Power Allocation in OFDM-based Cognitive Radio
Systems
Peng Wang, Ming Zhao, Limin Xiao, Shidong Zhou, Jing Wang
National Laboratory for Information Science and Technology, Tsinghua University
Beijing, P. R. China
email: p-w04@mails.tsinghua.edu.cn
Abstract— In this paper, we investigate the optimal power allo-
cation strategy that aims at maximizing the capacity in OFDM-
based cognitive radio systems. We show that the traditional
water-filling algorithm applied in general OFDM systems needs
to be modified due to the per subchannel power constraints
in such systems. An iterative partitioned water-filling algorithm
is proposed and proved to be optimal based on the convex
optimization theory.
Index Terms— Cognitive radio (CR), power allocation, OFDM,
water-filling.
I. I NTRODUCTION
Spectrum scarcity is becoming a serious problem as the
rapid developments of wireless communications. However,
one of the FCC documents has indicated that many licensed
frequency bands are severely underutilized in both the time
domain and the spatial domain [1]. As a result, the FCC and
other spectrum managements have considered allowing the
licensed channels to be reused by other users that do not cause
unacceptable interference to the licensees. The licensee of a
channel is often referred to as primary user (PU) and all the
other users of the channel as secondary users (SUs).
Cognitive radio [2], [3] is the key technology enabling
flexible, efficient and reliable spectrum use by adapting the
radio’s operating characteristics to the real-time conditions
of the environment. By these characteristics, the SUs with
cognitive radio technology are able to cleverly detect and
utilize the idle licensed frequency channels where the PUs are
absent sometime or somewhere and thus implement efficient
reuse of the licensed channels.
Orthogonal frequency division multiplexing (OFDM) is a
promising candidate for cognitive radio systems [4], [5]. With
OFDM, the SU has the ability to flexibly fill the spectral gaps
left by PUs. Also, the FFT components at the OFDM system’s
receiver may also be used for the SU to execute the channel
detection. In [4], Weiss and Jondral proposed that the band of
the SU covers multiple PUs’s licensed spectrum, then the SU
modulates zero on the subcarriers which belong to the detected
PUs’s licensed spectrum while utilizing other subcarriers for
transmission.
This work is partially supported by National Nature Science Foundation of
China-No. 90204001 and China’s 863 project-No. 2006AA01Z274.
4061
1930-529X/07/$25.00 © 2007 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.
In traditional OFDM systems, optimal power allocation
that maximizes the channel capacity is water-filling on the
subcarriers with the total power constraint [6]. In order to
avoid unacceptable interference to the PUs that can not be
detected due to the large distance from the SU, the SU should
limit its transmit power even when no PU is detected [7].
Because of this requirement, per subchannel power constraints
are introduced into the OFDM-based cognitive radio systems.
As a result, if we want to obtain optimal power allocation that
maximizes the capacity of the SU, the water-filling strategy in
traditional OFDM systems needs to be reconsidered, which is
the subject of this paper.
The main results of this paper are: the power allocation
problem in OFDM-based cognitive radio systems is modeled;
based on the convex optimization theory, we present and prove
the sufficient and necessary conditions that the optimal power
allocation should satisfy; compared to an intuitive algorithm, a
more efficient algorithm, named as iterative partitioned water-
filling is proposed and proved to be optimal.
The organization of the rest of this paper is as follows:
Section II gives the system model and formulates the problem.
The problem is discussed in detail in Section III. Section IV
presents our algorithm and the corresponding proof. Section
V concludes.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
A. Cognitive radio system and power constraint
One of the typical cognitive radio systems is shown in Fig.
1. A certain channel is licensed to the PU system. Since the
PU system does not occupy the channel anywhere at any time,
the channel is underutilized in both the spatial domain and the
time domain. A channel is said to be a spectrum opportunity
if the interference to the PU receivers caused by the SU’s
transmission is acceptable. The SU is permitted to access the
channel if the channel is detected to be a spectrum opportunity.
With the participance of the SU, the channel efficiency can be
improved.
To guarantee the interference acceptable is a crucial problem
in the implementation of cognitive radio systems. First, the
SU has to detect the channel to determine whether a spectrum
opportunity exists before transmission. The SU is permitted
to transmit unless the PU’s signal is seen as absent. A
 subchannel 1 subchannel 2 subchannel j subchannel M
SU

PU1 ... ...

d
... ... m j ... N
1 2

{
R Lj

PU Tx
PU2 SU Tx
Fig. 2. Spectrum of SU in OFDM-based cognitive radio systems.

PU Rx

SU Rx
Fig. 1. Cognitive radio system model.
detection probability has to be achieved to reduce the collisions
caused by detection errors. As in Fig.1, PU1 can be detected
when active and thus the interference caused by the SU’s
transmission to PU1 receiver can be avoided. Second, since
PU2 transmitter is far away from the SU, the SU is unable to
detect PU2 ’s signal with the required detection probability, as
in Fig.1. In this situation, as in [7], PU2 defines a protection
area whose radius is R and requires the interference power at
the margin of the area be lower than a certain value, say η.
Thus, the SU’s transmission power Ptx in this channel has to
subject to a power constraint, which is given by
Ptx ≤ η(d − R)β
where d is the distance between the SU transmitter and the
nearest undetectable PU transmitter, β is the path attenuation
factor. Note that for simplicity we only consider the distance-
based path loss here.
Actually, the constraint on the SU’s power destiny function
is more meaningful for the PU, i.e., the power constraint is
not enough to describe the SU’s impact on the communication
performance of the PU. However, for this preliminary work,
we only consider the power constraint below.
B. OFDM-based cognitive radio system and per subchannel
power constraint
In order to efficiently utilize the spectrum opportunities left
by different PU systems, a cognitive radio system needs to be
highly flexible with respect to the spectral shape of the trans-
mitted signal. OFDM modulation is a promising candidate for
such a flexible system because of its reconfigurable subcarrier
structure. Also, the FFT component of OFDM can be used by
the energy detector of SUs for channel detection.
As in Fig.2, in an OFDM-based cognitive radio system, the
spectrum that can be potentially used by the SU belongs to
M PU systems. Consequently, the SU’s subcarriers are divided
into M subchannels. Suppose the total number of subcarriers
is N and the jth subchannel which is licensed to the jth PU
system has Lj subcarriers.
Before transmission, the SU detects each subchannel first.
If the PU transmitter is detected in a subchannel, all the
subcarriers in this subchannel will be modulated by zero
during the transmission, i.e., the sum power of the subcarriers
in this subchannel is set to be zero. Otherwise, the SU can use
this subchannel but with a certain power constraint which has
been described in the last section. Assume Gj is the power
constraint on the jth subchannel after channel detection, we
have


0 PUj is detected
Gj
(2)
ηj (dj − Rj )βj PUj is not detected
where ηj is the interference power constraint of PUj at the
margin of the area whose radius is Rj , dj is the distance
between the SU’s transmitter and the nearest undetectable
PUj ’s transmitter and βj is the corresponding path attenuation
factor.
In traditional OFDM systems, given a certain power con-
straint, the optimal power allocation that aims at maximizing
(1) the sum capacity is the well-known water-filling algorithm.
In our cognitive radio system, besides total power constraint,
there are also power constraints with respect to each subchan-
nel. In order to maximize the sum capacity in this situation,
the traditional water-filling algorithm needs to be modified.
Suppose the received symbol on the ith subcarrier in some
OFDM block is given by

yi = hi xi + ni (3)
where xi is the signal transmitted on the ith subcarrier by the
SU transmitter, hi is the channel gain and ni is the additive
white Gaussian noise with mean 0 and variance 1. Assume
the total power constraint of an OFDM block is Ptotal and
the allocated signal power on the ith subcarrier is Pi , then the
optimal power allocation problem can be expressed as:
N
P ∗ = arg max 12 i=1 log (1 + hi Pi )
s.t. Pi ≥ 0 i = 1, 2, · · · , N
N (4)
P
i=1 i ≤ P total
F j ≤ Gj j = 1, 2, · · · , M
 mj
where Fj
i=m j−1 +1
Pi is the power allocated to the jth
subchannel and mj is the index of the last subcarrier in the
jth subchannel ( suppose m0 = 0 ).
III. O PTIMAL P OWER A LLOCATION AND E XHAUSTIVE
S EARCH A LGORITHM
M
In (4), if Ptotal ≥ j=1 Gj , the total power constraint is
actually meaningless. (4) is degenerated to the problem of
power allocation with only per subchannel power constraints

4062
1930-529X/07/$25.00 © 2007 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.
 TABLE I
and the solution is well-known as water-filling on each sub-
I TERATIVE PARTITIONED WATER - FILLING
channel respectively with the corresponding power constraint.
Therefore, in the rest ofthis paper, we only emphasize on the
M
situation that Ptotal < j=1 Gj . Initialization:
Theorem 1: A power allocation vector P is the solution for A = {j|j = 1, 2, · · · , M }, B = ∅, P̂1 = Ptotal , k = 1,
C = {i|the ith subcarrier belongs to the jth subchannel, j
(4) if and only if it satisfies: ∈ A}
 
1 Begin:
Pi = max 0, wj − (5) 1. Perform traditional water-filling on the subcarriers in C
hi with power P̂k and obtain all the Pi where i ∈ C. The
resulting water level is ŵk .
where i = 1, 2, · · · , N and j is the index of the subchannel  mj
2. Fj = i=m j−1 +1
Pi where j ∈ A,
which the ith subcarrier belongs to. Assume A
{j|Fj <
M D = {j|Fj ≥ Gj , j ∈ A}.

Gj }, B
{j|Fj = Gj }. If Ptotal < j=1 Gj , we have A = 3. A = A \ D, B = B ∪ D, P̂k+1 = P̂k − j∈D Gj ,
∅. Then wj is determined by update C.
1. for j ∈ A: 4. If D = ∅, k = k + 1 and goto 1.
5. For each jth subchannel that j ∈ B, perform traditional
water-filling on the corresponding subcarriers with power
Gj respectively and obtain all the Pi where i ∈ C̄. The
wj = ŵ (6a) resulting water level of the jth subchannel is wj .
 mj   
1
max 0, ŵ − = Ptotal − Gj (6b)
hi
j∈A i=mj−1 +1 j∈B nel, where j ∈ B, with power Gj respectively and obtain
2. if B = ∅, for j ∈ B: the resulting water level wj .
3) Perform traditional water-filling on all the subcarriers
mj
   that belong to the subchannels in A with power Ptotal −
1 
max 0, wj − = Gj (7a) j∈B Gj and obtain the common water level ŵ.
i=mj−1 +1
hi 4) Verify each partition whether satisfies Fj < Gj where
wj ≤ ŵ (7b) j ∈ A and wj ≤ ŵ where j ∈ B. According to
Proof: see Appendix. Theorem 1, there is only one available partition and the
From Theorem 1, we know that the subcarriers with the op- corresponding power allocation vector is the solution.
timal power allocation can be divided into two sets: the set A,
IV. I TERATIVE PARTITIONED WATER - FILLING
i.e., the subchannels whose allocated power is strictly smaller
than the corresponding per subchannel power constraint and In the last section, we have proposed an exhaustive search
the set B, i.e., the subchannels whose allocated power is equal algorithm to obtain the optimal power allocation vector. How-
to the corresponding per subchannel power constraint. For ever, the complexity of this algorithm is far too high. In the
subchannels in A, the allocated power is the result of water- extreme situation, we need to consider 2M kinds of partitions.
filling on all the subcarriers that belong to these subchannels As a result, we develop a more efficient algorithm, named as

with power Ptotal − j∈B Gj . Thus, all the subchannels in iterative partitioned water-filling.
A have a common water level ŵ. For each subchannel in B, As show in Table I, in the kth iteration, traditional water-
e.g., the jth subchannel where j ∈ B, the allocated power is filling with only total power constraint P̂k is done on the
the result of water-filling on the subcarriers that belong to the subcarriers in C. Then those subchannels whose power ex-
jth subchannel with power Gj . Therefore, each subchannel in ceeds his per subchannel power constraint are taken out from
B has a unique water level wj and satisfies wj ≤ ŵ, i.e., the A and will be performed water-filling with the corresponding
water level of the subchannel which has a unique water level per subchannel power constraints respectively at the last step.
is less or equal to the common water level. The iteration stops when all the subchannels in A satisfy their
Also, based on the sufficient property of Theorem 1 and power constraints.
the strict concavity of the objective function, there is only one The iterative partitioned water-filling is more efficient than
power allocation vector satisfying all the conditions. However, Mexhaustive search algorithm. Since we assume Ptotal <
the
as in the traditional water-filling problem, it is difficult to j=1 Gj , there must exist at least one subchannel such that
obtain the explicit expression of the optimal power allocation Fj < Gj , i.e., A = ∅. Therefore, let k̂ be the number of
vector. An intuitive algorithm directly obtained from Theorem iterations when convergent, we have k̂ ≤ M − 1.
1 is the exhaustive search algorithm, which is described in To explain this algorithm is also optimal, we need to prove
brief below: that the algorithm converges to the point that satisfies the
1) Divide all the subchannels into two sets, say A and B conditions (5) – (7) in Theorem 1. Actually, we only need
and there are 2M kinds of partitions in  total. to prove the condition (7b), i.e., wj ≤ ŵk̂ where j ∈ B, as
2) Remove such partitions that Ptotal − j∈B Gj ≤ 0. other conditions are inherently satisfied.
For each of the rest partitions, perform traditional water- Before the proof, we outline two lemmas about traditional
filling on the subcarriers that belong to the jth subchan- water-filling. Assume P is the power allocation vector of

4063
1930-529X/07/$25.00 © 2007 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.
 water-filling on N subcarriers with power 1T P and the water Power
Pi
level is w, where 1 is the column vector of all ones. 1/ hi
Lemma 1: If we take out n subcarriers from these N ŵ
w4
subcarriers and the corresponding power allocation vector of
these n subcarriers is P n , P n is also the power allocation w2
vector of water-filling on the n subcarriers with power 1T P n .
Lemma 2: If water-filling is done on these N subcarriers



with power P and the corresponding water level is w . We



have w ≤ w if P ≤ 1T P and vice versa.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Subcarriers
Proofs of these lemmas above are omitted here due to page
Subchannel 1 Subchannel 2 Subchannel 3 Subchannel 4
limit.
Theorem 2: The iterative partitioned water-filling algo- Fig. 3. Iterative partitioned water-filling.
rithm converges to the point satisfying wj ≤ ŵk̂ , where j ∈ B,
i.e., the unique water level is less or equal to the common water
level.
Proof: During the kth iteration, we obtain all the Pi , mizing the capacity with consideration of the per subchannel
where i ∈ C and the water level ŵk after step 1. If we find power constraints caused by the PUs’s interference limits.
some j that satisfy Fj ≥ Gj , then the jth channel needs In practical systems, the PU suffers interferences caused by
to be taken out from A. Water-filling will be done on the not only the SU’s transmission in the corresponding subchan-
corresponding subcarriers with power Gj at step 5 and the nel, but also the side-lobes of other subchannels, which was
resulting water level is wj . From Lemma 1, if water-filling not considered in this paper. In the future, we plan to inves-
is done on these subcarriers with power Fj , we also get the tigate the power allocation problem with further constraints
water level ŵk . From Lemma 2, since Fj ≥ Gj , we have introduced by the side-lobes of neighboring subcarriers.

wj ≤ ŵk (8) A PPENDIX

P ROOF OF T HEOREM 1
Based on Lemma 1, for the rest subchannels satisfying
Fj < Gj , if water-filling  is done on the corresponding The problem (4) can be reformulated into a standard convex
optimization form:
subcarriers with power P̂k − j∈D Fj , we also get the water
N
level ŵk . In the next iteration, we need to perform water- minimize − 12 i=1 log (1 + hi Pi )
filling on these subcarriers with power P̂k+1 = P̂k − j∈D Gj s.t. −Pi ≤ 0 i = 1, 2, · · · , N
N (10)
and the resulting water level is ŵk+1 . By Lemma 2, since i=1 Pi − Ptotal ≤ 0
P̂k − j∈D Fj < P̂k+1 , we have Fj − Gj ≤ 0 j = 1, 2, · · · , M
ŵk ≤ ŵk+1 (9) The constraint conditions obviously satisfy Slater’s con-
After k̂ iterations, the algorithm converges and the water ditions, so the Karush–Kuhn–Tucker(KKT) conditions are
level of the subchannels in A is ŵk̂ . From (9), we obtain sufficient and necessary for the optimal vector P [8]. The first
ŵk ≤ ŵk̂ , where k = 1, 2, · · · , k̂. Also, based on (8), it follows two KKT conditions are the constraint conditions of (10) and
∀j ∈ B, ∃k, so that wj ≤ ŵk . the others are given by
In conclusion, we have ∀j ∈ B, wj ≤ ŵk̂ , i.e., the unique
water level is less or equal to the common water level. λi ≥ 0 (11a)
So far, we have proved that the algorithm in Table I is
αj ≥ 0 (11b)
also optimal. Based on our algorithm, an iterative partitioned
water-filling result with four subchannels is illustrated in Fig.3. v≥0 (11c)
In each subchannel, it is just the traditional water-filling result. where i = 1, 2, · · · , N , j = 1, 2, · · · , M ,
Subchannel 1 and 3 have the same water level ŵ, as each
power is strictly lower than the corresponding per subchannel
power constraint. Subchannel 2 and 4 have their unique water λi Pi = 0 (12a)
levels w2 , w4 respectively, as each power is equal to the αj (Fj − Gj ) = 0 (12b)
corresponding per subchannel power constraint. Also, as in  N


Fig.3, we see that the unique water levels w2 , w4 are lower v Pi − Ptotal =0 (12c)
than the common water level ŵ. i=1

V. C ONCLUSION where i = 1, 2, · · · , N , j = 1, 2, · · · , M ,
In conclusion, we have proposed an iterative partitioned
water-filling algorithm applied to power allocation in OFDM- 1
− − λi + v + αj = 0 (13)
based cognitive radio systems. This algorithm aims at maxi- Pi + 1/hi

4064
1930-529X/07/$25.00 © 2007 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.
 where i = 1, 2, · · · , N and j is the index of the subchannel II-1. Based on (6b) and A = ∅, we have ŵ > 0. Define
which the ith subcarrier belongs to. v = 1/ŵ, then v > 0. From (6b), we can also obtain
I. First, we prove the only if parts hold. N
(13) can be written as i=1 Pi 
= j∈A Fj + j∈B Fj

1  mj 1
λi = v − + αj . (14) = j∈A i=mj−1 +1 max 0, ŵ − hi + j∈B Gj
Pi + 1/hi
= Ptotal .
Substituting (14) into (11a) and (12a) yields (22)
Therefore, (11c) and (12c) hold.
1
v + αj ≥ , (15) II-2. Define αj = 1/wj − 1/ŵ, from (6a) and (7b), we can
Pi + 1/hi conclude that


  αj = 0 where j ∈ A
1 αj ≥ 0 where j ∈ B
(23)
v− + αi Pi = 0. (16)
Pi + 1/hi
Thus, (11b) is satisfied. Since for j ∈ B, we have Fj = Gj ,
If v + αj < hi , we have Pi > 0 by (15). Then (16) leads to (12b) is also satisfied.
1
1 1 II-3. Define λi = v − Pi +1/hi
+ αj , where j is the index of
Pi = − . (17) the subchannel which the ith subcarrier belongs to.Then (13)
v + αj hi
is satisfied inherently.
1
If v + αj ≥ hi , from (15), we obtain v − Pi +1/h + αi > 0. If wj > 1/hi , based on (5), we have
i
Then based on (16), it follows that Pi = 0. 1
Let wj = 1/(v + αj ), then Pi = wj − . (24)
hi
 
1 Since we have wj = 1/(v + αj ) from the definitions of αj
Pi = max 0, wj − (18)
hi and v, (24) can be written as
where i = 1, 2, · · · , N and j is the index of the subchannel 1
v− + αj = 0. (25)
which the ith subcarrier belongs to. Pi + 1/hi
The subchannels are Mdivided into two sets A and B. Since Then λ = 0. Thus, given wj > 1/hi , (11a) and (12a) hold.
we assume Ptotal < j=1 Gj , there must exists at least such If wj ≤ 1/hi , from (5), we have Pi = 0. Then
one subchannel that Fj < Gj , i.e., A = ∅.
1
I-1. For all j ∈ A, since Fj < Gj , we have αj = 0 based λi = v − hi + αj = − hi ≥ 0. (26)
wj
on (12b). Let ŵ = 1/v, then wj satisfies the condition (6a)
wj = ŵ. Hence, given wj ≤ 1/hi , (11a) and (12a) also hold .
From (15), we obtain In conclusion, we have verified all of the KKT conditions
and thus the if parts also hold.
v + αj > 0, where j = 1, 2, · · · , M (19) This completes the proof of Theorem 1.
N
then v > 0. Consequently, (12c) leads to i=1 Pi = Ptotal . R EFERENCES
Thus, we come to the condition (6b) [1] Federal Communications Commission Spectrum Policy Task Force,
 m j
“FCC Report of the Spectrum Eifficiency Working Group”, November
1
i=mj−1 +1 max 0, ŵ − hi
2002.
j∈A
 [2] J.Mitola, “Cognitive radio for flexible mobile multimedia communica-
= Fj
j∈A
N  (20) tions,” In Proc. IEEE Int. Workshop on Mobile Multimedia Communi-
= i=1 Pi − j∈B Fj cations, pp. 3-10, 1999.
[3] S.Haykin, “Cognitive radio: Brain-Empowered Wireless Communica-
= Ptotal − j∈B Gj . tions,” In IEEE Journal on Selected Areas in Communicaitons, vol. 23,
pp. 201-220, February 2005.
I-2. For all j ∈ B, the condition (7a) is obviously satisfied. [4] T.Weiss and F.Jondral, “Spectrum pooling: an innovative strategy for the
Since αj ≥ 0, we have enhancement of spectrum efficiency,” IEEE Communications Magazine,
vol. 42, pp. S8-S14, March 2004.
1 1 [5] U.Berthold and F.Jondral, “Guidelines for designing OFDM overlay
wj = ≤ = ŵ (21) systems,” in Proc.of the first IEEE Syposium on New Frontiers in
v + αj v Dynamic Spectrum Access Networks, November 2005.
[6] D.TSE and P. Viswanath, C, “Fundamentals of Wireless Communica-
Therefore, the inequality (7b) also holds. tion,” Cambridge University Press, May 2005.
So far, we have proved the only if parts in Theorem 1. [7] Qing Zhao and B.M.Sadler, “A survey of dynamic spectrum access:
signal processing, networking, and regulatory policy,” IEEE Signal
II. To prove the if parts hold, we only need to prove that Processing Magazine, vol. 55, no. 5, pp. 2294-2309, May 2007.
all the KKT conditions can be satisfied. [8] S.Boyd and L. Vandenberghe, ”Convex Optimization”, Cambridge Uni-
The first two KKT conditions, i.e., the constraint conditions versity Press, 2004.
in (10), are obviously satisfied.

4065
1930-529X/07/$25.00 © 2007 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2007 proceedings.