130 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO.

1, JANUARY 2008

Distributed Rule-Regulated Spectrum Sharing

Lili Cao and Haitao Zheng

Abstract—Dynamic spectrum access is a promising technique In this paper, we consider the problem of distributed collab-
to use spectrum efficiently. Without being restricted to any orative sharing where a group of wireless nodes agree to share
prefixed spectrum bands, nodes choose operating spectrum on- available spectrum and maximize a predefined system utility.
demand. Such flexibility, however, makes efficient and fair
spectrum access in large-scale networks a great challenge. Prior This maps to scenarios where a single service provider deploys
work in this area focused on explicit coordination where nodes a large number of wireless nodes and enforces collaboration
communicate with peers to modify local spectrum allocation, and agreements among nodes. For example, future WiMAX and/or
may heavily stress the communication resource. In this paper, WiFi access points share spectrum to provide city-wide wire-
we introduce a distributed spectrum management architecture less coverage [1]. Equipped with cognitive radios, these future
where nodes share spectrum resource fairly by making indepen-
dent actions following spectrum rules. We present five spectrum access points can utilize multiple non-consecutive channels
rules to regulate node behavior and maximize system fairness concurrently to offer high-speed wireless access. Assuming
and spectrum utilization, and analyze the associated complexity nodes can sense and discover locally-available spectrum, we
and overhead. We show analytically and experimentally that the reduce the problem of spectrum sharing into a constrained
proposed rule-based approach achieves similar performance with channel allocation problem. We seek distributed solutions that
the explicit coordination approach, while significantly reducing
communication cost. efficiently and fairly distribute spectrum channels to nodes
such that any conflicting nodes are assigned with orthogonal
Index Terms—Cognitive radio, spectrum access, rule-based, channels. We model the requirement of fairness using the
distributed, poverty line, fairness.
proportional fairness [20] based system utility.
In this context, prior work [6] proposed a distributed
I. I NTRODUCTION algorithm using explicit communication. In this approach,
nodes self-organize into coordination groups on-demand and

W IRELESS innovation and deployment has been stifled

by the shortage of radio spectrum [13]. The shortage
problem comes from the current (and historical) spectrum
adjust spectrum allocation within each group to improve local
system utility. By regulating coordination procedures, this
approach converges quickly and produces spectrum alloca-
management policies that assign spectrum bands statically to tions that closely approximate the optimal solution. How-
wireless technologies in long-term licenses. While utilization ever, this approach also requires nodes to communicate and
of assigned spectrum can be as low as 5-10% [13], new exchange coordination message frequently. This explicit co-
technologies in search of market availability such as WiFi ordination model requires real-time reliable communication
are forced to utilize unlicensed spectrum bands, resulting in among nodes, which is not always available.
crowded spectrum utilization and unpredictable performance In this work, we propose an alternative distributed, collabo-
from aggressive deployments [2]. rative spectrum sharing approach with minimum coordination
The ideal and necessary solution to the artificial shortage overhead. Instead of exchanging coordination messages, nodes
problem is dynamic spectrum access. In this new model, act independently based on their local observation of spectrum
future wireless devices no longer operate on statically assigned condition. Nodes coordinate implicitly and regulate their ac-
spectrum, but acquire spectrum on-demand. Using cognitive tions by complying with a set of predefined rules [27]. In this
radios [8], [15], devices opportunistically utilize unused spec- case, while nodes act independently and tend to prioritize their
trum without disrupting operations of legacy spectrum owners, own performance, their compliance with the rules promotes
e.g. TV broadcasts. Peers share locally available spectrum to efficient and fair spectrum sharing. Compared to approaches
maximize spectrum utilization. with explicit coordination, this approach greatly simplifies im-
One key challenge in dynamic spectrum access networks plementation and significantly reduces coordination overhead.
is how to maintain efficient spectrum sharing among peers. This paper makes four contributions:
While maximizing spectrum utilization is the primary goal, we
also need good sharing mechanisms to provide fairness across • First, we design five rules that regulate nodes’ spectrum
wireless nodes. A node seizing spectrum without coordinating access to tradeoff fairness and utilization with communi-
with others can cause harmful interference to neighbors and cation cost and algorithm complexity.
hence reduce spectrum utilization. • Second, we prove analytically that a system under these
rules converges in finite iterations, and that each node has
Manuscript received March, 2007; revised September, 2007. Part of this a guaranteed amount of spectrum allocation.
work was published at IEEE DySPAN conference, November 2005. • Third, we design detailed algorithms to implement these
The authors are with the Department of Computer Science, rules and discuss practical issues.
University of California, Santa Barbara, CA 93106 U.S.A (e-mail:
{lilicao,htzheng}@cs.ucsb.edu). • Finally, we use extensive simulations to quantify the
Digital Object Identifier 10.1109/JSAC.2008.080112. impact of rules on spectrum access, while comparing
0733-8716/08/$25.00

c 2008 IEEE
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 131

allocation problem is known to be NP-hard [20]. In

this paper, we model the interference constraints using
the widely-used protocol interference model [11] and
represent interference constraints as a conflict graph. In
Section VIII, we also discuss practical considerations on
how to improve this model for more realistic characteri-
zation of interference.
• Fast system convergence: To support dynamic spectrum
access, nodes need to exercise an efficient allocation
algorithm to allocate resource in real-time, with only
local view of the network. In a distributed approach
where nodes independently adjust their spectrum usage,
the system must converge quickly and maintain stability.
Fig. 1. An illustrative scenario of distributed spectrum sharing. Access points
identify locally available channels that will not disrupt the operations of legacy B. Related Work on Spectrum Allocation
spectrum owners such as TV broadcasts. They also share these available There are multiple complementary approaches to allocate
spectrum with peers, and access spectrum to connect their subscribers.
spectrum, each designed for different scenarios. These in-
clude centralized allocation [23], [20], distributed game-theory
driven sharing [17], [19], [24] and distributed collaborative
implicit and explicit coordination approaches in terms of
sharing [6], [27].
efficiency and complexity. We also examine the impact of
In the centralized approach, a central server collects network
network topologies using both synthetic and real traces.
information and allocates channels to nodes to maximize a
The rest of the paper is organized as follows. We begin predefined system utility. The work in [4], [10], [21] proposed
in Section II with a brief discussion on spectrum allocation a central spectrum server model, and developed several heuris-
and existing solutions. In Section III we describe the spec- tics based centralized approximations for a limited number
trum allocation problem. Next in Section IV, we present the of users. The work in [20], [28] shows that the heuristic
rule-based spectrum allocation system and develop different based centralized algorithms perform similarly to the global
spectrum rules. We evaluate the system performance and optimum. While well-suited for small-scale networks, this ap-
complexity analytically in Section V. We then provide detailed proach suffers from lack of scalability in large-scale networks.
implementations of the proposed rules in Section VI. Next in In the game-theory driven approach, nodes compete to max-
Section VII, we conduct experiments to evaluate the proposed imize self benefits, e.g. competing wireless service providers
coordination based approaches. We discuss practical issues in access spectrum to maximize their service quality or rev-
VIII and conclude in Section IX. enue [10], [8], [18], [19], [21]. In this case, nodes act against
each other to maximize their own utility. These distributed
II. BACKGROUND AND R ELATED W ORK approaches are well-suited for scenarios where nodes are
In this section we introduce dynamic spectrum access sys- selfish and compete with each other.
tems, its key challenges and existing work. Figure 1 illustrates In the distributed collaborative sharing, nodes collaborate
a motivating scenario where a number of WiMAX/WiFi access to maximize a predefined system utility. Our prior work [6]
points share spectrum to connect their subscribers. These proposes a distributed collaborative approach using explicit
access points, hereby referred to as nodes, are equipped with coordination and communication. In this case, nodes self-
cognitive radios. They sense spectrum to identify usable chan- organize coordination groups, exchange control messages, and
nels without affecting operations of legacy spectrum users, determine the optimal spectrum allocation in each group to
such as TV broadcasts. For a given set of available channels, maximize the system utility. This approach is well-suited for
the spectrum allocation problem is to allocate channels to scenarios where nodes are controlled by a single entity who
access points so that conflicting access points are assigned can enforce a cooperation agreement among all the nodes [14],
with different channels. The optimal allocation maximizes a [22].
predefined system utility, e.g. proportional fairness or total In this paper, we focus on distributed collaborative ap-
spectrum utilization. proaches. However, our work differs from the explicit co-
ordination approach [6] because it does not require nodes
to communicate directly and negotiate spectrum allocations.
A. Challenges
Instead, nodes coordinate implicitly – they adjust self chan-
Solving the problem of spectrum allocation faces the fol- nel usage independently based on rules. By regulating node
lowing key challenges: behaviors using predefined rules, we show that this implicit
• NP-hard optimization under non-linear interference approach performs closely to that of the explicit approach
constraints: The optimization problem is subject to a while significantly reducing coordination overhead.
set of interference constraints. Nodes in close proximity
interfere with each other and can not use the same III. S PECTRUM A LLOCATION P ROBLEM
spectrum band while well-separated nodes can reuse In this section, we describe the problem of spectrum alloca-
the same spectrum. In general, the optimal spectrum tion using the scenario of WiMAX/WLAN access points. We
132 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

start with notations and assumptions of the network model, • Interference constraint – Let C = {cn,k |cn,k ∈
then define the channel allocation problem. {0, 1}}N ×N , a N by N binary matrix, represent the
interference constraints among nodes:


A. Assumptions 1, node n and k conflict with each other
cn,k =
As shown in Figure 1, we assume a network of N nodes 0, node n and k can reuse the same channel.
(access points) indexed from 0 to N − 1. Each node accesses Define the “neighbor”of node n as any node k with
N −1
spectrum to connect their subscribers, whose traffic is back- cn,k = 1, and d(n) = k=0 cn,k as the total number of
logged. We assume that the spectrum is divided into M non- conflicting neighbors of n.
overlapping orthogonal channels indexed 0 to M −1. Equipped
with cognitive radios, each node can communicate through • Conflict-free channel assignment – Let A =
multiple non-consecutive channels. We assume that each node {am,n |am,n ∈ {0, 1}}M×N , a M by N binary matrix,
transmits using a predefined combination of radio parameters represent a spectrum allocation:
(e.g. power and modulation), and focus on achieve conflict-

1, channel m is assigned to node n
free spectrum allocation by assigning conflicting nodes with am,n =
0, otherwise.
orthogonal channels1.
We assume nodes can reliably identify nearby legacy In addition, A satisfies all the constraints defined by C,
spectrum owners, and opportunistically utilize locally unused that is,
spectrum. In our model, each node observes a set of available am,n ·am,k = 0, if cn,k = 1, ∀ n, k < N, m < M. (1)
channels (defined by a channel availability matrix) that it can
use without disrupting legacy spectrum owners. C. The Optimization Problem
We assume nodes can reliably identify conflicting peers us-
ing interference detection techniques [22], [14]. We model the The problem of spectrum allocation is modeled by the
interference among nodes using the commonly-used protocol following constrained optimization problem:
model [11]. It models the interference condition as a pair-wise A∗ = argmax U (A), (2)
binary matrix – any two nodes either conflict and can not A
use the same channel simultaneously, or they do not conflict subject to Interference constraints defined by (1).
and can reuse the same channel. While this is a simplified U (.) represents the system utility function. In this paper, we
abstraction of the complex interference conditions, in practice, focus on the scenario where APs belong to a single entity
it makes the optimization problem tractable. In Section VIII such as an enterprise [14], [22]. The spectrum allocation
we discuss extensions to more complex interference models. is not driven by monetary gain but system-wide fairness.
Finally, we address the spectrum allocation problem from We use proportional fairness [16], [20], which is regarded
a MAC layer perspective. Because nodes (access points) are as a compromise between max-min fairness and maximum
static, we assume a pseudo-static interference environment throughput scheduling [12]:
without considering the impact of fast-scale channel fading.

N −1 
N −1 
M−1
In addition, we assume channels are homogeneous, i.e. they
Uf airness (A) = log RA (n) = log am,n · lm,n ,
offer the same bandwidth and have similar interference char-
n=0 n=0 m=0
acteristics. This assumption is well-suited for scenarios where (3)

nodes share a set of channels with similar carrier frequency. where RA (n) = M−1 m=0 a m,n · l m,n . represents the throughput
In the following, we assume that the channels offer an average that node n gets under assignment A.
bandwidth of 1, and the interference condition is uniform Because of the interference constraint and the non-linear
across all the channels. In Section VIII we also discuss optimization function, this optimization problem is known to
extensions to scenarios with heterogeneous channels. be NP-hard [20]. In this paper, we focus on heuristic based
distributed solutions. In the following section, we discuss the
B. System Model proposed rule-regulated spectrum allocation approach where
nodes adjust their spectrum usage A to maximize the system
We introduce the following notations to model the system: utility Uf airness , while subjecting to the interference con-
•Channel availability – Let L = {lm,n |lm,n ∈ straints.
{0, 1}}M×N , a M by N binary matrix, represent the
channel availability: IV. RULE -R EGULATED S PECTRUM S HARING


1, channel m available at node n, In the proposed system, nodes observe local conditions and
lm,n =
0, channel m occupied by a legacy user. neighbors’ actions and independently adapt their spectrum
M−1 usage. Their behavior is regulated by a set of rules defined by
Let L(n) = m=0 lm,n represent the total number of spectrum regulators. In contrast to the explicit coordination ap-
channels available at node n. proach [6], nodes tend to prioritize their own performance with
minimal regard to system utility. However, their compliance
1 While the proposed approach can be integrated with other parameter-
with the rules promotes efficient and fair spectrum sharing. In
tuning techniques such as power control, we will defer this investigation to a the following, we start by describing the high-level concept,
later study. and then discuss the detailed rule design.
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 133

unclaimed by conflicting peers. To provide fairness, the rules

limit the number of channels each node can access.
Rule A (Uniform Idle Preference): Each node adjusts its
spectrum usage to Ω = minn P L(n) number of idle channels.
Service providers can optimize the value of Ω for the entire
network. However, nodes experiencing intensive interference
from legacy nodes (small L(n)) or other peers in a crowded
area (large d(n)) can limit the the value of Ω, leading to less
than ideal spectrum utilization. Therefore, adapting Ω to each
node’s interference condition is preferred. Next we propose
a rule which gives each node n exactly P L(n) number of
channels.
Fig. 2. The proposed rule-regulated spectrum allocation. Each node observes Rule B (Poverty Exact Idle Preference): A node n selects
L(n)
spectrum usage status and identifies whether to adjust spectrum usage. If so, exactly P L(n) =  d(n)+1  idle channels. If the number of idle
they adjust based on a predefined spectrum rule.
channels < P L(n), it “seizes” channels from “richer” nodes
without affecting “poorer” nodes. A node conflicting with a
“poor” node will sense the conflict and give up the channel
A. System Overview
and switch to other channels following the same procedure.
Figure 2 illustrates the operation at each node to select To n, a neighbor is “richer” if it uses more channels than
channels. Each node n performs spectrum sensing to identify n; otherwise it is “poorer”. Rule B requires that each node
its spectrum usage, i.e. RA (n). Using spectrum rules, each has knowledge of the number of neighbors d(n), and the
node checks whether it needs to update its channel selections. channel selection of each neighbor in order to identify “richer”
If an update is required, nodes rely on rules to determine the nodes. To “grab” non-idle channels, a node n marks the
appropriate channels to use. channels occupied by “poorer” neighbors as busy, and the
The key challenge in this design is how to define the spec- rest as idle. node n then selects a set of channels from the
trum rules. The rules specify how many and which channels “idle” channels until its channel occupancy reaches P L(n).
a node should use, such that fairness and utilization can be The efficiency of grabbing depends on the set of channels
achieved. The estimation should not be overly aggressive and selected. In Section VI, we present the detailed procedure.
bring excessive contention, or overly conservative and result A limitation of Rule B is that each node only attempts
in spectrum under-utilization. Further, in a distributed system, to use P L(n) channels. Since P L(n) represents a lower
each node can only act based on limited local view of the bound on spectrum usage derived using a collaboration based
system. approach [6], Rule B could under-utilize available spectrum.
We design rules based on the observation from the collabo- Therefore, we devise an improved rule that guarantees the
rative spectrum sharing system with explicit coordination [6]. poverty line for each node while letting some nodes go beyond
In this prior work, we have shown that when the system their poverty lines.
converges, the number of channels a node obtains is bounded.
This lower bound, defined as “poverty line,” represents the Rule C (Poverty Guided Idle Preference): A node n selects
minimum amount of spectrum a node is entitled to. If a node channels from idle channels. Only if there are not enough idle
n has L(n) available channels and d(n) conflicting neighbors, channels to reach P L(n) does node n “grab” channels from
its poverty line is “richer” neighbors. The number of channels it can grab from
  any “richer” node r, is max{0, min{C(r)−P L(n), P L(n)−
L(n) C(n)}} where C(n) and C(r) are the current spectrum usage
P L(n) = .
d(n) + 1 of node n and r.
Results in [6] also show that the poverty line is very tight, Rule C allows nodes who have attained their poverty line
making it a relatively accurate estimation of the number of to seize additional idle channels. It still allows nodes below
channels a node should get in order to maximize fairness. In their poverty line to take channels from “richer” neighbors.
this paper, we propose to use this measure as the basis of However, each grabbing can not reduce a “richer” node’s
spectrum rules. spectrum below the grabber’s poverty line, avoiding cycles of
Next, we propose five different rules that tradeoff between nodes grabbing channels from each other in turn. In particular,
performance and complexity for different application scenar- a node n can collect all the channels used by its “richer”
ios. neighbors but not “poorer” neighbors into a channel pool,
reserve P L(n) channels for each “richer” neighbor and “grab”
from the rest of the pool. Similarly, the channel selection
B. Rules for Conflict Free Channel Assignment procedure impacts the efficiency of the grabbing, which will be
We begin by describing rules that lead to conflict free described in Section VI. We note that Rule C does not require
spectrum usage. Conflict free channel usage is one of the each node to have knowledge of its neighbors’ poverty line.
scheduling methods that allow for explicit and guaranteed We note that the performance of conflict free channel
throughput provisioning and control over packet delay. In assignments such as Rules B and C depends on the granularity
this case, nodes always select idle channels, i.e. channels of spectrum partition, i.e. the number of channels M . When
134 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

M is small compared to the number of neighbors d(n), V. A NALYTICAL B OUNDS

some nodes may have a poverty line of zero, and hence no In this section, we derive analytical bounds on the perfor-
performance guarantee. In this case, the system can increase mance and complexity of the proposed rules.
granularity by partitioning time, e.g. a channel is defined as a
frequency band at a particular time slot. A. Conflict-Free Rules
First, it is straightforward to show that Rule A provides a
conflict-free spectrum allocation.
C. Rules for Contention-based Channel Assignment Theorem 1: Rule A guarantees a conflict free spectrum allo-
cation. (Proof in appendix A).
Rule B and C require knowledge of neighbors’ spectrum Next, we illustrate the performance and complexity bounds
usage, which can be achieved by periodic broadcast. However, of Rules B and C. Assuming no two neighbors modify their
broadcasting spectrum usage to neighbors might be unde- spectrum/channel usage simultaneously, the system will reach
sirable for a number of reasons, including privacy concerns equilibrium after a finite number of iterations. Equilibrium
and protection against jamming from malicious nodes. For is the state where nodes have no incentive to adjust their
these reasons, we now describe two rules that do not require spectrum usage.
knowledge of neighbors’ spectrum usage.
Theorem 2: Using Rule B or C, the system reaches an
In this approach, on each channel, nodes follow a set equilibrium after an expected number of at most O(N 2 ) node
of random access rules such as CSMA to compete fairly spectrum modifications. In equilibrium, there is no conflict in
for channel access and avoid conflict. Each node performs spectrum usage, and each node’s spectrum usage is no less
contention detection, i.e. listens to the channel before initiating than its Poverty Line P L(n) (equal to P L(n) for Rule B).
any transmission. It initiates the transmission only when the (Proof in appendix B).
channel is idle for some given time. Otherwise, it backs off
and delays the action for a short period. Because channels B. Contention-based Rules
have different contention conditions, nodes should invoke
Theorem 3: Using Rule D or E, the system will reach an
independent contention detection and backoff process on each
equilibrium after at most Λ×M node spectrum modifications.
channel. The penalty of such random access is the overhead
Λ is bounded by O(N 2 ). (Proof in appendix C).
of contention detection even if there is only one node on the
The choice of Ψ and α depends on specific random access
channel.
mechanisms. To analyze their impact, we use a simple model
The rules specify the number of channels nodes should use to characterize channel sharing. A node contending with m
and how to select these channels. We propose the following other nodes on a channel gets 1/(λ · (m + 1)) of the channel
two rules depending on whether nodes have information about throughput, where λ is the contention penalty. We refer to this
their poverty line. model as the (λ, m) model. When λ = 1.8, this model matches
Rule D (Selfish Spectrum Contention): Each node n can the experimental test and analytical results for CSMA-based
use up to the Ψ channels that provide the highest throughput. IEEE 802.11b systems at 11Mbps in [9].
Communication on each channel is through CSMA based time We show that Rule D guarantees a lower bound on node
contention. throughput when |L(n)| = M (See Theorem 4). The bound
is only tight when Ψ = M , however, and cannot be used to
We note that the poverty line concept can provide a ref-
derive the optimal value of Ψ. Theoretical analysis of Rules
erence for choosing different value of Ψ for different nodes.
D and E is ongoing work.
Since the poverty line represents throughput attainable from
conflict free spectrum usage, Ψn should be larger than P L(n) Theorem 4: Using Rule D and (λ, m) model, a node
to account for channel contention. Note that in the random n’s spectrum usage is lower-bounded by LB(n, Ψ) =
1 Ψ
access scenario, P L(n) can still be computed using only the d(n) , (Ψ = 1); Ψ×d(n) , (1 < Ψ < M );
λ{ M +1} λ{ +2}
number of neighbors, which can be estimated by listening to M

λ{d(n)+1} , (Ψ = M ). (Proof in appendix D)

M
MAC control packets.
Rule E (Poverty Guided Selfish Spectrum Contention): This VI. RULE I MPLEMENTATIONS
rule is the same as Rule D except the number of channels each
In this section, we provide implementation details on how
node n can use is limited by Ψn = max(α · P L(n), 1), α ≥ 1.
each node applies the proposed rules to select channels. Nodes
Both rules encourage nodes to act selfishly. Nodes monitor execute these procedures after detecting a conflict or that its
channel conditions and switch to channels that provide the best actual usage is less than its poverty line. Table I summarizes
throughput, even if such a switch might reduce performance the information required at each node to execute each of the
for other neighbors. One question is how to choose the best five rules. They provide an intuitive measure of each rule’s
channels with maximum capacity and minimum contention. required communication complexity.
For the purpose of illustration, we use the number of compet-
ing nodes as an indicator of channel quality. Hence, following A. Implementing Rule B and C
Rule D or E, nodes always switch to channels with the We start with Rule B and C that lead to conflict free channel
least number of competing nodes. This also makes both rules assignment. Implementation of Rule B and C requires the
efficient. following two mechanisms:
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 135

TABLE I TABLE II
R EQUIRED I NFORMATION TO E XECUTE THE RULES C HARACTERIZATION OF CHANNELS

channel No. neighbors each neighbor’s channel Set Name Definition (Rule B) Definition (Rule C)
idle/busy dn channel usage utilization R Reserved Channels occupied channels occupied by
Rule A
by poor neighbors poor neighbors of n and
Rule B


of n the P L(n) smallest-
Rule C


ID channels of each
Rule D

“richer” neighbor of n
Rule E


O Conflicting Channels occupied by neighbors of n but
not in set R
I Idle Channels not occupied by any neighbor of
n
• Neighbor Spectrum Usage Discovery.
Rule B and C require knowledge of neighbors’ spec- Algorithm 1 Rule C for node n
trum usage. In particular, nodes (especially those below 1: Begin
their poverty line) need to know the set of channels 2: R ← Empty
3: O ← Empty
each neighbor currently occupies. This can be done by 4: I ← Empty
each node broadcasting their channel usage embedded 5: N ← Empty {New channel occupation}
in beacon broadcasts [26]. These broadcasts are simple 6: P ← P L(n) {Number of remaining channels to fill}
7: {Phase 1: Compute sets R, O, and I}
to implement, and their corresponding overhead is much 8: for every neighbor n
do
smaller compared to that of explicit coordination based 9: if |C(n
)| ≤ P L(n) then
approaches [6], [20]. An alternative is to use a sophisti- 10: R ← R ∪ C(n
)
11: else
cated spectrum analyzer. 12: R ← R∪ (the upper P L(n) channels in C(n
))
• Isolated Adjustment. 13: O ← O∪ (the rest of the channels in C(n
))
Rule B and C assume no simultaneous spectrum adjust- 14: end if
15: end for
ments by neighboring nodes. Since nodes can indepen- 16: O ←O\R
dently update their spectrum usage, a conflict occurs if 17: I ← channels not in R nor O
two neighbors simultaneously switch to the same idle 18: {Phase 2: Sequential channel occupation}
19: N ←I
channel. Therefore, after a node decides to switch to a 20: P ← P − |I|
new channel, it computes a short random wait time before 21: if P > 0 then
starting a transmission. If it detects activity on the channel 22: if P ≤ |O ∩ C(n)| then
23: N ← N ∪ (P channels in (O ∩ C(n)))
during the wait time, it marks the channel as busy, and 24: else
cancels the channel switch. 25: N ← N ∪ (O ∩ C(n))
26: P ← P − |O ∩ C(n)|
Next, we describe the detailed procedure. Using the conflict 27: N ← N ∪ (P channels in (O \ C(n)))
discovery module, each node n derives d(n), L(n) and poverty 28: end if
line P L(n). If n’s channel usage fells below P L(n) or upon 29: end if
30: Give up C(n) \ N ; Occupy N \ C(n)
observing a conflict, it adjusts spectrum usage in two phases.
31: End
• Phase I: Channel Classification.
In this phase, node n classifies its available channels
into three disjoint sets: reserved set R, conflicting set O, O ∩ C(n), and O \ C(n). For Rule C, the node seizes
and idle set I, defined in Table II. The reserved channel all channels in I, and if it hasn’t reached P L(n), then it
set includes all the channels that are not eligible for sequentially chooses its channel usage from O ∩ C(n),
n, preventing n from grabbing channels from “poorer” and O\C(n). During the channel selection, a node always
neighbors or over-grabbing from “richer” neighbors. The starts from the smallest channel IDs in a channel set. We
conflicting set and idle set identify all the channels list the detailed procedure in Algorithms 1 and 2.
eligible for n to grab. The channels in the idle set have the
highest priority since the corresponding “grabbing” does B. Implementing Rule D and E
not disturb any neighbor. It should be noted that before Essential to Rule D and E is the metric to determine
spectrum modification, the node n’s channel usage C(n) the “best” channels. In Section V, we use the number of
may intersect with R, O, and I; the cardinality of C(n) competing nodes as the channel quality indicator, assuming
could be below or above P L(n). each node’s traffic is backlogged. This metric can be inac-
• Phase II: Channel Selection. curate because in practice, nodes carry different and time-
In this phase, node n selects channels to use from sets I varying traffic load. Using IEEE 802.11 CSMA/CA devices,
and O. It always starts from the set I to avoid disturbing we propose to use busy time ratio (BTR) as the channel
neighbors. Only if |I| < P L(n) will it select from quality indicator [25]. Each node constantly monitors the
the conflicting set O. When selecting channels from a available channels, and records the percentage of channel
channel set (I or O), node n should select from its being busy. This new metric accounts for the impact of node
current channel occupation with higher priority, reducing contentions, traffic heterogeneity, transmission failures and
disturbance to neighbors and avoiding meaningless chan- retransmissions. Furthermore nodes in close proximity can
nel usage adjustments. For Rule B, the node sequentially perform collaborative sensing to reduce complexity and energy
chooses P L(n) channels from sets I ∩ C(n), I \ C(n), consumption associated with channel monitoring/sensing [25].
136 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

Algorithm 2 Rule B for node n 20 200

20 users 20 users
1: Begin

Utilization
Fairness
2: R ← Empty, O ← Empty, I ← Empty
3: N ← Empty {New channel occupation}
4: P ← P L(n) {Number of remaining channels to fill}
10 C 100 C
5: {Phase 1: Compute sets R, O, and I} B B
6: for every neighbor n
do A A
7: if |C(n
)| ≤ P L(n) then 0 0
8: R ← R ∪ C(n
) 0 500 0 500
9: else Topology Index Topology Index
10: O ← O ∪ C(n
) 200 40 users
11: end if 40 users
20

Utilization
12: end for

Fairness
13: O ←O\R
14: I ← channels not in R nor O C 100 C
10
15: {Phase 2: Sequential channel occupation} B B
16: if P ≤ |I ∩ C(n)| then
17: N ← N ∪ (P channels in (I ∩ C(n)))
A A
0 0
18: Goto 43 0 500 0 500
19: else Topology Index Topology Index
20: N ← N ∪ (I ∩ C(n))
21: P ← P − |I ∩ C(n)| Fig. 3. Comparison of of Rule A and B in terms of the spectrum utilization
22: end if (sum of total node throughput) and fairness (defined by Eq. (3)) assuming 20
23: if P ≤ |I \ C(n)| then channels and 20-40 nodes.
24: N ← N ∪ (P channels in (I \ C(n)))
25: Goto 43
26: else random network. We use this topology to examine the
27: N ← N ∪ (I \ C(n)) impact of conflict degree on the rule performance.
28: P ← P − |I \ C(n)|
• Real network trace. We extract a set of actual AP
29: end if
30: if R ≤ |O ∩ C(n)| then deployments using data traces collected by Placelab.
31: N ← N ∪ (P channels in (O ∩ C(n))) (http://www.placelab.org/).
32: Goto 43
33: else We simulate both small and large networks. We use small
34: N ← N ∪ (O ∩ C(n)) networks to compare the performance of implicit coordination
35: P ← P − |O ∩ C(n)| to that of centralized optimization. We use large networks
36: end if
37: N ← N ∪ (P channels in (O \ C(n))) to compare the performance of implicit coordination to that
38: Give up C(n) \ N ; Occupy N \ C(n) of explicit coordination. The real network trace contains 200
39: End nodes located in a 200m x 2000m area. Hence, we construct
random and clustered network in the same area with 200
nodes. For small networks, we deploy 20-40 nodes in an area
This allows each node to obtain status of entire channel set of 1000mx1000m, who contend for 20 channels.
by monitoring only a small fraction of the set. To avoid We design rules to maximize fairness utility, defined by
simultaneous channel adjustment at neighboring nodes, each Eq. (3). We evaluate the rule performance using both fairness
node applies a random wait scheme described earlier before and spectrum utilization2 . For randomly generated topologies,
starting transmission on the newly acquired channels. we run the experiments over 500 topologies. For Rule C, D
and E, the system has multiple equilibriums with different
VII. E XPERIMENTAL R ESULTS spectrum usage. We record the average performance over
In this section, we conduct experimental simulations to many equilibriums. By default, we assume λ = 1.8 for Rule
quantify the performance of the rule based spectrum access. D and E and set duty cycle to 1. By default, we assume no
We compare the proposed approach to those using explicit primary nodes.
coordination and centralized optimization. We compare the performance of Rule A, B, C, D and
E based implicit coordination schemes to that of explicit
A. Experiment Setup coordination [6] (marked as CA-EC) and centralized assign-
We assume a network of many WiFi and WiMAX access ment [20], [28] (marked as CA) schemes.
points. We assume each AP serves a large number of sub- B. Comparing Different Rules
scribers and has backlogged traffic. We use a simple binary
We start by examining the effectiveness of poverty line guided
interference model to construct the conflict graph – two nodes
design by comparing Rules A to B, C, and D to E. We assume
conflict if they are within distance of D. By default, we set
a random network topology.
D to 100m. We use this assumption to simplify the conflict
graph construction. However, it will not limit the scope of Rule A vs. B vs. C Figure 3 compares system utilities with
the proposed rule design. To examine the impact of network Rule A, B and C. Clearly, Rule B and C significantly outper-
topology, we use both randomly generated topologies and form Rule A. As the node density increases, the performance
measured AP deployment traces to validate our results. of the system using Rule A is limited by the node with the
worst conflict condition and the lowest poverty line. On the
• Random network. We place nodes randomly in an area.
• Clustered network. We simulate a hotspot scenario by 2 The spectrum utilization is the sum of total node throughput:
PN−1
deploying a set of nodes densely in a small area of the n=0 RA (n).
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 137

25 25

20 20

Fairness

Fairness
15 15 Rule E
Rule D
CA−EC
CA−EC

10 10
0 0.5 1 1 1.5 2 2.5 3
γ α
200 200

150 150
Utilization

Utilization
100 100
Rule D Rule E
CA−EC CA−EC

50 50
0 0.5 1 1 1.5 2 2.5 3
γ α

Fig. 4. Comparison of Rule D and E with the explicit coordination scheme. γ = Ψ/M , average results of 40 nodes sharing 20 channels.

200
Rule D. In the following, we will use α = 1.8 as the default
180 setting of Rule E. We leave the derivation of the optimal α to
CA
a future study.
Number of spectrum adjustments

160

140 Complexity In Figure 5, we measure the complexity of

120
different algorithms in terms of the total number of spectrum
adjustments before the system reaches an equilibrium. We use
100
a small network with 20 channels and 40 nodes. Compared
80 to the explicit coordination approach, Rule C provides more
CA−EC
60 than 60% reduction in complexity, and thus much faster
40
Rule E adaptation to topology changes. Rule E offers similar number
Rule B of adjustments as the explicit coordination approach.
20
Rule C
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 C. Comparing to Explicit Coordination
Culmulative Fraction
In Figure 6 we compare Rule B, C and E to CA and
Fig. 5. Comparison of centralized (CA), explicit coordination (CA-EC), Rule CA-EC, using small random networks. As a reference, we
B, C and E in term of complexity (the total number of spectrum adjustments),
plotted as the cumulative distributive function of 500 topologies. We assume also include the performance of centralized assignments that
20 channels shared by 40 access points. maximize spectrum utilization rather than system fairness.
Similarly, there is a noticeable performance gap between rules
and CA, CA-EC. This is as predicted because centralized
optimization or distributed explicit coordination can improve
other hand, Rule B and C use local poverty line and adapt
system performance at the cost of higher complexity. However,
gracefully to the node density. Rule C outperforms Rule B
we see that the proposed simple, implicit coordination leads
by allowing some nodes to obtain more spectrum than their
to a graceful degradation in both utilization and fairness.
poverty line.
In Figure 5, we also compare the complexity of different
Rule D vs. E Figure 4 compares the performance of Rule approaches. For the collaboration based approaches (CA and
D and E using two system utilities averaged over 500 random CA-EC), a 4-way handshaking is required for each spectrum
topologies under different γ and α values. We see that Rule adjustment, so that the message overhead is 4 times the
D is very sensitive to γ – small γ leads to under-utilization of number of adjustments. For Rule C, each node broadcasts its
spectrum while large γ (aggressive spectrum usage) results in spectrum usage to neighbors after the adjustment, resulting
excessive interference. On the other hand, Rule E is relatively in one message per adjustment. In terms of coordination
insensitive to α. Using α > 1.5, Rule E always outperforms overhead, Rule C leads to 240% reduction compared to the
138 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

240

220
CA designed to maximize 28

200 spectrum utilization

25
CA & CA−EC
Spectrum Utilization

180 CA & CA−EC

Proportional Fairness
160 20 Rule E (α=1.8)
Rule E(α = 1.8)
140 Rule C
Rule C
15
120

100
Rule B 10 Rule B
80

60 5
0 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 0 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
Cumulative Fraction Cumulative Fraction

(a) Utilization (b) Fairness

Fig. 6. Comparison of centralized (CA), explicit coordination (CA-EC), Rule B, C and E in term of system utilization and fairness utilities, plotted as the
cumulative distributive function of 500 topologies. We assume 20 channels shared by 40 access points.

explicit coordination approach. Because of its minor degrada- coordination divided by the one using Rule C (CA-EC/RuleC).
tion in system utility and significant reduction in complexity, All three topologies show that there are noticeable gaps at indi-
Rule C is a favorable low-complexity alternative to explicit vidual node’s spectrum assignment. We notice that some nodes
collaboration based approaches. For Rule E, if each node can get less spectrum assignments using explicit coordination in
measure channel utilization effectively, no message exchange order to improve the proportional fair based utility function.
is required. Given its low communication requirements, Rule E However, we confirm both explicit and implicit coordination
is also a low-complexity alternative that requires time-domain schemes always guarantee poverty line for all the nodes.
contention. System Utility. Figure 11 compares the system fairness
utility of Rule C and CA-EC for random, clustered and
D. Impact of Network Topology
measured networks with 200 nodes. We vary the conflict
Next, we examine the impact of network topology on node distance D from 20 to 200 meters. Without any explicit
spectrum assignments, system utility and complexity. Figure 7 communication, rule C leads to a minor degradation compared
illustrates three sample topologies corresponding to large scale to CA-EC. The difference increases with the conflict distance.
random, clustered and measured networks.
Complexity. Figure 12 plots the number of spectrum ad-
Node Spectrum Assignment. Figure 8 plots the number justments at each node for the system to reach an equilibrium,
of channels assigned to each node using Rule C. We observe versus the node’s conflict degree. Since our scheme always
that a node n’s spectrum usage is inversely proportional to its starts from an empty assignment, each node needs at least
conflict degree d(n), as predicted by the poverty line. one adjustments. We see that most nodes need only 2-3
To further investigate this dependency, we plot in Figure 9 adjustments, although there are 100 channels.
the amount of spectrum assigned at each node divided by its
Next, in Figure 13 we examine the number of spectrum
poverty line, as a function of the conflict degree d(n). Clearly,
adjustments required by CA-EC divided by those required
there are two types of nodes: one has assignments very close
by Rule C. We see that Rule C can reduce coordination
to their poverty line, and another has assignments much higher
complexity by 4-10 times. We note that the complexity of
than the poverty line. Nodes of the first type in general have
CA-EC scales with the number of channels. Overall, we see
clique-like local conflict graph. They are located in dense areas
that Rule C is highly favorable for networks with large number
where neighbors all conflict with each other. In this case, the
of channels.
poverty line is very tight. On the other hand, nodes of the
second type have star-like local conflict topology, and their
poverty line is loose. E. Impact of Primary/Legacy Devices
Figure 10 shows the cumulative distribution function of
the assignment using Rule C divided by its poverty line We now examine spectrum usage in the presence of pri-
(RuleC/PL) for all three types of topologies. We see that al- mary/legacy devices who own the spectrum. We randomly
most 70% of nodes obtain spectrum assignment similar to their deploy a set of primary devices in the network, each occupying
poverty line. Overall, we see that the poverty line provides a one channel. Each node obtains its spectrum availability
very reasonable estimation of node spectrum assignment and and follows the rules to select channels. In Figure 14, we
it is critical to incorporate poverty line into our rule design. measure the fairness and utilization of allocations derived from
Figure 10 also compares the performance of Rule C based Rules C and E and the explicit coordination approach (CA-
implicit coordination scheme to CA-EC. We examine the cu- EC). The utilities are averaged over 100 randomly deployed
mulative distribution function of the assignment using explicit primary nodes. We see that as primary devices increase their
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 139

2000
1800 Spectrum Allocated
0.5
1600
0.5 0.4
1400 0.4 0.3
1200 0.3 0.2
Meters

0.2
1000 0.1 0.1
800
0 0
600
0 2000
400
50 1500
200
100 1000
0 Meters Meters
0 20 40 60 80 100 120 140 160 180 200 150 500
Meters

2000
Spectrum Allocated
1800 0.5
1600 0.5 0.4
1400
0.4 0.3
0.3 0.2
1200 0.2
Meters

0.1 0.1
1000 0 0
800
600 0 2000
400 50 1500
200 100 1000
Meters 500 Meters
0 150
0 20 40 60 80 100 120 140 160 180 200
Meters

2000 Spectrum Allocated

0.5
1800
0.4
0.5
1600 0.4 0.3
0.3
1400 0.2
0.2

1200 0.1 0.1

Meters

0
0
1000
800
2000
0
600
1500
50
400
1000
100
200 Meters
500
Meters
150
0
0 20 40 60 80 100 120 140 160 180 200
Meters Fig. 8. The amount of spectrum allocated to each node using Rule C
(RruleC (n)) for (left) random network, (center) clustered network and (right)
Placelab network shown in Figure 7.
Fig. 7. Sample topology of (left) random network, (center) clustered network
and (right) Placelab network.

• Access points can broadcast “hello” beacons [26] peri-

spectrum usage, the performance of both explicit and implicit odically and help peers construct the interference matrix.
coordination approaches degrade gracefully. • Clients associated with APs can assist the interference
detection by sensing radio signals and provide feedback
on findings of interfering access points [14]. This mecha-
VIII. D ISCUSSION
nism has been shown to help refine the interference map.
In this section, we discuss practical issues associated with
Extending to complex interference characterizations. In
our network model and problem definition.
this paper, we use a binary matrix to model the pair-wise
Identifying channel availability and interference con- interference among access points. This model is widely used
straint. The proposed spectrum allocation system requires in existing works (e.g. [11], [14], [20]). Other models are
nodes (access points) to keep track of the channel availability based on aggregated SNR measurements, referred to as the
matrix L and the interference constraint C. We list three physical model [11]. However, under this model the corre-
complementary mechanisms to obtain this information. sponding channel allocation problem becomes significantly
• Access points perform spectrum measurements to de- complex, and there are no existing analytical bounds on indi-
termine L and C, using techniques like spectrum sens- vidual node’s spectrum usage. We are currently investigating
ing [5], [7] and RSSI measurement-based probing [22]. spectrum allocation schemes under this complex interference
140 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

10 10
RuleC/PL:Random
CA-EC/RuleC:Random
8 8
RuleC/PL

6 6

Y
4 4

2 2

0 0
0 10 20 30 40 50 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Degree Probability of Gain > Y
10 10
RuleC/PL:Cluster
CA-EC/RuleC:Cluster
8 8
RuleC/PL

6 6
Y
4 4

2 2

0 0
0 10 20 30 40 50 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Degree Probability of Gain > Y
10 10
RuleC/PL:Placelab
CA-EC/RuleC:Placelab
8
8
RuleC/PL

6
6
Y

4
4

2
2

0
0 5 10 15 20 25 30 35 40 45 50 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Degree
Probability of Gain > Y
Fig. 9. The amount of spectrum assigned to each node using Rule C divided
by its poverty line ( RruleC
P L(n)
(n)
), versus node conflict degree d(n) for (left) Fig. 10. Comparing the spectrum allocation per node using Rule C to
R (n)
random, (center) clustered and (right) Placelab networks. poverty line: ruleC
P L(n)
(marked by RuleC/PL), using explicit coordination
R (n)
over those using Rule C: RCA−EC(n) (marked by CA-EC/RuleC) for
ruleC
(left) random,(center) clustered and (right) Placelab networks, plotted as the
cumulative distributive function over 200 nodes.
model, from which we seek to derive analytical bounds on
node’s spectrum usage and devise spectrum rules accordingly.
Finally, in this paper, we assume interference conditions across channels to define the interference constraint and apply
are homogeneous across channels. Our model is a reasonable the proposed solution, i.e. any two nodes conflict if they
abstraction of the scenario where channels are divided from conflict on at lease one channel. However, this reduction
a single band pool and have carrier frequencies close to each is overly conservative and can degrade spectrum utilization.
other. When interference conditions are heterogeneous across An alternative is to make the interference constraint channel-
channels, we can use an aggregated measure of interference dependent, i.e. cn,k,m . Our prior work in [20] has developed
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 141

10
400
RuleC Random

Number of Usage Modifications

CA-EC Random
Utility (Proportional Fairness)

350
8
300
6
250

200 4
150
2
100

50 0
20 40 60 80 100 120 140 160 180 200 0 5 10 15 20 25
Distance (Meters) Degree
400 10
RuleC Cluster

Number of Usage Modifications

CA-EC Cluster
Utility (Proportional Fairness)

350
8
300

250 6

200
4
150

100 2

50
20 40 60 80 100 120 140 160 180 200 0
0 5 10 15 20 25 30 35 40 45 50
Distance (Meters)
Degree
350
RuleC PlaceLab 10
CA-EC PlaceLab
Number of Usage Modifications
Utility (Proportional Fairness)

300
8
250

6
200

150 4

100
2
50
20 40 60 80 100 120 140 160 180 200 0
Distance (Meters) 0 5 10 15 20 25 30 35 40 45 50
Degree
Fig. 11. The system fairness utility of Rule C and CA-EC versus the conflict
distance D, for (left) random,(center) clustered and (right) Placelab networks.
Fig. 12. The number of spectrum adjustments at each node versus the node
conflict degree d(n), for (left) random, (center) clustered and (right) Placelab
networks.
centralized approaches to optimize spectrum allocations under
these channel-dependent constraints. We are currently inves-
tigating analytical bounds on individual node spectrum usage
and applying these bounds to design rules for the proposed significantly less communication and computational overhead.
distributed implicit coordination approach. We propose five rules that tradeoff performance with imple-
mentation and communication complexity. We show that rules
guided by a lower bound on nodes’ spectrum usage (poverty
IX. C ONCLUSION line) provide superior performance. Experimental results show
We propose a distributed spectrum management scheme that rule-based approaches perform slightly worse than the
using implicit coordination. Nodes adjust spectrum usage previously proposed collaborative approaches, but have much
independently while complying with spectrum rules, leading to lower complexity and communication overhead.
142 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

40 24
Random
35 Clustered 23
Placelab CA−EC
30 22

Proportional Fairness
25 21
Rule E
20
Y

20

15 19
Rule C
18
10
17
5
16
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 15
2 4 6 8 10 12 14 16 18 20
Prob (complexity of CA-EC/ RuleC) >Y Number of Primary Users

Fig. 13. The number of spectrum adjustments at each node using CA-EC (a) Fairness
divided by those using Rule C, for (left) random, (center) clustered and (right) 170
Placelab networks, plotted as the cumulative distributive function over 200
nodes. 165

CA−EC
160

Spectrum Utilization
A PPENDIX
A. Proof of Theorem 1
155
Rule C
A node i observes its neighbors’ spectrum usage when
selecting channels. The total number of channels occupied 150
by its neighbors is at most O(i) = Ω · d(i). Obviously, Rule E
Ω ≤ L(i) − O(i). Hence, node i can always find at least 145
Ω idle channels that will not lead to any conflict.
140
B. Proof of Theorem 2
135
B.1 Proof for Rule B 2 4 6 8 10 12 14 16 18 20
Number of Primary Users
Definition 1: We define a node n as “qualified” if |C(n)| =
(b) Utilization
P L(n), and n does not conflict with any node who uses
less number of channels than P L(n). Otherwise, the node Fig. 14. Utilization and Fairness comparison of different rules assuming 20
is “disqualified”. channels, 40 secondary nodes and 20 primary nodes.

We start with a lemma that describes how the status of a node

switches by running Rule B. To see n must be “disqualified” before the spectrum ad-
Lemma 1: After a node n performs a spectrum adjustment justment, assume on the contrary that it is “qualified” before
based on rule B, it becomes “qualified”. the spectrum adjustment. According to the rule, the only
Proof: It is equivalent to prove that the sets O and I are motivation that n modifies its channel usage is there are more
large enough to fill P L(n) channels, i.e. |O| + |I| ≥ P L(n) idle channels (i.e. set I \ C(n) is not empty) so n seizes them
in Algorithm 2. and gives up the same amount of channels under conflict (i.e.
There are at most d(n) neighbors of n whose channel usage in set O ∩ C(n)). This modification does not alter |C(n)|
is less than P L(n), denoted as n1 , n2 , · · · , ni , i ≤ d(n). and does not create new conflicts, so n won’t change from
Hence, Eq. 4. “qualified” to “disqualified”.
Therefore, |O| + |I| = L(n) − |B| ≥ P L(n). From lemma 2, node n can only convert neighbors with
higher Poverty Line from “qualified” to “disqualified”, and
The following lemma describes how a node affects the
a “qualified” node won’t change any of its neighbors to
status of its neighbors by running Rule B.
“disqualified”.
Lemma 2: When a node n adjusts its spectrum usage and
converts any of its neighbors, i.e. node n from “qualified” to Lemma 3: For a given system with N nodes and any
“disqualified”, then P L(n) < P L(n ). Moreover, n must be initial channel assignment, after at most an expected number
“disqualified” before its spectrum adjustment. of O(N 2 ) spectrum adjustments, all the nodes will become
“qualified”.
Proof: Since n is “qualified” before n’s adjustment, n ’s
channel usage before adjustment is |C(n )| = P L(n ). If Proof: From lemma 2, if spectrum adjustment is carried
P L(n) > P L(n ), then after the spectrum adjustment, n’s out in a “structured” format such that “disqualified” nodes
channel usage will not conflict with that of n , so n won’t with lowest poverty line among its neighbors have the highest
change from “qualified” to “disqualified”. priority in adjusting spectrum, then each “qualified” node will
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 143


i 
i
d(n)
|B| = | C(ns )| ≤ |C(ns )| ≤ d(n) × P L(n) ≤ L(n) (4)
s=1 s=1
d(n) + 1

never become a “disqualified” node. As a result, the total Lemma 5: When a node n adjusts its spectrum and converts
number of spectrum adjustment is at most N for all nodes any of its neighbors, i.e. node n from “qualified” to “disqual-
to become “qualified”. ified”, then P L(n) < P L(n ). In addition, n is “disqualified”
However, if the order of spectrum adjustment is random, before its spectrum adjustment.
then a “disqualified” node with the lowest poverty line among
its neighbors will wait for an expected number of at most Proof: We first assume that P L(n) ≥ P L(n ). Note n is
N other adjustments to get the opportunity. This is mainly “qualified” before the spectrum adjustment. Therefore, there
due to at each adjustment slot, the number of nodes who are two possible cases.
 
needs to adjust spectrum is bounded by N . Therefore the • If |C(n )| > P L(n ), then according to Rule C n

number of total spectrum adjustments until the system reaches will avoid conflict with the upper P L(n) (≥ P L(n ))
an equilibrium is bounded by O(N 2 ). channels thus n is still “qualified”.
  
Now, we can use the above lemmas to prove theorem 2 for • If |C(n )| = P L(n ), then P L(n) ≥ |C(n )|. Then
Rule B. according to the rule, after the modification n won’t
Proof of Theorem 2 for Rule B: conflict with n , so n won’t change from “qualified” to
By lemma 3, after at most an expected number of O(N 2 ) “disqualified”.
channel modifications all nodes are “qualified”. Therefore, P L(n) ≥ P L(n ) is impossible, and we have
According to definition 1, when all nodes are “qualified”, proved P L(n) < P L(n ).
for every n, n is on its Poverty Line, and for every neighbor Next, we need to prove that n must be “disqualified”
n s.t. P L(n) ≥ |C(n )| = P L(n ), n doesn’t conflict with before the spectrum adjustment. Let’s first assume that n
n . This implies the global assignment is conflict-free. It is is “qualified” before the adjustment. From definition 2, n’s
just the status of equilibrium. spectrum usage can be divided into the following two cases.
B.2 Proof for Rule C • If |C(n)| > P L(n), and the upper P L(n) channels

Definition 2: We define a node n as “qualified” if one of occupied by n do not generate any conflict with any
the following conditions is satisfied. neighbor. During the spectrum adjustment, n will simply
1) |C(n)| > P L(n), and the P L(n) channels with the exit from the conflicted channels since the number of idle
smallest channel IDs occupied by n (hereby referred to channels is enough for n to reach its Poverty Line. After
as upper channels) do not generate conflict with any that n won’t conflict with its neighbor n , which means
neighbor of n. that n won’t become “disqualified”.
2) |C(n)| = P L(n), and for every neighbor n , n doesn’t • If |C(n)| = P L(n). Since n is “qualified”, the conflicting

conflict with n on the upper min(P L(n), |C(n )|) set B is disjoint with C(n). According to the rule, the
channels occupied by n . only motivation that n modifies its channel usage is
Otherwise, the node is “disqualified”. there are more idle channels (i.e. set I \ C(n) = ∅).
During spectrum adjustment, n seizes a number of idle
We start with a lemma that describes how the status of a node channels and exits from the same amount of channels that
switches by running Rule C. are conflicting (i.e. channels in the set O ∩ C(n)). This
Lemma 4: , After n performs a channel modification based modification does not alter |C(n)| and does not create
on rule C, n becomes “qualified”. new conflicts. Now n ’s spectrum usage can be divided
into two cases:
Proof: If |I| ≥ P L(n), then after applying Rule C n will
If |C(n )| > P L(n ) and the upper P L(n ) channels of
occupy all idle channels and won’t conflict with any neighbor,
n are free of conflict, then after n modifies its channel
so the qualification is trivial. Otherwise if |I| = P L(n)
usage those channels are still free of conflict thus n is
because Rule C is reserving the upper min(P L(n), |C(n )|)
still “qualified”; If |C(n )| = P L(n ), because P L(n ) >
channels for each neighbor, the only thing remains to prove
P L(n) and P L(n) = |C(n)|, it is straightforward that
is sets O and I are enough for providing PL(n) channels, i.e.
|C(n )| > |C(n)|. Because n is “qualified” before the
|O| + |I| ≥ P L(n) in algorithm 1.
modification of n, n doesn’t conflict with n. This means
There are at most d(n) neighbors of n. Based on Rule C,
that after n’s spectrum adjustment, n and n do not
node n reserves at most P L(n) channels into set B for each
conflict with each other. Hence, n is still “qualified”.
neighbor. Therefore
In conclusion, if n is “qualified”, then its channel modifi-
|B| ≤ d(n) × P L(n) ≤
d(n)
L(n) cation won’t transfer n into “disqualified”.
d(n) + 1 From lemma 5, node n can only convert neighbors with
Therefore, |O| + |I| = L(n) − |B| ≥ P L(n). higher Poverty Line from “qualified” to “disqualified”, and
a “qualified” node won’t change any of its neighbors to
The following lemma describes how a node affects the status “disqualified”.
of its neighbors by running Rule C.
144 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 26, NO. 1, JANUARY 2008

Lemma 6: For a given system with an arbitrary initial Since the switch will only impact on the values of CFm (.)
assignment, through at most an expected number of O(N 2 ) and CFm
(.), it is obvious that CF (A ) < CF (A). Hence,
channel modifications all nodes will become “qualified”. CF strictly decreases each time a node modifies its channel
usage. CF (A) ≤ |E| × M and CF ∈ Z, the switch will end
Proof: From lemma 5, the proof is quite similar as the
after at most |E| × M iterations, and the system will reach an
proof for lemma 3.
equilibrium.
Now, we can use the above lemmas to prove theorem 2 for
Rule C. D. Proof of Theorem 4
By lemma 6, after at most an expected number of O(N 2 )
channel modifications all nodes are “qualified”. At this point, Proof for Ψ = 1: By Pigeonhole Principle [3], there
some “qualified” nodes can still modify their channel usage. must exist a channel where a node n observes less or equal
We hereby refer to a “qualified” node whose spectrum usage is than  d(n)
M  neighbors. At an equilibrium, each node selfishly
above its poverty line (i.e. |C(n)| > P L(n), as the first clause selects the best channel for itself, and conflicts with at most
in definition 2) as type I node, and a node whose spectrum  d(n)
M  neighbors.
usage is equal to its poverty line(i.e. the second clause in
definition 2) as type II nodes. Proof for 1 < Ψ < M : Assuming that a node n is assigned
When all nodes are “qualified”, there may still exist some with Ψ channels. By an inductive argument using Pigeonhole
conflicts. However two type II nodes are not possible to Principle, there must exist Ψ channels, denoted c1 , c2 , · · · , cΨ ,
conflict: the one with higher or equal Poverty Line will avoid such that
conflict with the other. So conflict only exists between one Ψ × d(n)
Type I node and a Type II node, or between two Type I nodes. dc1 (n) + dc2 (n) + · · · + dcΨ (n) ≤ Ψ × (  + 1).
M
Type I nodes modify their spectrum usage by exiting from
channels that conflict with neighbors, so their modification Here we use dc (n) to denote the number of neighbors of n
eliminates the conflicts with their neighbors. So after at most who are occupying channel c. Now if node n switches to these
O(N 2 ) expected iterations all type I nodes modified their Ψ channels, its utility will be
 
channel usage and the system is conflict-free. 1 1 1 1
When spectrum usage in a system becomes conflict-free + + ···+
λ dc1 (n) + 1 dc2 (n) + 1 dcΨ (n) + 1
and all nodes are “qualified”, the only spectrum adjustment 1 Ψ
at a node is to seize idle channels. The number of this ≥
λ ((dc1 (n) + 1)(dc2 (n) + 1) · · · (dcΨ (n) + 1)) Ψ1
spectrum adjustment is at most N before the system enters
1 Ψ
an equilibrium. Therefore, the expected number of spectrum ≥ (d (n)+1)+(d (n)+1)+···+(d
λ( 1 cΨ (n)+1) Ψ 1
adjustment for a system with N nodes to reach an equilibrium
c c 2
Ψ ) Ψ
is at most O(N 2 ). 1 Ψ2
=
λ dc1 (n) + 1 + dc2 (n) + 1 + · · · + dcΨ (n) + 1
C. Proof of Theorem 3
1 Ψ2
We start the proof by defining a conflict graph G = {U, E}, ≥
λ Ψ( Ψd(n)  + 1) + m
where U represents the list of vertices (the secondary nodes), M
(|U | = N ) , and E represents the edges between vertices. An Ψ
= .
edge exists between two vertices if the corresponding nodes λ{ Ψd(n)
M  + 2}
conflict with each other if using the same channel. Λ = |E|
represents the number of edges in the conflict graph. It is Proof for Ψ = M : Straightforward following the interfer-
obvious that the number of edges is maximized when the ence model. For each node, there are d(n)+1 nodes competing
network is all connected, where Λ = N (N − 1)/2. Let A for all M channels.
characterize the network’s spectrum usage, i.e. the channel(s)
used by each secondary node/vertex. We now define the R EFERENCES
conflict factor of A as
[1] Business 2.0 magazine: Get ready for googlenet, 2005.

M−1
[2] A KELLA , A., J UDD , G., S ESHAN , S., AND S TEENKISTE , P. Self-
CF (A) = CFm (A), management in chaotic wireless deployments. In Proc. MobiCom (Aug.
m=0 2005).
[3] B RUALDI , R. A. Introductory Combinatorics, 3rd ed. Prentice Hall,
where CFm (A) is the number of the pair of conflicting nodes Englewood Cli s, N.J., 1999.
who are using channel m. Following Rules D and E, a node [4] B UDDHIKOT, M. M., K OLODY, P., M ILLER , S., RYAN , K., AND
n switches from channel m to m only if it can gain more E VANS , J. DIMSUMNet: New directions inwireless networking using
throughput from m than from m. Based on the interference coordinated dynamic spectrum access. In Proc. IEEE WoWMoM (June
2005).
model, the number of nodes conflicting with n on channel m [5] C ABRIC , D., M ISHRA , S. M., AND B RODERSEN , R. W. Implementa-
is smaller than that on channel m. Therefore, we can derive tion issues in spectrum sensing for cognitive radios. In Proc. Asilomar
the following. conference on signals, systems and computers (2004).
[6] C AO , L., AND Z HENG , H. Spectrum allocation in ad hoc networks via
CFm
(A ) − CFm
(A) < CFm (A) − CFm (A ) ⇒ local bargaining. In Proc. SECON (Sep. 2005).
[7] G. G ANESAN , Y. L. Implementation issues in spectrum sensing for
CFm
(A ) + CFm (A ) < CFm (A) + CFm
(A). cognitive radios. In Proc. IEEE DySPAN 2005 (Nov. 2005).
CAO and ZHENG: DISTRIBUTED RULE-REGULATED SPECTRUM SHARING 145

[8] H AYKIN , S. Cognitive radio: Brain-empowered wireless communica- [25] YANG , G., Z HENG , H., Z HAO , J., AND L I , V. Adaptive channel
tions. IEEE J. Select. Areas Commun. 23, 2 (Feb. 2005), 201–220. selection through collaborative sensing. In Proc. ICC (June 2006).
[9] H EUSSE , M., ROUSSEAU , F., B ERGER -S ABBATEL , G., AND D UDA , A. [26] Z HAO , J., Z HENG , H., AND YANG , G. Distributed coordination in
Performance anomaly of 802.11b. In Proc. INFOCOM (March 2003). dynamic spectrum allocation networks. In Proc. IEEE DySPAN (Nov.
[10] I LERI , O., S AMARDZIJA , D., AND M ANDAYAM , N. B. Demand 2005).
responsive pricing and competitive spectrum allocation via spectrum [27] Z HENG , H., AND C AO , L. Device-centric spectrum management. In
server. In Proc. IEEE DySPAN (Nov. 2005). Proc. IEEE DySPAN (Nov.2005).
[11] JAIN , K., PADHYE , J., PADMANABHAN , V., AND Q IU , L. Impact [28] Z HENG , H., AND P ENG , C. Collaboration and fairness in opportunistic
of interference on multi-hop wireless network performance. In Proc. spectrum access. In Proc. ICC (June 2005).
Mobicom 2003.
[12] L. J IANG , S. C. L. Proportional fairness in wireless lans and ad hoc
networks. In Proc. IEEE WCNC (2005).
[13] M C H ENRY, M. Spectrum white space measurements. New America
Foundation Broadband Forum (June 2003).
[14] M ISHRA , A., ET AL . Distributed channel management in uncoordinated Lili Cao received the B.S. and M.S. in Computer
wireless environments. In Proc. MobiCom (Sep. 2006). Science from Shanghai Jiaotong University, Shang-
[15] M ITOLA III, J., AND M AGUIRE J R ., G. Q. Cognitive radio: making hai, China, in 2002 and 2005, respectively. He
software radios more personal. IEEE Pers. Commun. 6, 4 (Aug. 1999), is currently a PhD student in the Department of
13–18. Computer Science, University of California, Santa
[16] N ANDAGOPAL , T., T.K IM , X.G AO , AND B HARGHAVAN , V. Achieving Barbara.
MAC layer fairness in wireless packet networks. In Proc. MobiCom His research interests include cognitive radio sys-
(Aug. 2000). tems, spectrum allocation, and distributed algorithms
[17] N EEL , J., AND R EED , J. Performance of distributed dynamic frequency for large-scale networks.
selection schemes for interference reducing networks. In Proc. Milcom
(Oct. 2006).
[18] N EEL , J., R EED , J., AND G ILLES , R. The role of game theory in the Haitao Zheng received her B.S.degree (with high-
analysis of software radio networks. In Proc. Software Defined Radio est honor) from Xian Jiaotong University in July
Forum Technical Conference and Product Exhibition (SDR 02) (San 1995, her M.S.EE and Ph.D degree in Electrical and
Diego, CA, USA, Nov. 2002), vol. 2, pp. pp. NP–3–02. Computer Engineering from University of Mary-
[19] N IE , N., AND C OMANICIU , C. Adaptive channel allocation spectrum land, College Park, in May 1998 and July 1999,
etiquette for cognitive radio networks. In Proc. IEEE DySPAN (Nov. respectively. She joined wireless research lab, Bell-
2005). Labs, Lucent Technologies as a member of techni-
[20] P ENG , C., Z HENG , H., AND Z HAO , B. Y. Utilization and fairness cal staff in August 1999, and moved to Microsoft
in spectrum assignemnt for opportunistic spectrum access. Mobile Research Asia as a project lead and researcher, in
Networks and Applications (MONET) 11 (May 2006), 555–576. March 2004. Since Sept. 2005, she has been an
[21] R AMAN , C., YATES , R. D., AND M ANDAYAM , N. B. Scheduling assistant professor in Computer Science Department,
variable rate links via a spectrum server. In Proc. IEEE DySPAN (Nov. University of California, Santa Barbara. Dr. Zheng was recently named as the
2005). 2005 MIT Technology Review Top 35 Innovators under the age of 35 for her
[22] R EIS , C., M AHAJAN , R., RODRIG , M., W ETHERALL , D., AND Z A - work on cognitive radios. Her work was selected by MIT Technology Review
HORJAN , J. Measurement-based models of delivery and interference in as one of the 10 Emerging Technologies in 2006. She also received 2002 Bell
static wireless networks. In Proc. ACM SIGCOMM (2006). Laboratories President’s Gold Award from Lucent Bell-Labs, and 1998-1999
[23] S TEENSTRUP, M. Opportunistic use of radio- frequency spectrum: a George Harhalakis Outstanding Graduate Student Award from Institute of
network perspective. In Proc. IEEE DySPAN (Nov. 2005). System Research, University of Maryland, College Park.
[24] X ING , Y., C HANDRAMOULI , R., M ANGOLD , S., AND S HANKAR , S. Dr. Zheng’s research interests include wireless systems and networking and
Dynamic spectrum access in open spectrum wireless networks. IEEE J. multimedia computing.
Select. Areas Commun. 24, 3 (March 2006), 626–637.