Sunteți pe pagina 1din 5

A Distributed Algorithm for Resource Allocation in

OFDM Cognitive Radio Systems


Yonghong Zhang and Cyril Leung
The University of British Columbia
Vancouver, B.C., Canada V6T 1Z4
Email: joyz@ece.ubc.ca, cleung@ece.ubc.ca

AbstractWe study the problem of allocating subchannels,


bits, and powers for variable rate services in an OFDM based
cognitive radio (CR) system, in which available system resources
are highly dynamic. In a resource-limited situation under which
the nominal rate requirements of users cannot be satisfied, it is
desirable to provide fair degradation among users. In a situation
with abundant resources, we may choose to maximize system
throughput while ensuring that user nominal rate requirements
are met. The problem is formulated as a single objective nonlinear optimization problem using techniques from goal programming. A distributed resource allocation algorithm is proposed and
simulation results are obtained which show that the proposed
distributed algorithm provides good fairness and aggregate bit
rates close to (within 8% of) optimal values.
Index Termscognitive radio, resource allocation, distributed
algorithm, OFDM, ad-hoc systems, multi-cell systems, goal programming.

I. I NTRODUCTION
It has been suggested [1] that the looming communications
spectrum shortage crisis can be alleviated using cognitive radio
(CR) technology [2]. It allows unlicensed users to use licensed
frequency bands when the licensed users are not active. The
highly dynamic nature of the bandwidth available to a CR
system makes resource allocation very challenging.
Due to its flexibility in dynamically allocating radio resources to multiple users and low interference between adjacent subcarriers, orthogonal frequency division multiplexing
(OFDM) is considered an appropriate modulation scheme for
CR systems [3]. Due to the time-varying resources available
in a CR system, it is possible that the nominal bit rates for
CR users cannot be provided. In resource-limited situations, it
is desirable to provide an equitable level of rate degradation
among the CR users. The proportional rate schemes in [4]
and [5] provide fair degradation among data users in OFDM
systems by guaranteeing specified rate ratios. However, they
consider that the rate ratios should be maintained even after
user nominal rate requirements have met; this often limits
efficient use of resources.
The above-mentioned algorithms are centralized. Distributed algorithms may be more suitable in multi-cell systems or ad hoc systems. In such systems, subcarriers are
simultaneously shared among several users and cochannel
interference has to be considered in allocating resources. In
[6], the cochannel interference in an OFDM-based digital
subscriber loop (DSL) system is treated as noise and the power

978-1-4244-1722-3/08/$25.00 2008 IEEE.

allocation problem is viewed as a non-cooperative game. A


distributed iterative waterfilling (ITWF) algorithm is proposed
which requires knowledge of the user highest achievable target
rates (HATRs) in order to obtain the optimal power allocation
solution. Obtaining this knowledge is difficult for a distributed
algorithm in a time-varying wireless channel environment. In
[7], a distributed algorithm is designed to maximize system
throughput for an OFDM system with no nominal user bit
rates. The distributed algorithm for multi-cell systems in [8]
can handle nominal user bit rates but does not consider fair
user degradation in resource-limited scenarios.
In this paper, we consider subcarrier, power, and bit allocation for user applications with nominal bit rates in an
OFDM-based CR system. In contrast to conventional OFDM
systems, the available resources in a CR system are highly
dynamic. When resources are scarce, the goal is to provide fair
degradation among all CR users. When resources are abundant,
we try to maximize system throughput while ensuring that the
nominal rates for all users are satisfied.
II. S YSTEM M ODEL
Consider a CR system, with bandwidth W , in which the
(licensed) primary users (PUs) are not active all the time. Assume that there are altogether K CR transceiver pairs (CRPs),
all of which use OFDM. The CRPs may use a PU frequency
band whenever the PU is sensed not to be active. There are
M subbands (subchannels) and the nominal bandwidth of
subband m, m = {1, 2, . . . , M } ranges from fc + (m 1)f
to fc + mf . The subbands (or subchannels) are modelled
in discrete-time, with the time-varying gain for subchannel
m
CRP is transmitter
 to CRP js receiver denoted by
 from
m . Let the power gains g m be outcomes of independent
gj,i
j,i 

m
random variables (rvs), and let gi,j
, m = 1, 2, , M be
independent identically distributed (i.i.d.) rvs for any given
values of i and j. It is assumed that there is no inter-carrier
interference (ICI). Each subchannel is shared among CRPs,
and the signal of any CRP is considered to be interference to
other CRPs. We note that in general the interference powers
experienced at the two transceivers of a given CRP are not
equal.
In order to reduce the mutual interference between the CRPs
and the PUs, the subchannels adjacent to PU bands are not
used by CRPs [9]. The PU active frequency bands, PU guard
bands, and spectrum holes are shown in Figure 1. Because

Ma


sk,m

Ik,m

Sk

(6)

m=1
K


m
gk,l
sl,m

(7)

l=1,l=k

Fig. 1.

Primary users active frequency bands and spectrum holes.

the active PU bands vary over time, the number, Ma , of


subchannels available to CRPs can range from 0 to 32. For
the example shown in Figure 1, Ma = 15 at this time at this
location.
The rapidly time-varying nature of available resources complicates resource allocation, especially if CRPs have nominal
bit rate targets. Suppose that the nominal rate requirement for
CRP k is RkN OM . When the number of available subchannels
is low, not every CRPs nominal rate may be achieved, i.e.,
for some CRP k, Rk < RkN OM where Rk is the bit rate
over all subchannels for CRP k. In this case, we ensure
fair degradation among CRPs, i.e. allocate resources so as
to equalize the fractional rate degradation, dk = (RkN OM
Rk )/RkN OM , RkN OM > Rk for all CRPs. The system degradation fairness is measured using the fairness index in [10],
i.e.
K
( k=1 dk )2
.
(1)
f= 
K
K k=1 dk 2
When all K CRP rate degradations are the same, i.e. dk =
, k, the fairness index attains its maximum value of 1. The
lowest value is f = 1/K. The f -value roughly reflects the
fraction of CRPs that receive similar service.
Under resource-limited conditions, the goal is to maximize
some
K function h of the fairness index, f , and the sum rate,
k=1 Rk , for CRPs. When resources are adequate, i.e., in
a resource-abundant situation, the goal is to maximize the
throughput of the CR system while ensuring that the nominal
user bit rate requirements are satisfied.
The optimization problem OP1 can thus be formulated as
 

K
K
( k=1 dk )2 
O1 : max h
,
Rk if Rk < RkN OM for some k
K
2
sk,m

O2 : max
sk,m

K


k=1

dk

k=1

Rk if Rk RkN OM for all k

(2)

k=1

subject to


dk

Rk

N OM
Rk
Rk
N OM
Rk

1
Ma

m=1

rk,m

= log2

Rk < RkN OM

Rk RkN OM

rk,m

m
sk,m
gk,k
1+
(Nk,m + Ik,m )

(3)
(4)


(5)

where subscripts k and m refer to CRP k and subchannel m


respectively, Sk is the total transmit power constraint for CRP
k, is a SNR gap parameter which indicates how far the
system is operating from capacity, sk,m is the transmit power,
Nk,m is the noise power, and Ik,m is the interference power
from other CRPs.
Problem OP1 is a two-objective non-linear optimization
problem, which is generally hard to solve. Here, we study the
case in which fair degradation is guaranteed, i.e., f = 1. In
a resource-abundant situation, the condition Rk RkN OM , k
implies that f = 1 from (3). In a resource-limited situation,
to ensure that f = 1, we need
Rk
= 1, 0 1, k = 1, 2, , K; if k, Rk < RkN OM .
RkN OM
(8)
As a result, when f = 1 is guaranteed, Problem OP1 can
be transformed to
OP2 :

max
sk,m

K


Rk

(9)

k=1

subject to constraints (4)-(7), and (8).


Because Problem OP2 involves constraint (8), which only
applies in the resource-limited situations, it is still hard to
solve. We can simplify Problem OP2 using a goal programming approach [11]. There are usually two steps involved
in goal programming, i.e., (1) introduce slack and surplus
variables to combine the difference requirements in resourcelimited situations (require (8)) and resource-abundant situations (require Rk RkN OM , k); (2) rewrite the objective
function in (9) to reflect the goals of the original problem.
Then, the objective of Problem OP2 becomes
OP3 :

max
sk,m

K


pk

k=1

K


RkN OM

(10)

k=1

and constraint (8) is replaced with


Rk + RkN OM pk = RkN OM , 0 1, pk 0, pk = 0
(11)
where RkN OM and pk are the slack and surplus variables for
CRP k. The slack variable, RkN OM , is the difference between
RkN OM and Rk when Rk < RkN OM . The surplus variable, pk ,
is the difference between Rk and RkN OM when Rk > RkN OM .
Since the slack value for CRP k is proportional to CRP ks
nominal rate requirement, the resulting fairness index will be
equal to 1.
We will refer to the set of user rates {Rk , k = 1, 2, . . . , K}
which results from solving problem (10) as the highest achievable target rate (HATR) set and Rk as the individual HATR
for CRP k. The optimization problem in (10) can be solved
using a non-linear single objective optimizer. However, the

computational complexity for such an approach is generally


very high. We next propose a distributed algorithm with lower
complexity which is especially suited for ad hoc or multi-cell
infrastructure systems.
III. D ISTRIBUTED A LGORITHM
As mentioned in [6], if each transmitter knows its HATR,
then an iterative waterfilling algorithm yields a Nash equilibrium if we view the allocation problem as a non-cooperative
game. In the following subsections, we examine the following
issues: (1) how can CRPs determine if a set of target rates
is achievable, (2) how can a CRP determine its individual
HATR when the system is resource-limited, (3) how can a CRP
determine its individual HATR when system are abundant,
and (4) how to design the distributed algorithm for resource
allocation. Since in practical systems, the number of bits has
to be an integer, this constraint is taking into account in this
section.
A. Determining Achievability of Target Rates
Let us first consider the case with only one subchannel.
Assume that each receiver of a CRP sends back its current
interference level to its corresponding transmitter. Suppose
that the target rate for CRP k is rk . Since the subchannel
is shared by multiple CRPs, each trying to maintain its
target rate, an increase in the power of any CRP transmitter
will result in an increase in interference to the other CRP
receivers, which would in turn cause an increase in the other
CRP transmitter powers. These higher powers will generate
more interference to the original CRPs receiver, causing its
transmitter to increase power again. This process of growing
CRP transmitter powers continues until either an equilibrium
point is attained or some of the CRP transmitters reach their
power limits unable to meet their target rates. A CRP thus
knows that if it is unable to reach its target rate, given the
current interference power level and it has reached its transmit
power limit, then that target rate is not achievable.
This strategy can be extended to multiple subchannels by
letting each CRP perform iterative water-filling. Each CRP
allocates power to subchannels by water-filling and treating
other CRP signals as interference. If the resulting total rate
over all subchannels is still below the CRPs target rate RkT AR
when its transmit power limit is reached, then the CRP knows
that RkT AR is not achievable. The CRP then broadcasts a
Not achievable (NA) message to inform other CRPs that
the current set of target rates is not achievable.
A pseudo-code listing for the algorithm is given in Algorithm 1. The parameter T is the time needed to determine
whether a given HATR set is achievable or not. The input
parameter RkDAT A is the rate at which CRP k can transmit
data.
B. Determining HATR in a Resource-limited Situation
Initially, each CRP set RkN OM as its target rate. Once it is
determined that the target rate set {RkN OM , k = 1, 2, . . . , K}
is not achievable, then the goal is to determine the HATR set

Algorithm 1 Achievable Rate Algorithm.


input: RkT AR , RkDAT A
output: achievable
DAT A to 0, m = 1, 2, , M
(1) initialize rm and rm
a
(2) if tstart T , exit; otherwise, do steps (3) to (8)
(3) measure current interference power Ik,m for each subchannel
DAT A
by using water-filling
(4) get rm
Ma and rmT AR
(5) if
r

R
, achievable = true;
m
k
m=1
otherwise achievable = f alse, broadcast message NA
(6) if receive message NA, set achievable = f alse.
DAT A bits to subchannel m
(7) load rm
(8) goto step (2)

while ensuring fair degradation. For this purpose, we propose


to use a binary search approach. The idea is based on the
binary search algorithm [12] for finding a particular value in
a sorted list.
Each CRP k keeps track of two individual target rates,
Rk,l and Rk,h : the value of Rk,l is the current highest ATR
plus 1, whereas Rk,h is the current lowest non-achievable
TR minus 1. The actual value of HATR is in the set R =
{Rk,l 1, Rk,l , Rk,l + 1, . . . , Rk,h }.
Initially, Rk,l is set to 1 and Rk,h is set to RkN OM . Then
at each round, the CRP sets its target rate to rk = (Rk,l +
Rk,h )/2. If the target rate turns out to be achievable, Rk,l is
set to rk + 1, otherwise, Rk,h is set to rk 1. This process
continues until the rate Rk,l exceeds Rk,h .
At each round, the size of the set R is halved, since either
the top half is removed if the target rate is not achievable or the
lower half is removed if the target rate is achievable. CRP k
can determine its individual HATR in O(log(RkN OM )) rounds.
A pseudo-code listing for the binary rate search algorithm is
given in Algorithm 2: Rk,min specifies the minimum rate that
should be used for its individual HATR. In a resource-limited
situation, Rk,min is set to 0.
Algorithm 2 Binary Rate Search Algorithm.
input Rk,l , Rk,h , Rk,min
output RkHAT R
(1) if Rk,l = Rk,h , then let RkHAT R = Rk,l and exit
(2) if Rk,l Rk,h , then do steps (2.1) to (2.3); otherwise, goto (3)
(2.1) rk = (Rk,l + Rk,h )/2
(2.2) call Algorithm 1 with input RkT AR = max(rk , Rk,min )
RkDAT A = Rk,l 1
(2.3) if the output of Algorithm 1 is achievable, then Rk,l = rk + 1;
otherwise, Rk,h = rk 1. Goto step (2)
(3) RkHAT R = Rk,l 1

C. Determining HATR in a Resource-abundant Situation


If the HATR for each CRP k is at least RkN OM , the system
is resource-abundant. Based on the current interference level
at the rate set {RkN OM , k = 1, 2, , K}, the HATR can be
M AX using water-filling and assuming that the
estimated as R
k
interference power level will not be raised. Since the level will
M AX , k = 1, 2, . . . , K} is
actually increase if the rate set {R
k
M
AX

used, R
is an upperbound on the HATR.
k

Flow chart of the distributed allocation algorithm.

Since plain equal power allocation, in which power is


equally allocated to all subchannels, yields similar results as
water-filling [13], we can estimate RkM AX as

Ma
m

Sk gk,k
M AX

(12)
Rk
=
log2 1 +
Ma (Nk,m + Ik,m )
m=1
where Ik,m is the interference power that results when all
CRP rates are set to their nominal values. The HATR for
CRP k can be found using Algorithm 2 with Rk,l = 1 and
M AX . By setting Rk,min to RN OM , we ensure that
Rk,h = R
k
k
the HATR is no less than RkN OM .
D. The Proposed Distributed Algorithm
The proposed distributed algorithm for finding a suboptimal solution to problem (10) can be stated as follows.
Each CRP k first sets its target rate at RkN OM . Using the
procedure in Subsection III-A, it determines if the target rate
set {RkN OM , k = 1, 2, , K} is achievable.
If the target rate set is not achievable, CRP k then uses
Algorithm 2 to find its HATR between 1 and its RkN OM . If
the target rate set is achievable, CRP k then uses Algorithm 2
M AX .
to find its HATR between RkN OM and R
k
CRP k then uses its HATR as the target rate and performs
iterative waterfilling. Figure 2 shows a block diagram of the
algorithm.
IV. S IMULATION R ESULTS
To evaluate the effectiveness of the proposed distributed
algorithm summarized in Fig. 2, simulations were performed
on a system covering a 3 km 3 km area. The transmitters
are uniformly distributed within that area and each receiver
is uniformly distributed within a 300 meter circle centered on
its corresponding transmitter. The path loss exponent is 3, the
variance of the shadowing effect is 10 dB, and the multipath
fading is assumed to be Rayleigh. The CRP bit error rate
requirement is 104 , which corresponds to a value of 5.
The number, Ma , of subchannels available to CRPs ranges
from 2 to 16 and each subchannel has a noise power Nk,m
of 1013 W.
For performance comparison, we obtained the optimal solution for Problem OP3 for each of a number of different
realizations of CRP locations using a commercial optimization
software package. Similarly, the optimal solution with a fixed
fractional rate degradation for all CRPs in resource-limited as
well as resource-abundant situations, hereafter referred to as

Average number of bits per OFDM symbol per CRP

Fig. 2.

80
70

Optimal (real bits)


Proportional rate (real bits)
Proposed (integer bits)

60
50
40
30
20
10
2

6
8
10
12
Ma, Number of available subchannels

14

16

Fig. 3. Average number of bits per OFDM symbol duration per CRP as a
function of the number of available subchannels with Sk = 104 W, K = 3,
R1N OM = 25, R2N OM = 30, R3N OM = 35.

proportional rate, was also obtained. The proportional rate


constraint is equivalent to
N OM
.
R1 /R1N OM = R2 /R2N OM = = RK /RK

(13)

Fig. 3 shows the average number of bits per OFDM symbol


per CRP as a function of the number of available subchannels.
The results for the optimal solution is obtained by solving
Problem OP3, and the results for the optimal proportional
rate solution is obtained by solving Problem OP2 but with
constraint (8) replaced by (13). To keep the running times
for obtaining optimal solutions reasonable, the value for each
point in Fig. 3 is averaged over 1000 realizations of CRP
locations.
It can be seen from Figure 3 that the proposed distributed
algorithm provides an average bit rate (throughput) close to
(within 8% of) optimal. The throughput with the proportional rate constraint is considerably lower because it limits
efficient use of the extra available system resources. The
throughput improvement with the proposed algorithm relative
to the proportional rate solution increases with the number of
available subchannels from 3% with 4 subchannels to 56%
with 16 subchannels. The fairness indices in all three cases
were found to be very close to 1.
To explain the relative performances of the optimal, proportional rate and proposed algorithms, we consider a system
with four subchannels and a randomly selected realization of
CRP locations. Table I lists the subchannel gains from CRP js
transmitter to CRP is receiver.
The three CRPs have equal transmit power limits, namely
Sk = 0.1 mW. The total (summed over all four subchannels)
number of bits per OFDM symbol duration for each CRP
is shown in Table II. Results are given for each of the
three algorithms and four different sets of nominal CRP rate
requirements. Note that in Table II the results for the optimal
and the proportional rate algorithms have been rounded to the
nearest integer.
For the first set of nominal CRP rate requirements, i.e.

i
1

j
1
2
3
1
2
3
1
2
3

sub 1
919
1.77
1.23
0.051
140
0.038
0.078
0.322
180

sub 2
7723
1.32
0.598
0.07
1696
0.052
0.083
0.198
193

sub 3
2058
4.46
0.567
0.038
2227
0.032
0.052
0.009
135

sub 4
6910
2.19
0.324
0.042
566
0.065
0.013
0.269
85

TABLE I
S UBCHANNEL GAINS (1010 ) FROM CRP j S TRANSMITTER TO
CRP i S RECEIVER .

RkN OM

= 10 bits per OFDM symbol duration, the system


is resource-limited and all three CRPs cannot achieve their
nominal rates. The goal is to allocate as many bits as possible
while keeping the rate ratios Rk /RkN OM for the three CRPs
as close as possible. All three algorithms yield CRP rates of
8 bits per OFDM symbol duration and a fairness value of 1.
For the second set of nominal CRP rate requirements, i.e.
R1N OM = 10 bits, R2N OM = 15 bits, R3N OM = 20 bits, the
CR system is even more resource-limited. The CRP sum rate
and the fairness index for the proposed algorithm is slightly
lower than the optimal and proportional rate values. We note
that the optimal and proportional rate algorithms produce the
same solutions when the system is resource-limited because
then the constraint in (13) applies to both.
For the third and fourth sets of nominal CRP rate requirements, the CR system is resource-abundant. The goal is to
maximize the overall system throughput while ensuring that
all CRP nominal rates are met. The proposed algorithm gives
a fairness value of 1 and sum rates which are within 4% of
the optimal values. When the system is resource-abundant, the
proportional rate algorithm does not make efficient use of the
surplus resources and its sum rates are lower than those of the
proposed algorithm by 32% and 74%, respectively.
V. C ONCLUSION
The subchannel, bit, and power allocation problem for
users with nominal bit rate requirements in an OFDM-based
CR system was formulated as a multi-objective non-linear
optimization problem. The goal is to provide a fair bit rate
degradation among users in resource-limited situations and to
maximize system throughput while satisfying user nominal
rate requirements when resources are plentiful. A goal programming approach was used to transform the problem into a
single objective non-linear optimization problem. A distributed
algorithm was designed to solve the problem. Simulation
results were provided which show that the proposed algorithm
has a performance which is within 8% of optimal.
Acknowledgment: This work was supported in part by the
Natural Sciences and Engineering Research Council of Canada
(NSERC) under Grant OGP0001731, an NSERC postgraduate
scholarship and by the UBC PMC-Sierra Professorship in
Networking and Communications.

R1N OM = 10, R2N OM = 10, R3N OM = 10


Algorithm
CRP 1
CRP 2
CRP 3
sum rate
Optimal
8
8
8
24
Proposed
8
8
8
24
Proportional rate
8
8
8
24

fair index
1
1
1

R1N OM = 10, R2N OM = 15, R3N OM = 20


Algorithm
CRP 1
CRP 2
CRP 3
sum rate
Optimal
4
6
8
18
Proposed
3
6
8
17
Proportional rate
4
6
8
18

fair index
1
0.997
1

R1N OM = 6, R2N OM = 6, R3N OM = 4


Algorithm
CRP 1
CRP 2
CRP 3
sum rate
Optimal
24
17
8
49
Proposed
23
17
7
47
Proportional rate
12
12
8
32

fair index
1
1
1

R1N OM = 1, R2N OM = 2, R3N OM = 4


Algorithm
CRP 1
CRP 2
CRP 3
sum rate
Optimal
24
17
8
49
Proposed
23
17
7
47
Proportional rate
2
4
8
14

fair index
1
1
1

TABLE II
N UMBER OF BITS PER OFDM SYMBOL AND FAIRNESS INDEX FOR EACH
OF THE THREE ALGORITHMS AND FOUR DIFFERENT SETS OF NOMINAL
RATE REQUIREMENTS .

R EFERENCES
[1] Federal Communications Commission, Facilitating opportunities for
flexible, efficient, and reliable spectrum use employing cognitive radio
technologies, FCC 03-322, 2003.
[2] S. Haykin, Cognitive radio: brain-empowered wireless communications, IEEE Journal on Selected Areas in Communications, vol. 23,
no. 2, pp. 201220, 2005.
[3] T. A. Weiss and F. K. Jondral, Spectrum pooling: an innovative strategy
for the enhancement of spectrum efficiency, IEEE Communications
Magazine, vol. 42, no. 3, pp. S8S14, 2004.
[4] Z. Shen, J. G. Andrews, and B. L. Evans, Adaptive resource allocation
in multiuser OFDM systems with proportional rate constraints, IEEE
Transactions on Wireless Communications, vol. 4, no. 6, pp. 27262737,
2005.
[5] G. Yu, Z. Zhang, Y. Chen, P. Cheng, and P. Qiu, Subcarrier and bit
allocation for OFDMA systems with proportional fairness, in Proc. of
IEEE WCNC 2006, vol. 3, 2006, pp. 17171722.
[6] W. Yu, G. Ginis, and J. M. Cioffi, Distributed multiuser power
control for digital subscriber lines, IEEE Journal on Selected Areas
in Communications, vol. 20, no. 5, pp. 11051115, 2002.
[7] G. Kulkarni, S. Adlakha, and M. Srivastava, Subcarrier allocation and
bit loading algorithms for OFDMA-based wireless networks, IEEE
Transactions on Mobile Computing, vol. 4, no. 6, pp. 652662, 2005.
[8] Z. Han, Z. Ji, and K. J. R. Liu, Non-cooperative resource competition
game by virtual referee in multi-cell OFDMA networks, IEEE Journal
on Selected Areas in Communications, vol. 25, no. 6, pp. 10791090,
2007.
[9] T. Weiss, J. Hillenbrand, A. Krohn, and F. K. Jondral, Mutual interference in OFDM-based spectrum pooling systems, in Proc. of IEEE 59th
Vehicular Technology Conference (VTC 2004-Spring), vol. 4, 2004, pp.
18731877.
[10] R. Jain, D. Chiu, and W. Hawe, A quantitative measure of fairness
and discrimination for resource allocation in shared computer systems,
DEC (Digital Equipment Corporation) Research Report TR-301, 1984.
[11] M. Schniederjans, Goal programming methodology and applications.
Kluwer, 1995.
[12] D. Knuth, The Art of Computer Programming, Volume 3: Sorting and
Searching. Addison-Wesley, 1997.
[13] Y. Zhang and C. Leung, Performance of equal power subchannel
loading in multiuser OFDM systems, in Proc. of IEEE PACRIM 2007,
2007, pp. 526529.

S-ar putea să vă placă și