IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO.

12, DECEMBER 2004 3081

On Adaptive Transmission for Energy Efficiency in

Wireless Data Networks
Elif Uysal-Biyikoglu, Member, IEEE, and Abbas El Gamal, Fellow, IEEE

Abstract—This paper investigates the problem of energy-effi- [15], [20]), and approximated by practical adaptive coding/mod-
cient transmission of data packets in a wireless network by jointly ulation schemes ([12], [14]). These schemes, however, assume a
adapting to backlog and channel condition. Specifically, we con- continuous stream of data. If the average data generation rate is
sider minimum-energy scheduling problems over multiple-access
channels, broadcast channels, and channels with fading, when known, these schemes would keep transmission rate close to the
packets of all users need to be transmitted before a deadline . data arrival rate and be energy efficient. However, in many wire-
Earlier work has considered a similar setup and demonstrated less data applications, the rate at which data is generated and
significant transmission energy saving by adapting to backlog needs to be transmitted is unknown and varies with time (e.g.,
for channels that are time invariant and when transmission is wireless web sessions or a sensor network where data gets gen-
restricted to time-division. For concreteness, throughout the
paper, rates and powers corresponding to optimal coding over erated at random times at each node). Ignoring this variability
discrete-time additive white Gaussian noise (AWGN) channels are and adapting solely to the channel can be inefficient in terms of
assumed. The results, however, hold for more general channels transmit power and bandwidth.
and coding schemes where the total transmitted power is convex in
To understand how inefficiency may arise, consider the fol-
the transmission rates. The offline scheduling problems for all the
channels considered are shown to reduce to convex optimization lowing generic data communication situation: the transmitter
problems with linear constraints. An iterative algorithm, referred and receiver engage in a session which may be a video confer-
to as FlowRight, that finds optimal offline schedules is presented. ence or a web session, or may alternate between the two. Dif-
A heuristic online algorithm that we call look-ahead water-filling, ferent types of applications generate data at different rates, and
which jointly adapts to both channel fading state and backlog is
the data packets are collected in the transmitter’s buffer to be
described. By the use of a small buffer which introduces an almost
fixed delay, this algorithm achieves a considerable reduction in sent to the receiver. Let the rate at which packets arrive into the
energy relative to water filling solely on channel states. transmitter’s buffer at time be packets per second. These
Index Terms—Adaptive transmission, broadcast, energy-effi- packets are transmitted to the receiver at a rate packets per
cient transmission, iterative algorithm, multiple-access, power second. Now, assume we set , a constant that is large
control, scheduling, time-division, wireless networks. enough, say, for a high-rate streaming video session. When the
required rate drops, for example because the user switches to
a lower rate web session where , the transmitter will
I. INTRODUCTION
idle a significant fraction of time and unnecessarily transmit at

E NERGY efficiency in computation, signal processing, and

information transmission has been a topic of increasing in-
terest motivated by applications in mobile computing, wireless
a high rate for the rest of the time.
Schemes that adapt solely to the channel state can maximize
the throughput for a given energy constraint; but since they
data, ad hoc and sensor networks. Among the various problems cannot track the value of , they do not have control over
related to energy efficiency, optimizing transmission power is of delay. In order to guarantee finite average delay, they need to be
special importance, since the scalability of a wireless network set for the largest possible value of , which causes them to
is fundamentally limited by transmission power. be energy inefficient.
Power control for the purpose of maximizing rate of reli-
In order to achieve better performance in terms of energy and
able communication in the presence of interference and fading,
delay, one must also adapt to the variation in the data rate. To our
both characteristic of wireless networks, is well studied. Op-
knowledge, the earliest appearance of joint queue state/channel
timal rate and power adaptation schemes for the single-user and
state adaptive power control is in [6]. A more comprehensive
multiple-access fading channels have been developed (e.g., [11],
treatment appears in [2], where the objective is to obtain the op-
timal power–delay tradeoff curve and develop algorithms that
Manuscript received June 26, 2002; revised June 29, 2004. The work was minimize power while keeping the buffer size below a certain
supported in part under SNRC Multilayer Mobile Networking Project and by a
Stanford Graduate Fellowship. level. In [23], the minimum-delay power control problem (under
E. Uysal-Biyikoglu is with the Department of Electrical Engineering and a power constraint) is posed. It is shown that under the two ex-
Computer Science, Massachusetts Institute of Technology, Cambridge, MA tremes of fast fading and slow fading, the optimal power con-
02139 USA (e-mail: [email protected]).
A. El Gamal is with the Information Systems Laboratory, Department of trol policy goes to the two extremes of water-filling in time [20],
Electrical Engineering, Stanford University, Stanford, CA 94305 USA (e-mail: and channel inversion, respectively. Médard et al. [16] showed
[email protected]). that the capacity region of the time-slotted ALOHA system with
Communicated by L. Tassiulas, Associate Editor for Communication Net-
works. power-constrained users is the same as the capacity region of the
Digital Object Identifier 10.1109/TIT.2004.838355 multiple-access channel.
0018-9448/04$20.00 © 2004 IEEE
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3082 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

Delay-optimal scheduling of data packets from different

queues is studied in the context of satellite transmission sys-
tems in [17], [9]. In [17], it is shown, using Lyapunov theory,
that the throughput-optimal schedule for a system of several
queues is a maximum weight matching where the weight of
each queue is the backlog multiplied by the rate at which it
can transmit under the current channel conditions, for a certain
power constraint. A related study, [9], considers optimal energy
allocation and admission control for communication satellites
in earth orbit. The goal is to choose which data requests to serve
at a given time, in order to maximize expected total reward
while the energy that can be stored is finite and is replenished at 2
Fig. 1. The top plot shows a sequence of packet arrivals (“ ”) and channel
gains. The lower three plots show the instantaneous rates used by the online
regular intervals. An optimal policy is obtained using dynamic algorithms water-filling (WF) and look-ahead water-filling (LW), and the
programming. optimal offline schedule, respectively, as they run on this sequence of packet
In this paper, we build a simple framework that captures the arrivals. The average energy per packet values (E) are normalized for a noise
power of unity. Average delay (in time units) per packet (D) is also given.
discrete nature of packet arrivals, on which we are able to study
scheduling (assigning rates and transmission times to users)
in a multiuser setting with respect to energy and delay. The by three scheduling algorithms. The first is water-filling in
insights obtained are then used to propose good scheduling time [11], which adapts optimally to the channel to achieve
algorithms. The work here extends [19] and [8], where it was a time average rate of packets/time unit. The second is
shown that adaptation to the packet arrival process can result the look-ahead water-filling algorithm. The third is the optimal
in significant transmission energy savings. In [19] and [8] (see offline algorithm, which has perfect information of all future
Section II for a review), the channel was assumed to be time packet arrival times and channel states at time . Notice how
invariant such that the energy as a function of transmission the water-filling algorithm uses a high rate and idles for a major
duration does not depend on when a packet is transmitted. fraction of the time, while the second online algorithm, which
Here, we consider the more realistic wireless communication adapts its rate to its current backlog, spreads the rate more uni-
scenario where the channels (hence, the energy functions) formly across time, imitating the optimal offline schedule. As
are time varying due to interference and fading. Specifically, a result, the online algorithm reduces the average transmission
we consider minimum-energy scheduling problems over mul- energy per packet by a factor of , to a value which is almost
tiple-access channels, broadcast channels, and channels with within a factor of of the offline optimal.
fading. For concreteness, throughout the paper, we assume rates The rest of the paper is organized as follows. First we con-
and powers corresponding to optimal coding over discrete-time sider transmission schedules for the multiple-access channel.
additive white Gaussian noise (AWGN) channels. Our results, The MoveRight algorithm (proposed in [8]) can find the best
however, hold for more general channels and coding schemes time-division solution to the offline multiple-access problem.
where the total transmitted power is convex in the transmission However, to find the optimal solution, we need to consider gen-
rates. eral multiple-access coding schemes, where the users can simul-
The key results in the paper are: i) showing that for each of taneously transmit. In the following section, we define the mul-
these channels, offline scheduling reduces to a convex optimiza- tiple-access offline scheduling problem and show that it can be
tion problem with linear constraints, ii) devising an algorithm, cast as a convex optimization problem with linear constraints. In
FlowRight, that iteratively finds the optimal offline schedules Section III-B, we present FlowRight, which solves this problem.
(Section III-B), and iii) devising a heuristic online algorithm, In Section III-E, we present an online scheduling algorithm that
look-ahead water-filling, that by jointly adapting to both channel uses a look-ahead buffer and compare its performance in terms
fading state and backlog achieves energy efficiency close to the of energy and average delay per packet to time-division and to
optimal offline schedule (Section IV-B). The following example the optimal offline schedule. FlowRight is also shown to op-
illustrates the potential energy saving achieved by such joint timally solve the offline scheduling problem for the broadcast
adaptation. channel. In Section IV, we turn to channels with fading, and
show that FlowRight can optimally solve the offline scheduling
Example problem in a slow-fading channel with perfect CSI at the trans-
Data packets of size bits arrive at a trans- mitter and the receiver. We then present the look-ahead water-
mitter’s buffer at a rate known to be at most filling algorithm, which adapts jointly to the channel state and
packets time unit.1 The packets are transmitted over an data arrival rate, and observe that it is significantly more energy-
AWGN channel with noise power and slow ergodic fading, efficient than the well-known water-filling algorithm that adapts
where the transmitter and the receiver have perfect channel solely to the channel state. The look-ahead water-filling algo-
state information (CSI). Fig. 1 shows the channel gain and rithm is shown (through simulations) to also perform closely to
the arrival instants, and the rate (bits per transmission) used the offline optimal benchmark provided by FlowRight. The re-
sults are shown to generalize to multiple-access and broadcast
1The arrivals in this example are a realization of a bursty arrival process at channels. We also show that scheduling in a fast-fading channel
average rate  = 0:2. reduces to the single-user problem of [19].
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
UYSAL-BIYIKOGLU AND EL GAMAL: ON ADAPTIVE TRANSMISSION FOR ENERGY EFFICIENCY IN WIRELESS DATA NETWORKS 3083

Since this paper extends the work in [22] and [8], in the fol- transmission durations is quite intuitive. Consider the following
lowing section we first briefly review the key results in these example. Suppose all packets were available at time . In
papers. this case, it would be optimal to set for (as
a consequence of the monotonicity and convexity of the energy
II. PREVIOUS WORK function .) Now, suppose the th packet arrives at time
, where , while all previous packets arrive at
The work in [22] and [8] is based on the observation that with
. Setting would violate causality (the th
many channel coding schemes, the energy required to transmit a
packet would be starting transmission before it arrived.) Clearly,
packet over units of time, , is monotonically decreasing
the solution is to set , while
and strictly convex in . The minimum-energy packet sched-
uling problem for a single transmitter–receiver pair in a wireless
network is formulated in [19]. The setup is as follows. Suppose
that packets arrive at the transmitter’s buffer in the interval
at times . The node is required to As in this simple example, the optimal algorithm tries to equate
transmit all packets within the interval . 2 The question arrival times and make them as large as possible, within the con-
is, how should the packet transmissions be scheduled to mini- straints of causality. This solution was further explored in [22],
mize the total energy required to transmit the packets. If we let [18] and the formulation was combined with other constraints.
be the transmission time for packet , , the offline In [8], the minimum energy scheduling problem for a mul-
version of the problem can be stated as follows. tiple-user channel, e.g., uplink and downlink, involving several
transmitters and receivers where time-division is used is investi-
Problem 1: Single-Transmitter Single-Receiver Offline gated. The goal is to minimize the total energy for all users, and
Scheduling [19]: Given a vector of packet arrival times this results in the setup being identical to that of the previous
, where , , and , problem except that packets can have different energy functions.
and an energy function that is strictly monotonically de- Again, energy functions are convex and decreasing in the trans-
creasing and convex, find a schedule so as to minimize mission duration, and this is essentially all that is assumed about
the total transmission energy: subject to causality3 the channel, transmitters, and receivers. The offline time-divi-
and deadline constraints. sion scheduling problem is formulated as follows.
Note that this is a convex optimization problem with linear Problem 2: Multiple-User Offline Time-Division Scheduling
constraints. The following explicit solution was found in [19]. [8]: Given a vector of packet arrival times ,
Let ’s be the packet inter-arrival times. Define , and where , , and , and energy functions
that are strictly monotonically decreasing and convex,
find a schedule that minimizes the total transmission energy:
subject to causality and deadline constraints.
This is also a convex optimization problem but does not in
general admit a simple closed-form solution. By exploiting
Note that is the index at which the running average of the ’s the special features of the problem, an algorithm, MoveRight,
is maximized for the first time. We set to which finds the global optimal schedule efficiently, is devel-
oped. MoveRight iteratively moves the start times of packet
this average value . Next, we set transmissions one at a time, so that each move locally optimizes
the energy function. The algorithm was shown to solve other
scheduling problems, such as when packets have individual
(1) deadlines, and when the transmit buffer is finite. MoveRight
also leads to an online algorithm that uses a simple look-ahead
buffer. The transmitter buffers the packets for a specified
In general, is set to
length of time (the look-ahead window). At the end of the
look-ahead window, the packets in the buffer are scheduled
using a faster version of the MoveRight algorithm for trans-
(2)
mission from to . Meanwhile, the arrivals from to
are buffered, to be transmitted in the following time window.
where is the largest integer such that . This Hence, at the expense of incurring a delay of , packets are
is the optimal schedule. Note that the schedule contains bands scheduled optimally. The average energy per packet given by
of equal transmission times between breakpoints at positions . the look-ahead algorithm was shown through simulations to be
The fact that the optimal solution is in the form of bands of equal quite close to that of the offline optimal schedule, using only a
small look-ahead buffer.
2The imposition of a strict deadline T , by which all transmissions had to
Throughout the paper, the well-known terms ”time-division”
terminate, was intended to capture several realistic wireless scenarios (see [22]
for further details).
and ”multiple-access” will be used to make the following
3In our setting, causality corresponds to the obvious constraint that no distinction: In time-division, packets do not overlap in time,
packet’s transmission can start before its arrival time. whereas in multiple-access scheduling, users’ packets interfere
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3084 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

Fig. 2. Packet arrivals in [0; T ).

with each other but can, with appropriate multiuser coding and
decoding, still be resolved at the receiver.

III. SCHEDULING FOR THE MULTIPLE-ACCESS CHANNEL

Consider the discrete-time AWGN multiple-access channel
with transmitters and a single receiver. Data packets are gen-
erated at each transmitter’s buffer at arbitrary times , in the
interval . Fig. 2 shows an example sequence of packet ar- Fig. 3. Feasible region of (P ; P ) for a given (R ; R ). The multiple-access
rival times for two users, where packet arrival times of users region is bounded below by the solid boundary and the time-division region is
1 and 2 are marked by crosses and circles, respectively. These bounded below by the dashed boundary.
packets must be transmitted reliably to the receiver in the time
interval . The received signal at time is Note that the time-division boundary always touches
the boundary of the multiple-access region at the point
(3) , where .
We refer to the sequence of arrivals of the th user as stream
where is user ’s signal, and are inde- and merge the two streams into one sequence , as shown in
pendent and identically distributed (i.i.d.) zero-mean Gaussian Fig. 2, where stream 1 arrivals are marked by crosses and stream
noise with variance . For now, we assume the ’s to be 2 arrivals are marked by circles. We denote the inter-arrivals of
constant in time. Later we consider ’s that vary with to this new sequence by data epochs, or in short, epochs, and mark
model frequency-flat fading. them , . Without loss of generality, we assume
It is well known that for the AWGN multiple-access channel that a packet (from either one of the two users) is received at
the set of feasible average received powers , for a given time , so .
set of rates , is given by (see [20]) Before we present the multiple-access offline scheduling
problem, we make the following two key observations (Lemmas
1,2), the first of which can be obtained from [20, Lemma 3.3 ],
but is included here for completeness.
for all . To achieve points on the boundary Lemma 1: In the symmetric case , and bits
of the region, one needs to use optimal codes with block lengths can be transmitted in time units with minimum energy by time
approaching infinity. However, for long enough packets, one can sharing between the users (i.e., with time-division). In the asym-
come arbitrarily close to the boundary using codes with finite metric case , time-division is strictly suboptimal, and
block lengths and achieving reasonable level of reliability. To the unique optimal scheme is the corner point of the multiple-ac-
simplify expressions, here, and throughout the paper, we assume cess energy region where is at its minimum possible value.
codewords are long enough so that points on the boundary are In the AWGN case, this point corresponds to successive cancel-
basically achievable, but much shorter than the time window by lation where users are decoded in decreasing order of ’s.
which they must be transmitted. We restrict our discussion to Proof: Since the duration of the interval and the number
two users, set , and define . The results of bits to be transmitted in that interval by each user are fixed,
can be readily extended to more than two users. the average rates will be fixed at , and . Given
For , the feasible region of received powers is simply this average rate pair, we wish to minimize the total transmitted
energy . The solution can be readily seen from
the multiple-access achievable powers region in Fig. 3. When
, any power pair on the line
achieves the minimum. In particular, the point where the time-
division boundary touches the multiple-access boundary mini-
This is plotted in Fig. 3. Note that when the received power is
mizes the total transmitted power for the time-division scheme.
, the transmitted power is . Time-division, i.e., one user
When , the minimum is attained at one of the two corner
transmitting at rate for a fraction of the time and the other
points. For example, when , the optimal power pair is
transmitting at rate for a fraction of the time, yields
. In the AWGN channel case, this
the region with the dashed boundary specified by
is where , which
can be achieved by decoding user 2, subtracting its signal from
and
the received signal, and then decoding 1.
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
UYSAL-BIYIKOGLU AND EL GAMAL: ON ADAPTIVE TRANSMISSION FOR ENERGY EFFICIENCY IN WIRELESS DATA NETWORKS 3085

In the rest of this section, for ease of notation we assume that time-division is optimal and thus the problem can be optimally
, . solved using MoveRight. For convenience, define
Lemma 2: In an optimal multiple-access offline schedule, the
rate of a user need not change during an epoch. and note that is convex in and .
Proof: By definition, new data can only arrive at the start
of a data epoch. So, at the beginning of an epoch, the two users Theorem 1: In the symmetric case, there exists a schedule
together have a certain number of bits to be transmitted, and no that achieves minimum energy by time-division between the
new bits are added to this during the epoch. Now, let us focus on packets of users. In the asymmetric case, time-division is strictly
a generic data epoch in the optimal schedule. Assume without suboptimal.
loss of generality that the epoch starts at and ends at Proof: Consider the symmetric case first. We show that
. Assume that in this schedule the epoch is divided into any schedule can be converted into a time-division schedule
intervals, , where the with equal or lower energy. First, note that from Lemma 2, it
rates of both users are constant during an interval. Denote the suffices to consider the schedule as a sequence of rate pairs
first user’s rate in interval by , and the second’s by , one pair for each epoch. Also note that we can limit at-
. At optimal power settings (see Lemma 1) the total trans- tention to the case where the received powers are the
mitted energy for the data epoch is given by optimal corner point of the feasible region for rates (if
not, the energy can be reduced without changing the schedule).
Consider epoch . From Lemma 1, in the symmetric case, there
is a point on the time-division curve that achieves the average
rates with minimum total energy. We can move to this
By convexity of , it is easy to see that the total energy can be time-division point by letting the first user transmit alone in a
decreased by using the average rates fraction of the total interval, i.e., for a duration , using a
rate , and the second user transmit in the remaining with rate
, where . Proceeding like this with other
epochs, the schedule we were provided with has been converted
and to a time-division schedule of equal or lower energy. Now, con-
sider a packet from user 1 that is being transmitted across the
epochs , as chunks, with instantaneous rates

throughout the epoch.

We are now ready to state the minimum energy offline sched-
This packet has bits, so . The same
uling problem for the multiple-access channel. For simplicity,
amount of data can be transmitted by averaging user 1’s rate
consider equal-sized packets each with bits. The formula-
over the pieces, setting
tion and the results we obtain, however, can be readily general-
ized to packets of unequal size. Define the sequences and
as the number of bits that have arrived at the beginning of where
epoch for users 1 and 2, respectively. In the case of constant
sized packets, this means that for , By convexity of the power function , energy is reduced. Now
if there is a stream arrival at the beginning of data epoch , that the rate has been averaged out, one can always collect these
and otherwise. By Lemma 2, the optimal multiple-access of- pieces together to transmit the packet of user 1 as a whole.
fline scheduling problem reduces to finding a rate pair sequence All of the above can be repeated for user 2. The result is time-
that minimizes the total division between the packets of user 1 and user 2, where rates
energy. The problem is then as follows. (and powers) are set independently.
Now, consider the asymmetric case and suppose a time-divi-
Problem 3: Multiple-Access Channel Offline Scheduling: sion schedule is given. Take any interval that is divided between
two users. By Lemma 1 one can convert4 this epoch’s rates to
Minimize the optimal corner point on the multiple-access boundary, and
strictly decrease energy.
subject to Problem 3 is a convex optimization problem and considering
the conditions on , it is easy to see that it has a unique
solution. However, except for the symmetric case, the problem
has no closed-form solution. Standard convex optimization
methods could be employed to compute solutions. Such a
general approach, though, is unlikely to provide as much
insight as an approach that notices the special structure of the
Thus, similar to Problems 1 and 2, multiple-access channel
problem. The next section describes such an algorithm that
scheduling is a convex optimization problem with linear con-
straints. We now show that in the symmetric case 4If causality does not permit this, pick another time-division interval.
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3086 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

will be called ”FlowRight.” The FlowRight algorithm performs ;

simple iterations starting from a feasible initial schedule. Each }
iteration strictly improves the schedule (decreases the total if ( and )
energy), which ultimately converges to the unique optimal. {
;
A. FlowRight: An Algorithm for Optimal Offline Scheduling }
FlowRight is an iterative algorithm. In the beginning, the }
transmission time of each packet is set to precisely the data
epoch at the beginning of which the packet arrived. That is, in In each pass, the function is run times, i.e., once
the starting schedule, packets begin transmission when they for each consecutive pair of data epochs. It locally minimizes the
arrive, and end transmission when the next data epoch starts. energy of the two epochs in the following way. Consider the th
Let the rates obtained in this way be and , such that pass when is running on epochs , . The problem
, , . is that of choosing that minimize the total energy of
The FlowRight algorithm performs local optimizations on epochs and , while the total number of bits on each stream
pairs of epochs in the following way. Consider the first two is fixed at
data epochs. The total number of bits transmitted by users 1
and 2 in these two data epochs are and
, respectively. Keeping the number of bits and
fixed at and , we update to , where
is the allocation of rates to the first data epoch that
respectively.
minimizes the overall energy of the pair of data epochs. Obvi-
The following definition will be useful in the proofs later:
ously, when are decreased (i.e., at least one compo-
nent is decreased and neither is increased), will in-
crease, since the bits that leave the first epoch go to the second.5
Note that and , since, from the initial condi-
tion, information can only flow to the right (otherwise, causality
would be violated.) We therefore have to reset to new
Note that is the total energy of epochs
values which are larger (or equal to) their initial values.
and , and that it is convex in the rates and .
Moving to the second pair of data epochs, this time opti-
In order to prove that this algorithm finds the optimal of-
mally decrease to , and reset the values
fline schedule, we make the two observations described later in
of . Proceed in this way to obtain for
Lemmas 3 and 4. Consider two data epochs, 1 and 2, with du-
. This completes the first pass of the algorithm. It
rations and . The first stream needs to transmit a total of
is easy to see that in the first pass, information can only flow
bits in the two data epochs and the second stream needs to
to the right. Interestingly, we will later prove that information
transmit bits. Of the bits, are available in the first
always flows right in the algorithm, and consequently, after
data epoch, and of the bits, are available in the first
each iteration the rates are closer to the optimal solution than
data epoch. Let the rates chosen for the first epoch be ,
they were before that iteration.
hence, the total energy is .
After the first pass is complete we start from the beginning
The first observation to make is that after runs on
and update the rates two data epochs at a time similarly to the
this epoch pair, the partial derivatives of are either zero or
above. Terminate after pass , where
negative (which corresponds to a causality constraint being met
and with equality.) This is made precise in Lemma 3.
A pseudocode for the algorithm is given as follows. Lemma 3: The optimal rate pair is unique, and sat-
; isfies and , where and
for are the partial derivatives of with respect to the first and
{ second components, respectively. If , then a
; new packet is starting transmission at epoch 2 on stream 1, and
if , a new packet starts on stream 2.
;
} The proof of Lemma 3 is given in the Appendix . The second
; observation is about what happens if we take out (eject) some
while ( ) bits from the second epoch or put in (inject) some extra bits into
{ the first epoch and run again. This is the crucial step in
; proving the convergence of FlowRight, as such bit ejections/in-
for jections happen from one iteration to the next. We record the
{ observation in Lemma 4 (proven in the Appendix ).
Lemma 4: Suppose, after has run on the epochs
5We are allowing fractional numbers of bits to move between epochs. resulting in the situation in Lemma 3, some additional bits
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
UYSAL-BIYIKOGLU AND EL GAMAL: ON ADAPTIVE TRANSMISSION FOR ENERGY EFFICIENCY IN WIRELESS DATA NETWORKS 3087

are injected into epoch 1 on one or both streams, that is, set From Lemma 4, part 1,
, and , and .
Also, some bits are ejected from the second epoch, i.e., set
, and , , and . and
For the change to be nontrivial, we require that at least one of
is nonzero. Notice that the constraint
space of the problem has changed. In this case we have the Now, as the th pass progresses, update performs
following. a local optimization on , and this, by hy-
1) If and , then after the pothesis, results in a right push (i.e., a push from
injection/ejection of the new bits, when is run onto ), which changes to
again, there will be a right push, i.e., a nonnegative amount , where for
of information will move from epoch 1 to epoch 2 on both . Continuing to the present time, on the th pass there
streams. is a right push (again, by the induction hypothesis), from
to resulting in , where
2) If , the first stream was limited by
for . By part 2 of Lemma 4, there can only be a
causality before the injection/ejection, hence it is limited
right push (if any) from to on the th iteration.
by causality again (because no information has crossed
from epoch 1 into epoch 2). If , 2) Consider , i.e., the total number of
there will be a push on stream 1. Otherwise, due to bits of stream 1 on epoch 1 after the th iteration. Since
causality, reoptimization will not result in any move all pushes are to the right, is monotonically nonin-
on that stream (i.e., no push or pull). Similarly, if creasing. Also, it is obviously bounded from below by
, on the second stream there may only be zero. Therefore, , and therefore . Simi-
a push or no movement at all. larly, (the total number of bits in epochs 1 and
2 on stream 1) is monotonic nonincreasing and bounded
Now, let and be the pair of optimal rate se- below by zero. Hence, this sum tends to a limit;
quences. We shall sometimes use the shorthand to refer to
this pair. Such a unique solution exists because of the convexity
of the problem and the compactness of the search space. In the
following, it is proved that the algorithm FlowRight results in Therefore, , and . Similarly, since
and . In order to show this, we first argue that the converges and converges, so
algorithm stops at the pair of sequences and . We does . Proceeding like this, we will see that ,
then show that this is identical to . The following results are , for all . Hence, the sequences of rates converge.
proved similarly to Theorem 1 in [8]. 3) First, observe that and
for all . To see why this is true, suppose it is not. That
Theorem 2: The following statements hold.
is, let for some . Then, if we run
1) As the algorithm FlowRight runs, information always FlowRight on this sequence, there will be a right push.
flows right. This contradicts the fact that is a fixed point.
2) FlowRight stops, and returns two sequences and Hence, we have and
. for all , and using that we will show that sat-
3) and . isfies the Karush–Kuhn–Tucker (KKT) conditions [4].
Recall that our problem is a convex problem with linear
Proof:
inequality constraints, so the KKT conditions are suffi-
1) The claim is that throughout the running of FlowRight, all cient for optimality in this case. The first of those con-
pushes are to the right. We now prove this by induction. ditions is feasibility, of course, but we already know that
In the first pass, the claim is trivially true, since all left is a feasible solution (FlowRight always re-
pushes are impossible due to causality. Now, suppose that spects feasibility). Then we need only check if for our so-
we are on the th pass of the algorithm, and so far update lution there is a set of Lagrange multipliers with the prop-
has operated on all epoch pairs up to and including the erties specified by the KKT conditions. Differentiating the
pair , and all pushes so far have been to the right. Lagrangian for the problem provides us with equa-
We will show that the next push will be to the right. On the tions
th run, performed a local optimization on
epochs . Let us call the two pairs of rates resulting
from this optimization , and .
The number of bits transmitted in the th epoch on streams and
1 and 2 are and . Similarly, de-
fine and to be the bits transmitted the next
epoch, and define
where , are the Lagrange multipliers.
and Now we need to inquire about the values of these La-
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3088 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

grange multipliers. By subtracting the th equation from

the th, we obtain

Now, substitute for

, and by what we showed above, we
obtain . Further, using Lemma 4,
if the th constraint is active (i.e., all the bits that
are present by that time have been transmitted by the end
of epoch on stream 1), and otherwise.
Proceeding this way, we obtain that if the th
constraint is active (i.e., again, all the bits that are present
Fig. 4. Comparison of offline time-division and multiple-access schedules
by that time have been transmitted by the end of epoch as obtained by the MoveRight and FlowRight algorithms for a two-user
on stream 1), and otherwise, for , and multiple-access channel. The users’ packets arrive according to two independent
if the th constraint is active (i.e., causality Poisson processes with identical rates, and the combined arrival process is at
rate . The energy values correspond to 10 -bit packets. The signaling rate
is met at the end of epoch on stream 2), and is 10 transmissions/s, and the nominal rate is 6 bits/transmission—that is
otherwise, for . But with these, we have obtained when a packet takes 1 time unit, i.e., seconds, to transmit.
the KKT conditions completely satisfied. This proves
that is the globally optimal solution of our zero-mean Gaussian noise with variance . The capacity re-
problem. gion of the channel (see [7]) assuming , is the set of rate
pairs such that
B. Time-Division Versus Optimal Multiple-Access
As we have pointed out before, one can calculate the best
time-division offline schedule by running the MoveRight algo-
rithm on the joint sequence of packet arrivals of all users. It is
interesting to find out how time-division compares to the op-
timal solution. To that end, in this subsection we compare the for some . Now, we express the minimum average
average energies consumed by MoveRight and FlowRight on power for a given rate pair, where the above inequalities are
the same arrival sequence. replaced by equalities, as follows. Rewrite the first equality as
The experiment setup is as follows. The two users’ packets ar- . Hence, .
rive according to two independent Poisson processes with iden- Substituting into the second inequality and rearranging we ob-
tical rates, with a combined rate of arrivals per unit time (unit tain
time is a symbol time). For each value of , 1000 arrivals are
generated, and is set to . Then, both FlowRight
and MoveRight are run on this sequence, and the average en-
ergy per packet and average delay per packet for both users are Consider epochs defined in the same way as in Section III,
calculated. Here, and , these values were chosen as the packet inter-arrival times of the merged sequence. Again,
because the resulting average energy and delay values are very making the observation that in an optimal broadcast schedule
similar for the two users. In Fig. 4, the average energy and delay rates do not need to change during an epoch, the offline sched-
values in both time-division and optimal solutions are plotted. uling problem is as follows.
The results suggest that the energy per packet can be reduced
Problem 4: Broadcast Channel Offline Scheduling:
significantly by using multiple-access codes, especially at high
rates. But note that when is small, and consequently transmis-
sion rates are low, time-division is almost as good. Minimize

C. Extension to the Broadcast Channel

subject to
Consider an AWGN broadcast channel with one sender and
two receivers. The sender has two streams of packet arrivals,
with one stream destined to each receiver. As before (see Fig. 2),
we merge the packets into a single sequence.
The received signal to the th receiver at time is given by

(4) Note that this is the same as Problem 3 in Section III, except
that the objective function is different. But the objective function
where is the transmitted signal with average power con- is still convex, monotonically increasing, and differentiable in
straint , is the channel gain, and the ’s are i.i.d. both and . Hence we get the following.
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
UYSAL-BIYIKOGLU AND EL GAMAL: ON ADAPTIVE TRANSMISSION FOR ENERGY EFFICIENCY IN WIRELESS DATA NETWORKS 3089

Theorem 3: FlowRight finds the unique solution to

Problem 4.

D. Online Scheduling
The optimal offline algorithm provides a lower bound on en-
ergy for all possible online algorithms, as it finds the minimum
possible energy per packet under complete knowledge of the fu-
ture. Fortunately, the optimal offline algorithm naturally lends
itself to online use by means of a simple look-ahead buffer [8].
We buffer all packets which arrive in the interval and op-
timally schedule them for departure in the interval . Note
that FlowRight does not need to perform any iterations—the rate
of each user is set to the number of bits it has in the buffer di-
vided by and transmission powers are chosen optimally ac-
cording to the feasible power region for the given rates. Con-
tinuing with the schedule, the packets that arrive in are
Fig. 5. Comparison of the online algorithms look-ahead time-division and
buffered to be transmitted in , and so on. look-ahead multiple-access for a two-user multiple-access channel. The users’
In Fig. 5, we compare the online algorithm obtained in this packets arrive according to two independent Poisson processes with identical
way, with the online algorithm obtained by using the look-ahead rates. The window size is 25 time units. The energy values correspond to 10 -
bit packets. The signaling rate is 10 transmissions/s, and the nominal rate is
buffer with MoveRight. The experiment setup is similar to the 6 bits/transmission achieved when a packet takes 1 time unit, i.e., seconds,
one in Section III-C. The only difference is that now we have to transmit.
a look-ahead buffer. In the experiment whose results are shown
in the figure, the look-ahead window size was held at 25
Proof: Noting that the channel state and the number of
time units.6 Hence, as our online algorithm “adapts” to the ar-
available bits are constant during an epoch by definition, the
rival rate, the average delay remains around 25 time units. The
proof follows very similarly to the proof of Lemma 2. Suppose
resulting average energy per packet is reasonably close to the
the rate is in the first time units of an epoch of length ,
optimal, which is also plotted in Fig. 5, and which, of course,
and during the remaining . The transmit energy in this
has much lower delay. Therefore, we see that by incurring some
epoch is then , where is the fading
fixed delay, it is possible to perform very energy efficiently.
state during the epoch. The same number of bits can also be
transmitted using the uniform rate for the
IV. SCHEDULING OVER SLOW-FADING CHANNELS whole time . This new rate results in a total energy
Consider the AWGN channel as specified by (3). We make the , which, by convexity of , is strictly lower than
block-fading assumption where the power gain changes previous, unless .
every channel uses (a“coherence window”). Further assume
From Lemma 5, ( is the number of epochs) is suf-
that fading is slow with respect to codeword lengths. Initially,
ficient to characterize the optimal schedule.
consider the single-user case, i.e., . We assume that both
the transmitter and the receiver have perfect channel state infor- Problem 5: Offline Scheduling for the Slow-Fading
mation at the beginning of each coherence window. Channel:
As before, consider packets coming at arbitrary instants in
, all of which need to be transmitted within this same time
Minimize
period. The optimal offline schedule is the one that minimizes
the total packet transmission energy given perfect knowledge of
the packet arrival instants and channel state values for the entire subject to
duration , at time .
Define an “epoch” to be a time interval that begins with either
a packet arrival or a change in the channel state, and continues
until the next arrival or state change. The first epoch starts at
, and continues until or , whichever is smaller, at
which point the second epoch starts, and so forth. Let denote
the number of bits that have arrived at the beginning of epoch , This convex optimization problem can be solved by the
so if the th epoch starts with a packet arrival, and FlowRight algorithm. Initially, the rates are set to ,
otherwise. Let the duration of epoch be . . The first two epochs are then considered.
Lemma 5: In an optimal schedule, rate is constant during an The total number of bits transmitted in these two data epochs
epoch. is . Keeping the total number of bits fixed, is
updated to , the value that minimizes the total energy of the
6In [8], a window of 20–30 time units was shown to be a good choice for L, first pair of data epochs. Note that , since from their
in the sense that most of the achievable decrease in energy is obtained. initial condition information can only be moved to the right
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3090 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

(otherwise, causality would be violated.) Therefore, is reset where is the average transmit power, and is the noise
to a new value that is larger than (or equal to) its initial value. variance. Furthermore, capacity is achieved by a “water-filling”
Moving to the second pair of epochs, this time is optimally power allocation to states. This means that transmit power is
decreased to , and the value of is reset. Proceeding in this set to if is larger than the cutoff value , and if
way we obtain for . This completes the first not, there is no transmission until changes. Hence the instan-
pass of the algorithm. The algorithm then repeats the same taneous transmit power is
procedure and terminates after passes, where
if
(5)
where otherwise.
for small enough . To use water-filling adaptation in our setting, we set the av-
erage rate equal to (packets/time unit) (bits/packet)
Theorem 4: The following statements hold.
(time units/symbol) to ensure stability. We then calculate
1) As the algorithm FlowRight runs on , information (hence, determine the average power) such that the capacity
always flows to the right. is equal to this target average rate. Before each packet transmis-
2) FlowRight stops, and returns a sequence . sion, the instantaneous power and rate are set according to the
3) . channel gain , which is assumed constant during packet trans-
mission.
The only difference between this theorem and Theorem 2 is 2) Look-Ahead Water-Filling Algorithm: The water-filling
that the energy function, due to scaling with channel gain, does scheduling algorithm presented above optimally adapts to the
not have the same form for each epoch pair channel state, and is energy optimal if the average rate of packet
arrivals is close to . But this algorithm can be wasteful
when the instantaneous packet arrival rate is much lower than
Note, however, that this does not affect any of the steps of the . Now we describe an online algorithm, which we refer to
proof of Theorem 2, and therefore the proof of this theorem as look-ahead water-filling algorithm, that adapts jointly to the
follows from the proof of Theorem 2. channel and backlog.
The algorithm is as follows: suppose just before time , a
A. Online Scheduling packet transmission ended. Let the backlog at time be .
We assume that the packet input process into the transmitter If , then we begin transmitting the packet at the head of
buffer is stationary and ergodic. The time average arrival rate the queue at time (otherwise, wait until there is a packet in the
queue). We set the target transmission rate to

packets/time unit
is bounded such that with probability . We are in- for some constant . Given , we determine the instanta-
terested in schedules that are stable, i.e., scheduling algorithms neous transmission rate according to water-filling. That is, the
that ensure that the number of packets in the buffer is finite with optimal cutoff value is computed as in Section IV-B1, which
probability . corresponds to an average power for which the capacity is
Future arrivals, channel states, or are not known. The (Using the concavity of the logarithm, the value of is cal-
channel has slow ergodic fading with known statistics where culated iteratively.) The current power and rate are then deter-
the power gains of different coherence windows are i.i.d. The mined from (5). We transmit the packet at the head of the queue
transmitter knows the present value of the channel gain just with this rate. The following pseudocode summarizes the algo-
before transmitting a packet. The bound on packet arrival rate rithm.
is also known. We first describe an online scheduling al-
gorithm based on water-filling in time that is known to achieve calculate rate estimate
the capacity of the channel. Next, we describe look-ahead ;
water-filling, an algorithm that simultaneously adapts to both find for which
the channel and the data arrival rate. if ( )
1) Water-Filling in Time: It is well known (see [11]) that the {
capacity of the AWGN channel with ergodic fading and with the ;
channel gain known at both the transmitter and receiver is given }
by else
{
bits/transmission
}
;
where is the probability density function of the channel
if
gain , and is the solution to the equation
{
;
;
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
UYSAL-BIYIKOGLU AND EL GAMAL: ON ADAPTIVE TRANSMISSION FOR ENERGY EFFICIENCY IN WIRELESS DATA NETWORKS 3091

}
else
{
;
}
repeat

Note that, in the look-ahead water-filling algorithm, the target

packet transmission rate never exceeds , yet the queue is
stable. This is shown in the following lemma.
Lemma 6: The look-ahead water-filling algorithm is stable,
i.e., given any , with probability one there exists ,
, such that .
Proof of Lemma 6: Suppose the claim is false. Consider
the Markov process7 where the state at time is , and state
transitions occur in the underlying Markov chain whenever Fig. 6. Energy per packet as arrival rate changes for = 1 in Rayleigh fading;
L = 25.
there is a packet arrival or departure. Let . Defining
and as the counting processes of the number of
arrivals and departures, respectively, from to

(6)

Since is finite and increases at most linearly with ,

the expectation is defined for every . Then

(7)

Our hypothesis that the queue is not stable implies that the un-
derlying Markov chain is transient, hence, eventually any fi-
nite set of states has probability zero. Consequently, the event
will eventually have zero probability. But,
referring to the algorithm, in the event , the
expected transmission rate (where the expectation is over the
fading process) is , and herefore in the limit (referring to the al-
gorithm) departures will happen at rate packets/time unit.
Dividing by and taking the limit in (7), we obtain Fig. 7. Average energy per packet as arrival rate changes for = 2 in Rayleigh
fading; L = 25.

(8) other two algorithms, thus quickly finishes its backlog and idles
a significant amount of the time. Look-ahead water-filling,
This implies , which is a contradiction. on the other hand, spreads its rate more uniformly over time,
almost as uniformly as OPT which has the lowest rate trans-
To compare the look-ahead water-filling algorithm to water-
mission.
filling, we perform the following experiment: Let 1-kbit packets
Figs. 6 and 7 explore the energy and delay performance
arrive at the buffer at a rate 1 arrivals/ time unit. A time
of these algorithms. Note that the water-filling schedule has
unit is 1/6 ms, which corresponds to the transmission duration
constant energy for all arrival rates, since the rate it assigns to
of a packet if it is transmitted at 6 bits/symbol (symbol
packets is independent of . This energy is much higher than
rate is constant at symbols/s). The packet arrival process
the average energy values achieved by look-ahead water-filling
is a Markov-modulated Poisson process for which
when is small; both for bursty and nonbursty arrival processes.
with probability , and otherwise. The
Of course, the energy efficiency is achieved at the expense of
parameter is chosen such that the process is ergodic with
an increase in delay. The delay of look-ahead water-filling
expected rate . Note that when , the arrival process is
is essentially lower-bounded by , as it allows this time to
bursty, and for it reduces to a Poisson process at rate .
monitor the arrival process. However, as can be observed from
Fig. 1 shows an example run of bursty packet arrivals at
Fig. 6, the variation of its delay is much smaller than that of
, scheduled by the three algorithms WF (water-filling),
water-filling. In the figure, the delay of water-filling varies by
LW(look-ahead water-filling), and OPT (optimal offline). No-
about 7000% as is varied from to , while the delay
tice that water-filling transmits with much higher rate than the
of look-ahead water-filling varies only about 60%. The fact that
7Sometimes referred to as a semi-Markov process [10]. the delay jitter is so much smaller makes the backlog-adaptive
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3092 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

algorithm attractive for data applications, especially streaming alizations. In this setting, it is natural to assume that the trans-
media. mitter does not have CSI; by the time feedback from the receiver
about channel state reaches the receiver, the channel has already
B. Extension to Multiple-Access and Broadcast Channels With changed. It is well known (see, e.g.,[5]) that the ergodic capacity
Fading of this (single-user) Gaussian fading channel is given by
In this subsection, we extend the slow-fading single-user of-
fline and online scheduling results to multiple-access and broad- (9)
cast channels. To formulate the offline scheduling problem we
first merge all users’ packet arrival sequences and the times at where denotes expectation over the channel state , and is
which channel states change to obtain epochs and note as the average transmit power.
before that in an optimal schedule rates do not need to change Notice that is a monotonically increasing and concave
during an epoch. function in . Hence, it is invertible, with inverse
monotonically increasing and convex in . Reliable commu-
Problem 6: Offline Scheduling for the Slow Fading Multiple- nication at rate is possible if received power is in the set:
Access Channel: .
Therefore, the power needed to communicate at rate is
Min. . This problem can be written in the style of Problem
5, by defining epochs as packet inter-arrival times, of durations
, .
s.t.
Problem 7: Offline Scheduling for the Fast-Fading Channel:

Minimize

where
and subject to

and where if a packet for user arrives at the beginning

of epoch , and otherwise.
This is a convex optimization problem with linear constraints
and can be solved by FlowRight. Note that the definition of Problem 7 does not involve channel
Recall that in the single-user case, the optimal adaptation to states. This is because the transmitter does not track the channel
the channel state is given by the water-filling solution. In the state, which is assumed to vary over a codeword resulting in
multiple-user case, analogous results exist. When the fading channel capacity to be constant (see (9)). This is unlike the
processes of users are i.i.d., and the goal is to maximize the sum slow-fading case, where we assumed the transmitter can obtain
rate with respect to a total power constraint, the important result channel state information and code accordingly, and thus the ca-
of Knopp and Humblet [15] says that the optimal power control pacity changes in time. Clearly, this problem can be solved using
scheme allows only the user with the best channel to transmit FlowRight. Note, however, that it is much simpler than Problem
at any given time. The rate of that user is then determined by 5, since we do not need to consider channel state change in-
water-filling across the channel states. Tse and Hanly [20] ex- stants when defining epochs. In fact, this problem is identical to
hibit the optimal power control when users are not necessarily Problem 1 and thus its closed-form solution is given by (2).
symmetric, and the goal is to maximize a weighted sum of the At this point, it is quite clear that online scheduling in the
rates. They propose a “greedy algorithm” which also solves the fast-fading case can also be done by the look-ahead algorithm:
dual problem, i.e., achieves a given vector of average rates with set the rate at time to bits/channel use. This is
minimum power. The greedy algorithm is an optimal online al- more straightforward than the slow-fading case as no adaptation
gorithm, as long as the transmitter knows the fading state. to the channel state needs to be performed.
A “look-ahead greedy” online schedule that uses a
look-ahead buffer to adapt to both backlog and channel V. CONCLUSION
state (similar to look-ahead water-filling) can be obtained as Progress in wireless networking has greatly increased the
follows. Each user’s required rate is estimated from the current need for transmitting information with minimum energy
backlogs. The power allocation is then determined using the under reasonable delay constraints. In the study [19], the min-
greedy algorithm in [20]. imum-energy packet transmission offline scheduling problem
Finally, note that the broadcast scheduling problem in the under deadline constraint was formulated and solved for a
slow-fading channel can be stated and solved similarly. single transmitter–receiver pair. The multiple-user case, when
transmission is restricted to time-division was studied in [8]. In
C. Fast-Fading Channels this paper, we extended the work in [19] and [8] to transmission
Now, suppose that fading is fast with respect to our codeword scenarios where the channel is time-varying due to interference
lengths, so that a codeword will experience many channel re- and fading. We showed that offline scheduling for classes
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
UYSAL-BIYIKOGLU AND EL GAMAL: ON ADAPTIVE TRANSMISSION FOR ENERGY EFFICIENCY IN WIRELESS DATA NETWORKS 3093

of multiple-access and broadcast channels with and without Define for . Due to strict con-
fading can be reduced to convex optimization problems with vexity (hence, the monotonicity of the derivative) it can be easily
linear constraints and devised an algorithm, FlowRight, that shown that
finds the optimal offline schedule. Using FlowRight, we were
able to find the minimum energy per packet achievable by any
algorithm in the uplink scenario. Through simulations, it was
observed that the significance of multiple-access coding over and that the inequality is strict unless , ,
using time-division increases as the system gets more loaded. , and . But . Hence, energy is
We devised a heuristic online algorithm, look-ahead water- uniquely minimized by a point that satisfies
filling, which adapts to both the channel variation and backlog. and . This solution results from
It was demonstrated through simulations that significant energy pushing a nonzero amount of information right, from data epoch
saving can be achieved by such joint adaptation. 1 to data epoch 2. In case b)
In this paper, we addressed the point-to-point, multiple-ac-
cess and broadcast settings. An interesting direction for future or
work would be to investigate energy-efficient scheduling for
multihop networks. This is of particular interest due to the So when the injection and subtraction is done, these derivatives
increasing practical importance of ad hoc sensor and mobile can remain negative or become positive. In the case that they
networks, where energy conservation is a key design criterion. become positive, there will be a right push. If either of these, say
Finding optimal energy-efficient scheduling algorithms for , remains negative, then a right push (i.e., decreasing ) on
multihop settings, however, is nontrivial. The optimal trans- that stream can only increase the total energy. That stream was
mission rates, which we used in deriving the optimal offline shown in part 1 to be limited by causality, and it still is, because
schedules, are not known even for the simplest such setting injection brought only bits from the left. So we cannot pull any
with a single sender–receiver pair and a single-relay node. bits from epoch 2 to epoch 1 on this stream (i.e., increase ).
Another direction for future work is formulating and solving So, there will be no move on this stream.
the question of optimal online scheduling. Recently, [1] has
considered the single transmitter–receiver case where the trans- ACKNOWLEDGMENT
mitter has a finite buffer, and solved the problem of dynam- The authors would like to thank Balaji Prabhakar, Sina Za-
ically assigning rates/powers to packets in order to minimize hedi, and Mayank Sharma for their useful comments, and the
the long-term average transmission energy subject to an upper anonymous reviewers for their careful reading.
bound on the buffer overflow probability. It would be interesting
to pursue the generalization of such a dynamic control formula-
REFERENCES
tion to multiuser settings.
[1] B. Ata, “Dynamic power control in a wireless static channel subject to
a quality of service constraint,” Oper. Res., to be published.
APPENDIX [2] R. Berry, “Power and delay trade-offs in fading channels,” Ph.D. disser-
PROOFS OF LEMMAS 3 AND 4 tation, MIT, Cambridge, MA, 2000.
[3] R. Berry and R. Gallager, “Buffer control for communication over fading
A. Proof of Lemma 3 channels,” in Proc. Int. Symp. Information Theory, Sorrento, Italy, June
2000, p. 409.
For simplicity, we shall drop the superscript and refer to [4] D. P. Bertsekas, Nonlinear Programming. Belmont, MA: Athena Sci-
as . This function is strictly convex entific, 1995.
[5] E. Biglieri, J. Proakis, and S. Shamai (Shitz), “Fading channels: In-
in both variables, and is the result of minimizing formation-theoretic and communications aspects,” IEEE Trans. Inform.
it over a bounded region. Hence, the solution is unique. The Theory, vol. 44, pp. 2619–2692, Oct. 1998.
solution, , is either at the boundaries of the region [6] B. Collins and R. Cruz, “Transmission policies for time varying channels
with average delay constraints,” in Proc. 1999 Allerton Conf. Commu-
defined by , , or inside. If it nication, Control and Computing, Monticello, IL, 1999.
is inside, it must satisfy and . [7] T. M. Cover and J. A. Thomas, Elements of Information Theory. New
Due to convexity, partial derivatives of are monotonic York: Wiley, 1991, Wiley Series in Telecommunications.
[8] A. El Gamal, C. Nair, B. Prabhakar, E. Uysal-Biyikoglu, and S. Zahedi,
increasing. At the point , and are “Energy-efficient scheduling of packet transmissions over wireless net-
both negative (this can be seen by substituting the values). If works,” in Proc. IEEE INFOCOM, vol. 3, New York, June 2002, pp.
is not achieved in the region, then due to monotonicity, 1773–1782.
[9] A. Fu, E. Modiano, and J. Tsitsiklis, “Optimal energy allocation and
for all permissible values of . In this admission control for communications satellites,” in Proc. IEEE IN-
case, increasing the rate further than the boundary would FOCOM, vol. 2, New York, June 2002, pp. 648–650.
decrease total energy, but this cannot be done due to causality [10] R. G. Gallager, Discrete Stochastic Processes. Boston, MA: Kluwer
Academic, 1995.
constraints, so . Similarly, if is not [11] A. Goldsmith, “The capacity of downlink fading channels with variable
achieved inside the region, then . rate and power,” IEEE Trans. Veh. Technol., vol. 46, pp. 569–580, Aug.
1997.
[12] A. Goldsmith and S.-G. Chua, “Variable-rate variable-power MQAM for
B. Proof of Lemma 4 fading channels,” IEEE Trans. Commun., vol. 45, pp. 1218–1230, Oct.
1997.
First consider case a), i.e., [13] S. Hanly and D. N. C. Tse, “Multi-access fading channels: Part II: Delay-
limited capacities,” IEEE Trans. Inform. Theory, vol. 44, pp. 2816–2831,
and Nov. 1998.
Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.
3094 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 50, NO. 12, DECEMBER 2004

[14] K. J. Hole, H. Holm, and G. E. Øien, “Adaptive multidimensional coded [19] B. Prabhakar, E. Uysal-Biyikoglu, and A. El Gamal, “Energy-efficient
modulation over flat fading channels,” IEEE J. Select. Areas Commun., transmission over a wireless link via lazy packet scheduling,” in Proc.
vol. 18, pp. 1153–1158, July 2000. IEEE INFOCOM, vol. 1, Anchorage, AK, Apr. 2002, pp. 386–394.
[15] R. Knopp and P. A. Humblet, “Information capacity and power control [20] D. N. C. Tse and S. Hanly, “Multi-access fading channels: Part I: Poly-
in single-cell multiuser communications,” in Proc. Int. Conf. Communi- matroid structure, optimal resource allocation and throughput capaci-
cations, vol. 1, Seattle, WA, June 1995, pp. 331–335. ties,” IEEE Trans. Inform. Theory, vol. 44, pp. 2796–2815, Nov. 1998.
[16] M. Médard, S. P. Meyn, J. Huang, and A. J. Goldsmith, “Capacity of [21] E. Uysal-Biyikoglu and A. El Gamal, “Energy-efficient packet transmis-
time-slotted ALOHA packetized multiple-access systems,” in Proc. sion over a multiaccess channel,” in Proc. Int. Symp. Information Theory,
IEEE Int. Symp. Information Theory, Sorrento, Italy, June 2001, p. 407. Lausanne, Switzerland, June/July 2002, p. 153.
[17] M. Neely, E. Modiano, and C. Rohrs, “Power and server allocation in [22] E. Uysal-Biyikoglu, B. Prabhakar, and A. El Gamal, “Energy-effi-
a multi-beam satellite with time varying channels,” in Proc. IEEE IN- cient packet transmission over a wireless link,” IEEE/ACM Trans.
FOCOM, vol. 3, New York, June 2002, pp. 1451–1460. Networking, to be published.
[18] P. Nuggehalli, V. Srinivashan, and R. R. Rao, “Delay constrained energy [23] W. S. Yoon and T. E. Klein, “Delay-optimal power control for wireless
efficient transmission strategies for wireless devices,” in Proc. IEEE IN- data users with average power constraints,” in Proc. 2002 Int. Symp.
FOCOM, vol. 3, New York, June 2002, pp. 1765–1772. Information Theory, Lausanne, Switzerland, June/July 2002, p. 53.

Authorized licensed use limited to: Muranga University College. Downloaded on February 15,2025 at 13:27:25 UTC from IEEE Xplore. Restrictions apply.