Professional Documents
Culture Documents
Abstract—Designing mobiles to harvest ambient energy such as This paper considers a mobile ad hoc network (MANET)
kinetic activities or electromagnetic radiation will enable wireless where transmitters are modeled as a homogeneous Poisson
networks to be self-sustaining. In this paper, the spatial throughput point process (PPP). Energy arrives randomly at a transmitter
of a mobile ad hoc network powered by energy harvesting is ana-
lyzed using a stochastic-geometry model. In this model, transmit-
with a fixed average rate, called the energy-arrival rate. The
ters are distributed as a Poisson point process and energy arrives at energy-arrival process is modeled as an i.i.d. sequence of
each transmitter randomly with a uniform average rate called the random variables and different processes are assumed indepen-
energy arrival rate. Upon harvesting sufficient energy, each trans- dent. Each transmitter deploys an energy harvester that stores
mitter transmits with fixed power to an intended receiver under an arriving energy in a rechargeable battery. Upon harvesting
outage-probability constraint for a target signal-to-interference- sufficient energy, a transmitter transmits with fixed power to an
and-noise ratio. It is assumed that transmitters store energy in bat-
teries with infinite capacity. By applying the random-walk theory, intended receiver under an outage-probability constraint for a
the probability that a transmitter transmits, called the transmission target signal-to-interference-and-noise ratio (SINR). Based on
probability, is proved to be equal to the smaller of one and the ratio the above model, the network spatial throughput is maximized
between the energy-arrival rate and transmission power. This re- by optimizing transmission power for a given energy-arrival
sult and tools from stochastic geometry are applied to maximize rate.
the network throughput for a given energy-arrival rate by opti-
mizing transmission power. The maximum network throughput is
shown to be proportional to the optimal transmission probability, A. Prior Work and Motivation
which is equal to one if the transmitter density is below a derived
function of the energy-arrival rate or otherwise is smaller than one
The fluctuation in harvested energy due to random energy
and solves a given polynomial equation. Last, the limits of the max- arrivals requires redesigning existing transmission algorithms
imum network throughput are obtained for the extreme cases of for wireless communication systems. Assuming infinitely back-
high energy-arrival rates and sparse/dense networks. logged data, existing work focuses on adapting transmission
Index Terms—Energy harvesting, mobile ad hoc networks, power to channel states and the temporal profile of energy ar-
mobile communication, power control, stochastic processes, rivals to maximize the system throughput [2]–[5]. For single-
throughput. user systems, the optimal power-control policies are shown to
be variations of the classic water-filling policy such that the
causality of energy arrivals and finite battery capacity are ac-
I. INTRODUCTION counted for [2], [3]. Adaptive transmission for broadcast chan-
nels with energy harvesting has been also investigated [4], [5].
with only data queues, namely minimizing the average trans- as the probability that a transmitter transmits and the network in-
mission power under a performance constraint such as fixed terference temperature as the maximum active transmitter den-
packet-transmission delay for single-user systems [11], [12], sity under the outage-probability constraint. Let denote the
allowed queuing delay for downlink systems [13], and given transmitter density, the energy-arrival rate, the transmis-
traffic load in wireless networks [14]. sion power, and a nonnegative random variable representing
Wireless networks with energy harvesting have been studied the amount of energy harvested by a typical harvester in an arbi-
[15]–[17]. For a wireless sensor network with energy harvesting trary slot.1 Note that and the density of active trans-
and based on a simple channel model that omits channel noise mitters is equal to .
and path loss, the probability that a sensor successfully trans- The main contributions of this paper are summarized as
mits a data packet to a fusion center is analyzed in [15] for dif- follows.
ferent multiple-access protocols including time-division mul- 1) Assume infinite battery capacity. Using the law of large
tiple access and Aloha like random access. Managing traffic numbers and random-walk theory, it is proved that is
load in time and space is important for wireless sensor net- equal to the smaller of and one. It is worth men-
works to be self-sustaining through energy harvesting. There- tioning that the tractable analysis relies on assuming
fore, distributive strategies are proposed in [16] for adapting the suboptimal fixed-power transmission. To the best
traffic load to the spatial-and-temporal energy profile and evalu- of the author’s knowledge, the aforementioned result
ated using a network prototype. For a two-user interference net- is unknown from existing work that mostly focuses on
work with energy harvesting, the data-and-energy arrivals are designing the optimal adaptive transmission algorithms
modeled as Bernoulli processes and the stability region is char- [2]–[5], [7]–[10].
acterized such that it comprises all data-rate pairs under the con- 2) Consider the case of finite battery capacity. Bounds on
straint of finite data-queue lengths [17]. In view of prior work, are derived, which converge to the results stated above as
there are few results that quantify the tradeoff between the net- the battery capacity increases. Moreover, two special cases
work throughput and the energy-arrival rate though such results are considered. If is bounded and no larger than , it is
specify the fundamental limit of the network performance. This shown that is equal to so long as the battery ca-
tradeoff is investigated in the sequel using a stochastic-geom- pacity is larger than . If is a discrete random variable,
etry approach. is analyzed using Markov-chain theory.
Stochastic geometry provides a set of powerful mathemat- 3) Assume infinite battery capacity. By applying the said re-
ical tools for modeling and designing wireless networks [18]. sult on transmission probability and tools from stochastic
MANETs based on random access and carrier-sensing multiple geometry, the network throughput is maximized by opti-
access have been modeled using the PPPs [19], [20] and Matern mizing for given . Consider the condition that is
hard-core processes [21], respectively. Cellular networks have smaller than the network interference temperature evalu-
been shown to be suitably modeled using the Poisson Voronoi ated for equal and . If this condition holds, the max-
tessellation [22]. Models of coexisting networks can be con- imum throughput is shown to be
structed by superimposing multiple point processes [23], [24].
Stochastic-geometry models of wireless networks have been
employed to quantify the network-performance gains due to
physical-layer techniques such as opportunistic transmission where is the target SINR. If the aforementioned condition
[25], bandwidth partitioning [26], successive interference is not satisfied
cancellation [27], and multiantenna techniques [28]–[32].
The performance metric typically considered in the literature
is the network spatial throughput under a constraint on the
outage probability for a target SINR, which is also adopted where the optimal transmission power is larger than
in this paper. Using this metric, most prior work focuses on and solves a derived polynomial equation.
deriving the outage probability using techniques such as the 4) Furthermore, the limits of the maximum network
Laplace transform [19], [22] and probabilistic inequalities throughput are obtained for the extreme cases of high
[20], [25]. This paper considers a MANET with Poisson energy-arrival rates ( ) and dense networks
distributed transmitters similar to the existing literature (see ( ). Specifically,
e.g., [20]). However, the transmitters in the current network
model are powered by energy harvesting instead of reliable
power supplies as in prior work. The consideration of energy
harvesting introduces several new design issues including the
aforementioned tradeoff between the network throughput and
energy-arrival rate, the corresponding optimization of trans- where is a positive constant determined by the max-
mission power, and the effect of finite energy storage, which imum outage probability.
are investigated in the sequel. The remainder of this paper is organized as follows. The net-
work model and performance metric are described in Section II.
B. Contributions and Organization The transmission probability is analyzed in Section III. The
results are applied to maximize the network throughput in
For exposition, a few definitions and notations are provided
as follows. Time is slotted. Define the transmission probability 1A typical point is selected from a spatial point process by uniform sampling.
HUANG: SPATIAL THROUGHPUT OF MOBILE AD HOC NETWORKS POWERED BY ENERGY HARVESTING 7599
(1)
B. Performance Metric
Assume infinitely backlogged and packetized data. The trans-
mission probability can be written as
Fig. 1. Single-antenna transmitters in the MANET are modeled as a homoge-
neous PPP in the horizontal plane. Each transmitter is powered by an energy har-
vester and transmits to an intended single-antenna receiver at an unit distance. (2)
III. TRANSMISSION PROBABILITY amount of energy stored in the typical harvester in excess of a
threshold in the th slot:
A. Infinite Battery Capacity
Deriving transmission probability requires analyzing the dis-
(14)
tribution of battery levels at energy harvesters. By substituting
into (1), the battery level at the typical energy harvester
with infinite battery capacity evolves as where represents the probability density function of .
The function can be bounded as shown in Lemma 3,
(8) which is proved in Appendix C using Lemma 2.
Lemma 3: Given infinite battery capacity and , the
The distribution of can be related to the threshold-crossing energy-overshoot function satisfies
probability for a random walk as follows. Denote the in-
stants when the battery level crosses the threshold from
below as , namely that and
for . These time instants are grouped into the with .
set . Moreover, define the random variable Using Lemma 3, the main result of this section is readily ob-
and two random processes and as tained as shown below.
Theorem 1: Given infinite battery capacity, the transmission
(9) probability is
(10)
(12)
(16)
where as defined in Assumption 1 with is the
positive root of the cumulant generating function of . Since , it follows that given , there exist
Lemma 1: Given infinite battery capacity, the battery level and such that for all . Using
satisfies this fact and substituting (16) into (15),
(13)
(17)
with .
The proof of Lemma 2 is given in Appendix B. Define the As is arbitrary and , the desired result for the case of
energy-overshoot function as the expected follows from (17) and letting .
HUANG: SPATIAL THROUGHPUT OF MOBILE AD HOC NETWORKS POWERED BY ENERGY HARVESTING 7601
Next, consider the case of . The expected total the case of , can be bounded by the same upper
amounts of harvested and transmitted energy by the th slot bound on (see Lemma 3) as shown below.
differ by the battery level in the th slot, namely Lemma 4: Given finite battery capacity and , the
energy-overshoot function satisfies
(18)
with .
Since , Lemma 4 is proved in Appendix D. Next, the tail probability
of can be bounded for the case of as shown in the
following lemma, which is proved in Appendix E.
Lemma 5: Given finite battery capacity and , the
(19) distribution of the battery level satisfies
(22)
with .
2) If
(23)
As a result, with .
3) If
(21)
(25) Fig. 2. Markov chain modeling the battery level at the typical transmitter for
the special case of binary energy arrivals and .
(26) where .
Proof: By expanding (1),
Fig. 3. Relation between the nominal node density and the maximum outage probability .
probability matrix with the th element given by and The probability for battery overflow is zero and the transmission
the steady-state-probability row vector with the th element probability is .
given by . Applying Perron–Frobenius theorem [33],
IV. NETWORK THROUGHPUT
(32)
In this section, using results on the transmission probability
The stationary probabilities can be computed by solving derived in the preceding section, the network throughput
(32) under the constraint . Given , the trans- is maximized by optimizing transmission power under the
mission probability is obtained as . outage-probability constraint and assuming infinite battery
The stationary probabilities can be derived in closed capacity. It is straightforward though tedious to extend the net-
form for the simple case of binary energy arrivals, namely that work-throughput analysis to the case of finite battery capacity
. Then energy-arrival rate is using related results from the last section.3 Such an extension
. The corresponding transition probabilities are modified from provides little new insight and thus is omitted.
(31) as
A. Maximum Network Throughput
or
or To characterize the network throughput, transmission power
otherwise. is related to the active transmitter density under the outage-
probability constraint. To this end, we define a parameter ,
Based on above transition probabilities, the Markov chain called the nominal node density, as the density of a homoge-
is illustrated in Fig. 2. The stationary distribution satis- neous PPP such that
fies the following equations obtained from (32) and the above
expression for :
(34)
Fig. 4. Admissible set is sketched as the shaded region that contains all combinations of active transmitter density and transmission power that satisfy
the outage-probability constraint. Given infinite battery capacity and fixed transmitter density , is a monotone-decreasing function of as plotted with
the thick lines based on Theorem 1 for the cases of and , where the intersections are indicated by and , respectively. Note
that and an arbitrary value in give the optimal transmission power that maximizes the network throughput for the corresponding cases.
Lemma 6: Consider the homogeneous PPP . Remark 5: The network interference temperature
The process with is a homogeneous specifies the maximum density of interferers a link can tol-
PPP with density . erate without violating the outage-probability constraint. This
Define the admissible set as all combinations of quantity is analogous to the interference temperature in cog-
that satisfy the outage-probability constraint nitive-radio systems that measures the maximum amount of
additional interference for a certain frequency band without
(35) significantly degrading the reliability of communications
A combination is admissible if it belongs to . To derive therein [37]. On one hand, the network for large is interfer-
, since follows the same distribution as , in (5) ence-limited and further increasing does not contribute any
can be rewritten as network-throughput gain. Correspondingly, it can be observed
from (38) that saturates as :
(39)
that is fixed at for all and a strictly-mono- . Combining and shows that the
tone-decreasing function for . Let correspond to the assumption of violates the constraint in (44), proving
intersection between the curve and the admissible set that . Then, applying Theorem 1 yields the desired
. Then, under the outage-probability constraint, is max- result in (41). Last, since and are strictly-mono-
imized at , which corresponds to based on the defi- tone-decreasing and strictly-monotone-increasing functions, re-
nition in (6) since it is proportional to . Note that the claim is spectively, the solution for the problem in (44) must satisfy
based on the assumption or else need not hold (see Re- or equivalently solves the polynomial
mark 10). Depending on if is larger or smaller than , equation in (42), completing the proof.
and have different expressions as shown in Theorem 2 that Remark 8: For the case of , the network is rela-
states the main result of this section. tively sparse and is sufficiently large such that it is optimal as
Theorem 2: Given infinite battery capacity, the maximum well as feasible for all transmitters to transmit with probability
network throughput and the optimal transmission power 1, resulting in the network throughput in (40). For the case of
are specified as follows. , the network is relatively dense and high transmis-
1) If , sion power is required for satisfying the outage-probability con-
(40) straint. Consequently, not all transmitters can transmit simulta-
neously, corresponding to the optimal transmission probability
and is an arbitrary value in the range smaller than one and the network throughput in (41).
. Remark 9: With and the last coefficient at the right-
2) If , hand side being negative, the polynomial equation in (42) has
at least one strictly positive solution that gives for the case
(41) of . For the special case of , the polynomial
equation in (42) is quadratic and solving it gives in closed
form as shown below.
where and solves the following polynomial
Corollary 1: Given infinite battery capacity, and
equation:
, the optimal transmission power is
(42)
The inequality in the theorem statement can be proved since the outage-probability constraint requires that (see
by contradiction as follows. Assume that . This as- Lemma 7)
sumption results in by applying Theorem 1. More-
over, is obtained using the aforementioned
monotonicity of , and hence given that (46)
7606 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 59, NO. 11, NOVEMBER 2013
where (46) follows from (39) and that is a monotone-in- density without violating the outage-probability constraint. For
creasing function. Combining the two upper bounds on in this reason, the network-throughput limit in (48) for the cur-
(43) and (45) gives rent case is proportional to the transmitter density. However,
for the case of , active transmitter density reaches
(47) the network-interference temperature and cannot be further in-
creased. Consequently, the corresponding network-throughput
limit in (48) is independent of the transmitter density.
For a high energy-arrival rate, equality is achieved in (47) as
Consider a sparse network ( ). It is optimal for each
shown below.
transmitter to transmit with probability 1 by setting the transmis-
Proposition 4: Given infinite battery capacity, as the energy-
sion power if or otherwise with probability
arrival rate , the maximum network throughput con-
by setting to be equal to (see Theorem 1). The cor-
verges as
responding network throughputs are and
, respectively, of which both diminish as
(48) . Next, for a dense network ( ), the maximum
network throughput is specified as follow.
Proof: First, consider the case of . Set Proposition 5: Given infinite battery capacity, as the trans-
with . This results in according to Theorem mitter density , the maximum network throughput con-
1. Consequently, and hence from the verges as
assumption about . Combining this inequality and (39) yields (52)
that as along with . It follows
that as , the combination is Proof: Set as the following function of :
admissible according to Lemma 7. This proves the equality in
(47) for the current case. (53)
Next, consider the case of . Given this strict
inequality, there exists such that that is shown shortly to achieve the limit of in (52). Given
(53), there exists such that for all .
(49) Therefore, it follows from (7) and Theorem 1 that
Set as
(50)
(56)
Letting proves the equality in (48) for the current case,
completing the proof. By comparing (55) and (56), there exists such that
Remark 11: Given a high energy arrival rate and infinite bat- for all . This proves the admis-
tery capacity, in the steady state, transmitters always have suffi- sibility of in (53) and in (55) as , com-
cient energy for transmission. Therefore, the expression in (48) pleting the proof.
also specifies the maximum network throughput of a MANET Remark 13: The rate of total energy harvested per unit area
with reliable power supplies instead of energy harvesting. is as . The linear growth of the rate with
Remark 12: For the case of , the active trans- increasing is due to the fact that the harvester density is equal
mitter density is below the network-interference temperature to . However, more aggressive energy harvesting in a dense
even though all transmitters transmit with probability 1. There- network does not continuously increase the network throughput
fore, there is margin for further increasing active transmitter that saturates at high transmitter power as the network becomes
HUANG: SPATIAL THROUGHPUT OF MOBILE AD HOC NETWORKS POWERED BY ENERGY HARVESTING 7607
Fig. 5. Maximum network throughput versus energy-arrival rate for the optimal transmission power, infinite battery capacity, and the transmitter density
.
Fig. 6. Maximum network throughput versus transmitter density for the optimal transmission power, infinite battery capacity, and the energy-arrival rate
.
interference limited (see Proposition 5). This issue may be re- density is equal to . Based on this setup, the values of
solved by using an alternative multiple-access protocol such as are computed using the Monte Carlo method that yields the plot
frequency-hopping multiple access that reduces the density of in Fig. 3. In addition, all numerical results are based on the SINR
simultaneous cochannel transmitters. threshold and the path-loss exponent .
The distribution of the energy arrival process is
V. NUMERICAL RESULTS specified as follows. Let denote an i.i.d. sequence of
The nominal node density is fixed as 0.05 for all numer- random variables following the chi-squared distribution with
ical results, corresponding to the maximum outage probability degrees of freedom (DoF) and mean equal to
. The relation between and is obtained by simula- . Let and hence has mean and variance
tion based on the following procedure (see e.g., [36]). The sum- . The chosen distribution of allows its variance to be
mation over the PPP in (34) is approximated by the signal controlled by varying while the mean of is fixed. Note that
power measured at the origin due to unit-power transmissions converges to a constant in probability as by the
by transmitters uniformly distributed in a disk. The number of law of large numbers.
transmitters follows the Poisson distribution with mean 200 and Infinite battery capacity is assumed for the numerical results
the disk radius is adjusted such that the expected transmitter presented in Figs. 5 and 6. In Fig. 5, the maximum network
7608 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 59, NO. 11, NOVEMBER 2013
Fig. 7. Comparison between the average tail probability of the battery level, , evaluated by simulation and its upper bound com-
puted based on Lemma 2 for infinite battery capacity, the DoF of the energy-arrival process , the energy-arrival rate , and the transmission power
.
Fig. 8. Transmission probability versus transmission power for finite battery capacity , the DoF of the energy-arrival process ,
the energy-arrival rate , and the transmitter density .
throughput computed using Theorem 2 is plotted against compared in Fig. 7 with its upper bound from Lemma 2 given
the increasing energy-arrival rate for the transmitter den- infinite battery capacity, , , and . The bound
sity . It can be observed from Fig. 5 is observed to be loose but sufficient for the analysis. The
that grows as increases and saturates for large . The similar observation and remark also hold for the upper bound
limits agree with those computed using Proposition 4, namely on the energy-overshoot function as given in Lemma 3
0.04 bit/(s Hz unit-area) for and 0.048 bit/(s Hz unit- and the numerical results are omitted for brevity.
area) for . In addition, Fig. 5 shows that in Next, consider the case of finite battery capacity. In Fig. 8,
a denser network (i.e., ), reaches its limit more the transmission probability obtained by simulation is plotted
rapidly as increases. against increasing transmission power for finite battery ca-
Fig. 6 shows the curves of versus for , pacity , , , and
which are obtained using Theorem 2. As increases and . It is found that is sufficiently large such that the
regardless of the value of , is observed to converge to values of closely match those for the case of infinite battery ca-
the limit 0.048 bit/(s Hz unit-area) predicted by Proposition 5. pacity as computed using Theorem 1. As observed from Fig. 8,
Moreover, it is observed from Fig. 6 that larger results in finite battery capacity degrades significantly only when and
faster convergence of to its limit as increases. hence are relatively small; as and increase, rapidly ap-
The average tail probability of the battery level, proaches the counterpart for the case of infinite battery capacity
, is evaluated by simulation and (or that for ).
HUANG: SPATIAL THROUGHPUT OF MOBILE AD HOC NETWORKS POWERED BY ENERGY HARVESTING 7609
Fig. 9. Transmission probability versus transmission power for both the cases of finite ( ) and infinite battery capacity. The DoF of the energy-arrival
process is , the energy-arrival rate , and the transmitter density .
Fig. 10. Maximum network throughput versus energy-arrival rate for the optimal transmission power and both the cases of finite ( ) and infinite battery
capacities. The DoF of the energy-arrival process is and the transmitter density .
Fig. 9 displays the curves of versus obtained by simula- smaller for larger because of less randomness in harvested
tion for , , , and . energy and hence smaller battery-overflow probability.
For comparison, the curve for the case of infinite battery ca-
pacity is also plotted. As observed from Fig. 9, reducing the VI. CONCLUSION
randomness of the energy arrival process by increasing leads The energy dynamics in a mobile ad hoc network have been
to smaller battery-overflow probability and hence higher . The characterized in terms of transmission probability. Assuming
effect of on diminishes as (and hence battery capacity) infinite battery capacity, it has been found that the transmis-
increases and converges to its counterpart for the case of infi- sion probability is equal to 1 when the energy-arrival rate ex-
nite battery capacity. ceeds transmission power or otherwise is equal to their ratio.
Last, we investigate the effect of the DoF of the energy arrival Moreover, for the case of finite battery capacity, bounds on
process on the network throughput. To this end, Fig. 10 shows the transmission probability have been obtained and exact ex-
the curves of versus obtained by simulation for pressions have been derived for the special cases of bounded
, , and . The curve for the or discrete energy arrivals. The results on transmission proba-
case of infinite battery capacity is also plotted for comparison. It bility have been applied to derive the maximum network spa-
can be observed from Fig. 10 that finite battery capacity causes tial throughput for a given energy-arrival rate and optimized
significant throughput loss especially for large . Such loss is transmission power. It has been shown that it is optimal and
7610 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 59, NO. 11, NOVEMBER 2013
feasible for all transmitters to transmit with probability one if By bounding the first term in (60) using Chernoff bound [33]
the transmitter density is below a threshold that depends on the and the second using (12)
energy-arrival rate; otherwise, each transmitter should transmit
with probability smaller than one.
There are several potential directions for extending this
work. Coexisting wireless networks may harvest electromag-
(61)
netic (EMR) energy from each others’ transmissions. Modeling
and designing coexisting networks with EMR energy har- where (61) results from setting . By choosing such
vesting give rise to many new research issues ranging from that the exponents of the two terms in (61) are equal, the desired
algorithm design to throughput analysis. The current work result follows.
focuses on ad hoc networks with random access and can be
extended to other types of networks such as cellular networks or APPENDIX C
other medium-access-control protocols such as carrier-sensing PROOF OF LEMMA 3
multiple access. Last, it is interesting to investigate the effects From the definition in (14),
of bursty data arrivals and more sophisticated power control
on the throughput of wireless networks powered by energy
harvesting.
APPENDIX A (62)
PROOF OF LEMMA 1
If , it follows from (10) that and hence the (63)
inequality in the lemma statement holds since . Next,
consider the case of and let denote the time where (62) and (63) are obtained using integration by parts and
instant closest to but smaller than . It follows that the random Lemma 2, respectively. The desired result follows from (63).
walk does not cross the threshold from below in the time
slots . Therefore, .
Then, can be upper bounded as APPENDIX D
PROOF OF LEMMA 4
The definitions of the random processes and in
(57) (9) and (10) are modified for the case of finite-battery capacity.
Specifically, is redefined as
It can be obtained from (9) that
(64)
(58) and is as given in (10) but with the battery-level evolution
following (1). Given finite battery capacity , the inequality
Since and from (10), combining (57) and can be proved using induction as follows. This
(58) proves the inequality in the lemma statement for the case inequality holds for since and .
of , completing the proof. Assume that . Consider the case of . It
APPENDIX B follows from the definition of in (10) that .
PROOF OF LEMMA 2 Therefore, since from (64).
Next, consider the case of . Based on the evolution
Using Lemma 1 and for , of in (1),
(65)
.
(67)
(60) (68)
HUANG: SPATIAL THROUGHPUT OF MOBILE AD HOC NETWORKS POWERED BY ENERGY HARVESTING 7611
where (67) applies , and (68) uses (64) and By inspecting (71), the event is equivalent
given that . Combining (65), (66), and (68) proves to the one . Then, the
that if . It follows that inequality in (72) can be rewritten as
for all . Furthermore, it can be shown
by expanding (64) that is no larger than that for
the case of infinite battery capacity. Using these results and fol-
lowing the same procedures as for proving Lemma 2 and 3, it
can be shown that (13) also holds for the case of finite battery
capacity and
APPENDIX E
PROOF OF LEMMA 5
(73)
Define the random process such that
(70) Using the definition of , the last term in (73) can be rewritten
as
By expanding (69)
(71)
(76)
ACKNOWLEDGMENT
The author thanks R. Zhang for helpful discussion that moti-
vated this research, and the anonymous reviewers whose com-
(72) ments have significantly improved the quality of this paper.
7612 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 59, NO. 11, NOVEMBER 2013