Robust Distributed Power Control in Cognitive

Radio Networks

arXiv:1105.2989v2 [cs.IT] 17 May 2011

Saeedeh Parsaeefard, Student Member, IEEE and Ahmad R. Sharafat, Senior Member, IEEE

Abstract
We propose a robust distributed uplink power allocation algorithm for underlay cognitive radio
networks (CRNs) with a view to maximizing the total utility of secondary users (SUs) when channel
gains from SUs to primary base stations, and interference caused by primary users (PUs) to the SUs base
station are uncertain. In doing so, we utilize the worst-case robust optimization to keep the interference
caused by SUs to each primary base station below a given threshold, and satisfy the SUs quality of
service for all realizations of uncertainty. We model each uncertain parameter by a bounded distance
between its estimated and exact values, and formulate the robust power allocation problem via protection
values for constraints. We demonstrate that the convexity of our problem is preserved, and in some cases
converts into a geometric programming problem, which we solve via a distributed algorithm by using
Lagrange dual decomposition. To reduce the cost of robustness, defined as the reduction in the total
utility of SUs and the increase in message passing, we utilize the D-norm approach to trade-off between
robustness and optimality, and propose a distributed power allocation algorithm with infrequent message
passing. Simulation results validate the effectiveness of our proposed approach.

Index Terms
Cognitive radio network, D-norm approach, geometric programming, infrequent message passing,
price of robustness, robust power allocation, uncertain channel gain, underlay spectrum sharing.

Manuscript received June 01, 2010, revised May 10, 2011. This work was supported in part by Tarbiat Modares University,
and in part by Iran Telecommunication Research Center, Tehran, Iran under PhD Research Grant 88-11-124.
The authors are with the Department of Electrical and Computer Engineering, Tarbiat Modares University, P. O. Box 141554838, Tehran, Iran. Corresponding author is A. R. Sharafat (e-mail: sharafat@modares.ac.ir)

Robust Distributed Power Control in Cognitive

Radio Networks
I. I NTRODUCTION
A. Motivation
Cognitive radio network (CRN) is a promising multi-user wireless communication system for improving
spectrum utilization. In a CRN, a cognitive (secondary/unlicensed) user (SU) can access the frequency
spectrum licensed to primary users (PUs) so long as its transmissions do not cause harmful interference to
PUs. For this opportunistic and non-cooperative access to spectrum, two paradigms have been introduced
[1], [2], namely, the interweave and the underlay. In the interweave paradigm, SUs utilize the vacant spots
in the frequency spectrum to the extent that no PU is prevented from using the spectrum; and in the
underlay paradigm, simultaneous utilization of frequency spectrum by PUs and SUs is possible provided
that the SUs interference to PUs base stations (PBSs) are kept below a permissible level (interference
threshold). In this paper, we focus on the underlay approach due to its ease of implementation and its
efficiency in utilizing the frequency spectrum.
An important task in the implementation of underlay multi-user CRNs is to dynamically allocate the
transmit power to SUs in such a way that the interference threshold at each PBS is adhered to, and the
total utility of SUs is maximized subject to maintaining a predefined quality of service (QoS) for SUs.
To achieve these two objectives, one needs to have the values of channel gains between SUs and PBSs
and the interference caused by PUs on the SUs base station (SBS).
However, given the random and sometimes erratic nature of many wireless channels, it is unlikely to
have exact values of channel gains. This is aggravated by the fact that PBSs do not have any obligation
to provide any information to SUs on the channels between SUs and PBSs. This type of uncertainty
may cause the total aggregated interference of SUs at PBSs to exceed the acceptable threshold, which
would increase the outage probability form PUs point of view, or increase the probability of violating the
interference threshold from SUs perspective [3]. In this paper, since we formulate the power allocation
problem form SUs perspective, we also use the notion of violation probability as in [3].
Admission control for PUs is completely different from that of SUs, and SUs transmissions are not
coordinated with PUs. These two issues cause uncertainties on interference levels on SBSs made by PUs
transmitters, which in turn, reduce the actual signal-to-interference-plus-noise ratio (SINR) of each SU at

SBSs below an acceptable threshold, i.e., increases the outage probability of SUs. Hence, transmissions
would become unreliable form PUs and SUs points of view. We wish to obtain the SUs optimal
power levels in an uncertain environment as stated above with a view to simultaneously maintaining the
interference threshold of PBSs, and satisfying the SUs SINR threshold.
B. Related Works
While there has been a significant amount of research on distributed power allocation in cellular
networks in the past two decades [4][6], considering uncertainty in system parameters and devising
robust distributed power allocation schemes has been a relatively under-explored area of research. This
problem was first studied in [7] to devise a scheme for preventing new users to deny service (causing
outage) to existing users; and was later extended in [8] for trading off between power consumption and
robustness. Subsequently, new paradigms were introduced in [9], [10] to deal with different sources of
uncertainty in the system.
Besides, optimal power allocation in CRN is a challenging research area in wireless networks, and
has been considered in the literature from different perspectives. As an instance, for multiple SUs in the
underlay paradigm and assuming that the channel information is available with certainty, power allocation
algorithms that guarantee PUs QoS are introduced in [11][13]. In contrast, by assuming the stochastic
nature of fading channels, the throughput of one SU transmitter-receiver pair is optimized in the presence
of one PU receiver in [14], [15] by considering the peak and average interference thresholds. The same
problem is solved in [3] where data rates and transmit power levels for multiple SUs are obtained via
a centralized approach when each channels gain is considered as the average gain multiplied by the
short-term fading factor.
In [16], to access the spectrum, multiple SUs should keep the outage probability of the PU bellow
a given threshold, where the channel is Rayleigh fading, and PUs listen to the PUs feedback channel
to obtain the outage probability of PU. By assuming no interference between SUs, they proposed a
distributed power control algorithm for SUs. For the interweave paradigm, optimal power levels for SUs
subject to an upper-bound on PUs outage probability and predefined probabilities of false-alarm and
miss-detection are obtained in [17][20].
C. Our Contribution in This Paper
In this paper, we use the robust optimization theory to address the power allocation problem in the
underlay paradigm by searching for a solution that satisfies the constraints and remains near-optimal

[9], [10], [21], [22] when optimization parameters of our problem are perturbed. Based on the available
information on uncertainty, two approaches for robustness exist: A stochastic approach that assumes
statistical knowledge of uncertainties, where constraints are statistically satisfied; and the worst-case
approach that assumes uncertainty in any parameter value is bounded in an uncertainty set, and constraints
are satisfied for all realizations of errors within that set.
Since the main concern of SUs in the underlay paradigm is to keep the interference to PBSs below a
predefined level for all realizations of uncertainty in parameter values, we utilize the worst-case approach,
which results in the following two major contributions in this paper: We simplify the robust power
allocation problem in such a way that its convexity is preserved and can be solved as the non-robust power
allocation in a distributed, simple, and efficient manner; and demonstrate that introducing robustness to
simultaneously deal with such uncertainties affects the SUs allocated power levels in opposite directions.
In what follows, we describe the above contributions in more detail.
To preserve the convexity of the robust power allocation problem, we consider each uncertainty as the
distance between the exact and the estimated values via the general norm. To reduce the computational
complexity of the problem, we apply the protection values instead of the uncertainty set in the constrains.
This would convert the robust problem into a geometric programming (GP) optimization problem, where
we use the linear norm, and utilize the Lagrange dual decomposition to solve the problem in a distributed
manner. However, this approach needs more message passing as compared to the nominal (no uncertainty)
power allocation problem, which can be considered as the cost of introducing robustness in an uncertain
environment. To reduce this cost, we propose a distributed scheme with infrequent message passing, and
prove its convergence to the optimal point when the maximum delay between two consecutive message
passings is upper bounded.
As stated earlier, there are two sources of uncertainty in the system, namely, uncertainty in the measured
interference level at the SBS, and uncertainty in channel gains between SUs and PBSs. Introducing
robustness to simultaneously deal with both of the above sources of uncertainty affects the allocated
power levels in opposite directions, i.e., the first one needs more transmit power, and the second one
requires less. To get an insight into the nature of this problem, we compare the feasibility conditions for
the robust optimization problem with those of the nominal (no uncertainty) optimization problem, as well
as their respective values of total utility; and demonstrate that uncertainty would shrink the feasibility
region, and consequently, reduces the SUs total utility, which is another cost of introducing robustness.
We also propose a trade-off mechanism between introducing robustness to PUs and maximizing the SUs
throughput by adjusting the protection values using the D-norm approach. To measure the effectiveness of

our trade-off scheme, we compare the conditions for feasibility of robust power allocation in the D-norm
approach with those where uncertainty is represented by an ellipsoid.
Compared to pervious works on power allocation in CRNs, we develop the worst-case robust approach
to simultaneously guarantee the QoS of both PUs and SUs in multi-user CRNs. This has not been
considered in the existing literature. Our approach does not assume any cooperation between SUs and
PUs, and does not need the statistics of uncertainties. In doing so, we use the general norm that covers
all sources of uncertainty in the system. In this way, the problem can be solved via GP, and can be
implemented easily and effectively in a distributed manner, as compared to the stochastic or Bayesian
approaches.
The rest of this paper is organized as follows. Section 2 contains the system model and formulation of
the robust power allocation problem for the worst-case approach, its reformulation by using protection
values, and derivation of its feasibility condition. The proposed distributed robust power allocation
algorithm that utilizes Lagrange dual decomposition is presented in Section 3. The infrequent-message
passing distributed algorithm for obtaining robust transmit power levels is introduced in Section 4,
followed by obtaining the cost of robustness, and a trade-off mechanism for its reduction in Section
5. Sections 6 and 7 contain simulation results and conclusions, respectively.
II. S YSTEM M ODEL

AND

P ROBLEM F ORMULATION

Consider an uplink of a cellular CRN in which the coverage area of a cognitive cell partially overlaps
with coverage areas of neighboring primary cells, sharing the same frequency band (e.g., CDMA). Each
user communicates with its base station as shown in Fig. 1. The set of PBSs is denoted by P =
{1, , P }, and the set of SUs serviced by the SBS is S = {1, , S}. Cooperation between the SBS

and PBSs is not assumed, but the SBS knows the predefined interference threshold of PBSs. Time is
divided into different frames whose durations for PUs and SUs are TP and TS , respectively. We denote
the direct channel gain between the ith SU and its SBS by hi , and the interference channel gain between
the ith SU and the j th PBS by gij .
For each frame, there are three time slots denoted by TP = [t1P , t2P , t3P ], and TS = [t1S , t2S , t3S ].
In the first time slot, the SBS and PBSs estimate interference levels, noise power, and hi for each user;
also the channel gains between each SU and PBSs (i.e., gij ) are estimated by that SU and sent to the
SBS1 . This information is utilized in the second time slot when the transmit power levels of SUs are
1

To satisfy the interference threshold constraint, i.e., C2 in (2), during t1S , SUs are polled by SBS and each SU transmits

the value of gij at power level pi = min{pmax

i , minjP

ITj
gij

}.

Fig. 1.

A cellular CRN with primary and secondary users

calculated in a distributed manner, and the transmit power levels of PUs are calculated in a distributed
manner by PUs or centrally by PBSs. In the third time slot, data is transmitted at the calculated power
levels. No coordination is assumed between the SBS and PBSs; and time slots for SUs and PUs do not
need to be identical, but the first two time slots are assumed to be much shorter than the third one, as
the former ones are used for signalling only. The SINR of the ith SU at its SBS is
i =

pi hi
IP +

k6=i,kS

pk hk + 2

(1)

max
th
where pi [pmin
i , pi ] is the transmit power of the i SU, IP is the aggregate interference of PUs to the
P
SBS, j6=i,kS pk hk is the interference caused by other SUs to their BS, and 2 is the channel noise

power. The objective is to adjust the transmit power levels of SUs in such a way that the following two
goals are simultaneously satisfied:

1) The total utility of SUs is maximized while their required SINRs are maintained above a given
value i , and
2) Interference to PBSs is maintained below a given threshold ITj .
In doing so, the transmit power levels of SUs can be obtained by solving the following optimization
problem
max

max

pmin
i pi pi

C :
1
subject to
C :
2

ui (i ),

(2)

iS

i i
P
iS gij pi ITj

i S,
j P,

where the utility function ui (i ) is the SINR of each SU at the SBS, C1 is on the desired SINRs
of SUs, and C2 is on the interference threshold at the j th PBS. In general, (2) is non-convex in pi .

However, if the utility i is concave, strictly increasing, twice continuously differentiable over (0, ),

and i ui (i )/ui (i ) = 1, the problem (2) can be transformed to a convex optimization problem via GP
and logarithmic transformations ([23], Chapter 5 in [24]). To satisfy the above conditions for convexity,
let the utility function be
ui = wi ln i

(3)

where wi > 0 is a per user coefficient. For a valid value of utility in (3), the minimum transmit power
> 0, which can be satisfied by a very small value for pmin
of SUs should be pmin
i , i.e., by no effective
i

transmission by SUs. We now utilize GP and an auxiliary variable qi associated with the ith SU, and
rewrite (2) as
max

max
pmin
i pi pi

1
subject to
C2

ui (hi pi qi1 ),

(4)

iS

: qi i hi pi
P
: iS gij pi ITj j P,
P
: (IP + k6=i,kS pk hk + 2 ) qi

By applying the logarithmic transformation pi = eyi and qi = ezi , the problem (4) changes to a convex
optimization problem [23], [25].

A. Robust Counterpart of Power Control Problem

In our system model, the uncertain parameters are the channel gains between SUs and PBSs, i.e, gij ,
and the interference levels caused by PUs to the SBS, i.e., IP , which directly affect the linear constraints
C2 and C3 in (4). To deal with such uncertainties, we use the robust optimization method introduced

for affine constraints convex optimization in [21], and consider the uncertainty set for both uncertain
parameters as the distance between their actual (uncertain) and nominal (no uncertainty) values using the
general definition of norm (Section A.1.2 in [26]). In such cases, the uncertainty set for the interference
caused by PUs to the SBS is
RIP = {IP |kIP IP k }

(5)

where IP and IP are the actual (uncertain) and estimated (nominal) interference levels caused by PUs at
the SBS, respectively, is the upper bound on the uncertainty region, and kxk is the general definition
of norm. Similarly, the uncertainty set for the channel gain is
Rg = {gj |kgj gj k j }

j P

(6)

where gj and gj are the actual (uncertain) and estimated (nominal) channel gain vectors between
transmitting SUs and the j th PBS, whose ith elements are gij and gij , respectively, and j is the
upper bound on the uncertainty set. Note that the values of j and , as well as the definition of norm
depend on size of error and sources of uncertainties. Also, for uncertainties caused by stochastic error
in measurements, one can derive the upper bound on uncertainty and its related norm [27], [28] (see
Appendix A).
For the uncertain parameters, the solution to (4) remains robust in the worst-case approach if for any
realization of gi Rg and IP RIP , the optimal solution satisfies C2 and C3 . Hence, this adds another
constraint C4 to the robust counterpart of (4), i.e.,
max

max
pmin
i pi pi

C1

C
2
subject to

C3

C
4

ui (hi pi qi1 ),

(7)

iS

: qi i hi pi
P
: iS gij pi ITj j P,
P
: (IP + k6=i,kS pk hk + 2 ) qi

: gj Rg , IP RIP

i and j,

Recall that the problem (4) is the nominal problem for (7), in which uncertainty is not considered, i.e., the
estimated (nominal) values are considered as the exact values. Furthermore, the impact of the uncertainty
sets in C4 of (7) on its convexity needs to be investigated [21]. To do so, we rewrite (7) by using
protection values.
Statement 1. When Rg and RIP are compact and convex sets, (7) is a convex optimization problem.
P
Proof: The above statement is true because C2 and C3 are satisfied if and only if maxgj Rg iS gij pi
P
P
ITj , and maxIP RIP (IP + k6=i,kS pk hk + 2 ) qi . We rewrite these conditions as ( iS gij pi +
P
P
maxgj Rg iS (gij gij )pi ) ITj , and (IP + maxIP RIP (IP IP ) + k6=i,kS pk hk + 2 ) qi .
Since the max function over a convex set is a convex function (Section 3.2.4 in [26]), the convexity of

C1 and C2 (and consequently (7)), is preserved.

P
The two terms violate
= maxgj Rg iS (gij gij )pi and I = maxIP RIP (IP IP ) are called
j

protection values against variations in channel gains and interference levels, respectively [9], [10], [29].
Using the above, the robust optimization problem can be written as
max

max
pmin
i pi pi

X
iS

ui (hi pi qi1 ),

(8)

C : q hi pi

1 i i
P
subject to
) ITj j P
C2 : ( iS gij pi + violate
j

C : (I +
2
3
P
k6=i pk hk + I + ) qi

Statement 2. When the uncertainty set is stated in terms of the general norm, protection values are
I = kIP k ,

violate
= j kpk ,
j

where p = [p1 , , pS ] and kxk is the dual norm of kxk (Definition 1 in [29]).
Proof: See Appendix B.
Based on Statements 1 and 2, convexity of the robust power allocation problem is preserved. In addition,
pP
xp , the dual norm is the linear
when the uncertainty set is a linear norm with order p, i.e., kxkp = p
norm with order q = 1 +

1
p1 .

Therefore, in this case, the protection value becomes a deterministic

function of optimization variables in (4), i.e., pi , and the non-linear function max in violate
and I is
j
eliminated from C2 and C3 . Hence, the robust power allocation problem is changed to the standard form
of convex optimization, which can be solved very efficiently.
Since channel uncertainties in wireless networks have a stochastic nature, the uncertainty set can be
represented by an ellipsoid, i.e., by the linear norm with p = 2 [9], [30][32], which is commonly used
in CRNs to model uncertainty in channel gains between secondary transmitters and primary receivers
[28], [33], [34]. We do the same in the rest of this paper, and use ellipsoid uncertainty regions to derive
robust formulations and algorithms, resulting in
I = kIP k2 ,

violate
= j kpk2
j

and the problem (8) becomes

max
min

pi pi pmax
i

ui (hi pi qi1 ),

(9)

iS

C : qi i hi pi

1 P
qP
2
subject to C2 : ( iS gij pi + j
iS pi ) ITj j P

C : (I + P
2
3
P
kS,k6=i pk hk + I + ) qi

Note that similar to (4), there are only two optimization variables in (9), i.e., pi and qi , and by logarithmic
transformations pi = eyi and qi = ezi , the convexity of the problem is preserved (see Appendix C).

10

B. Feasibility of Robust Counterpart Problem

Compared to the nominal problem, uncertainties in C2 and C3 in the robust counterpart, i.e., (9),
affect the SUs transmit power levels in opposite directions. On one hand, the SUs transmit power levels
should be decreased to satisfy PUs QoS because of the uncertainty in gj , and on the other, they should
be increased to achieve SUs SIRs due to uncertainty in IP at the SBS. Besides, increasing the SUs
power levels to achieve their SIRs in the presence of uncertainty increases interference. Each of the above
may result in infeasibility of the power allocation problem, meaning that a power allocation vector that
simultaneously satisfies the constraints in (9) may not exist. In what follows, we investigate that under
what conditions, the robust power allocation problem remains feasible when (4) is feasible.
Let = [ h1 n1 0 , , ShSn0 ]T , n0 = IP + 2 , and F be the normalized gain matrix of SUs whose elements
are

0
Fij =
h /h
i j
i

if

i=j

if

i 6= j

and G is the S P channel gain matrix between SUs and PBSs with [G]ij = gij , IT = [IT1 , , ITP ] is
max T
the vector of interference thresholds for PBSs, and pmax = [pmax
1 , , pS ] . The problem (4) is feasible

if the following conditions hold [11]

1) (F) < 1
2) b pmax
3) bT G IT
where is the spectral radius of F, and b = (I F)1 . Now we state the feasibility conditions for (9).
Proposition 1. The robust problem (9) is feasible if
1) (F) < 1
2) b pmax

+ kb k2 IT
3) bT G

is the estimated channel gain matrix between

where b = (I F)1 ( + ), = [ h11 , , hSS ]T , G

SUs and PBSs, and = [1 , , P ].

Proof: See Appendix D.
Increasing the values of and shrinks the feasibility set as compared to that of (4). The effect of
uncertainties in on the feasibility set of (9) is shown in Fig. 2 for a simple CRN that consists of
one PBS and one SU in the cognitive cell when C2 is considered and IP = . In this setup, without
considering C2 and with no uncertainties in , i.e., when = 0 and = 0, the SU reaches its for

11

100%

50%

p / p

max

75%

(9) is infeasible

25%

(9) is feasible
0
0
2%
4%

(=4%, IT/g=9, p=0.633 pmax) 6%

Fig. 2.

8%
10%

10

IT / g
1

Feasibility set of (9) versus and IT .

0.51 pmax . As can be seen in Fig. 2, by increasing IT /g and , the constraint C2 shrinks the feasibility

set of (9), until the feasible power level falls bellow 0.51 pmax , meaning that for those values of IT /g
and that correspond to power levels below 0.51 pmax , there is no power level for the SU to satisfy
both C1 and C2 in (9), i.e., infeasibility of the robust power allocation problem.
On the other hand, by considering without C2 , the required power for reaching is pi = 0.51
(1 + 0.5)pmax , meaning that SU should increase its power to guarantee . Consider the case of = 80%

, IT /g = 9, and = 4% in Fig. 2. To guarantee the SIR, the SU should transmit at p = 0.714 pmax ,
but to satisfy the interference threshold, p should be less than 0.633 pmax . This simple example shows
that when the nominal problem is feasible, the robust power allocation problem may also be feasible if
only one uncertainty is considered, but when both and have non-zero values, they affect the robust
transmit power of the SU in opposite directions, which may lead to infeasibility of robust power allocation
problem.
Note that increasing for any given value of IT /g causes infeasibility of (9), but the value of that
causes infeasibility of the robust power allocation problem, depends on the value of IT /g. In a tight
interference threshold, e.g., when IT /g < 2, the above mentioned value of is smaller as compared to
that of IT /g 2. This is evident from the sensitivity analysis of the optimization problem (Section 5.6
in [26]), and from Lemma 2 later in this paper. Introducing robustness to the power allocation problem
shrinks its feasibility region, which in turn may reduce the total throughput of SUs in (9) as compared
to that of (4). We consider both of the above effects, i.e., shrinking the feasibility region and decreasing
the total utility of SUs, as the cost of robustness form the SUs point of view.

12

III. ROBUST D ISTRIBUTED A LGORITHM

Considering the convexity of (9), its optimal solution can be obtained by the Lagrangian dual function
(Section 5.2 in [26]) as
L(i , i , , ezi , eyi ) =
X
X
ezi i

ui (hi eyi zi ) +
i ( yi 1)
hi e
iS
iS
p
P
2yi
X
gij eyi + j
iS e
+
1)
j (
ITj
jP
X
X
+
i (ezi (IP +
eyk hk + I + 2 ) 1)
kS,k6=i

iS

(10)

where i , j , and i are Lagrange multipliers for C1 , C2 , and C3 in (9), respectively. This Lagrangian
function has three features that are very useful for proposing distributed algorithms: 1) The utility function
of each user only depends on that users primal variables, 2) dual variables can be divided into local dual
variables of each SU, i.e., i and i , and global dual variables of all SUs, i.e., j , and 3) the protection
values have a simple and deterministic form, and by using a broadcasted scaler value, each user can
calculate its own protection value for the related constraints.
Based on the above features, we now utilize the gradient-based algorithm to solve (9) in a distributed
manner. The gradient update formulations of primal and dual variables are
y max

yi (t + 1) = [yi (t) yi L/yi ]yimin

(11)

zi (t + 1) = [zi (t) zi L/zi ]+

(12)

i (t + 1) = [i (t) + i (

ezi yi i
1)]+
hi

i (t + 1) = [i (t)+
i(ezi (IP +

(13)
(14)

jS,j6=i

j (t + 1) = [j (t) + j(

eyj hj + I + 2 ) 1)]+

P
2yi (t)
gij eyi (t) +j
iS e
ITj

1)]+

(15)

13

where yi , zi , i , i , and j are small step sizes for Lagrange multipliers, and x+ = max{0, x}.
The partial derivative of the Lagrange dual function with respect to yi and zi are
ezi i
hi eyi
X
) + eyi
k hk ezk

L/yi = ui (hi eyi zi )hi eyi zi i

+

X j
e2yi
(
gij eyi + j pP
2yi
ITj
iS e

jP

and

(16)

kS,k6=i

ezi i
L/zi = ui (hi eyi zi )hi eyi zi + i yi
hi e
X
zi
yk
i e (IP +
e hk + I + 2 ).

(17)

kS,k6=i

P
Based on (11)-(17), each SU needs the two global scalar variables b1 (t) = iS i hi ezi (t) and b2 (t) =
qP
2yi (t) to update all its primal and local dual variables, provided that it knows its SINR at the
iS e

SBS and receives its j and . The global dual variable j is updated at the SBS, as we need to know
all the values of SUs power levels, channel gains gij , and ITj . The distributed power control algorithm
based on Lagrange dual decomposition is as follows.
Distributed Algorithm 1
Initialization
At t = 0, all primal variables and Lagrange multipliers are set to non-negative feasible values for each
SU and the SBS.
Algorithm at ith SU:
At each iteration t = 1, 2,
1: The SU receives b1 (t), b2 (t) and j (t) from the SBS.
2: Locally calculates its zi (t), yi (t) from (11)-(12), and its Lagrange multipliers from (13)-(14).
3: Transmits i (t)hi ezi (t) and yi (t) to the SBS.
Algorithm at the SBS:
At each iteration t = 1, 2,
1: The values of b1 (t), b2 (t), and j (t) are updated based on the values of SUs parameters, and
broadcasted by the SBS to all SUs.
The above message passing in time slot t2S between SUs and the SBS are carried out in each iteration
using the power level obtained for each SU in pervious iteration. Algorithm 1 is a distributed scheme
for utility-based power control for GP and convex optimization [13], [23], [35].

14

Statement 3: Algorithm 1 converges to the optimal value of (9) when the condition of Proposition 1
holds and the step sizes are sufficiently small.
Proof: See Appendix E, Part 1.
Note that distribution of calculations amongst SUs is a key advantage of the distributed algorithm as
compared to a centralized scheme in which power levels are determined by the SBS and sent to SUs. This
is because in the latter, the SBS needs to calculate Lagrange multipliers i , i , and i at each iteration for
each SU. However, the distributed algorithm needs message passing in each iteration between SUs and
the SBS. If such messages are allowed to be updated at a slower rate without affecting the convergence
and optimality of power allocation, efficiency of the distributed algorithm would be improved.
IV. I NFREQUENT M ESSAGE PASSING
Now we explain how to reduce message passing and guarantee the convergence of the distributed
algorithm to its optimal point. Let D be the time difference between two successive broadcasting times
for b1 (t), b2 (t), and j (t). We propose the following distributed power control algorithm with infrequent
message passing.
Distributed Algorithm 2
Initialization
At t = 0, all primal variables and Lagrange multipliers are set to non-negative feasible values for each
SU and the SBS.
Algorithm at the ith SU:
1: For each value of n = 0, 1, , at consecutive iterations t = nD + for 0 < D , and for
b1 (nD), b2 (nD), and j (nD), each SU locally calculates its zi (t), yi (t), and Lagrange multipliers.

2: At t = D, 2D, , each SU transmits i (t)hi ezi (t) and yi (t) to its SBS.
Algorithm at the SBS:
1: For each value of n = 0, 1, , at consecutive iterations t = nD + for 0 < D , and for
yi (nD) and i (nD)hi ezi (nD) received from all users, the value of j (t) is updated by (15).

2: At t = D, 2D, , the values of b1 (t) and b2 (t) are updated using i (nD)ezi (nD) and yi (nD),
and together with j (t) are broadcasted to SUs.
Note that SUs send the values of i (nD)hi ezi (nD) to the SBS at power levels calculated at n =
0, 1, during t2S . Since SUs use the outdated j , and allocated power is still not converged to its

optimal value, the PUs interference threshold may be violated by SUs. However, since t2S t3S , we

15

do not take the respective violation probability into account. Note also that a higher value for D leads
to less message passing in the system, but may cause non-convergence of the distributed algorithm. As
such, we need to find the maximum value of D for which Algorithm 2 converges to the optimal solution
of (9).
Lemma 1. Let N be the number of total dual and primal variables of the Lagrange dual function (10),
and be the vector of all step sizes of the gradient algorithm in (11)-(15). Algorithm 2 converges to
the optimal value of (9) if
D(

1
1
1)
kk
N 1

(18)

Proof: See Appendix F.

V. C OST

OF

ROBUSTNESS ,

AND

T RADE - OFF

As stated in Proposition 1 and shown in Fig. 2, satisfying the interference threshold of PUs while
considering uncertainties in gj , reduces the feasibility domain of (8) as compared to that of (4) and
forces SUs to reduce their transmit power, which results in reduced SUs total utility when Proposition
1 holds (see Appendix I, Part 2). This is not desirable from SUs point-of-view, and calls for a trade-off
between increasing the SUs total throughput and preserving the interference threshold for all instances of
channel uncertainties 2 . This can be achieved via a suitable scheme for robustness, where the uncertainty
region is chosen in such a way that the probability of violating the interference threshold is kept bellow
a predefined level and the SUs total throughput is kept close to the optimal value of the non-robust case.
The reduction in the total utility of SUs is the cost of robustness, measured by
d = ku u k2

(19)

where u and u are the optimal utility values of (4) and (8), respectively.
Lemma 2. Let and be the optimal values of Lagrange multipliers for C2 and C3 in (4),
respectively. For all values of violate
and I we have
j
d

jP

j violate
+
j

i I

(20)

iS

Proof: See Appendix G.

2

It would be mention that the worst-case robust optimization is conservative approach, because in the real situation of system

operation, uncertainty dose not always happened in the worst situation, which is one of the disadvantageous worst-case approach,
and in this context, moderation of its conservative aspect is one of the important issues [21], [36].

16

By adjusting the size of violate

, one can control d . In the context of the worst-case robust optimization
j
theory, there are different approaches for adjusting the protection values [21], among which, we use the
D-norm approach [36], because of its simplicity and effectiveness.
In the D-norm approach, the uncertainty in each channel gain is |
gij | ij with a symmetric (e.g.,
uniform) distribution, i.e.,
gij [
gij gij , gij + gij ]

i S,

j P

(21)

By choosing a non-negative integer violate |S|, the protection value [36] is

violate
(violate , pi ) =
j
max

ej S,|ej |=violate ,|
gij |ij

gij pi ,

iej

jP

(22)

where ej is the subset of all SUs that affect the protection value, the total number of which is violate .
Note that the value of violate adjusts the protection value, which means a trade-off between robustness
and throughput. If violate = 0, no protection is considered, and increasing violate means more protection.
If we utilize the D-norm approach, the robust counterpart of (4) is similar to (8), except that we use
(22) for the protection value violate
, i.e.,
j
max

max
pmin
i pi pi

ui (hi pi qi1 ),

(23)

iS

C1 : qi i hi pi

C : (P
ij pi + violate
) ITj j P
2
j
iS g
subject to
P
C : (I +
2

3
P

k6=i pk hk + I + ) qi

violate = max
ij pi .
ej S,|ej |=violate |
gij |ij
j
iej g

Since the protection value in the D-norm approach is a special form of the norm function [29], the
convexity of (23) is preserved. The distributed algorithm for this approach is similar to the Lagrange
dual decomposition in Section 3, except that in each iteration, b2 changes to b2 = [b21 , , b2P ] and
the SBS calculates b2j by adding the values of violate that pertain to the largest values of gij pi , and
broadcasts b2 to all SUs. Feasibility of (23) and convergence of its distributed algorithm are guaranteed
when Proposition 2 holds (See Appendix E, Part 2) and the step size is small enough.
Proposition 2. Let ij = j , i S . The robust problem (23) is feasible if
1) (F) < 1
2) b pmax

17

+ max(kb k , kb k1 /) IT
3) bT G

where b = (I F)1 ( + ), = [ h1 , , hS ]T , = [1 , , P ], and k k is a norm.

1

Proof: See Appendix H.

If we use the D-norm approach, the robust solution with infrequent message passing can be obtained
similar to Algorithm 2, and its convergence to the optimal value of (23) is guaranteed provided that
Proposition 2 holds and the upper bound of D in Lemma 1 (See Appendix F) is satisfied.
Now we show that under what conditions, the domain of robust optimization in the D-norm approach
is larger than that of the ellipsoid uncertainty region, which leads to a higher utility for SUs.
Lemma 3. When the uncertainty boundaries in the ellipsoid uncertainty region and the D-norm model
are equal, the optimal utility of the power allocation problem (7) for the D-norm uncertainty model is
greater than or equal to the optimal utility for the ellipsoid uncertainty model if
violate

Proof: See Appendix I.

|S|

(24)

Note that the effect of reducing the protection value from the PUs point of view is that the interference
constraint may not be satisfied for all instances of uncertainties, and the probability of violating the
protection value is
Prob(violation) =
X
X
Prob(
gij pi > ITj |(
gij pi + violate
) ITj )
j
iS

(25)

iS

The upper bound for the violation probability for a symmetric probability distribution [36] is
Prob(violation) (1 (

violate 1
p
)),
|S|

(26)

where is the cumulative Gaussian probability distribution function. By fixing the probability of violating
C2 to jviolate , the lower bound of violate is

p
( |S|1 (1 jviolate ) 1) violate .

(27)

Fig. 3 shows the relationship between jviolate , the number of SUs, and violate . Note that reducing jviolate
results in a higher value for violate , and vice versa. Also note that when the number of SUs increases,
the SBS needs a higher value of violate to preserve the violation probability.
Recall that in contrast to (7), which keeps interference to PBSs below a given threshold in a hard
manner for all instances of uncertainties, by using the D-norm in (23), interference is robustly controlled

18

16
violate

14

<0.001

violate

<0.005

violate<0.01

12

violate

violate

<0.05

violate

10

<0.1

8
6
4
2
0

10

15

20

25

30

Number of SUs

Fig. 3.

The relationship between violate for j = 1, the number of SUs, and violate .

in a soft (probabilistic) manner. By using (27), the SBS can calculate the value of violate for any value
of violation probability for PUs and use it to calculate the protection value. In practice, the value of
jviolate can be assigned by the regulatory agency for trading off between the throughput of SUs and the

violation probability of PUs.

VI. S IMULATION R ESULTS
Now we provide simulation results to get an insight into the performance of our proposed scheme in
different uncertainty sets. We consider a CRN with two partially overlapping cells, one for PUs and one
for SUs. The PBS is located at the center of a circular cell whose radius is 2 Km. The SBS is located
0.5 Km from the PBS and covers a circular cell with a radius of 1 Km. There are three active SUs in
the cognitive cell, deployed at d = [150, 200, 350] m from the SBS and D = [550, 300, 400] m from the
PBS. The power range for each SU is [0.001, 1] Watts. Assuming i = 40 dBm and wi = 1 for each
SU, the AWGN power at the PBS is 2 = 113 dBm, and the estimated interference caused by PUs on

the SBS is IP = 2 . The direct path gain for each user is obtained from hi =
gain from the ith SU to the j th PBS is gij =

B
d4ij ,

1
d4i ,

and the channel

where B is the scattering coefficient equal to 128. In

our simulations, we express and in percentages as =

kg
gk2
k
gk2

and =

kIP IP k2
,
kIP k2

and the total SUs

utility is called the network utility.

First, we show the effect of robustness on the network utility for different values of , , and IT in
Figs. 4 and 5. Note that increasing uncertainty in both gj and IP monotonically decreases the network
utility as we expect form Lemma 2. This is because for an ellipsoid uncertainty region, the upper bound

19

16
IT = 2
IT = 1
IT =0 .5
IT= 0.1

14

Network utility

12
10
8
6
4
2

Fig. 4.

10%

20%

30%

40%

50%
60%
Error Bound ()

70%

80%

90%

100%

90%

100%

Network utility versus in Algorithm 1.

16
14

Network utility

12
10
IT= 2
IT= 1
IT= 0.5
IT= 0.1

8
6
4
2

10%

20%

30%

40%

50%

60%

70%

80%

Error bound ()

Fig. 5.

Network utility versus in Algorithm 1.

of (20) is
d

jP

j j kpk2 +

X
iS

i kIP k2 .

(28)

As can be seen from (28), increasing and monotonically reduces the network utility in robust power
allocation as compared to that of the nominal power allocation, i.e., for = 0 and = 0. Besides, when
IT is small, changing has more influence on the reduction in network utility as shown in Fig. 4 for
IT = 0.1 . This is because for small values of IT , the constraint C2 severely shrinks the feasibility

region of the power allocation problem, and a small increase in shrinks the feasibility region even
further, and may cause infeasibility of the problem. For example, in Fig. 4, the power allocation problem
becomes infeasible for > 70% when IT = 0.1 .
Next, we investigate how the D-norm approach can trade-off between the SUs total utility and
robustness in controlling the interference via violate . In Fig. 6, we see the effects of increasing violate on
the networks utility in different uncertainty regions and for different values of IT . Note that increasing

20

16

IT=
14

violate=1, IT=

12

Network utility

violate=2, IT=
violate=3, IT=

10

violate=1, IT= .1
violate=2, IT= .1

IT= .1

violate=3, IT= .1

Fig. 6.

10%

20%

30%

40%

50%
Error bound ()

60%

70%

80%

90%

100%

Network utility versus violate and .

14

12

10
8
6
4
2

200

400

600

800

1000

1200

1400

1600

1800

2000

Iteration number
(a) Algorithm 1

power of each user

1
0.8
0.6
User 3
User 2
User 1

0.4
0.2
0

200

400

600

800

1000

1200

1400

1600

1800

2000

Iteration number
(b) Algorithm 1

Fig. 7.

Convergence of Algorithm 1, all step sizes are 0.5.

violate reduces the SUs total utility. The slope of utility degradation for higher values of uncertainty

(e.g., = 70%) is more as compared to that for lower values of uncertainty (e.g., = 10%). Also for
low value of IT , decreasing the value of violate results in a higher total throughput of SUs. Again, this
is in line with Lemma 2, i.e., for low values of IT , the value of is large, and hence increasing or
decreasing violate
via increasing or decreasing violate , impacts d more profoundly.
j
Finally, we show the convergence of power levels and j in the proposed distributed Algorithms 1 and
2 for three users in Figs. 7, 8, and 9, respectively, where both and are 0.1%. Note that convergence

21

80

60
40
20
0

200

400

600

800

1
power of each user

1000

1200

1400

1600

1800

2000

Iteration number
(a) Algorithm 2
User 1
User 2
User 3

0.8
0.6
0.4
0.2
0

200

400

600

800

1000

1200

1400

1600

1800

2000

Iteration number
(b) Algorithm 2

Fig. 8.

Non-convergence of Algorithm 2, when all step sizes of the gradient algorithm are 0.5 and D = 20.

of Algorithm 1 is relatively fast. But for Algorithm 2 with a large step size, e.g., 0.5, and a large delay,
e.g., D = 20, the algorithm cannot converge to its optimal value as we see in Fig. 8. By decreasing the
step size to 0.01 and delay to D = 5 in Fig. 9, convergence is guaranteed but it may take an order of
magnitude more iterations to converge as compared to Algorithm 1, which is in line with our analysis
of Algorithm 2 in Lemma 1.
VII. C ONCLUSIONS
We proposed a robust scheme for distributed power control in underlay CRNs that takes into account
the two major sources of uncertainty, namely the channel gains between SUs and PBSs, and variations
in interference caused to the SBS by PUs. Such uncertainties may increase the probability of violating
the interference threshold of PBSs and reduce the actual SINR of SUs below an acceptable level. To
mitigate this, we utilized worst-case optimization theory and the notion of protection value in our robust
distributed power allocation scheme for each source of uncertainty.
To get some insight into the impact of introducing robustness on the SUs allocated power levels
(via protection values) and their achieved utilities, we compared the feasibility set of the nominal power
allocation problem with that of its robust counterpart, and showed that the reduction in the total utility
can be considered as the cost of robustness. To moderate this cost, we proposed to trade-off between

22

20

15

10

200

400

600

800

1000
1200
Iteration number
(a) Algorithm 2

1400

1600

1800

2000

1800

2000

power of each user

1
0.8
User 3
User 2
User 1

0.6
0.4
0.2
0

Fig. 9.

200

400

600

800

1000
1200
Iteration number
(b) Algorithm 2

1400

1600

Convergence of Algorithm 2, when all step sizes of the gradient algorithm are 0.01 and D = 5.

the total utility of SUs and the level of robustness in controlling interference to PBSs. Our proposed
distributed power allocation algorithm is based on Lagrange dual decomposition, and needs additional
message passing to deal with protection values. In order to reduce this added signalling, we developed a
distributed scheme for power allocation with infrequent message passing and obtained the condition for
its convergence to the optimal point.
A PPENDIX A
M ODELING U NCERTAINTY

BY

N ORM F UNCTION

In this appendix, we explain how the norm function can be used to model all sources of uncertainty.
To do so, we show the relationship between the size and the shape of the uncertainty region, and the
stochastic nature of the uncertain parameters by utilizing Section 8.5.5. in [27] and [28]. Let x be the
= xx
, where x
is the nominal (or estimated) value and x
is
uncertain parameter for which we have x

the error. Assume that error has a specific probability distribution function (pdf) f (
x), such as Gaussian.
One can find the uncertainty region in such a way that with probability P , any realization of uncertain
parameter falls in the uncertainty region. The generic formulation of this statement is
Z
)dx P
f (x x
kx
xk

(29)

23

For different types of uncertainty, each with a different pdf, a different uncertainty region can be obtained.
This approach is applied to model the uncertainty caused by Gaussian noise in estimation, uniformly
distributed quantization error, and also for fading channel [27], [28]. The shape, i.e., the type or order
of the norm function, can be obtained based on the source of uncertainty. For example, the spherical
uncertainty region is proposed to model uncertainty due to Gaussian channel, which can be considered
as norm-2. Also, the cubic uncertainty region is considered for quantization error that can be expressed
by k k 3 . As such, considering the general norm covers all types of uncertainty in channel gains and
system parameters in our system model.
A PPENDIX B
P ROOF

OF

S TATEMENT 2

.
Recall that the dual norm of kxk is kxk = maxksk1 sT x. For the uncertainty region in Rg , we use
yj =

j gj
g
j

to rewrite the uncertainty region as

Rg = {yj |k
yj yj k 1}

j P

(30)

We also rewrite the protection value as

max

gj Rg

X
iS

(gij gij )pi = j max

yj Rg

X
iS

(yij yij )pi

(31)

which corresponds to the definition of dual norm. The same is true for the uncertainty region pertaining
to interference of PUs on SBSs [29].
A PPENDIX C
C ONVEXITY

OF

(9)

By using logarithmic transformation of the optimization variable in (9), the constraints C1 , C2 , and
C3 as well as the power limit of each SU change to
1)

yi
1
pmin
i e

2)

1 yi
(pmax
i ) e 1

3)

zi yi
1
i h1
i e
X
ezi (IP +
eyk hk + 2 ) 1

4)

k6=i

Maximum or Chebychev norm of vector w is k w k = max wi [26].

24

The left hand side of the above constraints are the affine nonnegative sum of exponential variables, and
constitute convex functions ( [13] and Section 3.2 in [26]). From the above, the interference function
changes to
5)

yi

gij e + j

iS

sX
iS

e2yi ITj

The left hand side of the above is a vector of exponential variables with p = 2, which is convex (Sec.
3.2.4 in [26]). Note that (3) is a concave function under logarithmic transformation [13], [24], hence (9)
is a convex optimization problem.
A PPENDIX D
P ROOF

OF

P ROPOSITION 1

The problem (7) is feasible if there is at least one power vector that simultaneously satisfies all
constraints. The feasibility of problem (7), which is a GP problem, can be written (Sec. 2.1.3 in [23]) as
min

max
1s,pmin
i pi pi

(32)

C1 : f1i (zi , yi ) = ezi yi hii s

C2 : f2i (zi , yi ) = IT1 j ( iS gij eyi ) s j P,

subject to C3 : f3i (zi , yi ) =

ezi (IP + k6=i,kS eyk hk + 2 ) s

C : g R , I R ,
4
g P
IP
j

where s is the auxiliary variable.

The feasible solution of this problem is equal to s = max{1, maxi {f1i (zi , yi ), f2i (zi , yi ), f3i (zi , yi )}}

for any zi and yi . By solving (32) and obtaining (s , p = [p1 , , pS ]), when s = 1, the set of
posynomial constraints in C1 , C2 and C3 are feasible, and the associated power vector of all SUs, i.e.,
p = [p1 , , pS ] is a feasibility set of (7). Otherwise, the set of posynomial constraints and consequently

(7) are infeasible [23]. Based on the above, to obtain the feasibility conditions of (7), we set s = 1 and
obtain p = [p1 , , pS ]. In doing so, we divide the problem (32) into two subproblems, one for C1 , C3
and IP RIP , and the other for C2 and gj Rg .
Subproblem 1. This problem is
min

max
1s,pmin
i pi pi

(33)

25

C : ezi yi hii s

1
P
subject to
C3 : ezi (IP + k6=i,kS eyk hk + 2 ) s

C :I R ,
4
P
IP

If we use I instead of the uncertainty set in (8) and consider s = 1, this problem becomes a standard
power allocation problem [5] with a protection value. The solution to (33) satisfies
+ (I F)p

(34)

where p is the power vector of all SUs. This inequality is satisfied [5] if
(F) < 1

(35)

where is the spectral radius of F, minimized for b = ( + )(I F)1 . Also the power vector
obtained in (34) should satisfy the maximum power limit of each user via
( + )(I F)1 pmax

(36)

Subproblem 2. For C2 , we obtain the feasibility condition by solving

min

max
s1, pmin
i pi pi

yi
C : 1 (P
2 ITj
iS gij e ) s
subject to
C :g R ,
4
g
j

(37)
j P,

which is equivalent to solving the following linear programming problem

gj p + j max rp IT j
krk1

(38)

where r = (gj gj )/j . Note that the second term in (38) is the dual norm [26]. For (9), the dual norm
of a linear norm with order 2 is a linear norm with order 2, hence
gj p + j kpk2 IT j

(39)

When (35), (36), and (39) hold for b , (9) is feasible. Hence, when the feasibility conditions are satisfied,
the power allocation problem is feasible.

26

A PPENDIX E
P ROOF

OF

S TATEMENT 3

Part 1: We show that a descent-based algorithm for the Lagrange dual function converges to the
optimal point, because the gradient of the Lagrange dual function is Lipschitz continuous [23]. The
function F (x) is Lipschitz continuous if
kF (x) F (y)k2 kx yk2

(40)

when > 0. For Lipschitz continuity, it is enough to prove that the Hessian (with respect to the primal
variables yi , zi ) of the Lagrange dual function (10) evaluated at the optimal Lagrange multipliers is
positive definite for all yi , zi R2S [13]. In our problem, the Hessian is
2 L(i , i , , ezi , eyi ) := H

2
H
H12
yy L y z L
:= 11

2
H21 H22
z y L zz L

(41)

The matrices H11 and H22 are diagonal whose elements are
[H11 ]ii = i

ezi i
+ eyi
hi eyi

k hk ezk +

kS,k6=i

P
X j
2e2yi kS,k6=i e2yk
yi
pP
(
gij e + j
),
3/2
2yi
ITj
iS e

jP

[H22 ]ii = i

X
ezi i
zi
(
I
+
+

e
eyk hk + I + 2 )
P
i
hi eyi
kS,k6=i

both of which are positive. For H12 we have

ezi yi
i hi e i
[H12 ]ij =
ezi eyk h
i

if

i=j

if

k 6= i

The sum of each row of H11 + H12 is a positive value. For H21 we have

ezi yi
if i = j
i hi e i
[H21 ]ij =
,
ezi eyk h
if k 6= i
i
k

The sum of each row of H21 + H22 is positive. Since all rows of H are positive, H is positive definite.
On the other hand, for any matrix X we have kXk2 kXk1 kXk. Hence, we can choose =
P
max i j=1:s hi max hi where hi is the ith row of H, and hi is the ith column of H. Hence, the

Lagrange dual function is Lipschitz continuous, and the distributed Algorithm 1 converges to the optimal
value when Proposition 1 holds.

27

Part 2: All the above can be directly applied for (23), except that H11 changes for the D-norm
approach because of the definition of protection value. In this case, if user i belongs to ej , its protection
value is gij pi , otherwise its protection value is 0. Hence, H11 changes to
X
ezi i
yi
+
e
[H11 ]ii = i
k hk ezk +
y
hi e i
kS,k6=i
X j
(
gij eyi + ij ij ),
ITj
jP

where ij is obtained by

g eyi
ij
ij =
0

if

i ej

otherwise

For D-norm, the matrix H is also positive definite, and the Lagrange dual function is Lipschitz continuous.
Hence, the D-norm based distributed algorithms converge to the optimal value when Proposition 2 holds
and step sizes are appropriately small.
A PPENDIX F
P ROOF

OF

L EMMA 1

Let xi be an iterative vector with ni steps that can be obtained from (11)-(15) as

[y , z , , ]
if
iS
i i i
i
xi =

if
iP
i
P
Algorithm 1 can be decomposed into N = iS+P ni block components x1 , , xS+P . Note that in

Algorithm 1, ni = 4 for i S and ni = 1 for i P . Besides, for the iterative primal and dual variables

obtained from (11)-(15), the followings hold: 1) all primal and dual variables are positive; 2) The gradient
algorithm descends along each direction (Lemma 5.1, Section 7.5 in [35]); and 3) The Lagrange dual
function is Lipschitz continuous. Thus, assumptions 5.1-5.2 of Section 7 in [35] hold. Hence, Algorithm
1 is a block-iterative algorithm and Algorithm 2 is partially asynchronous implementation of Algorithm
1 with maximum delay D (Section 7.1 in [35]). As such, there is a step size 0 so that if all step sizes
are smaller than 0 , Algorithm 2 converges to the optimal point of Algorithm 1. The value of 0 depends
on D and N (Propositions 5.1 and 5.3, Section 7 in [35]), i.e.,
0

1
1+D+N D

(42)

Hence (18) follows form some rearrangements. Note that the above conditions are also valid for the
D-norm approach, therefore the same constraint (18) can be applied for infrequent message passing in
the D-norm approach.

28

A PPENDIX G
P ROOF

OF

L EMMA 2

We note that (8) is a perturbed version of (4) with protection values in C2 and C3 . To obtain the
relationship between d and the protection value, we use the local sensitivity analysis of the optimization
problem by perturbing its constraints [8], [26], [37]. Let
u (a, b) = inf{

min

max

max

pi pi pi

ui (hi pi qi1 )|,

iS

X
qi i hi pi , (
gij pi + violate
) ITj ,
j
iS

(IP +

k6=i,kS

where a is a vector whose j

th

pk hk + I + 2 ) qi }

element is violate
, and b is a vector whose elements are all equal to I .
j

When violate
and I are small, u (a, b) is differentiable with respect to the perturbation vectors a and
j
b [37]. Using Taylor series, we write
u (a, b) = u (0, 0)+
X u (0, b) X u (a, 0)
+
bi
+o
aj
aj
bi
iS

jP

where

(43)

is the optimal value for (4), and 0 is the all zero vector. Note that u (a, b) and u (0, 0) are

equal to u and u , respectively. Since (4) is convex, from the sensitivity analysis in [37], we have
u (0,b)
aj

j and

u (a,0)
bi

i for all j P and i S . Hence

u u

jP

j violate

i I

(44)

iS

Since j and i are non-negative Lagrange multipliers, the SUs total throughput is reduced as compared
to the case in which exact channel gains are known.
A PPENDIX H
P ROOF

OF

P ROPOSITION 2

The GP problem (23) is feasible (Sec. 2.1.3 (2.9) in [23]) when the following optimization problem
is feasible
min

max
1s,pmin
i pi pi

(45)

29

C1 : eyi zi hii s

C : ezi (P
) s j P
ij eyi + violate
2
j
iS g
ITj
subject to
P

C3 : ezi (IP + k6=i,kS eykk hk + I + 2 ) s

violate = max
ij eyi .
ej S,|
gij |ij ,|ej |=violate
j
iej g

This problem can be divided into two parts. The first part is similar to Proposition 1. For the second
part, assuming that (38) holds for any definition of norm, the domain of C2 in (23) is
gj p + j kpk IT j

where kpk = maxej S,|ej |=violate

[29] by

kej

(46)

pk , and kpk is the dual norm. The dual of this norm is define

kpk = max(kpk ,

kpk1
),

(47)

where k k is the maximum element of the vector. As such, (47) is a feasibility condition for (23).
Thus, feasibility conditions for (23) are (35), (36), and (47).
A PPENDIX I
P ROOF

OF

L EMMA 3

xk1
Part 1: Since min{1, violate }kxk2 max(kxk , kviolate
), the feasibility set of (23) is greater than that

|S|

of (9), provided that (24) holds. Hence, there is a power allocation vector for (23) denoted by pD-norm
such that it satisfies all the constraints of the robust power allocation in (9), and
pD-norm pEllipsoid

(48)

where pEllipsoid is the allocated power vector for (9) in the ellipsoid uncertainty region, meaning that
the ellipsoids feasibility set for (9), i.e., FEllipsoid is a subset of the D-norm feasibility set of (23), i.e.,
FD-norm .

Part 2: Let the optimal solution to (23) subject to FD-norm be uD-norm and the optimal solution to (9)
subject to FEllipsoid be uEllipsoid . We assume that

p1 FEllipsoid ,

uEllipsoid (p1 ) > uD-norm (p2 )

p2 FD-norm

(49)

Since FEllipsoid FD-norm , we have p1 FD-norm , and from the assumption in (49) and convexity of
robust power allocation for the general definition of norm, the value of uEllipsoid (p1 ) is also the optimal

30

solution to (23) subject to FD-norm , i.e., uEllipsoid (p1 ) = uD-norm (p1 ). However, from the assumption in (49)
we have uEllipsoid (p1 ) > uEllipsoid (p1 ). This contradiction implies that our assumption was wrong, and thus
uD-norm uEllipsoid . The above holds generally for any form of convex optimization problem. For example,

to compare the total utility achieved by (9) and (4), the total utility of (9) is always less than that of (4),
because the feasible domain of (9) is smaller that that of (4).
R EFERENCES
[1] A. Goldsmith, S. A. Jafar, I. Maric, and S. Srinivasa, Breaking spectrum gridlock with cognitive radios: an information
theoretic perspective, Proceedings of the IEEE, vol. 97, no. 5, pp. 894914, May 2009.
[2] I. F. Akyildiz, W. Y. Lee, M. C. Vuran, and S. Mohanty, Next generation dynamic spectrum access cognitive radio wireless
networks: a survey, Computer Networks, vol. 50, no. 13, pp. 21272159, May 2006.
[3] D. I. Kim, L. B. Le, and E. Hossain, Joint rate and power allocation for cognitive radios in dynamic spectrum access
environment, IEEE Transactions on Wireless Communications, vol. 7, no. 12, pp. 55175527, Dec. 2008.
[4] G. Foschini and Z. Miljanic, A simple distributed autonomous power control algorithm and its convergence, IEEE Trans.
Veh. Technol., vol. 42, pp. 641646, Nov. 1993.
[5] R. D. Yates, A framework for uplink power control in cellular radio systems, IEEE J. Select. Areas Comm., vol. 17,
no. 7, pp. 13411347, Sep. 1995.
[6] M. Chiang, P. Hande, T. Lan, and C. W. Tan, Power control in wireless cellular networks, Foundations and Trends in
Networking, vol. 2, no. 4, pp. 381533, July. 2008.
[7] N. Bambos, C. Chen, and G. J. Pottie, Channel access algorithms with active link protection for wireless communication
networks with power control, IEEE/ACM Trans. Networking, vol. 8, no. 5, pp. 583 597, Oct. 2000.
[8] C. W. Tan, D. Palomar, and M. Chiang, Energy robustness tradeoff in cellular network power control, IEEE/ACM
Transactions on Networking, vol. 17, no. 3, pp. 912925, May 2009.
[9] K. Yang, J. Huang, Y. Wu, X. Wang, and M. Chiang, Distributed robust optimization part I: framework and example, submitted to Springer Journal of Optimization and Engineering, 2009. [Online]. Available:
http://www.princeton.edu/chiangm/DRO1.pdf.
[10] Y. Wu, K. Yang, J. Huang, X. Wang, and M. Chiang, Distributed robust optimization part II: wireless
power controle, submitted to Springer Journal of Optimization and Engineering, 2009. [Online]. Available:
http://www.princeton.edu/chiangm/DRO2.pdf.
[11] Y. Xing, C. N. Mathur, M. A. Haleem, R. Chandramouli, and K. P. Subbalakshmi, Dynamic spectrum access with QoS
and interference temperatureconstraints, IEEE Trans. Mobile Comp., vol. 6, no. 4, pp. 423433, April 2007.
[12] L. B. Le and E. Hossain, Resource allocation for spectrum underlay in cognitive wireless networks, IEEE Transactions
on Wireless Communications, vol. 7, no. 12, pp. 53065315, Dec. 2008.
[13] N. Gatsis, A. G. Marques, and G. B. Giannakis, Power control for cooperative dynamic spectrum access networks with
diverse QoS constraints, IEEE Trans. on Communications, vol. 58, no. 3, pp. 933 944, March 2010.
[14] R. Zhang, On peak versus average interference power constraints for protecting primary users in cognitive radio networks,
IEEE Trans. on Wireless Communications, vol. 8, no. 4, pp. 21122120, April 2009.

31

[15] X. Kang, Y. Liang, A. Nallanathan, H. K. Garg, and R. Zhang, Optimal power allocation for fading channels in cognitive
radio networks: Ergodic capacity and outage capacity, IEEE Transactions on Wireless Communications, vol. 8, no. 2, pp.
940950, Feb. 2009.
[16] S. Huang, X. Liu, and Z. Ding, Decentralized cognitive radio control based on inference from primary link control
information, submitted to IEEE Journal on Selected Areas in Communications,, 2010.
[17] S. Srinivasa and S. Jafar, Soft sensing and optimal power control for cognitive radio, in IEEE Global Telecommunications
Conference (GLOBECOM), 2007, p. 13801384.
[18] Y. Liang, Y. Zeng, E. Peh, and A. T. Hoang, Sensing-throughput tradeoff for cognitive radio networks, IEEE transaction
on wireless communciatoin, vol. 7, no. 4, pp. 13261337, April 2008.
[19] W. Ren, Q. Zhao, and A. Swami, Power control in cognitive radio networks: how to cross a multi-lane highway, IEEE
Journal on Selected Areas in Communications,, vol. 27, no. 7, p. 1283, Sep. 2009.
[20] R. Fan, H. Jiang, Q. Guo, and Z. Zhang, Joint optimal cooperative sensing and resource allocation in multichannel
cognitive radio networks, IEEE Transactions on Vehicular Technology, vol. 60, no. 2, pp. 722 729, Feb. 2011.
[21] A. Ben-Tal and A. Nemirovski, Selected topics in robust convex optimization, Mathematical Programming, vol. 1, no. 1,
pp. 125158, July 2007.
[22] G. Calafiore and M. Campi, Uncertain convex programs: randomized solutions and confidence levels, Math. Program.,
vol. A, no. 102, pp. 2546, 2005.
[23] M. Chiang, Geometric programming for communication systems, Foundations and Trends in Communications and
Information Theory, vol. 2, no. 1/2, 2006.
[24] S. Stanczak, M. Wiczanowski, and H. Boche, Resource Allocation in Wireless Networks: Theory and Algorithms. Berlin,
Germany: Springer, 2006.
[25] M. Chiang, C. W. Tan, D. Palomar, D. ONeill, and D. Julian, Power control by geometric programming, IEEE
Transactions on Wireless Communications, vol. 6, no. 7, pp. 2640 2651, July 2007.
[26] S. Boyd and L. Vandenberghe, Convex Optimization.

Cambridge University Press, 2004.

[27] A. B. Gershman and N. D. Sidiropoulos, Space-time processing for MIMO communications. John Wiley and Sons, 2005.
[28] G. Zheng, K.-K. Wong, and B. Ottersten, Robust cognitive beamforming with bounded channel uncertainties, IEEE
Transactions on Signal Processing, vol. 57, no. 12, pp. 4871 4881, Dec. 2009.
[29] D. Bertsimas, D. Pachamanova, and M. Sim, Robust linear optimization under general norms, Operations Research
Letters, vol. 4, no. 32, pp. 510516, 2004.
[30] Y. Yao and G. Giannakis, Rate maximization power allocation in OFDM based on partial channel knowledge, IEEE
Transaction on Wireless Communication, vol. 4, no. 3, pp. 10731083, May 2005.
[31] A. P. Iserte, D. P. Palomar, A. I. P. Neira, and M. A. Lagunas, A robust MAXIMIN approach for MIMO communications
with imperfect channel state information based on convex optimization, IEEE Transaction on Signal Processing, vol. 46,
no. 1, pp. 346360, Jan. 2006.
[32] A. Anandkumar, S. Lambotharan, and J. A. Chambers, Robust rate-maximization game under bounded channel
uncertainty, submitted IEEE Trans. Signal Processing, Nov. 2010.
[33] G. Zheng, K.-K. Wong, and B. Ottersten, Robust cognitive beamforming with bounded channel uncertainties, IEEE
Trans. Signal Processing, vol. 57, no. 12, pp. 4871 4881, Dec. 2009.
[34] J. Wang, C. Scutari, and D. P. Palomar, Robust MIMO cognitive radio via game theory, IEEE Transaction on Signal
Processing, 2011.

32

[35] D. P. Bertsekas and J. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods.

Prentice Hall, 1999.

[36] D. Bertsimas and M. Sim, The price of robustness, Operations Research, vol. 52, no. 1, pp. 3553, Feb. 2004.
[37] D. G. Cacuci, Sensitivity and Uncertainty Analysis. Chapman and Hall/CRC, 2003.