A Power Control Game with Smooth Reduction of SINR Objectives
Omur Ozel Elif Uysal-Biyikoglu
Department of Electrical and Computer Engineering
University of Maryland College Park MD USA 20742
omur@umd.edu
AbstractWhen rate expectations of users in a wireless
network cannot all be satisfied, one choice is to discard some
users from the system, in a mechanism called admission control.
However, in a data network, users have a certain tolerance to
occasional rate outages. In this paper we argue that it may be
preferable for users to reduce their rate objectives smoothly,
by considering an outage probability tolerance, rather than
not be provided any service at all. We propose a distributed
utility based algorithm for doing this. The smoothness of
reactions is obtained by making transmit power reaction curves
absolutely subhomogeneous functions of interference. This is
done using an objective reduction factor, in addition to a
linear price. We first provide conditions for a unique Nash equi-
librium. Assuming that distributed nodes use gradient based
optimization, convergence and error sensitivity of gradient
based iterative algorithms are analyzed. Lastly, the continuous-
time counterpart of the problem is considered and a stability
condition is established for the system.
I. INTRODUCTION
The number of users that can be supported in a wireless
network in a given bandwidth depends highly on the rate
expectations of these users, their power constraints, and
channel gains. In an interference network, depending on the
number of users in the network, the rate expectations of users
can sometimes be unrealistic. And yet, these users may prefer
to lower their expectations rather than be provided no service.
If the network has a mechanism through which the transmit
powers of users (therefore the interference on each other) can
be controlled, it may be possible to reach operating points
where the number of users that are supported at a satisfactory
quality of service is maximized.
There is a large body of literature on power control mech-
anisms. Among them, utility-based power control algorithms
play an important role, as a natural method to generate a
distributed algorithm. A utility based power control algorithm
basically relies on users egoistically adjusting their own
transmit power to maximize a net utility. Results of such
behavior can be analyzed in depth using equilibrium concepts
of economic models. It is interesting to see that the game can
be set up such that the egoistic behavior leads to what is in
effect cooperation.
Noncooperative power control has been studied in a
multitude of previous work [1][4] with an emphasis on
mechanism design. A common point of them is the pricing
as a punishment mechanism to alleviate inefficiency due
to non-cooperative behaviour. However, the prices are not
appropriate for a fully distributed system since the coefficents
of prices are determined by another agent (a center). This
point is addressed in [5], where the pricing coefficients are
also determined distributedly to reach the socially optimum in
an ad-hoc wireless network. Actually, prices are not suitable
Department of Electrical and Electronics Engineering
Middle East Technical University Ankara Turkey 06531
elif@eee.metu.edu.tr
for explaining the cooperation mechanism as it has economic
implications and it necessitates price collector(s) who bene-
fit(s) from the collected amount. In [6], it is observed that
the cooperative effect of pricing is due to reduction of target
SIR value as the interference increases. Hence, pricing turns
out to be an implicit method to drive users to reduce their
rate objectives as interference increases. In this paper, we
attempt to view the problem directly in terms of reactions to
interference, whereby users explicitly reduce their objectives
in reaction to interference.
Users objectives are stated in terms of an outage prob-
ability. Outage probability formulations for power control
have appeared in the literature relatively recently; a notable
example is the work of Alpcan et al. [7] where a power
control game based on outage probabilities in multicell
wireless networks is control-theoretically analyzed. Defining
the utility as a concave function of non-outage probability,
stability conditions of associated dynamical system and con-
vergence conditions for gradient based iterative power update
algorithms are developed.
The notion of objective reductions was proposed in Ozel
et al. [8] to obtain a smooth power control scheme that
allows every user to stay in the network. The main goal of
this paper is to investigate the smooth cooperative mechanism
of [8] from the control theoretic perspective of [7]. We will
analyze a game in which the smoothness of reaction curves
is ensured through the use of absolutely subhomogeneous [9]
reactions. We first provide conditions for a unique Nash
equilibrium. Then, assuming that nodes employ distributed
gradient based optimization, convergence of gradient based
iterative algorithms are analyzed. Lastly, the continuous time
counterpart of the problem is considered and a stability
condition is established for the system.
The following is the organization of the rest of the pa-
per: The next section introduces the system model. In III,
the utility function and the objective reduction factor are
presented. In IV, existence and uniqueness conditions for
NE are given. A gradient based distributed power control
algorithm in which each user uses its own utility function
is analyzed in V and the analysis of the same algorithm,
carried to continuous domain is given in VI. Numerical
illustrations are provided in VII. The paper is concluded
with an emphasis on main points in VIII.
II. SYSTEM MODEL
We model an ad-hoc wireless network with a number of
links {i = 1, . . . , K}, each link corresponding to a distinct
transmitter. Link i has a transmitter t(i), and receiver r(i).
The receivers r(i), i = 1, . . . , K are not necessarily distinct
(-for example, the uplink of a centralized network having
 t1
a single base station or access point can be modeled by
G
taking all r(i)s as the same node.) In the sequel, the terms G14 = G
13
11 G
12

transmitter, user and link will be used interchangably. The r1

set of link indices will be denoted by < K >. G21 G22
r2
Unlike in a wired network, link i and link j interact due
t2 G =G
23
r3 = r4
to the interference one causes to the other. The channel gain 24
G32 G 42

between t(i) and r(j) is Gij hij . That is, due to a signal of G
33
power Pi transmitted by the sender of link i, the receiver t3
G
43
= G44

of link j receives a signal (interference if i 6= j) of power

Gij hij Pi . hij represent fading coefficients with unit variance,
while {Gij } are mean value of the random channel gains due
to the specific wireless environment, channel allocation and
coding scheme used. This model is quite general in that,
by proper choice of the coefficients {Gij }, it can model
centralized or distributed network architectures, as well as
different channel access mechanisms such as CDMA, TDMA
and FDMA.
For convenience, in the rest of the paper, a Rayleigh-
Rayleigh [10] environment is assumed, i.e., all signal and
interference terms are subject to Rayleigh fading. Accord-
ingly, the hij are independent exponentially distributed ran-
dom variables with unit variance. Gij are assumed constant
during the operation. For convenience, additive noise and
interference are modeled as Gaussian random variables with
zero mean and variance 2 .
Single user decoders are used in the receivers. Depending
on the channel access method, cross channel gains may be
suppressed by additional processing gains. We assume that
additional gains are also represented in the channel gains.
The signal to interference plus noise ratio (SINR1 ) of user i
is given as:
Gii hii Pi
i = 2 P (1)
+ j6=i Gji hji Pj
Without loss of generality, we consider that each user is
subject to the power constraint Pmin Pi Pmax i. The
upperbound on power models physical limitations while the
lower limit can be considered as a variable system parameter
(which will be discussed later in the paper).
Average SIR with respect to all hji is SIRiavg :
Gii Pi
SIRiavg ,
P Ii
where Ii = G P + 2
is the average interference
j6=i ji j
experienced by user i: The cross channel gains and power
values determine the range ofPinterference Imin i
Ii
i i 2
Imax where Imin = + j6=i Gji Pmin and Imax i
= of a user may depend on the type of application as well as
2
P constraints such as energy. Packet success probability (or a
+ j6=i Gji Pmax . Link i has interference suppression Si ,
defined by Si = P GiiGji .
j6=i
We let Oi be the probability that a wireless link ex-
periences outage: Oi , P r(i < th ) where th is the
minimum SIR required for communication. th is determined
by several system properties such as rate, modulation and
receiver structure, which is outside the scope of this paper
(e.g. see [11], [12]). In the rest, we assume the same th for
each user2 .
1 Since the interference limited case is considered, in the rest of the paper
the abbreviation is SIR
2 The results straightforwardly extend to the case of different th s for
users, which can model e.g.different receiver structures.
t4
Fig. 1. The ad-hoc wireless network model: A set of interfering links
Let xji = Eh (Gji hji Pj ) = Gji Pj be the average interfer-
ence power of user j on user i and xii = Eh (Gii hii Pi ) =
Gii Pi be the average received signal power of user i.
The following expression for Oi in the Rayleigh/Rayleigh
environment is derived in [13]:
2 th Y 1
Oi = 1 exp( ) th xji
(2)
xii
j6=i 1 + xii
In [10], outage probability in Rayleigh faded noiseless multi-
ple access channel is upper bounded by an exponential term
and this result was extended in [12] for the noisy case as
follows: th
Pr(i < th ) 1 exp( ) (3)
SIRiavg
The above upper bound well approximates outage probability,
especially when this probability is below 20%, and in fact the
approximation becomes quite tight for both extremes [10].
In a wireless data network, users will have a certain
tolerance to outage events: through link or transport layer
control mechanisms such as ARQ, they can ask for repetition
of lost packets, and maintain sufficient quality of service even
though the channel is not completely reliable. If utility is the
eventual quality of service, one could define it in terms of the
required outage probability tolerance, as a function of average
transmit power. In the next section, the utility function will
be made precise.
III. UTILITY FUNCTION AND OBJECTIVE
REDUCTION FACTOR
In game-theoretic terms, a utility functionQis a mapping
K
from the action space into real numbers ui : j=1 Aj R.
Aiming at developing distributed power control algorithms,
the action space used in the game theoretic analysis will be
defined as a set of users power levels. The actual utility
function thereof) has been a popular choice of utility func-
tion, considering an ARQ-type link layer scheme and energy
efficiency as the main concern [4], [14]. This motivates a
step function or a sharply rising sigmoid of SIR as a model
of utility for constant-rate real-time applications.
Following our work reported in [8], the derivation of our
utility function will be based on each user having an outage
probability objective, Oobj that it wants to stay below. For
simplicity, we assume that this target outage probability is
the same for all users. Utility is determined by the relative
value of non-outage probability 1 Oi with respect to the
objective 1 Oobj . Given an outage objective, we can find
 i
analysis, hence the following definitions are made: Hmin =
i i i
1
minx[Imini i
,Imax ] H (x), H max = max i i
x[Imin ,Imax ] H (x),
i i
0.8 Hmax = maxi Hmax and Hmin = mini Hmin .
0.6
The selection of fdi will be such that the functions
[Hi (x)]1/2 i are absolutely subhomogeneous, which further
fdi (x)

0.4
0.2
0
0 1 2 3 4 5
x A formal definition of absolute subhomogeneity proposed
Fig. 2. An example of objective reduction factor fdi
SIRobj such that SIRiavg SIRobj ensures that the outage
objective is reached [13]. From (3), we choose SIRobj such
th
that exp SIRobj = 1Oobj . We start by the rather simplistic
utility given below:
SIRobj
ui = (4)
SIRiavg
It can be deduced by a quick investigation that the game in
which users non-cooperatively maximize the utility in Eq. (4)
has an inefficent NE: Each user transmits with power Pmax
in NE. Hence, a more sophisticated utility function that leads
to a cooperative mechanism is required.
In [8], we proposed that node i cooperates by reducing its
SIR objective with a factor fdi . User is reaction to interfering
power is shaped by fdi (.). The factor fdi (.) is basically a non-
increasing continuous function defined from non-negative
real numbers into the range [0, 1]. An example function fdi (.)
is illustrated in Fig.2. In general, we assume that fdi is first
order differentiable except at some points, the set of which
has finite cardinality.
The concept of objective reduction factor will be central
to the development of the utility function here too, albeit
with a difference: here, a unit price pi is subtracted from
utility to obtain a net utility. It will be observed later in the
paper although users reduce their objectives as interference
escalates, inefficency in NE is not alleviated unless this price
term exists. The net utility function N Ui is the following:
SIRobj fdi (Ii )
N Ui (P) = pi (5)
SIRiavg
Our net utility function employs a linear price, combined with
an explicit objective reduction factor whereby SIR objectives
are effectively reduced in response to increasing interference.
This is akin to the approach of [3], [6] where, reduction of
objectives is not explicit, but adaptation to interference is
made within a price coefficient. In fact, if net utility were
defined as below, the resulting equilibria would be identical:
SIRobj 1 dx
N Ui = i pi (6)
SIRiavg fd (Ii )
This point will be made clearer during the analysis of reaction
curves in the next section. We will devise a distributed
mechanism by setting up the power control problem as
a non-cooperative game and analyze the properties of the
resulting equilibrium. In particular, selection of fdi functions
is key to the equilibrium and dynamic of the system. In the
analysis, the function H i (x) = xfdi (x) plays an important
role. H i (x) is continuous and piecewise continuously dif-
ferentiable. Optimum values of H i (x) will be used in the
implies that [xfdi (sx)]1/2 for all positive s are absolutely sub-
homogeneous. We will next define absolute subhomogeneity
and point important implications of this property.
A. Absolutely Subhomogeneous Functions
by Nuzman [9] is:
Definition 1: A real valued function g is called absolutely
subhomogeneous (AS) if the following inequalities hold for
all x R+ and arbitrary a
e|a| g(x) g(ea x) e|a| g(x)
Absolutely subhomogeneous (AS) functions involve a sense
of smoothness observed in limited rate of change. If value
of the function increases (as its argument is increased), the
rate of increase is less than linear. If the value decreases, its
absolute rate of decrease is smaller than reciprocal of linear.
A collection of useful results in [9] about AS functions
are presented in the next theorem.
Theorem 1: Let P(t + 1) = F (P(t)) be MSS and
let H(P) = (F (P)). If vector function is in the
form (x1 , x2 , ..., xK ) = [1 (x1 ), 2 (x2 ), ..., K (xK )] with
i (xi ) absolutely subhomogeneous, then P(t+1) = H(P(t))
has a unique fixed point.
Theorem 1 suggests that if an AS function is applied to each
component of an MSS update algorithm, then there exists a
unique fixed point of the new algorithm. This result justifies
the choice of fdi such that Hi are AS (see [8]).
Continuity will be main assumption in our analysis. We
present a useful implication of AS property under continuity
in the next lemma.
Lemma 1: Let g : + + be a bounded continuous
absolutely subhomogeneous function and let r(x) = g n (x)
for some positive integer n. Subdifferential of r, r(x)
satisfies: r(x) r(x)
r(x) [n ,n ]
x x
Proof: Assume that g(.) is differentiable at point x with
a positive derivative and let x + h = ea x. For h < 0, a < 0
and for h > 0, a > 0. Using the right inequality of AS
definition, we have,
dr(x) r(x + h) r(x) (en|a| 1)r(x)
= lim lim
h0 h a0 (ea 1)x
Hence, provided a positive derivative, dr(x) dx n r(x)
x . If
we had a negative derivative, we can use a similar argu-
ment and left inequality of AS definition to conclude that
dr(x)
dx n r(x)x . The results straightforwardly extend for
points where the derivative is undefined which can only be
due to inequality of left and right limits. We use the definition
of subdifferential in terms of left and right derivatives and a
similar treatment of the problem yields
r(x) r(x)
r(x) [n ,n ]
x x
 IV. EXISTENCE AND UNIQUENESS OF NASH and N Ui
Pi < 0 at Pi = Pmax . As NE is the intersection
EQUILIBRIUM of reaction curves, these necessary boundary conditions can
be expressed in terms of reaction curves. In order not to
By assumed properties on reduction factor fdi (.), the have an equilibrium on the boundary, we must have Pmin <
net utility function N Ui (P) in Eq. (5) is continuous and ri (Pi ) < Pmax . Pluging the value, we reach a condition for
piecewise continuously differentiable with respect to its argu- an inner NE:
ments. In particular, the derivative of N Ui (.) with respect to SIRobj
Pmin < [ Ii fd (Ii )]1/2 < Pmax
Pi is continuous while there may be discontinuities at some Gii
The following two boundary conditions are assumed to hold
breaking points in the derivative of N Ui with respect to Pj ,
in the rest of the paper,
j 6= i. Assumption 1:
The game is = [U, {Si }, {N Ui }] where Si =
[Pmin , Pmax ] and each users optimization is SIRobj Hmini
SIRobj Hmax
i
2 > 1 and 2
<1
Gii Pmin Gii Pmax
max N Ui (Pi , Pi ) (7)
Pi Si We state our findings about NE of the game in the
By direct evaluation of the second derivative, we have following theorem
Theorem 3: If boundary conditions in assumption 1 hold
2 N Ui SIRobj Ii fd (Ii ) and if [Hi (x)]1/2 are AS i, then the game has a unique
2 = 2
Pi Gii Pi3 inner NE.
Proof: The existence of NE is guaranteed. Unique-
and except some points at which derivative is not defined,
ness result follows from Theorem 1. It is well known that
we have
Pi (t + 1) = eIi (t) =: Fi (P(t)) where e is a positive
2 N Ui SIRobj Gji constant, is an MSS update algorithm. After a few algebraic
= [Ii fd (Ii ) + fd (Ii )]
Pi Pj Gii Pi2 manipulations, one can show that the reaction curve in Th. 2
Note that N Ui is strictly concave with respect to its action can be expressed as
2
variable Pi as PN Ui
< 0. Hence, there exists a solution to ri (Pi ) = min{Pmax, max{Pmin , Pbi }}
2

where Pbi = vHi ( Fi (P(t))

i
each users optimization. Moreover, the cartesian product of c ) for some v > 0. Finally, AS
action sets form a non-empty, compact and convex set, which functions are closed under clipping operations (see Lemma
guarantees existence of a NE (due to Theorem 4.4 p.176 in 1 in [8]). Boundary conditions in assumption 1 assure that
[15]). NE is not on the boundary.
NE can be defined in terms of reaction curves. Reaction
curve (or best response strategy) of user i, ri (Pi ) is defined V. ITERATIVE POWER UPDATES AND
as CONVERGENCE
r (P ) =
i i max N U (P , P ) i i i In this section, we will investigate certain distributed
Pi [Pmin ,Pmax ]

NE point P is such a point that Pi
= ri (Pi ) for all i. power iterations in which transmitters use a gradient based
The explicit expression of reaction curve for our particular algorithm to optimize the net utility with step size as
problem is given as in the next theorem: follows: N Ui (P)
Pi (n + 1) = Pi (n) + (8)
Theorem 2: For the net utility function defined in Eq. (5), Pi
the reaction curve is We consider the update algorithm I(P) with Ii (P(n)) =
SIRobj Pi (n + 1). In our analysis, we will follow the framework
ri (Pi ) = min{Pmax , max{Pmin , [ Ii fd (Ii )]1/2 }}
Gii in [7].
Proof: Given Pi , maximum of N Ui (Pi , Pi ) occurs
either at boundary points Pi = Pmin , Pi = Pmax or at the A. Analysis of Synchronous Updates
point where the derivative wrt Pi is zero. Direct evaluation Let unique NE of the game be P . A set of functions for
of derivative of N Ui yields: analysis, ci (.) : [0, 1] are defined as follows:
N Ui SIRobj Ii fd (Ii ) ci ( ) , Pi + (1 )Pi + i ( P + (1 )P ) (9)
= 1
Pi Gii Pi2 We will reach a sufficient condition so that the algorithm has
Pbi that satisfies N Ui
= 0 is Pbi = [ SIR
obj the contraction property under max norm:
1/2
Pi Gii Ii fd (Ii )] .
Hence, kI(P) I(P )k kP P k for some 0 < < 1
max N Ui (Pi , Pi ) = min{Pmax , max{Pmin , Pbi }} We have the following bound Z 1
Pi [Pmin ,Pmax ] dci ( )
|Pi Pi | = |ci (1) ci (0)| = | d |
Note that the reaction curve would be the same if the utility 0 d
Z 1
function were defined as in Eq. (6). The reaction curve dci ( )
| d |
in above theorem can be expressed in terms of H i (x) as 0 d
follows: dci ( )
SIRobj i max | |
ri (Pi ) = min{Pmax , max{Pmin , [ H (x)]1/2 }} [0,1] d
Gii
i ( )
In order to guarantee an inner NE, N Ui
Pi > 0 at Pi = Pmin Hence, in order to reach the condition, we will bound | dcd |.
 dci ( ) i X i
| | = |(1 + )(Pi Pi ) + (Pj Pj )| lossy environment, we take an illustrative example: the error
d Pi Pj terms are coalesced into a single random variable as a linear
j6=i
By triangle inequality and after some manipulations, factor to the update algorithm.
dci ( ) i X i e
N Ui (P)
| | (1 + ( + | |)) kP P k Pei (n + 1) = Pei (n) + i (n) e i ) (10)
=: Ii (P,
d Pi Pj Pi
j6=i
provided that | P i (P)
| < 1 In the notation of the update algorithm, Pei (n) is the random
i
i P i process representing ith users power level and i (n) is the
If P + |
j6=i Pj | < 0, then the contraction property
i
i
error random process. i (n) are assumed i.i.d uniform in the
is guaranteed. We use Lemma 1 to bound | P j
| for j 6= i interval [1 , 1 + ] for all n and i. We will follow a similar
and eliminate common terms. The following conditions are analysis to the one presented for synchronous updates. In
sufficient for < 1: this case, the norm is defined as kxk , maxi E(|xi |) where
X 3
mini Gii Pmin
i x is the vector of random variables and E(.) is the expected
Pmax Gji < Hmin and <
2SIRobj Hmax value. Using the bounds obtained for deterministic Pi and i
j6=i
Note that we can obtain a bound on the stepsize by values, we have the following:  
 e 
considering the boundary conditions in assumption 1. We E(Ii (P, i ) Pi )
state our findings in the following theorem.
   
 
i e
X  
Theorem 4: The synchronous power update algorithm in E 1 + i |Pi Pi | + i  i  |Pej Pj |
Eq. (8) converges to the unique NE P if Pi   Pj 
j6=i
i 3
Hmin mini Gii Pmin
i
> 1 i and < 2 Assume
  that (1 + )Bi1 < 1 where Bi1 is a lower bound
Imax 2 maxi Gii Pmax  i 
on  P i
 in [Pmin , Pmax ]. Independence of i (n) for all n
B. Extension for Asynchronous Updates
The convergence conditions of Bertsekas and Tsitsik- implies independence of Pei (n) and i (n) for all n, which in
lis [16] for totally asynchronous updates, stated in the next turn follows:
theorem, will be the main guide to extend the convergence E(|Ii (P, i ) Pi |)


result of previous section for asynchronous updates. 1 + E(i )(Bi1 + (K 1)Bi2 ) kP P k
Theorem 5: Totally asynchronous fixed point iterations  
 i 
PW (t) (t + 1) = I(P(t)) where Bi2 is an upper bound on  P j
 j 6= i. In order
where W (t) U is an arbitrary subset of user indices to satisfy contraction property, Bi + (K 1)Bi2 < 0 is
1

that update their power at time t, converge to the unique sufficient.

fixed point of the synchronous iterations if the following two One can show that the bounds can be selected as
conditions hold 2SIRobj Hmini
2SIRobj maxj Gji
K Bi1 = 3
and B 2
i = 2
If the sequence of sets Z(k) [Pmin , Pmax ] such Gii Pmax Gii Pmin
that {Z(k) I(Z(k 1))} satisfy ... Z(k) Z(k Following similar steps to derive a sufficient condition as in
1)... Z(1) = [Pmin , Pmax ]K . the case for synchronous updates, and eliminating common
If there exist bounded sets Qi (k) such that the set terms in Bi1 and Bi2 , a sufficient condition for contraction
i
Z(k) can be expressed as P3 Hmin
property is: Pmax
2 < (K1) max j Gji
i
min
Z(k) = Q1 (k) Q2 (k) ... QK (k)
Random power updates defined in Eq. (10) converge to
In Theorem 5, the first condition is called Synchronous
P almost surely under derived conditions. In order to show
Convergence Condition and the second one is called Box
this, we will make use of a similar technique as in [7]. Using
Condition. Since we made use of contraction property in max
Markov inequality first and then definition of max norm, we
norm, asynchronous version of power updates in Eq. (8) also
have:
converges to unique inner NE, P . X X
E(|Pk (t) Pk |)
Let (t) := kI(P(t)) P k for t 1. Define Si (t) = P r(|Pk (t) Pk | > )

[Pi (t), Pi + (t)]. Then S(t) = S1 (t) ... SK (t)
t=1 t=1

satisfy the synchronous convergence condition by definition 1X
of max norm and due to contraction property. Hence, under kP(t) P k
t=1
the same conditions given in Th. 4, totally asynchronous
version of gradient based power updates in Eq. (8) converge 1X t
kP(0) P k
to the unique inner NE by Theorem 5. t=1
C. Error Sensitivity and Conditions for Almost Sure Conver- kP(0) P k

gence (1 )
In the above analysis, we assumed that users have perfect
information before processing and they can calculate the for arbirtary > 0 and kP(0) P k is a non-negative
derivative perfectly. But, errors are involved before and constant. P r(A) represents the probability of event A and
during processing in a real life application of the algorithm. the one before the last inequality follows from contraction
In order to illustrate how the algorithm would operate in a property (that is guaranteed by the conditions). The sequence
 K
X XK X
of partial sums is bounded above and thus convergent. From . SIRobj 2 SIRobj
V (P) = 2 i i + 2 ij i j
Borel-Cantelli lemma, the random update scheme in Eq. (10)
i=1
Gii Pi2 i=1
Gii Pi2
j6=i
converges a.s. to the unique NE. We collect overall results where
of the analysis in the following theorem: 2f i (Ii )Ii df i (Ii )
i := d and ij := Gij [Ii d + fdi (Ii )]
Theorem 6: The stochastic power updates in Eq. (10) con- Pi dIi
verges to the unique NE P if the following two conditions Using the identities ij |ij | and 2i + 2j 2i j 0,
hold i: we reach a condition !
XK X K
3
Pmax i
Hmin 3
mini Gii Pmin . SIRobj
2 < , < 2 V (P) 2
(2 min k +K max |ij |) 2i
Pmin (K 1) maxj Gji 2(1 + ) maxi Gii Pmax m=1
Gmm Pm k ij
i=1
Theorem 7: A sufficent condition for global asymptotic
VI. POWER UPDATES AS A CONTINUOUS TIME
stability of the system in Eq. (11) is
DYNAMICAL SYSTEM
Hmin K
Consider now the iterative power updates of previous >
Pmax maxi,j6=i Gji 2
section with small slot duration and step size tending to
Note that the interference characteristic plays a key role in the
zero. In this case, we obtain continuous time counterpart of
stability condition of the system. Besides, it is inconclusive
unilateral utility maximization of users in Eq. (7). Mathemat-
about the stability of the system if the condition in Th. 7 is
ically, we now have a dynamical system that is governed by
not satisfied.
the following differential equations: VII. NUMERICAL STUDY
. N Ui A. Selection of Objective Reduction Factor
Pi = := i (P) i (11)
Pi
The functional shape of objective reduction factor fdi
We will investigate the stability of the system in Eq. reflects users attitude towards utilizing network resources.
(11). In particular, the notion of stability will be global If the application running on the user requires constant
asymptotic stability in the sense of Lyapunov [17]. NE is performance throughout the operation, then fdi is set to 1
the candidate to be a stable point of the system as it is a for all interference levels, and this reflects a rather non-
solution of the nonlinear system N Ui
Pi = 0 i (as an inner cooperative behaviour. On the other hand, the user behaves
NE is guaranteed). We will find sufficient condition(s) on the in cooperation with other users of the network if it decreases
stability of NE in defined dynamical system . its objective as the interference blows up.
It should be noted that i (P) is differentiable almost Various user behaviour can be addressed by defining
everywhere and thus it is Lipschitz continuous [17] although different fdi . In this paper, fdi has a specific generic form
it is not differentiable at a set of points where (fdi ) (.) with piecewise definitions in three regions. In particular,
is possibly discontinuous. We assume (fdi ) (.) is not con- users are allowed to be non-cooperative up to a certain
tinuous at finite number of points {dji } for all i. As the interference level Li1 . After this level, users cooperate by
surfaces Ii = dji form a set of Lebesgue measure zero in means of decreasing their objectives. The rate of decrease
[Pmin , Pmax ]K , i (P) is differentiable almost everywhere. is bound to satisfy absolute subhomogeneity of [xfdi (x)]1/2 .
Hence, it is legitimate to apply Lyapunov Theory. An additional constraint comes from boundary conditions and
Boundary conditions in assumption 1 immediately guar- i
that the condition on Hmax turns to a condition on the non-
antee that the power values Pi (t) evolve in such a way 2
Gii Pmax
that they bounce back to inside of [Pmin , Pmax ] whenever cooperative interference region: Li1 < SIR obj .

Pi (t) = Pmin or Pi (t) = Pmax . Hence, the power levels of User i reacts to interference level higher than Li1 by
users always reside inside the set [Pmin , Pmax ]. decreasing its objective. The decreasing region of fdi is
A candidate Lyapunov function V : K is to be constrained by absolute subhomogeneity assumption of
guessed now in order to establish sufficent conditions for [xfdi (x)]1/2 . An optimal way of cooperation under AS as-
stability of the system in Eq. (11): sumption can be described using least AS upper bound [9]
XK of the reaction that decreases to zero sharply at interference
V (P) = 2i (P) (12) level Li . One can show that the decreasing region of least AS
i=1 upper bound of sharp decrease is proportional to 1/x3 . The
Note that V (P) is actually defined in [Pmin , Pmax ]M . Since
proportionality constant is determined such that the function
NE is inner and unique, i (P) = 0 i if and only if
is continuous.
P = P . Hence,. V (P) is positive for all P other than NE.
The condition V (P) < 0 is sufficent for stability of NE, if A further constraint on the decreasing region is proposed
satisfied at points other than NE and if the derivative is well as each user must stay in the system. Users reaction to
defined3 . interference is imposed to be higher than its reaction to the
i i
. minimum interference it experiences: Hmin = Imin .
We begin the analysis by calculating V (P) and then 1
bounding it The region decreasing with proportion to x3 stops when
the reaction level reaches H i (Ii ) = Imin . The interference
3 If derivative is not well defined, analysis can be extended using subdif- level Ii at which H i (Ii ) = Imin is Li2 . fdi decreases
ferentials proportional to 1/x and H i (Ii ) = Imin for all Ii > L2i .
 i
The general form of fd is as follows: The threshold SINR is set to th = 9 dB for all experi-
i

1, Imin < x < Li1 ments. The particular value of the threshold is chosen since
i
L it guarantees a certain communication quality5 . Users goals
fdi (x) = ( x1 )3 , Li1 < x < Li2 (13)

i1 )3 1 , Li < x < I i .
(L are to stay below an outage probability of Oobj = 10%,
(Li2 )2 x 2 max
which corresponds to SIRobj = 18.7 dB or SIRobj = 75 in
We allow the non-cooperative region to be as large as possi- normal scale.
2
Gii Pmax
ble, hence Li1 = SIR i
obj . Note that L1 are set proportional In the simulated system, suppression level of users are
to the direct channel gain Gii . This way, users having around Si = 15 dB. Note that the exact values of channel
higher direct channel gains favor the network resources more gains and noise power are not important in terms of operation.
and thus a network opportunistic mechanism is established Gii takes arbitrary value in the interval [1, 2] for all i and
and throughput efficiency of the network is supported. For exact values of Gij are chosen randomly (say distributed
Li uniformly in a relatively small interval than its mean value) so
continuity, Li2 = (I i 1)1/3 . Introduction of the third region
min
in definition of fdi brings a notion of fairness to the control that around 15 dB suppression is maintained. Throughout the
algorithm as it maintains users remaining in the system. simulations, we set maximum transmit power as Pmax = 1
Hence, by defining fdi as in (13), users react to interference and noise power as 2 = 0.001. P
ForPthe simulated setting, maxi j6=i Gji = 0.034,
with a balance between throughput efficiency and fairness.
There may be exceptional cases that require modification mini j6=i Gji = 0.024, maxij Gji = 0.01, maxi Gii = 1.5
in the general form of fdi given above. If Imin i
> Li1 , then and mini Gii = 1. Interference suppression capability Si
i 1
fd would have only the x decreasing region and the reaction ranges from 30 to 60. Perron-Frobenius eigenvalue of the
would be a constant for all interference levels. In this extreme system is P F = 0.0227, which implies that SIRobj = 75
case, though user i transmits with Pmin + , it is not allowed is not feasible. Actually, users can achieve a common SIR
to control its power and is in a sense discarded from the level of at most P1F = 43 in normal scale.
i In the first experiment, we investigate the NE by means of
system4 . Another exception occurs if Imin < Li1 but Imax
i
<
convergence of iterating reaction curves. Transmitters have
Li2 . In this case, x1 decreasing region does not appear. Note
i i instantaneous access to the perfect information of interfer-
that the imposed condition Hmin = Imin is satisfied in both
ence level at the intended receiver. Minimum power is taken
cases. Possible reaction curves in several different cases are
as Pmin = 0.3. Evolutions of power level of transmitters and
depicted in Fig. 3.
Pmax Pmax
achieved SIR level in their corresponding receivers are shown
in Fig. 4. Convergence is observed in about 5 iterations.
1 80
NE
r2 (P1 ), P2
r2 (P1 ), P2

NE 0.9 70

0.8 60
Power

0.7 50
SIR

0.6 40

Pmin r1 (P2 ), P1 Pmax Pmin r1 (P2 ), P1 Pmax 0.5 30

0.4 20

Pmax Pmax
0.3 10
2 4 6 8 10 12 14 2 4 6 8 10 12 14
Iterations Iterations

Fig. 4. Evolutions of power and SIR are illustrarted for Pmin = 0.3.
r2 (P1 ), P2
r2 (P1 ), P2

Users have perfect knowledge of congestion level at their intended receiver

NE and synchronously react by changing their power levels.
NE

In the next experiment, synchronous iterative power update

algorithm in Eq. (8) is simulated. Choices of Pmin and step
Pmin r1 (P2 ), P1 Pmax Pmin r1 (P2 ), P1 Pmax size determine the convergence of the update algorithm.
Fig. 3. Possible Reaction Curves Minimum power is taken as Pmin = 0.5. The step size is
B. Simulation Setting and Results chosen as = 0.05 so that (along with the chosen value for
We will simulate an ad-hoc network with 5 transmitter- Pmin ) the synchronous convergence conditions given in Th.
receiver pairs. Direct and cross channel gains vary with 4 are satisfied. Evolutions of power and SIR are illustrated
Rayleigh fading. As the system model can represent a wide in Fig. 5. Convergence is observed in about 20 steps.
range of communication scenarios, we do not specify phys- Now, we change the minimum power to Pmin = 0.3 and
ical models about link gains and multiple access schemes = 0.01, in which case Pmin violates the convergence
used. Keeping the generality of the model, the average condition in Th. 4. However, the update algorithm is observed
channel gains Gij are determined arbitrarily with Gii Gij to converge in Fig. 6. Convergence is relatively slow in this
for all i. case due to smaller step size. Note that the equilibrium points
4 Note that 0 < P 5 9 dB threshold ensures BER less than 106 for QPSK modulation and
min . This makes sure that the equilibrium does
not lie on the boundary BER is in acceptable levels for higher order modulations. See [11].
 1 80
objective reduction factor. This is inherently an egoistic
0.95

0.9
70 cooperation mechanism which serves to improve overall
0.85 60 network performance. Each users reaction to interference is
0.8
assumed to be absolutely subhomogeneous so that a unique
Power

50

SIR
0.75

0.7
40 Nash equilibrium of the game is guaranteed. Iterative power
0.65 30
update algorithms based on gradient optimization techniques
0.6
20
have been investigated and convergence conditions have been
0.55

0.5 10
derived. Finally, power updates are considered as the result
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100

Iterations Iterations of a continuous time dynamic system, which is a limiting

Fig. 5. Evolutions of Power and SIR in synchronous update algorithm of
Eq. (8) with Pmin = 0.5 and = 0.05.
in Fig. 4 and Fig. 6 are exactly same because the gradient
based algorithm converges to the NE of the game. It can
be concluded that the convergence results in Th. 4 are not
necessary conditions though they provide a useful guideline
for selection of the parameters. Moreover, it turns out that the
conditions obtained for step size are necessary for invariance
of iterative power update algorithms.
1 80
0.9
0.8
0.7
Power
case for small step size. Sufficient conditions for stability of
the system have been established. Numerical results verify
the sufficency of the convergence conditions of iterative
algorithms, while cases are shown exhibiting these conditions
are not necessary for convergence.
R EFERENCES
[1] D. Falamori, N. Mandayam, and D. Goodman, A New Framework for
Power Control in Wireless Data Networks: Games, Utility, and Pricing,
vol. Volume 524. Online: Wireless Multimedia Network Technologies,
Springer US, 1999.
[2] D. J. Goodman and N. B. Mandayam, Power control for wireless
70
data, IEEE Personal Communications Magazine, vol. 7, pp. 4854,
60 April 2000.
[3] T. Alpcan, T. Basar, R. Srikant, and E. Altman, Power control in

50
SIR

0.6
CDMA as a non-cooperative game, Kluwer Academic Publishers,
40 Wireless Networks, vol. 8, pp. 659670, 2002.
0.5

30
[4] C. U. Saraydar, N. B. Mandayan, and D. J. Goodman, Efficient power
0.4
control via pricing in wireless data networks, IEEE Transactions on
0.3 20 Communications, vol. 50, pp. 291303, February 2002.
0.2 10
[5] J. Huang, R. Berry, and M. Honig, Distributed interference com-
0 50 100 150 200 250 300 0 50 100 150 200 250 300

Iterations
Fig. 6. Evolutions of Power and SIR in synchronous update algorithm of
Eq. (8) with Pmin = 0.3 and = 0.01.
Finally, we address error sensitivity of the synchronous
power update scheme for the same wireless environment.
Pmin = 0.5, = 0.05 and = 0.5 are the parameters
taken for the simulation. Note that the step size satisfies
the convergence condition in Th. 6, but Pmin does not satisfy
the respective sufficient condition. Yet, in this case too, the
erronous updates converge to the same point as the error free
updates, as can be observed by comparing Fig.s 5 and 7.
1 80
0.95
0.9
0.85
SIR
Iterations pensation for wireless networks, IEEE Journal on Selected Areas in
Communications, vol. 24, pp. 10741084, May 2006.
[6] M. Xiao, N. B. Shroff, and E. K. P. Chong, A utility-based power
control scheme in cellular wireless systems, IEEE/ACM Transactions
on Networking, vol. 11, pp. 210221, April 2003.
[7] T. Alpcan, T. Basar, and S. Dey, A power control game based
on outage probabilites for multicell wireless data networks, IEEE
Transactions on Wireless Communications, vol. 5, pp. 890899, April
2006.
[8] O. Ozel and E. Uysal-Biyikoglu, Distributed power control using non-
monotonic reaction curves, in Proc. 2009 Intl. Conf. on Game Theory
for Networks, pp. 6067, May 2009.
[9] C. J. Nuzman, Contraction approach to power control with non-
monotonic applications, in Proc. IEEE Global Telecommunications
Conference (Globecom), November 2007.
[10] S. Kandukuri and S. Boyd, Optimal power control in interference lim-
ited fading wireless channels with outage probability specifications,
70
vol. 1, pp. 4655, January 2002.
[11] K. A. Hamid, On the reliability of outage models, in Proc. 2004
60
IEEE 60th Vehicular Technology Conference, vol. 6, pp. 43494351,
September 2004.

0.8
Power

50
0.75 [12] J. Papandriopoulos, J. Evans, and S. Dey, Optimal power control in
40
0.7 CDMA networks with constraints on outage probability, in in Proc.
0.65 30 WiOpt03, INRIA Sophia Antipolis, pp. 279284, 2003.
0.6 [13] J. Papandriopoulos, J. Evans, and S. Dey, Optimal power control
20
0.55 for rayleigh-faded multiuser systems with outage constraints, IEEE
0.5
10 20 30 40 50 60 70 80 90 100
10
10 20 30 40 50 60 70 80 90 100
Transactions on Wireless Communications, vol. 4, pp. 27052715,
Iterations Iterations November 2005.
Fig. 7. Evolutions of Power and SIR in synchronous update algorithm of [14] S. Dey and J. Evans, Optimal power control in wireless data net-
Eq. (10) with erroneous feedback information. Pmin = 0.5, = 0.5 and works with outage-based utility guarantees, in Proc. 2003 42nd IEEE
= 0.05. Conference on Decision and Control, vol. 1, pp. 570575, December
VIII. CONCLUSION 2003.
[15] T. Basar and G. J. Olsder, Dynamic Noncooperative Game Theory.
This paper presented a game theoretic analysis of utility Philedelphia, PA: SIAM, 1999.
based distributed power control with smooth reactions to in- [16] D. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Compu-
tation: Numerical Methods. Upper Saddle River, NJ: Prentice Hall,
terference. The utility function has been defined as a measure 1989.
of relative performance with respect to rate objectives. Users [17] H. K. Khalil, Non-linear Systems. Upper Saddle River, NJ: Prentice
decrease their rate objectives as a reaction to interference, Hall, 1996.
exhibited in the definition of net utility by means of an