harvested into the battery. Therefore, what
power constrained in the following way:
is transmitted, i.e., Xi , affects the amount of energy in the
1 n 2 battery in the next channel use, and how much energy there
∑ Xi ≤ P
n i=1
(1) is in the battery, i.e., Bi , affects the allowable set of symbols
via the instantaneous amplitude constraint Xi2 ≤ Bi . Here we
for very large n, where Xi denotes the ith element of the note for future reference that: the battery state Bi will be a
transmitted codeword. Consider now that the energy arrives highly correlated random process over time even when the har-
(is harvested) stochastically at the transmitter as a stationary vesting process Ei is i.i.d.; actions of the transmitter (i.e., what
and ergodic random process Ei , with an average recharge rate it sends) affect the future of the battery state; and the transmitter
E[Ei ] = P, as shown in Fig. 1(a). We therefore introduce a naturally knows the battery state, but the receiver does not.
canonical model of an energy harvesting system as a commu- The capacity of the energy harvesting channel is known
nication channel augmented with an energy harvesting battery only in the cases of unboundedly large battery (Emax = ∞)
(energy queue) as shown in Fig. 1(a). In this initial model, we [23], no battery (Emax = 0) [24], and for a unit-sized battery
assume that the battery has an unbounded capacity as depicted (Emax = 1) over a binary noiseless link [25]. We will see that,
by an open ceiling in Fig. 1(a). At each channel use, Xi2 units for a Gaussian channel, the channel capacities for Emax = ∞ and
of energy are depleted from the battery, and Ei units of energy Emax = 0 are very different and they are achieved with vastly
enter the battery. For a codeword to be transmitted without different strategies. In particular, when Emax = ∞, Gaussian
any energy outages, we need to satisfy the energy causality codebooks achieve capacity as in the classical setting, whereas
constraints at every channel use when Emax = 0, discrete signaling is observed numerically to
be optimal. The result for the case of unit-sized battery over
k k
a binary noiseless link in [25] shows the richness of and the
∑ Xi2 ≤ ∑ Ei , k = 1, . . . , n (2)
challenges posed by this problem, and presents an interesting
i=1 i=1

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
362 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 2. (a) System with a finite-sized battery, (b) equivalent timing representation.

connection to timing channels [59]. The capacity for general use index, implying that only finitely many mismatches occur.
noisy channels with finite capacities (i.e., for any Emax ) is Another important point to note regarding these two schemes
an important open research problem. Some progress has been is that neither the transmitter nor the receiver need to know the
made in this direction, in terms of developing lower and upper energy arrival process or the current battery energy state. This is
bounds to the capacity, in recent work [60], [61]. again due to the unlimited nature of the battery which smoothes
out the randomness in the stochastic energy arrival process, and
any battery state information at the transmitter or receiver does
A. Capacity With an Unlimited-Sized Battery, Emax = ∞ not improve the achievable rates. On the other hand, as we will
see, in the case of 0 ≤ Emax < ∞ it is crucial that the transmitter
We first note that each codeword satisfying the constraints (2)
has causal information of the energy arrival to achieve a non-
also satisfies the average power constraint in (1) automatically
zero reliable rate. However, we note that it is natural that the
by the strong law of large numbers as 1n ∑ni=1 Ei → P. Therefore,
transmitter has causal knowledge of the energy arrival process,
the constraints of the energy harvesting system are stricter,
because it observes the incoming energies into its battery.
and hence the capacity of the energy harvesting system is
Finally, we note that [62] considers the same problem, and
upper bounded by the classical capacity with an average power
proves the same capacity result in [23] by using a different proof
constraint equal to the average recharge rate.
technique which relies on Asymptotically Mean Stationary
Reference [23] gives two schemes to achieve this upper
(AMS) sequences. In addition, [63], [64] extend these results
bound. In the first scheme, called save-and-transmit, the
to multiple access channels with energy harvesting transmitters
transmitter saves energy in the first h(n) channel uses (does
with unlimited-sized batteries; see also [23, Section VI].
not transmit anything), and transmits data in the remaining
n−h(n) channels uses by using a codebook generated with i.i.d.
Gaussian samples of power equal to the average recharge rate P.
B. Capacity With no Battery, Emax = 0
By letting both h(n) and n − h(n) go to ∞, and choosing h(n)
as o(n) such as log(n), we can achieve the AWGN capacity. We now consider the other extreme where there is no battery
This can be seen as follows: By saving an infinite amount of to store and save energy for future use. In this case, the
energy in the saving phase we prevent energy outages in the channel inputs are instantaneously amplitude constrained as in
data transmission phase with probability one, and by choosing (3). However, different from the existing literature [65], these
h(n) as o(n) we make the rate-hit taken by not transmitting amplitude constraints are not deterministic and constant, but are
any data during the saving phase negligible. In the second time-varying and stochastic. The transmitter knows the energy
scheme, called best-effort-transmit, we start data transmission arrival profile causally, and the receiver does not know it. In this
right away without a saving period. We construct a Gaussian case, the transmitter can choose the code symbols according to
codebook with average power that is ε (small enough) less the observed energy, which is the state of the system. Reference
than the average recharge rate. At any given channel use, if we [24] combines the works of Smith [65], which considers the
have sufficient energy in the battery, we send the corresponding static amplitude-constrained AWGN, and Shannon [66], which
code symbol, otherwise, we send a zero symbol (i.e., not send considers the capacity of state-dependent channels with causal
anything). This creates mismatches between what is in the state information available at only the transmitter, to deter-
codebook, and what is actually transmitted, but the number mine the capacity of this channel model. In the solution, the
of mismatches remains finite from the strong law of large transmitter sends the channel input T1 when the energy arrival
numbers, and therefore such mismatches are inconsequential state is e1 and T2 when the energy arrival state is e2 , where
√ √
with joint typical decoding. T1 , T2 are jointly distributed random variables in [− e1 , e1 ] ×
√ √
It is important to note that the availability of an unlimited [− e2 , e2 ]. Reference [24] observes experimentally that the
battery is essential in both achievable schemes. In particular, in support set of this distribution is finite. This is reminiscent of
the save-and-transmit scheme, the unlimited battery enables us Smith’s result for the static amplitude constrained case, where
to save essentially an unlimited amount of energy in the saving the input distribution resides in one dimension. In addition, if e1
phase to prevent any energy outages in the data transmission and e2 are sufficiently small, then symmetric binary distribution
√ √ √ √
phase. In the best-effort-transmit scheme, the unlimited battery with masses at ( e1 , e2 ) and (− e1 , − e2 ) is optimal. As
enables the energy queue size to blow up sooner or later, e1 and e2 are increased, the optimal distribution begins to
preventing any energy shortages after a large enough channel have higher number of mass points, e.g., ternary, quaternary,

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 363

Fig. 3. (a) Energy harvesting system. (b) Energy feasibility tunnel.

etc. Reference [24] shows that the capacity with an unlimited Then, we choose to wait a Vi number of channel uses to send
battery is significantly larger than the capacity with no battery. information. Here Vi serves as our channel input that is to be
Reference [67] extends these results to the case of an energy optimized. Reference [25] combines Anantharam–Verdu’s bits
harvesting multiple access channel with no batteries. through queues [59] and Shannon’s state-dependent channels
with causal state information available at only the transmitter
[66] to find a single-letter capacity expression for the capacity
C. Capacity With Unit-Sized Battery, Emax = 1 of this equivalent channel. We note that the channel input Vi is
We now summarize the very recent work in [25] which a function of the message (denoted by Ui in [25]) and the state
takes a step towards understanding the capacity of an energy Zi which is causally known to the transmitter. Shannon strategy
harvesting link with a finite-sized battery, i.e., 0 < Emax < ∞; [66] is optimal here, because the state Zi is i.i.d. in time. The
see Fig. 2(a). With a finite-sized battery, the channel inputs capacity expression involves an auxiliary random variable U,
are instantaneously amplitude-limited to the (square root of and its optimization is difficult. For this reason, [25] determines
the) current amount of energy in the battery. From [24], [65] an achievable rate based on a certain selection of this auxiliary
(and the previous sub-section), we know that, when the channel random variable. This selection resembles the concentration
inputs are constant amplitude-constrained, or i.i.d. stochastic idea in [69], and may be interpreted as a lattice-type coding
amplitude-constrained, over a Gaussian channel, the optimum for the timing channel. Reference [26] provides an n-letter
input distributions are discrete. However, these discrete mass expression for the noisy-channel version of this problem with
points are arbitrary real numbers, and it is hard to track the dy- an arbitrary battery size, conjectures that it is the capacity, and
namics of the energy queue, if it is served with codebooks gen- evaluates it using techniques in [70].
erated by arbitrary real mass points. For a tractable abstraction
of the system, [25] models energy arrivals as multiples of a fixed III. O FFLINE E NERGY M ANAGEMENT FOR
quantity, and correspondingly, considers a physical layer which T HROUGHPUT M AXIMIZATION
has a discrete alphabet based on this fixed quantity. For further
analytical tractability, [25] assumes that the physical layer is In this section, we take a communication theory and network-
a noiseless binary channel, energy arrivals are binary, and the ing approach to the energy harvesting communication problem.
battery is unit-sized. Even in this simple model, unavailability We first consider the basic single-user channel, and then present
of the battery state at the receiver, memory of the battery state extensions to multi-user settings and practical considerations
in time, and the fact that the state evolves based on the previous such as processing costs and battery imperfections.
channel inputs, render the problem challenging. The fact that
channel inputs affect future states is reminiscent of action
A. Single-User Channel
dependent channels in [68]. While Shannon strategy, which
is optimal in the zero-battery case in [24], yields achievable Consider the single-user fading channel with additive Gaus-
rates for the finite-battery case, the transmitter may utilize the sian noise as shown in Fig. 3(a). The transmitter has two queues,
memory in the battery state to achieve higher rates. the data queue where the data packets are stored, and an energy
Reference [25] shows that this noiseless binary channel with queue where the arriving (harvested) energy is stored. The goal
a unit-battery can equivalently be modeled as a timing channel here is to schedule the transmission of data packets in the data
[59], where information is transmitted by timings between 1s, queue using the energy in the battery. We relate the instanta-
as opposed to the actual places of 1s and 0s. This converts the neous power and rate through a monotone increasing concave
problem into a timing channel with additive geometric noise function. While we can use an arbitrary monotone concave
(service time), where the service time is causally known to relationship, for simplicity and convenience, we assume the
the transmitter. This is explained in Fig. 2(b), where circles following familiar power–rate relationship: R = 12 log(1 + hP).
represent energies harvested and triangles represents 1s put to Therefore, whenever we send a signal with power P in an epoch
the channel. Here, after sending a 1, we have to wait a random of duration , 2 log(1 + hP) bits of data are served out from
Zi number of channel uses to receive an energy into the battery. the data backlog with the cost of P units of energy depletion

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
364 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 4. The optimal policy is the tightest string that connects the two ends. (a) Emax = ∞. (b) Finite Emax .

from the energy queue. With this model in mind, we solve for constant in between energy harvests; therefore, the dimension
the optimum power control policy P(t) in time as a function of the optimization problem is reduced to the finite number
of the energy arrival profile, the data backlog profile and the of epochs in an interval. Geometrically, this means that the
channel fading profile, in order to minimize the time by which feasible energy consumption profiles which are candidates to
all of the packets are successfully transmitted. Minimizing be optimum must be piece-wise linear. The optimal policy is
the transmission completion time for a given number of bits shown to be the tightest string that lies in the energy feasibility
is equivalent to maximizing the number of bits transmitted tunnel [12], [13]. This solution aims to keep longest stretches
in a given duration. Therefore, in the following, we consider of constant power periods subject to energy causality and no-
maximizing the number of bits delivered by a deadline T . energy-overflow constraints, as the concavity of the power–rate
The optimization problem is subject to the energy causal- relationship favors constant powers to the extent possible. An
ity constraint on the harvested energy, and the finite-storage example of the optimum energy consumption curve is shown
constraint on the rechargeable battery. In particular, the en- in Fig. 4(a) for Emax = ∞, i.e., there is no energy overflow
ergy causality constraint requires that the energy that has not concerns, and in Fig. 4(b) for a finite Emax .
arrived yet (has not been harvested yet) cannot be used. The An alternative approach to the feasible tunnel approach is
finite-storage constraint, on the other hand, requires that no the directional water-filling algorithm presented in [14]. The
energy is wasted because of battery being full at the time directional water-filling algorithm aims to distribute the water
of energy arrivals; we also call this constraint the no-energy- (energy) equally over time, subject to energy causality con-
overflow constraint. Assume that energies of {E0 , E1 , . . . , EN−1 } straints, which introduce the directionality of water (energy)
are harvested, and epoch lengths are {1 , . . . , N−1 }. Due to flow. The directional water-filling algorithm requires walls at
the concavity of the rate–power relationship, power must be the points of energy arrival, with right permeable water taps in
kept constant between energy harvests [12]. This reduces the each wall which allows water to flow only to the right. This
power control policy of P(t) to a sequence of constant powers implements the energy causality constraint, i.e., energy can be
{p1 , . . . , pN }. The energy causality constraints become [12] saved and used in the future, but the energy that will arrive
k k−1
in the future cannot be used before it has arrived. In addition,
∑ i pi ≤ ∑ Ei , k = 1, . . . , N (5) these taps allow at most Emax amount of water to flow to the
right. This implements the finite-capacity battery constraint by
i=1 i=0
avoiding overflows. These are based on the KKT optimality
and the no-energy-overflow constraints become [13] conditions found from the corresponding convex optimization
k k problem [14]
∑ Ei − ∑ i pi ≤ Emax , k = 1, . . . , N − 1 (6)
p∗i = 
1
 − 1, i = 1, . . . , N (7)
i=0 i=1
∑ j=i λ j − ∑Nj=i µ j
N+1
We illustrate these two constraints on the energy consump-
tion policy in Fig. 3(b). The upper staircase is the cumulative where λi are the Lagrange multiplier that enforce energy causal-
energy arrival profile which provides the energy causality up- ity and µi are the Lagrange multipliers that enforce no-energy-
per bound, and the lower staircase is the no-energy-overflow overflow conditions. In the implementation of the directional
curve which provides a lower bound. Any feasible energy water-filling algorithm, first, the taps are kept off, and transfer
consumption curve must lie in between. We note that the energy from one epoch to the other is not allowed. Then, the taps
causality constraint forces the energy consumption to slow are turned on one by one, and at most Emax − Ei units of
down not to exceed the harvested amount, while the no-energy- energy transfer from past to the i + 1st epoch is allowed. An
overflow constraint forces energy consumption to speed up to example run of the algorithm is shown in Fig. 5, for a case of
open up space in the battery for new energy arrivals. Although 5 epochs. Four energy arrivals occur during the course of the
the optimization is over all monotonically non-decreasing time transmission, in addition to the energy available at time t = 0.
functions for the energy consumption curve, by the concavity We observe that the energy level equalizes in epochs 1 and 2.
of the objective function, the optimal power policy must remain The energy arriving at the beginning of epoch 3 cannot flow

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 365

Fig. 5. Directional water-filling algorithm. (a) Initial water (energy) levels. (b) Final water levels.

left due to the energy causality constraint, which are enforced relay optimizes its throughput subject to its own energy profile
by right permeable taps. We observe that the excess energy in and the data profile coming from the source. In [33]–[38], half-
epoch 3 cannot flow right either, due to the Emax constraint. duplex two-hop and more general two-way relay settings are
In addition, the energy arriving at the beginning of epoch 5 studied.
cannot flow left, again due to the energy causality constraint. In Reference [39], [40] consider the case of processing costs,
the case of a fading channel, additionally, the strengths of the where the transmitter spends a constant amount of energy per
channel states play an important role in the directional water- unit time for the circuitry when the transmit power is non-
filling algorithm [14]. See also [15] for a staircase water-filling zero, and solve the throughput maximization problem subject
algorithm. to energy causality, no-energy-overflow and processing cost
conditions. The solution is characterized as a directional glue-
pouring algorithm. There is a threshold power level p∗ that is
B. Multi-User Channels and Practical Considerations found by solving the following fixed-point equation:
This concludes the summary of the basic findings in the case log(1 + p∗ ) 1
= (8)
of a single-user channel. In the case of a broadcast channel p∗ + ε 1 + p∗
with an energy harvesting transmitter, [17], [19] showed that
the optimal total transmit power management policy is the Glue-pouring is performed such that the power level is always
same as the optimal single-user counterpart summarized above, higher than p∗ whenever it is non-zero and the glue level is
and this optimum total transmit power is distributed among calculated accordingly. In particular, the optimal transmission
signals going to the users according to a cut-off structure; policy is bursty in the sense that the length of a transmission
only the optimum total transmit power that is above this cut- schedule is not allowed to be arbitrarily long due to the pro-
off level goes to the weaker user. For the case of a broadcast cessing cost incurred per unit time. In the directional glue-
channel, an iterative algorithm is developed in [18], and these pouring algorithm, harvested energies are allocated into the
approaches are generalized to a fading and MIMO case in [20]. corresponding epochs first where energy is viewed as the glue
In the case of a multiple access channel with energy harvesting in [73]. Then, the glue is allowed to flow to the right only
transmitters, [16] uses a combination of directional water-filling and the equilibrium glue levels are determined. Reference [41]
together with a generalized water-filling in [71] and iterative extends the results in [39] by considering a broadband fading
water-filling in [72] to obtain the optimum energy management energy harvesting communication system with processing cost,
schemes to maximize the region of departed bits in a given and shows that when energy is limited, additional bandwidth
duration. For the case of an interference channel, using some may not be utilized.
recent advances in the sum-capacity of the interference channel, References [21], [22], [27] address practical imperfections in
[21] develops sum-rate optimal transmission policies for a energy storage and retrieval. Energy storage units may foster
class of two-user interference channels with energy harvesting imperfections such as losses during charging/discharging, leak-
transmitters. In particular, using concavity properties of the age of available energy over time, etc. In particular, [27] con-
sum-rate expressions, sum-rate optimal transmission policies siders offline throughput maximization for energy harvesting
of the users are found by using directional iterative water-filling systems with energy leakage over time. The presence of energy
algorithms, where one user’s power profile determines the noise leakage is reflected in the modification of the shape of the
profile of the other user. energy feasibility tunnel. In particular, the cumulative energy
References [28], [29] consider the full duplex end-to-end harvesting curve decreases in between two energy harvests due
communication over a two-hop relay channel where the source to energy leakage and the power levels have to be determined
and the relay node harvest energy from nature. The source sends accordingly. Reference [21] considers the imperfection that
data by using its harvested energy and the relay forwards the occurs instantaneously at the time of charging and discharging
data coming from the source to the destination using its own of the battery. In particular, due to the inefficiency in charging/
harvested energy. References [28], [29] show that the optimal discharging, [21] shows that it may be more advantageous to
policy is in general non-unique and that there exists a separable immediately use the harvested energy without first storing it
optimal policy in which the source optimizes its throughput in the battery, and determines an optimum double-threshold
without regard to the relay energy harvesting profile, and the policy. Recent [22] considers the case where the portion of the

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
366 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

energy not stored in the inefficient battery can be saved in an so that the decision will always be compatible with the amount
efficient, but size-limited, super-capacitor, yielding a hybrid en- of energy that is available in the battery.
ergy storage unit composed of a large-sized but inefficient bat- Two different models have been studied in this case. In
tery and small-sized but perfect super-capacitor. Reference [22] the first model [77], each packet has an associated random
finds the optimum power management policy by applying the importance value, Vk , which represents the reward that is gained
directional water-filling algorithm in multiple stages. Reference if the packet is transmitted. In this case, the action is binary and
[42] considers the outage probability as opposed to throughput is represented by the decision about whether or not to transmit
as the design metric, and determines the corresponding optimal the packet. Such decision can be implemented by a simple
power allocation policy. threshold policy, where the packet is sent if its importance value
is above a certain state-dependent threshold, and discarded
otherwise. The long-term average reward per slot can be com-
IV. O NLINE E NERGY M ANAGEMENT FOR G ENERAL
puted as
R EWARD M AXIMIZATION
 
K−1
In this section, we consider the case in which devices with 1
R = lim inf E ∑ QkVk (9)
energy harvesting capabilities send data packets to a receiver K→∞ K
k=0
according to some transmission policy. Specifically, time is
slotted, and in every slot a data packet is produced, which can be and the optimal strategy can be numerically found using stan-
either transmitted in the same slot or discarded (i.e., no queu- dard techniques, such as the Policy Iteration Algorithm (PIA)
ing is considered). The transmission of a packet corresponds [78]. In some cases, it is possible to approximate the optimal
to some reward, and maximization of the long-term average policy through some heuristics which, while being very easy
reward per slot is sought. We start by considering the case with a to compute and to store in the device, provide a level of
single device. The treatment here is based on [74], [75]. Related performance that is very close to the optimum. A notable
work that studies the throughput of TDMA and carrier sense example, which is asymptotically optimal for Emax → ∞ and
multiple access protocols can be found in [76]. very good already for quite modest values of Emax , is the so-
The device operates in a completely on-line fashion, i.e., called balanced policy, where the threshold is such that in
it has only causal knowledge of the evolution of the system. each slot the average consumed energy is equal to the expected
In this case, the device must make intelligent decisions about harvested energy.
whether or not to transmit based on the system state, which In the second model, different power levels can be used
includes the amount of energy stored in the battery, and possibly for transmission, which correspond to different rewards (no
some state of the harvesting process. A useful framework to importance value needs to be considered in this case). The
study these types of problems is provided by Markov Decision optimization now looks for the best strategy in selecting the
Processes. For simplicity, assume that the harvesting process is transmit power so as to make the best use of the available energy
i.i.d.,1 so that the state of the system is limited to the current while accounting for the specific relationship between transmit
contents of the battery. Each system state is assigned a set of power and reward gained.
possible actions (e.g., idle and transmit, possibly with different As discussed in [75], the above model can be extended to
power levels), so that transitions from a certain state depend on the case in which the harvesting process is correlated, which in
external events (e.g., whether or not some energy is harvested) most cases is a much more accurate model of what may happen
as well as on the decision made (e.g., whether or not the packet in reality (e.g., solar energy). In this case, if the harvesting
is transmitted). Since each policy (i.e., the definition of the process is itself regulated according to some underlying Markov
probability distributions of the various actions as a function of chain, we can include it in the system state and use the same
the state) will result in a different evolution of the underlying approach as before. Note that in this case the standard numerical
Markov chain and correspondingly a different value of the procedures can still be applied (at least as long as the complex-
overall reward, the objective is to come up with the policy for ity of the model is manageable), but the additional complexity
which the reward is maximized. may make it more difficult (and in some cases impossible)
A quite general model that describes the evolution of the bat- to obtain closed-form results. An interesting observation in
tery status in slot k, Bk , is Bk+1 = min{[Bk − Qk ]+ + Ek , Emax }, this case is that the performance is no longer determined by
where Qk is the amount of energy used in slot k as a result of the battery size alone, Emax , but rather depends on the ratio
the action chosen, Ek is the amount of energy harvested during between Emax and the dynamics of the harvesting process (e.g.,
slot k, and Emax is the finite storage capacity of the battery. for slow harvesting processes in which periods of low harvest
can potentially last long, a large battery is needed for good
performance).
A. Optimal Transmission Policies With Perfect Knowledge of
the State-of-Charge
B. Optimal Transmission Policies With Imperfect Knowledge
Consider the case in which the device has full access to the
of the State-of-Charge
information regarding the exact energy level in its own battery.
In this case, the battery state evolution above can be observed, Although it may seem quite natural that a microprocessor is
able to access the battery and to accurately know the amount of
1 Extensions to non-i.i.d. harvesting are discussed in [75]. energy it contains, several studies have shown that this may not

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 367

always be the case. For example, uncertainty in the parameters rechargeable batteries hold the promise of breaking this barrier,
of the components of the circuitry may lead to errors as large as enabling perpetual operation if the consumption and harvesting
30% [79]. In addition, there is a trade-off between the accuracy of energy can be balanced in some ways. However, this vision
in the knowledge of the battery State-of-Charge and the amount overlooks an important aspect, which is familiar to all cell
of time and energy invested in gaining such knowledge, so that, phone users: after a battery has been recharged many times, it
even if it were possible, obtaining accurate information may starts losing its capability of storing energy and/or it discharges
be too expensive, especially in simple and resource constrained more quickly. In fact, a battery cannot last forever, and some
devices. It is therefore of interest to study the case in which the aging phenomena make it less and less capable as a function
optimization of the transmission policy needs to be carried out of the number of recharge cycles it went through. Perpetual
under the constraint of uncertain information about the battery operation may not be possible, after all.
status. In order to study this phenomenon, one can still use the MDP
The underlying process that describes the evolution of the model described above, but needs to introduce an additional
battery status is still a Markov chain, but in this case the state of layer of memory which tracks the degradation of the battery,
the chain is known with some error, and therefore the Markov which in this case translates in a time-varying characterization
Decision Process framework no longer applies. Rather, this new of its parameters. For example, after a certain number of cycles,
formulation falls in the realm of the so-called Partially Observ- c, a battery that originally was capable of storing an amount of
able MDPs, which model the lack of perfect observability of energy equal to Emax may exhibit a reduced capacity α(c)Emax ,
the system state. In this case, the optimal policy can be found, where α(c) ≤ 1 is a non-increasing function of c. In order to
but full knowledge of the past history is required (unlike for the avoid the very large number of states that would result from
Markov case, where the most recent state is sufficient), making this direct formulation, one can characterize this degradation
the policy very difficult to compute in general. If resource process using some probabilistic technique, in which several
constrained devices are considered, it may make sense to study degradation stages are identified and the battery jumps from
simplified policies, in which for example only the latest state one to the next with some (small) probability in every slot.
is used. For some cases of interest, it has been shown that Such formulation, described in detail in [83], makes it possi-
such suboptimal policies actually provide very good results, and ble to carefully study these effects while keeping the model
may be very close to the true optimum [80]. For the case of complexity within reasonable limits. Using this model, one
correlated harvesting, it has been shown that overall it is more can characterize the total amount of energy that a battery can
important to know the harvesting state rather than the energy provide through all the recharging cycles it can support during
state [81]. its entire life.

C. Optimal Transmission Policies for Two Users

E. Sensing Policies
As a first step towards multi-user energy harvesting systems,
A formulation that is similar to the above but has a different
consider a small network with only two devices who send
flavor is presented in [84]. Here the model focuses on the energy
their data to a Fusion Center (FC) through a collision channel,
consumed for sensing (instead of transmitting), and the policy
where individual transmissions are correctly received whereas
makes a decision about whether a sensor should be activated in
simultaneous transmissions lead to the loss of both packets. As
a given slot (with the corresponding energy consumption and
an upper bound to the performance of such a system, consider
accrued benefit) or it should remain idle (to save energy but at
the case in which a central controller (e.g., the FC itself) has
the risk of missing important events).
full knowledge of the energy state and of the importance values
For correlated events, the optimal strategy leverages on the
of the packets at the two devices, and decides which (if any)
ability to predict whether or not an event is likely to occur in
is going to transmit. This case can be studied by extending
the next slot, and on some thresholding rule on the amount of
the previous formulation to a two-dimensional model, which
available energy: intuitively, if an event is expected and there is
is conceptually similar to the one-dimensional case (and can
enough energy, the sensor will be activated.
similarly be solved numerically using the PIA), but does not
lend itself to analysis and closed-form results. What can be
numerically observed in this case is that as long as the batteries
F. Optimal Random Multi-Access for a Network of Energy
are not very small, a balanced policy is able to provide good
Harvesting Devices
results which lose very little compared to the true optimal
solution [82]. We now turn to the case in which multiple devices try to
access the channel randomly in order to send a data packet of
random importance to the Fusion Center (FC). The scheme is
D. Battery Degradation Phenomena
very simple in that there is no coordination enforced by the
In traditional energy conservation schemes, it is assumed FC. The model for each device would still obey the recursive
that a finite amount of energy is initially stored in a non- formulation for the battery state above. The problem here,
rechargeable battery, and therefore needs to be judiciously unlike in the simple two-user example given above, is that no
used in order to maximize the device lifetime or some mea- coordination is available, which makes the model much more
sure of energy efficiency. Energy harvesting technologies and realistic but also prone to collision errors and inefficiencies.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
368 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 6. (a) Frames (orange) and slots (green) for TDMA and DFA MAC protocols in the nth inventory round (IR); (b) Trade-off between delivery and time
efficiencies for different harvesting rates µH for the MAC protocols TDMA, FA and DFA.

The reward model for a given user u can be rewritten as above is very good unless the battery capacity of the devices
  is very small. More details about the above formulation as well
K−1
1
R(u) = lim inf E ∑ Qu,kVu,k ∏(1 − Qi,k ) (10) as numerical results and discussions can be found in [85].
K→∞ K
k=0 i=u

where the user index has been explicitly added to the decision V. M EDIUM ACCESS C ONTROL FOR E NERGY
variable, Qu,k ∈ {0, 1}, and to the importance value, Vu,k , and H ARVESTING N ETWORKS
U is the total number of users. Equation (10) shows that user
In this section, we address medium access control (MAC)
u gains its importance value if it decides to transmit while all
protocols for single hop networks in which a fusion center
other users are silent. The overall system reward is therefore
collects data from energy harvesting devices in its surround-
expressed as the sum of R(u) over all users u. Using the Markov
ings. We consider the case in which the devices generate data
formulation, it is possible to write the objective function as
periodically, when timed measurements of a given quantity of
U interest need to be reported.
R= ∑ G(ηu ) ∏ (1 − P(ηi )) (11) We investigate how the performance and design of standard
u=1 i=u
MAC protocols, such as TDMA, framed-ALOHA (FA) and
where ηu is the policy for user u, which specifies for each level dynamic-FA (DFA) [86], are influenced by the discontinuous
of battery energy, e = 0, 1, . . . , Emax , the transmission proba- energy availability in the energy harvesting devices in the pres-
bility or, equivalently, the threshold on the importance value ence of periodic data generation; see [74], [87], [88]. Consider
used in making a decision about whether or not to transmit, a single hop network with a fusion center surrounded by M
and G(ηu ) and P(ηi ) are the corresponding average gain and energy harvesting devices. The fusion center retrieves data, e.g.,
average transmit probability for users u and i, respectively. If measurements of a given phenomenon of interest, from the
we assume a symmetric system with fairness constraints, (11) devices via periodic inventory rounds (IRs). Each IR is started
becomes by the fusion center by transmitting an initial query command,
which provides both synchronization and instructions to the de-
R = UG(η) (1 − P(η))U−1 (12) vices on how to access the channel. Time is slotted. In every IR,
each device has a packet, e.g., a new measurement, to transmit
The overall optimization problem can then be formulated as with a given probability, independent of the other devices and
the maximization of R within the set of admissible policies. previous IRs. Each measurement is the payload of a packet,
Unfortunately, unlike in the case U = 1, the problem is non- whose transmission fits within the slot duration. The goal of the
convex, and therefore the global maximum cannot be easily fusion center in each IR is to collect as many packets as possible
found. Through a game theoretical formulation, the problem within the constraints imposed by the energy availability at the
can be addressed as follows: (i) we assume that the devices devices.
play a game, each trying to increase its own reward; (ii) we Each IR is organized into frames, each of which is composed
look for Nash Equilibria in this game, i.e., situations in which of a number of slots that is selected by the fusion center; see
no user has an incentive to deviate; (iii) in particular, we look Fig. 6(a). Depending on the adopted MAC protocol, any device
for symmetric Nash Equilibria, where all users adopt the same that needs to transmit in a frame either chooses or is assigned
policy and they all have no incentive to do otherwise; (iv) it a single slot within the frame for transmission. Moreover,
is possible to show that the game has a unique symmetric Nash after a device has successfully transmitted its packet to the
Equilibrium, and that such equilibrium is a local optimum of the fusion center, it first receives an acknowledgment of negligible
original problem; (v) while there is of course no guarantee that duration from the fusion center and then becomes inactive for
the local optimum is also a global optimum, some numerical the remaining of the IR. The fusion center knows neither the
investigation has shown that the solution found as described number of devices with a new measurement to transmit nor the

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 369

state of the devices’ batteries. Assuming flat fading channels the estimated backlog. We define the proportionality constant
and that transmission is successful if the signal-to-interference as ρ. It is known that the optimal ρ for non-energy harvesting
ratio is large enough, any slot can be: empty when it is not devices in terms of time efficiency equals 1 [86]. However,
selected by any device; collided when it is chosen by more than following the discussion above (see also [87]), this is no longer
one device but no device is received successfully; successful the case for energy harvesting enabled networks. Finally, FA is
when a device transmits successfully, while possibly in the a special case of DFA where only one single frame of size L1 (n)
presence of other (interfering) devices. Each device has a finite is announced and hence the retransmission of collided packets
battery, and a fixed amount of energy is consumed at each is not allowed.
packet transmission. During the time between two successive The trade-off between the delivery efficiency and the time ef-
IRs, each device harvests a random amount of energy with ficiency is shown in Fig. 6(b) for the considered MAC protocols
average, normalized by the transmission energy, defined as µH . and for different values of the harvesting rate µH . For TDMA,
No energy is harvested within an IR, as its duration is assumed the trade-off consists of a single point on the plane, whereas
to be much smaller than the time needed to harvest sufficient FA and DFA allow for more flexibility via the selection of the
energy for a transmission. parameter ρ. Specifically, when ρ is increased more devices can
We measure the system performance in terms of the trade- report their measurements to the fusion center, thus increasing
off between the time efficiency, which measures the rate of data the delivery efficiency, but at the cost of lowering the time
collection at the fusion center and the delivery efficiency, which efficiency.
accounts for the number of packets successfully reported to the
fusion center. The time efficiency is a standard measure for the
VI. J OINT W IRELESS E NERGY AND
performance of MAC protocols, and is calculated as the ratio
I NFORMATION T RANSFER
between the overall number of packets successfully received
by the fusion center and the total number of slots allocated In the preceding sections, we considered wireless nodes that
by the MAC protocol. The delivery efficiency is instead the acquire their operational energy from external sources. In this
fraction of devices that are able to successfully report their realm of energy self-sustaining networks, a new dimension that
payload to the fusion center within a given IR. This criterion can be envisioned is sharing and transferring energy between
is specifically relevant for energy harvesting networks, since the nodes just like information. Energy and information transfer
energy harvesting devices may run out of energy before being between nodes can be made simultaneously or by separate
able to transmit successfully. We observe that, with contention- means/technologies. In this section, we review the state of the
based MACs such as ALOHA and variations thereof, there is a art in energy and information transfer for wireless networks.
trade-off between time and delivery efficiencies. Increasing the
delivery efficiency requires the fusion center to allocate a larger
A. An Information-Theoretic View
number of slots in an IR, as this reduces packet collisions and
hence the amount of energy wasted in unsuccessful transmis- The possibility of harvesting energy from the communicated
sions; however, a larger number of slots per IR decreases the signal itself has been discussed in the circuits literature, es-
time efficiency. pecially in the context of biomedical implants [89]. The first
With the TDMA protocol, each device is pre-assigned an explicit information-theoretic formulation of harvesting energy
exclusive slot in each IR, irrespective of whether it has a from the communicated signal is given in [90]; prior to this
packet to deliver or enough energy to transmit. Recall that work, [91] studied a related problem, where capacity with
such information is not available at the fusion center. Any receiver-side power constraints were considered. Reference
IR is thus composed by one frame with M slots as seen in [90] modeled the problem as one of maximization of mutual
Fig. 6(a). Since TDMA is free of communication errors in the information under a minimum requirement of energy delivered
considered interference-limited scenario, its delivery efficiency to the receiver:
is only limited by the energy availability at the devices and it
is thus an upper bound on the delivery efficiency for ALOHA- C(B) = max I(X;Y ), s.t. E[Y 2 ] ≥ B (13)
f (x)
based MACs. However, TDMA might not be time efficient due
to the many empty slots when the probability of having a new where B denotes the energy harvested by the receiver, and C(B)
measurement and/or the energy harvesting rate are small. denotes the capacity at this harvesting level. This reference
With the DFA protocol, the nth IR is organized into a set of also provided tradeoffs between capacity and average energy
frames as shown in Fig. 6(a). The backlog for the kth frame delivered for some example cases. The information-energy
is the set composed of all sensors that simultaneously satisfy transfer formulation is also mathematically similar to the clas-
the following three conditions: i) have a new measurement sical problem of cost-constrained capacity in [92, Thm. 6.11]
to transmit in the nth IR; ii) have transmitted unsuccessfully, and [93, Thm. 3.7.2]. For finite alphabets, by defining the “cost”
because of collisions, in the previous k − 1 frames; iii) have of a particular input symbol as the difference of the energy
enough energy left in the battery to transmit in the kth frame. All delivered by the symbol, and the maximum energy delivered
the devices in the backlog set attempt transmission during frame over all symbols, one can reduce the information-energy trans-
k. To make this possible, the fusion center allocates a frame of fer problem to the cost-constrained capacity problem.
Lk (n) slots, where Lk (n) is selected based on an estimate of the For energy and information transfer using inductively cou-
backlog size. A typical choice is to make Lk (n) proportional to pled circuits (see Fig. 7), assuming an additive Gaussian noise,

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
370 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 7. (a) Coupled-inductor circuit commonly used for wireless power transfer. (b) Frequency-faded channel transfer function.

[94] observed that this problem can be viewed as transfer energy utilization is modeled as a binary Markov stochastic
of information and energy over a frequency-faded channel. process. Due to the finite capacity of the battery, there may
Fig. 7(a) shows a coupled-inductor circuit commonly used be battery overflows and underflows. An overflow event takes
for wireless power transfer. The same circuit models near- place when energy is received but the battery is full; instead, an
field communication commonly used in RFIDs and medical underflow event occurs when energy is required by the receiver
implants [89]. Fig. 7(b) shows a transfer function of H( f ) but the battery is empty. Note that the probability of overflow
and white noise z (of power spectral density N( f ) = 1); the measures the efficiency of energy transfer by accounting for
curve is the plot of N( f )/|H( f )|2 . Reference [94] derived the the energy wasted at the receiver. In contrast, the probability
tradeoffs between capacity and energy delivery. For this system, of underflow is a measure of the fraction of the time in which
the optimal information-transfer strategy is water-filling [95]; the application run at the receiver is in outage due to the lack of
whereas the optimal power-transfer strategy is to send a single energy.
sinusoidal tone at the resonant frequency. However, the optimal Constrained run-length limited (RLL) codes are defined by
info-power transfer tradeoff is not achieved by a strategy that constraints on the minimum and maximum duration of bursts
time-shares between these two. Reference [94] showed that of “1” or “0” symbols. Specifically, type-0 (d, k)-RLL codes
the optimal strategy outperforms any strategy that time-shares are such that the runs of 0s have length at most k, while the
between water-filling and transmission of a single tone at the runs of 0s between successive 1s have length at least d; see
resonant frequency. Fig. 8(a). As a result, type-0 (d, k)-RLL codes are suitable for
The simultaneous energy and information transfer literature overflow-limited regimes in which controlling overflow events
has been extended to MIMO broadcast [96], fading [97], MIMO is most critical. Type-1 (d, k)-RLL codes are similarly defined
interference [98] channels, and also considering practical cir- by substituting “1” for “0”, and are hence appropriate for
cuit implementations [99] keeping in mind the current practical underflow-limited regimes in which it is necessary to ensure
limitations on receivers which may not be able to decode the presence of bursts of energy. Constrained RLL codes have
information and harvest energy simultaneously. been traditionally studied for applications related to magnetic
and optical storage [101]. The first application to the problem
of energy transfer has been reported in the context of point-to-
B. Coding for Energy and Information Transfer
point RFID systems in [102].
In this section, we concentrate on the problem of code design We now provide a numerical example to compare the per-
for systems with joint information and energy transfer. Classical formance of unconstrained and constrained codes. In Fig. 8(b),
codes, which are designed with the only aim of maximizing we observe the trade-off between the information rate and the
the information rate, are unstructured, i.e., random-like. As a performance in terms of energy transfer. The latter is measured
result, they do not allow to control the timing of the energy by the minimum between the probability of underflow Pu f and
transfer. This is a critical drawback, since, in most scenarios the probability of overflow Po f . In the example, the receiver
of interest, arbitrary energy transmission patterns may lead wishes to use energy periodically once every two time slots.
to inefficiencies, such as battery overflows or underflows. In A full description of the simulation set-up can be found in
contrast, as reviewed here, constrained, rather than classical [100]. The figure shows that, when the desired rate is small, it
unconstrained, codes allow the energy transfer properties of the is sufficient to use a type-0 RLL code with a small k, since,
code to be better adjusted to the receiver’s energy utilization with this choice, the resulting pattern of 0s and 1s in the
requirements [100]. codewords matches well the periodic requests of energy by the
We consider a point-to-point link and assume binary trans- energy harvesting receiver. The resulting improvement in terms
mission, in which “1” symbols carry energy while no energy is of energy transfer efficiency is significant. As the rate grows
carried by “0” symbols. Barring channel losses, which happen larger, one needs to increase the value of k, while keeping d as
with a given probability, the receiver can harvest the energy car- small as possible in order to increase the number of available
ried by the “1” symbols and store it in a battery. The receiver’s codewords and hence the rate [101].

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 371

Fig. 8. (a) Finite-state machine that defines the constraints that the codewords of a type-0 (d, k)-RLL codes must satisfy. (b) Maximum between probability of
underflow and overflow for unconstrained and type-0 constrained codes versus the information rate.

C. Energy Cooperation relay node at the end of the transmission session. This is the
rationale behind the concept of energy cooperation introduced
Consider a three-node relay network as shown in Fig. 9(a).
in [103]; see also [104]. In this scheme, we observe two types
Here, both the source node and the relay node harvest energy
of cooperation: the relay node cooperates at the signal level
from nature in the amounts Ei and Ēi in slot i. In addition,
by forwarding the source’s data packets to the destination
there is a wireless energy transfer unit that transfers δi amount
[105], and the source node cooperates at the energy level by
of energy from the source’s battery into the relays battery in
transferring some of its harvested energy to the relay node.
slot i. When the source node sends δi amount of energy, the
These two cooperation schemes combined yield the optimum
relay node receives αδi amount of energy into its battery,
scheme for the overall system. Note that, even though there was
where α < 1 accounts for the inefficiency of wireless energy
energy transfer inefficiency in the system represented by α < 1,
transfer. The goal of the network is to deliver the data packets
so long as the source node had sufficiently more energy than the
of the source node to the destination node, and in the process,
relay node, it is worth transferring some of the source energy to
maximize the end-to-end throughput of the system, from the
the relay.
source node to the destination node.
The concept of energy cooperation in two-way and multi-
The problem of power control to maximize the end-to-end
ple access channels have been studied in [106], [107], where
data throughput of this relay system without any energy transfer
was studied in [28], [29]. They showed that the optimum users transmit information over two-way and multiple access
scheme should work in the following way: The source node channels, and there is one-way energy transfer from one user
should transmit as many data packets as possible to the relay to the other, as motivated by practical scenarios such as RFID
node by using its harvested energy. This will cause a certain networks, where there is two-way information exchange but
packet arrival profile at the relay node. The relay node then one-way energy transfer from the reader to the RFID node.
should transmit as many of these data packets as possible to the References [106], [107] develop two-dimensional directional
destination node by using its harvested energy. One can note water-filling, where energy is distributed directionally due to
now that some of the packets that were delivered by the source energy causality and flows only from the past to the future
node to the relay node may not be forwarded to the destination and from the energy transferring user to the energy harvesting
node if the relay node does not have a sufficient amount of user. This creates two dimensions for energy flow, in time
harvested energy. Therefore, some of the data packets delivered and over users; and also directionality due to energy causality.
to the relay node will have no utility for the overall system Fig. 10 shows an example run of two-dimensional directional
(i.e., will not contribute towards the end-to-end throughput) water-filling algorithm; (a) shows the initial energy allocation
since they are not delivered to the destination node. Note that and (b) shows the final energy allocation. More recent works
the source node had spent some of its energy to deliver these [108], [109] generalize this approach to consider the case with
eventually undelivered data packets to the relay node. two-way energy cooperation in two-way and multiple access
One could think that the source node might as well not channels, and develop a separation-based approach that opti-
transmit those data packets to the relay node and keep its mizes wireless energy transfer and temporal power allocation
remaining energy. However, a better solution would be obtained separately to yield a global optimum solution for the overall
if the source node sent some of its remaining energy to the problem. As a final remark, we note that in energy cooperation
relay node via wireless energy transfer and used the rest of the described in this section, energy and information are sent by
remaining energy to send as many data packets as possible to different signals over orthogonal channels. This is reminiscent
the relay node. The source node will determine the amount of of the power splitting approach implemented at the receivers,
energy to be transferred to the relay node such that there will see for example [99], however, we implement power splitting
be no remaining energy and no remaining data packets at the at the transmitter as opposed to the receiver.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
372 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 9. Energy cooperation via wireless energy transfer. (a) Relay channel. (b) Two-way channel.

D. Interactive Exchange of Energy and Information Fig. 11(b) compares the achievable sum-rate obtained with
the mentioned scheme to an upper bound derived in [111]
In this section, we consider a multi-hop topology in which
versus the total number of energy units. Specifically, for the
the harvested energy can be reused for communications. Multi-
achievable sum-rate, we consider both a conventional codebook
hop networks with distinct source and destinations are consid-
design, in which all the codebooks have the same fraction of
ered under this assumption in [110]. Here, instead, we study a
0s and 1s irrespective of the energy state (labeled as “non-
two-way communication system, in which two nodes interact
adaptive” in the figure), and one in which the probabilities are
for the exchange of information and can harvest the received
optimized (labeled as “adaptive” in the figure). It can be seen
energy; see Fig. 11(a). To enable analysis and insights, we
that using conventional codebooks, which only aim at maxi-
assume that the two parties involved have a common clock
mizing information flow on a single link, leads to substantial
and that, at each time, a node can either send a “1” symbol,
performance loss. Instead, the proposed strategy with optimized
which carries one unit of energy, or a “0” symbol, which does
probabilities, which account also for the need to manage the
not carry any energy. Each node communicates in a full-duplex
energy flow in the two-way communication system, performs
manner, that is, at a given instant, it can simultaneously send
close to the upper bound. The latter is indeed achieved when
and receive an energy unit. The channel in one direction is
the number of energy units is large enough.
orthogonal to the channel in the other direction, and hence the
full-duplex channel is an ideal composition of two independent
unidirectional channels. We consider here the case in which VII. E NERGY H ARVESTING W IRELESS
the two nodes start with a given number of energy units in S ENSOR N ETWORKS
their batteries, which can neither be lost or replenished from Wireless sensor networks differ from the wireless systems
outside, and the binary channel in either direction is noiseless. considered thus far in that the devices need not only to cater to
Extensions can be found in [111]. the requirements of data transmission but also to those of source
To see that even this simple scenario offers relevant re- acquisition. Specifically, each sensor runs a source acquisition
search challenges, we observe the following. If there were that involves sensing, sampling and compression, and these
no limitation on the number of energy units, the nodes could operations often entail an energy cost that is comparable with
communicate 1 bit per channel use in either direction, given that of radio transmission; see, e.g., [112]–[115]. Therefore,
that the channels are ideal. However, if there is, say, one a proper allocation of the limited energy resources to source
energy unit available in the system, only the node that currently acquisition and transmission is necessary. As a result, the
possesses the energy unit can transmit a “1”, whereas the other techniques studied above that only adapt to the temporal varia-
node is forced to transmit a “0”. Therefore, the design of the tions of the energy harvesting process and of the transmission
communication strategy at the nodes should aim not only at channel must be revised in order to account for the time-varying
transferring the most information to the counterpart, but also properties of the source acquisition systems, e.g., for the quality
to facilitate energy transfer to enable communication in the of the measurements taken by the sensor.
reverse direction. In [111], a coding strategy is proposed for In order to concentrate on the main aspects of the problem,
this set-up that employs codebook multiplexing. The basic idea we focus on a system, studied in [116], in which a single
is that each node utilizes a different codebook for each energy sensor communicates with a single receiver. Time is slotted.
state. An energy state is defined by the current distribution The energy harvested in each time slot is assumed to follow an
of the energy units between the two nodes. Note that, under ergodic stationary process. As shown in Fig. 12(a), the sensor is
the given assumptions, both nodes are aware of the current equipped with a battery (energy queue) in which the harvested
energy state. Whenever a given energy state takes place, each energy is stored. The battery is assumed here to be of infinite
node sends the next bit from the corresponding codebook. size for simplicity of analysis. We observe that the case with
The key point is that, when the number of available energy multiple sensors is studied in [116], while multi-hop sensor
units at the node is large, the node should use a codebook networks are addressed in [117]. Moreover, a scenario with
with a larger fraction of “1” symbols in order to facilitate delay constraints is investigated in [118], where the optimal
energy transfer; instead, when the available energy is scarce, offline resource allocation policy is derived.
a codebook with a larger fraction of “0” symbols should In each time slot, the sensor acquires a time sequence
be used. of the phenomenon of interest. This is characterized by a

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 373

Fig. 10. Two-dimensional directional water-filling. (a) Initial energy allocation. (b) Final energy allocation.

measurement SNR, which evolves across the time slots as an classical notion of “throughput optimal” policies, where only
ergodic stationary process. The measurements are compressed the stability constraint is imposed. As shown in [116], a class
and stored into a data queue. The quality of compression is of optimal strategies exists that performs separate resource
defined by a distortion measure, such as the mean squared error. allocation optimizations for the source acquisition and data
Following [112], [113], we assume that the bit rate produced transmission. More precisely, without loss of optimality, the
by this source acquisition step depends on the energy allocated battery can be divided into two subcomponents, one used for
for source acquisition, on the desired distortion and on the source acquisition and one for data transmission: the energy
measurement SNR. In every time slot, the sensor also transmits harvested in each slot is split according to a fixed factor between
a number of bits from the data queue to the receiver over the two batteries. Moreover, the energy allocated to source
a fading channel with a given instantaneous channel SNR. acquisition from the corresponding battery in a given time slot
The channel SNR evolves across the time slots according to only depends on the measurement SNR, and not on the channel
an ergodic stationary process. The number of bits that are SNR; and, similarly, the energy used for data transmission
successfully transmitted depends on the energy allocated to data depends only on the channel SNR.
transmission and on the channel SNR as per, e.g., Shannon’s Fig. 12(b) shows the average distortion of the source recon-
channel capacity. struction at the receiver versus the variance of the harvesting
The energy management problem of interest is the following. process; see [116] for a full description of the set-up. The
Based on the statistics of the energy harvesting process, and performance of the optimal policy discussed above is compared
based on the current states of the measurement SNR, channel to a number of suboptimal policies that either use the energy
SNR and data queue, the energy management unit (EMU) in a greedy fashion or do not perform adaptation to the cur-
decides the compression distortion and how much energy to rent SNR states. Specifically, we consider strategies that use
allocate to source acquisition and data transmission (Encoder). immediately all the harvested energy with and without optimal
The performance criterion considered here is the stability of energy allocation between source acquisition and transmission
the data queue, which separates the data acquisition unit from depending on the SNR states (“no battery” and “no battery,
the data transmission block, under a constraint on the average adapt”, respectively); and strategies that use the battery to adapt
distortion of the measurements recovered at the receiver. the operation only of source acquisition (“source-only”) or of
For a given desired average distortion level, we wish to data transmission (“transmission-only”). It can be seen that
identify a class of policies that is able to stabilize the data an increased uncertainty about the harvested energy, i.e., a
queue while satisfying the distortion constraint whenever pos- larger variance, degrades the quality of the reconstruction at
sible. Note that this definition of optimality generalizes the the receiver. Moreover, adapting the energy usage to the current

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
374 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 11. (a) Two-way system with energy reuse. (b) Achievable sum-rates with non-adaptive and adaptive codebooks and upper bound.

measurement and channel SNRs, especially when leveraging ment, the energy arrival processes at different transmitters are
the possibility to store energy in the battery, leads to significant assumed to be i.i.d. random sequences with mean λe called the
performance gains. energy-arrival rate. For simplicity, the batteries of harvesters
are assumed to have infinite capacity.
VIII. L ARGE -S CALE W IRELESS N ETWORKS In the steady state, depending on the energy availability, each
W ITH E NERGY H ARVESTING transmitter is turned on or off with probability ρ and (1 − ρ),
respectively. The probability is shown in [120] to have a simple
The preceding sections focus on algorithms and protocols form: ρ = min(1, λe /P). This allows the network throughput to
for point-to-point links or small-scale systems powered by be written as
energy harvesting. From the perspective of network designers
and operators, it is interesting to study the effects of en- R = λ0 ρ log(1 + θ). (14)
ergy harvesting on large-scale networks based on models that
capture network architectures, node distributions and realistic Note that λ0 ρ is the active-transmitter density. The expression
channel/interference models, which is the theme of this section. of ρ suggests that the density can be controlled by varying P. On
In particular, we model mobile ad hoc networks (MANETs) and one hand, large density and power can cause strong interference
cellular networks powered by energy harvesting and relate their and as a result violate the outage constraint. On the other
performance to the characteristics of energy arrival processes or hand, too sparse transmitters reduce network throughput and
energy fields. Typical renewable energy fields such as wind and too low power leads to incorrect decoding. Thus, transmission
solar power exhibit both temporal and spatial variations and are power should be optimized under the outage constraint and
commonly modeled as random fields [119]. Preceding sections the criterion of maximum throughput. As shown in [120], for
address the temporal variation and how to counteract it by relatively sparse networks with a sufficiently high energy arrival
adaptive transmission. On the other hand, the spatial variation rate, the optimal power is one that allows all transmitters to be
of renewables can significantly degrade network coverage and active with probability one. Otherwise, the probability should
hence has to be accounted for in network modeling and design. be smaller than one as derived.
To this end, a tractable model for renewable-energy fields is
discussed in the sequel and applied to investigate the coverage B. Cellular Networks With Energy Harvesting
of renewable powered networks.
Reference [121] addresses the effect of the spatial variation
of the (renewable) energy field, e.g., solar or wind power, on
A. Mobile Ad Hoc Networks With Energy Harvesting
the coverage of a cellular network. To this end, a tractable
The spatial throughput of a MANET with energy harvest- energy field model is proposed in [121] where the energy
ing is analyzed in [120]. The transmitters are modeled as a intensity at a particular location is given by spatial combining
homogeneous Poisson point process (PPP) with density λ0 in of Poisson distributed “energy centers” with fixed maximum
the horizontal plane. Each communicates with an affiliated intensities, which is known as a Boolean random function.
receiver located at a unit distance. Given a fixed encoding rate The combining factors are determined by an exponential-decay
log(1 + θ), reliable decoding of a data packet at a receiver function of squared distance. Such a function is commonly
requires the receive SINR to exceed a threshold θ except for used in spatial interpolation for atmospherical mapping [122]
a small outage probability, called an outage constraint. The and solar-field estimation [123]. The advantage of the energy-
transmission powers are assumed to be identical and equal to field model is that its distribution is controlled by only two
P. Nodes in the MANET are small devices like sensors and parameters, namely the energy-center density denoted as λe and
wearable computing devices. Given their deployment environ- the exponential rate of the weight function denoted as ν. It can

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 375

Fig. 12. (a) System model for an energy harvesting wireless sensor link. (b) Average distortion D of the source reconstruction at the receiver versus the variance
of the harvesting process.

be observed from the illustrations in Fig. 13 that the energy field Also adopting the approach of using stochastic geometry
is almost flat for large λe and ν or otherwise highly random. for network modeling and analysis, the requirements on the
Next, the downlink network is modeled using the traditional MPT network deployment are analyzed for different MPT
hexagonal-cell model but with renewables powered BS’s. The technologies.
network is assumed to operate in the noise-limited regime Last, heterogeneous cellular networks with energy harvesting
that is most interesting for renewable powered network since is modeled and studied in [126] where multi-tier renewable
network performance in the regime is sensitive to transmission powered base stations are modeled as independent PPP’s. Con-
power and hence harvested energy. In addition, an outage con- sidering small base stations deployed in an urban area, energy
straint is applied on downlink transmissions. We consider two arrival processes for BS’s are suitably assumed to be indepen-
scenarios of harvester deployments. First, each BS is powered dent unlike the energy field discussed earlier. This allows the
by a single on-site harvester. Based on the energy field model, on/off states of BS’s to be modeled as independent Bernoulli
the outage probability can be related to the coverage probability random variables and then their impact on the network coverage
of a Boolean model comprising random disks. The result shows performance can be quantified mathematically using stochastic
that the outage probability decreases exponentially with the geometry.
product λe ν.
Next, consider the scenario where energy harvested by dense
IX. E NERGY C ONSUMPTION M ODELS
harvesters is collected by aggregators that distribute power
to nearby BS’s. The aggregators effectively perform spatial Because energy-harvesting systems typically operate at short
averaging of the energy field and consequently stabilize the distances (a few meters or less), the energy consumed in
transmission power of BS’s. It is shown in [121] that as the transmitter/receiver circuitry can be comparable, or can even
number of energy harvesters connected to a single aggregator dominate, the energy consumed in transmissions. Thus, it is im-
increases, the transmission power converges to a constant by portant to understand energy harvesting communications in the
virtue of the law of large numbers. This renders the outage context of total (transmit + circuit) energy minimization. The
probability to be either one or zero. Nevertheless, since har- difficulty in developing a comprehensive theory of total energy
vesters cannot afford high-voltage power transmission, total minimization lies in obtaining models for energy consumed in
energy-transmission loss can become significant as the number circuitry that are simple enough for analysis, and yet accurate
of harvesters for energy aggregation increases. Such loss can enough to yield relevant estimates of energy consumption. How
be regulated by increasing the voltage at harvesters following a can we abstract the energy consumed by various possible circuit
derived scaling law. algorithms, implementation architectures, and implementation
technologies? As a first step, the authors in [73], [127], model
the transmitter and receivers as black-boxes that consume a
C. Other Types of Energy Harvesting Networks
fixed amount of energy per unit time powered “on”. Although
The interesting idea of opportunistic energy harvesting is the formulations in [73], [127] are considerably different, their
explored in [124] for cognitive radio networks where passive conclusions are the same: since keeping systems powered “on”
secondary nodes not only opportunistically access the spectrum consumes circuit energy, transmissions should be “bursty,” i.e.,
of primary nodes but also harvest energy from radiation by the both receiver and transmitter are turned “off” for some time in
latter whenever it is possible. Coexisting networks are modeled order to reduce circuit energy. However, the energy required in
as overlaid spatial point processes. Based on this model, the transmission increases exponentially with burstiness (because
transmission capacity of the secondary network adopting the capacity scales logarithmically in power), so the transmission
strategy of opportunistic energy harvesting is characterized and cannot be too bursty. That the optimal burstiness is non-zero
optimized over the node density and transmission power. is surprising: traditional transmit power analysis [128] for a
A closely related piece of work is presented in [125] for non-fading channel predicts that the transmission rate should
cellular networks where mobiles are wireless recharged by be made as small as possible, and the signals least bursty, for
dedicated power stations via microwave power transfer (MPT). minimum energy consumption.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
376 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

Fig. 13. The energy field. (a) λe = 1 and ν = 0.1. (b) λe = 10 and ν = 1.

While insightful, the black-box model has its shortcomings: to communicate at a constant gap from the Shannon limit as
because it lumps together all of the power used in processing the Pe → 0. Numerical evaluation of regular LDPC codes shows
signal, the black-box model does not yield much insight into that they can achieve order optimal total energy in this model
code choice or decoder design. Does the code-choice matter? of energy consumption; see Fig. 14(b).
Recent empirical work has shown that energy consumption of However, the empirical work in [129] showed that the wiring
the decoding circuitry changes substantially with the choice of energy (ignored in [131]–[133]) can be a significant fraction
the code [129]. Thus a more detailed model of circuit energy of the circuit energy, and can further change substantially with
is needed in order to guide the choice of the communication change in the code design ([129] uses LDPC codes where the
strategy as well as the circuit design. One model that has code-degree and code-length is kept constant but the code-girth
commonly been used to understand tradeoffs in circuit com- is varied). Using the information-friction model to abstract the
putations is the “VLSI model of computation,” developed by wiring energy, [130] (and the ensuing work in [134]) showed
Thompson in 1980. The model, illustrated in Fig. 14(a), was that the wiring energy diverges to infinity much faster than
used by Thompson to derive fundamental limits on the tradeoffs the node energy, an observation consistent with that of circuit
between the required total wiring-length and the number of practitioners. Further, the total-energy optimal communication
clock-cycles. The underlying idea is: when the computation to strategy is to communicate farther and farther away from the
be performed is such that a substantial amount of information channel capacity as the error-probability converges to zero, and
needs to be moved around on the circuit, one has to either make thus keeping the rate close to capacity can actually increase the
wires long so that information can be communicated farther in total power.
the same clock-cycle, or increase the number of clock-cycles. Thus the total energy minimizing communication strategy,
These wiring and clock-cycle tradeoffs provide an approximate as well as the scaling limits on total energy as Pe → 0, change
understanding of the required energy: if we assume that each drastically when circuit energy is also incorporated. We be-
wire is used in each clock-cycle, the product of the total wiring- lieve it is important to model and understand the effects of
length and the number of clock-cycles yields a bound on the circuit energy (both node and wiring energy) in the context of
required energy. Even this approximate understanding has been energy-harvesting as well for the following reasons: Ignoring
made more accurate in the more recent “information-friction” circuit energy can yield optimistic results and strategies that
model [130], which also generalizes the VLSI model so that it is may significantly underestimate total energy requirements in
applicable to asynchronous models of computation, and to other practice, and be highly inefficient in a total (transmit + circuit)
substrates of computation (e.g., circuit links made of carbon energy sense. In order to understand the energy required in a
nanotubes, optical fibers, physical matter transport, or wireless, computation, it is not sufficient to compute the Turing com-
or even axons that connect neurons in the nervous system). plexity (i.e., count the number of operations) or node energy.
The VLSI model has been used to understand energy and Understanding wiring energy (i.e., the energy required to move
complexity of encoding and decoding in the system. Focusing information between circuit elements) can be key because it
on the energy consumed in the computational nodes at the tends to dominate the node energy in asymptotics. Thompson’s
decoder (and ignoring the wiring energy), [131]–[133] showed VLSI model, and ensuing improvements, can prove useful in
that the required number of clock-cycles at the decoder (im- estimating energy requirements of circuitry in an order sense,
plemented in the VLSI model) diverges to infinity as Pe → 0, which can be extremely helpful in deciding the communication
and as the communication rate approaches the channel capacity. strategy. Energy-harvesting circuit might itself need energy, and
This is used to obtain a lower bound on total energy. It turns when systems use joint information and energy transfer, there
out that to optimize this lower bound, the optimal strategy is might be a further increase in cost for circuit energy in order

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 377

Fig. 14. (a) Thompson’s VLSI model of computing. (b) Energy with a model of computation that only considers energy consumed in computational nodes. The
upper bounds are obtained using a node model on a (4,6) regular LDPC code.

to harvest energy from the same signal. Models for energy- R EFERENCES
harvesting circuits can thus reveal much about performance of [1] J. A. Paradiso and T. Starner, “Energy scavenging for mobile and wire-
these systems in the real world. less electronics,” IEEE Pervasive Comput., vol. 4, no. 1, pp. 18–27,
Jan.–Mar. 2005.
[2] V. Raghunathan, C. Schurgers, P. Sung, and M. B. Srivastava, “Energy-
aware wireless microsensor networks,” IEEE Signal Process. Mag.,
vol. 19, no. 2, pp. 40–50, Mar. 2002.
X. C ONCLUSION AND F ORWARD L OOK [3] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes: Sur-
vey and implications,” IEEE Commun. Surveys Tuts., vol. 13, no. 3,
In this article, we have summarized recent advances that have pp. 443–461, 3rd Quart. 2011.
taken place in the broad area of energy harvesting wireless [4] S. Roundy, “Energy scavenging for wireless sensor nodes with a focus
communication networks. We have covered a variety of topics on vibration to electricity conversion,” Ph.D. dissertation, Dept. EECS,
Univ. Berkeley, Berkeley, CA, USA, May, 2003.
ranging from information-theoretic physical layer performance [5] J. Rabaey, M. J. Ammer, J. L. Silva, D. Patel, and S. Roundy, “Picaradio
limits to scheduling policies and medium access control proto- supports ad hoc ultra-low power wireless networking,” IEEE Comput.,
vol. 33, no. 7, pp. 42–48, Jul. 2000.
cols, as well as the newly emerging paradigm of energy transfer [6] A. Kansal, J. Hsu, S. Zahedi, and M. B. Srivastava, “Power management
and cooperation that occur in addition or simultaneous with in- in energy harvesting sensor networks,” ACM Trans. Embeded Comput.
formation transfer for such networks. Models and results under Syst., vol. 6, no. 4, p. 32, 2007.
[7] Microstrain Press Release. Microstrain wins navy contract for develop-
a variety of network structures, those with single and multiple ing energy-harvesting wireless sensors, Jul. 2006. [Online]. Available:
hops as well as small and large scale have been addressed. We http://www.microstrain.com/news/article-65.aspx
have also presented models for total energy consumption. [8] I. A. Ieropoulos, J. Greenman, C. Melhuish, and J. Hart, “Comparative
study of three types of microbial fuel cells,” Enzyme Microbial Technol.,
It is worth noting that energy harvesting wireless networks vol. 37, no. 2, pp. 238–245, Jul. 2005.
simultaneously present new theoretical challenges and those [9] J. Lei, R. Yates, and L. Greenstein, “A generic model for optimizing
that stem from physical phenomena and practical concerns. single-hop transmission policy of replenishable sensors,” IEEE Trans.
Wireless Commun., vol. 8, no. 2, pp. 547–551, Feb. 2009.
As such, the area provides a rich set of possibilities for ob- [10] M. Gatzianas, L. Georgiadis, and L. Tassiulas, “Control of wireless
taining design insights from mathematical formulations which networks with rechargeable batteries,” IEEE Trans. Wireless Commun.,
take practical considerations into account. These considerations vol. 9, no. 2, pp. 581–593, Feb. 2010.
[11] V. Sharma, U. Mukherji, V. Joseph, and S. Gupta, “Optimal energy
include such physical properties as storage imperfections, con- management policies for energy harvesting sensor nodes,” IEEE Trans.
sumption models, processing costs, as well as realistic mod- Wireless Commun., vol. 9, no. 4, pp. 1326–1336, Apr. 2010.
[12] J. Yang and S. Ulukus, “Optimal packet scheduling in an energy har-
eling such as causal energy harvesting profiles. Additionally, vesting communication system,” IEEE Trans. Commun., vol. 60, no. 1,
the area of energy and information transfer provides exciting pp. 220–230, Jan. 2012.
possibilities to further adapt the network operation and improve [13] K. Tutuncuoglu and A. Yener, “Optimum transmission policies for bat-
tery limited energy harvesting nodes,” IEEE Trans. Wireless Commun.,
its performance. The possible improvement therein is closely vol. 11, no. 3, pp. 1180–1189, Mar. 2012.
tied to the efficiency of energy transfer and hence to the device [14] O. Ozel, K. Tutuncuoglu, J. Yang, S. Ulukus, and A. Yener, “Transmis-
and circuit technologies, connecting the theory again to the sion with energy harvesting nodes in fading wireless channels: Optimal
policies,” IEEE J. Sel. Areas Commun., vol. 29, no. 8, pp. 1732–1743,
real world. To this end, we conclude by stating that the future Sep. 2011.
challenges for energy harvesting wireless networks lie not only [15] C. K. Ho and R. Zhang, “Optimal energy allocation for wireless com-
in advancements in various layers of network design starting munications with energy harvesting constraints,” IEEE Trans. Signal
Process., vol. 60, no. 9, pp. 4808–4818, Sep. 2012.
from signal processing and communications physical layer all [16] J. Yang and S. Ulukus, “Optimal packet scheduling in a multiple access
the way to the networking layer, but also in embracing the truly channel with energy harvesting transmitters,” J. Commun. Netw., vol. 14,
interdisciplinary nature of the energy harvesting wireless net- no. 2, pp. 140–150, Apr. 2012.
[17] J. Yang, O. Ozel, and S. Ulukus, “Broadcasting with an energy harvest-
works integrating with the advances from circuits and devices ing rechargeable transmitter,” IEEE Trans. Wireless Commun., vol. 11,
that harvest and transfer energy. no. 2, pp. 571–583, Feb. 2012.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
378 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

[18] M. A. Antepli, E. Uysal-Biyikoglu, and H. Erkal, “Optimal packet [44] C. Tapparello, O. Simeone, and M. Rossi, “Dynamic compression–
scheduling on an energy harvesting broadcast link,” IEEE J. Sel. Areas transmission for energy-harvesting multihop networks with correlated
Commun., vol. 29, no. 8, pp. 1721–1731, Sep. 2011. sources,” IEEE/ACM Trans. Netw., vol. 22, no. 6, pp. 1729–1741,
[19] O. Ozel, J. Yang, and S. Ulukus, “Optimal broadcast scheduling for an Dec. 2014.
energy harvesting rechargeable transmitter with a finite capacity bat- [45] Y. Luo, J. Zhang, and K. B. Letaief, “Training optimization for en-
tery,” IEEE Trans. Wireless Commun., vol. 11, no. 6, pp. 2193–2203, ergy harvesting communication systems,” in Proc. IEEE GLOBECOM,
Jun. 2012. Dec. 2012, pp. 3365–3370.
[20] O. Ozel, J. Yang, and S. Ulukus, “Optimal transmission schemes for par- [46] M. Gregori and M. Payaró, “Efficient data transmission for an en-
allel and fading broadcast channels with an energy harvesting recharge- ergy harvesting node with battery capacity constraint,” in Proc. IEEE
able transmitter,” Comput. Commun., vol. 36, no. 12, pp. 1360–1371, GLOBECOM, 2011, pp. 1–6.
Jul. 2013. [47] S. Luo, R. Zhang, and T. J. Lim, “Diversity analysis of an outage mini-
[21] K. Tutuncuoglu and A. Yener, “Optimal power policy for energy har- mizing energy harvesting wireless protocol in Rayleigh fading,” in Proc.
vesting transmitters with inefficient energy storage,” in Proc. 46th Annu. SPCOM, 2012, pp. 1–5.
Conf. CISS, Mar. 2012, pp. 1–6. [48] S. Lee, K. Huang, and R. Zhang, “Cognitive energy harvesting
[22] O. Ozel, K. Shahzad, and S. Ulukus, “Optimal scheduling for energy and transmission from a network perspective,” in Proc. IEEE ICCS,
harvesting transmitters with hybrid energy storage,” in IEEE ISIT, Nov. 2012, pp. 225–229.
Jul. 2013, pp. 1784–1788. [49] I. Ahmed, A. Ikhlef, R. Schober, and R. K. Mallik, “Power alloca-
[23] O. Ozel and S. Ulukus, “Achieving AWGN capacity under stochastic tion in energy harvesting relay systems,” in Proc. IEEE 75th VTC, 2012,
energy harvesting,” IEEE Trans. Inf. Theory, vol. 58, no. 10, pp. 6471– pp. 1–5.
6483, Oct. 2012. [50] P. Castiglione, O. Simeone, E. Erkip, and T. Zemen, “Energy-harvesting
[24] O. Ozel and S. Ulukus, “AWGN channel under time-varying amplitude for source–channel coding in cyber–physical systems,” in Proc. 4th
constraints with causal information at the transmitter,” in Proc. Asilomar IEEE CAMSAP, 2011, pp. 189–192.
Conf., Nov. 2011, pp. 1–5. [51] S. Reddy and C. R. Murthy, “Dual-stage power management algorithms
[25] K. Tutuncuoglu, O. Ozel, A. Yener, and S. Ulukus, “Binary energy for energy harvesting sensors,” IEEE Trans. Wireless Commun., vol. 11,
harvesting channel with finite energy storage,” in Proc. IEEE ISIT, no. 4, pp. 1434–1445, Apr. 2012.
Jul. 2013, pp. 1591–1595. [52] R. Srivastava and C. E. Koksal, “Basic performance limits and trade-
[26] W. Mao and B. Hassibi, “On the capacity of a communication system offs in energy-harvesting sensor nodes with finite data and energy
with energy harvesting and a limited battery,” in Proc. IEEE ISIT, storage,” IEEE/ACM Trans. Netw., vol. 21, no. 4, pp. 1049–1062,
Jul. 2013, pp. 1789–1793. Aug. 2013.
[27] B. Devillers and D. Gunduz, “A general framework for the optimization [53] A. Nayyar, T. Basar, D. Teneketzis, and V. V. Veeravalli, “Optimal strate-
of energy harvesting communication systems with battery imperfec- gies for communication and remote estimation with an energy harvesting
tions,” J. Commun. Netw., vol. 14, no. 2, pp. 130–139, Apr. 2012. sensor,” IEEE Trans. Autom. Control, vol. 58, no. 9, pp. 2246–2260,
[28] D. Gunduz and B. Devillers, “Two-hop communication with energy Sep. 2013.
harvesting,” in Proc. 14th IEEE CAMSAP, Dec. 2011, pp. 201–204. [54] P. Blasco, D. Gunduz, and M. Dohler, “A learning theoretic approach to
[29] C. Huang, R. Zhang, and S. Cui, “Throughput maximization for the energy harvesting communication system optimization,” in Proc. IEEE
Gaussian relay channel with energy harvesting constraints,” IEEE J. Sel. Globecom Workshop, Dec. 2012, pp. 1657–1662.
Areas Commun., vol. 31, no. 8, pp. 1469–1479, Aug. 2013. [55] R. Vaze, “Competitive ratio analysis of online algorithms to minimize
[30] S. Luo, R. Zhang, and T. J. Lim, “Optimal save-then-transmit protocol packet transmission time in energy harvesting communication system,”
for energy harvesting wireless transmitters,” in Proc. IEEE ISIT, 2012, in Proc. IEEE INFOCOM, Apr. 2013, pp. 115–1123.
pp. 955–959. [56] P. Mitran, “On optimal online policies in energy harvesting systems
[31] S. Reddy and C. R. Murthy, “Duty cycling and power management with a for compound Poisson energy arrivals,” in Proc. IEEE ISIT, 2012,
network of energy harvesting sensors,” in Proc. 4th IEEE Int. Workshop pp. 960–964.
CAMSAP, 2011, pp. 205–208. [57] D. Gunduz, K. Stamatiou, N. Michelusi, and M. Zorzi, “Designing
[32] K. Huang, “Throughput of wireless networks powered by energy har- intelligent energy harvesting communication systems,” IEEE Commun.
vesting,” in Conf. Rec. 45th Asilomar Conf. Signals, Syst. Comput., 2011, Mag., vol. 52, no. 1, pp. 210–216, Jan. 2014.
pp. 8–12. [58] C. E. Shannon, “A mathematical theory of communication,” Bell. Syst.
[33] O. Orhan and E. Erkip, “Optimal transmission policies for energy har- Tech. J., vol. 27, no. 3, pp. 379–423, Jul. 1948.
vesting two-hop networks,” in Proc. 46th Annu. Conf. CISS, Mar. 2012, [59] V. Anantharam and S. Verdu, “Bits through queues,” IEEE Trans. Inf.
pp. 1–6. Theory, vol. 42, no. 1, pp. 4–18, Jan. 1996.
[34] O. Orhan and E. Erkip, “Energy harvesting two-hop networks: Optimal [60] V. Jog and V. Anantharam, “An energy harvesting AWGN channel with
policies for the multi-energy arrival case,” in Proc. 35th IEEE Sarnoff a finite battery,” in Proc. IEEE ISIT, Jun. 2014, pp. 806–810.
Symp., May 2012, pp. 1–6. [61] Y. Dong and A. Ozgur, “Approximate capacity of energy harvesting
[35] O. Orhan and E. Erkip, “Throughput maximization for energy harvesting communication with finite battery,” in Proc. IEEE ISIT, Jun. 2014,
two-hop networks,” in Proc. IEEE ISIT, Jul. 2013, pp. 1596–1600. pp. 801–805.
[36] B. Varan and A. Yener, “Two-hop networks with energy harvesting: The [62] R. Rajesh, V. Sharma, and P. Viswanath, “Capacity of Gaussian channels
(non-)impact of buffer size,” in Proc. IEEE GLOBALSIP, Dec. 2013, with energy harvesting and processing cost,” IEEE Trans. Inf. Theory,
pp. 399–402. vol. 60, no. 5, pp. 2563–2575, May 2014.
[37] B. Varan, K. Tutuncuoglu, and A. Yener, “Energy harvesting two-way [63] R. A Raghuvir, D. Rajan, and M. D. Srinath, “Capacity of the multiple
half-duplex relay channel with decode-and-forward relaying: Optimum access channel in energy harvesting wireless networks,” in Proc. IEEE
power policies,” in Proc. 18th Int. Conf. DSP, Jul. 2013, pp. 1–6. WCNC, Apr. 2012, pp. 898–902.
[38] B. Varan and A. Yener, “The energy harvesting two-way decode-and- [64] R. Rajesh, P. K. Deekshith, and V. Sharma, “Capacity of a Gaussian
forward relay channel with stochastic data arrivals,” in Proc. IEEE MAC with energy harvesting transmit nodes,” in Proc. Inf. Theory Appl.
GLOBALSIP, Dec. 2013, pp. 371–374. Workshop, Feb. 2012, pp. 338–343.
[39] O. Orhan, D. Gunduz, and E. Erkip, “Throughput maximization for an [65] J. G. Smith, “The information capacity of amplitude and variance-
energy harvesting system with processing cost,” in Proc. IEEE ITW, constrained scalar Gaussian channels,” Inf. Control, vol. 18, no. 3,
Sep. 2012, pp. 84–88. pp. 203–219, Apr. 1971.
[40] J. Xu and R. Zhang, “Throughput optimal policies for energy harvesting [66] C. E. Shannon, “Channels with side information at the transmitter,” IBM
wireless transmitters with non-ideal circuit power,” IEEE J. Sel. Areas J. Res. Develop., vol. 2, no. 4, pp. 289–293, Oct. 1958.
Commun., vol. 32, no. 2, pp. 322–332, Feb. 2014. [67] O. Ozel and S. Ulukus, “On the capacity region of the Gaussian
[41] O. Orhan, D. Gunduz, and E. Erkip, “Energy harvesting broadband MAC with batteryless energy harvesting transmitters,” in Proc. IEEE
communication systems with processing energy cost,” arXiv:1312.0054 GLOBECOM, Dec. 2012, pp. 2385–2390.
2013. [68] T. Weissman, “Capacity of channels with action-dependent
[42] C. Huang, R. Zhang, and S. Cui, “Optimal power allocation for outage states,” IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5396–5411,
probability minimization in fading channels with energy harvesting con- Nov. 2010.
straints,” IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 1074–1087, [69] F. M. J. Willems, “Signaling for the Gaussian channel with side informa-
Feb. 2014. tion at the transmitter,” in Proc. IEEE ISIT, Jun. 2000, p. 348.
[43] R. Vaze, “Transmission capacity of wireless ad hoc networks with energy [70] D. M. Arnold, H.-A. Loeliger, P. O. Vontobel, A. Kavcic, and W. Zeng,
harvesting nodes,” in Proc. IEEE GLOBALSIP, Dec. 2013, pp. 353–358. “Simulation-based computation of information rates for channels with

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 379

memory,” IEEE Trans. Inf. Theory, vol. 52, no. 8, pp. 3498–3508, [97] L. Liu, R. Zhang, and K.-C. Chua, “Wireless information transfer
Aug. 2006. with opportunistic energy harvesting,” IEEE Trans. Wireless Commun.,
[71] O. Kaya and S. Ulukus, “Achieving the capacity region of fading CDMA vol. 12, no. 1, pp. 288–300, Jan. 2013.
channels via generalized iterative waterfilling,” IEEE Trans. Wireless [98] J. Park and B. Clerckx, “Joint wireless information and energy transfer
Commun., vol. 5, no. 11, pp. 3215–3223, Nov. 2006. in a two-user MIMO interference channel,” IEEE Trans. Wireless Com-
[72] W. Yu, W. Rhee, S. Boyd, and J. M. Cioffi, “Iterative water-filling for mun., vol. 12, no. 8, pp. 4210–4221, Aug. 2013.
Gaussian vector multiple-access channels,” IEEE Trans. Inf. Theory, [99] X. Zhou, R. Zhang, and C. K. Ho, “Wireless information and power
vol. 50, no. 1, pp. 145–152, Jan. 2004. transfer: Architecture design and rate-energy tradeoff,” IEEE Trans.
[73] P. Youssef-Massaad, L. Zheng, and M. Medard, “Bursty transmis- Commun., vol. 61, no. 11, pp. 4754–4767, Nov. 2013.
sion and glue pouring: On wireless channels with overhead costs,” [100] A. M. Fouladgar, O. Simeone, and E. Erkip, “Constrained codes for joint
IEEE Trans. Wireless Commun., vol. 7, no. 12, pp. 5188–5194, energy and information transfer,” IEEE Trans. Commun., vol. 62, no. 6,
Dec. 2008. pp. 2121–2131, Jun. 2014.
[74] F. Iannello, O. Simeone, P. Popovski, and U. Spagnolini, “Energy group- [101] B. H. Marcus, R. M. Roth, and P. H. Siegel, An Introduction to Coding
based dynamic framed ALOHA for wireless networks with energy har- for Constrained Systems, 2001. [Online]. Available: http://www.math.
vesting,” in Proc. 46th Annu. Conf. CISS, Mar. 2012, pp. 1–6. ubc.ca/~marcus/Handbook/
[75] N. Michelusi, K. Stamatiou, and M. Zorzi, “Transmission policies for [102] A. I. Barbero, E. Rosnes, G. Yang, and O. Ytrehus, “Constrained codes
energy harvesting sensors with time-correlated energy supply,” IEEE for passive RFID communication,” in Proc. ITA, 2011, pp. 1–9.
Trans. Commun., vol. 61, no. 7, pp. 2988–3001, Jul. 2013. [103] B. Gurakan, O. Ozel, J. Yang, and S. Ulukus, “Energy cooperation
[76] V. Sharma, U. Mukherji, and V. Joseph, “Efficient energy management in energy harvesting wireless communications,” in Proc. IEEE ISIT,
policies for networks with energy harvesting sensor nodes,” in Proc. 46th Jul. 2012, pp. 965–969.
Annu. Allerton Conf. Allerton Commun., Control, Comput., Sep. 2008, [104] B. Gurakan, O. Ozel, J. Yang, and S. Ulukus, “Energy cooperation in
pp. 375–383. energy harvesting communications,” IEEE Trans. Commun., vol. 61,
[77] N. Michelusi, K. Stamatiou, and M. Zorzi, “On optimal transmis- no. 12, pp. 4884–4898, Dec. 2013.
sion policies for energy harvesting devices,” in Proc. ITA, 2012, [105] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation
pp. 249–254. diversity—Part I: System description,” IEEE Trans. Commun., vol. 51,
[78] D. Bertsekas, Dynamic Programming and Optimal Control. Belmont, no. 11, pp. 1927–1938, Nov. 2003.
MA, USA: Athena Scientific, 2005. [106] B. Gurakan, O. Ozel, J. Yang, and S. Ulukus, “Two-way
[79] J. Chiasson and B. Vairamohan, “Estimating the state of charge of a and multiple access energy harvesting systems with energy
battery,” IEEE Trans. Control Syst. Technol., vol. 13, no. 3, pp. 465–470, cooperation,” in Conf. Rec. 46th Asilomar Signals, Syst. Comput.,
May 2005. Nov. 2012, pp. 58–62.
[80] N. Michelusi, K. Stamatiou, L. Badia, and M. Zorzi, “Operation poli- [107] B. Gurakan, O. Ozel, J. Yang, and S. Ulukus, “Energy cooperation
cies for energy harvesting devices with imperfect state-of-charge knowl- in energy harvesting two-way communications,” in Proc. IEEE ICC,
edge,” in Proc. IEEE ICC, 2012, pp. 5782–5787. Jun. 2013, pp. 3126–3130.
[81] N. Michelusi, L. Badia, R. Carli, K. Stamatiou, and M. Zorzi, [108] K. Tutuncuoglu and A. Yener, “Multiple access and two-way
“Correlated energy generation and imperfect state-of-charge knowl-
channels with energy harvesting and bidirectional energy cooperation,”
edge in energy harvesting devices,” in Proc. 8th IWCMC, Aug. 2012,
in Proc. ITA, Feb. 2013, pp. 1–8.
pp. 401–406.
[109] K. Tutuncuoglu and A. Yener, “Cooperative energy harvesting
[82] D. Del Testa, N. Michelusi, and M. Zorzi, “On optimal transmission
communications with relaying and energy sharing,” in Proc. IEEE ITW,
policies for energy harvesting devices: The case of two users,” in Proc.
Sep. 2013, pp. 1–5.
10th ISWCS, 2013, pp. 1–5.
[110] A. M. Fouladgar and O. Simeone, “Information and energy flows in
[83] N. Michelusi, L. Badia, R. Carli, L. Corradini, and M. Zorzi, “Energy
graphical networks with energy transfer and reuse,” IEEE Wireless
management policies for harvesting-based wireless sensor devices with
battery degradation,” IEEE Trans. Commun., vol. 61, no. 12, pp. 4934– Commun. Lett., vol. 2, no. 4, pp. 371–374, Aug. 2013.
4947, Dec. 2013. [111] P. Popovski, A. M. Fouladgar, and O. Simeone, “Interactive joint trans-
[84] N. Jaggi, K. Kar, and A. Krishnamurthy, “Rechargeable sensor activa- fer of energy and information,” IEEE Trans. Commun., vol. 61, no. 5,
tion under temporally correlated events,” in Proc. IEEE WiOpt, 2007, pp. 2086–2097, May 2013.
pp. 1–10. [112] Z. He and D. Wu, “Resource allocation and performance analysis of
[85] N. Michelusi and M. Zorzi, “Optimal random multiaccess in wireless video sensors,” IEEE Trans. Circuits Syst. Video Technol.,
energy harvesting wireless sensor networks,” in Proc. IEEE ICC, 2013, vol. 16, no. 5, pp. 590–599, May 2006.
pp. 463–468. [113] X. Lu, E. Erkip, Y. Wang, and D. Goodman, “Power efficient multimedia
[86] F. C. Schoute, “Dynamic frame length ALOHA,” IEEE Trans. Commun., communication over wireless channels,” IEEE J. Sel. Areas Commun.,
vol. COM-31, no. 4, pp. 565–568, Apr. 1983. vol. 21, no. 10, pp. 1738–1751, Dec. 2003.
[87] F. Iannello, O. Simeone, and U. Spagnolini, “Medium access control [114] V. Joseph, V. Sharma, U. Mukherji, and M. Kashyap, “Joint power con-
protocols for wireless sensor networks with energy harvesting,” IEEE trol, scheduling and routing for multicast in multihop energy harvesting
Trans. Commun., vol. 60, no. 5, pp. 1381–1389, May 2012. sensor networks,” in Proc. ICUMT, Oct. 2009, pp. 1–8.
[88] F. Iannello, O. Simeone, and U. Spagnolini, “Optimality of myopic [115] L. Lin, N. B. Shroff, and R. Srikant, “Asymptotically optimal energy-
scheduling and whittle indexability for energy harvesting sensors,” in aware routing for multihop wireless networks with renewable energy
Proc. 46th Annu. CISS, Mar. 2012, pp. 1–6. sources,” IEEE/ACM Trans. Netw., vol. 15, no. 5, pp. 1021–1034,
[89] R. Sarpeshkar, Ultra Low Power Bioelectronics: Fundamentals, Oct. 2007.
Biomedical Applications, and Bio-inspired Systems. Cambridge, U.K.: [116] P. Castiglione, O. Simeone, E. Erkip, and T. Zemen, “Energy manage-
Cambridge Univ. Press, 2010. ment policies for energy-neutral source–channel coding,” IEEE Trans.
[90] L. R. Varshney, “Transporting information and energy simultaneously,” Commun., vol. 60, no. 9, pp. 2668–2678, Sep. 2012.
in Proc. IEEE ISIT, Jul. 2008. [117] C. Tapparello, O. Simeone, and M. Rossi, “Dynamic compression–
[91] M. Gastpar, “On capacity under receive and spatial spectrum-sharing transmission for energy-harvesting multihop networks with correlated
constraints,” IEEE Trans. Inf. Theory, vol. 53, no. 2, pp. 471–487, sources,” IEEE/ACM Trans. Netw., vol. 22, no. 6, pp. 1729–1741,
Feb. 2007. Dec. 2014.
[92] I. Csiszar and J. Korner, Information Theory: Coding Theorems for [118] O. Orhan, D. Gunduz, and E. Erkip, “Delay-constrained distortion min-
Discrete Memoryless Systems. New York, NY, USA: Academic, imization for energy harvesting transmission over a fading channel,” in
1981. Proc. IEEE ISIT, 2013, pp. 1794–1798.
[93] T. S. Han, Information-Spectrum Methods in Information Theory. New [119] N. A. C. Cressie, Statistics for Spatial Data. Hoboken, NJ, USA: Wiley,
York, NY, USA: Springer-Verlag, 2003. 1993.
[94] P. Grover and A. Sahai, “Shannon meets Tesla: Wireless information and [120] K. Huang, “Spatial throughput of mobile ad hoc networks with energy
power transfer,” in Proc. IEEE ISIT, Jul. 2010, pp. 2363–2367. harvesting,” IEEE Trans. Inf. Theory, vol. 59, no. 11, pp. 7597–7612,
[95] T. M. Cover and J. A. Thomas, Elements of Information Theory, 1st ed. Nov. 2013.
New York, NY, USA: Wiley, 1991. [121] K. Huang, M. Kountouris, and V. O. K. Li, “Renewable powered cellular
[96] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wire- networks: Energy field modeling and network coverage,” submitted to
less information and power transfer,” IEEE Trans. Wireless Commun., IEEE Trans. Wireless Commun. [Online]. Available: http://arxiv.org/abs/
vol. 12, no. 5, pp. 1989–2001, May 2013. 1408.3198.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
380 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 33, NO. 3, MARCH 2015

[122] S. L. Barnes, “A technique for maximizing details in numerical weather Aylin Yener (S’91–M’00–SM’13–F’14) received
map analysis,” J. Appl. Meteorol., vol. 3, no. 4, pp. 396–409, Aug. 1964. the B.Sc. degree in electrical and electronics en-
[123] Z. Sen, “Solar energy in progress and future research trends,” Progr. gineering and the B.Sc. degree in physics from
Energy Combustion Sci., vol. 30, no. 4, pp. 367–416, 2004. Bogazici University, Istanbul, Turkey, and the M.S.
[124] S. Lee, R. Zhang, and K. Huang, “Opportunistic wireless energy har- and Ph.D. degrees in electrical and computer engi-
vesting in cognitive radio networks,” IEEE Trans. Wireless Commun., neering from Rutgers University, New Brunswick,
vol. 12, no. 9, pp. 4788–4799, Sep. 2013. NJ, USA. She has been a Professor of electrical
[125] K. Huang and V. K. N. Lau, “Enabling wireless power transfer in cel- engineering at The Pennsylvania State University,
lular networks: Architecture, modelling and deployment,” IEEE Trans. University Park, PA, USA, since 2010, where she
Wireless Commun., vol. 13, no. 2, pp. 902–912, Feb. 2014. joined the faculty as an Assistant Professor in 2002.
[126] H. S. Dhillon, Y. Li, P. Nuggehalli, Z. Pi, and J. G. Andrews, “Fundamen- During the academic year 2008–2009, she was a
tals of heterogeneous cellular networks with energy harvesting,” IEEE Visiting Associate Professor with the Department of Electrical Engineering,
Trans. Wireless Commun., vol. 13, no. 5, pp. 2782–2797, May 2014. Stanford University, Stanford, CA, USA. Her research interests are in informa-
[127] S. Cui, A. J. Goldsmith, and A. Bahai, “Energy constrained modulation tion theory, communication theory, and network science with recent emphasis
optimization,” IEEE Trans. Wireless Commun., vol. 4, no. 5, pp. 1–11, on green communications and information security.
Sep. 2005. Dr. Yener has been on the Board of Governors of the IEEE Information The-
[128] S. Verdu, “On channel capacity per unit cost,” IEEE Trans. Inf. Theory, ory Society as its Treasurer (2012–2014). She served as the Student Committee
vol. 36, no. 5, pp. 1019–1030, Sep. 1990. Chair for the IEEE Information Theory Society in 2007–2011 and was the co-
[129] K. Ganesan, P. Grover, and J. M. Rabaey, “The power cost of overde- founder of the Annual School of Information Theory in North America co-
signing codes,” in Proc. IEEE SiPS, Oct. 2011, pp. 128–133. organizing the school in 2008, 2009, and 2010. She was a Technical (Co)-Chair
[130] P. Grover, “‘Information-friction’ and its impact on minimum energy per for various symposia/tracks at IEEE ICC, PIMRC, VTC, WCNC, and Asilomar
communicated bit,” in Proc. IEEE ISIT, Jul. 2013, pp. 2513–2517. (2005–2014). She served as an Editor of the IEEE T RANSACTIONS ON
[131] P. Grover, “Bounds on the tradeoff between rate and complexity for C OMMUNICATIONS (2009–2012), an Editor and an Editorial Advisory Board
sparse-graph codes,” in Proc. IEEE ITW, 2007, pp. 196–201. Member of the IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS
[132] P. Grover and A. Sahai, “Green codes: Energy-efficient short-range com- (2001–2012), a Guest Editor of the IEEE T RANSACTIONS ON I NFORMATION
munication,” in IEEE ISIT, Jul. 2008, pp. 1178–1182. F ORENSICS AND S ECURITY (2011), and a Guest Editor of the IEEE J OURNAL
[133] P. Grover, K. A. Woyach, and A. Sahai, “Towards a communication- ON S ELECTED A REAS IN C OMMUNICATIONS (2015). She was a recipient of
theoretic understanding of system-level power consumption,” IEEE J. the NSF CAREER Award in 2003, the Best Paper Award in Communication
Sel. Areas Commun., vol. 29, no. 8, pp. 1744–1755, Sep. 2011. Theory at the IEEE International Conference on Communications in 2010, the
[134] C. Blake and F. R. Kschischang, “Energy of decoding algorithms,” in Penn State Engineering Alumni Society (PSEAS) Outstanding Research Award
Proc. 13th Can. Workshop Inf. Theory, 2013, pp. 1–5. in 2010, the IEEE Marconi Prize Paper Award in 2014, the PSEAS Premier
Research Award in 2014, and the Leonard A. Doggett Award for Outstanding
Writing in Electrical Engineering at Penn State in 2014.

Sennur Ulukus (S’90–M’98) received the B.S. and Elza Erkip (S’93–M’96–SM’05–F’11) received the
M.S. degrees in electrical and electronics engineer- B.S. degree in electrical and electronics engi-
ing from Bilkent University, Ankara, Turkey, and the neering from Middle East Technical University,
Ph.D. degree in electrical and computer engineer- Ankara, Turkey, and the M.S. and Ph.D. degrees
ing from Rutgers University, New Brunswick, NJ, in electrical engineering from Stanford University,
USA. She is a Professor of electrical and computer Stanford, CA, USA. She is currently a Professor
engineering at the University of Maryland (UMD), of electrical and computer engineering with New
College Park, MD, USA, where she also holds a joint York University Polytechnic School of Engineering,
appointment with the Institute for Systems Research Brooklyn, NY, USA. Her research interests are in
(ISR). Prior to joining UMD, she was a Senior information theory, communication theory, and wire-
Technical Staff Member at AT&T Labs-Research. less communications.
Her research interests are in wireless communication theory and networking, Dr. Erkip is a member of the Science Academy Society of Turkey and is
network information theory for wireless communications, signal processing among the Thomson Reuters 2014 Edition of Highly Cited Researchers. She
for wireless communications, information theoretic physical layer security, and was a General Chair for the IEEE International Symposium of Information
energy harvesting communications. Theory in 2013, a Technical Program Chair for the International Symposium
Dr. Ulukus served as the TPC Co-Chair for the 2014 IEEE PIMRC, Commu- on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks
nication Theory Symposium at 2014 IEEE Globecom, Communication Theory (WiOpt) in 2011, a Technical Program Chair for the IEEE GLOBECOM
Symposium at 2013 IEEE ICC, Physical-Layer Security Workshop at 2011 Communication Theory Symposium in 2009, the Publications Chair for the
IEEE Globecom, Physical-Layer Security Workshop at 2011 IEEE ICC, 2011 IEEE Information Theory Workshop, Taormina, in 2009, the Technical Area
Communication Theory Workshop (IEEE CTW), Wireless Communications Chair for the MIMO Communications and Signal Processing track of Asilomar
Symposium at 2010 IEEE ICC, Medium Access Control Track at 2008 IEEE Conference on Signals, Systems, and Computers in 2007, and a Technical
WCNC, and Communication Theory Symposium at 2007 IEEE Globecom. Program Chair for the IEEE Communication Theory Workshop in 2006. She
She was the Secretary of the IEEE Communication Theory Technical Com- is currently a Member of the Board of Governors of the IEEE Information
mittee (CTTC) in 2007–2009. She served as an Associate Editor of the Theory Society and a Guest Editor of the IEEE J OURNAL ON S ELECTED
IEEE T RANSACTIONS ON I NFORMATION T HEORY (2007–2010) and IEEE A REAS IN C OMMUNICATIONS. She was a Distinguished Lecturer of the
T RANSACTIONS ON C OMMUNICATIONS (2003–2007). She served as a Guest IEEE Information Theory Society from 2013 to 2014, an Associate Editor of
Editor of the IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS the IEEE T RANSACTIONS ON I NFORMATION T HEORY from 2009 to 2011,
for the special issue on wireless communications powered by energy harvesting an Associate Editor of the IEEE T RANSACTIONS ON C OMMUNICATIONS
and wireless energy transfer (2015), Journal of Communications and Net- from 2006 to 2008, a Publications Editor of the IEEE T RANSACTIONS ON
works for the special issue on energy harvesting in wireless networks (2012), I NFORMATION T HEORY from 2006 to 2008, and a Guest Editor of the IEEE
IEEE T RANSACTIONS ON I NFORMATION T HEORY for the special issue on S IGNAL P ROCESSING M AGAZINE in 2007. She was a recipient of the NSF
interference networks (2011), and IEEE J OURNAL ON S ELECTED A REAS IN CAREER award in 2001, the IEEE Communications Society Stephen O. Rice
C OMMUNICATIONS for the special issue on multiuser detection for advanced Paper Prize in 2004, the IEEE ICC Communication Theory Symposium Best
communication systems and networks (2008). She was a recipient of the Paper Award in 2007, and the IEEE Communications Society Award for
2003 IEEE Marconi Prize Paper Award in Wireless Communications, a 2005 Advances in Communication in 2013. She co-authored a paper that received the
NSF CAREER Award, the 2010–2011 ISR Outstanding Systems Engineering IEEE International Symposium on Information Theory Student Paper Award
Faculty Award, and the 2012 George Corcoran Education Award. in 2007.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.
ULUKUS et al.: ENERGY HARVESTING WIRELESS COMMUNICATIONS 381

Osvaldo Simeone (S’02–M’05–SM’14) received the Pulkit Grover (S’01–M’13) received the B.Tech.
M.Sc. degree (with honors) and the Ph.D. degree and M. Tech. degrees from the Indian Institute of
in information engineering from the Politecnico di Technology Kanpur, Kanpur, India, in 2003 and
Milano, Milan, Italy, in 2001 and 2005, respectively. 2005, respectively, and the Ph.D. degree from the
He is currently with the Center for Wireless Com- University of California (UC), Berkeley, CA, USA,
munications and Signal Processing Research, New in 2010. He is an Assistant Professor at Carnegie
Jersey Institute of Technology, Newark, NJ, USA, Mellon University (CMU), Pittsburgh, PA, USA.
where he is an Associate Professor. His current re- Prior to joining CMU in 2013, he was a Postdoctoral
search interests concern wireless communications, Researcher at Stanford. He is interested in inter-
information theory, and machine learning. He cur- disciplinary research directed toward developing a
rently serves as an Editor for the IEEE T RANS - science of information for making decentralized
ACTIONS ON I NFORMATION T HEORY . He was a co-recipient of Best Paper sensing, communication, and computing systems (including biomedical sys-
Awards at the IEEE SPAWC 2007 and IEEE WRECOM 2007. tems) energy-efficient and stable. He was the recipient of the 2010 Best Student
Paper Award at the IEEE Conference in Decision and Control (CDC); a Finalist
at the 2010 Best Student Paper Award at the IEEE International Symposium on
Information Theory (ISIT); the 2011 Eli Jury Award from UC Berkeley; the
2012 Leonard G. Abraham Best Paper Award from the IEEE Communications
Society; a 2014 Best Paper Award at the International Symposium on Integrated
Circuits (ISIC); and a 2014 NSF CAREER Award.

Kaibin Huang (M’08–SM’13) received the B.Eng.

degree (with first-class honors) and the M.Eng. de-
gree from the National University of Singapore,
Michele Zorzi (S’89–M’95–SM’98–F’07) received Singapore, in 1998 and 2000, respectively, and the
the Laurea and Ph.D. degrees in electrical engineer- Ph.D. degree from The University of Texas at Austin
ing from the University of Padova, Padova, Italy, in (UT Austin), TX, USA, in 2008, all in electrical
1990 and 1994, respectively. During academic year engineering.
1992–1993, he was on leave at the University of Since January 2014, he has been an Assistant
California, San Diego (UCSD), working on multi- Professor with the Department of Electrical and
ple access in mobile radio networks. In 1993, he Electronic Engineering (EEE), The University of
joined the faculty of the Dipartimento di Elettronica Hong Kong, Pokfulam, Hong Kong. He is an Adjunct
e Informazione, Politecnico di Milano, Milan, Italy. Professor in the School of EEE, Yonsei University, Korea. He used to be a
After spending three years with the Center for Wire- faculty member in the Department of Applied Mathematics (AMA), Hong
less Communications at UCSD, in 1998 he joined Kong Polytechnic University (PolyU), and the Department of EEE, Yonsei
the School of Engineering, University of Ferrara, Italy, where he became a University. He had been a Postdoctoral Research Fellow in the Department
Professor in 2000. Since November 2003, he has been on the faculty of the of Electrical and Computer Engineering, Hong Kong University of Science
Department of Information Engineering, University of Padova. His current and Technology, from June 2008 to February 2009 and an Associate Scientist
research interests include performance evaluation in mobile communications at the Institute for Infocomm Research in Singapore from November 1999 to
systems, random access in wireless networks, ad hoc and sensor networks, July 2004. His research interests focus on the analysis and design of wireless
Internet-of-Things, energy constrained communications protocols, cognitive networks using stochastic geometry and multi-antenna techniques.
networks, and underwater communications and networking. Dr. Huang is an Elected Member of the SPCOM Technical Committee of
Prof. Zorzi served as a Member-at-Large of the Board of Governors of the the IEEE Signal Processing Society. He frequently serves on the Technical
IEEE Communications Society from 2009 to 2011 and is currently its Director Program Committees for major IEEE conferences in wireless communications.
of Education. He was the Editor-in-Chief of the IEEE W IRELESS C OMMU - He was the Chairman for the Communication Theory Symposium of IEEE
NICATIONS from 2003 to 2005 and the Editor-in-Chief of the IEEE T RANS - GLOBECOM 2014 and the Advanced Topics in Wireless Communications
ACTIONS ON C OMMUNICATIONS from 2008 to 2011 and is the founding Symposium of IEEE/CIC ICCC 2014, the Technical Co-Chair for IEEE CTW
Editor-in-Chief of the IEEE T RANSACTIONS ON C OGNITIVE C OMMUNICA - 2013, and the Track Chair for IEEE PIMRC 2015, IEE VTC Spring 2013,
TIONS AND N ETWORKING . He has also been an Editor of several journals and Asilomar 2011, and IEEE WCNC 2011. He is a Guest Editor of the IEEE
a member of the Organizing or the Technical Program Committee for many J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS and an Editor of the
international conferences, as well as a Guest Editor for special issues in IEEE IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS, IEEE W IRELESS
P ERSONAL C OMMUNICATIONS (“Energy Management in Personal Com- C OMMUNICATIONS L ETTERS, and IEEE/KICS J OURNAL OF C OMMUNI -
munications Systems”), IEEE N ETWORK (“Video over Mobile Networks”), CATION AND N ETWORKS . He was a recipient of the Outstanding Teaching
and IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS (“Mul- Award from Yonsei, Motorola Partnerships in Research Grant, the University
timedia Network Radios” and “Underwater Wireless Communications and Continuing Fellowship from UT Austin, and a Best Paper Award from IEEE
Networking”). GLOBECOM 2006 and PolyU AMA in 2013.

Authorized licensed use limited to: IEEE Xplore. Downloaded on April 21,2021 at 21:14:26 UTC from IEEE Xplore. Restrictions apply.