Professional Documents
Culture Documents
CR
CR
Sm
[p
m
|H
(p)
m
|
2
(1 S
m
)] = p
m
2
m
(1 E[S
m
|
S
m
]) T
0
(1)
where H
(p)
m
is the channel gain from the secondary base station to the primary receiver on the
m-th subcarrier,
2
m
= E(|H
(p)
m
|
2
) is the path loss from the secondary base station to the primary
receiver on the m-th subcarrier, p
m
is the transmit power of the secondary base station on the
m-th subcarrier and
S
m
= { s
k,m
|k {1, K}}.
C. System Utility for the Secondary System
In this paper, we shall consider average weighted sum throughput as the performance measure
of the secondary system. Let r
k,m
be the instantaneous data rate of k-th secondary user on the
m-th subcarrier. Let
P = {p
k,m
0|k {1, K}, m {1, M},
m
p
k,m
P
0
} (2)
be the set of power allocation for all K users in all M subcarriers (where P
0
is the total transmit
power constraint) and A = {(A
1
, ...A
M
)|A
m
{1, K}} be the set of user assignment to the M
subcarriers where A
m
denotes the user index assigned to the m-th subcarrier. Since the secondary
base station is assumed to have perfect CSI knowledge, the largest possible achievable data rate
r
k,m
is given by the mutual information between the transmitted symbol X
k,m
and the received
symbol Y
k,m
:
r
k,m
= log
2
(1 +p
k,m
k,m
)
where
k,m
= |H
k,m
|
2
. As a result, the total instantaneous weighted throughput is given by:
U(A, P, S, H) =
M
m=1
S
m
Am
r
Am,m
=
M
m=1
S
m
Am
log
2
(1 +p
Am,m
Am,m
) (3)
where
k
is the weight of the k-th secondary user, H = {H
k,m
|k {1, K}, m {1, M}} and
S
m
indicates whether the subcarrier is available or not.
2
In practice, we consider the situation where primary users can transmit on all potential subcarriers. For example, in IEEE
802.22 scenario, the primary users may be Part 74 devices (such as wireless microphone) can operate on any 200kHz BW on
the entire UHF/VHF spectrum. Hence, it is natural to consider common average interference constraint over all subcarrier.
February 14, 2008 DRAFT
8
Since the secondary base station has knowledge on the CSI H and the RSI
S only as well.
As a result, we shall formally dene the power allocation policy P and the subcarrier allocation
policy A below.
Denition 1 (Power Allocation Policy): The power allocation policy P = {P(H,
S) : H
C
KM
,
S {0, 1}
KM
} denes the power control actions P for any feasible realization of CSI
H and RSI
S such that
k,m
p
k,m
P
0
(4)
Denition 2 (Subcarrier Allocation Policy): The subcarrier allocation policy A = {A(H,
S) :
H C
KM
,
S {0, 1}
KM
} denes the subcarrier allocation actions A for any feasible
realization of CSI H and RSI
S.
Given the power allocation policy P and subcarrier allocation policy A, the average system
throughput U(P, A) is given by:
U(P, A) = E
S,S,H
[U(A, P, S, H)] = E
S,S,H
_
M
m=1
S
m
Am
log
2
(1 +p
Am,m
Am,m
)
_
(5)
IV. JOINT CROSS-LAYER SCHEDULING FOR OFDMA-BASED CR SYSTEMS
In this section, we shall rst give the problem formulation, and then derive a centralized
solution and the asymptotic performance.
A. Problem Formulation
The cross-layer design problem is equivalent to selecting the optimal power adaptation and
subcarrier allocation policies P and A so as to maximize the average weighted system throughput
U subject to the total transmit power constraint in (2) and the average interference constraint in
(1). This is cast into the following problem.
Problem 1 (Joint Cross-Layer Optimization Problem):
U
= max
P,A
U(P, A) (6)
subject to the total transmit power constraint in (2) and the average interference constraint in
(1).
February 14, 2008 DRAFT
9
From (5), we dene
U to be the conditional system utility which is given by
U(P, A, H,
S) = E
S
[
M
m=1
S
m
Am
log
2
(1 +p
Am,m
Am,m
) |
S, H]
=
M
m=1
Am
log
2
(1 +p
Am,m
Am,m
) (7)
where
m
= E[S
m
||
S
m
|] = Pr[S
m
= 1||
S
m
|] =
(1 q
p
) Pr[|
S
m
||S
m
= 1]
(1 q
p
) Pr[|
S
m
||S
m
= 1] +q
p
Pr[|
S
m
||S
m
= 0]
(8)
and |
S
m
| denotes the number of
1
in the set
S
m
. In fact,
m
represents the condence level that
the BS believes the m-th subcarrier is available. Since a policy is dened to be the set of actions
for each realization of CSI H and RSI
S
[
U(.)], maximizing the average weighted system throughput U with respect to the policies
P and A is equivalent to maximizing the conditional average weighted system throughput
U(.)
(7) for each realization of CSI and RSI. Hence, we have the following equivalent optimization
problem:
Problem 2 (Equivalent Optimization Formulation): Given a CSI realization H and a RSI re-
alization
S, nd the optimal user selection action A and the optimal power allocation action P
such that the conditional average system utility
U(P, A, H,
=max
P,A
M
m=1
Am
log
2
(1 +p
Am,m
Am,m
)
=max
P
M
m=1
max
Am
_
Am
log
2
(1 +p
Am,m
Am,m
)
_
(9)
s.t.
M
m=1
p
Am,m
P
0
(10)
p
Am,m
2
m
(1
m
) T
0
m {1, M} (11)
February 14, 2008 DRAFT
10
B. Centralized Solution
In this part, we shall derive a centralized algorithm to solve the Problem 2, which involves
both real and combinatorial variables, using dual optimization techniques. Specically, we dene
a dual problem of Problem 2 as:
Problem 3 (Dual Problem):
d
= min
,
g(, ) = min
,
_
M
m=1
_
max
Am
[ max
p
Am,m
l
Am,m
(p
Am,m
,
Am,m
,
m
)] +
m
T
0
_
+P
0
_
s.t. 0
m
0 m {1, M} (12)
where
l
Am,m
(p
Am,m
,
Am,m
, ,
m
) =
m
Am
log
2
(1 +p
Am,m
Am,m
) p
Am,m
m
p
Am,m
2
m
(1
m
)
(13)
and as well as is the Lagrange multiplier.
The dual problem can be decomposed into three subproblems:
Subproblem 1 (Power Scheduling): Given the multipliers (, ), for any A
m
{1, K} and
any m {1, M} optimize the transmit power p
Am,m
such that l
Am,m
(13) is maximized. Denote
p
Am,m
and l
Am,m
be the optimized variable, that is:
p
Am,m
(
Am,m
, ,
m
) = arg max
p
Am,m
l
Am,m
(p
Am,m
,
Am,m
, ,
m
)
=
_
m
Am
[ +
m
2
m
(1
m
)] ln 2
1
Am,m
_
+
and
l
Am,m
(
Am,m
, ,
m
) = l
Am,m
(p
Am,m
,
Am,m
, ,
m
)
=
_
_
0 p
Am,m
= 0
Am
log
2
_
m
Am
Am,m
[+m
2
m
(1m)] ln 2
_
m
Am
ln 2
+
+m
2
m
(1m)
Am,m
otherwise
where [x]
+
= max{0, x}.
Subproblem 2 (User Selection): Given the optimized l
Am,m
(.) where m {1, M} and A
m
{1, K}, nd the optimal user of each subcarrier, thus,
A
m
(,
m
) = arg max
Am
l
Am,m
(
Am,m
, ,
m
) (14)
February 14, 2008 DRAFT
11
and
l
m
(,
m
) = max
Am
l
Am,m
(
Am,m
, ,
m
) (15)
Subproblem 3 (Multiplier Update): Find the optimal multipliers (, ) which minimize the
dual function g(.), thus,
{
} = arg min
0
0
M
m=1
[l
m
(,
m
) +
m
T
0
] +P
0
(16)
We shall adopt the subgradient algorithm to solve for the optimal multipliers (
). Specically,
the subgradient of the dual function g(, ) in (12) is given by the vector:
(, ) =
_
_
_
_
_
_
_
P
0
M
m=1
p
m
,m
(,
m
)
T
0
2
1
(1
1
)p
1
,1
(,
1
)
.
.
.
T
0
2
M
(1
M
)p
M
,M
(,
M
)
_
_
_
_
_
_
_
(17)
As a result, the iterative subgradient search is given by
(n+1)
= [
(n)
s
n
(
(n)
,
(n)
)]
+
(18)
where
(n)
= [
(n)
,
(n)
1
, ...,
(n)
M
]
T
is the vector of Lagrange multipliers during the n-th iteration
and {s
n
} is a sequence of scalar step size. The iterative algorithm terminates when |g(
(n+1)
)
g(
(n)
)| where is a terminating threshold. The subgradient update is guaranteed to converge
to the optimal multipliers (
) as long as s
n
is chosen correctly[15]
3
.
One important issue is whether solving the dual problem is equivalent to solving the original
primal problem(or whether the duality gap is zero[16]). Notice that the optimization problem in
(9) includes both the real and combinatorial variables, and hence, the duality gap p
may
not be always zero for general weights {
1
, ..,
K
} in the system utility
4
. In Lemma 1, we shall
summarize that under a similar condition as in [17], the duality gap p
= 0 for general
weights {
1
, ..,
K
}.
Lemma 1: (Duality Gap for the Mixed Real and Combinatorial Problem) Suppose each pri-
mary user will span 1/Q of the total spectrum. For sufciently large M such that min(M/Q, M/L)
, we have the duality gap between (9) and (12) p
0.
3
In this paper s
n
is chosen as
n
where is a constant.
4
If we solve the optimization problem in (9) by solving for A and then solve the optimization with respect to power p (which
is a real variable), the power optimization problem may not be a convex problem for any utility weights {1, 2, ..., K}
February 14, 2008 DRAFT
12
Proof: This result is an extension of [17] and can be proved using similar techniques. We
shall omit it here due to the page limit.
The condition in Lemma 1 is quite mild and can be satised in most practical systems. For
example, we have M = 2048, Q 30 (FCC Part 74 device) and L 6 in WRAN (IEEE
802.22) systems with B = 6MHz bandwidth. In the following, we shall refer every adjacent
min(M/Q, M/L) subcarriers to one independent subband. All the subcarriers within one inde-
pendent subband have the highly correlated channel fading and primary activities.
C. Asymptotic Throughput Performance
In this part, we shall discuss the asymptotic performance of the joint sensing and cross-layer
scheduling design. We consider sum-rate utility (
k
= 1 k {1, K}) as an example for
illustration. Problem 2 can be reduced to
max
P
M
m=1
m
log
2
(1 +p
m
m
) (19)
s.t.
M
m=1
p
m
P
0
(20)
p
m
2
m
(1
m
) T
0
m {1, M} (21)
where
m
= max
Am
Am,m
and p
m
= p
A
m
,m
to simplify the notations.
(1) Analysis for Large P
0
: Intuitively, the average system throughput
U is an increasing
function of the total power constraint P
0
. However,
U will saturate at high SNR P
0
due to the
interference constraint (21). We shall derive the closed-form of the saturated throughput in this
part. First, we introduce the following lemma:
Lemma 2 (Asymptotic System Throughput for High SNR): When the total SNR P
0
is suf-
ciently large, i.e.
P
0
M
m=1
T
0
2
m
(1
m
)
, (22)
the optimal power for each subcarrier is given by:
p
m
=
T
0
2
m
(1
m
)
m (23)
February 14, 2008 DRAFT
13
and the average system throughput approaches:
U =
M
m=1
K
n=0
Pr(|
S
m
| = n)
_
K
m
ln 2
K1
k=0
(1)
k1
_
K 1
k
_
e
(k+1)(1m)
2
m
T
0
k + 1
Ei[
(k + 1)(1
m
)
2
m
T
0
]
_
(24)
where Ei(.) is the exponential integral,
m
depends on |
S
m
| and
Pr(|
S
m
| = n) =
_
K
n
_
(1 q
f
)
n
q
Kn
f
(1 q
p
) +
_
K
n
_
(1 q
d
)
n
q
Kn
d
q
p
(25)
Proof: Please refer to Appendix A.
Figure 3 illustrates the average system throughput U versus q
p
for different mobile sensing
accuracy. We observe that the average system througput is a monotonic decreasing function of
q
p
and better mobile sensing accuracy leads to better throughput performance.
Analysis for Large K: In this part, we shall analyze the asymptotic performance of our
joint design when the number of mobiles K tends to innity. We rst introduce the following
Proposition:
Proposition 1: When the number of mobiles K tends to innity, the spectrum sensing become
perfect: (1) when there is no primary transmission on one subcarrier (say the m-th subcarrier),
m
1 when K + (26)
(2) when there exists primary transmission on one subcarrier (say the m-th subcarrier),
m
0 when K + (27)
Proof: Please refer to Appendix B.
As a result, the base station has perfect sensing on each subcarrier for sufciently large K
and we have the following lemma on the asymptotic throughput analysis for large K.
Lemma 3 (Asymptotic System Throughput for Large K): For sufciently large K, the average
system throughput per channel use (b/s/Hz) has the order of growth O[(1 q
p
) log
2
ln K].
Proof: Please refer to Appendix C.
From this lemma, we observe that despite the false alarm and mis-detection probabilities,
the system throughput enjoys a multi-user diversity gain of log log K and is proportional to the
percentage of idle period (1 q
p
) in the primary system.
February 14, 2008 DRAFT
14
V. DISTRIBUTIVE CROSS-LAYER IMPLEMENTATION
In this section, we shall derive a distributive implementation on the cross-layer solution using
primal-dual decomposition [18].
A. Distributed Algorithm based on Decomposition
In the previous section, the dual problem 3 is decomposed into three subproblems, where
the subproblem 1 can be calculated in each user and the remaining two subproblems should be
solved at the base station. Hence, we have the following distributed algorithm.
Algorithm 1: (Distributed Algorithm)
1) In the sensing slot, each user performs spectrum sensing on each subcarrier, and then,
determine whether each subcarrier is occupied by the primary transmitter.
2) In the control slot, each user (say the k-th user) feeds back the RSI { s
k,m
|m {1, M}}
to the base station on a control channel. The base station rst calculates {
m
|m
{1, M}} according to (8) and then broadcasts {
m
} as well as the initial multiplier
(1)
= [
(1)
,
(1)
1
, ...,
(1)
M
]
T
to the mobiles.
3) In the n-th iteration, each user (say the user k) performs the optimization described in
the subproblem 1 (section IV-B), and feeds back {l
k,m
|m {1, M}} as well as the
corresponding optimized power {p
k,m
|m {1, M}} to the base station.
4) The base station selects the user for each subcarrier according to (14), and update the
multiplier
(n)
to
(n+1)
according to (17) and (18).
5) If the difference |g(
(n+1)
)g(
(n)
)| is less than a threshold then terminate the algorithm;
otherwise, jump to step (3).
Note that the distributive algorithm offers a few advantage from implementation perspective.
Firstly, it offers exibility as the base station does not need to know about the utility function
of the mobiles in the system. Hence, the scheduling operation is completely autonomous as
new users join and quit their sessions. Moreover, the computational loading can be shared by
the mobile users and hence, the complexity order of the distributive algorithm is O(ML) versus
O(KML) of the centralized solution. Furthermore, since the objective calculation is done locally
at the mobile side, the distributive implementation facilitates reduction of CSIT feedback as we
shall illustrate in the next section. On the other hand, one important issue about the distributive
February 14, 2008 DRAFT
15
algorithm is on the number of iterations between the master and the subproblems. In our problem,
the condition number of the Hessian of the objective function is small and hence, the subgradient
search converges very fast. As we shall illustrate in the results section, only a small number of
iterations are needed for signicant performance gain over isolated design.
B. Distributed Algorithm with Reduced Feedbacks
In algorithm 1, the feedback overhead grows linearly with the number of users K, hence, the
feedback overhead becomes very large when the system has a large number of users. In this
section, we shall propose a generalized threshold-based mechanism to reduce the feedback over-
head. Threshold-based feedback has been proposed in [19] for cross-layer scheduling to reduce
the feedback overheads. In this paper, we shall extend the idea to consider a generalized feedback
threshold where the threshold is compared with a generalized objective function at a user (rather
than CSI). Specically, we assume there are N QoS classes with weights [
1
,
2
, ...,
N
] and
percentage of users in the N QoS classes are [R
1
, R
2
, ..., R
N
]. Without loss of generality, we
assume
1
2
...
N
. The proposed algorithm is illustrated in Figure 4 and is summarized
below.
Algorithm 2: Distributed Algorithm with Generalized Threshold:
1) In the sensing slot, each user performs spectrum sensing on each subcarrier, and then,
determine whether each subcarrier is occupied by the primary transmitter.
2) In the rst iteration, the base station broadcasts the initial {
m
} where
1
=
2
= ... =
M
= , initial multiplier
(1)
= [
(1)
,
(1)
1
, ...,
(1)
M
]
T
where
(1)
1
= ... =
(1)
M
as well as a
system feedback threshold to the K mobiles.
3) In the rst iteration, each user (say the user k) performs the optimization described in the
subproblem 1. Each user attempts to feedback the local objective value {l
k,m
|m {1, M}}
only for those subcarriers l
k,m
. In the rst iteration, the user will also feed back
the optimized power {p
k,m
|m {1, M}, l
k,m
} and the sensing report { s
k,m
|m
{1, M}, l
k,m
} to the base station.
4) For m = 1 : M, the base station picks up one best user for each of the N QoS classes (user
with the largest l
k,m
among the QoS class) and notify the selected users. Furthermore, the
base station updates the multiplier
(1)
to
(2)
according to (17) and (18).
February 14, 2008 DRAFT
16
5) In the n-th iteration (n > 1), for each subcarrier only the mobiles notied in the step (4)
(say the user k) performs the optimization described in the subproblem 1, and feeds back
objective values l
k,m
and the corresponding optimized power p
k,m
to the base station.
6) The base station selects the user for each subcarrier according to (14), and update the
multiplier
(n)
to
(n+1)
according to (17) and (18).
7) If the difference |g(
(n+1)
)g(
(n)
)| is less than a threshold then terminate the algorithm;
otherwise, jump to step (5).
Clearly, in the above algorithm, the feedback overhead of the rst iteration depends on the
system feedback threshold , and the feedback overhead of the subsequent iterations is only N
which is usually much smaller than K. Hence, the feedback overhead of this algorithm is much
smaller than Algorithm 1. In general, this algorithm is suboptimal in throughput performance
because (1) not all the mobiles will give the sensing feedback and this may jeopardize the sensing
performance compared with Algorithm 1; (2) its possible that some subcarriers are abandoned
as there may be no feedback on these subcarriers due to the overlarge . We refer the case
where no user gives feedback on certain subcarrier/subcarriers as feedback outage. Obviously,
there is a tradeoff between feedback outage probability and feedback overhead. However, when
K is sufciently large, we can achieve both innitely small feedback outage and innitely small
feedback overhead per user.
Lemma 4 (Asymptotic Feedback Overhead for N-class CR systems): Given any sensing ac-
curacy at the mobiles (q
d
> 0, q
f
< 1) and for sufciently large K, the proposed algorithm 2
for N-classes of users achieves optimal average throughput performance, wherein the system
threshold is given by
= ln ln
_
K
ln K
_
for some constant [
1
,
N
], and the corresponding average feedback overhead (total number
of feedback attempts) per independent subband is O(ln(K)), where the independent subband
means the aggregation of adjacent subcarriers having highly correlated channel fading and
primary user activities.
Proof: Please refer to Appendix D.
As a result, for any non-degenerating sensing sensitivity at the mobiles (q
f
, q
d
), the proposed
algorithm is asymptotically optimal for large K with average feedback cost (number of feedback
February 14, 2008 DRAFT
17
attempts per user per independent subband) given by O(ln(K)/K) which tends to 0 for large
K.
VI. SIMULATION RESULTS AND DISCUSSIONS
In this section, we shall compare our joint sensing and cross-layer scheduling design with
several baseline references. Baseline 1 refers to the isolated design where the HSI per subband
is generated from the RSI and the cross-layer scheduler obtains the optimal power and subcarrier
allocation based on the CSI and the HSI. In baseline 2, we consider a reference system with
round robin scheduling only. In our simulation, we consider two QoS classes (N = 2) with the
weights
1
= 1 and
2
= 0.5 respectively. The number of subcarrier is M = 2048 and the
number of independent subband is 4. The average interference constraint T is 0dB. Each point
in the gures is obtained by averaging over 2000 independent fading realization.
A. Joint Design versus Isolated Design
In Figure 5, we compares the average system throughput versus SNR of our proposed joint
design and baseline 1 and 2 at moderate and high primary user activities q
p
= [0.5, 0.8]. The
simulations are done at q
d
= 0.5, q
f
= 0.2 and K
1
= K
2
= 10, where K
1
and K
2
are the number
of users belong to the rst and second QoS classes respectively. In the gure, we can see that
the average weighted throughput of the isolated design becomes saturated much earlier than the
joint design. For example, the isolated design saturates at P
0
/N
0
= 4dB and 2dB at q
p
= 0.3
and 0.8 due to the average interference constraint. However, the joint design continue to increase
with SNR (and shall saturate at a higher SNR). This signicant performance gain demonstrates
that the joint design could exploit the gaps in the primary user activity prole more effectively.
In fact, the power allocation in the isolated design follows the traditional water-lling approach
(only on the subcarriers which are sensed to be clean). Hence, large power will be allocated to
the clean subcarriers if the CSIT is strong. However, if in case the sensing is wrong, such
power allocation will induce potentially large interference to the primary users. Moreover, the
isolated approach will be too conservative in order to satisfy the interference constraint to the
primary user and it failed to exploit the temporal primary user activity gaps. On the other hand,
in the combined design, we will try to allocate power to a subcarrier by carefully weighting
the effect of CSIT and how likely this subcarrier is occupied. E.g., the power allocated on
February 14, 2008 DRAFT
18
a subcarrier may be small even though the subcarrier may have good CSIT when the sensing
results indicate high likelihood of being occupied by primary users. Hence, the joint design can
exploit the activity of the primary user much better, which leads to signicant performance gain.
Furthermore, it can be observed that both baseline 1 and joint design perform much better than
the baseline 2 and this demonstrates the signicant cross-layer gain.
Moreover, the complexity of the proposed combined design is very similar with that of the
isolated design. This is because we have obtained closed form expression for the power allocation
weighting factor due to sensing in the (8). Table I also illustrates the complexity comparison
(from matlab proler) between the two schemes, where we have included the computation for
power, rate, subcarrier selection as well as the processing of sensing information and the numbers
indicate the CPU time spent (in matlab) in processing a frame of data.
We have also done simulations (with K
1
= K
2
= 10, q
f
= 0.2, M = 4, q
d
= 0.5, q
p
= 0.5)
and veried that only very small iteration numbers (< 6 iterations) is needed to achieve over
92% of the optimal throughput of joint design as illustrated in Table II. In other words, using
small number of iterations already could achieve signicant gain w.r.t. the isolated design.
B. Asymptotic Performance for Large P
0
or K
Figure 6 illustrates the average throughput performance versus the primary user activities at
q
f
= [0.2, 0.3, 0.4], q
d
= 0.5, K
1
= K
2
= 5 at sufciently large SNR P
0
5
. The simulation results
match the closed-form expression derived in Lemma 2 closely at high SNR. Furthermore, we can
also observe that the average throughput is a monotonically decreasing function of the primary
user activity q
p
and false alarm q
f
.
Figure 7 illustrates the the spectral efciency versus the total number of users K at P
0
= 10dB,
q
f
= 0.2, q
d
= 0.5 for low, moderate and high primary activity q
p
= [0.3, 0.5, 0.8] respectively.
It can be observed that the simulation results matches the closed-form expression closely, which
justies the ln ln(K) growing order of the average system throughput at sufciently large K at
various q
p
.
5
Note that under large P0, the system is interference limited, meaning that the base station is not able to utilize all the available
power and the system performance is limited by the interference requirement.
February 14, 2008 DRAFT
19
C. Tradeoff of Performance and Feedback Overhead in the Threshold-Based Algorithm
Figure 8 illustrates the average weighted throughput versus the average feedback overhead by
implementing algorithm 2 at different primary user activity q
p
= [0.3, 0, 5, 0.8] where P
0
= 10dB,
K1 = K2 = 10, q
f
= 0.2, q
d
= 0.5 and the initial is 1. For each q
p
, we use different
feedback thresholds to generate different average throughput performance with different feedback
overhead. The optimal average throughput for q
p
= [0.3, 0, 5, 0.8] is also marked in the gure as
a reference. It is observed that for all q
p
, the throughput performance of algorithm 2 is very close
to the optimal performance at an reduced average feedback overhead of 15. Compared to the
algorithm 1 (optimal) where the average feedback overhead is 80, the proposed threshold-based
algorithm saves up to 80% feedback overhead with negligible performance loss. On the other
hand, when average feedback overhead too small, the throughput performance of algorithm 2
will degrade accordingly, illustrating a natural tradeoff.
VII. SUMMARY
In this paper, we proposed a combined spectrum sensing and cross-layer scheduling design
for downlink OFDMA-based CR systems, where we utilize both the RSI and CSI directly in
the cross-layer OFDMA power and subcarrier adaptations to maximize the weighted system
throughput for the secondary system under an average interference constraint at the primary
receivers. Compared to the conventional isolated design, the secondary users in the CR system
could exploit the spectrum holes of the primary users more effectively and therefore, achieving
a signicantly higher system throughput. To obtain more design insights, we have derived a
closed-form expression for the average system throughput of the proposed design. We have
also proposed a distributive implementation where the power control is distributed to individual
mobiles to reduce the computational loading at the base station. Simulations illustrate that our
proposed design has a signicant performance gain over the isolated design. In addition, for any
non-degenerating sensing sensitivity at the mobiles (q
f
, q
d
), we can achieve both asymptotically
zero feedback overhead per user and optimal system throughput for sufciently large K.
February 14, 2008 DRAFT
20
APPENDIX A: PROOF OF LEMMA 2
Equation (23) is obvious since the constraint (20) is a superset of constraint (21). Hence, given
the power for each subcarrier, the average system throughput is give by:
U = E
H,
S
[
M
m=1
m
log
2
(1 +
m
T
0
2
m
(1
m
)
)] =
M
m=1
E
H,|
Sm|
_
m
log
2
[1 +
m
T
0
2
m
(1
m
)
]
_
(28)
Notice that the CDF of
m
is (1 e
x
)
K
, we have
U =
M
m=1
E
|
Sm|
_
K
m
ln 2
K1
k=0
(1)
k1
_
K 1
k
_
e
(k+1)(1m)
2
m
T
0
k + 1
Ei[
(k + 1)(1
m
)
2
m
T
0
]
_
(29)
Since |
S
m
| is a discrete random variable with probability mass function (25), we can obtain the
equation (24).
APPENDIX B: PROOF OF PROPOSITION 1
From (8), we have
m
=
1
1 +
qp
1qp
_
1q
d
1q
f
_
|
Sm|
_
q
d
q
f
_
K|
Sm|
=
1
1 +
qp
1qp
__
1q
d
1q
f
_
|
Sm|/K
_
q
d
q
f
_
1|
Sm|/K
_
K
We rst prove the case where there is no primary transmission on the m-th subcarrier. When
the number of users K tends to innity, |
S
m
|/K (1 q
f
), hence,
_
1 q
d
1 q
f
_
|
Sm|/K
_
q
d
q
f
_
1|
Sm|/K
=
_
1 q
d
1 q
f
_
1q
f
_
q
d
q
f
_
q
f
< 1
Since K tends to innity, we have (26). Following similar approach, its easy to prove (27).
APPENDIX C: PROOF OF LEMMA 3
When K tends to innity, the average system throughput is give by:
U = E
S,H
[
U] E
S,H
[
M
m=1
S
m
log
2
(ln K)] = M(1 q
p
) log
2
(ln K)
where the approximation is because K is sufciently large and
m
ln K with probability 1
for large K [20]. Notice that U gives the average system throughput per one OFDM symbol and
that each OFDM symbol includes M channel uses, the average system throughput per channel
user (b/s/Hz) is U/M = (1 q
p
) log
2
(ln K).
February 14, 2008 DRAFT
21
APPENDIX D: PROOF OF LEMMA 4
We rst introduce the following proposition:
Proposition 2: Let the target feedback outage probability be T(K). If T(K) satises T
1
K
(K)
1 when K +, for sufciently large K the feedback overhead per one subcarrier of the
proposed algorithm 2 in the rst iteration is K[1T
1
K
(K)], and corresponding system threshold
is ln ln[
1
1T
1
K (K)
] for some [
1
,
N
].
Proof: In the following proof, we simplify the notation l
k,m
to l
i
when the user k belongs
to the i-th QoS class, and simplify
k,m
to . Let q
i
be the probability that the mobiles of the
i-th QoS class give feedback on one subcarrier. Hence,
q
i
= Pr[l
i
() ] = Pr(
i
) = e
i
where
i
= l
1
i
() and l
1
i
(.) is the inverse function of l
i
(.). The last equality is because is
exponentially distributed. From the denition of T(K), we have
N
n=1
(1 e
n
)
Rn
= T
1
K
(K)
N
n=1
(1 R
n
e
n
) = T
1
K
(K)
N
n=1
R
n
e
n
= 1 T
1
K
(K)(30)
e
1
1 T
1
K
(K) and e
N
1 T
1
K
(K)
1
ln[
1
1 T
1
K
(K)
] and
N
ln[
1
1 T
1
K
(K)
]
where the rst two steps is because e
n
must be very small (T
1
K
(K) 1 for K tends to
innity). Notice that for large , we have
n
= e
n
. Hence,
1
ln ln[
1
1 T
1
K
(K)
]
N
ln ln[
1
1 T
1
K
(K)
] = ln ln[
1
1 T
1
K
(K)
]
where is certain constant between
1
and
N
. Notice that the average feedback overhead in
the rst iteration is given by
N
n=1
KR
n
e
n
, substitute (30) into
N
n=1
KR
n
e
n
we can get
the closed-form expression of average feedback overhead.
Let T(K) = 1/K, we get the feedback overhead per subcarrier is ln K and the feedback
threshold is ln ln[
K
ln K
]. Notice that the subcarriers in one independent subband have the same
channel gain and primary activity, the scheduling for each subcarrier within one subband is the
same. Hence, mobiles only need to give feedbacks on one subcarrier of a subband. Thus, the
feedback overhead per independent subband is also ln K.
The optimality on average system throughput is justied as follows. First of all, since the
feedback outage probability is 1/K 0, the effect of feedback outage can be neglected.
February 14, 2008 DRAFT
22
Secondly, notice that the number of sensing report is ln K. When K tends to innity, ln K
also tends to innity. Quoting the conclusion of Proposition 1, the calculated in the rst
iteration of Algorithm 2 is able to achieve the actual status S
m
. Hence, the performance of
Algorithm 2 will achieve that of Algorithm 1.
REFERENCES
[1] S. Haykin, Cognitive radio: brain-empowered wireless communications, IEEE Journal on Selected Areas in Communi-
cations, vol. 23, pp. 201220, Feb. 2005.
[2] P. Kolodzy et al., Next generation communications: Kickoff meeting, Proc. DARPA, Oct. 17, 2001.
[3] J. Mitola and G. Q. J. Maguire, Cognitive radio: making software radios more personal, IEEE Personal Communications,
vol. 6, pp. 1318, Aug. 1999.
[4] J. Mitola, Cognitive radio: An integrated agent architecture for software dened radio, Doctor of Technology, Royal Inst.
Technol. (KTH), Stockholm, Sweden, 2000.
[5] N. Devroye, P. Mitran, and V. Tarokh, Achievable rates in cognitive radio channels, IEEE Transactions on Information
Theory, vol. 52, pp. 18131827, May 2006.
[6] P. Mitran, N. Devroye, and V. Tarokh, On compound channels with side information at the transmitter, IEEE Transactions
on Information Theory, vol. 52, pp. 17451755, April 2006.
[7] P. Cheng, G. Yu, Z. Zhang, H. H. Chen, and P. Qiu, On the achievable rate region of gaussian cognitive multiple access
channel, IEEE Communications Letters, vol. 11, pp. 384386, May 2007.
[8] Draft Standard for Wireless Regional Area Networks Part 22: Cognitive Wireless RAN Medium Access Control (MAC)
and Physical Layer (PHY) specications: Policies and procedures for operation in the TV Bands, IEEE Std. 802.22,
P802.22/D0.2, November 2006.
[9] Q. Zhao, L. Tong, A. Swami, and Y. Chen, Decentralized cognitive MAC for opportunistic spectrum access in ad hoc
networks: A POMDP framework, IEEE Journal on Selected Areas in Communications, vol. 25, pp. 589600, April 2007.
[10] Q. Zhao, S. Geirhofer, L. Tong, and B. M. Sadler, Opportunistic Spectrum Access via Periodic Channel Sensing, to
appear in IEEE Transactions on Signal Processing, 2008, 2008.
[11] J. Huang, R. Berry, and M. L. Honig, Auction-based Spectrum Sharing, ACM Mobile Networks and Applications Journal
(MONET), vol. 11, pp. 405418, June 2006.
[12] S. Sharma and D. Teneketzis, An Externality-based Decentralized Optimal Power Allocation Scheme for Wireless Mesh
Networks, 4th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc Communications and
Networks, 2007. SECON 07., 2007.
[13] Y. Xing, C. N. Mathur, M. A. Haleem, R. Chandramouli, and K. P. Subbalakshmi, Dynamic Spectrum Access with QoS
and Interference Temperature Constraints, IEEE Transactions on Mobile Computing, vol. 6, pp. 423433, April 2007.
[14] Y. Chen, Q. Zhao, and A. Swami, Joint Design and Separation Principle for Opportunistic Spectrum Access, Fortieth
Asilomar Conference on Signals, Systems and Computers, 2006. ACSSC 06., pp. 696700, Oct. 2006.
[15] D. Bertsekas, Nonlinear Programming. Belmont, MA: Athena Scientic, 1999.
[16] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
[17] Y. Wei and R. Lui, Dual methods for nonconvex spectrum optimization of multicarrier systems, IEEE Transactions on
Communications, vol. 54, pp. 13101322, July 2006.
February 14, 2008 DRAFT
23
[18] D. P. Palomar and M. Chiang, Alternative decompositions and distributed algorithms for network utility maximization,
Proc. IEEE GLOBECOM, vol. 5, Nov. 2005.
[19] V. Hassel, M.-S. Alouini, G. E. Oien, and D. Gesbert, Rate-optimal multiuser scheduling with reduced feedback load and
analysis of delay effects, Proc. EUSIPCO-2005, September 2005.
[20] P. Viswanath, D. Tse, and R. Laroia, Opportunistic beamforming using dumb antennas, IEEE Transactions on Information
Theory, vol. 48, pp. 12771294, June 2002.
February 14, 2008 DRAFT
24
Centralized Hard
Decision Sensing
Cross-Layer
Module
Raw Sensing
Information (RSI)
from mobiles
CSI feedbacks
Hard Sensing
Information (HSI)
Power and
Subcarrier
Adaption
Joint Sensing and
Cross-Layer
Scheduling
Raw Sensing
Information (RSI)
from mobiles
CSI feedbacks
Power and
Subcarrier
Adaption
Isolated Design
Joint Design
Fig. 1. The block diagrams of joint design and isolated design (In the isolated design, the sensing modular passes HSI, which
means 1-bit sensing decision on each subband, to the cross-layer modular, and the scheduling module then performs traditional
power waterlling and subcarrier selection only on the set of clean subcarriers. On the other hand, in the joint design, the
system utilizes the RSI and CSI directly).
TABLE I
COMPLEXITY COMPARISON.
Joint Design Isolated Design
Scheduling processing time / frame 0.012ms 0.011ms
Centralized sensing processing time / frame 0.0013ms 0.0004ms
February 14, 2008 DRAFT
25
Secondary Cell
Secondary BS
Primary Receiver
Secondary Cell
Secondary BS
Primary Receiver
Primary Transmitter
Coverage of Primary Transmitter
Fig. 2. Illustration of secondary cell.
TABLE II
EFFECT OF ITERATIONS.
Number of iterations 4 5 6 7 8 9 10
Performance 72% 81% 92% 96% 98% 98% 99%
February 14, 2008 DRAFT
26
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
5
10
15
20
25
30
Probability of Primary Transmisson q
p
A
v
e
r
a
g
e
S
y
s
t
e
m
T
h
r
o
u
g
h
p
u
t
q
f
= 0.1; q
d
= 0.9
q
f
= 0.2; q
d
= 0.8
q
f
= 0.4; q
d
= 0.6
Fig. 3. Average system throughput U versus primary user activity qp at K = 10, M = 1, T = 0dB and sufciently large P0.
February 14, 2008 DRAFT
27
Start
Broadcast initial
and
(1)
Receive the
feedback from the
mobiles
Select mobiles for
each subcarrier &
Update and
Broadcast the user
selection, and
(2)
Receive the
feedback from the
mobiles
Update
(n)
to
(n+1)
|g(
(n+1)
)-
g(
(n)
)|<
End
Yes
No
Broadcast
(n+1)
n=n+1
Start
Receive initial and
(1)
Calculate p*k,m and l*k,m
Receive
(n)
from the BS
l*k,m<?
Feed back sensing
result, p*k,m and l*k,m
Calculate and feed back
p*k,m and l*k,m on the
selected subcarrier
n=n+1
Is there terminate
signal?
End
No
Yes
Yes
No
Selected mobile on
certain subcarriers?
Yes
No
Flowchart of the BS Flowchart of the Mobile
Fig. 4. The owchart of base station procedure and mobile user procedure in algorithm 2.
February 14, 2008 DRAFT
28
1 2 3 4 5 6 7 8 9 10
0
0.5
1
1.5
2
2.5
3
3.5
4
SNR P
0
/N
0
(dB)
A
v
e
r
a
g
e
W
e
i
g
h
t
e
d
T
h
r
o
u
g
h
p
u
t
(
b
/
s
/
H
z
)
Joint design; q
p
= 0.5
Isolated design; q
p
= 0.5
Joint Design; q
p
= 0.8
Isolated design; q
p
= 0.8
round robin scheduling, q
p
=0.5
round robin scheduling, q
p
=0.8
Fig. 5. Average weighted throughput U versus transmit SNR P0/N0 at K1 = K2 = 10, q
f
= 0.2, q
d
= 0.5, M = 4 and
T = 0dB. Isolated design refers to the design where the base station determines the availability of each subcarrier by majority
voting and performs power and subcarrier scheduling on the available subcarriers. round robin scheduling refers to the design
where the user for each subband is selected with equal probabilities (round robin algorithm)
February 14, 2008 DRAFT
29
0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Probability of Primary Transmisson q
p
A
v
e
r
a
g
e
S
y
s
t
e
m
T
h
r
o
u
g
h
p
u
t
(
b
/
s
/
H
z
)
Closedform, q
f
=0.4
Closedform, q
f
=0.3
Closedform, q
f
=0.2
Simulation, q
f
=0.2
Simulation, q
f
=0.3
Simulation, q
f
=0.4
q
f
=0.2
q
f
=0.3
q
f
=0.4
Fig. 6. Average system throughput U versus primary user activity qp at q
d
= 0.5, K = 10, M = 4, T = 0dB and sufciently
large P0.
February 14, 2008 DRAFT
30
100 150 200 250 300 350 400 450 500
1
1.5
2
2.5
3
3.5
4
4.5
5
Total Number of Users K
A
v
e
r
a
g
e
S
y
s
t
e
m
T
h
r
o
u
g
h
p
u
t
p
e
r
C
h
a
n
n
e
l
U
s
e
(
b
/
s
/
H
z
)
q
p
=0.3
q
p
=0.5
q
p
=0.8
Fig. 7. Average system throughput U versus the number of users K at P0 = 10dB, q
f
= 0.2, q
d
= 0.5, M = 4 and T = 0dB.
February 14, 2008 DRAFT
31
0 5 10 15 20 25 30 35
0.5
1
1.5
2
2.5
3
3.5
4
Average Feedback Overhead
A
v
e
r
a
g
e
W
e
i
g
h
t
e
d
T
h
r
o
u
g
h
p
u
t
(
b
/
s
/
H
z
)
q
p
=0.8
q
p
=0.5
q
p
=0.3
Optimal Throughput =3.5b/s/Hz at q
p
=0.3
Optimal Throughput=2.5b/s/Hz at q
p
=0.5
Optimal Throughput=1b/s/Hz at q
p
=0.8
Fig. 8. Average weighted throughput U versus average feedback overhead at P0 = 10dB, K1 = K2 = 10, q
f
= 0.2, q
d
= 0.5,
M = 4 and T = 0dB.
February 14, 2008 DRAFT