Jointly Optimal Power Control and Routing for a

Single Cell, Dense, Ad Hoc Wireless Network*

Venkatesh Ramaiyan Anurag Kumar Eitan Altman
Dept. of ECE Dept. of ECE INRIA
Indian Institute of Science Indian Institute of Science Sophia-Antipolis
Bangalore, India Bangalore, India France
[email protected] anuraggece.iisc.ernet.in eitan.altmangsophia.inria.fr

Abstract-We consider a dense, ad hoc wireless network con- scheme is used in order to schedule transmissions; for example,
fined to a small region, such that direct communication is possible the CSMA/CA based distributed coordination function (DCF)
between any pair of nodes. The physical communication model is of the IEEE 802.11 standard for wireless local area networks
that a receiver decodes the signal from a single transmitter, while
treating all other signals as interference. Data packets are sent be- (WLANs). We assume that all nodes can decode all the con-
tween source-destination pairs by multihop relaying. We assume tention control transmissions (i.e., there are no hidden nodes),
that nodes self-organise into a multihop network such that all hops and only one successful transmission takes place at any time in
are of length d meters, where d is a design parameter. There is a the network. In this sense we say that we are dealing with a sin-
contention based multiaccess scheme, and it is assumed that every gle cell scenario. Thus our work in this paper can be viewed as
node always has data to send, either originated from it or a transit
packet (saturation assumption). In this scenario, we seek to maxi- an extension of the performance analysis presented in [2] and
mize a measure of the transport capacity of the network (measured extended in [1]. We further assume that, during the exchange
in bit-meters per second) over power controls (in a fading environ- of contention control packets, pairs of communicating nodes
ment) and over the hop distance d, subject to an average power are able to estimate the channel fading between themselves and
constraint. are thus able to perform power control per transmission.
We first argue that for a dense collection of nodes confined to a
small region, single cell operation is efficient for single user decod- There is a natural tradeoff between using high power and
ing transceivers. Then, operating the dense ad hoc network (de- long hop lengths (single hop direct transmission between
scribed above) as a single cell, we study the optimal hop length and the source-destination pair), versus using low power and
power control that maximizes the transport capacity for a given shorter hop lengths (multihop communication using interme-
network power constraint. More specifically, for a fading channel
and for a fixed transmission time strategy (akin to the IEEE 802.11 diate nodes), with the latter necessitating more packets to be
TXOP), we find that there exists an intrinsic aggregate bit rate transported in the network. The objective ofthe present paper is
(Oopt bits per second, depending on the contention mechanism and to study optimal routing, in terms ofthe hop length, and optimal
the channel fading characteristics) carried by the network, when power control for a fading channel, when a single cell network
operating at the optimal hop length and power control. The op- (such as that studied in [1]) is used in a multihop mode. Our
timal transport capacity is of the form dopt(Pt) x Oopt with dopt
_ 1 _ objective is to maximise a certain measure of network transport
scaling as Pt 7, where Pt is the available time average transmit capacity (measured in bit-meters per second; see Section IV),
power and q1 is the path loss exponent. Under certain conditions
on the fading distribution, we then provide a simple characterisa- subject to a network power constraint. A network power con-
tion of the optimal operating point. straint determines, to a first order, the lifetime of the network.
Index Terms-Multihop Relaying, Optimal Power Control, Self- Situations and considerations such as those that we study
Organisation, Fixed Transmission Time could arise in a dense ad hoc sensor network. Ad hoc sen-
sor networks are now being studied as possible replacements
for wired measurement networks in large factories. For exam-
I. INTRODUCTION
ple, a distillation column in a chemical plant could be equipped
We consider a scenario in which there is a large number of with pressure and temperature sensors and valve actuators. The
stationary nodes (e.g., hundreds of nodes) confined to a small sensors monitor the system and communicate the pressure and
area (e.g., spatial diameter 30m), and organised in to a multi- temperature values to a central controller which in turn actu-
hop ad hoc wireless network. Source-destination pairs are cho- ates the valves to operate the column at the desired operating
sen randomly and we assume that the traffic in the network is point. Direct communication between the sensors and actuators
homogeneous. We assume that data packets are sent between is also a possibility. Such installations could involve hundreds
source-destination pairs by multihop relaying with single user of devices, organised into a single cell ad hoc wireless network
decoding and forwarding of packets, i.e., we assume that sig- because of the physical proximity of the nodes. There would be
nals received from nodes other than the intended transmitter many flows within the network and there would be multihop-
are treated as interference. A distributed multiaccess contention ping. We wish to address the question of optimal organisation
*This research was supported by the Indo-French Centre for the Promotion of such an ad hoc network so as to maximise its transport capac-
of Advanced Research (IFCPAR), under project 2900-IT ity subject to a power constraint. The power constraint relates
to the network life-time and would depend on the application. (with single user decoding receivers) is bounded, i.e., SINR <
In a factory situation, it is possible that power could be sup- (NK( ))where IK(n,) denotes the interference at a node due
plied to the devices, hence large power would be available. In to spatial reuse. Using the finite (and fixed) area assumption, we
certain emergencies, "transient" sensor networks could be de- bound the minimum interference by any simultaneous transmis-
ployed for situation management; we use the term "transient" sion by P(K(n) Hence, SINR < (K(n))
as these networks are supposed to exist for only several minutes (v2A)7 N±(K(n) l1)P(K(-)).Th
(2A) 2
or hours, and the devices could be disposable. Such networks aggregate capacity (bits/slot) achieved in such a network is now
need to have large throughputs, but, being transient networks, bounded above by
the power constraint could again be loose. On the other hand
sensor networks deployed for monitoring some phenomenon in
a remote area would have to work with very small amounts of C(K(n)): K((n) log ± N ± (K(n)-1)P )
power, while sacrificing transport capacity. Our formulation
aims at providing insights into optimal network operation in a
range of such scenarios. Clearly, C(K(n)) is uniformly bounded for any K(n). Since
r(K(n)) < 2A, we see that the transport capacity (in bit-
A. Preview ofContributions metres per second, see Section IV) achieved in the network,
bounded above by C(K(n))r(K(n)), is also finite, indepen-
We motivate the definition of the transport capacity of the dent of the number of nodes and power P. Also, we ex-
network as the product of the aggregate throughput (in bits per pect that the transmitter-receiver separation (bounded above by
second) and the hop distance (in meters). For random spatio- r(K(n))) would decrease to 0 as K(n) increases to oc (finite
temporal fading, we seek the power control and the hop dis- area assumption). Hence, limK(n), C(K(n))r(K(n)) = 0.
tance that jointly optimise the transport capacity, subject to a This implies that there exists an optimal K(n), 1 < K(n) <
network average power constraint. For a fixed data transmis- oo, which maximises the transport capacity in the network, i.e.,
sion time strategy (discussed in Section III-B), we show that the optimum spatial reuse is finite. However, we note that, with-
the optimal power allocation function has a water pouring form out spatial reuse, a simple TDMA scheme with direct trans-
(Section V-A). At the optimal operating point (power control missions between the source and the destination with transmit
and hop distance) the network throughput (O opt, in bits per sec- power nP (and hence, an average power P), achieves log(n)
ond) is shown to be a fixed quantity, depending only on the con- order transport capacity. As seen above, with spatial reuse, the
tention mechanism and fading model, but independent of the system becomes interference limited, and hence, becomes inef-
network power constraint (Section V-B). Further, we show that ficient both for large n and for large P. Thus, we conclude that
the optimal transport capacity is of the form dopt (Pt) x (0opt, single cell operation (as defined earlier) is efficient for dense
with dopt scaling as Pt , where Pt is the available time average networks with single user decoding. In the context of sen-
transmission power, and Ti is the power law path loss exponent sor networks, log(n) scaling has been achieved with maximum
(Theorem V.2). Finally, we provide a condition on the fading node power constraints as well, using cooperative transmission
density that leads to a simple characterisation of the optimal techniques ([6]).
hop distance (Section V-C). With the above motivation, in this work, we study the trans-
port capacity of power constrained dense ad hoc networks op-
II. MOTIVATION FOR SINGLE CELL OPERATION erated as a single cell. More recently, El Gamal and Mam-
In this context, the seminal paper by Gupta and Kumar [4] men [7] have shown that, if the transceiver energy at each hop
would suggest that each node should communicate with neigh- is factored in, then the operating regime studied in [4] is nei-
bours as close as possible while maintaining network connec- ther energy efficient nor delay optimal. Fewer hops between
tivity. This maximises network transport capacity (in bit-metres the transmitter and receiver (and hence, less spatial reuse) re-
per second), while minimising network average power. It has duce the energy consumption and lead to a better throughput-
been observed by Dousse and Thiran [5], however, that if, un- delay tradeoff. While optimal operation of the network might
like [4], the practical model of bounded received power for fi- suggest using some spatial reuse (finite, as discussed above),
nite transmitter power is used, then the increasing interference coordinating simultaneous transmissions (in a distributed fash-
with an increasing density of simultaneous transmitters is not ion), in a constrained area, is extremely difficult and the asso-
consistent with a minimum SINR requirement at each receiver. ciated time, energy and synchronisation overheads have to be
The following discussion motivates a single cell operation in accounted for. In view of the above discussions, in this paper,
our framework, i.e., only one transmission exists in the network we assume that the multiple access control (MAC) is such that
at any time. only one transmitter-receiver pair communicate at any time in
Consider a dense planar wireless network with n nodes in a the network.
square of fixed area A. Let K(n) denote the spatial reuse in the
network (number of simultaneous transmissions) with r(K(n)) A. Outline of the Paper
bounding the transmitter-receiver separation. Let P(K(n)) be In Section III we describe the system model and in Section IV
the transmitter power per node (with a network average power we motivate the objective. We study the transport capacity of
P, as in Section V) and let N be the receiver noise power. a single cell multihop wireless network, operating in the fixed
The maximum SINR achievable per link in such a network transmission time mode, in Section V. Section VI concludes
the paper and discusses future work. The proofs of all lemmas to transmission, but remains constant over any packet transmis-
and theorems, if not in the paper, are provided in [10]. sion duration. Since successive transmissions can take place
between randomly selected pairs of nodes (as per the outcome
of the distributed contention mechanism) we are actually mod-
III. THE NETWORK MODEL
eling a spatio-temporal fading process. We assume that this
There is a dense collection of immobile nodes that use mul- fading process is stationary in space and time with some given
tiaccess multihop radio communication with single user decod- marginal distribution H. Let the cumulative distribution of
ing and packet forwarding to transport packets between various H be A(h) (with a p.d.f. a(h)), which by our assumption
source-destination pairs. of spatio-temporal stationarity of fading is the same for all
* All nodes use the same contention mechanism with the transmitter-receiver pairs and for all transmissions. We assume
same parameters (e.g., all nodes use IEEE 802.11 DCF a flat and slow fading channel with additive white Gaussian
with the same back-off parameters). noise of power ca2. And, Tc, the channel coherence time ap-
* We assume that nodes send control packets (such as plicable to all the links in the network, upper bounds the time
RTS/CTS in IEEE 802.11) with a constant power (i.e., taken to complete any data transmission in the network. We
power control is not used for the control packets) during assume that H and Tc are independent of the hop distance d.
contention, and these control packets are decodable by ev- When a node transmits to another node at a distance d (in
ery node in the network. As in IEEE 802.11, this can be the transmitting antenna's far field), using transmitter power P,
done by using a low rate, robust modulation scheme and by with channel power gain due to fading, h, then we assume that
restricting the diameter of the network. This is the "single the transmission rate given by Shannon's formula is achieved
cell" assumption, also used in [1], and implies that there over the transmission burst; i.e., the transmission rate is given
can be only one successful ongoing transmission at any by
time.
* During the control packet exchange, each transmitter hPa
c=wlog(1±cT2d)
C

learns about the channel "gain" to its intended receiver,

and decides upon the power level that is used to transmit where W is the signal bandwidth and a is a constant account-
its data packet. For example, in IEEE 802.11, the chan- ing for any fixed power gains between the transmitter and the
nel gain to the intended receiver could be estimated during receiver. Note that this requires that the transmitter has avail-
the RTS/CTS control packet exchange. Such channel in- able several coding schemes of different rates, one of which is
formation can then be used by the transmitter to do power chosen for each channel state and power level.
control. In our paper, we assume that such channel esti-
mation and power control is possible on a transmission- B. Fixed Transmission Time Strategy
by-transmission basis.
* In this work, we model only an average power constraint We focus on a fixed transmission time scheme, where all data
and not a peak power constraint. transmissions are of equal duration, independent of the bit rate
* We assume that the traffic is homogeneous in the network achieved over the wireless link. This implies that the amount
and all the nodes have data to send at all times; these could of data that a transmitter sends during a transmission opportu-
be locally generated packets or transit packets (saturation nity is proportional to the achieved physical link rate. Let T
assumption). (< Tc, the channel coherence time), be the data transmission
time. Upon a successful control packet exchange, the chan-
nel (between the transmitter, that "won" the contention, and its
A. Channel Model: Path Loss, Fading and Transmission Rate intended receiver) is reserved for a duration of T seconds in-
The channel gain between a transmitter-receiver pair for a dependent of the channel state h. This is akin to the "TxOP"
hop is a function of the hop length and the multipath fading (transmission opportunity) mechanism in the IEEE 802.1 1 stan-
"gain" (h). Based on our dense network and traffic homogene- dard. Thus, when the power allocated during the channel state
ity assumption, we further make the following assumption. h is P(h), and P(h) > 0, then data transmission occupies
* The nodes self-organise so that all hops are of length d, the channel for the duration T seconds, sending C(h)T bits
i.e., a one hop transmission always traverses a distance of across the channel, where C(h) = W log (I + P(2 dT). If
d meters. This hop distance, d, will be one of our optimi- P(h) = 0, we assume that the channel is left idle for the next
sation variables. T seconds.
The path loss for a hop distance d is given by d, where Ti is the The optimality of a fixed transmission time scheme, for
path loss exponent, chosen depending on the propagation char- throughput, as compared to a fixed packet length scheme, can
acteristics of the environment (see, for e.g., [8]). This variation be formally established (see [10]), but, due to lack of space,
of path loss with d holds for d > do, the far field reference dis- we only provide an intuition here. When using fixed packet
tance; we will assume that this inequality holds (d > do), and lengths, a transmitter may be forced to send the entire packet
will justify this assumption in the course of the analysis below even if the channel is poor, thus taking longer time and more
(see Theorem V.2). power. On the other hand, in a fixed transmission time scheme,
We assume that for each transmitter-receiver pair, the chan- we send more data when the channel is good and limit our in-
nel gain due to multipath fading may change from transmission efficiency when the channel is poor.
 IV. MULTIHOP TRANSPORT CAPACITY to vary d with fading, as routes cannot vary at the fading time
Let d denote the common hop length and {P(h) } a power scale. Hence, we propose to optimize first over {P(h)} for a
allocation policy, with P(h) denoting the transmit power used given d, and then optimize over d, i.e., we seek to solve the
when the channel state is h. We take a simple model for the following problem,
random access channel contention process. The channel goes
through successive contention periods. Each period can be ei- max
d
max
ffP(h)l:F(fP(h)})<P}
O({P(h)}, d) (2)
ther an idle slot, or a collision period, or a successful trans-
mission with probabilities pi, PC and ps respectively. Under the where the network average power, 1P({P(h) }), is given by,
node saturation assumption, the aggregate bit rate carried by
the system, (2) T({P(h)}, d), for the hop distance d and power 'P(fP(h)l) :=
allocation {P(h)}, is given by (see [2], or [1]) piEi +pcEc + ps (Eo + Tf0 P(h) dA(h) )
piTi +pcTc + p5(To + T) (3)
T ({P(h) } ((f0 l (h)± p,d A(h)
d):= p1T1Ps ±p,T, )
(T0, ±T) (1) Ei,EC and E correspond to the energy overheads associ-
ated with an idle period, collision and successful transmission.
where L(h) = C(h)T, and, Ti, T, and To are the average time Thus, Ei denotes the total energy expended in the network over
overheads associated with an idle slot, collision and data trans- an idle slot, E, denotes the total average energy expended by
mission. For e.g., in IEEE 802.11 with the RTS/CTS mecha- the colliding nodes, as well as the idle energy of the idle nodes,
nism being used, a collision takes a fixed time independent of and E denotes the average energy expended in the success-
the data transmission rate. We note that Pi, p, Pc, Ti, To, and T, ful contention negotiation between the successful transmitter-
depend only on the parameters of the distributed contention receiver pair, the receive energy at the receiver (in the radio and
mechanism (MAC protocol), and not on any of the decision in the packet processor), and the idle energy expended by all
variables that we consider. the other nodes over the time To + T.
With OT({P(h)},d) defined as in (1), we consider For a given d and power allocation {P(h) }, define the time
OT( ({P(h) }, d) x d as our measure of transport capacity of average transmission power, Pt ({P(h) }, d), and the time aver-
the network. This measure can be motivated in several ways. age overhead power, PO (which does not depend on {P(h) } or
OT({P(h)}, d) is the rate at which bits are transmitted by the d), as
network nodes. When transmitted successfully, each bit tra-
verses a distance d. Hence, OT( ({P(h) }, d) x d is the rate Ps (f0ooP (h) dA (h) )T
of spatial progress of the flow of bits in the network (in bit- Pt({P(h)},d) -=
p1Ti +±pcTc + ps(To + T)
metres per second). Viewed alternatively, it is the weighted p ._ piEi +pcEc +psEo
average of the end-to-end flow throughput with respect to the °piTi + PCTC + Ps (To + T)
distance traversed. Suppose that a flow i covers a distance
Di with Di hops (assumed to be an integer for this argu-
ment). Let /i3OT({P(h)}, d) be the fraction of throughput of Then the network power constraint can be rewritten as
the network that belongs to flow i. Then, OiET({P(h)},d) is the Pt({P(h)},d)<P -Po
d
end-to-end throughput for flow i and OiET({P(h),d) X Di
d where the right hand side does not depend on {P(h)} or d.
/iOT({P(h)} , d) x d is the end-to-endflow throughput for flow Observe that Pt(:= P -PO) is the time average transmission
i in bit-metres per second. Summing over all the flows, we have power constraint.
OT({P (h) }, d) x d, the aggregate end-to-flow throughput in
bit-metres per second. A. Optimization over {P(h)} for afixedd
With the above motivation, our aim in this paper is to max-
imise the quantity OT({P(h)}, d) x d over the hop distance d Consider the optimization problem
and over the power control {P(h)}, subject to a network aver-
max O({P(h)},d) (4)
age power constraint, P. We use a network power constraint {{P(h) }:P({P(h) }) <P}
that accounts for the energy used in data transmission as well
as the energy overheads associated with communication. The denominators of OT(., ) in (1) and of P in (3) are in-
dependent of d and the power control {P(h)}. Thus, with
V. OPTIMISING THE TRANSPORT CAPACITY d fixed, the optimization problem simplifies to maximizing
For a given {P(h) } and d, and the corresponding throughput fo7 L(h) dA(h) or,
OT({P(h)}, d), the transport capacity in bit-meters per second,
which we will denote by 0 ({P(h)}, d), is given by log (I+± (h)h) dA(h)
b({P(h)}, d) := OT({P(h)}, d) x d subject to the average power contraint,
Maximizing b(., -) involves optimizing over d, as well as j
{P(h)}. However, we observe that, it would not be possible P(h) dA(h) < Pt
where Pt is given by, 2) When the objective function (5) is unbounded, the opti-
mal solution occurs at d = oc (from the continuity re-
(piTi+PcTc+Ps(To+T))p- sults).
Pt': p,T P
3) We note that, in practice, Ti > 2.
Notice that Pt' is also independent of {P(h)} or d and is the Theorem V2: The following hold for the problem in (5),
average transmit power constraint averaged only over the trans- 1) Without the constraint d > do, the optimum hop distance
mission periods. dopt scales as (Pt') .
Without a peak power constraint, this is a well-known prob- 2) There is a value Ptin such that, for Pt > Ptlin,
lem whose optimal solution has the water-pouring form (see dopt > do, and hence the optimal solution obeys the scal-
[3]). The optimal power allocation function {P(h) } is given by ing shown in 1).
3) For Pt' > Pt$mi, the optimum power control {P(h)} is
P(h) ( of the water pouring form and scales as Pt'.
ha 7 4) For Pt' > Ptmi, the optimal transport capacity scales as
where A is obtained from the power constraint equation (Pt ) .

Proof:
1) Let dopt be optimal for Pt' > 0. We claim that, for x >
I >f2d7 a(h)P(h)dh = Pt 0, x n dopt is optimal for the power constraint xPt'. For
suppose this was not so, it would mean that there exists
The optimal power allocation is a nonrandomized policy, where d > 0 such that
a node transmits with power P(h) every time the channel is in
state h (whenever P(h) > 0), or leaves the channel idle for h
such that P(h) = 0.
1
tz T7dopt F
( -Ptp
x
)
17 d JtJ7 dr(xPt'
< dF t~ )

B. Optimization over d or, equivalently,

By defining 4?(h) : P(h), the problem of maximising the
throughput over power controls, for a fixed d, becomes
dptF(F P)) < X dr( A
\\opt (- 77 d)'n
Vx
max log (I+ ah2 (h)) a(h)dh
which contradicts the hypothesis that d,pt is optimal for
subject to Pt
2) Using the path loss model d , we see that for d < do, the
j D(h)a(h)dh < d received power is scaled more than P, due to the factor
d77 and an do factor in a, i.e., the model over-estimates
the received power and the transport capacity. Hence, the
Denoting by F (P-) the optimal value of this problem, the achieved transport capacity for d < do is definitely less
problem of optimisation over the hop-length now becomes than d F(f') . The result now follows from the scaling
result in 1).
maxd x F () 3) It follows from 1) that, if Pt scales by a factor x, then
d \doJ(5 the optimum d scales by xr , so that, at the optimum, P-'I
Theorem VI: In the problem defined by (5), the objective is unchanged. Hence the optimal {4 (h) } is unchanged,
which means that {P(h)} must scale by x. The water
dxF (Pt'), when viewed as a function of d, is continuously pouring form is evident.
differentiable. Further, when the channel fading random vari- 4) Again, by 1) and 2), if Pt scales by a factor x, then the
able, H, has a finite mean (E(H) < oo), then optimum d scales by . ,so that, at the optimum, P- is
1) limd,o d x F( ) = O and, unchanged. Thus F ( -) is unchanged, and the optimal
2) if in addition, Ti > 2, 12a (h) is continuously differ- transport capacity scales as the optimum d, i.e., by the
entiable and P(H > h) = 0 (12 ) for large h, then, factor x n .
limd, x
ddF(t ) o,
Proof: The proofs of continuous differentiability of d x Remarks V2:
Pt' ) 1) and 2) are provided in [10]. P The above theorem yields the following observations for the
fixed transmission time model.
Remarks VI:
1) As an illustration, withT7 = 3, in order to double the op-
1) Under the conditions proposed in Theorem V. 1, it follows timal transport capacity, we need to use 23 times the Pt '.
that d x F (Pt) is bounded over d C [0, oo) and achieves This would result in a considerable reduction in network
its maximum in d E (0, oo). lifetime, assuming the same battery energy.
 More generally, and still pursuing the discrete case, let N de-
note the set of fading states when the fading random variable is
Is discrete with a finite number of values; NX1 denotes the cardi-
nality of N.
Theorem V3: There are at most 2 NE I1 stationary points of
d F(wF(d)) inO < d < oo.
Proof: See [10] for the related analysis and the proof of
this theorem. P
We conclude from the above discussion that it is difficult to
characterise the optimal solution when there are multiple sta-
iol
tionary points. Hence we seek conditions for a unique positive
log(d) stationary point, which must then be the maximising solution.
In Appendix A, we have shown that the equation characterising
Fig. 1. Plot of d x F ( 3) (linear scale) vs. d (log scale) for a channel with
two fading states hl, h2. The fading gains are hi = 100 and h2 = 0.5, with
the stationary points, F(wF) -TrjA(-F) = 0, can be rewritten as
probabilities ahl = 0.01 = 1-ah2 The function has 3 non-trivial stationary
points.
1
A2
(log(y) .TI(y 1))-f dy
A
0 (6)
y2 y
2) We observe that as the power constraint Pt' scales, the op- for (x)
f a (ffj) 5, the density of the random variable
timal bit rate carried in the network, F (Pt), stays con-
H '. Notice that wF does not appear in this expression. The solu-
stant, but the optimal transport capacity increases since tion directly yields the Lagarange multiplier of the throughput
the optimal hop length increases. Further, because of the maximisation problem for the optimal value of hop length. The
way the optimal power control and the optimal hop length following theorem guarantees the existence of atmost one sta-
scale together, the nodes transmit at the same physical bit tionary point of (6).
rate in each fading state; see the proof of Theorem V.2
part 3). P Theorem V4: If for any A1 > A2 > °, f( Ay,) is a strictly
monotonic decreasing function of y, then the objective function
C. Characterisation of the Optimal d d x F (P) has at most one stationary point dopt, 0 < dopt <
By the results in Theorem V. 1 we can conclude that the opti- 00.
mal solution of the maximisation in (5) lies in the set of points Proof: The proof follows from Lemmas A. 1, and A.2 in
for which the derivative of d x F (Pt) is zero. For fixed Pt , Appendix. P
Corollary VI: If H has an exponential distribution and 7i >
define -F(d) :-Pt' Differentiating d x F(wF(d)), we obtain, 2, then the objective in the optimisation problem of (5) has a
(see Appendix A) unique stationary point dopt E (0, oc), which achieves the max-
imum.
d (d F (w(d)) F (w(d)) -Ir(d)A(w(d)) Proof: a(h) is of the form ,ue-ph and the monotonic-
ity hypothesis in Theorem V.4 holds for a(h). Also, from
where A(wF) is the Lagrange multiplier for the optimisation Theorem V.1, we see that limd,o d x F1 7 = 0 and
problem that yields F(7r(d)). Since d appears only via 7r(d),
we can view the right hand side as a function of -w. We are in- limdO d x F (Pt)' 0.
terested in the zeros of the above expression. Clearly, -w = 0 is Remarks V4: 1) Hence, for Ti > 2, for the Rayleigh fad-
a solution. This solution corresponds to the case d = oc; How- ing model there exists a unique stationary point which
ever, we are interested only in solutions of d E (0, oc), and corresponds to the optimal operating point.
hence, we seek positive solutions of -w of F(-w) -rTrA(wF) = 0. 2) For P ' > in, and for the conditions in Theorem V.1
Remarks V3: The above analysis has been done for a contin- and V.4, let 7Fopt denote the unique stationary point of
uously distributed fading random variable H. The analysis can (6). Then define F(wFopt) = Oopt. It follows from The-
be done for a discrete valued fading distribution as well, and orem V.2 that the optimal transport capacity takes the
1
we provide this analysis in [10]. The following example then
form Pt ) " 0 opt, where 0 opt depends on a(h) and the
illustrates that, in general, the function 1(-w) -rTjA(wF) = 0
can have multiple solutions. Consider a fading distribution that MAC parameters but not on P (or Pt).
takes two values: h1 100 and h2 = 0.5, with probabilities 3) Figure 2 numerically illustrates our results for Rayleigh
ah1 = 0.01 1
1-ah2. Figure 1 plots d x F ( ) for the sys- fading and Ti = 2. Scaling Pt by 4 scales the transport
tem with Ti 3. Notice that there are 3 stationary points other capacity from 2.3 to 4.6, i.e., by 4 = v4 and similarly
than the trivial solution d = oo (which is not shown in the fig- for scaling Pt' by 9.
ure). Also, the maximising solution is not the first stationary The uniqueness results guarantees that a distributed imple-
point (the stationary point close to 0). If, on the other hand, mentation of the optimization problem, if it converges, shall
ahl = 0.001 = 1 - ah2, we again have 3 stationary points, but converge to the unique stationary point, which is the optimal
the optimal solution now is the first stationary point. solution.
 . +.: +.: -.:*+-:,~-. ~:.+~: -~=.:~-lp: .
7
[4] P. Gupta and P.R. Kumar, The Capacity of Wireless Networks, IEEE
6
Trans on IT, 2000.
[5] 0. Dousse and P. Thiran, Connectivity vs capacity in dense ad-hoc net-
E.

~ ~ ~.
works, Proc IEEE Infocom, 2004.
5
[6] Hesham El Gamal, On the scaling laws of dense wireless sensor networks:
the data gathering channel, ToIT, 2005.
4
[7] Abbas El Gamal and James Mammen, Optimal Hopping in Ad hoc Wire-
less Networks, Proc IEEE Infocom, 2006.
.,
,.. ... ... .,,,,_.....

3 ,+ . . .. , ,,, , ,,.. . . ~4 [8] T.S. Rappaport, Wireless Communications: Principles & Practice, Pren-
...... , , , , ,., .,.,., .,-., -.E
, ,+-.,._~~~~~~~~tt
, .,,.,.... 41Y tice Hall, 2002.
2 [9] J.P. Aubin, Applied Functional Analysis, John Wiley & Sons, 2000.
[10] V. Ramaiyan et al., Jointy Optimal Power Control and Routing
for a single Cell, Dense, Ad hoc Wireless Network, Tech Rep. at,
http://ece.iisc.ernet.in/anurag/papers/anurag/ramaiyan-etalo6singlecell-
u
-2
;

10-1 100
;

101
;;;;;

102
multihop.pdf.gz, 2006.

APPENDIX
Fig. 2. Plot of d x F (d) (linear scale) vs. Xr (= d ) (log scale) for a fading A. Stationary Points of d x F (7r (d))
channel (with exponential distribution). We consider 3 power levels (Pt', 4Pt'
and 9Pt') and 2. The function has a unique optimum 'JUvpt (iTvpt 0.2)
Recall that we defined -r(d) : P= t . Further, F(wr(d)) was
for all the 3 cases. defined by
Oc):la ( 2 P(h)
)ah
VI. CONCLUSION
F(F()):=maxj log (I+ a(h)dh
2 d (7)
In this paper we have studied a problem of jointly optimal where the maximum is over all power controls {P(h) } satisfy-
power control and self-organisation in a single cell, dense, ad ing the constraint
hoc multihop wireless network. The self-organisation is in
terms of the hop distance used when relaying packets between aO(h)Odh < 7r(d)
J(h) (8)
source-destination pairs.
We formulated the problem as one of maximising the trans- For ease of notation, let us use the substitution := 'h. Write x

port capacity of the network subject to an average power con-

straint. We showed that, for a fixed transmission time scheme,
~(X) =(( ) P(r) and f(x) a (Ja)2 Note that .

f (.) is the probability density ofthe random variable X: CH.

there corresponds an intrinsic aggregate packet carrying capac-
Then, equations (7) and (8) can be rewritten as
ity at which the network operates at the optimal operating point,
independent of the average power constraint. We also obtained
the scaling law relating the optimal hop distance to the power 17(w) = max log(1 + xz(x))f (x)dx
constraint, and hence relating the optimal transport capacity to and
the power constraint (see Theorem V.2). Because of the way
the power control and the optimal hop length scale, the optimal j (x) f (x)dx < wF

physical bit rate in each fading state is invariant with the power This optimisation problem is one of maximising a convex func-
constraint. In Theorem V.4 we provide a characterisation of tional of { (x)}, subject to a linear constraint. The optimal so-
the optimal hop distance in cases in which the fading density lution of the problem has water-pouring form, and the optimal
satisfies a certain monotonicity condition. solution is given by,
One motivation for our work is the optimal operation of sen-
sor networks. If a sensor network is supplied with external +
power, or if the network is not required to have a long life- where (ibai A("w) I
J
time, then the value of the power constraint, P, can be large, where A (w) is obtained from
and a long hop distance will be used, yielding a large transport 00
1
capacity. On the other hand, if the sensor network runs on bat- I) f (x)dx:
teries and needs to have a long life-time then P would be small, X(7) VA(T)
yielding a small hop length. In both cases the optimal aggregate Further, the derivative of the optimum value F(wF), w.r.t. IF, i.e.,
bit rate carried by the network would be the same. ar(w) = A(wF) (see Aubin [9]).
Future work on this topic will include developing a dis- B7r

Let us now reintroduce the dependence on d, and consider

tributed algorithm for nodes to adapt themselves towards the the problem of optimising d x F(wF(d)) over d. Differentiating
optimal operating point. d F(wF(d)) w.r.t. d, we get,
a

REFERENCES
a

(d F(-r(d)) F (w(d)) + d F(-F(d))

[1] A. Kumar, E. Altman, D. Miorandi and M. Goyal, New Insights from
a fixed point analysis of single cell IEEE 802.11 Wireless LANs, Proc F (-(d)) + d
O
(wr(d)) x (d)
Infocom, 2005.
[2] G. Bianchi, Performance Analysis of the IEEE 802.11 Distributed Coor- /P
F (w(d)) + d F'(7r(d)) dR+
f

dination Function, IEEE JSAC, 2000.

[3] A.J. Goldsmith and P.P. Varaiya, Capacity of Fading Channels with Chan-
nel Side Information, IEEE Trans on IT, 1997. F (-(d)) -TI(d)F'(w(d))
where F'(wF) := 9F(w). Substituting F'(-F) = A(wF), we have, (17) provides the I (actually I(wF)) to be substituted in (16). Sub-
stituting for 17(w) (from (16)), and for I (from (17)), into the
a (dF(-(d)) = F(-(d))-T(d)A(-(d)) q) right hand side of (9), dividing across by P(Z < 1), and using
the definition of El (.), we have,
The stationary points of d x F (w(d)) are now obtained by equat-
ing the right hand side of (9) to zero. Note that since d appears (P(w
log <(Z)))-E
± -1(log(Z)) -
7± G (Z)
in this equation only as wr(d), we need only study the roots of
the equation
log (pz<I + El (Z)) -El(log(Z))- wr ± 7jw ( 0
F(wF) - TI-wA(wF) = 0 (10) G, (Z)
We now proceed to obtain a characterisation ofthe stationary
log (G?z) + I) El (Z)] + log (e- (log(z)))
points. Substituting the optimal solution in the expression of
F(wF) and A(wF), and suppressing the argument wF in A(wF), we Tw+Z 0
get, 7r + G, (Z)
F(w) j log (iD f (x)dx (1 1) Rearranging terms, we get,
with A being given by
log (w G±G(Z)) + log (El(Z)e -El(log(Z)))
I
Ja (
x/
ff(x)dx (12) 0
7r + G, (Z)
Using the substitution z 1, I 1, and defining g(z) Denote b, = log (Ei (Z)-Ej (log(Z)) ). Then, we have,
2f (1),we get,
log (w G (Z)) + b Tw 0
F(w) jlog(-)g(z)dz (13) 7r + G, (Z)
Using (17), we have
with I (actually, I(wF)) being given by
G, (Z) G, (Z) El (Z)
G, (Z) IP(Z < 1) I
wr (I z)g(z)dz (14) 7r +

which, with the previous equation, yields

We note that 9(.) is the density of the random variable Z
1 _
log(E(Z)) + b El (Z))
2

17(1 0
X ctH
For t(.) of the random variable Z, define the oper-
a function
ators El(.) and Gi(.) as Recall that I is actually I (wF). We now find that wF appears in the
equation only as I(wF). Hence we can view this as an equation
fo t(z)g(z)dz in the variable I (= X ). Rearranging terms, we get
El (t(Z))
fog(z)dz
1 -log( El (Z)) ±17El (Z) -(b1-TI)
G1 (t(Z)) a t(z) g(z) dz
Exponentiating both sides, and substituting back for bl, yields
Lemma A. 1. The roots of(10) are equivalentto obtaining the
roots of the equation El (Z) El (Z)
El(Z)c- El (log (Z)) C-77
rTG 1 (AZ - 1) = G 1 (log(AZ)) (15)
On cancelling El (Z), and transposing terms, we next obtain
with 7 then being given by
( El (Z) _i) c- El Oog( 'i ))
= -)f (x)dx
or,
Proof: Using the definitions of El(.) and Gi(.), (13) and e O(El ( 1
)) c- El Oog( zi ))
(14) simplify to
Taking log on both sides, we have,
F (w) = log(l) P(Z < l) - G1 (log(Z)) (16)
z I
7w = IP(Z < I) - G1(Z) (17) Ti El
I El (log (h))
In terms of G, (.), this is equivalent to Hence,

TIG, G, (log( )) >1

f,Y c(y) Ibx (y)dy f1, c(y)bx (y)dy

which is the desired result after writing I = . X
We next address the question of a unique positive solution i.e.,
of (15). The following lemma guarantees the existence of a f c(y) bx2 (y)dy f> 'I c(y) bx (y)dy
unique positive solution, when f (.), the density of fH, satisfies
a certain monotonicity condition.
fy, c(y)bx2 (y)dy fy, c(y)bxl (y)dy
Lemma A.2: (15) has at most one positive solution if for any i.e., the ratio of the negative area of the integral to the positive
fA2) area ofthe integral is a strictly monotonic function of A. Hence,
A1 > A2 > 0, f ( ) is a strictly monotone decreasing function as A decreases, the integral can cross 0 at most once, or, there
of y. exists at most one (non-trivial) solution for (18). A
Proof: Expanding G 1 (.), (15) becomes,

Iq (AZ 1)g(z)dz -j log(Az)g(z)dz= 0

Rewriting the equation in terms of f (.), we have,
1

0
0
T z
)- o( z)) Z/d
In this last equation change the variable to y := Az, yielding
1
A2 (A d
(log(y) rjI(y 1))-yf A-Jdy 0 (18)
y2 \y
Define c(y) := (log(y) -i(y -1)) y and bx(y) : A

8)
Thus, we are interested in a positive A that solves
1
c(y)bA (y)dy = °
Observe that limy-o c(y) =-oc and c(1) = 0. Further, there
exists a unique y' such that c(y) < 0 for all 0 < y < y' and
c(y) > 0 for all y' < y < 1. Since bx(y) > 0 for all y and A,
we have c(y)bA(y) < 0 for all 0 < y < y' and c(y)bA(y) > 0
for all y' < y < 1.
Consider A1, A2 such that A1 > A2 > 0. By hypothesis,
bA2 (y) iS a strictly monotone decreasing function of y. Hence,
bcl(y)b()
c(y)bA2 (y) is also a strictly monotone decreasing function of y.
We then have,

c(y) bx (y)dy IC(y) 2

2 bx (y)dy

fA c(y) bx1 (y)dy f' c(y) bx1 (y)dy

bx2 (y')
>bx, (yl)'
And,

f,
c(y)bx2 (y)dy , c(y) 6j {8} blx (y)dy
f, c(y)bx1 (y)dy f, c(y)bx1 (y)dy
bx2 (y')
bx1 (yl)