Stochastic Analysis and Optimization of Heterogeneous ... · Stochastic Analysis and Optimization...

Stochastic Analysis and Optimization of Heterogeneous WirelessNetworks

by

Wei Bao

A thesis submitted in conformity with the requirementsfor the degree of Doctor of Philosophy

Graduate Department of Electrical and Computer EngineeringUniversity of Toronto

c⃝ Copyright 2016 by Wei Bao

Abstract

Stochastic Analysis and Optimization of Heterogeneous Wireless Networks

Wei Bao

Doctor of Philosophy

Graduate Department of Electrical and Computer Engineering

University of Toronto

2016

In order to improve the performance of mobile networks, one widely promoted approach is to install

a diverse set of small-cells overlaying the current macrocell network to form a multi-tier heterogeneous

wireless network (HWN). In this thesis, I propose new stochastic approaches to evaluate and design

HWNs by investigating user load, interference patterns, and user mobility, the results of which provide

new analytical insights and design guidelines to improve future HWNs.

In the first part of this thesis, I focus on the evaluation of user load through characterizing the

joint distribution of users in different cells in an HWN with arbitrary user movement trajectories and

dependently distributed user channel holding times. Through developing a new stochastic network

analysis framework, I derive a closed-form expression for the joint user distribution, which is only related

to the average arrival rate and the average channel holding time of each cell, and hence it is irrelevant to

the general user movement patterns and distributions of channel holding times. This property suggests

that accurate evaluation of the user distribution and other associated metrics such as the system workload

can be achieved with low complexity, without the need to collect a large amount of user location data.

The multi-tier architecture of HWNs introduces complicated interference patterns in the system. In

the second part of this thesis, I introduce a stochastic analytical framework to compare the performance

of open and closed access modes in a two-tier network with macrocells and femtocells, with regard

to uplink interference and outage at both the macrocell and femtocell levels. A stochastic geometric

approach is employed as the basis for the analysis to characterize the distributions of uplink interference

and the outage probabilities. I further derive sufficient conditions for open and closed access modes to

outperform each other in terms of the outage probability at either the macrocell or femtocell level. This

leads to closed-form expressions to upper and lower bound the difference in the targeted received power

between the two access modes.

In the third part of this thesis, I study the resource allocation and user association problem in HWNs

with random distributed users and BSs for optimizing the average user data rate. Both the user load and

interference patterns are considered. I first derive the average user data rate through stochastic geometric

analysis. The expression is employed as the objective function of the optimization problem, which is

ii

non-convex in nature and cannot be solved with a standard method. Then, I propose an innovative

approach, solving the optimization problem optimally for low user density, and asymptotically optimally

for high user density.

The deployment of small-cell BSs in HWNs leads to a higher user data rate, but it also introduces

more handoffs to the users. In the fourth part of this thesis, I present a new stochastic geometric analysis

framework on user mobility in HWNs, which captures the spatial randomness and various scales of cell

sizes in different tiers. I derive analytical expressions for the rates of all handoff types experienced by

an active user with arbitrary movement trajectory. Noting that the data rate of a user also depends on

the set of cell tiers that it is willing to use, I also provide guidelines for tier selection under various user

velocity, so that an optimal tradeoff between the handoff rate and the data rate can be achieved.

iii

Acknowledgements

First, I would like to express my deepest gratitude to my supervisor, Prof. Ben Liang, for his

patient guidance, constant encouragement, and excellent advice throughout my PhD study. Without his

invaluable help, this work would not be possible.

I would like to thank Prof. Wei Yu for being my thesis committee member and for his collaboration

with me on part of my research. His suggestions significantly improved the quality of the thesis. I would

also like to thank committee member Prof. Jorg Liebeherr for his careful proof reading and insightful

suggestions on this thesis. In addition, Prof. Jianping Pan graciously agreed to be my external examiner,

and his feedback in the final stage was very helpful. I am also very grateful to Dr. Stefen Valentin, who

offered me the precious research intern opportunity at the Bell Labs and helped a lot in my research.

I am very grateful to Dr. Wei Wang for his help in improving my research skills and planning my

future career. His strong support and encouragement kept me motivated and confident. I would also like

to thank Dr. Yicheng Lin for his collaboration with me on our IEEE JSAC paper. His visions on future

mobile networks greatly inspired me. Special thanks also go to Dr. Sun Sun, Jaya Prakash Champati,

Yuhan Zhou, and Binbin Dai, who provided me with precious assistance during my PhD study. Thanks

to all my colleagues who offered helpful discussions: Dr. Mahdi Hajiaghayi, Ali Ramezani, Meng-Hsi

Chen, Dr. Ruhallah Ali Hemmati, Yujie Xu, Sowndarya Sundar, Qiang Xiao, Dr. Honghao Ju, Juan

Wen, Caiyi Zhu, Wanyao Zhao, Samer Fouad Zakhary, and others.

I must also mention the joy moments I had on weekends and holidays, such as badminton games,

board games, travels, picnics, and so on. They will definitely become special memories of mine when

looking back to my PhD life. Thanks to Dr. Wei Wang, Dr. Yicheng Lin, Yu Xiao, Binbin Dai, Dan

Fang, Dr. Tony Liang Liang, Jessica Yihua Hu, Yuhan Zhou, Caiyi Zhu, Wanyao Zhao, Dr. Qi Zhang,

and Siyu Liu who have brought happiness to my life. I must also mention Yifei Hao, Yu Xia, Dr. Wenbo

Shi, Candy Tian Yu, and Yichun Qiu. Thanks for their encouragement, support, and most of all their

humor. They kept things light and me smiling.

I must acknowledge with tremendous thanks to my wife, Zhi Zeng. Through your love, patience,

support, and unwavering belief in me, I’ve been able to complete my long PhD journey. Thank you with

all my heart and soul. I love you and am forever indebted to you for giving me your love and heart.

Finally, I take this opportunity to express my profound gratitude to my beloved parents for under-

standing, support, and endless love during my study in Canada. To you I dedicate this thesis.

This research was partially funded by Bell Canada and the Natural Sciences and Engineering Research

Council (NSERC) of Canada.

iv

Contents

1 Introduction 1

1.1 New Challenges in Analysis and Design of HWNs . . . . . . . . . . . . . . . . . . . . . . . 1

1.1.1 Understanding User Distribution in HWNs . . . . . . . . . . . . . . . . . . . . . . 1

1.1.2 Characterization of Interference Patterns . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.3 Design of User Association Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.4 Quantification of Handoff Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Thesis Outline and Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.1 User Distribution under General Mobility and Session Patterns . . . . . . . . . . . 4

1.2.2 Uplink Interference Comparison of Open Access and Closed Access . . . . . . . . . 5

1.2.3 Optimal Spectrum Allocation and User Association . . . . . . . . . . . . . . . . . 5

1.2.4 Handoff Rate Analysis in HWNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Related Works 7

2.1 User Mobility Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.1 Queueing Networks Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.2 Insensitivity Property of Queueing Networks . . . . . . . . . . . . . . . . . . . . . 7

2.1.3 Mobility Modeling with Cell Geometry . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2 Interference Analysis of HWNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.1 Stochastic Geometry as a Basic Tool . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.2 Downlink Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.2.3 Uplink Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.3 Spectrum Allocation and User Association . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4 A Brief Review on Queueing Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4.1 Queueing Network under Consideration . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4.2 Routing Balance Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4.3 Stationary Distributions of Jackson Networks . . . . . . . . . . . . . . . . . . . . . 11

2.5 A Brief Review on Stochastic Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.5.1 Interference Analysis Based on Poisson Point Process . . . . . . . . . . . . . . . . 12

2.5.2 Random Fibre Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.6 Publications Related to this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Insensitivity of User Distribution in HWNs 14

3.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.2 Stationary User Distribution in Single-Route Network . . . . . . . . . . . . . . . . . . . . 16

v

3.2.1 Queueing Network Model for Single-Route Network . . . . . . . . . . . . . . . . . 16

3.2.2 Reference Single-Route Memoryless Network . . . . . . . . . . . . . . . . . . . . . 17

3.2.3 Insensitivity of Single-Route Network . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.3 Stationary User Distribution in Multiple-Route Network . . . . . . . . . . . . . . . . . . . 21

3.3.1 Queueing Network Model for Multiple-Route Network . . . . . . . . . . . . . . . . 21

3.3.2 Insensitivity of π(x) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.3.3 Insensitivity of π1(y) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4 Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.4.1 Requirements and the Dartmouth Traces . . . . . . . . . . . . . . . . . . . . . . . 23

3.4.2 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.4.3 Trace Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.4.4 Marginal User Distribution at a Single AP . . . . . . . . . . . . . . . . . . . . . . . 26

3.4.5 KL Divergence and Entropy Gap for Multiple APs . . . . . . . . . . . . . . . . . . 26

3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4 Uplink Interference Analysis: Open Access versus Closed Access 31

4.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.1.1 Two-tier Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.1.2 Open Access versus Closed Access . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.1.3 Path Loss and Power Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.1.4 Outage Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.2 Open Access vs. Closed Access at the Macrocell Level . . . . . . . . . . . . . . . . . . . . 34

4.2.1 Open Access Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4.2.2 Closed Access Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.2.3 Parameter Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.2.4 Open Access vs. Closed Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.3 Open Access vs. Closed Access at the Femtocell Level . . . . . . . . . . . . . . . . . . . . 39

4.3.1 Open Access Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.3.2 Closed Access Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.3.3 Open Access vs. Closed Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.4 Numerical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5 Rate Maximization through Spectrum Allocation and User Association 47

5.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.1.1 Multi-tier Wireless Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.1.2 Power and Path Loss Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.1.3 Spectrum Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.1.4 Coverage Probability and UE Data Rate . . . . . . . . . . . . . . . . . . . . . . . . 48

5.1.5 Flexible User Association . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.1.6 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5.2 Closed-form Average UE Data Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5.3 Optimization Problem and SSAUA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.3.1 Density Thresholding Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

vi

5.3.2 SSAUA under∑K

i=1 ai > 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53


i=1 ai < 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5.3.4 Computational Complexity Comparison . . . . . . . . . . . . . . . . . . . . . . . . 56

5.4 Nash Equilibrium for SSAUA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

5.5 The Multiple-MCS Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.5.1 Average UE Data Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.5.2 SSAUA in the Multiple-MCS Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5.6 Numerical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.6.1 Average UE Data Rate under Different UE Densities . . . . . . . . . . . . . . . . . 60

5.6.2 Average UE Data Rate under Different Path Loss Exponents . . . . . . . . . . . . 62

5.6.3 Association Bias Values and Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5.6.4 Run Time Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.6.5 Performance of SSAUA in the Multiple-MCS Case . . . . . . . . . . . . . . . . . . 63

5.6.6 Performance Comparison under More Realistic Network Topologies . . . . . . . . . 64

5.7 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

6 Stochastic Geometric Analysis of User Mobility in HWNs 67

6.1 Handoff Rate Analysis in HWNs with Poisson Patterns . . . . . . . . . . . . . . . . . . . 67

6.1.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.1.2 Handoff Rate Analysis in Multi-tier HWNs . . . . . . . . . . . . . . . . . . . . . . 69

6.1.3 UE’s Data Rate and Tier Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

6.1.4 Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

6.2 Handoff Rate Analysis in HWNs with Poisson and Poisson Cluster Patterns . . . . . . . . 81

6.2.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6.2.2 Handoff Rate Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6.2.3 Numerical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

6.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

7 Conclusions 91

A Proofs of Theorems in Chapter 4 93

A.1 Proof of Theorem 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93




A.5 Proof of Lemma 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

B Properties and Proofs of Chapter 5 100

B.1 Useful Properties of Mk(Ak) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

B.2 Proof of Lemma 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

B.3 Proof of Lemma 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

B.4 Useful Properties of Mk(Ak) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

B.5 Proof of Lemma 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

B.6 Proof of Theorem 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

B.7 Some Properties Used for Exhaustive Search . . . . . . . . . . . . . . . . . . . . . . . . . . 103

vii

B.8 Proofs of Properties of Mk(·) in the Multiple-MCS Case . . . . . . . . . . . . . . . . . . . 103

B.8.1 Property (M-1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

B.8.2 Property (M-2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

B.8.3 Property (M-3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

B.8.4 Property (M-4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

B.9 Numerical Verification of Properties (M-2’) and (M-5’) of Mk(Ak) in the Multiple-MCS

Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

C Derivations and Proofs of Theorems in Chapter 6 106

C.1 Proof of Theorem 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106


C.3 Derivation of P(Rk > R0) when k ∈ KC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110




C.7 Proof of Lemma 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

Bibliography 115

viii

List of Tables

3.1 Number of stages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

ix

List of Figures

1.1 (a) Three separated tiers of BSs. From top to bottom: macrocell, picocell, and femtocell

BSs. (b) A three-tier HWN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 An example of a three-tier HWN. Macrocell BSs, picocell BSs, and femtocell BSs are

represented by squares, circles, and triangles, respectively; users are represented by dots;

blue solid lines show cell boundaries; dashed lines represent connections between users

and BSs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 An example of a three-tier HWN and selective tier association. Tier-1, 2, and 3 BSs are

represented by squares, circles, and triangles, respectively; blue curves show intra-tier cell

boundaries; green curves show inter-tier cell boundaries. . . . . . . . . . . . . . . . . . . . 4

3.1 System model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2 Single-route network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.3 Decoupled network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.4 Reference memoryless decoupled network. . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.5 Multiple-route network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.6 The pdf of channel holding time in different stages. . . . . . . . . . . . . . . . . . . . . . . 26

3.7 Comparison of distributions for single APs. Real distributions are in solid lines; analytical

distributions are in dashed lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.8 Hkl and Hreal under the influence of non-Poisson arrivals. . . . . . . . . . . . . . . . . . . 28

3.9 Hgap and Hreal under the influence of non-Poisson arrivals. . . . . . . . . . . . . . . . . . 28

3.10 Hkl, Hgap and Hreal under the influence of distance restriction. . . . . . . . . . . . . . . . 28

3.11 Hkl, Hgap and Hreal under the influence of one-stage sessions. . . . . . . . . . . . . . . . . 28

4.1 Two-tier network with macrocells and femtocells. . . . . . . . . . . . . . . . . . . . . . . . 32

4.2 Macrocell outage probability under different λ, with µ = 4 units/km2/sub-band and T = 0.1. 43

4.3 Femtocell outage probability under different λ, with µ = 4 units/km2/sub-band and T = 0.1. 43

4.4 Macrocell outage probability under different µ, with λ = 4 units/km2/sub-band and T = 0.1. 43

4.5 Femtocell outage probability under different µ, with λ = 4 units/km2/sub-band and T = 0.1. 43

4.6 Macrocell outage probability under different T , with λ = 8 units/km2/sub-band and µ = 8

units/km2/sub-band. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.7 Femtocell outage probability under different T , with λ = 8 units/km2/sub-band and µ = 8

units/km2/sub-band. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.8 ρ∗ under different R at the macrocell level. . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.9 ρ∗∗ under different R at the femtocell level. . . . . . . . . . . . . . . . . . . . . . . . . . . 45

x

4.10 ρ∗∗ under different xB , xB = (xB , 0), at the femtocell level. . . . . . . . . . . . . . . . . . 45

4.11 ρ∗∗ under different xB , xB = (0, yB), at the femtocell level. . . . . . . . . . . . . . . . . . 45

5.1 Average UE data rate under different UE density µ. . . . . . . . . . . . . . . . . . . . . . 60

5.2 Comparison of different schemes under different UE density µ. . . . . . . . . . . . . . . . 60

5.3 Average UE data rate under different path loss exponent γ. . . . . . . . . . . . . . . . . . 61

5.4 Comparison of different schemes under different path loss exponent γ, analytical case. . . 61

5.5 Comparison of different schemes under different path loss exponent γ, simulation case. . . 62

5.6 Designed association bias values under different UE density µ. “o” represents the opti-

mality region, and “a. o.” represents the asymptotic-optimality region. . . . . . . . . . . . 62

5.7 Pricing values under different UE density µ. “o” represents the optimality region, and “a.

o.” represents the asymptotic-optimality region. . . . . . . . . . . . . . . . . . . . . . . . . 63

5.8 Comparison of run time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.9 Comparison of different schemes in the multiple-MCS case. . . . . . . . . . . . . . . . . . 64

5.10 Comparison of different schemes under more realistic network settings, (λ′1, λ

′2, λ

′3) =

(1, 2, 3) unit/km2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.11 Comparison of different schemes under more realistic network settings, (λ′1, λ

′2, λ

′3) =

(1, 5, 10) unit/km2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

6.1 The blue curves show T(1); and the region within red dashed curves shows T(2)(∆d). . . . 70

6.2 Cumulative distribution function of the latitude and longitude. . . . . . . . . . . . . . . . 77

6.3 Two-tier case: comparison of analytical and empirical handoff rates. . . . . . . . . . . . . 77

6.4 One-tier case: comparison of analytical and empirical handoff rates. . . . . . . . . . . . . 77

6.5 Two-tier case: handoff rates under different λ1 values. . . . . . . . . . . . . . . . . . . . . 78

6.6 Three-tier case: handoff rates under different λ2 values. . . . . . . . . . . . . . . . . . . . 78

6.7 Two-tier case: handoff rates under different B1 values. . . . . . . . . . . . . . . . . . . . . 79

6.8 Three-tier case: handoff rates under different B2 values. . . . . . . . . . . . . . . . . . . . 79

6.9 Two-tier case: overall utility comparison of different tier selections. . . . . . . . . . . . . 79

6.10 Three-tier case: overall utility comparison of different tier selections. . . . . . . . . . . . . 80

6.11 Two-tier case: tier selection velocity threshold. . . . . . . . . . . . . . . . . . . . . . . . . 81

6.12 Three-tier case. The set of dashed and solid curves show the velocity threshold values for

tier selection, under UR = 1 and UR = 1.05, respectively. Each set of curves separates the

plane into multiple regions, and each region corresponds to a specific optimal tier selection

scheme (labeled in the region). For example, a vertical line at λ3 = 16 has 3 intersections

with the set of solid curves, illustrating that under λ3 = 16 and UR = 1.05, tier selections

1, 2, 3, 1, 3, 1, 2, and 1 are optimal in four different velocity ranges separated by

the 3 intersections. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6.13 An example of a two-tier HWN with Poisson and Poisson cluster patterns. Tier-1 and 2

BSs are represented by “” and “”, respectively. Tier-2 BSs are clustered in four disk

regions. Blue, red, and green curves show cell boundaries within tier-1, between tier-1

and tier-2, and within tier-2, respectively. The magenta arrow represents the trajectory

of an active UE. The UE makes two handoffs at the intersections between its trajectory

and the set of cell boundary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6.14 Geometric patterns of two intersecting circles. . . . . . . . . . . . . . . . . . . . . . . . . . 85

xi

6.15 Accuracy of PCP handoff rate analysis under different µ2 values, for µ2 · ν2 = 0.5. For

comparison, black dashed lines indicate analytical results assuming all PPP BSs. . . . . . 88

6.16 Handoff rates under different λ1 values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

6.17 Handoff rates under different B2 values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

6.18 Handoff rates, two P-tiers and two C-tiers. . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

B.1 Diagrams of Mi(·), Mj(·), Mi(·), and Mj(·). . . . . . . . . . . . . . . . . . . . . . . . . . 101

B.2 [Mj(aj)− Mj(aj +D)]− [Mi(ai)− Mi(ai +D)] versus D (curves) and Mj(aj)− Mi(ai)

(dots). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

C.1 The region (shaded part) of Skj(∆d). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

xii

Chapter 1

Introduction

Traditional single-tier macro-cellular networks provide wide coverage for mobile users, but they are

insufficient to satisfy the exploding demand for high bandwidth access driven forth by modern mobile

traffic, such as multimedia transmissions and cloud computing tasks. In order to resolve this issue, one

efficient means is to provide more serving stations in a geographical area, i.e., installing a diverse set

of small-cells, such as picocells [1], femtocells [2], hotspots [3], and cooperative relays [4], overlaying the

macrocells, to form a multi-tier heterogeneous wireless network (HWN) [5, 6], as illustrated in Fig. 1.1.

Each small-cell is equipped with a shorter-range and lower-cost base station (BS) or access point (AP),

to provide nearby users with higher bandwidth network access with lower power usage, and to offload

heavy data traffic from macrocells. In next generation wireless networks, this heterogeneity in network

structure is expected to become a key feature to improve network coverage and throughput. In practice,

commercial deployment of small-cells has attracted increasing attention. For example, AT&T Inc. has

released its femtocell product [7]. It has also deployed WiFi hotspots in a number of metropolitan areas

with dense population [8], to offload data traffic from macro-cellular networks.

1.1 New Challenges in Analysis and Design of HWNs

In the presence of the multiple tiers of BSs in wireless networks, new challenges are introduced due

to the random spatial patterns of BSs and random movement of users. First, small-cell BSs are often

installed incrementally and irregularly, with a high level of spatial randomness. Second, different tiers of

BSs communicate at different power levels, causing various scales of cell sizes, and complex cell shapes.

Finally, the movement of users may be individually arbitrary, without following any common mobility

pattern. As a consequence, many important mechanisms to operate the networks, such as mobility

management, user association, and interference control should be carefully reconsidered and redesigned

to accommodate such complex randomness.

1.1.1 Understanding User Distribution in HWNs

The distribution of active users is an important factor in the management and planning of wireless net-

works. However, the analytical modeling of user distribution is burdened with many technical challenges

due to the complex random spatial patterns of BSs and random movement of users. Moreover, the

1

Chapter 1. Introduction 2

(a) (b)

Figure 1.1: (a) Three separated tiers of BSs. From top to bottom: macrocell, picocell, and femtocellBSs. (b) A three-tier HWN.

session durations of users have a general probability distribution, and the channel holding times of users

at different cells are correlated, which further complicates the analysis of user distribution.

1.1.2 Characterization of Interference Patterns

An important challenge that limits the throughput in an HWN is the interference. Interference patterns

are more complicated in HWNs compared with conventional single-tier networks, as the deployment of

small-cell BSs induces more interference sources. Small-cell BSs are often randomly deployed. In the

downlink, it is not straightforward to quantify the sum interference from all interfering BSs. Uplink

interference analysis is even more challenging compared with the downlink case because (1) the spatial

patterns of interfering users are more complicated, and (2) uplink power control further burdens the

analysis by introducing a coupling effect between users and BSs. Therefore, in order to improve user

performance in HWNs, we are motivated to characterize the interference levels in the system and then

design new mechanisms to decrease them.

1.1.3 Design of User Association Rules

In order to improve the user data rate in the system, it is important to wisely design user-BS association

schemes in HWNs. An appropriate user association scheme should jointly consider the signal quality

and interference from the users’ perspective and load balancing from the BSs’ perspective, since a user’s

data rate is relevant to both its spectrum efficiency and the fraction of spectrum it can access.

A most direct approach is association by maximum received power, in which users are associated

with the BS (in any tier) with the highest received power. However, in this case, because small-cell

BSs transmit at lower power levels, only a small number of users close to them will connect with them,


−1 −0.5 0 0.5 1−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1(a)

x coordinate (km)

y c

oord

inate

(km

)

(a) Association by maximum received power.

−1 −0.5 0 0.5 1−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1(a)

x coordinate (km)

y c

oord

inate

(km

)

(b) Flexible user association.

Figure 1.2: An example of a three-tier HWN. Macrocell BSs, picocell BSs, and femtocell BSs arerepresented by squares, circles, and triangles, respectively; users are represented by dots; blue solid linesshow cell boundaries; dashed lines represent connections between users and BSs.

while most other users crowd in macrocells, leading to degraded performance. An example is shown in

Fig. 1.2(a), in which many users occupy the macrocells, while some small-cells are nearly empty.

In order to resolve this issue, a flexible user association approach (also called ranged expansion in some

literature) may be employed [9–13], in which each tier of BSs is assigned a user association bias value, and

a user is associated with a BS with the maximum received power multiplied by the bias value. If small-

cell BSs are assigned larger association bias values, the small-cells are “expanded” accordingly. This can

result in a more balanced mobile traffic pattern and thus better network performance. Fig. 1.2(b) shows

an example of flexible user association. However, if the association bias values for small-cell BSs are too

large, it will cause improper expansion of small-cells such that users at their cell-edge may suffer from

inadequate received power. As a consequence, we are motivated to properly design the association bias

values so that the overall network performance is optimized despite the randomness in BS location, user

location, channel condition, and interference levels.

1.1.4 Quantification of Handoff Patterns

User movement in HWNs introduces vertical handoffs, i.e., handoffs made between two BSs in different

tiers [14]. Compared with horizontal handoffs, i.e., handoffs made between two BSs in the same tier,

vertical handoffs impact both the users and the overall system in more complicated ways. For example,

extra traffic latency and additional network signaling are incurred during channel setup and tear down

when a vertical handoff is made; more user power may be consumed due to simultaneously active network

interface to multiple tiers; call drops or degraded quality of service (QoS) could also be experienced by

users due to the lack of radio resource after handoff. As a prerequisite to performance evaluation and

system design in HWNs, it is essential to quantify the rates of different handoff types. However, the

analysis on handoff rates in HWNs is complicated by the irregularly shaped multi-tier network topologies

introduced by the small-cell structure. It is difficult to characterize the cell boundaries and to track

boundary crossings made by users (i.e., handoffs) in the system.

Another design concern is the tradeoff between the handoff rates and the data rate. As shown in the

example in Fig. 1.3(a), a user starts a call at X and terminates it at Y . By choosing to access all of tier-1,

2, and 3 cells, it experiences one horizontal handoff at B1 and two vertical handoffs at B2 and B3. As


-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

x coordinate (km)

y c

oord

inate

(km

)

X

Y

B1

B2

B3

(a) A user starts a call at X and terminates it at Y . Itexperiences one horizontal handoff at B1 and two verticalhandoffs at B2 and B3.

-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

x coordinate (km)

y c

oord

inate

(km

)

C1

C2

X

Y

(b) Same BS locations and user trajectory. Tier-3 BSs arenot accessed. The same user experiences one horizontalhandoff at C1 and one vertical handoff at C2.

Figure 1.3: An example of a three-tier HWN and selective tier association. Tier-1, 2, and 3 BSs arerepresented by squares, circles, and triangles, respectively; blue curves show intra-tier cell boundaries;green curves show inter-tier cell boundaries.

shown in Fig. 1.3(b), by choosing to access tier-1 and 2 cells only, it experiences one horizontal handoff

at C1 and only one vertical handoff at C2. However, in the latter case, the user misses the opportunity

to access a high-bandwidth tier-3 BS. Thus, a user could choose to access high-bandwidth small-cell

BSs to improve data rate, but this may also lead to more frequent vertical handoffs, which potentially

deteriorates the service quality. Therefore, we are motivated to optimize the user tier selection scheme,

by taking both the handoff rates and the data rate into consideration.

1.2 Thesis Outline and Main Contributions

The goal of this thesis is to develop new methods and models to evaluate and design HWNs by investigat-

ing user load, interference patterns, and user mobility, the results of which provide new analytical insights

and design guidelines that will help improve future HWNs. The main contributions are summarized as

follows.

1.2.1 User Distribution under General Mobility and Session Patterns

In Chapter 3, we study the joint stationary distribution for the number of users in all cells in an

HWN. Prior studies have proposed several analytical models to estimate the user distribution with

various degrees of detail and generality [15–19]. Instead, we consider general mobility and session

patterns, only requiring that the new session arrivals form a Poisson process, which is well supported

by experimental data [19–21]. We model the user mobility with a general system with multiple routes,

each representing one type of users with a specific movement pattern. A general probability distribution

is used to represent the session durations. Moreover, the channel holding times at different cell sites are

no longer independent.

Through a decomposition-composition approach, we derive a closed-form expression for the joint

stationary distribution for the number of users in all cells. We observe five important conclusions on the

stationary user distribution: first, it is insensitive to how the users move through the system; second, it


is insensitive to the general distribution of channel holding times; third, it is insensitive to the correlation

among the channel holding times; fourth, it depends only on the average arrival rate and average channel

holding time at each cell; and fifth, it is perfectly captured by an open Jackson network with M/M/∞queues.

The conclusion of this chapter has important consequence to performance analysis and practical

system design. It suggests that accurate calculation of the user distribution, and other associated

metrics such as the system workload, can be achieved with much lower requirement for system parameter

estimation than previously expected.

1.2.2 Uplink Interference Comparison of Open Access and Closed Access

In Chapter 4, we focus on a two-tier wireless network with macrocells and femtocells. In the network,

two different access mechanisms may be applied: closed access and open access. Under closed access, a

femtocell BS only provides service to its local users, without further admitting nearby macrocell users.

In contrast, under open access, all nearby macrocell users are allowed to access the femtocell BS. We

study how the two access modes may affect both macrocell users and local femtocell users, in terms of

the uplink interference and outage probabilities. We present a stochastic geometric analysis framework

to derive numerical expressions for the uplink interference and outage probabilities of open access and

closed access by modeling macrocell BSs as a regular grid, macrocell users as a Poisson point process

(PPP), and femtocell users as a two-level clustered Poisson point process, which captures the spatial

patterns of different network components. However, uplink interference analysis is notoriously complex

for the two-tier network under consideration, our analysis yields non-closed forms requiring numerical

integrations. This motivates us to further develop closed-form sufficient conditions for open access and

closed access to outperform each other, at both the macrocell and femtocell levels.

Based on the above analysis, we are able to extract a key factor that influences the performance

difference between open access and closed access: the power enhancement factor ρ, which is defined as

the ratio, of the uplink targeted received power of an open access user in the femtocell, to its original

targeted received power in the macrocell. We investigate the threshold value ρ∗ (resp. ρ∗∗) such that

macrocell (resp. femtocell) users may benefit through open access if ρ < ρ∗ (resp. ρ < ρ∗∗) as we apply

open access to replace closed access. Upper and lower bounds of ρ∗ are derived in closed forms, and the

bounds of ρ∗∗ can be found by numerically searching through closed-form equations, providing system

design guidelines with low computational complexity.

1.2.3 Optimal Spectrum Allocation and User Association

In Chapter 5, our objective is to study jointly optimal spectrum allocation and user association in an

HWN with multiple tiers of BSs. First, we develop a stochastic geometric model to study the network

performance analytically. A closed-form expression for the average downlink user data rate is derived,

which is then employed as the objective function for jointly optimizing the spectrum allocation among

tiers and the user association bias values. This resultant optimization problem is of non-convex pro-

gramming in nature and cannot be solved with a standard method. Instead, we explore two important

structures in solving the problem. Referred to as the density thresholding structure, we show that the

problem can be studied separately over two regions, divided by a parameter specific user density thresh-

old. Referred to as the priority ordering structure, we show that a tier with higher BS density should


have higher priority in spectrum allocation. Based on these observations, we propose a computationally

efficient Structured Spectrum Allocation and User Association (SSAUA) approach to solve the problem

optimally and asymptotically optimally in the two density regions respectively. Finally, toward practi-

cal implementation of SSAUA, we propose a Surcharge Pricing Scheme (SPS), such that the designed

association bias values can be achieved in Nash equilibrium. Hence, each user is incentivized to adopt

the proposed design with individual rationality.

1.2.4 Handoff Rate Analysis in HWNs

In Chapter 6, we contribute to user mobility modeling and network access optimization in HWNs by

providing new technical tools to quantify the rates of horizontal and vertical handoffs, under random

multi-tier BSs, arbitrary user movement trajectory, and flexible user-BS association. A new stochastic

geometric analysis framework on user mobility is proposed. In this framework, different tiers of BSs are

firstly modeled as Poisson point processes (PPPs) to capture their spatial randomness in Section 6.1. To

model flexible scaling of cell sizes in different tiers, we consider the biased user association scheme dis-

cussed in Section 1.1.3. Through stochastic and analytic geometric analysis, we derive exact expressions

for the rates of all handoff types experienced by an active user with arbitrary movement trajectory. In

addition, as a study on the application of the above handoff rate analysis, after calculating the downlink

data rate of an active user given the set of BS tiers that it chooses to use, we further study optimal tier

selection by the user, considering both the handoff rates and the data rate. Finally, motivated by the fact

that some BSs are likely to aggregate around highly populated geographical regions (e.g., urban areas,

attractions, etc.), in Section 6.2, we extend the above handoff analysis by allowing BSs to form Poisson

cluster processes (PCPs) [22–24], in order to capture their non-uniform and dependent aggregation in

space.

Chapter 2

Related Works

In this chapter, we summarize the prior research in stochastic analysis and optimization of HWNs, and

discuss the relation between this thesis and the prior works.

2.1 User Mobility Model

2.1.1 Queueing Networks Models

In order to characterize user mobility, one common category of previous works employed queueing

systems to model wireless networks. In this case, cells are modeled as queues, users are modeled as units

in the queues, and handoffs correspond to unit transfers among queues. Through this approach, user

distributions were derived in different environments, such as wireless multimedia networks [16], vehicular

ad-hoc networks [17], and WLANs [18, 19]. However, these works did not allow arbitrary mobility or

arbitrary user session patterns. In terms of user movement, [16], [17], and [19] assumed that users move

from one cell to another probabilistically and memorylessly, while [18] focused on scattered single cells, so

that user movement among multiple cells was not discussed. None of them considered the arbitrary user

movement patterns. In terms of channel holding times, [16] used the sum of hyper-exponentials or the

Coxian distribution to approximate arbitrary distributions; [18] assumed generally distributed channel

holding times but concerned only a single cell; and [17] and [19] considered generally but independently

distributed channel holding times in different cells. Compared with our proposed model in Chapter 3,

none of the above works considered the dependence among channel holding times.

2.1.2 Insensitivity Property of Queueing Networks

The insensitivity of queueing networks indicates the situation where the stationary distribution remains

unchanged while the distribution of service times takes arbitrary forms. When the service times are

assumed independent among different queues, there are several well known conditions for insensitivity.

For example, networks with symmetric queues are insensitive [25]. In [26] and [27], the partial balance

of probability flows was shown to be a sufficient condition for insensitivity. In [28], partial reversibility is

shown to be a necessary and sufficient condition for insensitivity. However, none of these known results

considered the case where the service times between different queues are dependent. For example, the

queueing network closely related to our model in Chapter 3 is one with M/G/∞ queues. It is known to

7

Chapter 2. Related Works 8

be insensitive when the service times are independent [25], but to the best of our knowledge, there is no

further general result for dependent service times.

2.1.3 Mobility Modeling with Cell Geometry

In order to characterize the geometric patterns of network topologies, another category of works modeled

the shape of cells, mostly in non-random regular grids. Zonoozi and Dassanayake [29] modeled a one-

tier cellular network as a hexagonal grid. Anpalagan and Katzela [30] studied a two-tier network by

modeling small-cells as hexagons, and each macrocell as a cluster of neighbouring small-cells. Shenoy

and Hartpence [31] studied a two-tier network by modeling WLAN small-cells as squares, and macrocells

as larger squares, each covering 5× 5 WLAN cells. Hasib and Fapojuwo [32] studied a two-tier cellular

network including one hexagonal macrocell and a predetermined number of circular microcells.

To further capture the spatial randomness of network topologies, Lin et al. [33] conducted a pioneering

study on the user mobility in a one-tier cellular network with randomly distributed BSs, where the BSs

are modeled as a homogeneous Poisson point process (PPP), and cell splitting is modeled as a standard

Poisson Voronoi. Our work in Section 6.1 extends the above study to the multi-tier case, where each

tier of BSs is modeled as a homogeneous PPP, and the resultant cell splitting is modeled as a weighted

Poisson Voronoi. Note that our studies in Section 6.1 follow conventional stochastic geometric analysis

of HWNs, where the PPP is commonly used to model the distribution of BSs to capture their spatial

randomness (e.g., [34–36]). However, in reality, a higher density of BSs are often installed in more

populated regions, hence the PPP assumption may not be accurate in such a scenario. In order to

resolve this issue, in Section 6.2, as an extension of Section 6.1, we further model some tiers of BSs as

Poisson cluster processes (PCPs) [22–24], in order to capture the non-uniform and dependent aggregation

of BSs.

2.2 Interference Analysis of HWNs

2.2.1 Stochastic Geometry as a Basic Tool

Stochastic geometry [37–41] is an important mathematical tool to characterize random spatial patterns

of wireless networks. In recent years, this tool has succeeded to develop tractable models to analyze

interference and provide design guidelines for HWNs. Through standard stochastic geometric approach,

interferers are modeled as a Poisson point process (PPP), and their interference can be analyzed as the

shot noise on the two-dimensional Euclidean space R2. Then, the Laplace transform of the shot noise

can be derived directly from the Laplace functional [37,38] or the generating functional [39] of the PPP.

2.2.2 Downlink Interference

The downlink interference and outage performance in wireless networks have been extensively studied

using the stochastic geometric approach. Dhillon et al. [35, 42] analyzed the downlink performance of

heterogeneous networks with multiple tiers by assuming the signal-to-interference plus noise ratio (SINR)

threshold is greater than 1. Keeler et al. [43] extended the work by allowing the SINR threshold to be

less than 1. Kim et al. [44] studied the maximum tier-1 user and tier-2 cell densities under downlink

outage constraints. Dhillon et al. [45] studied the downlink interference considering load balance. Singh


et al. [13] studied the downlink user achievable rate in a heterogeneous network considering both SINR

and spatial user distributions. Jo et al. [46] studied open access versus closed access in femtocell networks

in terms of downlink performance.

2.2.3 Uplink Interference

The analysis of uplink interference in multi-tier networks is more challenging compared with the downlink

case. For uplink analysis, the interference generators are the set of users, which are more complicatedly

distributed compared with the interference generators (i.e., BSs) in downlink analysis. Under closed

access, without considering random spatial patterns, Kishore et al. [47] studied the uplink performance

of a single tier-1 cell and a single femtocell, while the same authors [48] extended it to the case of multiple

tier-1 cells and multiple femtocells. An and Pianese [49] studied the co-channel uplink interference in

LTE-based multi-tier cellular networks, considering a constant number of femtocells in a macrocell.

However, none of [47–49] considered the random spatial patterns of users or femtocells.

By considering random spatial patterns, Novlan et al. [50] analyzed the uplink performance of cellular

networks, but it was limited to the one-tier case. The work was extended in [51] where frequency reuse

among edge users and inner users were accommodated. Chakchouk and Hamdaoui [52] studied the

two-tier case with one macrocell and randomly spatially distributed femtocells and femtocell users.

Chandrasekhar and Andrews [53] evaluated the uplink performance of two-tier networks considering

multiple macrocells, femtocells, macrocell users, and femtocell users. However, several interference

components were analyzed based on approximations, such as BSs see a femtocell as a point interference

source, and femtocell users transmit at the maximum power at the edge of cells. Cheung et al. [54]

studied both uplink and downlink interference of femtocell networks based on a Neyman-Scott Process.

However, it assumed that each user transmits at the same power and femtocell users are uniformly

distributed in an infinitesimally thin ring around the femtocell BS. With a more general system model,

we [55] derived the uplink interference in a two-tier network with multiple types of users and small-cell

BSs, but no closed-form result was obtained. Note that [53–55] considered only the closed access case.

Compared with the closed access mode, the uplink interference analysis of the open access mode

is even more complicated. This is because the model for open access needs to capture the impact of

the users disconnecting from the original macrocell BS and connecting to a small-cell BS. In order

to achieve mathematical tractability, the previous analysis of open access used simplified assumptions.

Xia et al. [56] compared the performance of open access and closed access based on a model with

one macrocell, one femtocell, and a given number of macrocell users, and Tarasak et al. [57] used a

model with one macrocell, a constant number of macrocell users, and randomly distributed femtocells.

Although [56] and [57] provide useful insights into the performance comparison between open access

and closed access, their limited system models cannot account for the challenging issues brought by the

diverse spatial patterns of BSs and users. Zeinalpour-Yazdi et al. [58] studied a model with one macrocell

and randomly spatially distributed macrocell users and femtocells. The model in [58] is still limited to

a single-macrocell scenario where the spatial patterns of macrocell BSs and inter-macrocell interference

cannot be accommodated. ElSawy and Hossain [59] studied the uplink outage performance of multi-tier

cellular networks with truncated channel inversion power control. Open access was considered in [59],

but approximations had to be used for mathematical tractability, e.g, the correlation of the locations of

interfering users and their transmit power levels were ignored. Focusing on a different scope from our

analysis in Chapter 4, [59] did not compare the performance of open access and closed access.


Another category of previous works employed geometrical probability models to characterize inter-

ference patterns in mobile networks. Geometrical probability models are employed when the cell shapes

are predetermined (e.g., triangles [60], rhombuses [61], hexagons [62], trapezoids [63], and other polygon-

s [64]). On the other hand, stochastic geometry models are more advanced when the network topology

is random.

2.3 Spectrum Allocation and User Association

Flexible user association is an important approach to improve user data rate in HWNs, which has

attracted much attention in recent years. Through flexible user association approach, each tier of BSs

is assigned a user association bias value, and a user is associated with a BS with the largest received

power multiplied by the bias value [9–13]. If small-cell BSs are assigned larger association bias values,

the small-cells are “expanded” accordingly. An example is shown in Fig. 1.2.

Assuming a fixed number of users and BSs and without considering their random spatial patterns,

the authors of [10, 11] investigated performance benefit introduced by flexible user association; and the

authors of [65–70] studied optimal user association with deterministic utility optimization.

With a stochastic geometric approach, some prior studies considered either spectrum allocation or

user association separately. For example, Cheung et al. [36] studied optimal spectrum allocation, in

cellular networks limited to two tiers of BSs, without flexible user association. We [71, 72] studied

optimal spectrum allocation among different tiers of BSs in HWNs, without further optimizing user

association bias values. Jo et al. [12] presented the performance evaluation of flexible user association

model with bias values. They derived the coverage probability and user data rate, considering cross-tier

interference, but in non-closed forms. They did not provide a means to optimize the derived performance

metrics. This work was later extended in [13] to study optimal user association in a network with two

tiers of BSs, without considering spectrum allocation.

Through stochastic geometric analysis, spectrum allocation and user association were jointly studied

in [73–76]. Singh and Andrews [73] analyzed the network performance in terms of coverage probability

and data rate under different spectrum allocation and user association settings. However, the optimal

spectrum allocation and user association were not derived. A similar problem was also studied by Lin

and Yu [74], with frequency reuse instead of tiered spectrum division as the approach for spectrum

sharing.

Different from the above studies, in Chapter 5, we consider multiple tiers of BSs with disjoint spec-

trum, and provide optimal and analytically bounded asymptotically optimal solutions for joint spectrum

allocation and user association. The closest works to our analysis in Chapter 5 may be [75, 76], where

joint spectrum allocation and user association was also considered. However, [75, 76] aimed to optimize

the average log-utility per user, such that the logarithm function cancels the exponential term corre-

sponding to the Laplace transform of interference [37, 38]. This alternate objective led to substantial

difference in the optimization problem formulation and solution. Different from [75, 76], our model in

Chapter 5 aims to optimize the average user data rate directly, leading to an even more challenging

problem and requiring the proposed structured optimization solution. Furthermore, only the single

modulation and coding scheme (MCS) was considered in [75, 76], while we additionally study multiple

MCSs in this Chapter.


2.4 A Brief Review on Queueing Networks

In this section, we briefly review some preliminaries for queueing networks [77], which are employed in

the analysis in Chapter 3.

2.4.1 Queueing Network under Consideration

Consider an m-node queueing network with single unit movements. Let x = [x1, x2, . . . , xm]T be an

m-dimensional vector representing the state where there are xi units at node i, ∀i = 1, 2, . . . ,m. Let Sdenote the state space. The network may be one of the following types1:

• Closed network with A units: S = x : |x| = A.

• Open network with unlimited capacity: S = x : |x| < ∞.

• Open network with capacity A: S = x : |x| ≤ A.

In the queueing network, the units are moving in the node setM . In a closed network,M = 1, 2, . . . ,M.In an open network, M = 0, 1, 2, . . . ,M, where node 0 represents the outside of the network.

Let ei be the unit vector where the ith entry is 1. We define Tjkx , x − ej + ek if j = k, j = 0,

and k = 0; Tjkx , x + ek if j = 0 and k = 0; and Tjkx , x − ej if j = 0 and k = 0. In state x,

the time to the next movement of a single unit from node j to node k (i.e., transition from x to Tjkx)

is independently exponentially distributed with rate λjkϕj(x), where the term λjk denotes the routing

intensity from node j to node k, j = k, and the term ϕj(x) denotes the service rate at node j when

the network is in state x. Note that if the network is open, λ0kϕ0(x) represents the arrival rate to node

k from the outside. In a Jackson network, ϕj(x) is a function ϕj(xj) only of xj , ∀j = 1, 2, . . . ,m, and

ϕ0(·) ≡ 1 when the network is open.

2.4.2 Routing Balance Equations

Let wj , j ∈ M , denote a positive invariant measure that satisfies the routing balance equations as follows

wj

∑k∈M

λjk =∑k∈M

wkλkj . (2.1)

When the network is open, we adopt the convention that w0 = 1.

2.4.3 Stationary Distributions of Jackson Networks

The stationary distribution of a closed Jackson network with A units is

π(x) = cm∏j=1

wxj

j

xj∏n=1

ϕj(n)−1, ∀x ∈ S = x : |x| = A, (2.2)

where c is the normalizing constant given by

c−1 =∑x∈S

m∏j=1

wxj

j

xj∏n=1

ϕj(n)−1. (2.3)

1Note that |x| ,∑

i xi in Section 2.4.


Substituting the definition S = x : |x| ≤ A into (2.2) and (2.3), we derive the stationary distribution

of an open Jackson network with capacity A.

The stationary distribution of an open Jackson network with unlimited capacity is given as follows.

First, suppose we have

cj =∞∑

xj=0

wxj

j

xj∏n=1

ϕj(n)−1 < ∞, ∀j = 1, 2, . . . ,m. (2.4)

Then the stationary distribution is

π(x) = π1(x1)π2(x2) . . . πm(xm), ∀x ∈ S, (2.5)

where

πj(xj) = cjwxj

j

xj∏n=1

ϕj(n)−1. (2.6)

From (2.5), we notice that the joint distribution is the product of the marginal distributions among

individual queues.

2.5 A Brief Review on Stochastic Geometry

Stochastic geometry is an important mathematical tool to characterize random spatial patterns of wire-

less networks. In recent years, this tool has succeeded to develop tractable models to analyze perfor-

mance and provide design guidelines for HWNs. In this section, we briefly review some preliminaries for

stochastic geometry [37–39], which are employed in the analysis in Chapters 4-6.

2.5.1 Interference Analysis Based on Poisson Point Process

Suppose interferers are modeled as a Poisson point process (PPP) on the two dimensional Euclidean

space R2. Their sum interference can be analyzed as the shot noise of the PPP. The Laplace transform

of the shot noise can be derived directly from the Laplace functional or the generating functional of the

PPP.

Consider the following simple example: assume that mobile users form a homogeneous PPP Φ with

intensity λ on R2. Each user transmits at power level P . Let L(x,y) denote the path loss function from

coordinate x to y. Consider an observer located at y0, which regards all the mobile users as interferers.

In this case, because the mobile users form a PPP, which is a random point process, the overall

interference I is a random variable. To investigate the distribution of I, we can focus on its Laplace

transform L(s) = E(e−sI). Because I can be computed as the sum of interference from each user, we

have I =∑

x∈Φ PL(x,y0). Thus,

L(s) = E(e−

∑x∈Φ sPL(x,y0)

)= E

(∏x∈Φ

e−(sPL(x,y0))

). (2.7)

The generating functional of a PPP is defined as G(v) = E(∏

x∈Φ v(x)) for an arbitrary function v,


which can be computed as G(v) = exp(−λ∫R2(1− v(x))dx

). Substituting it into (2.7), we have

L(s) = exp

(−λ

∫R2

(1− e−sPL(x,y0))dx

). (2.8)

In the above steps, the sum form of the interference is converted to the product form through the

exponential function, which then matches the form of the generating functional of the PPP. Through

this way, the sum interference in the system is analytically characterized.

2.5.2 Random Fibre Process

A random fibre process is defined as a random collections of curves on R2. A fibre process is stationary

if its distribution is invariant under translation through any vector. It is isotropic if its distribution is

invariant under rotations about the origin.

Let Ψ denote a stationary and isotropic fibre process on R2. Then, the intensity of Ψ, LA(Ψ), is

defined as its mean line length per unit area.

LA(Ψ) = E(∣∣∣Ψ∩B

∣∣∣1

), (2.9)

where |·|1 represents one-dimensional Lebesgue measure, and B is an arbitrary region with a unit area.

Let T be an arbitrary curve on R2 with finite length L. Then, the expected number of intersections

between T and Ψ is equivalent to (a generalized version of the Buffon’s needle problem)

2

πLA(Ψ)L. (2.10)

2.6 Publications Related to this Thesis

A version of Chapter 3 is published in [78, 79]; a version of Chapter 4 is published in [80, 81]; a version

of Chapter 5 is published in [82, 83]; a version of Section 6.1 is published in [84, 85]; and a version of

Section 6.2 is published in [86].

Chapter 3

Insensitivity of User Distribution in

HWNs

In this chapter, we study the joint stationary distribution for the number of users in all cells in a

multicell network, which has important utilization in network management and planning. As discussed

in Section 2.1.1, to facilitate tractable analysis, existing studies often adopt simplified models. Instead,

we consider more general user mobility and session patterns, only requiring that the new session arrivals

form a Poisson process, which is well supported by experimental data [19–21]. We model the user

mobility with a general system with multiple routes, each representing one type of users with a specific

movement pattern. A general probability distribution is used to represent the session durations. The

channel holding times at different cell sites are generally distributed and are no longer independent.

Through a decomposition-composition approach, we derive a closed-form expression for the joint

stationary distribution for the number of users in all cells. In spite of the general user movement

patterns and complicated dependency and distribution of channel holding times, we observe that the

stationary user distribution (1) is insensitive to the user movement patterns, (2) is insensitive to general

and dependently distributed channel holding times, (3) depends only on the average arrival rate and

average channel holding time at each cell, and (4) is completely characterized by an open network with

M/M/∞ queues. We use the Dartmouth trace to validate our analysis, which shows that the analytical

model is accurate when new session arrivals are Poisson and remains useful when non-Poisson session

arrivals are also included in the data set. Our results suggest that accurate calculation of the user

distribution, and other associated metrics such as the system workload, can be achieved with much

lower complexity than previously expected.

3.1 System Model

Consider a multicell network with C cells. There are L unique routes, each defined as a finite ordered

sequence of cells. The jth stage on the lth route corresponds to the jth cell in the sequence, which is

denoted as c(l, j). Let Nl be the number of stages on the lth route. Each user of the lth route starts

a new session in cell c(l, 1); then it moves along the route through cells c(l, 1), c(l, 2) . . . c(l, Nl), as long

as the session remains active. The user is considered to have departed the network when its session

terminates or when it exits cell c(l, Nl). We allow an arbitrary number of arbitrary routes to cover all

14

Chapter 3. Insensitivity of User Distribution in HWNs 15

handoffT1

τ12 τ13τ11

t11 t12 t13

T1

Cell 3

Cell 4

Cell 6

Cell 2Route 3

Cell 5

Route 1

τ11 = t11τ12 = t12 t13

τ13

Route 2

start

Cell 1

handoff terminate

Figure 3.1: System model.

possible movement patterns.

For each route, we assume the arrivals of new sessions to form a Poisson process. Note that although

the arrivals of packets in the Internet may not form Poisson processes [87], the arrivals of new sessions

are at a much larger time scale and are well justified as Poisson [20,21]. Furthermore, in [19] and later in

Section 3.4, experimental data show that new sessions in the type of mobile networks under consideration

are indeed Poisson barring some extreme cases. We emphasize that only the new session arrivals are

Poisson, while the handoff arrivals at each cell have general statistics with complicated dependencies.

The session duration of a user on the lth route is modeled as an arbitrarily distributed random

variable Tl. Let λl0 be the new session arrival rate at the lth route. After a new session arrival, let

τl1 denote the residual cell dwell time of the user in the 1st stage on the lth route, which is arbitrarily

distributed. Let τlj , 2 ≤ j ≤ Nl, denote the cell dwell time of the user in the jth stage on the lth route,

which are also arbitrarily distributed. Then, the channel holding time of the jth stage on the lth route,

tlj , if it exists, can be represented as follows:

tlj =

minTl, τl1, if j = 1,

minTl −j−1∑i=1

τli, τlj, if Tl >∑j−1

i=1 τli, 2 ≤ j ≤ Nl.(3.1)

Fig. 3.1 shows an example network with 3 routes. Route 1 starts from cell 1 and passes cell 3, 4

and 6 (i.e., c(1, 1) = 1, c(1, 2) = 3, c(1, 3) = 4 and c(1, 4) = 6). A user starts a session in cell 1, and

the session is terminated in cell 4. The corresponding T1, τ11, τ12, τ13, t11, t12, and t13 are labeled in the

figure.

Note that a user is being active from the initiation of its session until the termination of its session.

Given a route l, we know the sessions start at cell c(l, 1) but the cell where they end is random since

the session duration is random. Furthermore, we do not assume independence between Tl and τlj , and

the channel holding times tlj are not independent either. Finally, each route defines the user movement


trace and the distribution of channel holding times, which implicitly characterizes the speed of users on

this route.

Let xlj , 1 ≤ l ≤ L, 1 ≤ j ≤ Nl, denote the number of active users in the jth stage on the lth route; let

yn, 1 ≤ n ≤ C, denote the number of active users in the nth cell. Let x = [xlj : 1 ≤ l ≤ L, 1 ≤ j ≤ Nl]T

and y = [y1, y2, . . . , yC ]T . We aim to derive π(x) and π1(y), the joint stationary user distributions for

active users x and y, respectively. Note that since π(x) and π1(y) are defined in the steady state, we

explicitly ignore any temporal fluctuation in these distributions.

3.2 Stationary User Distribution in Single-Route Network

We first derive the stationary user distribution on a single route. We construct a reference single-

route memoryless network, where all the channel holding times are independently and exponentially

distributed. We prove insensitivity by showing an equivalence between the original network and the

memoryless network in terms of stationary user distribution.

3.2.1 Queueing Network Model for Single-Route Network

...

. . .

(a) Single-route network.

λl0

tl1 tl2 tl3 tlNl

0

(b) Reference single-route memoryless network.

. . .λl0

pl1∑Nlj=1 plj

λl1

pl2∑Nlj=2 plj

λl2

0

λlNl

pl(Nl)

pl(Nl−1)+plNl

λl(Nl−1)

∑Nlj=2 plj∑Nlj=1 plj

λl1

∑Nlj=3 plj∑Nlj=2 plj

λl2

pl(Nl−1)

pl(Nl−1)+plNl

λl(Nl−1)

...

Figure 3.2: Single-route network.

Consider exclusively the lth route in the network. Throughout Section 3.2, we will carry the route

index l in most symbols, since they will be re-used in the analysis of multiple-route networks.

As shown in Fig. 3.2(a), we model the route as a tandem-like queueing network. The node labeled

with 0 represents the exogenous world. The jth queue, 1 ≤ j ≤ Nl, represents the jth stage of the route,

and units in this queue represent sessions in the jth stage. Each queue has infinite servers, since the


sessions are served in parallel with no waiting1.

The channel holding time of a session in the jth stage, tlj , is equivalent to the service time of the

jth queue. The handoff of a session from the jth stage to the (j + 1)th stage is equivalent to a unit

movement from the jth queue to the (j + 1)th queue. The termination of a session is equivalent to the

movement from a queue to node 0.

Let plk denote the probability that a session lasts for k stages. It is given by

plk = P[ k−1∑j=1

τlj < Tl ≤k∑

j=1

τlj

], for 2 ≤ k ≤ Nl − 1,

with pl1 = P [Tl ≤ τl1] and plNl= P

[∑Nl−1j=1 τlj < Tl

]. Note that we have

∑Nl

k=1 plk = 1. Given a session

in the kth stage, it enters the (k+1)th stage with probability∑Nl

j=k+1 plj∑Nlj=k plj

and terminates with probability

plk∑Nlj=k plj

.

3.2.2 Reference Single-Route Memoryless Network

We define a reference single-route memoryless network, as a Jackson network with the same topology

as the original single-route network, where each queue has infinitely many independent and exponential

servers. An illustration is shown in Fig. 3.2(b). By matching the mean service times in this memoryless

network with those of the original network, we see that its external arrival rate is λl0, the service rate

of the jth queue is λlj = 1tlj

. The routing probability from the kth queue to the (k + 1)th queue is the

probability that a session enters the (k+1)th stage conditioned on it is in the kth stage,∑Nl


. The

routing probability from the kth queue to node 0 is plk∑Nlj=k plj

. Thus, the service rate from the kth queue

to the (k+1)th queue is∑Nl


λlk, and the service rate from the kth queue to node 0 is plk∑Nlj=k plj

λlk.

Let w′lj denote the positive invariant measure of the jth queue that satisfies the routing balance

equations of the single-route memoryless network. w′0 is the positive invariant measure of node 0. We

adopt the convention that w′0 = 1. It can be derived from the topology of Fig. 3.2(b) that

λl0w′0 =w′

l1, (3.2)∑Nl

n=j pln∑Nl

n=j−1 plnw′

lj−1 =w′lj , 2 ≤ j ≤ Nl, (3.3)

which leads to

w′l1 =λl0, (3.4)

w′lj =λl0(1−

j−1∑n=1

pln), 2 ≤ j ≤ Nl. (3.5)

Because each queue has infinite servers, the departure intensity at the jth queue is λljxlj when there

1Users move into and out of each cell in parallel. Therefore, when considering the channel holding time as the servicetime of a queue that models mobility, this is equivalent to all users being served at the same time by its own dedicatedserver, which is the same as having infinite servers. This model is accurate for communication systems with no admissioncontrol (e.g., WiFi) and gives reasonable approximation to systems with many available channels.


tlNl21

λl110pl1λl0

tl2Ml21 tl2Ml22

......tl221 tl222

λl220

tl211 tl212

pl2λl0

...

λl120

λl210

λl1Ml10

tl1Ml11

tl111

tl121

0

. . .

tlNlMlNl1 tlNl2Nl

...

tlNlMlNl2 tlNlMlNl

3

tlNl22 tlNl23 tlNlMlNlNl

tlNl1Nl

.........

. . .

tlNl12 tlNl13

. . .

.... . .

plNlλl0

λlNl10

λlNlMlNl0

λlNl20

λl2Ml20

tlNl11

Figure 3.3: Decoupled network.

are xlj users in it. Let wlj =w′

lj

λlj. Then the stationary user distribution w.r.t. x of this network is [77]

π0(x) =

Nl∏j=1

e−wljwxlj

lj

1

xlj !. (3.6)

3.2.3 Insensitivity of Single-Route Network

For the original single route network, we employ a decomposition-composition approach to derive its

stationary user distribution.

Given that one session lasts for k stages, we denote the channel holding times as a k-dimensional

random vector tlk = tlk1, . . . tlkj , . . . , tlkk, where tlkj is the channel holding time at the jth stage.

We assume that tlk is an arbitrarily distributed discrete random vector with Mlk possible realizations2.

For any i, 1 ≤ i ≤ Mlk, we define a k-dimensional deterministic vector tlki = [tlki1, . . . , tlkij , . . . , tlkik]T

corresponding to the ith realization of tlk. Let qlki be the probability of the ith realization given that

the session lasts for k stages. Also, let Plki = plkqlki denote the probability that a session lasts for k

stages and it is in the ith realization.

By doing so, we decompose the original network into a multiple-branch queueing network as shown in

Fig. 3.3, which is referred to as the decoupled network. In this network, there are Nl main branches, where

the kth main branch represents the event that a session lasts for k stages. The kth main branch contains

2For a vector of continuous channel holding times, we can use a sequence of discrete distributions with decreasinggranularity to approach its distribution. The granularity does not influence the computational complexity of the stationaryuser distribution of the original multicell network. See Section 3.3 for more details.


λlNlMlNlNl

0

λl110pl1λl0

λl2Ml21 λl2Ml22

......λl221 λl222

λl220

λl211 λl212

pl2λl0

...

λl120

λl210

λl1Ml10

λl1Ml11

λl111

λl121

. . .

...

λlNlMlNl3

λlNl22 λlNl23

λlNl1Nl

.........

. . .

λlNl12 λlNl13

. . .

.... . .

plNlλl0

λlNl10

λlNlMlNl0

λlNl20

λl2Ml20

λlNl11

λlNl2NlλlNlMlNl

2

λlNl21

λlNlMlNl1

Figure 3.4: Reference memoryless decoupled network.

Mlk sub-branches, where the ith sub-branch represents the realization where tlk = tlki. Furthermore,

the jth queue in the ith sub-branch of the kth main branch represents the jth stage of the ith realization

of the sessions that last for k stages.

Hence, each queue of the decoupled network has infinite servers with deterministic service time, tlkij ,

for the jth stage of the ith sub-branch of the kth main branch. Furthermore, the arrival rate of the ith

sub-branch of the kth main branch is λlki0 = Plkiλl0. Let x = [xlkij : 1 ≤ k ≤ Nl, 1 ≤ j ≤ k, 1 ≤i ≤ Mlk]T be the vector of number of sessions in the jth stage of the ith sub-branch of the kth main

branch. Denote by πD(x) the stationary user distribution of the decoupled network.

Note that the stationary distribution of a Jackson network with infinite servers at each queue is

insensitive with respect to the distribution of the service times [26]. Therefore, πD(x) remains unchanged

if we create a reference Jackson network by replacing each queue with deterministic service time in the

decoupled network with a queue that has exponential distributed memoryless service time with the same

mean (e.g., the service rate at the jth queue of the ith sub-branch of the kth main branch λlkij =1

tlkij),

as shown in Fig. 3.4, which is referred to as the reference memoryless decoupled network.

Let w′lkij be the positive invariant measure of the jth queue of the ith sub-branch of the kth main

branch of the reference memoryless decoupled network, which satisfies the routing balance equations

with the convention that at node 0, w′0 = 1. Since each sub-branch is a chain network, we have

w′lkij = Plkiλl0. (3.7)

Let wlkij =w′

lkij

λlkij. Then the stationary user distribution of the decoupled network as well as the reference


memoryless decoupled network is

πD(x) =

Nl∏j=1

Nl∏k=j

Mlk∏i=1

e−wlkij wxlkij

lkij

1

xlkij !. (3.8)

The stationary user distribution of the original single route network, π(x), is the sum of πD(x)

satisfying xlj =∑Nl

k=j

∑Mlk

i=1 xlkij , ∀j. To derive π(x), we first introduce the following lemma.

Lemma 1. Consider a stationary open Jackson network with N queues each with an infinite number of

servers. Let xj be the number of units in the jth queue and x = [x1, . . . xN ]T . Suppose J1,J2, . . .JMis a set of mutually exclusive subsets of 1, 2, . . . , N. Let zi =

∑j∈Ji

xj , i = 1, 2, . . . ,M , denoting the

sum of units in the queues inside Ji. Then, the distribution of z = [z1, . . . zM ]T is

π(z) =

M∏i=1

e−vivzii1

zi!, (3.9)

where vi =∑

j∈Jiwj , and wj is the expected number of units in the jth queue.

Proof. For a Jackson network with infinite servers at each queue, the stationary queue lengths are

independent Poisson random variables with mean wj for the jth queue. Hence, zi is Poisson with mean

vi =∑

j∈Jiwj for all i. Furthermore, since Ji are mutually exclusive, zi are independent.

Next, we note that the expected service time spent in the jth stage given that the jth stage exists,

i.e., j ≤ k for the kth main branch, can be computed as

tlj =

∑Nl

k=j

∑Mlk

i=1 Plkitlkij∑Nl

k=j

∑Mlk

i=1 Plki

=

∑Nl

k=j

∑Mlk

i=1 Plkitlkij

1−∑j−1

n=1 pln. (3.10)

Combining this with (3.7), we have

Nl∑k=j

Mlk∑i=1

wlkij =

Nl∑k=j

Mlk∑i=1

λl0Plki

λlkij

=

Nl∑k=j

Mlk∑i=1

λl0Plkitlkij

= λl0(1−j−1∑n=1

pln)tlj

=λl0

λlj(1−

j−1∑n=1

pln)

= wlj . (3.11)


tLNL

. . .λ10

λL0

λ20...

· · ·

...

...t11 t1N1

t13t12

. . .

0

tL1 tL2 tL3

Figure 3.5: Multiple-route network.

Therefore, by Lemma 1, we have

π(x) =∑

x:xlj=∑Nl

k=j

∑Mlki=1 xlkij ,∀j

πD(x)

=

Nl∏j=1

e−wljw

xlj

lj

xlj !, (3.12)

which is restated as the following theorem:

Theorem 1. The single-route network has the same stationary user distribution as that of the corre-

sponding single-route memoryless network: π(x) = π0(x).

3.3 Stationary User Distribution in Multiple-Route Network

In this section, we study the general case with multiple routes. We first extend the results from the

previous section to show π(x) = π0(x) in a multiple-route network. We then derive the stationary user

distribution π1(y) with respect to cells and show its insensitivity.

3.3.1 Queueing Network Model for Multiple-Route Network

Since the L routes are independent, we model the multiple-route network as a paralleling of L single-

route networks, as shown in Fig. 3.5. Similar to Section 3.2, we consider a reference multiple-route

memoryless network, which is a paralleling of L corresponding single-route memoryless networks. Then,

we construct the decoupled multiple-route network, which is a paralleling of L corresponding single-route

decoupled networks.

3.3.2 Insensitivity of π(x)

Theorem 2. The multiple-route network has the same stationary user distribution as that of the

corresponding multiple-route memoryless network.

Proof. Since the routes are independent, the stationary user distribution of the multiple-route network


can be computed as the product of the stationary user distribution of single-route networks:

π(x) =

L∏l=1

Nl∏j=1

e−wljwxlj

lj

1

xlj !. (3.13)

Since the same holds for the multiple-route memoryless network, we have π(x) = π0(x).

3.3.3 Insensitivity of π1(y)

Let λn be the average total arrival rate to cell n, including both new and handoff arrivals. Let tn be the

average channel holding time in cell n, considering all routes and stages. Thus,

λn =∑

l,j:c(l,j)=n

Nl∑k=j

Mlk∑i=1

λl0Plki, (3.14)

tn =

∑l,j:c(l,j)=n

∑Nl

k=j

∑Mlk

i=1 λl0Plkitlkij∑l,j:c(l,j)=n

∑Nl

k=j

∑Mlk

i=1 λl0Plki

. (3.15)

Then from (3.11), we have

λntn =∑

l,j:c(l,j)=n

Nl∑k=j

Mlk∑i=1

λl0Plkitlkij

=∑

l,j:c(l,j)=n

Nl∑k=j

Mlk∑i=1

wlkij

=∑

l,j:c(l,j)=n

wlj . (3.16)

The joint stationary user distribution among all cells can be computed as a summation over those

entries of π0(x) satisfying yn =∑

l,j:c(l,j)=n xlj , ∀n. Then from Lemma 1, we obtain

π1(y) =∑

x:yn=∑

l,j:c(l,j)=n xlj ,∀n

L∏l=1

Nl∏j=1

e−wljwxlj

lj

1

xlj !

=∏n

e−(∑

l,j:c(l,j)=n wlj)

∑l,j:c(l,j)=n

wlj

yn

1

yn!

=∏n

e−(λntn) (λntn)yn 1

yn!. (3.17)

We make the following observations from (3.17):

• The marginal distribution within a single cell depends only on the average arrival rate and average

channel holding time at that cell.

• The number of users in each cell is independent and Poisson. This is in accordance with Theorem

9.27 in [77].

• The stationary user distribution depends only on the average arrival rates and average channel


holding times in individual cells, having the exact same form of an M/M/∞ open Jackson network.

It is insensitive with respect to the distribution of channel holding times, or the correlation among

them. Furthermore, it is insensitive with respect to movement patterns, since the exact routing in

the network is irrelevant.

In reality, λn and tn can be easily obtained at the base station or access point of cell n by tracking

user arrivals and departures at the cell. Then the base station or access point can report the two local

values λn and tn to a central controller. After knowing λn and tn values of all cells, the central controller

can easily obtain the joint user distribution by (3.17). Obviously, the above approach facilitates efficient

system management and planning in practice, helping to avoid the need for collecting a large amount of

user location data. Note that in Section 3.2.3, for continuous channel holding times, we use a sequence of

discrete distributions with decreasing granularity to approach its distribution. The granularity does not

influence the computational complexity of the stationary user distribution in (3.17), since the discrete

distributions are only used in an intermediate step to prove the insensitivity property. We do not need

to explicitly compute the user distributions from the discrete distributions.

3.4 Experimental Study

In this section, our analysis is validated via experimenting with real-world traces. We first present the

data source and experimental settings. We then compare the experimental and analytical results.

3.4.1 Requirements and the Dartmouth Traces

There are serval publicly available traces online, including the Dartmouth traces [88–90], the UCSD

traces [91], the IBM-Watson traces [92], and the Montreal traces [93]. To choose proper traces, we

need to consider the following requirements. First, there should be a large amount of sample points to

facilitate an estimation of the user distribution by relative frequency, which is to be compared with the

distribution derived by the proposed analysis. Note that the support of the user distribution increases

exponentially with the number of cells. Most available traces do not have a large enough data set.

Second, the location of cells should be close enough so that there is enough handoff traffic among them

to create strong dependency between channel holding times. Data from already independently operated

cells can be analyzed using exiting techniques and hence are not challenging enough to test our analytical

model. To the best of our knowledge, the Dartmouth traces are the most recent public traces satisfying

both requirements. They have been widely studied in the literature [19, 94–96]. We use data from the

academic area in the Dartmouth traces [90], a comprehensive record of network activities in a large

wireless LAN (using 802.11b) in Dartmouth College. The traces includes the data of 152 APs and more

than 5000 users, during a 17-week period (Nov. 1, 2003 to Feb. 28, 2004). Most users are students

walking on campus. We focus on the Simple Network Management Protocol (SNMP) logs of the traces,

which are constructed every five minutes, when each AP polls all users attached to it. Each polling

message includes the information such as the name of AP, timestamp, the MAC and IP addresses of

users attached to it, signal strength, and the number of packets transmitted. By analyzing such data, we

can derive the average arrival rate, average channel holding time, and the user distribution by relative

frequency.


3.4.2 Data Preprocessing

3.4.2.1 Data Extraction

Since the behavior of users may change greatly between daytime and nighttime, or workdays and holidays,

we focus on data accumulated from 9 am to 5 pm on Monday to Friday. We also discard the data

accumulated during the periods of holiday breaks, including Thanksgiving (Nov. 26, 2003 to Nov. 30,

2003) and Christmas and New Year (Dec. 17, 2003 to Jan. 4, 2004). In addition, for some APs, we

observed periods when they are temporally power off. If the total service time of an AP on a certain

day is less than 1/3 of its average value, we discard the data for this day.

3.4.2.2 Trace Gap Padding

The session duration is defined as the period of time during which a user is continuously connected to

the network. The user may move from one AP to another during a session. Occasionally, a user may

disappear from the SNMP report and soon reappear. This may be caused by the user departing and then

returning to the network, or due to the missing of an SNMP report. Following the solution proposed

in [19], we set a departure length threshold Td = 10 minutes. Only if a user disappears and reappears

within Td, it is regarded as staying in the network and the missing SNMP logs are padded.

3.4.2.3 Multiple Association and Ping-Pong Effect

We also observe that some users are simultaneously associated with multiple APs within a small time

interval. Some even ping-pong among multiple APs. We use two methods to offset these effects. First,

when multiple associations occur, we check the number of packets exchanged with the user. We deem

the user is associated with the AP which has exchanged the largest number of packets with the user

during its multiple association period. In addition, if a user leaves one AP and then returns within 5

minutes, it is regarded as having stayed in the AP.

3.4.2.4 Open Users

A fraction of the users may stay in the system during almost all working hours. These users are regarded

as closed users. Since our analytical model assumes an open network, the closed users are excluded in

our experiment. If a user stays for greater than or equal to 7.5 hours during working hours on a valid

day, it is regarded as a closed user. In our experiment, we observe that 9.91% of all users are closed

users. An analytical model for accommodating closed users is provided in [19], which can also be applied

to our work.

3.4.3 Trace Analysis

3.4.3.1 Poisson Arrivals

Analysis of the Dartmouth trace in [19] has shown that the overall new session arrivals into the network

are well modeled by a Poisson process. In this section, we further test the arrival process of new

sessions at each AP against the Poisson assumption. This is divided into two steps. In the first step,

we run an independence test, which indicates whether the numbers of arrivals in different time intervals

are approximately independent. Since it is not practical to account for all time intervals, we test the


Table 3.1: Number of stages.Stages 1 2 3 4 ≥ 5

Observations 80448 15767 7410 3553 6107

independence of arrivals in two consecutive hours at each AP. If the AP passes the test, we regard the

arrivals at this AP to be sufficiently independent. Let H2 denote the entropy in the number of new

arrivals in two consecutive hours and H1 denote the entropy in the number of arrivals in one hour. Let

η = 2H1−H2

H2be the normalized entropy gap. If η < 0.15, we regard the AP as passing the independence

test. We observe that 144 of the 152 APs pass the independence test.

In the second step, we run a Poisson distribution test, which indicates whether the number of arrivals

is approximately Poisson distributed in a fixed time interval. For each AP that passes the independence

test, we count the number of new arrivals in each hour and calculate its real distribution. Furthermore,

by using the actual average arrival rate per hour, we can determine the corresponding theoretical Poisson

distribution. Then, we compute the Kullback-Leibler (KL) divergence H0 between the real distribution

and the theoretical distribution3. Let θ = H0

H1be the normalized KL value. If θ < 0.15, we regard

the AP as passing the Poisson distribution test. We observe that 124 of the 144 APs pass the Poisson

distribution test.4

Those 124 APs are referred to as valid APs, as the new arrivals at these APs can be well approximated

as Poisson. The other 28 APs are referred to as invalid APs. In our experiments, we study the effects

of both including and excluding the non-Poisson new sessions. We emphasize that the Poisson test is

for new arrivals only. Even for those APs that pass the Poisson test, the overall session arrival process

includes both new arrivals and handoff arrivals and hence is non-Poisson.

From the SNMP logs, we observe that the invalid APs tend to have occasional bursty arrivals.

Since they are within the academic area, we conjecture that they correspond to large classrooms, which

experience periodic rushes at the beginning of lecture hours. Even though such APs do not match our

analytical model, their user distribution is likely easy to predict in practice.

3.4.3.2 Number of Stages and Channel Holding Times

We have collected the distributions of number of stages in each route, which is shown in Table 3.1. It

can be seen that there is a large percentage of sessions staying for just one stage. To rigorously test the

analytical stationary user distribution, we will later present different cases where one-stage sessions are

either included or excluded.

Note that if the channel holding times are independently exponentially distributed, our conclusions

on the stationary user distribution trivially holds. Therefore, more challenging channel holding times

(i.e., arbitrarily distributed and correlated) are necessary to test our analytical results. Fig. 3.6 shows

the real distributions of channel holding times in different stages. This figure illustrates that none of

3Kullback-Leibler (KL) divergence is a standard approach to measure the difference between two probability distribu-tions X and Y . It can be regarded as a measure of the information lost (in bits) when Y is used to represent X. When KLdivergence is 0, Y is exactly the same with X. If the KL divergence is small compared with the entropy of the distributionX, the distribution Y is a close approximation of that of X.

4In the experiment, we do not employ the statistical hypothesis test. This is because the session arrival data extractedfrom the Dartmouth trace are noisy (e.g., there are fluctuations and abnormal points), so that most APs cannot pass thetest if a typical significance level 0.05 is used. Therefore, we resort to the KL divergence to show if the number of arrivals isapproximately Poisson distributed. Note that in Sections 3.4.4 and 3.4.5, we only need to show that our analytical modelis accurate when new session arrivals are approximately Poisson.


0 50 100 1500

0.05

0.1

0.15

0.2

0.25

0.3

0.35

minutes

Stage 1 usersStage 2 usersStage 3 usersStage 4 users

Figure 3.6: The pdf of channel holding time in different stages.

them are exponentially distributed. Furthermore, we check the dependency of channel holding times in

different stages. The entropies of the distributions of channel holding times at stages 1, 2, 3 and 4 are

4.0657, 3.4172, 3.3942 and 2.9792, respectively, in bits. The entropy of their joint distribution is 10.2998

bits. Hence, the entropy gap is 4.0657 + 3.4172 + 3.3942 + 2.9792− 10.2998 = 3.5565 bits, much larger

than 0. This shows that the channel holding times at different stages are dependent.

3.4.3.3 AP Locations and Distance Constraint

APs that are far away are likely to have little effect on each other, regardless of the mobility and session

patterns. Therefore, to rigorously test the joint distribution of multiple APs, we are more interested

in selecting adjacent APs with spatial correlation. We set a distance constraint, under which APs are

located pairwisely less than 500 meters from each other. In the experiments, when we study the joint

distribution over multiple APs, this distance constraint is enforced by default, unless otherwise stated.

However, we will also present comparison results for cases with and without it.

3.4.4 Marginal User Distribution at a Single AP

We first show the marginal user distribution at individual APs. For this test, we applied all data after

the pre-processing described in Section 3.4.2, without further exclusions. We show a sampling of the

152 APs. In order to avoid selection bias, we choose APs according to their numeric identity. For each

building (with at least one AP), we select the AP with the smallest identity number (i.e., AP1 if it

exists; otherwise, we select AP2 if it exists; and so forth). There are 32 buildings with at least one AP,

and thus 32 APs are selected accordingly.

Fig. 3.7 shows a comparison between the real distributions and the analytical distributions of these

APs. Each subplot is labeled with Y or N, where Y indicates that the AP passes the two-step Poisson

test and N indicates the opposite. The figure illustrates that the real distributions and the analytical

distributions agree well with each other for those APs that pass the Poisson test.

3.4.5 KL Divergence and Entropy Gap for Multiple APs

In this section, we use KL divergenceHkl to compare the real and analytical joint distributions of multiple

APs. We also test the independence of the numbers of users in different cells by computing the entropy


0 5

0

0.5

1Y

0 20 40

0

0.5N

0 5

0

0.5

1Y

0 10 20

0

0.2

0.4Y

0 20 40

0

0.1

0.2N

0 20 40

0

0.1

0.2N

0 5 10

0

0.2

0.4Y

0 5 10

0

0.5Y

0 10 20

0

0.2

0.4Y

0 2 4

0

0.5

1Y

0 10 20

0

0.5

1N

0 2 4

0

0.5

1Y

0 5 10

0

0.2

0.4Y

0 5 10

0

0.5

1Y

0 5 10

0

0.5Y

0 1 2

0

0.5

1Y

0 20 40

0

0.1

0.2N

0 20 40

0

0.2

0.4Y

0 10 20

0

0.2

0.4Y

0 5 10

0

0.5Y

0 2 4

0

0.5

1Y

0 10 20

0

0.2

0.4N

0 10 20

0

0.2

0.4Y

0 50

0

0.1

0.2N

0 20 40

0

0.1

0.2Y

0 10 20

0

0.2

0.4Y

0 10 20

0

0.2

0.4Y

0 5 10

0

0.2

0.4Y

0 5 10

0

0.5

1Y

0 5 10

0

0.5Y

0 5 10

0

0.5Y

0 1 2

0

0.5

1Y

Figure 3.7: Comparison of distributions for single APs. Real distributions are in solid lines; analyticaldistributions are in dashed lines.

gap Hgap, between the sum of the entropies of real marginal distributions and the entropy of the real

joint distribution. The entropy of the real joint distribution Hreal is also presented for reference. Note

that if Hkl is much smaller than Hreal, the analytical distribution is a close approximation of the real

distribution; if Hgap is much smaller than Hreal, the numbers of users of single APs are approximately

independent.

Given n, the number of APs we aim to study, we randomly choose n different APs. Then we compute

Hkl, Hgap, and Hreal with respect to these APs. By running this procedure 100 times, we obtain the

sample mean and sample standard deviation of Hkl, Hgap, and Hreal. In subsequent studies, we plot the

sample mean versus n, along with bars showing one sample standard deviation, in Figs. 3.8-3.11. Note

that the plot points are slightly shifted to avoid overlaps. Because the sample space of user distribution

increases exponentially with the number of cells, and the real user distribution is counted through its

relative frequency, we limit n ≤ 5 in the experiment in order to ensure enough data are counted for each

sample point.


1 2 3 4 50

2

4

6

8

10

12

Number of APsV

alu

e(b

its)

Hkl, excluding invalid APs

Hreal, excluding invalid APs

Hkl, excluding invalid sessions

Hreal, excluding invalid sessions

Hkl, without exclusion

Hreal, without exclusion

Figure 3.8: Hkl and Hreal under the influence of non-Poisson arrivals.

1 2 3 4 50

2

4

6

8

10

12

Number of APs

Valu

e(b

its)

Hgap, excluding invalid APs

Hreal, excluding invalid APs

Hgap, excluding invalid sessions

Hreal, excluding invalid sessions

Hgap, without exclusion

Hreal, without exclusion

Figure 3.9: Hgap and Hreal under the influence of non-Poisson arrivals.

1 2 3 4 50

2

4

6

8

10

12

Number of APs

Valu

e(b

its)

Hkl, with distance constraint

Hgap, with distance constraint

Hreal, with distance constraint

Hkl, without distance constraint

Hgap, without distance constraint

Hreal, without distance constraint

Figure 3.10: Hkl, Hgap and Hreal under the influence of distance restriction.

1 2 3 4 50

2

4

6

8

10

12

Number of APs

Valu

e(b

its)

Hkl, with one-stage sessions

Hgap, with one-stage sessions

Hreal, with one-stage sessions

Hkl, without one-stage sessions

Hgap, without one-stage sessions

Hreal, without one stage sessions

Figure 3.11: Hkl, Hgap and Hreal under the influence of one-stage sessions.


3.4.5.1 Influence of Non-Poisson Arrivals

Clearly, excluding non-Poisson arrivals could improve the accuracy of the analytical model. We compare

Hkl, Hgap, and Hreal under the conditions of either including or excluding non-Poisson arrivals.

A direct method to exclude non-Poisson session arrivals is to remove from the data set all sessions

that are initiated at invalid APs. However, this will reduce the number of handoff session arrivals even

in valid APs, hence biasing the analysis. An alternate approach is to simply remove the invalid APs

from the data set, while allowing those non-Poisson sessions to be counted in the valid APs that they

pass through. In this way, accurate average arrival rates at the valid APs are maintained.

Thus, we study the following three cases: 1) Excluding sessions initiating at invalid APs (i.e., invalid

sessions); 2) Excluding invalid APs; and 3) Without exclusion. Fig. 3.8 illustrates Hkl compared with

Hreal for the three cases, and Fig. 3.9 illustrates Hgap compared with Hreal for the three cases. We

observe that both Hkl and Hgap are much smaller than Hreal, when we either exclude invalid sessions

or exclude invalid APs, illustrating that the real distributions are close to the analytical distributions,

and the numbers of users of single APs are approximately independent. When we do not exclude

invalid sessions or invalid APs, Hkl and Hgap become larger, showing that the analytical distribution

is influenced by the non-Poisson arrivals. However, Hkl and Hgap remain much smaller than Hreal,

illustrating that the analytical distribution is still valid to approximate the real distribution, even the

arrivals are not strictly Poisson.

In addition, excluding invalid sessions only brings small decrements in Hkl and Hgap compared with

excluding invalid APs. Note that when we exclude invalid sessions, both the one-stage and multiple-stage

non-Poisson arrival sessions are excluded; when we exclude invalid APs, only the one-stage non-Poisson

arrival sessions are excluded. This illustrates that multiple-stage non-Poisson arrival sessions have only

weak influence on the modeling accuracy.

3.4.5.2 Influence of Distance Constraint

Fig. 3.10 shows Hkl, Hgap, and Hreal with and without the distance constraint. For both cases, we

exclude the invalid APs. We observe that Hkl, Hgap, and Hreal are nearly unchanged with or without

the distance constraint, confirming our expectation that the distance constraint does not influence the

accuracy of the analytical model, since the analytical model predicts that the numbers of users of adjacent

APs are independent.

3.4.5.3 Influence of One-Stage Sessions

Fig. 3.11 shows Hkl, Hgap, and Hreal with and without the one-stage sessions. For both cases, we exclude

the invalid APs. We observe that when we exclude the one-stage sessions, Hkl and Hgap becomes smaller,

suggesting that our model is even more accurate in this case. This is an apparently counter-intuitive

result, since the analytical distribution trivially holds for one-stage sessions. An explanation for this is

the following. Since one-stage sessions are more likely to be new sessions corresponding to attending

lectures in a classroom, they are more likely to be non-Poisson. Since not all non-Poisson arrivals can

be excluded by removing the invalid APs, when we further exclude one-stage sessions, we obtain more

accurate analytical results.

Note that one-stage sessions can be analyzed as a single-queue model [18]. Thus, in practice, one may

separately analyze one-stage and multiple-stage sessions and combine the resultant user distributions.


3.5 Summary

In this chapter, we have studied the user distribution in multicell network by establishing a precise ana-

lytical model, considering arbitrary user movement and arbitrarily and dependently distributed channel

holding times. We have derived the stationary distribution of the number of users in each cell, which is

only related to the average arrival rate and the average channel holding time of each cell, and hence is

insensitivity with respect to the general movement and session patterns. We have used the Dartmouth

trace to validate our analysis, which shows that the analytical model is accurate when new session ar-

rivals are Poisson and remains useful when non-Poisson session arrivals are also included in the data

set.

Chapter 4

Uplink Interference Analysis: Open

Access versus Closed Access

As discussed in Chapter 1, in order to achieve higher capacity, better service quality, lower power usage,

and ubiquitous coverage in mobile networks, one effective approach is to install a second tier of smaller

cells, referred to as femtocells, overlapping the original macrocell network [97]. In the presence of

femtocells, whenever some user equipment (UE) is near a femtocell BS, two different access mechanisms

may be applied: closed access and open access. Under closed access, a femtocell BS only provides service

to its local users, without further admitting nearby macrocell users. In contrast, under open access, all

nearby macrocell users are allowed to access the femtocell BS.

In this chapter, we mainly discuss how the two access modes may affect both macrocell users and local

femtocell users, in terms of the uplink interference and outage probabilities. We present a stochastic

geometric analysis framework to derive numerical expressions for the uplink interference and outage

probabilities of open access and closed access by modeling macrocell BSs as a regular grid, macrocell

UEs as a Poisson point process (PPP), and femtocell UEs as a two-level clustered Poisson point process,

which captures the spatial patterns of different network components. Due to the complexity of uplink

interference, the expressions of outage probabilities for both the open and closed access cases are in

non-closed forms, requiring multiple levels of integration. Then, we further derive sufficient conditions

for open and closed access modes to outperform each other in terms of the outage probability at either

the macrocell or femtocell level. This leads to closed-form expressions to upper and lower bound the

difference in the targeted received power between the two access modes. Simulations are conducted to

validate the accuracy of the analytical model and the correctness of the bounds.

4.1 System Model

4.1.1 Two-tier Network

We consider a two-tier network with macrocells and femtocells as shown in Fig. 4.1. In this chapter,

we assume that the macrocells form an infinite hexagonal grid in the two-dimensional Euclidean space

R2. Macrocell BSs are located at the centers of the hexagons B = ( 32aRc,√32 aRc +

√3bRc)|a, b ∈ Z,

where Rc is the radius of the hexagon. Macrocell UEs are randomly distributed in the system, which are

31

Chapter 4. Uplink Interference Analysis: Open Access versus Closed Access 32

Macrocell UEs Femtocell UEs

Open access UEs

Macrocells

Femtocells

Rc

R

Figure 4.1: Two-tier network with macrocells and femtocells.

modeled as a homogeneous Poisson point process (PPP) Φ with intensity λ. We focus on interference

analysis in a single shared radio frequency sub-band (which is referred to as the reference sub-band)

[52,53,58,98,99]. The UEs (macrocell UEs and femtocell UEs) considered in this chapter are those using

the reference sub-band.

Femtocell BSs are randomly spatially deployed. We assume femtocell BSs form a homogeneous PPP

Θ with intensity µ. Each femtocell BS is connected to the core network by high-capacity wired links

that have no influence on our wireless performance analysis.

Each femtocell BS communicates with local femtocell UEs surrounding it, constituting a femtocell.

We assume R as the maximum communication radius of each femtocell BS. Given the location of a

femtocell BS at x0, we assume that its femtocell UEs, denoted by Ψ(x0), are distributed as a non-

homogenous PPP in the disk centered at x0 with radius R. Its intensity at x is described by ν(x− x0),

a non-negative function of the vector x− x0. Note that the UE intensity ν(x− x0) = 0 if |x− x0| > R.

The femtocell UEs in one femtocell are independent of femtocell UEs in other femtocells, as well as the

macrocell UEs. We assume the scale of femtocells is much smaller than the scale of macrocells [97],

R ≪ Rc.

To better understand the spatial distribution of femtocell BSs and femtocell UEs, the femtocell

BSs Θ can be regarded as a parent point process in R2, while femtocell UEs Ψ is a daughter process

associated with a point in the parent point process, forming a two-level random pattern (see Section

5.3 of [39]). Note that the aggregating of femtocell UEs around a femtocell BS implicitly defines the

location correlation among femtocell UEs.

Let 0 denote the origin of the plane. Let H(x) denote the hexagon region centered at x with radius

Rc; let B(x, R) denote the disk region centered at x with radius R; let BS(x) denote the hexagon center

nearest to x (i.e., BS(x) = x0 ⇔ x ∈ H(x0)).

4.1.2 Open Access versus Closed Access

If a macrocell UE is covered by a femtocell BS (i.e., within a distance of R from a femtocell BS), under

closed access, the UE still connects to the macrocell BS. Under open access, the UE is handed off to

the femtocell BS and disconnects from the original macrocell BS; the UE is then referred to as an open


access UE.

Given a femtocell BS located at x0, let Ω(x0) denote the point process corresponding to the open

access UEs connecting to it. We assume that the probability of two femtocells overlapping can be

ignored [53].1 Thus, Ω(x0) corresponds to points of Φ inside the range of the femtocell BS at x0, which

is a PPP with intensity λ inside B(x0, R).

Let ν =∫B(0,R)

ν(x)dx be the average number of local femtocell UEs inside a femtocell (using the

reference radio frequency sub-band). In the open access case, λ = πR2λ is the average number of open

access UEs inside a femtocell (using the reference radio frequency sub-band).

4.1.3 Path Loss and Power Control

Let Pt(x) denote the transmission power at x and Pr(y) denote the received power at y. We assume that

Pr(y) =Pt(x)hx,y

|x−y|γ , where |x − y|γ is the propagation loss function with predetermined constant γ > 2,

and hx,y is the fast fading term. Note that we focus on the scenario where different tiers in the system

use the same radio spectrum under a similar radio environment, and thus the path loss exponent values

of different tiers are identical [58,59,100]. Also, following a common assumption of stochastic geometric

modeling of wireless networks, we assume that hx,y is independently exponentially distributed with unit

mean (i.e., Rayleigh fading) [12, 13, 35, 50, 58, 59]. Let H(·) be the cumulative distribution function of

hx,y.

We also assume that uplink power control adjusts for propagation losses [53, 98, 101, 102]. The

targeted received power level of macrocell UEs, femtocell UEs, and open access UEs are P , Q, and

P ′, respectively2. Given the targeted received power PT (where PT = P , Q, or P ′), receiver at y,

and transmitter at x, the transmission power is PT |x − y|γ . Then, the resultant interference at y′ isPT |x−y|γhx,y′

|x−y′|γ .

We define ρ , P ′/P , which is the targeted received power enhancement if a macrocell UE becomes

an open access UE. In this chapter, we study the performance variation when open access is applied

to replace closed access. Therefore, as a parameter corresponding to open access UEs, ρ is regarded

as an important designed parameter. The other parameters, such as P, Q, T , and γ are considered as

predetermined system-level constants.

4.1.4 Outage Performance

In this chapter, the performance of macrocell UEs and femtocell UEs (under open access or closed

access) is examined through the outage probability, which is defined as the probability that the signal

to interference ratio (SIR) is smaller than a given threshold value T . Because we focus on interference

analysis, the noise is assumed to be negligible.

1Given a femtocell, the probability that it does not overlap with another femtocell (i.e., no other femtocell BSs arelocated 2R from it) is Pnonover = exp(−4πR2µ). We assume that R and µ are small enough such that Pnonover can beapproximated by 1.

2It is recommended (Section 7.3 of [103]) that the received power level at macrocell BSs is lower than that at femtocellBSs. Here, we assume a single fixed level of targeted received power at the macrocell or femtocell level for mathematicaltractability. We show that our model is still valid when the targeted received power is randomly distributed throughsimulations in Section 4.4.


4.2 Open Access vs. Closed Access at the Macrocell Level

In this section, we analyze the uplink interference and outage performance of macrocell UEs. Consider a

reference macrocell UE, termed the typical UE, communicating with its macrocell BS, termed the typical

BS. We aim to investigate the performance of the typical UE.

Due to the stationarity of point processes corresponding to macrocell UEs, femtocell BSs, and fem-

tocell UEs, throughout this section we will re-define the coordinates so that the typical BS is located at

0. Correspondingly, the typical UE is located at some xU that is uniformly distributed in H(0), since

macrocell BSs form a deterministic hexagonal grid (see page 60 of [37]).

Let Φ′ be the point process of all other macrocell UEs conditioned on the typical UE, which is a

reduced Palm point process [37] with respect to (w.r.t.) Φ. Because the reduced Palm point process of

a PPP has the same distribution as its original PPP, Φ′ is still a PPP with intensity λ [37]. Therefore,

for presentation convenience, we still use Φ to denote this reduced Palm point process.

4.2.1 Open Access Case

4.2.1.1 Interference Components

The overall interference in the uplink has three parts: from macrocell UEs not inside any femtocell

(denoted by I1), from open access UEs (denoted by I2), and from femtocell UEs (denoted by I3).

I1 can be computed as the sum of interference from each macrocell UE:

I1 =∑x∈Φ0

P |x− BS(x)|γhx,0

|x|γ, (4.1)

where Φ0 denotes the points of Φ not inside any femtocell.

I2 can be computed as the sum of interference from all open access UEs of all femtocells:

I2 =∑x0∈Θ

∑x∈Ω(x0)

P ′|x− x0|γhx,0

|x|γ. (4.2)

I3 can be computed as the sum of interference from all femtocell UEs of all femtocells:

I3 =∑x0∈Θ

∑x∈Ψ(x0)

Q|x− x0|γhx,0

|x|γ. (4.3)

The overall interference of open access is I = I1 + I2 + I3.

4.2.1.2 Laplace Transform of I

In this subsection, we study the Laplace transform of I, denoted by LI(s), which leads to the following

theorem:

Theorem 3. Under the open access mode, the Laplace transform of uplink interference at a typical

macrocell BS is given by

LI(s) =E( ∏

x∈Φ

u(x, s)

)· E

[ ∏x0∈Θ

E(∏

x∈Ω(x0)v(x,x0, s)

)E(∏

x∈Ω(x0)u(x, s)

) E( ∏

x∈Ψ(x0)

w(x,x0, s)))]

, (4.4)


where u(x, s) , exp(− sP |x−BS(x)|γhx,0

|x|γ

), v(x,x0, s) , exp

(− sρP |x−x0|γhx,0

|x|γ

), and w(x,x0, s) ,

exp(− sQ|x−x0|γhx,0

|x|γ

).

See Appendix A.1 for the proof.

4.2.1.3 Numeric Computation of LI(s)

In this subsection, we present a numeric approach to compute LI(s) derived in (4.4), which will facilitate

later comparison between open access and closed access. Let L0(s) = E(∏

x∈Φ u(x, s)), which is a

generating functional corresponding to Φ [37, 39]. It can be re-written in a standard integral form as

follows:

L0(s) = exp

(− λ

∫R2

(1−

∫R+

e−sP |x−BS(x)|γh

|x|γ H(dh)

)dx

)

=exp

(− λ

∫R2

sP |x− BS(x)|γ

sP |x− BS(x)|γ + |x|γdx

). (4.5)

Given the location of a femtocell BS at x0, let W(x0, s) = E

( ∏x∈Ψ(x0)

w(x,x0, s)

), which is a

generating functional corresponding to Ψ(x0). It can be expressed in a standard form through the

Laplace functional of PPP Ψ(x0):

W(x0, s) = exp

(−

∫B(0,R)

sQ|x|γ

sQ|x|γ + |x+ x0|γν(x)dx

). (4.6)

Similarly, let V(x0, s) = E(∏


), and U(x0, s) = E

(∏x∈Ω(x0)

u(x, s)), we have

V(x0, s) = exp

(− λ

∫B(0,R)

sρP |x|γ

sρP |x|γ + |x+ x0|γdx

), (4.7)

U(x0, s) = exp

(− λ

∫B(x0,R)

sP |x− BS(x)|γ

sP |x− BS(x)|γ + |x|γdx

). (4.8)

Let J (x0, s) = V(x0,s)U(x0,s)

W(x0, s), which is numerically computable through (4.6)-(4.8). Finally, we

note that

E

[ ∏x0∈Θ

E(∏


)E(∏

x∈Ω(x0)u(x, s)

) E( ∏

x∈Ψ(x0)

w(x,x0, s)))]

=E

[ ∏x0∈Θ

(V(x0, s)

U(x0, s)W(x0, s)

)]= E

( ∏x0∈Θ

J (x0, s)

)

=exp

(−µ

∫R2

(1− J (x0, s)) dx0

), (4.9)

where (4.9) is derived from the generating functional with respect to PPP Θ. Substituting (4.5) and


(4.9) into (4.4), we can numerically compute LI(s):

LI(s) = L0(s) · exp(−µ

∫R2

(1− J (x0, s)) dx0

). (4.10)

An intuitive explanation to the above is as follows. First, in terms of the Laplace transform, additive

interference is in the product form, and interference reduction is in the division form. Suppose that there

are no femtocells at the beginning, and L0(s) corresponds to the interference from macrocell UEs. Then,

we add femtocells to the system. Given a femtocell BS at x0, W(x0, s) corresponds to the interference

from local femtocell UEs inside the femtocell, V(x0, s) corresponds to interference from open access UEs

inside the femtocell, and U(x0, s) corresponds to interference reduction due to open access UEs as they

disconnect from their original macrocell BS. Thus, J (x0, s) = V(x0,s)U(x0,s)

W(x0, s) represents the overall

interference variation when a femtocell centered at x0 is added. Finally, exp(−µ∫R2(1− J (x0, s))dx0

)is the overall interference variation after adding all femtocells. As a consequence, the overall interference

can be computed in formula (4.10).

4.2.1.4 Outage Probability

Given the SIR threshold T , the outage probability of the typical UE can be computed as the probability

that the signal strength PhxU ,0 over the interference I is less than T :

P oout = P(PhxU ,0 < TI) = 1− LI(s)|s= T

P. (4.11)

The last equality above is due to hxU ,0 being exponentially distributed with unit mean. As a result,

P oout can be derived directly from LI(s) (see Section 16.2.2.1 of [37]).

4.2.2 Closed Access Case

Different from the open access case, the overall interference has only two parts: from macrocell UEs

(denoted by I1) and from femtocell UEs (denoted by I3).

I1 can be computed as the sum of interference from each macrocell UE:

I1 =∑x∈Φ

P |x− BS(x)|γhx,0

|x|γ. (4.12)

I3 is exactly the same as I3 in (4.3).

Then, the total interference can be computed as I = I1 + I3. Similar to Section 4.2.1.3, the Laplace

transform of I is

LI(s) =E

[ ∏x∈Φ

u(x, s)∏

x0∈Θ

∏x∈Ψ(x0)

w(x,x0, s)

]

=L0(s) · E

[ ∏x0∈Θ

(W(x0, s)

)]

=L0(s) · exp(−µ

∫R2

(1−W(x0, s))dx0

), (4.13)

where L0(s) is the same as (4.5), and W(x0, s) is the same as (4.6).


An intuitive explanation to the above is as follows. First, L0(s) corresponds to the interference of all

macrocell UEs. Given a femtocell BS at x0, W(x0, s) corresponds to interference from local femtocell UEs

inside the femtocell. Then, exp(−µ∫R2(1−W(x0, s))dx0

)is the overall interference from all femtocells.

As a consequence, the overall interference can be computed as formula (4.13).

Finally, the outage probability of the typical UE can be computed as

P cout = P(PhxU ,0 < T I) = 1− LI(s)|s= T

P. (4.14)

4.2.3 Parameter Normalization

From the above performance analysis of both open access and closed access, we see that one can normalize

the radius of macrocells Rc to 1, so that R is equivalent to the ratio of the radius of femtocells to that

of macrocells (R ≪ 1). Also, we can normalize the target received power of macrocell UEs P to 1, so

that Q is equivalent to the ratio of the target received power of femtocell UEs to that of macrocell UEs,

and P ′ = ρ. Therefore, in the rest of this section, without loss of generality, we set Rc = 1 and P = 1.

4.2.4 Open Access vs. Closed Access

We compare the outage performance of open access and closed access at the macrocell level. Due to the

integral form of the Laplace transform, the expressions of outage probabilities for both the open and

closed access cases are in non-closed forms, requiring multiple levels of integration. As a consequence,

we are motivated to derive closed-form bounds to compare open access and closed access.

Let Vmax , 4π2R4(Tρ)2γ

(18 + 1

4(γ+2) +1

(γ+2)(γ−2)

), Vmin , 2π2R4(Tρ)

2γ

(18 + 1

4(γ+2) +1

(γ+2)(γ−2)

),

and Cu ,∫R2

(T |x−BS(x)|γ

T |x−BS(x)|γ+|x|γ

)dx be a system-level constant determined by T and γ. We have the

following theorem:

Theorem 4. A sufficient condition for P oout < P c

out is

−Vmax + πR2Cue−ν > 0, (4.15)

and a sufficient condition for P oout > P c

out is

−πR2Cueλ +Vmine

−λ−ν > 0. (4.16)


Through Theorem 4, closed-form expressions can be used to compare the outage probabilities between

open access and closed access without the computational complexity introduced by numeric integrations

in (4.10) and (4.13).

In the following, we focus on the performance variation if open access is applied to replace closed

access. The parameter corresponding to open access UEs, ρ, is regarded as a design parameter. If

we fix all the other network parameters (including targeted received power level of macrocell UEs,

femtocell UEs), increasing ρ implies better performance for open access UEs, but it will also increase the

interference from open access UEs to macrocell BSs. As a consequence, we aim to derive the threshold

ρ∗ such that P oout = P c

out. At the macrocell level, macrocell UEs experience less outage iff ρ < ρ∗.

Thus, ρ∗ is referred to as the maximum power enhancement tolerated at the macrocell level. Thus, in the


deployment of open access femtocells, the network operator is motivated to limit ρ below ρ∗ to guarantee

that the performance of macrocell UEs under open access is no worse than that under closed access. One

way to derive ρ∗ is through numerical computation of (4.10) and (4.13) and numerical search, which

introduces high computational complexity due to the multiple levels of integration. A more efficient

alternative is to find the bounds of ρ∗ through Theorem 4. Simple algebra manipulation leads to

ρ∗min =1

T

Cue−ν

4πR2(

18 + 1

4(γ+2) +1

(γ+2)(γ−2)

)

γ2

, (4.17)

ρ∗max =1

T

Cueν+2λ

2πR2(

18 + 1

4(γ+2) +1

(γ+2)(γ−2)

)

γ2

, (4.18)

where ρ∗min and ρ∗max are the lower bound and upper bound of ρ∗, respectively. If the network operator

limits ρ < ρ∗min, the performance of macrocell UEs under open access can be guaranteed no worse than

their performance under closed access.

Furthermore, through (4.17) and (4.18), we observe that ρ∗min = Θ( 1Rγ ) and ρ∗max = Θ( 1

Rγ ), leading

to the following corollary:

Corollary 1.

ρ∗ = Θ(1

Rγ). (4.19)

Note that, in (4.19), Rc is normalized to 1 and R represents the ratio of the radius of femtocells to

that of macrocells. If Rc is not normalized, (4.19) should be re-written as ρ∗ = Θ((

Rc

R

)γ).

Intuitively, as a rough estimation, open access UEs have their distance to the BS reduced approxi-

mately by a factor of R, leading to the capability to increase their received power by the corresponding

gain in the propagation loss function, as their average interference level is maintained. However, Corol-

lary 1 cannot be trivially obtained from the above intuition. This is because the outage probability does

not only depend on the average interference, but also depends on the distribution of the interference

(i.e., the Laplace transform of the interference). Comparing (4.10) with (4.13), we note that if we switch

from closed access to open access, the distribution of the interference will change drastically. Corollary

1 can be derived only after rigorously comparing and bounding the Laplace transforms of interference

under open access and closed access.

From (4.17) and (4.18), the gap between the upper and lower bounds can be expressed asρ∗max

ρ∗min

=

2γ2 eγ(ν+λ). We emphasize that in this chapter, interference analysis is conducted for a single shared

radio frequency sub-band (i.e., the reference sub-band). The average number of UEs in a macrocell or

femtocell using the reference sub-band is usually no greater than one. Note that ν + λ is the average

number of UEs in a femtocell using the reference sub-band and is expected to typically be a small value.

For example, as indicated in the 3GPP guidelines in Annex A of [104], the number of femtocell UEs per

femtocell is much smaller than the number of macrocell UEs per macrocell. Therefore,ρ∗max

ρ∗min

is expected

to be small in practical systems.


4.3 Open Access vs. Closed Access at the Femtocell Level

In this section, we analyze the uplink interference and outage performance of femtocell UEs. Given a

reference femtocell UE, termed the typical femtocell UE, communicating with its femtocell BS, termed

the typical femtocell BS, we aim to study the interference at the typical femtocell BS. We also define

the femtocell corresponding to the typical femtocell BS as the typical femtocell, and the macrocell BS

nearest to the typical femtocell BS as the typical macrocell BS.

Similar to Section 4.2, we re-define the coordinate of the typical macrocell BS as 0. Correspondingly,

the typical femtocell BS is located at some xB that is uniformly distributed in H(0) [37]. Given the

typical femtocell centered at xB, let Θ′ denote the point process of other femtocell BSs conditioned

on the typical femtocell BS, i.e., the reduced Palm point process w.r.t. Θ. Then, Θ′ is still a PPP

with intensity µ [37]. For presentation convenience, we still use Θ to denote this reduced Palm point

process. Let Ψ(xB) denote the other femtocell UEs inside the typical femtocell conditioned on the typical

femtocell UE. Similarly, Ψ(xB) has the same distribution as Ψ(xB). Let Ω(xB) denote open access UEs

connecting to the typical femtocell BS.

4.3.1 Open Access Case

4.3.1.1 Interference Components

The overall interference in the uplink of the typical femtocell UE has five parts: from macrocell UEs

not inside any femtocell (I ′1(xB)), from open access UEs outside the typical femtocell (I ′2(xB)), from

femtocell UEs outside the typical femtocell (I ′3(xB)), from local femtocell UEs inside the typical femtocell

(I ′4(xB)), and from open access UEs inside the typical femtocell (I ′5(xB)). We have

I ′1(xB) =∑x∈Φ0

P |x− BS(x)|γhx,xB

|x− xB |γ, (4.20)

I ′2(xB) =∑x0∈Θ

∑x∈Ω(x0)

ρP |x− x0|γhx,xB

|x− xB |γ, (4.21)

I ′3(xB) =∑x0∈Θ

∑x∈Ψ(x0)

Q|x− x0|γhx,xB

|x− xB |γ, (4.22)

I ′4(xB) =∑

x∈Ψ(xB)

Qhx,xB, (4.23)

I ′5(xB) =∑

x∈Ω(xB)

ρPhx,xB . (4.24)

The overall interference is I ′(xB) =∑5

i=1 I′i(xB).

4.3.1.2 Laplace Transform of I ′(xB)

In this subsection, we study the Laplace transform of I ′(xB), denoted by LI′(xB, s). We have the

following theorem:

Theorem 5. Under the open access mode, the Laplace transform of uplink interference at a typical


femtocell BS located at xB is given by

LI′(xB , s) =E( ∏

x∈Φ

u′(x,xB , s)

)E

[ ∏x0∈Θ

(E( ∏

x∈Ω(x0)

v′(x,x0,xB , s))

E( ∏

x∈Ω(x0)

u′(x,xB , s)) E

( ∏x∈Ψ(x0)

w′(x,x0,xB , s)))]

E( ∏

x∈Ψ(xB)

w′(x,xB ,xB, s)

)E(∏

x∈Ω(xB) v′(x,xB,xB , s)

)E(∏

x∈Ω(xB) u′(x,xB , s)

) . (4.25)

where u′(x,xB , s) , exp(− sP |x−BS(x)|γhx,xB

|x−xB |γ

), v′(x,x0,xB, s) , exp

(− sρP |x−x0|γhx,xB

|x−xB |γ

), and

w′(x,x0,xB, s) , exp(− sQ|x−x0|γhx,xB

|x−xB |γ

).


4.3.1.3 Numeric Computation of LI′(xB , s)

First, similar to the derivations of (4.5)-(4.8) in Section 4.2.1.3, we have

L′0(xB , s) =E

( ∏x∈Φ

u′(x,xB , s)

)= exp

(− λ

∫R2

sP |x− BS(x)|γ

sP |x− BS(x)|γ + |x− xB |γdx

), (4.26)

W ′(x0,xB , s) =E( ∏

x∈Ψ(x0)

w′(x,x0,xB , s))= exp

(−

∫B(x0,R)

sQ|x− x0|γ

sQ|x− x0|γ + |x− xB |γν(x− x0)dx

),

(4.27)

V ′(x0,xB , s) =E( ∏

x∈Ω(x0)

v′(x,x0,xB , s))= exp

(− λ

∫B(x0,R)

sρP |x− x0|γ

sρP |x− x0|γ + |x− xB |γdx

), (4.28)

U ′(x0,xB , s) =E( ∏

x∈Ω(x0)

u′(x,xB , s))= exp

(− λ

∫B(x0,R)

sP |x− BS(x)|γ

sP |x− BS(x)|γ + |x− xB |γdx

), (4.29)

In addition, we can derive

W ′′(xB , s) =E( ∏

x∈Ψ(xB)

w′(x,xB ,xB, s)

)= exp

(− sQν

sQ+ 1

), (4.30)

V ′′(xB , s) =E( ∏

x∈Ω(xB)

v′(x,xB ,xB , s)

)= exp

(− sρPλ

sρP + 1

), (4.31)

U ′′(xB , s) =E( ∏

x∈Ω(xB)

u′(x,xB, s)

)= exp

(− λ

∫B(xB ,R)

sP |x− BS(x)|γ

sP |x− BS(x)|γ + |x− xB|γdx

). (4.32)

Then, following the same steps as (4.9), LI′(xB , s) is derived as

LI′(xB , s) = L′0(xB, s) exp

(− µ

∫R2

(1− V ′(x0,xB , s)W ′(x0,xB , s)

U ′(x0,xB , s)

)dx0

)W ′′(xB , s)V ′′(xB , s)

U ′′(xB , s).

(4.33)


An intuitive explanation to the above is as follows. First, L′0(xB , s) corresponds to the interference of

all macrocell UEs. Second, similar to the discussions in Section 4.2.1.3, V′(x0,xB ,s)W′(x0,xB ,s)U ′(x0,xB ,s) represents

the overall interference variation when a femtocell centered at x0 is added. Third, exp(− µ

∫R2

(1−

V′(x0,xB ,s)W′(x0,xB ,s)U ′(x0,xB ,s)

)dx0

)is the overall interference variation after adding all femtocells other than the

typical femtocell. Fourth, W′′(xB ,s)V′′(xB ,s)U ′′(xB ,s) represents the overall interference variation after adding the

typical femtocell. As a consequence, the overall interference can be computed as formula (4.33).

4.3.1.4 Outage Probability

Similar to (4.11), the outage probability (given xB) is

P oout(xB) = P(QhxU ,xB

< TI ′(xB)) = 1− LI′(xB , s)|s=T ′ , (4.34)

where xU is the coordinate of the typical femtocell UE (irrelevant to the result), T ′ = TQ , and T is

the SIR threshold. Because xB is uniformly distributed in H(0), the average outage probability can be

computed as∫H(0)

P oout(xB)dxB/|H(0)|, where |H(0)| = 3

√3R2

c

2 is the area of a macrocell.

4.3.2 Closed Access Case

The overall interference has three parts: from macrocell UEs (I ′1(xB)), from femtocell UEs outside the

typical femtocell (I ′3(xB)), and from femtocell UEs inside the typical femtocell (I ′4(xB)). I′1(xB) can be

computed as

I ′1(xB) =∑x∈Φ

P |x− BS(x)|γhx,xB

|x− xB |γ, (4.35)

and I ′3(xB) and I ′4(xB) are exactly the same as I ′3(xB) in (4.22) and I ′4(xB) in (4.23), respectively.

The overall interference is I ′(xB) = I ′1(xB)+ I ′3(xB)+ I ′4(xB). Then, the Laplace transform of I ′(xB)

is

LI′(xB , s) = L′0(xB, s) · exp

(−µ

∫R2

(1−W ′(x0,xB)) dx0

)· W ′′(xB , s). (4.36)

The outage probability (given xB) is

P cout(xB) = 1− LI′(xB , s)|s=T ′ . (4.37)

The average outage probability is∫H(0)

P cout(xB)dxB/|H(0)|. Similar to the discussion in Section 4.2.3,

we still can normalize Rc and P . Hence, in the rest of this section, without loss of generality, we set

Rc = 1 and P = 1.

4.3.3 Open Access vs. Closed Access

In this subsection, we compare the outage performance of open access and closed access at the femtocell

level.

LetV′max , 4π2R4(T ′ρ)

2γ

(18 + 1

4(γ+2) +1

(γ+2)(γ−2)

),V′

min , 2π2R4(T ′ρ)2γ

(18 + 1

4(γ+2) +1

(γ+2)(γ−2)

),

C ′u be a system-level constant shown in (A.35), Rmin and Rmax be as shown in (A.36) and (A.37) in the


proof of Theorem 6, which are in closed forms if γ is a rational number3. Then we have the following

theorem:

Theorem 6. Given xB , a sufficient condition for P oout(xB) < P c

out(xB) is

K1 , −µV′max + µπR2C ′

ue−ν − πR2T ′ρ

T ′ρ+ 1+Rmin > 0, (4.38)

and a sufficient condition for P oout(xB) > P c

out(xB) is

K2 , −µπR2C ′ue

λ + µV′mine

−ν−λ +πR2T ′ρ

T ′ρ+ 1−Rmax > 0. (4.39)


Through Theorem 6, the closed-form expressions can be used to compare the outage probabilities

between open access and closed access without the computational complexity introduced by numeric

integrations in (4.34) and (4.37).

Similar to the discussion in Section 4.2.4, let ρ∗∗ denote the threshold value of ρ such that P oout(xB) =

P cout(xB). At the femtocell level, given that a femtocell BS is located at xB (the relative coordinate

w.r.t. the nearest macrocell), its local femtocell UEs experience less outage iff ρ < ρ∗∗. Thus, ρ∗∗ is

referred to as the maximum power enhancement tolerated by the femtocell.

Instead of deriving ρ∗∗ through (4.34) and (4.37), which introduces high computational complexity

due to multiple levels of integration, we can find the lower bound ρ∗∗min and upper bound ρ∗∗max of ρ∗∗

through Theorem 6. Accordingly, ρ∗∗min is the value satisfying K1 = 0 and ρ∗∗max is the value satisfying

K2 = 0. Thus, ρ∗∗min and ρ∗∗max can be found by a numerical search approach w.r.t. the closed-form

expressions.

4.4 Numerical Study

We present simulation and numerical studies on the outage performance in the two-tier network with

femtocells. First, we study the performance of open access and closed access under different user densities,

femtocell densities, and SIR thresholds. Second, we present the numerical results of ρ∗ and ρ∗∗. Unless

otherwise stated, Rc = 500 m, R = 50 m, γ = 3; and fast fading is Rayleigh with unit mean. Each

simulation data point is averaged over 50000 trials.

First, we study the performance under different µ, λ, and T . The network parameters are as follows:

ν(x) = 80 units/km2/sub-band if |x| < R, and ν(x) = 0 otherwise; P = −60 dBm, and Q = P ′ = −54

dBm (ρ = 6 dB). Figs. 4.2 and 4.3 show the uplink outage probabilities of macrocell and femtocell

UEs under different λ; Figs. 4.4 and 4.5 show the same under different µ; and Figs. 4.6 and 4.7 show

the same under different T . The analytical results are derived from the exact expressions in Sections

4.2.1, 4.2.2, 4.3.1, and 4.3.2, without applying any bounds. The error bars show the 95% confidence

intervals for simulation results. For easier inspection, in Figs. 4.2-4.7, the plot points are slightly shifted

horizontally to avoid overlapping error bars. The figures illustrate the accuracy of our analytical results.

In addition, the figures show that the macrocell UE density strongly influences the outage probability

of both macrocell and femtocell UEs, while the femtocell density only has a slight influence. At the

3It is acceptable to assume γ as a rational number in reality, because each real number can be approximated by arational number with arbitrary precision.


3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8

0.3

0.35

0.4

0.45

0.5

0.55

0.6

λ [units/km2/sub-band]

Ou

tag

e p

rob

abil

ity

Open access, simulation

Open access, analytical

Closed access, simulation

Closed access, analytical

Figure 4.2: Macrocell outage probability under different λ, with µ = 4 units/km2/sub-band and T = 0.1.

3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 80.2

0.25

0.3

0.35

0.4

0.45

λ [units/km2/sub-band]

Ou

tag

e p

rob

abil

ity





Figure 4.3: Femtocell outage probability under different λ, with µ = 4 units/km2/sub-band and T = 0.1.

3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 80.35

0.355

0.36

0.365

0.37

0.375

0.38

0.385

0.39

µ [units/km2/sub-band]

Outa

ge

pro

bab

ilit

y





Figure 4.4: Macrocell outage probability under different µ, with λ = 4 units/km2/sub-band and T = 0.1.

3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 80.245

0.25

0.255

0.26

0.265

0.27

0.275

0.28

0.285

0.29

µ [units/km2/sub-band]

Ou

tag

e p

robab

ilit

y Open access, simulation




Figure 4.5: Femtocell outage probability under different µ, with λ = 4 units/km2/sub-band and T = 0.1.


−2 −1.5 −1 −0.5 0 0.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log10

(T )

Ou

tag

e p

rob

abil

ity





Figure 4.6: Macrocell outage probability under different T , with λ = 8 units/km2/sub-band and µ = 8units/km2/sub-band.

−2 −1.5 −1 −0.5 0 0.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log10

(T )

Ou

tag

e p

rob

abil

ity





Figure 4.7: Femtocell outage probability under different T , with λ = 8 units/km2/sub-band and µ = 8units/km2/sub-band.

macrocell level, increasing the density of femtocell leads to more proportion of macrocell UEs becoming

open access UEs, which gives higher performance gap between open access and closed access. At the

femtocell level, the interference is observed at femtocell BSs, and the average number of macrocell UEs

in a femtocell becomes a more important factor influencing the performance gap.

Next, we present the numerical results of ρ∗ and ρ∗∗. The network parameters are as follows: λ = 4

units/km2/sub-band; µ = 4 units/km2/sub-band; ν(x) = 20 units/km2/sub-band if |x| < R, and

ν(x) = 0 otherwise; P = −60 dBm, and Q = −54 dBm.

Fig. 4.8 presents the value of ρ∗ at the macrocell level. We compute the actual value of ρ∗ by numer-

ically searching for the value such that (4.11) is equal to (4.14). Through the closed-form expressions

in Theorem 4, we are able to derive the upper and lower bounds of ρ∗. Through simulation, we can

also search for the value of ρ∗ such that the simulated outage probability of open access is equal to that

of closed access. Furthermore, we also simulate a more general scenario, where the received power is

randomly distributed, rather than perfectly fixed to a single level. We study the scenario where the

received power level of macrocell UEs is randomly distributed among 0.5P , P , 1.5P , and 2P with equal

probability. If a macrocell UE is handed off to a femtocell, then its targeted received power is multiplied

by ρ. The figure shows that ρ∗ is indeed within the upper bound and the lower bound, and the simulated

ρ∗ agrees with the analytical ρ∗, validating the correctness of our analysis. Furthermore, this remains

the case when the targeted received power is random, indicating the usefulness of our analysis in more

practical scenarios.

Figs. 4.9, 4.10, and 4.11 present the value of ρ∗∗ at the femtocell level. Fig. 4.9 shows ρ∗∗ under

different R as we fixed xB = (0, 100m); Figs. 4.10 and 4.11 show ρ∗∗ under different xB as we fixed


5 10 15 20 25 30 35 40 45 5020

25

30

35

40

45

50

55

60

R [m]

ρ∗

(dB

)

Lower bound

Upper bound

Analytical result

Simulation, with fixed received power

Simulation, with random received power

Figure 4.8: ρ∗ under different R at the macrocell level.

5 10 15 20 25 30 35 40 45 5010

15

20

25

30

35

40

45

50

55

60

R [m]

ρ∗∗

(dB

)

Lower bound

Upper bound

Analytical result



Figure 4.9: ρ∗∗ under different R at the femtocell level.

0 50 100 150 200 250 300 350 400 450 5005

10

15

20

25

30

35

xB [m]

ρ∗∗

(dB

)

Lower bound

Upper bound

Analytical result



Figure 4.10: ρ∗∗ under different xB , xB = (xB , 0), at the femtocell level.

0 50 100 150 200 250 300 350 400 4505

10

15

20

25

30

35

yB [m]

ρ∗∗

(dB

)

Lower bound

Upper bound

Analytical result



Figure 4.11: ρ∗∗ under different xB , xB = (0, yB), at the femtocell level.


R = 50 m. The y-coordinates of xB , yB , are fixed to 0 in Fig. 4.10, and the x-coordinates of xB , xB,

are fixed to 0 in Fig. 4.11. The results show that ρ∗∗ is indeed within the upper and lower bounds, and

the simulated values of ρ∗∗ agree with their analytical values, validating the correctness of our analysis.

Furthermore, ρ∗∗ decreases in R at a rate slightly faster than that of ρ∗, while it increases in xB and

yB , until saturating when the femtocell BS is near the macrocell edge. This quantifies when femtocells

are more beneficial as they decrease in size and increase in distance away from the macrocell BS.

4.5 Summary

We have presented a theoretical framework to analyze the performance difference between open access

and closed access in a two-tier femtocell network. Through establishing a stochastic geometric mod-

el, we capture the spatial patterns of different network components. We derive the numerical outage

probabilities of open access and closed access at the macrocell and femtocell levels. As in most uplink in-

terference analysis, the outage probability expressions are in non-closed forms. Hence, we further derive

closed-form bounds for the maximum tolerated received power enhancement, to compare the two access

modes. Simulations and numerical studies are conducted, validating the correctness of the analytical

model as well as the usefulness of the bounds even when the received power is random.

Chapter 5

Rate Maximization through

Spectrum Allocation and User

Association

In this chapter, we study spectrum allocation and user association in HWNs with multiple tiers of

BSs. A stochastic geometric approach is applied as the basis to derive the average downlink user

data rate in a closed-form expression. Then, the expression is employed as the objective function in

optimizing spectrum allocation and user association, which is of non-convex programming in nature.

A computationally efficient Structured Spectrum Allocation and User Association (SSAUA) approach

is proposed, solving the problem optimally and asymptotically optimally in two regions divided by

a parameter specific threshold. A Surcharge Pricing Scheme (SPS) is also presented, such that the

designed association bias values can be achieved in Nash equilibrium. Simulations and numerical studies

are conducted to validate the accuracy and efficiency of the proposed SSAUA approach and SPS.

5.1 System Model

5.1.1 Multi-tier Wireless Network

We consider an HWN with randomly spatially distributed K ≥ 2 tiers of BSs. As in conventional

stochastic geometric modeling of multi-tier HWNs [12,13,35,73,74], each tier of BSs independently form

a homogeneous Poisson point process (PPP) in two-dimensional Euclidean space R2. Let Φk denote

the PPP corresponding to tier-k BSs, with intensity λk. Without loss of generality, we assume that

λ1 < λ2 . . . < λK . (If λi = λj , i = j in reality, we may approximate by setting λj = λi + ξ, where ξ is

arbitrarily close to 0.) UEs are also modeled as a homogeneous PPP Ψ with intensity µ, independent of

all BSs. We assume each BS is connected to the core network by separate high-capacity wired or wireless

links that have no influence on our performance analysis. In addition, because we focus on downlink

analysis, we assume that the downlink and uplink of the system are operated in different spectra, so

that uplink interference and capacity have no influence on downlink analysis.

47

Chapter 5. Rate Maximization through Spectrum Allocation and User Association 48

5.1.2 Power and Path Loss Model

We define the tiers of BSs by their transmission power. Let Pk be the transmission power of tier-k

BSs, which is a given parameter. Let x and y denote some two-dimensional coordinates throughout

this chapter. If Pt(x), Pt(x) ∈ P1, P2, . . . , PK, is the transmission power from a BS at x and Pr(y)

is the received power at y, we have Pr(y) =Pt(x)hx,y

|x−y|γ , where |x − y|γ is the propagation loss function,

γ > 2 is the path loss exponent, and hx,y is the fast fading term. We assume that γ is constant for

all tiers. Corresponding to common Rayleigh fading with power normalization, hx,y is independently

exponentially distributed with unit mean. Let h(·) be the probability density function of hx,y.

5.1.3 Spectrum Allocation

In order to avoid cross-tier interference, different tiers of BSs are allocated separated spectrum. Note

that such separated spectrum allocation approach is also advocated in practical systems (e.g., Section 5.1

of [105]). Assume the total spectrum bandwidth isW . We allocate ηkW to each tier-k BS, where ηk is the

spectrum allocation factor and∑K

k=1 ηk = 1. Let η = (η1, η2, . . . , ηK). Note that BSs in the same tier are

operated on the same spectrum. We additionally consider the possible constraints ηmin,k ≤ ηk ≤ ηmax,k,

for k = 1, 2, . . . ,K. Clearly, we have∑K

k=1 ηmin,k ≤ 1 ≤∑K

k=1 ηmax,k. Furthermore, we assume that

0 < ηmin,1 ≤ ηmin,2 . . . ≤ ηmin,K and 0 < ηmax,1 ≤ ηmax,2 . . . ≤ ηmax,K , i.e., the network operator is likely

(but not necessarily) to allocate more spectrum to a tier with higher BS density.

Given a specific tier-k BS, it is common to assume that all its associated UEs are equally allocated

spectrum [12, 36, 74]. Hence, the per-UE assigned spectrum bandwidth is βk = ηkW/Nk, where Nk is a

random variable denoting the number of UEs associated with the BS.

5.1.4 Coverage Probability and UE Data Rate

Following conventional stochastic geometric modeling [12, 36, 74], in Sections 5.2 and 5.3, we assume

that UEs employ a single modulations coding scheme (MCS). In this case, let T denote the minimum

required Signal-to-Interference Ratio (SIR) of UEs. The coverage probability of a UE is defined as the

probability that its SIR is no lower than T [37]. Initially, we assume that if a UE experiences coverage

probability P′ and is allocated spectrum bandwidth β′, its data rate is β′ log(1 + T ) if the SIR is no

lower than T , and its data rate is 0 if the SIR is lower than T (i.e., outage occurs). Thus, the overall

data rate of the UE is β′ log(1+T )P′. In Section 5.5, we present the extension of our model considering

multiple MCSs, where N different threshold values T1, . . . , TN are accommodated.

Note that log is in base 2 throughout this chapter. Also, we have assumed the system is interference

limited [13,36,74–76], such that noise is negligible.

5.1.5 Flexible User Association

Given that a UE is located at y, it associates itself with the BS that provides the maximum biased

received power [12, 13,74] as follows:

BS(y) = arg maxx∈Φk,∀k

BkPk|x− y|−γ , (5.1)


where BS(y) denotes the location of the BS associated with the UE, and Pk|x−y|−γ is the received power

from a tier-k BS located at x, and Bk is the association bias, indicating the connecting preference of a UE

toward tier-k BSs. In this case, the resultant cell splitting forms a generalized Dirichlet tessellation, or

weighted Poisson Voronoi [106], shown in Fig. 1.2(b). Note that for B1, B2, . . . , BK , their effects remain

the same if we multiply all of them by the same positive constant. Thus, without loss of generality, in

this chapter, we normalize them such that∑K

k=1 Bk = 1. Let B = (B1, B2, . . . , BK).

Let Ak denote the probability that a UE associates itself with a tier-k BS, and A = (A1, A2, . . . , AK).

As derived in [12], we have

Ak =λk(PkBk)

2γ∑K

j=1 λj(PjBj)2γ

, (5.2)

and thus

Bk =P−1k (Ak/λk)

γ2∑K

j=1 P−1j (Aj/λj)

γ2

. (5.3)

Hence, there is a one-to-one mapping between A and B, so we can view them interchangeably.

5.1.6 Problem Statement

We first aim to derive a closed-form expression for the average UE data rate. Then, our objective is

to maximize the average UE data rate by optimizing the spectrum allocation factors η and the user

association bias values B (or equivalently A). Finally, we give a pricing scheme to incentivize each UE

to adopt the designed B.

5.2 Closed-form Average UE Data Rate

In this section, we derive the average UE data rate via stochastic geometric analysis. Consider a reference

UE, termed the typical UE, communicating with its BS, termed the typical BS. Due to the stationarity of

UEs and BSs, throughout this section we will re-define the coordinates so that the typical UE is located

at 0. We are interested in the typical UE since the average UE performance in the system is the same

as the performance of the typical UE [37].

First, we study the coverage probability given that the typical UE is associating with a tier-k BS

and their distance is d. In this case, the overall interference to the typical UE is the sum interference

from all tier-k BSs other than the typical BS. Let Ik(d) denote such interference. Then

Ik(d) =∑x∈Φ′

k

Pkhx,0

|x|γ. (5.4)

where Φ′k is the reduced Palm point process corresponding to all tier-k BSs other than the typical BS,

given that the typical BS is located at a distance of d from the typical UE. It can be shown that Φ′k is

a PPP with intensity 0 in B(0, d) and intensity λk in R2\B(0, d), where B(0, d) denotes the disk region

centered at 0 with radius d [37].


The distribution of Ik(d) is derived through its Laplace transform as follows:

LIk(d, s) = E

exp−

∑x∈Φ′

k

sPkhx,0

|x|γ

=exp

(−λk

∫R2\B(0,d)

(1−

∫R+

e−sPkh

|x|γ h(h)dh

)dx

)(5.5)

= exp

(−λk

∫R2\B(0,d)

sPk

|x|γsPk

|x|γ + 1dx

)(5.6)

= exp

(−2πλk

∫ ∞

d

sPkr

sPk + rγdr

), (5.7)

where (5.5) is obtained from the Laplace functional of PPP Φ′k [37], (5.6) is because the fading term is

exponentially distributed with unit mean, and (5.7) is through a transformation to polar coordinates.

Let Pcover,k(d) denote the conditional coverage probability of the typical UE (given k and d). Then

Pcover,k(d) =P(PkhxB ,0

dγ≥ TIk(d)

)=LIk(d, s)|s=Tdγ

Pk

, (5.8)

where xB is the coordinate of the typical BS, and |xB| = d. Substituting (5.7) into (5.8), we have

Pcover,k(d) = exp

(−2πλk

∫ ∞

d

Tdγr

Tdγ + rγdr

)t= r2

T2/γd2= exp

(−πλkT

2γ d2

∫ ∞

( 1T )

2γ

1

1 + tγ2

dt

). (5.9)

Furthermore, the probability density function of the distance between the typical UE and its associ-

ated tier-k BS is

fk(d) =2πλk

Akd exp

−πd2K∑j=1

λj

(PjBj

PkBk

) 2γ

(5.10)

=2πλk

Akd exp

(−πd2

λk

Ak

), (5.11)

where (5.10) is derived in [12], and (5.11) is by substituting (5.2) into (5.10).

Hence, the coverage probability Pcover,k of the typical UE associated with a tier-k BS can be computed

as

Pcover,k =

∫ ∞

0

fk(d)Pcover,k(d)dd

=

∫ ∞

0

2πλk

Akd exp

(−πd2

λk

Ak

)exp

(−πλk (T )

2γ d2

∫ ∞

( 1T )

2γ

1

1 + tγ/2dt

)dd

=πλk

Ak

1

π λk

Ak+ πλk (T )

2γ∫∞( 1

T )2γ

11+tγ/2 dt


=1

1 +AkC, (5.12)

where C , (T )2γ∫∞( 1

T )2γ

11+tγ/2 dt is a system-level constant only related to γ and T . Note that the

coverage probability is given in non-closed form in [12] for a system where the spectrum is shared by

all tiers. Here we are able to obtain a closed-form expression, mainly as a consequence of different tiers

using separate spectrum.

Let E0(βk) denote the expected spectrum bandwidth allocated to the typical UE (connecting to a

tier-k BS). Following the model in Section 5.1.3, E0(βk) equals the spectrum bandwidth allocated to the

typical tier-k BS divided by the average number of UEs associated with it conditioned on the typical

UE, which is Akµ/λk + 1. Hence,

E0(βk) =ηkW

Akµ/λk + 1. (5.13)

Then, by Section 5.1.4, the conditional expected data rate of the typical UE, given it is associated with

a tier-k BS, can be computed as [12,74]

Rk = E0(βk) log(1 + T )Pcover,k. (5.14)

Note that by doing so, we slightly underestimate the average data rate because the coverage event

and βk are not completely independent. Although some efforts have been made to approximate their

correlation [13, 107], all of them are inexact but result in tremendous mathematical complexity. In

Section 5.6, we show that the resultant analysis is close to actual performance via simulation.

Finally, the average data rate of the typical UE, and hence the average data rate per UE in the

system, is

F =K∑

k=1

AkRk =K∑

k=1

AkE0(βk) log(1 + T )Pcover,k

=K∑

k=1

ηkW log(1 + T )

(Akµ/λk + 1)( 1Ak

+ C). (5.15)

Note that stochastic geometric analysis often leads to non-closed forms requiring numerical integra-

tions (e.g., [12, 13, 34, 35]), due to the integral form of the Laplace functional or generating functional

of PPPs applied in analysis [37,39]. Fortunately, our derived closed-form expression for the average UE

data rate facilitates the tractability of the resultant optimization problem.

5.3 Optimization Problem and SSAUA

We aim to maximize the average UE data rate F with respect to η and B. As there is a one-to-

one mapping between A and B, we study the optimization problem over (η,A) instead for analytical

convenience. This is formally stated as optimization problem P as follows:

maximizeη,A

F(η,A) =K∑

k=1

ηkMk(Ak)


subject to

K∑k=1

ηk = 1, ηmin,k ≤ ηk ≤ ηmax,k,∀k,

K∑k=1

Ak = 1, Ak ≥ 0, ∀k, (5.16)

where Mk(Ak) is defined as

Mk(Ak) =1

(Akµ/λk + 1)(

1Ak

+ C) . (5.17)

Problem P is non-convex and cannot be solved through standard methods. Instead, we investigate

into two important structures of the optimal solution, termed density thresholding and priority order-

ing, based on which we propose a computationally efficient Structured Spectrum Allocation and User

Association (SSAUA) approach to solve the problem.

5.3.1 Density Thresholding Structure

First, we define an important parameter

ak ,√λk/(µC). (5.18)

Note that Mk(Ak) is increasing on [0, ak] and decreasing on [ak,∞). We further observe several useful

properties of Mk(Ak), which are presented in Appendix B.1. Based on these properties, we obtain the

following lemma, whose proof is given in Appendix B.2.

Lemma 2. Consider a potential solution (η∗∗,A∗∗) to Problem P. If ∃i = j, such that A∗∗i < ai and

A∗∗j > aj , then (η∗∗,A∗∗) is not an optimal solution.

Lemma 2 suggests that, in an optimal solution, every Ak must be on the same side of ak. This

directly leads to the following theorem, which is fundamental to our optimization solution.

Theorem 7. (Density Thresholding) Let (η∗,A∗) be an optimal solution to ProblemP. If∑K

k=1 ak >

1, then ∀k,A∗k ≤ ak; if

∑Kk=1 ak < 1, then ∀k,A∗

k ≥ ak; if∑K

k=1 ak = 1, then ∀k,A∗k = ak.

Proof. If∑K

k=1 ak > 1, because∑K

k=1 A∗k = 1, ∃l such that A∗

l < al. This leads to A∗k ≤ ak, ∀k,

according to Lemma 2. The cases where∑K

k=1 ak < 1 and∑K

k=1 ak = 1 are similar.

Note that, the condition∑K

i=1 ai > 1 (i.e.,√

1C

(∑Ki=1

√λi

)>

√µ) is referred to as the optimality

region throughout this chapter, since the optimization problem P can be solved optimally in this case,

as shown later in Section 5.3.2. The condition∑K

i=1 ai < 1 (i.e.,√

1C

(∑Ki=1

√λi

)<

√µ) is referred to

as the asymptotic-optimality region throughout the chapter, since the optimization problem P can be

solved asymptotically optimally in this case, as shown later in Section 5.3.3. If∑K

i=1 ai = 1, Problem P

can be trivially solved and is ignored in the rest of our discussion. Note that because ak can be computed

directly from the given parameters, one can judge in which region Problem P falls before solving the

problem. Next, the solution to P will be investigated separately under∑K

i=1 ai > 1 and∑K

i=1 ai < 1.



i=1 ai > 1

In this case, the original Problem P becomes Problem P1 as follows:

maximizeη,A

F(η,A) =

K∑k=1

ηkMk(Ak)

subject to

K∑k=1


K∑k=1

Ak = 1, 0 ≤ Ak ≤ ak, ∀k. (5.19)

We first observe an important ordering property of the optimal solution to P1, as shown in the

following lemma, whose proof is given in Appendix B.3.

Lemma 3. (Ordering Property) If A∗ is optimal for Problem P1, then M1(A∗1) ≤ M2(A

∗2) ≤ . . . ≤

MK(A∗K).

Next, by sequentially computing η∗ as follows:

η∗K = min(1−∑K−1

k=1 ηmin,k, ηmax,K),

η∗K−1 = min(1− η∗K −∑K−2

k=1 ηmin,k, ηmax,K−1),

. . . ,

η∗l = min(1−∑K

k=l+1 η∗k −

∑l−1k=1 ηmin,k, ηmax,l),

. . . ,

η∗1 = min(1−∑K

k=2 η∗k, ηmax,1),

(5.20)

we have the following theorem:

Theorem 8. (Priority Ordering) If A∗ is optimal for Problem P1, then (η∗,A∗), where η∗ is

computed in (5.20), is an optimal solution to P1.

Proof. Consider Problem P1A as follows:

maximizeη

K∑k=1

ηkMk(A∗k)

subject to

K∑k=1

ηk = 1, ηmin,k ≤ ηk ≤ ηmax,k,∀k. (5.21)

It is a simple linear programming problem with ordered linear coefficients in the objective, since

M1(A∗1) ≤ M2(A

∗2) ≤ . . . ≤ MK(A∗

K) due to Lemma 3. Note that η∗ does not depend on the exact

values of A∗; it only requires the ordering property as shown in Lemma 3. Also, η∗ is in the feasible

region due to∑K

k=1 ηmin,k ≤ 1 ≤∑K

k=1 ηmax,k. It is easy to verify that (η∗,A∗) is an optimal solution

to P1.

In Theorem 8, we note that (5.20) indicates a priority ordering structure in spectrum allocation. Tier-

K has the highest priority in spectrum allocation, followed by tier-(K − 1), and so forth. Furthermore,


Theorem 8 provides a means to derive an optimal η∗ regardless of the A∗ values. We need one further

step to derive the corresponding optimal A∗ by solving the following Problem P1B:

maximizeA

K∑k=1

η∗kMk(Ak)

subject to

K∑k=1

Ak = 1, 0 ≤ Ak ≤ ak, ∀k. (5.22)

Note that P1B is a convex programming problem, since Mk(Ak) is concave on [0, ak]. Thus, A∗ can

be computed by a computationally efficient algorithm, such as the interior-point method. Hence both

steps to compute the optimal solution (η∗,A∗) have low computational complexity.

In summary, under∑K

i=1 ai > 1, the original optimization problem can be solved optimally, and thus∑Ki=1 ai > 1 is referred to as the optimality region.


i=1 ai < 1

In this case, the original Problem P becomes Problem P2 as follows:

maximizeη,A

F(η,A) =K∑

k=1

ηkMk(Ak)

subject to

K∑k=1


K∑k=1

Ak = 1, Ak ≥ ak, ∀k. (5.23)

Problem P2 is more complicated than Problem P1, as Mk(Ak) is not concave, but an S-shaped function,

in the feasible region. Hence, P2 generally incurs high computational complexity even if an optimal η∗

is given [108,109].

Therefore, instead of directly solving P2, we first approximate Mk(Ak) by Mk(Ak) defined as follows:

Mk(Ak) =1

(Akµ/λk)(

1Ak

+ C) . (5.24)

Note that this approximation is reasonable when Akµ/λk is much larger than 1, e.g., when µ is large.

This observation is also supported by the performance gap as derived in Section 5.3.3.2. Some useful

properties of Mk(Ak) are shown in Appendix B.4.

The approximated problem is referred to as Problem P2A, where we simply replace the objective

function of P2 by the following:

F′(η,A) =K∑

k=1

ηkMk(Ak). (5.25)


5.3.3.1 Solution to P2A

The important ordering property still holds for Problem P2A, as formalized in the following lemma,

whose proof is given in Appendix B.5.

Lemma 4. (Ordering Property) If A∗ is optimal for Problem P2A, then M1(A∗1) ≤ M2(A

∗2) ≤ . . . ≤

MK(A∗K).

We observe that with the same ordering property, (5.20) can again be adopted as an optimal solution

to P2A, leading to the following theorem:

Theorem 9. (Priority Ordering) If A∗ is optimal for Problem P2A, then (η∗, A∗), where η∗ is

computed the same way as η∗ in (5.20), is an optimal solution to Problem P2A.

Proof. The proof is similar to that of Theorem 8.

Given an optimal η∗ for P2A, we find the corresponding optimal A∗ for P2A by solving the following

Problem P2B:

maximizeA

K∑k=1

η∗kMk(Ak)

subject toK∑

k=1

Ak = 1, Ak ≥ ak,∀k. (5.26)

Unlike in the optimality region, here we have an explicit solution, as stated in the following theorem:

Theorem 10. Given an optimal η∗ for P2A (computed the same way as η∗ in (5.20)), the corresponding

optimal A∗ can be expressed as follows:A∗

k = ak, k ≥ 2

A∗1 = 1−

∑Kl=2 A

∗l .

(5.27)

See Appendix B.6 for the proof.

Note that both (5.20) and (5.27) can be computed with low computational complexity.

5.3.3.2 Bounding the Performance Gap

Since (η∗, A∗) is optimal for P2A rather than P2, we next quantify the performance gap between

(η∗, A∗) and an optimal solution (η∗,A∗) to P2.

The performance gap is defined as

E = F(η∗,A∗)− F(η∗, A∗). (5.28)

Because F(η∗, A∗) ≤ F(η∗,A∗) ≤ F′(η∗,A∗) ≤ F′(η∗, A∗), we have

E ≤ F′(η∗, A∗)− F(η∗, A∗) , E′. (5.29)


Substituting η∗ and A∗ into E′, we have

E′ =K∑

k=1

η∗kA∗

kµ

λk

(A∗

kµ

λk+ 1)(

1

A∗k

+ C) . (5.30)

Therefore, the relative performance gap is bounded:

ϵ , E

F(η∗,A∗)≤ E′

F(η∗, A∗)

=

∑Kk=1

η∗k

A∗kµ

λk

(A∗

kµ

λk+1

)(1

A∗k

+C

)∑K

k=1η∗k(

A∗kµ

λk+1

)(1

A∗k

+C

) (5.31)

(a)

≤ maxk

λk

A∗kµ

≤ maxk

λk

akµ

=√λKC/µ, (5.32)

where inequality (a) is obtained by observing the common factor in the summations in the numerator

and denominator of (5.31). This result implies that ϵ scales as O(√

λK/µ). Note that when µ ≫ λK ,

we have ϵ ≃ 0 and the performance of SSAUA is asymptotically optimal. Thus∑K

i=1 ai < 1 is referred

to as the asymptotic-optimality region.

5.3.4 Computational Complexity Comparison

In this subsection, we discuss the computational complexity of SSAUA and that of an exhaustive search

approach to solve the original optimization problem.

5.3.4.1 Computational Complexity of SSAUA

In the optimality region, the optimal η∗ can be derived with computational complexity O(K) through

(5.20). Given the optimal η∗, the remaining problem (5.22) is a convex optimization problem, which

can be solved using the interior-point method. Let ε denote the error bound between the output and

the optimal solution. According to Section 11.5 of [110], the interior-point method involves O(log(Kε )K)

Newton iterations; each Newton iteration involves a matrix inversion operation, which has computational

complexity O(K3). Therefore, the overall computational complexity is O(log(Kε )K4). In the asymptotic-

optimality region, the asymptomatically optimal η∗ and A∗ can be derived through (5.20) and (5.27),

respectively, with an overall computational complexity O(K).

5.3.4.2 Computational Complexity of Exhaustive Search

As explained in detail in Appendix B.7, we observe that at least one of the optimal solutions to P,

(η∗,A∗), has the following property: there is at most one k ∈ 1, 2, . . .K such that ηmin,k < η∗k < ηmax,k;

∀j = k, either η∗j = ηmin,j or η∗j = ηmax,j (i.e., at the boundary). Thus, the search for η∗ needs to be

performed at these boundary cases only, leading to a complexity of O(2K).

In the optimality region, if η is fixed, the remaining problem is a convex optimization problem. The

interior-point method takes another fold of computational complexity of O(log(Kε )K4). Consequently,


the overall computational complexity is O(2K log(Kε )K4). In the asymptomatic-optimality region, if η

is fixed, the remaining problem is still a non-convex optimization problem. A numerical search over

all solutions to the KKT condition is required, leading to another fold of computational complexity of

O( 2K

ε ). Consequently, the overall computational complexity is O( 4K

ε ).

5.4 Nash Equilibrium for SSAUA

Individual UEs may behave selfishly to derive unfair advantage despite our design of B∗ (or equivalently

A∗). Thus, in this section, we propose a Surcharge Pricing Scheme (SPS), such that the designed B∗ is

the natural outcome of a Nash equilibrium. We assume that the designed spectrum allocation factors

η∗ are centrally maintained by the network operator.

We consider a reference individual UE, whose association bias values are B′ = (B′1, B

′2, . . . , B

′K). Let

A′ = (A′1, A

′2, . . . , A

′K) be its corresponding association probabilities. For the other UEs, suppose they

all obey the association bias values B∗ assigned by the network operator. Similar to the discussions in

Sections 5.1 and 5.2, the average data rate of the reference UE is

F =K∑

k=1

η∗kW log(1 + T )

(A∗kµ/λk + 1)( 1

A′k+ C)

. (5.33)

If the reference UE performs an optimization on F with respect to A′, the resultant optimal A′∗ =

(A′∗1 , A

′∗2 , . . . , A

′∗K) is unlikely to be the same as A∗. Therefore, we add the following Surcharge Pricing

Scheme: the network operator applies a surcharge ck to each UE associated with a tier-k BS. Let

c = (c1, c2, . . . , cK). In this case, the average surcharge for the reference UE is∑K

k=1 ckA′k. Accordingly,

the reference UE will perform the following optimization Problem P3:

maximizeA′

F′ =

K∑k=1

η∗kW log(1 + T )

(A∗kµ/λk + 1)

(1A′

k+ C

) − ckA′k

subject to

K∑k=1

A′k = 1, A′

k ≥ 0. (5.34)

Different from P, it can be shown that P3 is a standard convex optimization problem. By the KKT

conditions, its optimal solution A′∗ satisfies

Hk

(1 + CA′∗k )

2− ck − ν + θk = 0, (5.35)

θkA′∗k = 0, θk ≥ 0, (5.36)

where Hk =η∗kW log(1+T )A∗

kµ/λk+1 , θk is a Lagrange multiplier corresponding to the inequality constraint A′k ≥ 0,

and ν is a Lagrange multiplier corresponding to the equality constraint∑K

k=1 A′k = 1.

Setting A′∗k = A∗

k, we have

ck =

∞, if A∗k = 0,

Hk

(1+CA∗k)

2 − ν, otherwise.(5.37)


Note that ν could be set arbitrarily due to the equality constraint. Without loss of generality, we

set ν = minkHk

(1+CA∗k)

2 so that the minimum surcharge among tiers is 0. As a consequence, a Nash

Equilibrium is achieved where every UE adopts the assigned B∗.

5.5 The Multiple-MCS Case

In this section, we discuss the usefulness of our proposed SSAUA in systems with multiple modulation

and coding schemes (i.e., the multiple-MCS case). Instead of considering only one SIR threshold T

(see Section 5.1.4), N SIR threshold values, T1, T2, . . . , TN , where T1 < T2 < . . . < TN , corresponding

to N MCSs, are accommodated. In this case, if a UE is allocated spectrum bandwidth β′, its data

rate is β′ log(1 + TN ), β′ log(1 + TN−1), . . ., β′ log(1 + T1), and 0, respectively, if its SIR is in [TN ,∞),

[TN−1, TN ), . . ., [T1, T2), and [0, T1).

5.5.1 Average UE Data Rate

First, similar to the derivations of (5.4)-(5.12), given that the typical UE is associated with a tier-k BS,

we can find its coverage probabilities under T1, T2, . . . , TN to be

P(SIRk ≥ Tn) =Pcover,k,n =1

1 +AkCn, n = 1, 2, . . . , N, (5.38)

where Cn , (Tn)2γ∫∞( 1

Tn)

2γ

11+tγ/2 dt. Note that because T1 < T2 < . . . < TN , we have C1 < C2 < . . . <

CN and Pcover,k,1 > Pcover,k,2 > . . . > Pcover,k,N .

Then, the conditional expected data rate of the typical UE, given it is associated with a tier-k BS,

is recomputed as

Rk =E0(βk)

[log(1 + TN )Pcover,k,N +

N−1∑n=1

log(1 + Tn)(Pcover,k,n − Pcover,k,n+1

)](5.39)

=E0(βk)

[log(1 + T1)Pcover,k,1 +

N∑n=2

(log(1 + Tn)− log(1 + Tn−1)

)Pcover,k,n

](5.40)

=ηkW

(Akµ/λk + 1)

[log(1 + T1)

1 +AkC1+

N∑n=2

log(1 + Tn)− log(1 + Tn−1)

1 +AkCn

](5.41)

=ηkW

(Akµ/λk + 1)

( N∑n=1

bn1 +AkCn

), (5.42)

where b1 , log(1 + T1), b2 , log(1 + T2)− log(1 + T1), . . . , bN , log(1 + TN )− log(1 + TN−1).

Similar to (5.15), the average data rate of the typical UE (unconditioned on k), and hence the average

data rate per UE in the system, is recomputed as

F =K∑

k=1

AkRk =K∑

k=1

ηkW

(Akµ/λk + 1)

( N∑n=1

bn1Ak

+ Cn

). (5.43)


The optimization problem (5.16) is updated correspondingly, such that Mk(Ak) is redefined as

Mk(Ak) =1

(Akµ/λk + 1)

( N∑n=1

bn1Ak

+ Cn

). (5.44)

5.5.2 SSAUA in the Multiple-MCS Case

In this subsection, we discuss the usefulness of our proposed SSAUA in solving the modified optimization

problem considering multiple-MCS.

5.5.2.1 Density Thresholding

We redefine ak as the unique positive solution to M ′k(Ak) = 0. The existence and uniqueness of ak

is shown in Appendix B.8. Properties (M-1) to (M-4) presented in Appendix B.1 still hold with the

redefined Mk(Ak) and ak. The proofs of these properties are also shown in Appendix B.8.

As a result, Lemma 2 and Theorem 7 still hold in the multiple-MCS case. Note that different from the

single-MCS case (5.18), ak cannot be represented in an explicit expression. As shown in Appendix B.8,

ak is the unique solution to fk(A) = 0, where fk(A) =∑N

n=1bn(1−A2Cnµ/λk)

(ACn+1)2 is a decreasing function. A

simple binary search method can be applied to compute ak, which has low computational complexity.

5.5.2.2 Optimality Region

When∑K

i=1 ai > 1, since Properties (M-1) to (M-4) still hold, Lemma 3 and Theorem 8 still hold in

the multiple-MCS case. Then, the same method presented in Section 5.3.2 can be applied to solve the

optimization problem.

5.5.2.3 Asymptotic-optimality Region

When∑K

i=1 ai < 1, Mk(Ak) is approximated by Mk(Ak) redefined as follows:

Mk(Ak) =1

Akµ/λk

[N∑

n=1

bn1Ak

+ Cn

]. (5.45)

If Properties (M-1’)-(M-6’) presented in Appendix B.4 still hold for the above redefined Mk(Ak), then

Lemma 4, Theorem 9, and Theorem 10 still hold in the multiple-MCS case, and the same method

employed in Section 5.3.3 can be applied to solve the optimization problem.

It is straightforward to verify Properties (M-1’), (M-3’), and (M-4’). Property (M-6’) is implied by

(M-4’) and (M-5’). However, Properties (M-2’) and (M-5’) are difficult to verify analytically, as ak is no

longer expressed in closed form. Therefore, we conduct numerical validation to check Properties (M-2’)

and (M-5’) under a wide range of parameter settings.

We set T = 10−1, 10−0.9, . . . , 100.9, 101, U = 101, 101.1, . . . , 103, Γ = 3, 3.5, 4, 4.5, 5. It is shownthat Properties (M-2’) and (M-5’) are true under N = 2, 3, and 4; ∀T1, T2, . . . , TN ∈ T (T1 < T2 < . . . <

TN ); ∀µ/λi, µ/λj ∈ U (µ/λi > µ/λj); and ∀γ ∈ Γ. Therefore, at least for the wide range of parameter

settings that are tested, the proposed SSAUA is still useful in the asymptotic-optimality region. See

Appendix B.9 for visualized verification results.


100 200 300 400 500 6000.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

µ (units/km2)

Dat

a R

ate

(Mb

ps)

Analytical (SSAUA)

Analytical (upper bound)

Analytical (exhaustive search)

Simulation (SSAUA)

Threshold value

optimalityregion

asymptotic−optimalityregion

Figure 5.1: Average UE data rate under different UE density µ.

100 150 200 250 300 350 400 450 500 550 600 6500.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

µ (units/km2)

Dat

aR

ate

(Mbps)

SSAUA (analytical)Scheme 1 (analytical)Scheme 2 (analytical)Scheme 3 (analytical)Scheme 4 (analytical)SSAUA (simulation)Scheme 1 (simulation)Scheme 2 (simulation)Scheme 3 (simulation)Scheme 4 (simulation)Threshold value


optimalityregion

Figure 5.2: Comparison of different schemes under different UE density µ.

5.6 Numerical Study

In this section, we present numerical studies on the performance of SSAUA. Unless otherwise stated,

we label the SSAUA solution as (η∗, B∗) and (η∗, B∗) in the optimality and asymptotic-optimality

regions, respectively. Note that (η∗, B∗) is optimal in the optimality region. We use (η∗,B∗) to label an

optimal solution obtained from exhaustive search in the asymptotic-optimality region. In this section,

each simulation point is derived as follows: in each round of simulation, UEs and BSs are generated on

a 10 km × 10 km square, and the UEs in the central 5 km × 5 km square are sampled for performance

evaluation (in order to remove the edge effect). Each simulation data point is averaged over all sampled

UEs during 1000 rounds of simulations.

5.6.1 Average UE Data Rate under Different UE Densities

In this subsection, we study the average UE data rate of SSAUA under different values of UE density

µ, and compare it with four reference schemes. The network parameters are as follows: K = 3, λ1 = 1

units/km2, λ2 = 5 units/km2, λ3 = 10 units/km2, P1 = 56 dBm, P2 = 46 dBm, P3 = 36 dBm,

ηmin,1 = 0.2, ηmin,2 = 0.25, ηmin,3 = 0.3, ηmax,1 = 0.35, ηmax,2 = 0.4, ηmax,3 = 0.45, γ = 4, W = 200

MHz, and T = 0.2.

The performance of SSAUA is shown in Fig. 5.1. A vertical line indicates the threshold value of µ,

as given in Theorem 7, separating the optimality and asymptotic-optimality regions. For both regions,

we show results of the analytical and simulated performance of SSAUA. Since SSAUA is not optimal in

the asymptotic-optimality region, we also add two sets of results accordingly: the optimal performance


3 3.5 4 4.5 50

0.5

1

1.5

2

2.5

3

γ

Data

Rate

(Mbps)

optimality region, µ = 100 units/km2

Anlytical(SSAUA)Simulation(SSAUA)

(a) Optimality region.

5 3 3.5 4 4.5 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

γ

Data

Rate

(Mbps)

asymptotic-optimality region, µ = 500 units/km2

Anlytical(SSAUA)Analytical (upper bound)Analytical (exhaustive search)Simulation (SSAUA)

(b) Asymptotic-optimality re-gion.

Figure 5.3: Average UE data rate under different path loss exponent γ.

3 3.5 4 4.5 50

0.5

1

1.5

2

2.5

3

γ

Data

Rate

(Mbps)


SSAUAScheme 1Scheme 2Scheme 3Scheme 4

(a) Asymptotic-optimality re-gion.

5 3 3.5 4 4.5 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

γ

Data

Rate

(Mbps)




Figure 5.4: Comparison of different schemes under different path loss exponent γ, analytical case.

F(η∗,B∗) through exhaustive search and its analytical upper bound F′(η∗, B∗). Fig. 5.1 illustrates that

the performance of SSAUA is very close to the optimal solution in the asymptotic-optimality region.

Next, we compare the performance of SSAUA with four reference schemes listed as follows:

• Scheme 1 employs equal spectrum allocation, and user association based on the maximum received

power.

• Scheme 2 employs equal spectrum allocation, and optimal user association as in SSAUA.

• Scheme 3 employs optimal spectrum allocation as in SSAUA, and user association based on the

maximum received power.

• Scheme 4 employs optimal spectrum allocation as in SSAUA, and user association based on a

simple range expansion scheme by setting Bk = 1Pk

, ∀k.

Fig. 5.2 shows that SSAUA outperforms these schemes. This illustrates that both the spectrum allocation

component and the user association component of SSAUA bring performance benefits. Note that due

to the approximation made in (5.14), the analytical data rate is slightly lower than the simulated one,

matching our discussion in Section 5.2.


3 3.5 4 4.5 50

0.5

1

1.5

2

2.5

3

γ

Data

Rate

(Mbps)



(a) Optimality region.

5 3 3.5 4 4.5 50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

γ

Data

Rate

(Mbps)




Figure 5.5: Comparison of different schemes under different path loss exponent γ, simulation case.

0 100 200 300 400 500 600 7000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

µ (units/km2)

Val

ue

B∗

1(o.)

B∗

2(o.)

B∗

3(o.)

B∗

1(a. o.)

B∗

2(a. o.)

B∗

3(a. o.)

B∗

1(a. o.)

B∗

2(a. o.)

B∗

3(a. o.)

Threshold

optimalityregion


Figure 5.6: Designed association bias values under different UE density µ. “o” represents the optimalityregion, and “a. o.” represents the asymptotic-optimality region.

5.6.2 Average UE Data Rate under Different Path Loss Exponents

Fig. 5.3 shows the optimal network performance under different values of the path loss exponent γ. The

network parameters are the same as those used in Fig. 5.1 except µ is fixed at 100 (i.e., optimality region)

and 500 (i.e., asymptotic-optimality region) units/km2 in Figs. 5.3 (a) and (b), respectively. This figure

further confirms the observations from Fig. 5.1. Furthermore, it shows that SSAUA is effective for a

wide range of path loss conditions.

Fig. 5.4 shows further analytical performance comparison of SSAUA and the four reference schemes

stated in Section 5.6.1. The results illustrate that in both the optimality and asymptotic-optimality

regions, SSAUA outperforms Scheme 2, Scheme 3, and Scheme 4, and both Scheme 2 and Scheme 3 out-

perform Scheme 1, confirming the observations from Fig. 5.2. Fig. 5.5 shows the simulated performance

comparison of SSAUA and the four reference schemes. The results show that the performance orders of

the five schemes agree with those in Fig. 5.4 in both the optimality region and the asymptotic-optimality

region.

5.6.3 Association Bias Values and Prices

Fig. 5.6 shows B∗, B∗, and B∗; and Fig. 5.7 shows their corresponding prices c∗, c∗, and c∗, under

different µ values. The other network parameters are the same as those used in Fig. 5.1. We observe

that the B∗ and c∗ computed based on SSAUA are close to their counterparts B∗ and c∗.


0 100 200 300 400 500 600 700

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

µ (units/km2)P

rici

ng

c∗

1(o.)

c∗

2(o.)

c∗

3(o.)

c∗

1(a. o.)

c∗

2(a. o.)

c∗

3(a. o.)

c∗

1(a. o.)

c∗

2(a. o.)

c∗

3(a. o.)

Threshold

optimalityregion


Figure 5.7: Pricing values under different UE density µ. “o” represents the optimality region, and “a.o.” represents the asymptotic-optimality region.

2 3 4 5 6 7 8 9 10

10−4

10−2

100

102

104

K

Tim

esp

ent

(s)

SSAUA, asymptotic−optimality region

SSAUA, optimality region

Exhaustive search, optimality region

Exhaustive search, asymptotic−optimality region

Figure 5.8: Comparison of run time.

5.6.4 Run Time Experiment

In this subsection, a run time experiment is presented to compare the computational complexity of

SSAUA with that of exhaustive search. The experiment is executed by Matlab R2011a on an ASUS PC

with Intel i7-3610QM 2.3GHz processor and 4GB RAM. The results are averaged over 1000 runs for

SSAUA and 10 runs for exhaustive search (both with randomly generated parameters). Fig. 5.8 shows

that the run time of SSAUA is almost negligible compared with exhaustive search. Note that the y-axis

is in log scale. When K increases, the run time of exhaustive search exhibits an exponential growth

tendency, while SSAUA remains computationally efficient.

5.6.5 Performance of SSAUA in the Multiple-MCS Case

In this subsection, a numerical experiment is presented to validate the usefulness of SSAUA in the

multiple-MCS case. The network parameters are as follows: N = 3, T1 = 0.1, T2 = 0.3, T3 = 0.5, K = 3,

λ1 = 1 units/km2, λ2 = 5 units/km2, λ3 = 10 units/km2, P1 = 56 dBm, P2 = 46 dBm, P3 = 36 dBm,

ηmin,1 = 0.2, ηmin,2 = 0.25, ηmin,3 = 0.3, ηmax,1 = 0.35, ηmax,2 = 0.4, ηmax,3 = 0.45, γ = 4, and W = 200

MHz.

Fig. 5.9 shows the comparison of SSAUA with the four reference schemes. Similar to the results in

5.6.1, in the multiple-MCS case, we can still observe that SSAUA outperforms Scheme 2, Scheme 3,

and Scheme 4, and both Scheme 2 and Scheme 3 outperform Scheme 1. Both the spectrum allocation

part and the user association part used in SSAUA bring performance benefits, and the joint spectrum

allocation and user association of SSAUA brings the greatest performance benefits in the Multiple-MCS


100 150 200 250 300 350 4001

2

3

4

5

6

7

8

µ (units/km2)

Data

Rate

(Mbps)

SSAUA (analytical)Scheme 1 (analytical)Scheme 2 (analytical)Scheme 3 (analytical)Scheme 4 (analytical)SSAUA (simulation)Scheme 1 (simulation)Scheme 2 (simulation)Scheme 3 (simulation)Scheme 4 (simulation)Threshold value

optimalityregion


Figure 5.9: Comparison of different schemes in the multiple-MCS case.

case.

5.6.6 Performance Comparison under More Realistic Network Topologies

In this subsection, we present simulation of SSAUA and the four reference schemes in more realistic

network settings. First, the UE and BS density settings follow the 3GPP simulation guidelines from

Annex A of [104], which recommend that there are between 1 to 10 small cell BSs, and 10 to 100 UEs

per macrocell. Second, BSs are non-Poisson, and UEs are inhomogeneously distributed.

The Matern hard core (MHC) point process is an alternative point process adopted in the literature

to counter the drawback of PPP modeling of HWNs [111, 112]. MHC point processes can additionally

capture the effect that two BSs are unlikely to be located very close to each other. In our simulation,

tier-k BSs are generated as an MHC point process as follows: First, we generate a PPP with intensity

λ′k. Each point in the PPP is associated with a “mark”, which is independently uniformly distributed

on [0, 1]. A point is retained in the point process if its mark is the largest among all the points within

a distance Dk from it (or there are no other points within this range); otherwise, the point is removed

from the point process. The remaining points form an MHC point process. Note that the distance

between any two points in the point process is no less than Dk. In our simulation, each tier of BSs

is independently generated as an MHC point process. Then, the spectrum allocation factors and user

association bias values are determined, by either SSAUA or one of the reference schemes, based on PPP

BSs with equivalent densities λk = 1−e−πD2kλ′

k

πD2kλ

′k

.

We also consider the scenario where UEs are likely to crowd near BSs. At the beginning, UEs follow

a PPP with intensity µ′. Then, each tier-k BS brings increment of µ0 to the UE density in the disk

region centered at the BS with a radius of Rk. In our simulation, the spectrum allocation factors and

user association bias values are determined, by either SSAUA or one of the reference schemes, based on

an equivalent UE density µ = µ′ + µ0

(∑Kk=1 λkπR

2k

).

In Figs. 5.10 and 5.11, we set γ = 4; (P1, P2, P3) = (56, 43, 33) dBm; ηmin,1 = 0.2, ηmin,2 = 0.25,

ηmin,3 = 0.3, ηmax,1 = 0.35, ηmax,2 = 0.4, ηmax,3 = 0.45; W = 50 MHz; N = 3, T1 = 1, T2 = 2, T3 = 3.

In Fig. 5.10, (λ′1, λ

′2, λ

′3) = (1, 2, 3) unit/km2, (D1, D2, D3) = (200, 80, 40) m, (R1, R2, R3) = (200, 80, 40)

m, and µ0 = µ′. The equivalent BS densities are (λ1, λ2, λ3) = (0.9397, 1.9603, 2.9775) unit/km2, and

the equivalent UE density is µ = 1.1725µ′. In Fig. 5.11, we set (λ′1, λ

′2, λ

′3) = (1, 5, 10) unit/km2,

(D1, D2, D3) = (200, 80, 40) m, (R1, R2, R3) = (200, 80, 40) m, and µ0 = µ′. The equivalent BS densities

are (λ1, λ2, λ3) = (0.9397, 4.7569, 9.7528) unit/km2, and the equivalent UE density is µ = 1.2628µ′. In

both figures, µ ranges from 10 to 100. The performance ratios of SSAUA to all four reference schemes

are shown in the figures. The result suggests that, under more realistic network topologies, SSAUA still


10 20 30 40 50 60 70 80 90 1001

1.05

1.1

1.15

1.2

1.25

1.3

1.35

1.4

1.45

µ (units/km2)

Per

form

ance

ratio

Gain over Scheme 1Gain over Scheme 2Gain over Scheme 3Gain over Scheme 4

Figure 5.10: Comparison of different schemes under more realistic network settings, (λ′1, λ

′2, λ

′3) = (1, 2, 3)

unit/km2.

10 20 30 40 50 60 70 80 90 1001

1.05

1.1

1.15

1.2

1.25

1.3

1.35

µ (units/km2)

Per

form

ance

ratio

Gain over Scheme 1Gain over Scheme 2Gain over Scheme 3Gain over Scheme 4

Figure 5.11: Comparison of different schemes under more realistic network settings, (λ′1, λ

′2, λ

′3) =

(1, 5, 10) unit/km2.

brings useful performance gain compared with the reference schemes.

5.7 Summary and Discussion

In this chapter, we provide a theoretical framework to study the optimization of spectrum allocation and

user association in HWNs. We establish a stochastic geometric model that captures the random spatial

patterns of BSs and UEs, and a closed-form expression of the analytical average UE data rate is derived.

We then consider the problem of maximizing the average UE data rate by optimizing the spectrum

allocation factors and user association bias values, which is non-convex programming in nature. We

propose the SSAUA approach to solve this problem with low computational complexity. We show that

SSAUA is either optimal or asymptotically optimal with a quantified bound scaling as O(√λK/µ). We

also propose a pricing scheme such that the designed association bias values can be achieved in Nash

equilibrium.

So far we have focused on the spectrum allocation under the constraints of ηmin,k ≤ ηk ≤ ηmax,k, ∀k.We note that our proposed SSAUA is effective when ηmin,k > 0 and ηmax,k < 1, which is necessary in

some practical scenarios. For one example, different BS tiers are operated by different network carriers.

For another example, spectrum has already been allocated to different BS tiers, and now some extra

spectrum is available to be further allocated. Note that if the spectrum allocation factors are given

(e.g., [13]), SSAUA gives an optimal or asymptotically optimal user association solution. However, if we

allow 0 ≤ ηk ≤ 1, ∀k, because the average user data rate (5.16) is linear with respect to the spectrum

allocation factors, the spectrum allocation component of SSAUA (5.20) leads to a simple solution, where


only the BS tier with the highest density will be allocated with 100% spectrum, which may not be

desirable in reality. This implies that the user data rate may not be the only design goal in this scenario.

In order to resolve this issue, in one of our follow-up works [76], we choose the average log-utility of

user data rate as our objective. The log-utility can capture a tradeoff between opportunism and user

fairness, by encouraging low-rate cell edge users to improve their rates while saturating the utility gain

of high-rate users. Through this way, all BS tiers will be allocated with non-zero potions of spectrum.

Of course, in future works, it will also be important to design user association and spectrum allocation

schemes to benefit network traffic through decreasing traffic delay, reducing data rate fluctuation, and

improving quality of experience.

Chapter 6

Stochastic Geometric Analysis of

User Mobility in HWNs

Horizontal and vertical handoffs have important ramifications for user mobility in multi-tier HWNs.

They directly impact the signaling overhead and quality of calls. However, they are difficult to analyze

due to the irregularly shaped network topologies introduced by multiple tiers of cells. In this chapter, a

stochastic geometric analysis framework on user mobility is proposed to capture the spatial randomness

and various scales of cell sizes in different tiers.

In Section 6.1, we model different tiers of BSs as Poisson point processes (PPPs) to capture their

spatial randomness. Then, we derive expressions for the rates of all handoff types experienced by an

active user with arbitrary movement trajectory. Furthermore, noting that the data rate of a user depends

on the set of cell tiers that it is willing to use, we provide guidelines for optimal tier selection under various

user velocity, taking both the handoff rates and the data rate into consideration. Empirical studies using

user mobility trace data and extensive simulation are conducted, demonstrating the correctness and

usefulness of our analysis.

In Section 6.2, motivated by the fact that some BSs are likely to aggregate around highly populated

geographical regions (e.g., urban areas, attractions, etc.), we extend the handoff analysis in Section 6.1

by allowing BSs to form Poisson cluster processes (PCPs), in order to capture their non-uniform and

dependent aggregation in space. We then derive analytical expressions for the rates of all handoff types

in this more realistic scenario.

6.1 Handoff Rate Analysis in HWNs with Poisson Patterns

In this section, we aim to derive the rates of all handoff types experienced by an active user with

arbitrary movement trajectory. Furthermore, we study the optimal tier selection under various user

velocity, taking both the handoff rates and the data rate into consideration. Note that we employ the

conventional assumption on HWNs that each BS tier forms a homogeneous PPP in this section.

67

Chapter 6. Stochastic Geometric Analysis of User Mobility in HWNs 68

6.1.1 System Model

In this subsection, we describe the network under consideration, clarifying the notions of cell boundaries

and handoffs.

6.1.1.1 Multi-tier Network

We consider an HWN with spatially randomly distributed K tiers of BSs. Let K = 1, 2, . . . ,K.In order to characterize the random spatial patterns of BSs, in this section, we use the conventional

assumption that each tier of BSs independently form a homogeneous Poisson point process (PPP) in

two-dimensional Euclidean space R2 [12, 13, 35, 37, 38, 73, 74]. Let Φk denote the PPP corresponding to

tier-k BSs, and let λk be its intensity.

6.1.1.2 Biased User Association

Different tiers of BSs transmit at different power levels. Let Pk be the transmission power of tier-k BSs,

which is a given parameter. If Pt(x), for Pt(x) ∈ P1, P2, . . . , PK, is the transmission power from a

BS at x and Pr(y) is the received power at y, we have Pr(y) =Pt(x)hx,y

|x−y|γ , where γ > 2 is the path loss

exponent, |x − y|γ is the propagation loss function, and hx,y is the fast fading term. Corresponding

to common Rayleigh fading with power normalization, hx,y is independently exponentially distributed

with unit mean.

In order to capture various scales of different cell sizes, biased user association considered in Section

5.1.5 is also employed in this chapter. Bk is the association bias of tier-k BSs, and the user association

rule follows formula (5.1). As a consequence, the resultant cell splitting forms a generalized Dirichlet

tessellation, or weighted Poisson Voronoi [106], an example of which is shown in Fig. 1.3. Let T(1)

denote the overall set of cell boundaries, and let T(1)kj denote the set of boundaries between tier-k cells

and tier-j cells, which is also referred to as type k-j cell boundaries in this chapter. Clearly, T(1)kj and

T(1)jk are equivalent.

Note that the effects of B1, B2, . . . , BK remain the same if we multiply all of them by the same positive

constant. For presentation convenience, we define βkj =(

PkBk

PjBj

)1/γ. Clearly, βkj = 1

βjk. Furthermore,

let Ak denote the probability that a UE associates itself with a tier-k BS. We still have

Ak =λk(PkBk)

2γ∑K

j=1 λj(PjBj)2γ

. (6.1)

6.1.1.3 UE Trajectory and Handoff Rate

We aim to study the rates of all handoff types of some active UE moving in the network. Let T0 denote

the trajectory of the UE, which is of finite length. The number of handoffs the UE experiences is equal

to the number of intersections between T0 and T(1), which is denoted by N (T0,T(1)). In this chapter,

a handoff made from a tier-k cell to a tier-j cell is called a type k-j handoff. The number of type k-j

handoffs is denoted by Nkj(T0,T(1)kj ).

If j = k, a type k-j (vertical) handoff is not equivalent to a type j-k handoff. When the UE

crosses some type k-j boundary, either a type k-j or a type j-k handoff is made, depending on its

direction of movement. Thus, the number of type k-j plus type j-k handoffs is equal to the num-

ber of intersections between T0 and T(1)kj , which is denoted by N (T0,T(1)

kj ). In other words, we have


N (T0,T(1)kj ) = Nkj(T0,T(1)

kj ) +Njk(T0,T(1)kj ).

If j = k, N (T0,T(1)kk ) = Nkk(T0,T(1)

kk ) indicates the number of type k-k (horizontal) handoffs.

In Section 6.1.2, we aim to study the rates of all handoff types, which correspond to the expected

numbers of handoffs experienced by an active UE per unit time.

6.1.2 Handoff Rate Analysis in Multi-tier HWNs

The proposed analysis of handoff rates consists of a progressive sequence of four components, which are

described in the following subsections.

6.1.2.1 Length Intensity of Cell Boundaries

Handoffs occur at the intersections between an active UE’s trajectory with cell boundaries. In order to

track the number of intersections, we need to first study the intensities of cell boundaries T(1) and T(1)kj .

Higher intensity of cell boundaries leads to greater opportunities for boundary crossing, and thus higher

handoff rates.

The set of cell boundaries T(1) is a fiber process [39] generated by Φ1,Φ2, . . . ,ΦK . T(1) also corre-

sponds to the set of points on R2, where the same biased power level is received from two nearby BSs,

and this biased received power level is no less than those from any other BSs. Mathematically, we have

T(1) =

x

∣∣∣∣∣∀k, j ∈ K, ∃x1 ∈ Φk,x2 ∈ Φj ,x1 = x2,

s.t. Pr =PkBk

|x1 − x|γ=

PjBj

|x2 − x|γ, and ∀i ∈ K,y ∈ Φi, Pr ≥ PiBi

|y − x|γ

.

Similarly, T(1)kj can be expressed as

T(1)kj =

x

∣∣∣∣∣∃x1 ∈ Φk,x2 ∈ Φj ,x1 = x2,

s.t. Pr =PkBk

|x1 − x|γ=

PjBj


|y − x|γ

.

Note that∪K

k=1

∪Kj=k T

(1)kj = T(1).

Let µ1

(T(1)

)denote the length intensity of T(1), which is the expected length of T(1) in a unit

square1 [39]:

µ1

(T(1)

)= E

(∣∣∣T(1)∩

[0, 1)2∣∣∣1

), (6.2)

where |L|1 denotes the length of L (i.e., one-dimensional Lebesgue measure of L). Similarly, let µ1

(T

(1)kj

)denote the length intensity of T

(1)kj :

µ1

(T

(1)kj

)= E

(∣∣∣T(1)kj

∩[0, 1)2

∣∣∣1

). (6.3)

1Because Φ1, . . . ,ΦK are stationary, T(1) is also stationary, and thus the unit square could be arbitrarily picked on R2.


−4 −3 −2 −1 0 1 2 3−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

2.5

3

3.5

T(1)

T(2)(∆d)∆d

Tier-1 BS

Tier-2 BS

Tier-2 BS

Figure 6.1: The blue curves show T(1); and the region within red dashed curves shows T(2)(∆d).

Note that we have µ1

(T(1)

)=∑K

k=1

∑Kj=k µ1

(T

(1)kj

).

6.1.2.2 ∆d-Extended Cell Boundaries

It is difficult to directly quantify the one-dimensional measures µ1

(T(1)

)and µ1

(T

(1)kj

)on the two-

dimensional plane. Instead, we first introduce ∆d-extended cell boundaries, which extends the one-

dimensional measures to two-dimensional measures.

The ∆d-extended cell boundaries of T(1), denoted by T(2)(∆d), is defined as

T(2)(∆d) =x∣∣∣∃y ∈ T(1), s.t. |x− y| < ∆d

. (6.4)

In other words, T(2)(∆d) is the ∆d-neighbourhood of T(1). A point is in T(2)(∆d) iff its (shortest)

distance to T(1) is less than ∆d, as shown in Fig. 6.1. Similarly, we define T(2)kj (∆d) as the ∆d-extended

cell boundaries of T(1)kj (i.e, ∆d-neighbourhood of T

(1)kj ):

T(2)kj (∆d) =

x∣∣∣∃y ∈ T

(1)kj , s.t. |x− y| < ∆d

. (6.5)

The area intensity of T(2)(∆d) is defined as the expected area of T(2)(∆d) in a unit square:

µ2

(T(2)(∆d)

)= E

(∣∣∣T(2)(∆d)∩

[0, 1)2∣∣∣) , (6.6)

where |S| denotes the area of S (i.e., two-dimensional Lebesgue measure of S). Similarly, the area

intensity of T(2)kj (∆d) is

µ2

(T

(2)kj (∆d)

)= E

(∣∣∣T(2)kj (∆d)

∩[0, 1)2

∣∣∣) . (6.7)

Because Φ1,Φ2, . . . ,ΦK are stationary and isotropic, T(2)(∆d) and T(2)kj (∆d) are also stationary and

isotropic. By definition, the area intensity of T(2)(∆d) (resp. T(2)kj (∆d)) is the average area of T(2)(∆d)

(resp. T(2)kj (∆d)) in a unit square, which is equal to the probability that an arbitrarily located point in

the unit square is in T(2)(∆d) (resp. T(2)kj (∆d)). Thus, we have

µ2

(T(2)(∆d)

)= P(0 ∈ T(2)(∆d)), (6.8)

µ2

(T

(2)kj (∆d)

)= P(0 ∈ T

(2)kj (∆d)). (6.9)


Note that 0 is the origin of the plane.

We observe that the probabilities in (6.8) and (6.9) are analytically tractable, which will be presented

in the next subsection.

6.1.2.3 Derivations of the Area Intensities

In this subsection, we present the derivations of P(0 ∈ T(2)(∆d)) and P(0 ∈ T(2)kj (∆d)). First, we study

the probability that the reference UE at 0 is in T(2)kj (∆d), given that it is associated with a tier-k BS

at a distance of R0 from it. By employing both analytic geometric and stochastic geometric tools, we

derive the following theorem:

Theorem 11. Suppose the reference UE is located at 0; it is associated with a tier-k BS; and their

distance is R. The conditional probability that 0 ∈ T(2)kj (∆d) given R = R0 is

P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)= 1− exp

(−2λj∆dR0F(βkj) +O(∆d2)

), (6.10)

where O represents big O notation and

F(β) , 1

β2

∫ π

0

√(β2 + 1)− 2β cos(θ)dθ. (6.11)

See Appendix C.1 for the proof.

Second, through stochastic geometric tools and deconditioning on R, we can derive the unconditioned

probabilities that the reference UE at 0 is in T(2)(∆d) and in T(2)kj (∆d):

Theorem 12. The area intensities of T(2)(∆d) and T(2)kj (∆d) are

(a)

µ2

(T(2)(∆d)

)=

K∑k=1

λk

(∑Ki=1 λi∆dF(βki)

)(∑K

i=1 λiβ2ik

) 32

+O(∆d2). (6.12)

(b)

µ2

(T

(2)kj (∆d)

)=

λk(λj∆dF(βkj))

(∑K

i=1 λiβ2ik)

32

+λj(λk∆dF(βjk))

(∑K

i=1 λiβ2ij)

32

+O(∆d2) if k = j,

λ2k∆dF(1)

(∑K

i=1 λiβ2ik)

32+O(∆d2) if k = j.

(6.13)


6.1.2.4 From Area Intensities to Handoff Rates

In this subsection, we derive handoff rates from area intensities derived in Theorem 12. This involves

two steps: (1) from area intensities µ2

(T(2)(∆d)

)and µ2

(T

(2)kj (∆d)

)to length intensities µ1

(T(1)

)and

µ1

(T

(1)kj

), and (2) from length intensities to handoff rates.

First, we derive the length intensity µ1

(T(1)

)(resp. µ1

(T

(1)kj

)) from the area intensity µ2

(T(2)(∆d)

)(resp. µ2

(T

(2)kj (∆d)

)) by taking ∆d → 0. Following [113] and Section 3.2 in [114], µ1

(T(1)

)=

lim∆d→0µ2

(T(2)(∆d)

)2∆d and µ1

(T

(1)kj

)= lim∆d→0

µ2

(T

(2)kj (∆d)

)2∆d . Therefore, we have


Theorem 13. The length intensities of T(1) and T(1)kj can be computed as follows:

(a)

µ1

(T(1)

)=

K∑k=1

λk

(∑Ki=1 λiF(βki)

)2(∑K

i=1 λiβ2ik

) 32

. (6.14)

(b)

µ1

(T

(1)kj

)=

λkλjF(βkj)

2(∑K

i=1 λiβ2ik)

32+

λjλkF(βjk)

2(∑K

i=1 λiβ2ij)

32

if k = j,

λ2kF(1)

2(∑K

i=1 λiβ2ik)

32

if k = j.(6.15)

Remark 1. Note that, if we consider the single-tier case by taking K = 1, we have F(1) = 4, and

µ1

(T(1)

)= µ1

(T

(1)11

)= 2

√λ1. This matches the length intensity of a standard Poisson Voronoi. See

Section 10.6 of [39].

Second, we can derive the expected number of handoffs of an active UE. Since T(1) and T(1)kj are

stationary and isotropic fiber processes with length intensity µ1

(T(1)

)and µ1

(T

(1)kj

), respectively, we can

derive the expected number of intersections between T0 and T(1) (resp. T(1)kj ) following the conclusion

in Section 9.3 of [39]. We have

Theorem 14. Let T0 denote an arbitrary UE’s trajectory on R2 with length |T0|1. Then, the expected

number of intersections between T0 and T(1) (resp. T(1)kj ) are

E(N (T0,T(1))

)=2

πµ1

(T(1)

)|T0|1, (6.16)

E(N (T0,T(1)

kj ))=2

πµ1

(T

(1)kj

)|T0|1, (6.17)

and the expected number of type k-j handoffs are

E(Nkj(T0,T(1)

kj ))=

12E(N (T0,T(1)

kj ))

if k = j,

E(N (T0,T(1)

kj ))

if k = j.(6.18)

Note that the expected number of type k-j handoffs is the same as the expected number of type j-k

handoffs, both of which are equal to half of E(N (T0,T(1)

kj )).

Let v denote the instantaneous velocity of an active UE, H(v) denote its overall handoff rate (i.e.,

sum handoff rate of all types), and Hkj(v) denote its type k-j handoff rate. Then we have the following

Corollary from Theorem 14:

Corollary 2.

H(v) =2

πµ1

(T(1)

)v, (6.19)

Hkj(v) =

1πµ1

(T

(1)kj

)v if k = j,

2πµ1

(T

(1)kj

)v if k = j.

(6.20)


Note that the above handoff rates are instantaneous rates. Thus, our analysis allows time-varying

velocity for the UEs, in which case the handoff rates are also time varying.

6.1.3 UE’s Data Rate and Tier Selection

In this subsection, we discuss optimal BS tier selection as an application of the above analysis, taking

both the handoff rates and the user data rate into consideration. Let S ⊂ K be the set of BS tiers that

an active UE chooses to use. We assume that 1 ∈ S (i.e., the UE always selects tier-1 macrocells).

6.1.3.1 UE Data Rate

6.1.3.1.1 Spectrum Allocation and Multiple Modulation and Coding Schemes In multi-

tier HWNs, in order to avoid cross-tier interference, and the prohibitive complexity in tracking and

provisioning for such interference especially with unplanned deployment of small cells, a disjoint spec-

trum mode is advocated (e.g., [36, 73, 115]), where different tiers of BSs are allocated non-overlapping

portions of the spectrum. In what follows, we consider this scenario as an illustrative example. However,

we emphasize that, since the proposed handoff rate analysis above is independent of the spectrum allo-

cation strategy, the subsequent study on optimal tier selection is applicable to more general scenarios of

spectrum sharing, as long as the average UE data rate can be derived (e.g., [12, 13,35]).

We assume that each active UE associated with a tier-k BS is allocated the same spectrum with

bandwidth Wk. The UE adaptively selects one of N available modulation and coding schemes (MCSs).

Each MCS corresponds to a specific Signal-to-Interference Ratio (SIR) Tn, for 1 ≤ n ≤ N . Without loss

of generality, we set T1 < T2 < . . . < TN . If a UE is allocated spectrum bandwidth W ′, its data rate

is W ′ log(1 + TN ), W ′ log(1 + TN−1), . . ., W′ log(1 + T1), and 0, respectively, if its SIR is in [TN ,∞),

[TN−1, TN ), . . ., [T1, T2), and [0, T1). Note that log is in base 2 throughout this chapter. Also, we

consider the common scenario where the system is interference limited, such that noise is negligible.

6.1.3.1.2 Average UE Data Rate Derivation An approach similar to one proposed in Chapter

5 is used to derive a closed-form expression for the average UE data rate.

Following (6.1), the probability that the active UE associates itself with a tier-k BS, for k ∈ S, is

Ak,S =λk(PkBk)

2γ∑

j∈S λj(PjBj)2γ

. (6.21)

Given that the UE is associated with a tier-k BS and their distance is d, the overall interference to

the UE is the sum interference from all tier-k BSs other than the BS serving the UE:

Ik,S(d) =∑x∈Φ′

k

Pkhx,0

|x|γ, (6.22)

where Φ′k is the reduced Palm point process [12] corresponding to all tier-k BSs other than the BS serving

the UE. It can be shown that Φ′k is a PPP with intensity 0 in B(0, d) and intensity λk in R2\B(0, d),

where B(x, r) denotes the disk region centered at x with radius r [37].


The distribution of Ik,S(d) is derived through its Laplace transform as follows:

LIk,S (d, s) = E

exp−

∑x∈Φ′

k

sPkhx,0

|x|γ

=exp

(−2πλk

∫ ∞

d

sPkr

sPk + rγdr

). (6.23)

Let Pcov,k,S(d, Tn) denote the conditional probability that the SIR of the active UE is larger than Tn

(given k and d). Then,

Pcov,k,S(d, Tn) =P(PkhxB ,0

dγ≥ TnIk,S(d)

)=LIk,S (d, s)|s=Tndγ

Pk

, (6.24)

where xB is the coordinate of the BS serving the UE, and |xB | = d. We have the second equality in

(6.24) because hxB ,0 is exponentially distributed. Substituting (6.23) into (6.24), we have

Pcov,k,S(d, Tn) = exp

(−2πλk

∫ ∞

d

Tndγr

Tndγ + rγdr

)t= r2

d2·T2/γn= exp

(−πλkT

2γn d2

∫ ∞

( 1Tn

)2γ

1

1 + tγ2

dt

). (6.25)

Note that by the change of variable in the second equality above, we are able to capture the fact that

the UE’s data rate is higher if it is closer to its serving BS.

Next, the probability density function of the distance between the UE and its serving BS is

fk,S(d) =2πλk

Ak,Sd exp

−πd2∑j∈S

λj

(PjBj

PkBk

) 2γ

(6.26)

=2πλk

Ak,Sd exp

(−πd2

λk

Ak,S

), (6.27)

where (6.26) is derived in [12], and (6.27) is obtained by substituting (6.21) into (6.26).

Let Pcov,k,S(Tn) denote the probability that the SIR of the active UE is larger than Tn (given that

it is associated with a tier-k BS). It can be computed as

Pcov,k,S(Tn) =

∫ ∞

0

fk,S(d)Pcov,k,S(d, Tn)dd

=1

1 +Ak,SCn, (6.28)

where Cn , (Tn)2γ∫∞( 1

Tn)

2γ

11+tγ/2 dt. Note that because T1 < T2 < . . . < TN , we have C1 < C2 < . . . <

CN .

Consequently, by using multiple MCSs, the average data rate of the UE (given that it is associated


with a tier-k BS) can be computed as

Rk,S =Wk

[log(1 + TN )Pcov,k,S(TN ) +

N−1∑n=1

log(1 + Tn)(Pcov,k,S(Tn)− Pcov,k,S(Tn+1)

)](6.29)

=Wk

[log(1 + TN )

1 +Ak,SCN+

N−1∑n=1

log(1 + Tn)

(1

1 +Ak,SCn− 1

1 +Ak,SCn+1

)]. (6.30)

Finally, the expected data rate of the active UE can be computed as

RS =∑k∈S

Ak,SRk,S . (6.31)

6.1.3.2 Optimal Tier Selection

Following the derivations in Section 6.1.2 and 6.1.3.1, we see that different tier selections lead to different

data rates and handoff rates. Let Ckj be the cost for one type k-j handoff, and UR be the utility value

for one bit of data transmission. We assume the UE also pays service charge Pk per second when it is

associated with a tier-k BS. Note that Ckj , UR, and Pk may be assigned arbitrarily, and Ckj and Cjkmay be different if k = j [14]. If the UE favors higher data rate, it could assign a larger value for UR; if

it favors lower handoff rates, it could assign larger values for Ckj .If the active UE’s tier selection is S, its overall average utility on data transmission per second is

URRS , overall average service charge per second is P(S) =∑

k∈S Ak,SPk, and overall average expense

on handoffs per second is

C(S, v) = 2

πv∑k∈S

λk

(∑i∈S

(Cki+Cik)2 λiF(βki)

)2(∑

i∈S λiβ2ik

) 32

, (6.32)

where (6.32) follows the conclusions of Theorem 13 and Corollary 2. Consequently, the overall average

utility per second of tier selection S is

G(S, v) = URRS − C(S, v)− P(S). (6.33)

Finally, the optimal tier selection is

Sopt = argmaxS∈S

(URRS − C(S, v)−P(S)

), (6.34)

where S is the set of all possible tier selections. Sopt can be determined through comparing all possible

tier selections.

6.1.4 Experimental Study

In this subsection, our analysis is validated via experimenting with real-world traces and simulations.

6.1.4.1 Yonsei Trace Data

We use the Yonsei Trace [116] to validate our analytical results. The trace was accumulated from 12

commercial mobile phones during an eight-month period in 2011 in the city of Seoul. An application


named SmartDC was running on the commercial mobile phones equipped with GPS, GSM, and WiFi.

For every 2 to 5 minutes, the application collected UE’s location information (latitude and longitude),

the MAC addresses of surrounding WiFi APs, and the cell IDs of nearby cellular BSs they could detect.

Each AP has a unique MAC address and each BS has a unique cell ID. By analyzing the data set, we are

able to determine which APs and BSs a UE could detect at the recorded coordinates and time instants.

In the following, we regard cellular BSs as tier-1 BSs and APs as tier-2 BSs.

6.1.4.2 Data Processing

6.1.4.2.1 Location Approximations of APs and BSs As the data set does not explicitly provide

the latitudes and longitudes of APs and BSs, we apply the following approach to approximate their

locations: for each AP (resp. BS), we list all the coordinates recorded by UEs when they are able to

detect the AP (resp. BS). Then, we approximate the location of the AP (resp. BS), by taking the

average of these recorded coordinates.

6.1.4.2.2 Reference Region In order to avoid the edge effect, we define a reference region, in which

most recorded coordinates are located. The UEs’ trajectories are only accounted inside the reference

region. By plotting the cumulative distribution function (cdf) of the latitude (resp. longitude) of

all recorded coordinates (shown in Fig. 6.2), we observe a sharp step upward between 37.48N and

37.58N (resp.126.9E and 127.1E). As a consequence, we employ the rectangle defined by 37.48N ,

and 37.58N , 126.9E, and 127.1E as the reference region.

6.1.4.2.3 UE Trajectory In the trace data, the coordinates of a UE are recorded only once every

few minutes. To recover its full trajectory, we regard it as moving in a straight line at a constant

velocity between two consecutive recorded coordinates. Thus, interpolation can be made to determine

the coordinate of the UE at any time. Note that only the trajectories inside the reference region are

used.

6.1.4.2.4 Handoff Rates Through the locations of BSs and APs, as well as the UE trajectories, we

are able to derive the empirical rates of all handoff types following the biased user association scheme

discussed in Section 6.1.1.2. We set P1 = 45 dBm, P2 = 20 dBm, B1 = 100B2, and γ = 3. If we ignore

all the APs, we can also derive the empirical handoff rates for the one-tier case.

6.1.4.2.5 BS and AP Intensities The AP (resp. BS) density is computed as the number of APs

(resp. BSs) over the area of the reference region, which is 455.1 unit/km2 (resp. 52.6 unit/km2). This

indicates an urban area with high population and BS densities.

6.1.4.2.6 Empirical Results We compare the handoff rates derived from our analysis and those

from our empirical study based on the Yonsei Trace. The empirical handoff rates are derived from the

steps in Sections 6.1.4.2.1 - 6.1.4.2.4. For the analytical results, we use the BS and AP intensities shown

in Section 6.1.4.2.5 as input parameters.

For the two-tier case (considering both APs and BSs), the comparison of analytical and empirical

handoff rates is shown in Fig. 6.3. For the one-tier case (by eliminating all the APs), the comparison is

shown in Fig. 6.4. Both figures illustrate the accuracy of our analysis. When the UE’s velocity is low,


36.5 37 37.5 38 38.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Latitude (N)

cdf

126 126.5 127 127.5 128 128.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Longitude (E)

cdf

Figure 6.2: Cumulative distribution function of the latitude and longitude.

0 20 40 60 80 100 120 140 160 180 2000

0.5

1

1.5

2

2.5

3

3.5

4

Velocity (m/minute)

Han

doff

rate

(unit/m

inute

)

1-1, analytical1-1, empirical1-2 (or 2-1), analytical1-2, empirical2-1, empirical2-2, analytical2-2, empirical

Figure 6.3: Two-tier case: comparison of analytical and empirical handoff rates.

0 20 40 60 80 100 120 140 160 180 2000

0.5

1

1.5

2

2.5

3

3.5

4

Velocity (m/minute)

Han

doff

rate

(unit/m

inute

)

analytical

empirical

Figure 6.4: One-tier case: comparison of analytical and empirical handoff rates.

empirical handoff rates are slightly greater than analytical handoff rates. This is because the locations

of APs and BSs are not strictly homogeneously distributed (e.g., some APs and BSs are crowding along

some streets, or at the center of the urban region). We also observe that UEs with lower velocity are

more likely to be sampled in regions with higher AP and BS densities. As a consequence, the empirical

handoff rates are higher than those computed by our analytical results.

Fig. 6.3 and Fig. 6.4 also show that type 1-1 horizontal handoff rates are almost the same in the

one-tier and two-tier cases, but extra type 1-2 and type 2-1 vertical handoffs are introduced in the

two-tier case. This agrees with our expectation that adding a second tier of APs brings more vertical

handoffs. In addition, as a validation of (6.18), type 1-2 and type 2-1 handoff rates are almost the same

in empirical results.


1 2 3 4 5 6 7 8 9 100

0.5

1

1.5

2

2.5

3

3.5

4

λ1 (unit/km2)

Han

doff

rate

(unit/m

inute

)

1-1, analytical

1-1, simulation

1-2 & 2-1, analytical

1-2 & 2-1, simulation

2-2, analytical

2-2, simulation

Figure 6.5: Two-tier case: handoff rates under different λ1 values.

1 2 3 4 5 6 7 8 9 100

0.5

1

1.5

2

λ2 (unit/km2)

Han

doff

rate

(unit/m

inute

)

1-1, analytical1-1, simulation1-2 & 2-1, analytical1-2 & 2-1, simulation1-3 & 3-1, analytical1-3 & 3-1, simulation2-2, analytical2-2, simulation2-3 & 3-2, analytical2-3 & 3-2, simulation3-3, analytical3-3, simulation

Figure 6.6: Three-tier case: handoff rates under different λ2 values.

6.1.4.3 Simulation Study

In this subsection, we present simulation results to further demonstrate our analysis in more complex

HWNs.

6.1.4.3.1 Simulation Setup The simulation procedure is as follows: in each round of simulation,

two or three tiers of BSs are generated on a 10 km × 10 km square. Then, we randomly generate 5

waypoints X1, . . . ,X5 in the central 5 km × 5 km square (uniformly distributed). The four line segments

X1X2,X2X3, . . . ,X4X5 form the trajectory of an active UE. In this way, we derive the simulated handoff

rates in one round of simulation. The above procedure is repeated 2000 rounds to derive one simulated

data point. Note that in this subsection, in order to avoid overlapping in figures, we only show the sum

rate of type j-k and type k-j (k = j) handoffs for easier inspection; the individual handoff rates are half

of the sum handoff rate.

6.1.4.3.2 Handoff Rates under Different BS Intensities We study handoff rates under different

BS intensities. Fig. 6.5 shows a two-tier case, with parameters as follows: (P1, P2) = (30, 20) dBm,

(B1, B2) = (1, 1), and λ2 = 1 unit/km2. Fig. 6.6 shows a three-tier case, with parameters as follows:

(P1, P2, P3) = (30, 20, 10) dBm, (B1, B2, B3) = (1, 1, 1), and (λ1, λ3) = (1, 1) unit/km2. The parameter

values γ = 3 and v = 60 km/h are used for both Fig. 6.5 and Fig. 6.6.

Fig. 6.5 illustrates that increasing λ1 leads to higher type 1-1 handoff rate but lower type 2-2 handoff

rate. Fig. 6.6 illustrates that increasing λ2 leads to higher type 2-2 handoff rate but lower type 1-1,

1-3, and 3-1 handoff rates. Both observations suggest that increasing the BS intensity of one tier causes

higher horizontal handoff rate within the tier, but lower handoff rates outside the tier.

6.1.4.3.3 Handoff Rates under Different Association Bias Values Next, we study handoff

rates under different association bias values. Fig. 6.7 shows a two-tier case, with parameters as follows:


101

102

103

104

105

0

0.2

0.4

0.6

0.8

1

1.2

1.4

B1/B2

Handoff

rate

(unit/m

inute

)

1-1, analytical

1-1, simulation

1-2 & 2-1, analytical

1-2 & 2-1, simulation

2-2, analytical

2-2, simulation

Figure 6.7: Two-tier case: handoff rates under different B1 values.

101

102

103

104

105

0

0.2

0.4

0.6

0.8

1

1.2

1.4

B2

Handoff

rate

(unit/m

inute

)

1-1, analytical1-1, simulation1-2 & 2-1, analytical1-2 & 2-1, simulation1-3 & 3-1, analytical1-3 & 3-1, simulation2-2, analytical2-2, simulation2-3 & 3-2, analytical2-3 & 3-2, simulation3-3, analytical3-3, simulation

Figure 6.8: Three-tier case: handoff rates under different B2 values.

5 10 15 20 25 30 35 40 45 500

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

velocity (km/h)

unit

utility

/sec

ond

Utility of tier selection 1,2, simulationUtility of tier selection 1, simulationAnalytical tier selection threshold

Figure 6.9: Two-tier case: overall utility comparison of different tier selections.

(P1, P2) = (30, 20) dBm, B2 = 1, and (λ1, λ2) = (1, 1) unit/km2. Fig. 6.8 shows a three-tier case, with

parameters as follows: (P1, P2, P3) = (30, 20, 10) dBm, (B1, B3) = (1, 1), and (λ1, λ2, λ3) = (1, 1, 1)

unit/km2. The parameter values γ = 3 and v = 60 km/h are used for both Fig. 6.7 and Fig. 6.8. These

figures suggest that, increasing the association bias value of one tier has a similar effect as increasing

the BS intensity of this tier, leading to higher horizontal handoff rate within the tier, but lower handoff

rates outside the tier.

6.1.4.3.4 UE’s Utility under Different Tier Selections Fig. 6.9 shows the simulated utility of

an active UE under different velocity values in a two-tier case. The parameters are as follows: (P1, P2) =

(40, 20) dBm, (B1, B2) = (1, 1), (λ1, λ2) = (1, 10) units/km2, γ = 3, N = 3, T1 = 0.2, T2 = 0.4, T3 = 0.8,

(W1,W2) = (2, 5) MHz, (C11, C12, C21, C22) = (10, 45, 35, 20), UR = 1, and (P1,P2) = (0.1, 0.1)2. The

vertical line shows the analytical tier selection threshold on the UE velocity, which is 28.82 km/h. This

figure demonstrates that the simulation results agree with the analysis of tier selection as discussed in

Section 6.1.3. Tier selection 1, 2 is optimal if the UE’s velocity is low, but its overall utility decreases

faster due to the higher handoff expense. Tier selection 1 is optimal if the UE’s velocity is greater

2In this section, Ckj is in unit utility/handoff, UR is in unit utility/Mbit, and Pk is in unit utility/s.


5 10 15 20 25 30 35 40 45 50 55 600

0.5

1

1.5

2

2.5

velocity (km/h)

unit

utility

/sec

ond

Utility of tier selection 1,2,3, simulationUtility of tier selection 1,3, simulationUtility of tier selection 1,2, simulationUtility of tier selection 1, simulationAnalytical threshold, 1,2,3 and 1,2Analytical threshold, 1,2 and 1

Figure 6.10: Three-tier case: overall utility comparison of different tier selections.

than 28.82 km/h.

Fig. 6.10 shows the simulated utility of an active UE under different velocity values in a three-

tier case. The parameters are as follows: (P1, P2, P3) = (40, 20, 10) dBm, (B1, B2, B3) = (1, 1, 1),

(λ1, λ2, λ3) = (1, 5, 20) units/km2, γ = 3, N = 3, T1 = 0.2, T2 = 0.4, T3 = 0.8, (W1,W2,W3) = (2, 5, 10)

MHz, (C11, C12, C21, C22, C13, C31, C23, C32, C33) = (10, 35, 25, 20, 45, 35, 55, 45, 30), UR = 1, and (P1,P2,P2)

= (0.1, 0.1, 0.1). Two vertical lines show the analytical tier selection thresholds, which are 27.15 km/h

and 42.15 km/h, respectively. This figure again shows that the simulation results agree with the analysis

of tier selection. When the velocity is in the range [0, 27.15) km/h, tier selection 1, 2, 3 is optimal;

when the velocity is in the range [27.15, 42.15) km/h, tier selection 1, 2 is optimal; when the velocity

is in the range [42.15,∞) km/h, tier selection 1 is optimal.

6.1.4.3.5 Velocity Threshold Fig. 6.11 and Fig. 6.12 show the computed velocity thresholds for

tier selection, under different BS densities. Fig. 6.11 shows a two-tier case. The parameters are as

follows: (P1, P2) = (40, 20) dBm, λ1 = 1 units/km2, γ = 3, N = 3, T1 = 0.2, T2 = 0.4, T3 = 0.8,

(W1,W2) = (3, 7) MHz, (C11, C12, C21, C22) = (10, 45, 35, 20), and (P1,P2) = (0.1, 0.1). Fig. 6.12 shows a

three-tier case. The parameters are as follows: (P1, P2, P3) = (30, 20, 10) dBm, (B1, B2, B3) = (1, 1, 1),

(λ1, λ2) = (1, 5) units/km2, γ = 3, N = 3, T1 = 0.2, T2 = 0.4, T3 = 0.8, (W1,W2,W3) = (2, 5, 10) MHz,

(C11, C12, C21, C22, C13, C31, C23, C32, C33) = (10, 35, 25, 20, 45, 35, 55, 45, 30), and (P1,P2,P2) = (0.1, 0.1, 0.1).

In the two-tier case, increasing λ2 or B2 improves the average UE data rate (as the UE has higher

probability to be associated with tier-2 BSs), but it could also cause higher handoff rates. Through

our theoretical analysis, we could observe that the latter factor has a stronger effect and the velocity

threshold value is lowered if λ2 increases; while the former factor dominates and the velocity threshold

value increases if B2 becomes greater. In addition, increasing UR leads to a higher weight in data rate,

so the threshold values increase.

In the three-tier case, increasing λ3 could have more complicated impact on optimal tier-selection.

We observe that the tier selection 1, 3 is broken into two separate regions. When λ3 is small, the

velocity range of tier selection 1, 2 is below the velocity range of tier selection 1, 3, while the former

range is above the latter one when λ3 becomes larger. Still, increasing UR leads to a higher weight in

data rate, thus the threshold values are increased (i.e., the solid curves are shifted upward compared

with the dashed curves).


1 2 3 4 5 6 7 8 9 10

35

40

45

50

55

60

65

λ2 (unit/km2)Sel

ection

thre

shold

(km

/h)

B1/B2 = 2, UR = 1B1/B2 = 2, UR = 1.05B1/B2 = 1, UR = 1B1/B2 = 1, UR = 1.05B1/B2 = 0.5, UR = 1B1/B2 = 0.5, UR = 1.05

Figure 6.11: Two-tier case: tier selection velocity threshold.

2 4 6 8 10 12 14 16 18 200

10

20

30

40

50

60

70

80

90

λ3 (unit/km2)

Sel

ection

thre

shol

d(k

m/h

)

1,3

1,3

1,2

1

UR = 1 UR = 1.05

1,2,3

1,2 performs best

1 performs best

1,2,3 performs best

1,3 performs best

Figure 6.12: Three-tier case. The set of dashed and solid curves show the velocity threshold valuesfor tier selection, under UR = 1 and UR = 1.05, respectively. Each set of curves separates the planeinto multiple regions, and each region corresponds to a specific optimal tier selection scheme (labeledin the region). For example, a vertical line at λ3 = 16 has 3 intersections with the set of solid curves,illustrating that under λ3 = 16 and UR = 1.05, tier selections 1, 2, 3, 1, 3, 1, 2, and 1 are optimalin four different velocity ranges separated by the 3 intersections.

6.2 Handoff Rate Analysis in HWNs with Poisson and Poisson

Cluster Patterns

In Section 6.1, we employed the conventional assumption on HWNs that each BS tier forms a homoge-

neous PPP. However, the PPP model cannot capture the non-uniform and dependent aggregation of BSs

in, for example, popular regions of the network where more BSs tend to be installed, as the positions

of the points in PPP are independent of each other. In this section, as an extension of Section 6.1, we

accommodate the non-uniform and dependent aggregation by modeling some BS tiers as Poisson cluster

processes (PCPs) [39]. Each BS cluster includes multiple nearby BSs, and multiple BS clusters are ran-

domly distributed in space. In this case, the resultant cell splitting is a tessellation generated by both

PPPs and PCPs. Through our improved stochastic and analytic geometric analysis, we derive exact

expressions for the rates of all handoff types experienced by an active user with arbitrary movement

trajectory.

6.2.1 System Model

We consider an HWN with randomly distributed K tiers of BSs. It includes KP Poisson tiers (P-tiers)

and KC Poisson cluster tiers (C-tiers), so that K = KP +KC . Let KP = 1, . . . ,KP denote the set of

P-tiers, and KC = KP + 1, . . . ,KP +KC denote the set of C-tiers. Let K = 1, . . . ,KP +KC.


-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

Figure 6.13: An example of a two-tier HWN with Poisson and Poisson cluster patterns. Tier-1 and 2BSs are represented by “” and “”, respectively. Tier-2 BSs are clustered in four disk regions. Blue,red, and green curves show cell boundaries within tier-1, between tier-1 and tier-2, and within tier-2,respectively. The magenta arrow represents the trajectory of an active UE. The UE makes two handoffsat the intersections between its trajectory and the set of cell boundary.

Let Φk denote the point process representing the locations of tier-k BSs. If k ∈ KP , Φk is a

homogeneous PPP with intensity of λk on the two-dimensional space R2. If k ∈ KC , Φk is a PCP defined

as follows: First, cluster centers are generated as a parent point process Θk, which is a homogeneous

PPP with intensity µk on the two-dimensional space R2. Second, for each parent point (cluster center) x

in Θk, a cluster of BSs form a child point process Ωk(x), which is a PPP with intensity νk in the region

B(x, Rk), where B(x, r) is defined as the disk region centered at x with radius r. The overall PCP Φk

is the union of all child point processes, i.e., Φk =∪

x∈ΘkΩk(x). Note that Φk does not include parent

points Θk. Given the location of some BS xBS in a tier-k PCP, let C(xBS) denote the cluster center of

the BS. Thus Ωk(C(xBS)) is the cluster it belongs to.

In this section, the assumptions on the biased user association, UE trajectory, and handoffs are

exactly the same with those in Sections 6.1.1.2 and 6.1.1.3. The definitions of Pk, Bk, βkj , T0, T(1),

T(1)kj , N (T0,T(1)), Nkj(T0,T(1)

kj ) are exactly the same as those defined in Sections 6.1.1.2 and 6.1.1.3.

Note that since a BS tier is modeled as a PPP or a PCP, the resultant cell splitting is a tessellation

generated by both PPPs and PCPs. An example cell splitting is shown in Fig. 6.13 (a two-tier scenario

including one P-tier and one C-tier).

6.2.2 Handoff Rate Analysis

Following Section 6.1.2, the proposed analysis of handoff rates consists of a progressive sequence of four

components, which are described in the following subsections.

6.2.2.1 Length Intensity of Cell Boundaries

Following Section 6.1.2.1, we first study the length intensity of different types of cell boundaries T(1)kj ,

which is defined as the expected length of T(1)kj in a unit square.

The set of cell boundaries T(1)kj is generated by all K tiers of BSs Φ1,Φ2, . . . ,ΦK . It corresponds to

the set of points on R2 where the same biased power level is received from a tier-k BS and a tier-j BS,

and this biased received power level is greater than those from any other BS in all tiers. Mathematically,


T(1)kj is expressed as

T(1)kj =

x

∣∣∣∣∣∃x1 ∈ Φk,x2 ∈ Φj ,x1 = x2,

s.t. Pr =PkBk

|x1 − x|γ=

PjBj


|y − x|γ

. (6.35)

A main challenge in this section is in characterizing handoff rates across the highly irregular cell

boundaries generated by BSs that form PCPs. If we consider the cell boundaries within tier-k, and

tier-k is a C-tier (i.e., k ∈ KC), then we can further classify the set of cell boundaries T(1)kk into the set

of intra-cluster cell boundaries (i.e., the two BSs that provide the largest biased received power belong

to the same cluster), which is denoted by T(1)kk,in, and the set of inter-cluster cell boundaries (i.e., the

two BSs belong to two different clusters), which is denoted by T(1)kk,out. We formally express

T(1)kk,in =

x

∣∣∣∣∣∃x1 ∈ Φk,x2 ∈ Φk,x1 = x2, C(x1) = C(x2), (6.36)

s.t. Pr =PkBk

|x1 − x|γ=

PjBj


|y − x|γ

.

Similarly, by replacing C(x1) = C(x2) by C(x1) = C(x2) in (6.36), we have the expression of T(1)kk,out.

Note that we have T(1)kk = T

(1)kk,in

∪T

(1)kk,out. Also, if k, j ∈ KC and k = j, we cannot classify T

(1)kj in

the same way, as the cell boundary is surely formed by two BSs in two different clusters in two different

tiers.

Let µ1

(T

(1)kj

)denote the length intensity of T

(1)kj , which is the expected length of T

(1)kj in a unit square.

Because Φ1, . . . ,ΦK are stationary, T(1)kj is also stationary, and thus the unit square could be arbitrarily

picked on R2. Hence, we have

µ1

(T

(1)kj

)= E

(∣∣∣T(1)kj

∩[0, 1)2

∣∣∣1

), (6.37)

where |L|1 denotes the length of a collection of curves L (i.e., one-dimensional Lebesgue measure of L).

Similarly, if k ∈ KC ,

µ1

(T

(1)kk,x

)= E

(∣∣∣T(1)kk,x

∩[0, 1)2

∣∣∣1

), (6.38)

where the subscript “x” indicates either “in” or “out”. Note that µ1

(T

(1)kk

)= µ1

(T

(1)kk,in

)+µ1

(T

(1)kk,out

).

6.2.2.2 ∆d-Extended Cell Boundaries

Following Section 6.1.2.2, we study the ∆d-neighbourhoods of T(1)kj , denoted by T

(2)kj (∆d).

T(2)kj (∆d) =

x∣∣∣∃y ∈ T

(1)kj , s.t. |x− y| < ∆d

. (6.39)

Similarly, T(2)kk,in(∆d) and T

(2)kk,out(∆d) are defined as the ∆d-neighbourhoods of T

(1)kk,in and T

(1)kk,out,

respectively.


The area intensity of T(2)kj (∆d) is defined as the expected area of T

(2)kj (∆d) in a unit square:

µ2

(T

(2)kj (∆d)

)= E

(∣∣∣T(2)kj (∆d)

∩[0, 1)2

∣∣∣) , (6.40)

where |S| denotes the area of some region S (i.e., two-dimensional Lebesgue measure of S).

The area intensity of T(2)kj (∆d) is equal to the probability that the reference UE at 0 is in T

(2)kj (∆d).

µ2

(T

(2)kj (∆d)

)= P(0 ∈ T

(2)kj (∆d)). (6.41)

Similarly, if k ∈ KC ,

µ2

(T

(2)kk,x(∆d)

)= P(0 ∈ T

(2)kk,x(∆d)), (6.42)

where the subscript “x” indicates either “in” or “out”.

6.2.2.3 Derivation of Area Intensities

In this subsection, we present the derivation of P(0 ∈ T(2)kj (∆d)). It consists of a progressive sequence of

four steps. In the first step, we recall a few important properties of two intersecting circles, which will be

frequently used in the subsequent steps. In the second step, we study the distribution of the distance from

the reference UE at 0 to the BS it is associated with, which is referred to as the reference BS throughout

the rest of this section. In the third step, we study the conditional probability of 0 ∈ T(2)kj (∆d) given

the distance from the reference UE to the reference BS. Based on step two and step three, we derive the

unconditioned probability of 0 ∈ T(2)kj (∆d) in the fourth step.

6.2.2.3.1 Geometric Patterns of Two Intersecting Circles We first recall a few important

properties of two intersecting circles (as shown in Fig. 6.14), which will be frequently used in the subse-

quent steps to derive P(0 ∈ T(2)kj (∆d)). Let r1 and r2 be the radii of these circles, and r be the distance

between their centers. We assume |r1 − r2| ≤ r ≤ r1 + r2.

As labeled in Fig. 6.14, let C(r1, r2, r) denote the area of the overlapping part of the two circles,

L(r1, r2, r) denote the arc length of circle 1 covered by circle 2, and θm(r1, r2, r) denote the half central

angle corresponding to the arc. The expressions of C(·), L(·), and θm(·) are all in closed form, which

are shown as follows:

C(r1, r2, r) = arccos((r21 + r2 − r22)/(2r1r)

)· r21 + arccos

((r22 + r2 − r21)/(2r2r)

)· r22

− 2√((r1 + r2 + r)/2) ((r1 + r2 + r)/2− r1) ·

√((r1 + r2 + r)/2− r2) ((r1 + r2 + r)/2− r),

(6.43)

L(r1, r2, r) = 2r1 arccos((r21 + r2 − r22)/(2r1r)

), (6.44)

θm(r1, r2, r) = arccos((r21 + r2 − r22)/(2r1r)

). (6.45)

Note that θm(r1, r2, r) = θm(r, r2, r1).


r1

r2

rC(r1, r2, r)

L(r1, r2, r)m(r1, r2, r)

Figure 6.14: Geometric patterns of two intersecting circles.

6.2.2.3.2 Distance Distribution of Reference UE-BS Pair Let Rk denote the distance from

0 to the nearest tier-k BS. If k ∈ KP , i.e., Φk is a PPP, by the Markovian property of PPPs, it is

straightforward to derive the complementary cumulative distribution function (ccdf) and probability

density function (pdf) of Rk:

ccdfRk(R0) =P(Rk > R0) = exp

(−πR2

0λk

), (6.46)

pdfRk(R0) =2πR0λk

(−πR2

0λk

). (6.47)

If k ∈ KC , the ccdf of Rk is more complex due to the dependent and non-uniform aggregation of

points in PCP. It can be shown in Appendix C.3 that

ccdfRk(R0) , P(Rk > R0) = (6.48) exp[− π(R0 −Rk)

2µk

(1− e−πR2

kνk

)−∫ R0+Rk

R0−Rk2πrµk

(1− e−C(R0,Rk,r)νk

)dr], if R0 ≥ Rk,

exp[− π(R0 −Rk)

2µk

(1− e−πR2

0νk

)−∫ Rk+R0

Rk−R02πrµk


)dr], if R0 < Rk.

By taking the first derivative, we find the pdf of Rk as

pdfRk(R0) = (6.49)


2µk

(1− e−πR2

kνk

)−∫ R0+Rk

R0−Rk2πrµk


)dr]

·µk

∫ R0+Rk

R0−Rk2πre−νkC(R0,Rk,r)νkL(R0, Rk, r)dr, if R0 ≥ Rk,


2µk

(1− e−πR2

0νk

)−∫ Rk+R0

Rk−R02πrµk


)dr]

·µk

[2π2(Rk −R0)

2R0e−πR2

0νkνk +∫ Rk+R0

Rk−R02πre−νkC(R0,Rk,r)νkL(R0, Rk, r)dr

], if R0 < Rk.

We note that the reference UE is placed at 0. Hence, Rk is also the distance from the reference UE to

its nearest BS in tier-k. However, the reference UE is associated with the BS providing the largest biased

received power (i.e., the reference BS), which may not be the nearest BS. Let R denote the distance

between the reference UE and the reference BS. The probability that the reference BS is a tier-k BS and

R is greater than some R0 is derived as follows:

P[R > R0, tier = k] = P[R > R0|tier = k]P[tier = k]

=

∫ ∞

R0

K∏i=1,i=k

P[Ri > βikr]pdfRk(r)dr. (6.50)


The last equality reflects the fact that the biased received power from the nearest BSs in all other tiers

should not exceed that from the reference BS.

Let pdfR|k(R0) denote the pdf of R given that the reference BS is in tier-k, and let fR,k(R0) ,pdfR|k(R0)P[tier = k], we have

fR,k(R0) =

K∏i=1,i =k

ccdfRi(βikR0) · pdfRk(R0). (6.51)

6.2.2.3.3 Conditional Probability of 0 ∈ T(2)kj (∆d) In this subsection, we study the conditional

probabilities that the reference UE at 0 is in T(2)kj (∆d), given that it is associated with a tier-k BS

(reference BS) at a distance of R0 from it, which is denoted as P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

). By

employing both analytic geometric and stochastic geometric tools, we derive the probability in different

cases, given in Theorems 15-18 below. We note that the handoff rates for P-tiers are already known in

Section 6.1. These theorems additionally address boundaries between P-tiers and C-tiers, boundaries

between different C-tiers, and inter-cluster and intra-cluster boundaries within a C-tier.

For brevity, we define the following quantities that will be used extensively in the rest of this section:

F(β) , 1

β2

∫ π

0

√(β2 + 1)− 2β cos(θ) dθ, (6.52)

and

Hkj(R0) ,1

πF(βkj)βkj

pdfRj(R0βjk)

ccdfRj (R0βjk). (6.53)

Theorem 15. If tier-j is a P-tier, i.e., j ∈ KP , for all k ∈ K (k = j is allowed), we have

P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)= 2λj∆dR0F(βkj) +O(∆d2). (6.54)


Theorem 16. If tier-j is a C-tier, i.e., j ∈ KC , for all k ∈ K, k = j, we have

P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)= Hkj(R0)∆d+O(∆d2). (6.55)


Theorem 17. If tier-k is a C-tier, i.e., k ∈ KC , we have

P(0 ∈ T

(2)kk,out(∆d)|R = R0, tier = k

)= Hkk(R0)∆d+O(∆d2). (6.56)

Proof. Given the reference tier-k BS located at xBS, we know that xBS belongs to some BS cluster

Ωk(C(xBS)). Because the point process of cluster centers Θk is a PPP, by the Slivnyak Theorem [37], the

set of all cluster centers other than C(xBS) remain a PPP with the same statistics as Θk. Consequently,

if we denote the set of all tier-k BSs other than the cluster Ωk(C(xBS)) as Φ′k, then Φ′

k remains a PCP

with the same statistics as Φk. Because the set of inter-cluster cell boundaries T(2)kk,out(∆d) is generated

by xBS and Φ′k, and Φ′

k is still a PCP, the proof of Theorem 17 is the same as that of Theorem 16.


The set of intra-cluster cell boundaries T(2)kk,in(∆d) is generated by xBS and Ωk(C(xBS))\xBS. We

have the following theorem:

Theorem 18. If tier-k is a C-tier, i.e., k ∈ KC , we have P(0 ∈ T

(2)kk,in(∆d)|R = R0, tier = k

)=

Gk(R0)∆d+O(∆d2), where Gk(R0) is expressed as follows:

If R0 ≥ Rk,

Gk(R0) =

∫ R0+Rk

R0−RkrνkR0 (16θm(r,Rk, R0)− 16 sin(θm(r,Rk, R0))) exp(−νkC(R0, Rk, r))dr∫ R0+Rk

R0−Rk2θm(r,Rk, R0)r exp(−C(R0, Rk, r)νk)dr

. (6.57)

If R0 < Rk,

Gk(R0) =π(Rk − R0)

2 exp(−πR2

0νk

)8νkR0 +

∫ Rk+R0Rk−R0

rνkR0 (16θm(r,Rk, R0) − 16 sin(θm(r, Rk, R0))) exp(−νkC(R0, Rk, r))dr

π(Rk − R0)2 exp(−πR2

0νk

)+

∫ Rk+R0Rk−R0

2θm(r,Rk, R0)r exp(−C(R0, Rk, r)νk)dr.

(6.58)


6.2.2.3.4 Unconditioned Probability of 0 ∈ T(2)kj (∆d) Through deconditioning on R, we derive

the unconditioned probabilities that the reference UE at 0 is in T(2)kj (∆d). If k = j, we have

P(0 ∈ T(2)kj (∆d)) =

∫ ∞

0

P(0 ∈ T(2)kj (∆d)|R = R0, tier = k)fR,k(R0)dR0 (6.59)

+

∫ ∞

0

P(0 ∈ T(2)kj (∆d)|R = R0, tier = j)fR,j(R0)dR0.

If k ∈ KP , we have

P(0 ∈ T(2)kk (∆d)) =

∫ ∞

0

P(0 ∈ T(2)kk (∆d)|R = R0, tier = k)fR,k(R0)dR0. (6.60)

If k ∈ KC , we have

P(0 ∈ T(2)kk,x(∆d)) =

∫ ∞

0

P(0 ∈ T(2)kk,x(∆d)|R = R0, tier = k)fR,k(R0)dR0, (6.61)

where the subscript “x” indicates either “in” or “out”.

6.2.2.4 From Area Intensities to Handoff Rates

Next, we compute the handoff rates using the area intensities derived in Section 6.2.2.3.

First, following Section 6.1.2.4, we derive the length intensity µ1

(T

(1)kj

)µ1

(T(1)

y

)= lim

∆d→0

µ2

(T

(2)y (∆d)

)2∆d

, (6.62)

where the subscript “y” indicates “kj”, “kk, in”, or “kk, out”.

Consequently, combining (6.51)-(6.62), we have


0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

1

2

3

4

5

6

7

8

9

10

11

µ2 (unit/km2)

Han

doff

rate

(unit/hour)

1 − 1, analytical1 − 1, simulation1 − 2 or 2 − 1, analytical1 − 2, simulation2 − 1, simulation2 − 2 intra-cluster, analytical2 − 2 intra-cluster, simulation2 − 2 inter-cluster, analytical2 − 2 inter-cluster, simulation2 − 2 all, analytical2 − 2 all, simulation

1 − 1, Poisson

1 − 2 or 2 − 1, Poisson

2 − 2, Poisson

Figure 6.15: Accuracy of PCP handoff rate analysis under different µ2 values, for µ2 · ν2 = 0.5. Forcomparison, black dashed lines indicate analytical results assuming all PPP BSs.

Case 1: If k, j ∈ KP and k = j,

µ1

(T

(1)kj

)=

∫ ∞

0

K∏i=1,i=k

ccdfRi(βikR0)pdfRk(R0)λjR0F(βkj)dR0 (6.63)

+

∫ ∞

0

K∏i=1,i=j

ccdfRi(βijR0)pdfRj

(R0)λkR0F(βjk)dR0.

Case 2: If k ∈ KP we have

µ1

(T

(1)kk

)=

∫ ∞

0

K∏i=1,i=k

ccdfRi(βikR0)pdfRk(R0)λkR0F(1)dR0. (6.64)

Case 3: If k ∈ KP and j ∈ KC , µ1

(T

(1)kj

)can be obtained by replacing the term λjR0F(βkj) with

12Hkj(R0) in (6.63).

Case 4: If k, j ∈ KC and k = j, µ1

(T

(1)kj

)can be obtained by replacing the term λjR0F(βkj) with

12Hkj(R0) and the term λkR0F(βjk) with

12Hjk(R0) in (6.63).

Case 5: If k ∈ KC , by replacing the term λkR0F(1) with 12Hkk(R0) and

12Gk(R0) in (6.64), we find

the expressions of µ1

(T

(1)kk,out

)and µ1

(T

(1)kk,in

), respectively.

Let v denote the instantaneous velocity of an active UE, and Hkj(v) denote its type k-j handoff rate.

Following Section 6.1.2.4, we have:

Hkj(v) =

1πµ1

(T

(1)kj

)v, if k = j,

2πµ1

(T

(1)kj

)v, if k = j.

(6.65)

6.2.3 Numerical Study

In this subsection, we present simulation studies to validate the accuracy and usefulness of the pro-

posed analytical framework. In each round of simulation, multiple tiers of BSs are generated on a 20

km × 20 km square. Then, we randomly generate 5 waypoints X1,X2, . . . ,X5 in the central 10 km

× 10 km square. The four line segments X1X2,X2X3, . . . ,X4X5 form the trajectory of an active UE

in one round of simulation. By tracking which BSs the UE is associated with along its trajectory, we

can obtain its handoff rates in this round of simulation. Each simulation data point is averaged over

2000 simulation rounds. Error bars in the figures show the 95% confidence intervals for simulation results.


0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20

2

4

6

8

10

12

14

λ1 (unit/km2)

Han

doff

rate

(unit/hour)

1 − 1, analytical1 − 1, simulation1 − 2 or 2 − 1, analytical1 − 2, simulation2 − 1, simulation2 − 2 intra-cluster, analytical2 − 2 intra-cluster, simulation2 − 2 inter-cluster, analytical2 − 2 inter-cluster, simulation

Figure 6.16: Handoff rates under different λ1 values.

10−1

100

101

0

2

4

6

8

10

12

B2

Han

doff

rate

(unit/hour)

1 − 1, analytical1 − 1, simulation1 − 2 or 2 − 1, analytical1 − 2, simulation2 − 1, simulation2 − 2 intra-cluster, analytical2 − 2 intra-cluster, simulation2 − 2 inter-cluster, analytical2 − 2 inter-cluster, simulation

Figure 6.17: Handoff rates under different B2 values.

6.2.3.1 Comparison with PPP Modeling

First, we focus on a two-tier HWN with one P-tier (representing macrocell BSs) and one C-tier (repre-

senting clustered femtocell BSs). Tier-1 is the P-tier, with λ1 = 1 unit/km2, P1 = 30 dBm, and B1 = 1;

tier-2 is the C-tier, with R2 = 1 km, P2 = 20 dBm, and B2 = 1; γ = 3.5; and v = 10 km/hour. In

Fig. 6.15, we study the handoff rates under different µ2 values while maintaining a constant µ2ν2 = 0.5

(i.e., the overall tier-2 BS intensity is a constant 0.5π unit/km2). For reference, we also show the handoff

rates (black dashed lines) if tier-2 is replaced by a P-tier with the same BS intensity of 0.5π unit/km2.

Fig. 6.15 shows that the actual handoff rates obtained from simulation match well with the proposed

analysis, while they are far from the dashed lines when µ2 is small (and ν2 is large). This interesting

observation suggests that using a simple PPP to model the BSs as in can lead to substantial numeri-

cal errors in computing the handoff rates in HWNs with clustered BSs. The PPP model gives a close

approximation only when the cluster intensities are high but per cluster BS intensities are low in the

C-tier. In general, we observe that the actual handoff rates vary drastically as µ2 and ν2 change, which

can be accounted for only by the proposed PCP analysis.

6.2.3.2 Effect of BS Densities and Association Bias Values

We further study the influence of different network parameters on the handoff rates. In Figs. 6.16 and

6.17, we consider a two-tier network with the same default parameter values as in Fig. 6.15, except

µ2 = 0.5 unit/km2 and ν2 = 1 unit/km2.

Fig. 6.16 shows handoff rates under different tier-1 BS intensities λ1. The figure illustrates that

increasing λ1 leads to a higher type 1-1 handoff rate but a lower type 2-2 handoff rate. Fig. 6.17


1,1 1,2/2,1 2,2 1,3/3,1 1,4/4,1 2,3/3,2 2,4/4,2 3,3,out 3,3,in 3,4/4,3 4,4,out 4,4,in0

1

2

3

4

5

6

Handoff type

Han

doff

rate

(unit/hour)

AnalyticalSimulation

Figure 6.18: Handoff rates, two P-tiers and two C-tiers.

shows the handoff rates under different tier-2 association bias values B2. We observe that increasing the

association bias value of one tier has a similar effect as increasing the BS intensity of this tier, leading

to a higher horizontal handoff rate within this tier, but lower handoff rates outside the tier. Moreover,

both figures illustrate that the simulation results agree with the analytical results, demonstrating the

correctness of our analysis for different types of handoff rates.

6.2.3.3 A More Complex Example

In Fig. 6.18, we study different types of handoff rates in a network with four tiers. The network pa-

rameters are as follows: Tier-1 and tier-2 are P-tiers, with (λ1, λ2) = (1, 1) unit/km2; tier-3 and tier-4

are C-tiers, with (µ3, µ4) = (0.5, 1) unit/km2, (R3, R4) = (0.9, 1.1) unit/km2, and (ν3, ν4) = (1, 0.5)

unit/km2; (P1, P2, P3, P4) = (30, 33, 20, 23) dBm; (B1, B2, B3, B4) = (1, 1, 1, 1); γ = 3.5; and v = 10

km/hour. In order to avoid redundancy, we only show the sum rate of type k-j and type j-k (k = j)

handoffs for easier inspection; the individual handoff rates are half of the sum handoff rate. We again

observe that the simulation results agree with the analytical results, validating the correctness of our

analysis. We further observe that the handoff rates of types 2-2, 2-3/3-2, and 2-4/4-2 are higher than

those of 1-1, 1-3/3-1, and 1-4/4-1, respectively, due to the higher transmission power of tier-2 BSs com-

pared with tier-1 BSs. The handoff rates of types 4-4, 1-4/4-1, and 2-4/4-2 are higher than those of 3-3,

1-3/3-1, and 2-3/3-2, respectively, due to the higher transmission power and higher intensity of tier-4

BSs, i.e., πR24µ4ν4, compared with tier-3 BSs.

6.3 Summary

In this chapter, we provide a theoretical framework to study user mobility in multi-tier HWNs. Through

a stochastic geometric framework, we capture the irregularly shaped network topologies introduced by

the small-cell structure. In Section 6.1, we are able to capture (1) the spatial randomness of BSs, and

(2) the various scales of cell sizes. In Section 6.2, we additionally capture (3) the non-uniform and

dependent aggregation of BSs, by allowing BSs to form Poisson cluster processes. Analytical expressions

for the rates of all handoff types experienced by an active UE with arbitrary movement trajectory are

derived. In addition, we study the downlink data rate of the UE under different tier selections. Based

on these results, we propose an optimal tier selection formulation considering both the handoff rates and

the data rate. Empirical study on the Yonsei Trace and extensive simulation are conducted, validating

the accuracy and usefulness of our analytical conclusions.

Chapter 7

Conclusions

In this thesis, we develop new methods and models to evaluate and design HWNs by investigating user

load, interference patterns, and user mobility, the results of which provide new analytical insights and

design guidelines that will help improve the performance of HWNs.

In Chapter 3, we focus on the evaluation of user load through characterizing the joint distribution of

users in different cells in an HWN with arbitrary user movement trajectories and dependently distributed

user channel holding times. Through developing a new queueing network model, we derive a closed-form

expression for the user distribution, revealing several desirable properties: the joint user distribution is

only related to the average arrival rate and the average channel holding time of each cell, and hence

it is insensitive to the general user movement patterns and distributions of channel holding times;

the marginal user distribution within a single cell depends only on the average arrival rate and the

average channel holding time at that cell. These properties suggest that accurate evaluation of the

user distribution and other associated metrics such as the system workload can be achieved with low

complexity, without the need to collect a large amount of user location data.

In Chapter 4, we study the interference patterns in HWNs with randomly spatially distributed users

and BSs. We additionally characterize the outage performance of closed access mode and open access

mode. However, in order to capture the complicated spatial patterns of users and BSs, the mathematical

expressions of the outage performance inevitably involve multiple folds of integrals. Therefore, we further

derive sufficient conditions in simple expressions for open and closed access modes to outperform each

other, at either the macrocell or the femtocell level. Further, we prove that the maximum allowable ratio

of the received power level of an open access user in the femtocell to its original received power level in

the macrocell is in the scale of Θ((

Rc

R

)γ)(where Rc, R, and γ represent the radius of macrocells, the

radius of femtocells, and path loss exponent, respectively), so that the network can benefit through the

replacement of closed access by open access.

In Chapter 5, we study the spectrum allocation and user association problem in HWNs with random

distributed users and BSs for optimizing the average user data rate. Both the user load and interference

patterns are considered. We first derive the average user data rate through stochastic geometric analysis.

The expression is employed as the objective function of the optimization problem, which is non-convex

in nature and cannot be solved with a standard method. Then, we propose an innovative approach,

solving the optimization problem optimally for low user density, and asymptotically optimally for high

user density. In this chapter, if the spectrum allocation factors are not constrained, SSAUA leads to a

91

Chapter 7. Conclusions 92

simple solution, where only the BS tier with the highest density will be allocated with 100% spectrum,

which may not be desirable in reality. Therefore, the user data rate may not be the only design goal.

It will also be crucial to benefit network traffic through improving network fairness, decreasing traffic

delay, reducing data rate fluctuation, and improving quality of experience, which are still challenging

open problems left for future research.

In Chapter 6, a stochastic geometric analysis framework on user mobility is proposed, which captures

the spatial randomness and various scales of cell sizes in different tiers. We derive theoretical expressions

for the rates of all handoff types experienced by an active user with an arbitrary movement trajectory.

In Section 6.1, we employ the conventional assumption on HWNs that each BS tier forms a homogeneous

PPP. In Section 6.2 we further extend the analysis by allowing BSs to form PCPs, in order to capture

their non-uniform and dependent aggregation in space. Then, we provide guidelines for tier selection

under various user velocity, so that an optimal tradeoff between the handoff rate and the data rate can

be achieved. The handoff rates in this chapter are derived under the assumption that the BSs in each

tier form PPPs or PCPs. In reality, BSs may be even more complicated distributed (e.g., Matern hard

core point process). Handoff analysis for more general BS models remains a challenging open problem

for future research.

Appendix A

Proofs of Theorems in Chapter 4

A.1 Proof of Theorem 3

LI(s) = E (exp(−sI)) = E

[ ∏x∈Φ0

u(x, s)∏

x0∈Θ

∏x∈Ω(x0)

v(x,x0, s)∏

x0∈Θ

∏x∈Ψ(x0)

w(x,x0, s)

](A.1)

=E

[E( ∏

x∈Φ0

u(x, s)

∣∣∣∣Θ)E( ∏x0∈Θ

∏x∈Ω(x0)

v(x,x0, s)

∣∣∣∣Θ)E( ∏x0∈Θ

∏x∈Ψ(x0)

w(x,x0, s)

∣∣∣∣Θ)]

(A.2)

=E

[E( ∏

x∈Φ0

u(x, s)

∣∣∣∣Θ)E( ∏

x∈Φ1

u(x, s)

∣∣∣∣Θ)E( ∏

x∈Φ1

u(x, s)

∣∣∣∣Θ)E( ∏

x0∈Θ

∏x∈Ω(x0)

v(x,x0, s)

∣∣∣∣Θ)E( ∏x0∈Θ

∏x∈Ψ(x0)

w(x,x0, s)

∣∣∣∣Θ)]

(A.3)

=E

[E( ∏

x∈Φ

u(x, s)

∣∣∣∣Θ)E(∏

x0∈Θ

∏x∈Ω(x0)

v(x,x0, s)

∣∣∣∣Θ)E(∏

x0∈Θ

∏x∈Ω(x0)

u(x, s)

∣∣∣∣Θ) E( ∏

x0∈Θ

∏x∈Ψ(x0)

w(x,x0, s)

∣∣∣∣Θ)]

(A.4)

=E( ∏

x∈Φ

u(x, s)

)E

[ ∏x0∈Θ

(E(∏


)E(∏

x∈Ω(x0)u(x, s)

) E( ∏

x∈Ψ(x0)

w(x,x0, s)))]

. (A.5)

Proof. The steps to derive Theorem 3 is shown in (A.1)-(A.5), where Φ0 is the point process correspond-

ing to macrocell UEs not inside any femtocell, Φ1 is the point process corresponding to macrocell UEs

inside some femtocell, and Φ is the aggregation of Φ0 and Φ1.

By the law of total expectation, we derive (A.2) from (A.1). Φ1 can be rewritten as the union of all the

open access UEs in each femtocell, so E(∏

x∈Φ1 u(x, s)

∣∣∣∣Θ) is equal to E(∏

x0∈Θ

∏x∈Ω(x0)

u(x, s)

∣∣∣∣Θ).In addition, because Φ is the aggregation of Φ0 and Φ1, E

( ∏x∈Φ0

u(x, s)

∣∣∣∣Θ) E( ∏

x∈Φ1

u(x, s)

∣∣∣∣Θ) is equal

to E(∏

x∈Φ u(x, s)

∣∣∣∣Θ). By considering the two equalities, we derive (A.4) from (A.3). Finally, we

obtain (A.5) from the law of total expectation.

93

Appendix A. Proofs of Theorems in Chapter 4 94


Proof. We use the fact that P and Rc can be normalized and set P = Rc = 1. Furthermore, we substitute

s = T into (4.6), (4.7), and (4.8) such that

W(x0, T ) = exp

(−

∫B(0,R)

TQ|x|γ

TQ|x|γ + |x+ x0|γν(x)dx

), (A.6)

V(x0, T ) = exp

(− λ

∫B(0,R)

Tρ|x|γ

Tρ|x|γ + |x+ x0|γdx

), (A.7)

U(x0, T ) = exp

(− λ

∫B(x0,R)

T |x− BS(x)|γ

T |x− BS(x)|γ + |x|γdx

). (A.8)

(a) A sufficient condition for P oout < P c

out

According to (4.10), (4.11), (4.13), and (4.14), P oout < P c

out iff

exp

(− µ

∫R2

(1− V(x0,T )

U(x0,T )W(x0, T ))dx0

)exp

(− µ

∫R2

(1−W(x0, T )

)dx0

) > 1, (A.9)

which is equivalent to ∫R2

(V(x0, T )

U(x0, T )− 1

)W(x0, T )dx0 > 0. (A.10)

Let

V (x0) =

∫B(x0,R)

Tρ|x− x0|γ

Tρ|x− x0|γ + |x|γdx =

∫B(0,R)

Tρ|x|γ

Tρ|x|γ + |x+ x0|γdx, (A.11)

U(x0) =

∫B(x0,R)

T |x− BS(x)|γ

T |x− BS(x)|γ + |x|γdx. (A.12)

Then, (A.10) becomes ∫R2

(exp(−λV (x0))

exp(−λU(x0))− 1

)W(x0, T )dx0 > 0. (A.13)

Because exp(A)−1 ≥ A for arbitraryA, andW(x0, T ) > 0, we have∫R2

(exp(−λV (x0))exp(−λU(x0))

− 1)W(x0, T )dx0 >∫

R2 (−λV (x0) + λU(x0))W(x0, T )dx0. Therefore, the following inequality is a sufficient condition for

(A.13): ∫R2

(−λV (x0) + λU(x0))W(x0, T )dx0 > 0. (A.14)

Let Wmin and Wmax be the lower bound and upper bound of W(x0, T ), respectively. According to


(4.6), Wmax = 1 and Wmin = e−ν . Thus, the following is a sufficient condition for (A.14):

−Wmax

∫R2

V (x0)dx0 +Wmin

∫R2

U(x0)dx0 > 0. (A.15)

Let V =∫R2 V (x0)dx0, we have the following lemma corresponding to the upper and lower bounds

of V.

Lemma 5. Suppose γ > 2. Let

Vmax = 4π2R4(Tρ)2γ

(1

8+

1

4(γ + 2)+

1

(γ + 2)(γ − 2)

), (A.16)

Vmin = 2π2R4(Tρ)2γ

(1

8+

1

4(γ + 2)+

1

(γ + 2)(γ − 2)

). (A.17)

Then Vmin ≤ V ≤ Vmax.


Hence, the following is a sufficient condition for (A.15):

−WmaxVmax +Wmin

∫R2

U(x0)dx0 > 0. (A.18)

In addition, we have∫R2

U(x0)dx0 =

∫R2

∫B(x0,R)

(T |x− BS(x)|γ

T |x− BS(x)|γ + |x|γ

)dxdx0

= πR2

∫R2

(T |x− BS(x)|γ


)dx = πR2Cu, (A.19)

where

Cu =

∫R2

(T |x− BS(x)|γ


)dx (A.20)

is only related to predetermined system-level constants T and γ.

As a consequence, (A.18) becomes

−WmaxVmax +WminπR2Cu > 0. (A.21)

(b) A sufficient condition for P oout > P c

out

According to (4.10), (4.11), (4.13), and (4.14), P oout > P c

out iff

exp

(− µ

∫R2

(1−W(x0, T )

)dx0

)exp

(− µ

∫R2

(1− V(x0,T )

U(x0,T )W(x0, T ))dx0

) > 1, (A.22)

Similar to the steps in (A.9)-(A.14), the following is a sufficient condition for (A.22):∫R2

(−λU(x0, T ) + λV (x0, T ))V(x0, T )

U(x0, T )W(x0, T )dx0 > 0. (A.23)


Let W ′min and W ′

max be the lower bound and upper bound of V(x0,T )U(x0,T )W(x0, T ), respectively. According

to (4.6), (4.7), and (4.8), W ′max = exp

(λ)and W ′

min = exp(−λ− ν

). Finally, the following is a sufficient

condition for (A.23):

−W ′maxπR

2Cu +W ′minVmin > 0. (A.24)


E(exp

(−s(I ′(xB))

))(A.25)

=E

[ ∏x∈Φ0

u′(x,xB , s) ·∏

x0∈Θ

∏x∈Ω(x0)

v′(x,x0,xB , s) ·∏

x0∈Θ

∏x∈Ψ(x0)

w′(x,x0,xB , s) (A.26)

·∏

x∈Ω(xB)

v′(x,xB ,xB , s) ·∏

x∈Ψ(xB)

w′(x,xB ,xB , s)

]

=E

[E( ∏

x∈Φ0

u′(x,xB , s)

∣∣∣∣Θ)E( ∏

x∈Φ1

u′(x,xB , s)

∣∣∣∣Θ)E( ∏

x∈Φ1

u′(x,xB , s)

∣∣∣∣Θ)E( ∏

x0∈Θ

∏x∈Ω(x0)

v′(x,x0,xB , s)

∣∣∣∣Θ) (A.27)

E( ∏

x0∈Θ

∏x∈Ψ(x0)

w′(x,x0,xB , s)

∣∣∣∣Θ)E( ∏x∈Ω(xB)

v′(x,xB ,xB , s)

∣∣∣∣Θ)E( ∏x∈Ψ(xB)

w′(x,xB ,xB , s)

∣∣∣∣Θ)]

=E

[E( ∏

x∈Φ

u′(x,xB , s)

∣∣∣∣Θ)E(∏

x0∈Θ

∏x∈Ω(x0)

v′(x,x0,xB , s)

∣∣∣∣Θ)E(∏

x0∈Θ

∏x∈Ω(x0)

u′(x,xB , s)

∣∣∣∣Θ)E(∏

x∈Ω(xB) v′(x,xB ,xB , s)

∣∣∣∣Θ)E(∏


∣∣∣∣Θ) (A.28)

E( ∏

x0∈Θ

∏x∈Ψ(x0)

w′(x,x0,xB , s)

∣∣∣∣Θ)E( ∏x∈Ψ(xB)

w′(x,xB ,xB , s)

∣∣∣∣Θ)]

=E( ∏

x∈Φ

u′(x,xB , s)

)E

[ ∏x0∈Θ

(E(∏

x∈Ω(x0)v′(x,x0,xB , s)

)E(∏

x∈Ω(x0)u′(x,xB , s)

) E( ∏

x∈Ψ(x0)

w′(x,x0,xB , s)))]

(A.29)

E( ∏

x∈Ψ(xB)

w′(x,xB ,xB , s)

)E(∏

x∈Ω(xB) v′(x,xB ,xB , s)

)E(∏


) .

Proof. The steps to derive Theorem 5 is shown in (A.25)-(A.29). Substituting (4.20)-(4.24) into (A.25),

we derive (A.26). According to the law of total expectation, we derive (A.27) from (A.26). Φ1 can be

rewritten as the union of all open access UEs in each femtocell (including the typical femtocell). Thus

E(∏

x∈Φ1 u′(x,xB , s)


x0∈Θ

∏x∈Ω(x0)

u′(x,xB, s)

∣∣∣∣Θ)·E(∏x∈Ω(xB) u′(x,xB, s)

∣∣∣∣Θ).Also, E

(∏x∈Φ0 u′(x,xB, s)

∣∣∣∣Θ)· E(∏x∈Φ1 u′(x,xB , s)


x∈Φ u′(x,xB , s)

∣∣∣∣Θ). By

considering the two equalities, we derive (A.28) from (A.27). Finally, we derive (A.29) from the law of


total expectation.


Proof. We use the fact that P and Rc can be normalized and set P = Rc = 1. Furthermore, we substitute

s = T ′ into (4.27)-(4.32) such that

W ′(x0,xB , T′) = exp

(−

∫B(x0,R)

T ′Q|x− x0|γ

T ′Q|x− x0|γ + |x− xB |γν(x− x0)dx

),

V ′(x0,xB , T′) = exp

(− λ

∫B(x0,R)

T ′ρ|x− x0|γ

T ′ρ|x− x0|γ + |x− xB |γdx

),

U ′(x0,xB , T′) = exp

(− λ

∫B(x0,R)

T ′|x− BS(x)|γ

T ′|x− BS(x)|γ + |x− xB |γdx

),

V ′′(xB , T′) =e

− T ′ρλT ′ρ+1 ,

U ′′(xB , T′) = exp

(− λ

∫B(xB ,R)

T ′|x− BS(x)|γ

T ′|x− BS(x)|γ + |x− xB |γdx

). (A.30)

(a) A sufficient condition for P oout(xB) < P c

out(xB)

According to (4.33), (4.34), (4.36), and (4.37), P oout(xB) < P c

out(xB) iff

exp

(− µ

∫R2

(1− V′(x0,xB ,T ′)

U ′(x0,xB ,T ′)W′(x0,xB , T

′))dx0

)exp

(− µ

∫R2

(1−W ′(x0,xB, T ′)

)dx0

) V ′′(xB, T′)

U ′′(xB, T ′)> 1. (A.31)

Let

V ′(x0,xB) =

∫B(x0,R)

T ′ρ|x− x0|γ

T ′ρ|x− x0|γ + |x− xB |γdx,

U ′(x0,xB) =

∫B(x0,R)

(T ′|x− BS(x)|γ

T ′|x− BS(x)|γ + |x− xB |γ

)dx,

R =

∫B(xB ,R)

T ′|x− BS(x)|γ

T ′|x− BS(x)|γ + |x− xB |γdx. (A.32)

Similar to (A.14), the following is a sufficient condition for (A.31):

µ

∫R2

(−λV ′(x0,xB) + λU ′(x0,xB))W ′(x0,xB, T′)dx0 −

λπR2T ′ρ

T ′ρ+ 1+ λR > 0. (A.33)

Let W ′′min and W ′′

max be the lower bound and upper bound of W ′(x0,xB , T′), respectively. According

to (4.27), W ′′max = 1 and W ′′

min = exp (−ν). Thus, the following is a sufficient condition for (A.33):

µ

∫R2

(−V ′(x0,xB)W′′max + U ′(x0,xB)W

′′min) dx0 −

πR2T ′ρ

T ′ρ+ 1+R > 0, (A.34)


where∫R2 V

′(x0,xB)dx0 =∫R2

∫B(x0,R)

T ′ρ|x−x0|γ|x|γ

T ′ρ|x−x0|γ|x|γ +1

dxdx0 is in the same form as (A.41). Thus, by

applying Lemma 5, we can derive its upper bound and lower bound as V′max and V′

min from (A.44) and

(A.47), respectively. Similar to the derivation of (A.19),∫R2 U

′(x0,xB)dx0 = πR2C ′u where

C ′u ,

∫R2

(T ′|x− BS(x)|γ

T ′|x− BS(x)|γ + |x− xB|γ

)dx. (A.35)

In addition, the lower bound Rmin and the upper bound Rmax of R can be derived as follows:

Rmin =

π∫ R

0T ′(|xB |)γr

T ′(|xB |)γ+rγ dr if |xB | ≤ R,

π∫ R

0T ′(|xB |−R)γr

T ′(|xB |−R)γ+rγ dr + π∫ R

0T ′(|xB |)γr

T ′(|xB |)γ+rγ dr if |xB | > R,(A.36)

and

Rmax =π

∫ R

0

T ′(|xB|+R)γr

T ′(|xB |+R)γ + rγdr + π

∫ R

0

T ′(√|xB|2 +R2)γr

T ′(√|xB |2 +R2)γ + rγ

dr. (A.37)

Note that∫

Brrγ+Bdr is in closed form when γ is a rational number. Therefore, both Rmin and Rmax

are expressed in closed forms.

Finally, the following is a sufficient condition for (A.34):

− µV′max + µπR2C ′

uW′′min − πR2T ′ρ

T ′ρ+ 1+Rmin > 0. (A.38)

(b) A sufficient condition for P oout(xB) > P c

out(xB)

Similar to the derivations of (A.31) and (A.33), P oout(xB) > P c

out(xB) iff

µ

∫R2

(−λU ′(x0,xB) + λV ′(x0,xB))W ′(x0,xB , T

′)V ′(x0,xB , T′)

U ′(x0,xB , T ′)dx0 +

λπR2T ′ρ

T ′ρ+ 1− λR > 0. (A.39)

Let W ′′′min and W ′′′

max be the lower bound and upper bound value of W′(x0,xB ,T ′)V′(x0,xB ,T ′)U ′(x0,xB ,T ′) , respec-

tively. According to (4.27)-(4.29), W ′′′max = exp

(λ)and W ′′′

min = exp(−λ− ν

). Then similar to the

derivation of (A.38), we see that the following is a sufficient condition for (A.39):

− µπR2C ′uW

′′′max + µV′

minW′′′min +

πR2T ′ρ

T ′ρ+ 1−Rmax > 0. (A.40)

A.5 Proof of Lemma 5

Proof. Upper Bound of V

V =

∫R2

∫B(x0,R)

Tρ|x−x0|γ|x|γ

Tρ|x−x0|γ|x|γ + 1

dxdx0 (A.41)


=

∫R2

∫B(x,R)

Tρ|x−x0|γ|x|γ

Tρ|x−x0|γ|x|γ + 1

dx0dx

=

∫ ∞

0

2πr1

∫ R

0

Tρrγ2rγ1

Tρrγ2rγ1

+ 12πr2dr2dr1 (A.42)

≤∫ ∞

0

2πr1

∫ R

0

1(Tρrγ2rγ1

≥ 1)2πr2dr2dr1+∫ ∞

0

2πr1

∫ R

0

1(Tρrγ2rγ1

< 1)Tρrγ2rγ1

2πr2dr2dr1 (A.43)

=4π2R4(Tρ)2γ

(1

8+

1

4(γ + 2)+

1

(γ + 2)(γ − 2)

). (A.44)

In (A.42), the integrated item is in the form of XX+1 , where X =

Tρrγ2rγ1

≥ 0. The bound of the integrated

item can be found as follows: if X ≥ 1, 12 ≤ X

X+1 ≤ 1; otherwise, if X < 1, X2 ≤ X

X+1 ≤ X. Accordingly,

we can separate the integration region into theTρrγ2rγ1

≥ 1 region and theTρrγ2rγ1

< 1 region. As a

consequence, the upper bound of (A.42) can be derived as (A.43).

Lower Bound of V

Following a similar approach as above, we have

V =

∫ ∞

0

2πr1

∫ R

0

Tρrγ2rγ1

Tρrγ2rγ1

+ 12πr2dr2dr1 (A.45)

≥∫ ∞

0

2πr1

∫ R

0

1(Tρrγ2rγ1

≥ 1)πr2dr2dr1+∫ ∞

0

2πr1

∫ R

0

1(Tρrγ2rγ1

< 1)Tρrγ2rγ1

πr2dr2dr1 (A.46)

=2π2R4(Tρ)2γ

(1

8+

1

4(γ + 2)+

1

(γ + 2)(γ − 2)

). (A.47)

Appendix B

Properties and Proofs of Chapter 5

B.1 Useful Properties of Mk(Ak)

(M-1) Mk(Ak) is increasing on [0, ak] and decreasing on [ak,∞).

(M-2) Mk(Ak) is concave on [0, ak].

(M-3) If λi < λj , then Mi(A) < Mj(A), ∀A > 0.

(M-4) If λi < λj , then Mj(A)−Mi(A) is strictly increasing on [0, aj ].

B.2 Proof of Lemma 2

Proof. Suppose A∗∗ is optimal, A∗∗i < ai, and A∗∗

j > aj . Consider that we increase A∗∗i by a small

value ∆ > 0 and decrease A∗∗j by ∆. According to property (M-1), η∗∗i Mi(A

∗∗i ) + η∗∗j Mj(A

∗∗j ) <

η∗∗i Mi(A∗∗i + ∆) + η∗∗j Mj(A

∗∗j − ∆). Thus, through replacing A∗∗

i and A∗∗j by A∗∗

i + ∆ and A∗∗j − ∆,

respectively, we find a better solution to P, leading to a contradiction.


Proof. Suppose ∃i < j such that Mi(A∗i ) > Mj(A

∗j ). This implies that A∗

i > A∗j . (Otherwise, if A∗

i ≤ A∗j ,

then we have Mi(A∗i ) ≤ Mi(A

∗j ) < Mj(A

∗j ), leading to a contradiction.) A corresponding diagram is

shown in Fig. B.1(a).

Case 1: η∗i ≤ η∗j .

Let A∗j = A∗

i and A∗i = A∗

j . Then we have

[η∗i Mi(A∗i ) + η∗jMj(A

∗j )]− [η∗iMi(A

∗i ) + η∗jMj(A

∗j )]

=[η∗i Mi(A∗j ) + η∗jMj(A

∗i )]− [η∗iMi(A

∗i ) + η∗jMj(A

∗j )]

=η∗j [Mj(A∗i )−Mj(A

∗j )] + η∗i [Mi(A

∗j )−Mi(A

∗i )]

≥η∗i [Mj(A∗i )−Mj(A

∗j ) +Mi(A

∗j )−Mi(A

∗i )] > 0, (B.1)

where (B.1) is due to property (M-4).

100

Appendix B. Properties and Proofs of Chapter 5 101

Ai*

Aj*

Mi(Ai*)

Mi(Aj*)

Mj(Aj*)

Mj(Ai*)

Mj

Mi

(a) Diagram of Mi(·) and Mj(·).Aj*

Mi(Ai*)

~

Mj(Ai*)

~

Mj(Aj*)

~

Mi(Aj*)

~ ~Mi

~Mj

ai aj Ai*~

~

~

~

~

~

(b) Diagram of Mi(·) and

Mj(·), Case 1.1 and 2.1.

Aj*

D

D

Mi(Ai*)

~ Mj(Aj*)

~

Mj(Aj*)

~Mi

~Mj

~ ^

Mi(Ai*)

~ ^

ai ajAi*~

~ ~

~

(c) Diagram of Mi(·) and

Mj(·), Case 1.2 and 2.2.

Figure B.1: Diagrams of Mi(·), Mj(·), Mi(·), and Mj(·).

As a consequence, if A∗i and A∗

j are replaced by A∗i and A∗

j , respectively, we obtain a larger F, leading

to a contradiction.

Case 2: η∗i > η∗j .

Let A∗j = A∗

i , A∗i = A∗

j , η∗j = η∗i , and η∗i = η∗j . (Note that because ηmin,i ≤ ηmin,j and ηmax,i ≤ ηmax,j ,

η∗j and η∗i are guaranteed to be in the feasible region.)

[η∗i Mi(A∗i ) + η∗jMj(A

∗j )]− [η∗iMi(A

∗i ) + η∗jMj(A

∗j )]

=[η∗jMi(A∗j ) + η∗i Mj(A

∗i )]− [η∗iMi(A

∗i ) + η∗jMj(A

∗j )]

=η∗i [Mj(A∗i )−Mi(A

∗i )] + η∗j [Mi(A

∗j )−Mj(A

∗j )]

>η∗j [Mj(A∗i )−Mi(A

∗i ) +Mi(A

∗j )−Mj(A

∗j )] > 0. (B.2)

Thus, if A∗i , A

∗j , η

∗i , and η∗j are replaced by A∗

i , A∗j , η

∗i , and η∗j , respectively, we can find a larger F,

leading to a contradiction.

B.4 Useful Properties of Mk(Ak)

(M-1’) Mk(Ak) is a decreasing convex function.

(M-2’) If λi < λj , then Mi(ai) < Mj(aj).

(M-3’) If λi < λj , then Mj(A)− Mi(A) is a strictly decreasing function.

(M-4’) Mk(A)− Mk(A+D) > Mk(A′)− Mk(A

′ +D), for any A′ > A ≥ ak and D > 0.

(M-5’) If λi < λj , then Mj(aj)− Mj(aj +D) > Mi(ai)− Mi(ai +D), for any D > 0.

(M-6’) If λi < λj , then Mj(aj) − Mj(aj + D) > Mi(A′) − Mi(A

′ + D), for any D > 0 and A′ > ai

(combining (M-4’) and (M-5’)).


Proof. Suppose that ∃i < j (i.e., λi < λj) such that Mi(A∗i ) > Mj(A

∗j ), which also implies that

ai ≤ A∗i < A∗

j . The corresponding diagrams are shown in Figs. B.1(b) and B.1(c).

Case 1: η∗i ≤ η∗j .


Case 1.1: A∗i ≥ aj.

Let A∗j = A∗

i and A∗i = A∗

j . We have

[η∗i Mi(A∗i ) + η∗j Mj(A

∗j )]− [η∗i Mi(A

∗i ) + η∗j Mj(A

∗j )]

=[η∗i Mi(A∗j ) + η∗j Mj(A

∗i )]− [η∗i Mi(A


∗j )]

=η∗j [Mj(A∗i )− Mj(A

∗j )] + η∗i [Mi(A

∗j )− Mi(A

∗i )]

≥η∗i [Mj(A∗i )− Mj(A

∗j ) + Mi(A

∗j )− Mi(A

∗i )] > 0, (B.3)

where (B.3) is due to property (M-3’).

Thus, if A∗i and A∗


j , respectively, we obtain a larger F′, leading to a

contradiction.

Case 1.2: A∗i < aj.

Let A∗j = aj , D = A∗

j − aj and A∗i = A∗

i +D. We have


∗j )]− [η∗i Mi(A


∗j )]

=[η∗i Mi(A∗i +D) + η∗j Mj(aj)]− [η∗i Mi(A


∗j )]

=η∗j [Mj(aj)− Mj(A∗j )] + η∗i [Mi(A

∗i +D)− Mi(A

∗i )]

≥η∗i [Mj(aj)− Mj(A∗j ) + Mi(A

∗i +D)− Mi(A

∗i )] > 0, (B.4)

where (B.4) is due to property (M-6’).

Thus, if A∗i and A∗


j , respectively, we obtain a larger F′, leading to a

contradiction.

Case 2: η∗i > η∗j .

Case 2.1: A∗i ≥ aj.

Let A∗j = A∗

i , A∗i = A∗

j , η∗j = η∗i , and η∗i = η∗j . (Note that because ηmin,i ≤ ηmin,j and ηmax,i ≤ ηmax,j ,

η∗j and η∗i are guaranteed to be in the feasible region.)


∗j )]− [η∗i Mi(A


∗j )]

=[η∗j Mi(A∗j ) + η∗i Mj(A

∗i )]− [η∗i Mi(A


∗j )]

=η∗i [Mj(A∗i )− Mi(A

∗i )] + η∗j [Mi(A

∗j )− Mj(A

∗j )]

>η∗j [Mj(A∗i )− Mi(A

∗i ) + Mi(A

∗j )− Mj(A

∗j )] > 0. (B.5)

Thus, if A∗i , A

∗j , η


i , A∗j , η

∗i , and η∗j , respectively, we obtain a larger F′,


Case 2.2: A∗i < aj.

Let A∗j = aj , D = A∗

j − aj , A∗i = A∗

i +D, η∗j = η∗i , and η∗i = η∗j . We have


∗j )]− [η∗i Mi(A


∗j )]

=[η∗j Mi(A∗i +D) + η∗i Mj(aj)]− [η∗i Mi(A


∗j )]

=η∗i [Mj(aj)− Mi(A∗i )] + η∗j [Mi(A

∗i +D)− Mj(A

∗j )]

>η∗j [Mj(aj)− Mi(A∗i ) + Mi(A

∗i +D)− Mj(A

∗j )] > 0. (B.6)


Thus, if A∗i , A

∗j , η


i , A∗j , η

∗i , and η∗j , respectively, we obtain a larger F′,


B.6 Proof of Theorem 10

Proof. Suppose ∃k ≥ 2 such that A∗k > ak. Let l = 1, A∗

k = ak, D = A∗k − ak, and A∗

l = A∗l +D. Note

that we have η∗k ≥ η∗l through Lemma 4. Thus,

[η∗kMk(A∗k) + η∗l Ml(A

∗l )]− [η∗kMk(A

∗k) + η∗l Ml(A

∗l )]

=[η∗kMk(ak) + η∗l Ml(A∗l +D)]− [η∗kMk(A

∗k) + η∗l Ml(A

∗l )]

=η∗k[Mk(ak)− Mk(A∗k)] + η∗l [Ml(A

∗l +D)− Ml(A

∗l )]

≥η∗l [Mk(ak)− Mk(A∗k) + Ml(A

∗l +D)− Ml(A

∗l )] > 0. (B.7)

As a consequence, if we replace A∗l and A∗

k by A∗l and A∗

k, respectively, we find a better solution to

Problem P2A, which leads to a contradiction.

B.7 Some Properties Used for Exhaustive Search

Lemma 6. Let (η∗,A∗) be an optimal solution to P. Suppose ∃i = j, such that ηmin,i < η∗i < ηmax,i

and ηmin,j < η∗j < ηmax,j . Then, let η∗i = ηmax,i and η∗j = η∗i + η∗j − ηmax,i if ηmax,i− η∗i ≤ η∗j − ηmin,j ; let

η∗i = η∗i + η∗j − ηmin,j and η∗j = ηmin,j otherwise. Let η∗ = (η∗1 , . . . , η∗i , . . . , η

∗j , . . . , η

∗K). Then (η∗,A∗) is

still an optimal solution to Problem P.

Proof. First, we have Mi(A∗i ) = Mj(A

∗j ). Otherwise, suppose Mi(A

∗i ) > Mj(A

∗j ) (without loss of

generality); then we can find a better solution by replacing η∗i and η∗j by η∗i + δ and η∗j − δ, respectively,

where δ > 0.

Given that Mi(A∗i ) = Mj(A

∗j ), the same F can be obtained when we replace η∗i and η∗j by η∗i and η∗j .

Lemma 6 demonstrates that if there are η∗i and η∗j not at the boundary, we can “push” one of them

to the boundary and maintain the optimization in P. If there are more than two terms in η∗ not at the

boundary, we can “push” them to the boundary one by one, until there is at most one term in η∗ not

at the boundary. Thus, Lemma 6 directly leads to the following Theorem:

Theorem 19. At least one of the optimal solutions to P, (η∗,A∗), has the following property: There

is at most one k ∈ 1, 2, . . .K such that ηmin,k < η∗k < ηmax,k, and ∀j = k, either η∗j = ηmin,j or

η∗j = ηmax,j .

B.8 Proofs of Properties of Mk(·) in the Multiple-MCS Case

In this appendix, we prove Properties (M-1) to (M-4) when we redefine Mk(A) as in (5.44). In the proof,

we set uk = µ/λk.


B.8.1 Property (M-1)

The first derivative of Mk(A) is

M ′k(A) =

1

(Auk + 1)2

[ N∑n=1

bn(1−A2Cnuk)

(ACn + 1)2

]. (B.8)

M ′k(A) is positive at A = 0, and is negative when A is sufficient large. Thus, there exists pos-

itive ak such that it is a solution to M ′k(A) = 0. In addition, we have

[∑Nn=1

bn(1−A2Cnuk)(ACn+1)2

]′=

−∑N

n=12bnCn(Auk+1)

(ACn+1)3 < 0, demonstrating that∑N

n=1bn(1−A2Cnuk)

(ACn+1)2 is a decreasing function. Thus, ak is

the unique solution to M ′k(A) = 0. M ′

k(A) is positive on [0, ak), and negative on (ak,∞). Note that a

simple binary search method can be applied to numerically search for ak.

Note that

N∑n=1

bn(1−A2Cnuk)

(ACn + 1)2> 0 (B.9)

on [0, ak), which will be used in the subsequent steps.


We have

M ′′k (A) =

[∑Nn=1 z

′n(A)

]v(A)−

[∑Nn=1 zn(A)

]v′(A)

(v(A))2, (B.10)

where zn(A) =bn(1−A2Cnuk)

(ACn+1)2 , and v(A) = (Auk + 1)2. Then, we have

z′n(A) = −2bnCn(Auk + 1)

(ACn + 1)3, (B.11)

which is negative.

Consequently, on [0, ak),∑N

n=1 z′n(A) is negative, v(A) is positive,

∑Nn=1 zn(A) is positive (due to

(B.9)), and v′(A) is positive. Thus, M ′′k (A) is negative. Property (M-2) is proved.


Trivially true.


Let λk > λj . Then we have uk < uj . Let ∆Mkj(A) = Mk(A)−Mj(A). Then

∆M ′kj(A) = M ′

k(A)−M ′j(A) (B.12)

=uk − uj

(Auk + 1)2(Auj + 1)2·[ N∑n=1

Abn(A3Cnukuj −Auk −Auj −ACn − 2)

(ACn + 1)2

](B.13)


0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.05

0.1

0.15

0.2

0.25

D

Valu

e

Figure B.2: [Mj(aj) − Mj(aj + D)] − [Mi(ai) − Mi(ai + D)] versus D (curves) and Mj(aj) − Mi(ai)(dots).

=uk − uj

(Auk + 1)2(Auj + 1)2·[ N∑n=1

A2bnuj(A2Cnuk − 1)

(ACn + 1)2−

N∑n=1

Abn(Auk +ACn + 2)

(ACn + 1)2

]. (B.14)

Note that∑N

n=1A2bnuj(A

2Cnuk−1)(ACn+1)2 = A2uj

∑Nn=1

bn(A2Cnuk−1)

(ACn+1)2 < 0 on [0, ak) (see (B.9)), and uk < uj .

Consequently, ∆M ′kj(A) > 0. Property (M-4) is proved.

B.9 Numerical Verification of Properties (M-2’) and (M-5’) of

Mk(Ak) in the Multiple-MCS Case

Fig. B.2 visualizes a typical example in our numerical verification of Properties (M-2’) and (M-5’). In

this figure, we narrow the parameter sets to T = 10−1, 10−0.5, 1, 100.5, 101, U = 101, 101.5, . . . , 103,and Γ = 4. For all T1, T2, T3 ∈ T (T1 < T2 < T3), ∀µ/λi, µ/λj ∈ U (µ/λi > µ/λj), and ∀γ ∈ Γ, we plot

[Mj(aj)− Mj(aj +D)]− [Mi(ai)− Mi(ai +D)] versus D, and values of Mj(aj)− Mi(ai). Consequently,

100 curves and points are plotted in Fig. B.2. The curves and points are all above zero, illustrating the

correctness of Properties (M-2’) and (M-5’) under the range of parameter settings that are tested. The

results for other parameter setting are similar and are omitted to avoid redundancy.

Appendix C

Derivations and Proofs of Theorems

in Chapter 6

C.1 Proof of Theorem 11

Proof. In this proof, the tier-k BS serving the reference UE is referred to as the reference BS. Without

loss of generality, we assume the reference BS is located at R0 = (R0, 0). Note that because the reference

UE receives the highest biased power level from the reference BS, there are no tier-j BSs, ∀j ∈ K, located

within B(0, R0

βkj), where B(x, r) denotes the disk region centered at x with radius r, and B(x, r) denotes

R2\B(x, r).Let x0 = (x0, y0) denote the position of some tier-j BS (other than the reference BS if j = k). Let

T (R0,x0, βkj) denote the curve satisfying the following condition:

T (R0,x0, βkj) =

(x, y)

∣∣∣∣∣ PkBk

((x−R0)2 + y2)γ/2

=PjBj

((x− x0)2 + (y − y0)2)γ/2

. (C.1)

Note that 0 ∈ T(2)kj (∆d) is equivalent to the event the distance from 0 to curve T (R0,x0, βkj) is less

than ∆d.

In the following, we discuss three cases respectively: βkj > 1, βkj = 1, and βkj < 1.

Case 1: βkj > 1.

In this case, we have

T (R0,x0, βkj) =

(x, y)

∣∣∣∣∣[x−

(β2kjx0 −R0

β2kj − 1

)]2+

[y −

β2kjy0

β2kj − 1

]2=

β2kj(R

20 + x2

0 + y20 − 2x0R0)

(β2kj − 1)2

,

(C.2)

which is a circle centered at(

β2kjx0−R0

β2kj−1

,β2kjy0

β2kj−1

)with radius

βkj

√(R2

0+x20+y2

0−2x0R0)

(β2kj−1)

. Thus, the distance

from 0 to the trace T (R0,x0, βkj) is

d(R0,x0, βkj) =

∣∣∣∣∣∣√(β2

kjx0 −R0)2 + (β2kjy0)

2 −√β2kj(R

20 + x2

0 + y20 − 2x0R0)

(β2kj − 1)

∣∣∣∣∣∣ . (C.3)

106

Appendix C. Derivations and Proofs of Theorems in Chapter 6 107

Thus, 0 ∈ T(2)kj (∆d) iff d(R0,x0, βkj) < ∆d, or equivalently, x0 ∈ Skj(∆d), where

Skj(∆d) =x0

∣∣∣d(R0,x0, βkj) < ∆d. (C.4)

Mathematical manipulations1 of (C.4) lead to

Skj(∆d) =

(x0, y0)

∣∣∣∣∣∣∣∣∣(x2

0 + y20)−R2

0

β2kj

∣∣∣∣ < ∆d

β2kj

· (C.5)

√2(β4kj + β2

kj

)(x2

0 + y20)− 8β2kjx0R0 + 2(β2

kj + 1)R20 +O(∆d2)

.

By converting (x0, y0) into polar coordinates (r, θ), (C.5) becomes

Skj(∆d) =

(r, θ)

∣∣∣∣∣∣∣∣∣r2 − R2

0

β2kj

∣∣∣∣ < ∆d

β2kj

· (C.6)

√2(β4kj + β2

kj

)r2 − 8β2

kjR0r cos(θ) + 2(β2kj + 1)R2

0 +O(∆d2)

.

Note that there are no tier-j BSs located inside B(0, R0

βkj). Let Skj(∆d) = Skj(∆d)

∩B(0, R0/βkj). As

a result, 0 ∈ T(2)kj (∆d) iff x0 ∈ Skj(∆d), where

Skj(∆d) =

(r, θ)

∣∣∣∣∣r ≥ R0

βkjand

∣∣∣∣r2 − R20

β2kj

∣∣∣∣ < ∆d

β2kj

· (C.7)

√2(β4kj + β2

kj

)r2 − 8β2


0 +O(∆d2)

.

Fig. C.1 shows the region Skj(∆d). The inner circle, middle circle, and outer circle showR2

0

β2kj

−

r2 = ∆dβ2kj

·√2(β4kj + β2

kj

)r2 − 8β2


0 +O(∆d2), r = R0

βkj, and r2 − R2

0

β2kj

=

∆dβ2kj

·√2(β4kj + β2

kj

)r2 − 8β2


0 +O(∆d2), respectively. Therefore, Skj(∆d)

corresponds to the shaded area shown in Fig. C.1. Given an angular coordinate θ, the difference be-

tween the radial coordinates of the outer circle and the middle circle is defined as ∆Dkj(θ) (i.e., the

thickness of Skj(∆d) at angular coordinate θ). We can observe that ∆Dkj(θ) is in the scale of O(∆d),

and thus ∀(r, θ) ∈ Skj(∆d), r = R0

βkj+O(∆d). Substituting it into (C.7) gives,

Skj(∆d) =

(r, θ)

∣∣∣∣∣r ≥ R0

βkjand

∣∣∣∣r2 − R20

β2kj

∣∣∣∣ < (C.8)

∆d

β2kj

·[2(β4kj + β2

kj

)( R0

βkj+O(∆d)

)2

− 8β2kjR0

(R0

βkj+O(∆d)

)cos(θ) + 2(β2

kj + 1)R20 +O(∆d2)

]1/2,

1Let A = (β2kjx0−R0)2+(β2

kjy0)2, B = β2

kj(R20+x2

0+y20 −2x0R0), and C = β2kj −1.

∣∣∣√A−√

BC

∣∣∣ < ∆d is equivalent to

|A−B| < ∆d√2AC2 + 2BC2 − C4∆d2, i.e., |A−B| < ∆d

√2AC2 + 2BC2 +O(∆d2), which is then equivalent to (C.5).

O represents big O notation.


−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

−0.4

−0.2

0

0.2

0.4

0.6

Skj(∆d)

θ′

r = R0/βkj

∆Dkj(θ′)

−r2 +R2

0

β2

kj

= ∆dβ2

kj

√

2(β4

kj + β2

kj )r2 − 8β2

kj R0r cos(θ) + 2(β2

kj + 1)R20

+O(∆d2)

r2−

R2

0

β2

kj

= ∆dβ2

kj

√

2(β4

kj + β2

kj )r2 − 8β2

kj R0r cos(θ) + 2(β2

kj + 1)R20

+O(∆d2)

Figure C.1: The region (shaded part) of Skj(∆d).

which leads to,

Skj(∆d) =

(r, θ)

∣∣∣∣∣r ≥ R0

βkjand

∣∣∣∣r2 − R20

β2kj

∣∣∣∣ <2∆dR0

β2kj

√(β2kj + 1

)− 2βkj cos(θ) +O(∆d2)

. (C.9)

The area of Skj(∆d) is

|Skj(∆d)| =2

∫ π

0

∫ √R2

0β2kj

+2∆dR0

β2kj

√(β2

kj+1)−2βkj cos(θ)+O(∆d2)

R0βkj

rdrdθ (C.10)

=2∆dR0

β2kj

∫ π

0

√(β2kj + 1

)− 2βkj cos(θ)dθ +O(∆d2).

Given the reference UE and BS, it can be shown that Φj2 is a PPP with intensity 0 in B

(0, R0

βkj

)and

intensity λj in B (0, R0/βkj) [37]. Because P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)is equal to the probability

that there is at least one point of Φj in Skj(∆d) (i.e., some x0 in Skj(∆d)), we have

P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)(C.11)

=1− exp (−λj |Skj(∆d)|)

=1− exp(−2λj∆dR0F(βkj) +O(∆d2)

),

which completes the proof of Case 1.

Case 2: βkj < 1. The proof is similar to that of Case 1.

Case 3: βkj = 1.

In this case, we have

T (R0,x0, 1) =

(x, y)

∣∣∣∣y0 (y − y02

)= −(x0 −R0)

(x− x0 +R0

2

), (C.12)

2If k = j, it is the reduced Palm point process [12] corresponding to all tier-k BSs other than the reference BS.


which is a line. Thus, the distance from 0 to T (R0,x0, 1) is

d(x0,R0, 1) =

∣∣∣y20

2 − (R0−x0)(R0+x0)2

∣∣∣√(R0 − x0)2 + y20

. (C.13)

Consequently, similar to (C.4), 0 ∈ T(2)kj (∆d) iff d(x0,R0, 1) < ∆d, or equivalently, x0 ∈ Skj(∆d),

where

Skj(∆d) =

(x0, y0)

∣∣∣∣∣∣∣∣∣y2

0

2 − (R0−x0)(R0+x0)2

∣∣∣√(R0 − x0)2 + y20

< ∆d

. (C.14)

After converting (x0, y0) into polar coordinate (r, θ),

Skj(∆d) =

(r, θ)

∣∣∣∣∣∣r2 −R20

∣∣ < 2∆d√R2

0 + r2 − 2R0r cos θ

, (C.15)

which is a special case of (C.6) with βkj = 1. Thus, following the same steps as (C.7)-(C.11), we can

still derive (6.10), which completes the proof of Case 3.


Proof. (a) Let Ei denote the event that there is at least one tier-i BS located in Ski(∆d). Then

P(0 ∈ T(2)(∆d)|R = R0, tier = k

)=1− P

(E1

∩E2

∩. . .∩

EK

∣∣R = R0, tier = k)

=1− exp

(−

K∑i=1

|Ski(∆d)|λi

)

=1− exp

(−

K∑i=1

2λi∆dR0F(βki) +O(∆d2)

)

=

K∑i=1

2λi∆dR0F(βki) +O(∆d2). (C.16)

Furthermore, according to the results in [12], the probability density function of the distance between

the reference UE and the reference BS is

fk(R0|tier = k) =2πλk

AkR0 exp

(−πR2

0

K∑i=1

λiβ2ik

). (C.17)

Also, we have P(tier = k) = Ak. Thus

P(0 ∈ T(2)(∆d)

)=

K∑k=1

∫ ∞

0

P(0 ∈ T(2)(∆d)|R = R0, tier = k)fk(R0|tier = k)P(tier = k)dR0


=

K∑k=1

∫ ∞

0

2πλkR0 exp

(−πR2

0

K∑i=1

λiβ2ik

)(K∑i=1

2λi∆dR0F(βki) +O(∆d2)

)dR0

=

K∑k=1

λk

(∑Ki=1 λi∆dF(βki) +O(∆d2)

)(∑K

i=1 λiβ2ik

) 32

, (C.18)

which completes the proof of (a).

(b)

Similar to (C.18), if k = j, we have

P(0 ∈ T

(2)kj (∆d)

)=

∫ ∞

0

P(0 ∈ T(2)kj (∆d)|R = R0, tier = k)fk(R0|tier = k)P(tier = k)dR0

+

∫ ∞

0

P(0 ∈ T(2)kj (∆d)|R = R0, tier = j)fj(R0|tier = j)P(tier = j)dR0

=λk

(λj∆dF(βkj) +O(∆d2)

)(∑Ki=1 λiβ2

ik

) 32

+λj

(λk∆dF(βjk) +O(∆d2)

)(∑Ki=1 λiβ2

ij

) 32

. (C.19)

Otherwise, if k = j, we have

P(0 ∈ T

(2)kk (∆d)

)=

λk

(λk∆dF(1) +O(∆d2)

)(∑Ki=1 λiβ2

ik

) 32

, (C.20)

which completes the proof of (b).

C.3 Derivation of P(Rk > R0) when k ∈ KC

In this appendix, we aim to compute the probability

P(Rk > R0) = P(Φk

∩B(0, R0) = ∅). (C.21)

Based on the definition in Section 6.2.1, Φk =∪

x∈ΘkΩk(x), where Θk is the parent point process,

and Ωk(x) is a child point process. In order to investigate the event Φk

∩B(0, R0) = ∅, we construct a

point process Θ′k, which is a thinned point process on Θk as follows

Θ′k = x ∈ Θk|Ωk(x)

∩B(0, R0) = ∅. (C.22)

By its definition, we have Θ′k = ∅ iff

(∪x∈Θk

Ωk(x))∩

B(0, R0) = ∅, or equivalently, Φk

∩B(0, R0) = ∅.

Thus, we have

P(Φk

∩B(0, R0) = ∅) = P(Θ′

k = ∅). (C.23)

Because Ωk(x) is a PPP with intensity νk in B(x, Rk), we have

pk(x) , P(Ωk(x)∩

B(0, R0) = ∅) =


1− exp(−νkC(R0, Rk, |x|)), if |Rk −R0| ≤ |x| < R0 +Rk,

1− exp(−νkπR2k)), if |x| < R0 −Rk and R0 ≥ Rk,

1− exp(−νkπR20)), if |x| < Rk −R0 and R0 < Rk,

0, if |x| ≥ Rk +R0.

(C.24)

Note that since Ωk(x) are independent with respect to different x, the events Ωk(x)∩B(0, R0) = ∅ are

also independent with respect to different x. As a result, Θ′k is an independently thinned point process

on Θk, with thinning probability pk(x) at x. Thus, Θ′k is a PPP with intensity pk(x)µk at x, and we

have

P(Θ′k = ∅) = exp

(∫R2

−µkp(x)dx

). (C.25)

Substituting (C.24) into (C.25), and transforming to polar coordinate, we derive (6.48). Note that

because p(x) is 0 in the region |x| ≥ Rk +R0, the value∫R2 −µkp(x)dx in (C.25) is finite.


Proof. In this proof, we study the probability P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

). Without loss of

generality, we assume the reference BS is located at R0 = (R0, 0). Then, following the similar steps in

(C.2)-(C.9) in the proof of Theorem 11, 0 ∈ T(2)kj (∆d) is equivalent to the event that there is at least

one point of Φj in the ring region Skj(∆d). Note that the ring region here is still expressed in (C.9) and

shown in Fig. C.1. Given the reference UE and BS, Φj3 is a PPP with intensity 0 in B

(0, R0

βkj

)and

intensity λj in B (0, R0/βkj) [37]. Because P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)is equal to the probability

that there is at least one point of Φj in Skj(∆d) (i.e., some x0 in Skj(∆d)), we have

P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)(C.26)

=1− exp (−λj |Skj(∆d)|)

=1− exp(−2λj∆dR0F(βkj) +O(∆d2)

)=2λj∆dR0F(βkj) +O(∆d2),

which completes the proof.



(2)kj (∆d)|R = R0, tier = k

). Without loss of

generality, we assume the reference BS is located at R0 = (R0, 0). Then, following the similar steps in

(C.2)-(C.9) in the proof of Theorem 11, 0 ∈ T(2)kj (∆d) is equivalent to the event that there is at least

one point of Φj in the ring region Skj(∆d), given that no point of Φj is in B(0, R0/βkj). Note that the

ring region here is still expressed in (C.9) and shown in Fig. C.1.

3If k = j, it is the reduced Palm point process [12] corresponding to all tier-k BSs other than reference BS.


As labeled in Fig. C.1, we also define ∆Dkj(θ) as the thickness of the “ring” region of Skj(∆d) at

angular coordinate θ. We have

∆Dkj(θ) =

√R2

0

β2kj

+2∆dR0

β2kj

√(β2kj + 1

)− 2βkj cos(θ) +O(∆d2)− R0

βkj(C.27)

=∆d

√β2kj + 1− 2βkj cos(θ)

βkj+O(∆d2).

Then, we have

P(0 ∈ T

(2)kj (∆d)|R = R0, tier = k

)(C.28)

=P(Φj

∩Skj(∆d) = ∅

∣∣∣Φj

∩B(0, R0/βkj) = ∅

)(C.29)

=1

π

∫ π

0

ccdfRj

(R0

βkj

)− ccdfRj

(R0

βkj+∆Dkj(θ)

)+O(∆d2)

ccdfRj (R0/βkj)dθ

=1

π

∫ π

0

∆Dkj(θ)dθpdfRj

(R0/βkj)

ccdfRj (R0/βkj)+O(∆d2) (C.30)

=1

π

∫ π

0

∆d

√β2kj + 1− 2βkj cos(θ)

βkjdθ

pdfRj(R0/βkj)

ccdfRj (R0/βkj)+O(∆d2)

=1

πF(βkj)βkj∆d

pdfRj(R0/βkj)

ccdfRj (R0/βkj)+O(∆d2), (C.31)

which completes the proof.

Note that because the location of the reference tier-k BS is given, whether tier-k is a P-tier or a

C-tier does not influence this proof. The proof is true for all k ∈ K.



(2)kk,in(∆d)|R = R0, tier = k

). Without loss of

generality, we assume the reference BS is located at xBS = R0 = (R0, 0). C(R0) is the cluster center

of the reference BS, and Ωk(C(R0)) is the cluster that the reference BS belongs to. Note that C(R0) is

surely located in B(R0, Rk) as its distance to the reference BS cannot be larger than Rk.

Following the similar steps in (C.2)-(C.9) in the proof of Theorem 15, 0 ∈ T(2)kk,in(∆d) is equivalent

to the event that there is at least one point of Ωk(C(R0)) in the ring region Skk(∆d), which is defined

in (C.9) with βkk = 1. Thus, we have

P(0 ∈ T(2)kk,in(∆d)|xBS = R0, tier = k) (C.32)

=P(Ωk(C(R0))

∩Skk(∆d) = ∅

∣∣xBS = R0, tier = k).

The event Ωk(C(R0))∩

Skk(∆d) = ∅ depends on the location of cluster center C(R0). Then we have

P(Ωk(C(R0))

∩Skk(∆d) = ∅

∣∣xBS = R0, tier = k)

(C.33)


=

∫B(R0,Rk)

P[Ωk(C(R0))

∩Skk(∆d) = ∅

∣∣C(R0) = x]pdfcen,k(x)dx

=

∫B(R0,Rk)

(1− exp

(−νk

∣∣∣B(x, Rk)∩

Skk(∆d)∣∣∣))pdfcen,k(x)dx

=

∫B(R0,Rk)

νk

∣∣∣B(x, Rk)∩

Skk(∆d)∣∣∣ pdfcen,k(x)dx+O(∆d2),

where pdfcen,k(x) is defined as the pdf of the location of the cluster center C(R0), given the reference

UE is associated to the reference tier-k BS located at R0.

The distribution pdfcen,k(x) is derived in the following lemma:

Lemma 7. Given that the reference UE located at 0 is associated with the reference tier-k BS located

at R0 = (R0, 0), the probability density function of the location of the cluster center C(R0) is given in

(C.34), where (r, θ) is the polar coordinate of the cluster center.

pdfcen,k(r, θ) = (C.34)

exp(−C(R0,Rk,r)νk)∫R0+RkR0−Rk

2θm(r′,Rk,R0)r′ exp(−C(R0,Rk,r′)νk)dr′, if R0 ≥ Rk and R0 −Rk ≤ r < R0 +Rk

and − θm(r,Rk, R0) ≤ θ ≤ θm(r,Rk, R0),

exp(−C(R0,Rk,r)νk)

π(Rk−R0)2 exp(−πR20νk)+

∫Rk+R0Rk−R0

2θm(r′,Rk,R0)r′ exp(−C(R0,Rk,r′)νk)dr′, if R0 < Rk and Rk −R0 ≤ r < Rk +R0,

exp(−πR20νk)

π(Rk−R0)2 exp(−πR20νk)+

∫Rk+R0Rk−R0

2θm(r′,Rk,R0)r′ exp(−C(R0,Rk,r′)νk)dr′, if R0 < Rk and r < Rk −R0,

0, Otherwise.


Based on Lemma 7, we derive (C.32) in two cases: R0 ≥ Rk and R0 < Rk.

Case one R0 ≥ Rk. After transforming to polar coordinate, (C.33) becomes

P(0 ∈ T(2)kk,in(∆d)|xBS = R0, tier = k)

=

∫ R0+Rk

R0−Rk

∫ θm(r,Rk,R0)

−θm(r,Rk,R0)

r · pdfcen,k(r, θ) · νk ·∣∣∣B((r, θ), Rk)

∩Skk(∆d)

∣∣∣dθdr +O(∆d2). (C.35)

Given (r, θ) as the polar coordinate of the cluster center of Ωk(C(R0)), we derive the area of over-

lapping region of B((r, θ), Rk) and Skk(∆d) as follows:∣∣∣B((r, θ), Rk)∩

Skk(∆d)∣∣∣ (C.36)

=∆dR0

∫ θ+θm(R0,Rk,r)

θ−θm(R0,Rk,r)

√2− 2 cos(θ′)dθ′ +O(∆d2)

=∆dR0

[− 4 cos

(θ + θm(R0, Rk, r)

2

)− 4 cos

(θ − θm(R0, Rk, r)

2

)+ 8

]+O(∆d2).

Substituting (C.36) into (C.35), we obtain Gk(R0)∆d+O(∆d2), where Gk(R0) is expressed as (6.57).

Case Two R0 < Rk. After transforming to polar coordinate, (C.33) becomes

P(0 ∈ T(2)kk,in(∆d)|xBS = R0, tier = k)

=

∫ Rk+R0

Rk−R0

∫ θm(r,Rk,R0)

−θm(r,Rk,R0)


∩Skk(∆d)

∣∣∣dθdr


+

∫ Rk−R0

0

∫ π

−π


∩Skk(∆d)

∣∣∣dθdr +O(∆d2). (C.37)

Given (r, θ) as the polar coordinate of the cluster center of Ωk(C(R0)), we derive the area of over-

lapping region of B((r, θ), Rk) and Skk(∆d). If r ≥ Rk −R0, we have∣∣∣B((r, θ), Rk)∩

Skk(∆d)∣∣∣

=∆dR0

∫ θ+θm(R0,Rk,r)

θ−θm(R0,Rk,r)

√2− 2 cos(θ′)dθ′ +O(∆d2)

=∆dR0

[− 4 cos

(θ + θm(R0, Rk, r)

2

)− 4 cos

(θ − θm(R0, Rk, r)

2

)+ 8

]+O(∆d2). (C.38)

If r < Rk −R0, Skk(∆d) is covered by B((r, θ), Rk), then we have∣∣∣B((r, θ), Rk)∩

Skk(∆d)∣∣∣

=∆dR0

∫ π

−π

√2− 2 cos(θ′)dθ′ +O(∆d2)

=8∆dR0 +O(∆d2). (C.39)

Substituting (C.38) and (C.39) into (C.37), we obtainGk(R0)∆d+O(∆d2), whereGk(R0) is expressed

as (6.58).

C.7 Proof of Lemma 7

Proof. Because the reference UE located at 0 is associated with the reference tier-k BS located at

R0, it is true that Ωk(C(xBS))∩B(0, R0) = ∅, as R0 is the closest point to 0 in the cluster. Let

Xc = C(xBS) denote the location of the cluster center. We aim to derive the pdf of Xc given xBS = R0

and Ωk(C(xBS))∩B(0, R0) = ∅. For simplicity, we use the term Ω0 to represent Ωk(C(xBS)) in the rest

of this proof.

pdfcen,k(x) =pdfXc

[x∣∣xBS = R0,Ω0

∩B(0, R0) = ∅

](C.40)

=pdfXc

[x,Ω0

∩B(0, R0) = ∅

∣∣∣xBS = R0

]P [Ω0

∩B(0, R0) = ∅|xBS = R0]

=pdfXc

[x∣∣xBS = R0]P[Ω0

∩B(0, R0) = ∅|Xc = x

]∫B(R0,Rk)

pdfXc[x′|xBS = R0]P [Ω0

∩B(0, R0) = ∅|Xc = x′] dx′ .

Consider a single BS cluster. BSs are homogeneously Poisson distributed in the disk region centered

at the cluster center with radius Rk. If given the location of one BS, it can be shown that the cluster

center is uniformly distributed in the disk region centered at the BS with radius Rk. pdfXc[x|xBS = R0]

follows a uniform distribution in the disk region centered at xBS with radius Rk:

pdfXc[x|xBS = R0] =

1

πR2k

, if x ∈ B(R0, Rk). (C.41)


Given the cluster center x, Ω0 is a PPP in the disk region B(x, Rk) with intensity νk. We have

P[Ω0

∩B(0, R0) = ∅|Xc = x

]=exp

(−νk|B(0, R0)

∩B(x, Rk)|

)(C.42)

=

exp

(−νkπ(min(R0, Rk))

2), if |x| < |Rk −R0|,

exp (−νkC(R0, Rk, |x|)) , if |Rk −R0| ≤ |x| < Rk +R0,

1, if |x| ≥ Rk +R0.

(C.43)

Substituting (C.41) and (C.43) into (C.40), and transforming into polar coordinate, we obtain (C.34).

Note that when a transformation to polar coordinates is made,∫B(R0,Rk)

(·)dx is converted to∫ R0+Rk

R0−Rk∫ θm(r,Rk,R0)

−θm(r,Rk,R0)r(·)dθdr if R0 ≥ Rk, and

∫ Rk+R0

Rk−R0

∫ θm(r,Rk,R0)

−θm(r,Rk,R0)r(·)dθdr +

∫ Rk−R0

0

∫ π

−πr(·)dθdr if R0 <

Rk.

Bibliography

[1] X. Lagrange, “Multitier cell design,” IEEE Communications Magazine, vol. 35, no. 8, pp. 60–64,

Aug. 1997.

[2] V. Chandrasekhar, J. Andrews, and A. Gatherer, “Femtocell networks: a survey,” IEEE Commu-

nications Magazine, vol. 46, no. 9, pp. 59–67, Sep. 2008.

[3] A. Doufexi, E. Tameh, A. Nix, S. Armour, and A. Molina, “Hotspot wireless LANs to enhance

the performance of 3G and beyond cellular networks,” IEEE Communications Magazine, vol. 41,

no. 7, pp. 58–65, Jul. 2003.

[4] J. Sydir and R. Taori, “An evolved cellular system architecture incorporating relay stations,” IEEE

Communications Magazine, vol. 47, no. 6, pp. 115 – 121, Jun. 2009.

[5] A. Damnjanovic, J. Montojo, Y. Wei, T. Ji, T. Luo, M. Vajapeyam, T. Yoo, O. Song, and D. Mal-

ladi, “A survey on 3GPP heterogeneous networks,” IEEE Wireless Communications, vol. 18, no. 3,

pp. 10–21, Jun. 2011.

[6] A. Ghosh, N. Mangalvedhe, R. Ratasuk, B. Mondal, M. Cudak, E. Visotsky, T. A. Thomas, J. G.

Andrews, P. Xia, H. S. Jo, H. S. Dhillon, and T. D. Novlan, “Heterogeneous cellular networks:

From theory to practice,” IEEE Communications Magazine, vol. 50, no. 6, pp. 54–64, Jun. 2012.

[7] “AT&T 3G Microcell,” http://www.att.com/standalone/3gmicrocell.

[8] “AT&T Wi-Fi hot spot locations,” https://www.att.com/maps/wifi.

[9] “Performance of eICIC with control channel coverage limitation,” 3GPP Standard Contribution,

R1-103264, NTT DOCOMO, May 2010.

[10] R. Madan, J. Borran, A. Sampath, N. Bhushan, A. Khandekar, and T. Ji, “Cell association and

interference coordination in heterogeneous LTE-A cellular networks,” IEEE Journal on Selected

Areas in Communications, vol. 28, no. 9, pp. 1479–1489, Dec. 2010.

[11] D. Lopez-Perez, X. Chu, and I. Guvenc, “On the expanded region of picocells in heterogeneous

networks,” IEEE Journal of Selected Topics in Signal Processing, vol. 6, no. 3, pp. 281–294, Jun.

2012.

[12] H.-S. Jo, Y. J. Sang, P. Xia, and J. G. Andrews, “Heterogeneous cellular networks with flexible cell

association: a comprehensive downlink SINR analysis,” IEEE Trans. on Wireless Communications,

vol. 11, no. 10, pp. 3484–3495, Oct. 2012.

116

Bibliography 117

[13] S. Singh, H. Dhillon, and J. G. Andrews, “Offloading in heterogeneous networks: Modeling, analy-

sis, and design insights,” IEEE Trans. on Wireless Communications, vol. 12, no. 5, pp. 2484–2497,

May 2013.

[14] M. Stemm and R. H. Katz, “Vertical handoffs in wireless overlay networks,” ACM/Springer Mobile

Networks and Applications, vol. 3, no. 4, pp. 335–350, Jan. 1998.

[15] C. Bettstetter, G. Resta, and P. Santi, “The node distribution of the random waypoint mobility

model for wireless ad hoc networks,” IEEE Trans. on Mobile Computing, vol. 2, no. 3, pp. 257 –

269, Jul. - Sept. 2003.

[16] F. Ashtiani, J. Salehi, and M. Aref, “Mobility modeling and analytical solution for spatial traffic

distribution in wireless multimedia networks,” IEEE Journal on Selected Areas in Communica-

tions, vol. 21, no. 10, pp. 1699 – 1709, Dec. 2003.

[17] G. Mohimani, F. Ashtiani, A. Javanmard, and M. Hamdi, “Mobility modeling, spatial traffic

distribution, and probability of connectivity for sparse and dense vehicular ad hoc networks,”

IEEE Trans. on Vehicular Technology, vol. 58, no. 4, pp. 1998 – 2007, May 2009.

[18] A. Ghosh, R. Jana, V. Ramaswami, J. Rowland, and N. Shankaranarayanan, “Modeling and

characterization of large-scale Wi-Fi traffic in public hot-spots,” in Proc. of IEEE INFOCOM,

Shanghai, China, Apr. 2011.

[19] Y. Chen, J. Kurose, and D. Towsley, “A mixed queueing network model of mobility in a campus

wireless network,” in Proc. of IEEE INFOCOM Mini-Conference, Orlando, FL, Mar. 2012.

[20] V. B. Iversen, Teletraffic Engineering Handbook. ITU-D SG 2/16 & ITC Draft 2001-06-20, 2001.

[21] P. Heegaard, “Empirical observations of traffic patterns in mobile and IP telephony,” in Next

Generation Teletraffic and Wired/Wireless Advanced Networking, ser. Lecture Notes in Computer

Science. Springer Berlin / Heidelberg, 2007, vol. 4712, pp. 26–37.

[22] Z. Jako and G. Jeney, “Outage probability in poisson-cluster-based lte two-tier femtocell networks,”

Wiley Wireless Communications and Mobile Computing, Jun. 2014.

[23] K. Gulati, B. Evans, J. Andrews, and K. Tinsley, “Statistics of co-channel interference in a field

of poisson and poisson-poisson clustered interferers,” IEEE Trans. on Signal Processing, vol. 58,

no. 12, pp. 6207–6222, Dec. 2010.

[24] Y. Zhong andW. Zhang, “Multi-channel hybrid access femtocells: A stochastic geometric analysis,”

IEEE Trans. on Communications, vol. 61, no. 7, pp. 3016–3026, Jul. 2013.

[25] F. P. Kelly, Reversibility and Stochastic Networks. Cambridge, 1979.

[26] D. Y. Burman, “Insensitivity in queueing systems,” Advances in Applied Probability, vol. 13, no. 4,

pp. 846 – 859, Dec. 1981.

[27] P. Whittle, “Partial balance and insensitivity,” Journal of Applied Probability, vol. 22, no. 1, pp.

168 – 176, Mar. 1985.

Bibliography 118

[28] T. Bonald, “Insensitive queueing models for communication networks,” in Proc. of International

Conference on Performance Evaluation Methodologies and Tools (VALUETOOLS), Pisa, Italy,

Oct. 2006.

[29] M. M. Zonoozi and P. Dassanayake, “User mobility modeling and characterization of mobility

patterns,” IEEE Journal on Selected Areas in Communications, vol. 15, no. 7, pp. 1239–1252, Sep.

1997.

[30] A. Anpalagan and I. Katzela, “Overlaid cellular system design, with cell selection criteria for mobile

wireless users,” in Proc. of IEEE Canadian Conference on Electrical and Computer Engineering,

Edmonton, Canada, May 1999.

[31] N. Shenoy and B. Hartpence, “A mobility model for cost analysis in integrated cellular/WLANs,”

in Proc. of International Conference on Computer Communications and Networks, Chicago, IL,

Oct. 2004.

[32] A. Hasib and A. Fapojuwo, “Mobility model for heterogeneous wireless networks and its application

in common radio resource management,” IET Communications, vol. 2, no. 9, pp. 1186–1195, Oct.

2008.

[33] X. Lin, R. Ganti, P. Fleming, and J. Andrews, “Towards understanding the fundamentals of

mobility in cellular networks,” IEEE Trans. on Wireless Communications, vol. 12, no. 4, pp.

1686–1698, Apr. 2013.

[34] J. G. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage and rate in cellular

networks,” IEEE Trans. on Communications, vol. 59, no. 11, pp. 3122–3134, Nov. 2011.

[35] H. Dhillon, R. Ganti, F. Baccelli, and J. G. Andrews, “Modeling and analysis of K-tier downlink

heterogeneous cellular networks,” IEEE Journal on Selected Areas in Communications, vol. 30,

no. 3, pp. 550–560, Apr. 2012.

[36] W. C. Cheung, T. Q. S. Quek, and M. Kountouris, “Throughput optimization, spectrum allo-

cation, and access control in two-tier femtocell networks,” IEEE Journal on Selected Areas in

Communications, vol. 30, no. 3, pp. 561–574, Apr. 2012.

[37] F. Baccelli and B. Blaszczyszyn, “Stochastic geometry and wireless networks, volume 1: Theory,”

Foundations and Trends in Networking, vol. 3, no. 3-4, pp. 249 – 449, 2009.

[38] ——, “Stochastic geometry and wireless networks, volume 2: Applications,” Foundations and

Trends in Networking, vol. 4, no. 1-2, pp. 1–312, 2009.

[39] D. Stoyan, W. Kendall, and J. Mecke, Stochastic Geometry and Its Applications, 2nd ed. Wiley,

1995.

[40] J. Andrews, R. Ganti, M. Haenggi, N. Jindal, and S. Weber, “A primer on spatial modeling and

analysis in wireless networks,” IEEE Communications Magazine, vol. 48, no. 11, pp. 156–163, Nov.

2010.

[41] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic geometry

and random graphs for the analysis and design of wireless networks,” IEEE Journal on Selected

Areas in Communications, vol. 27, no. 7, pp. 1029 – 1046, Sep. 2009.

Bibliography 119

[42] H. Dhillon, R. Ganti, F. Baccelli, and J. Andrews, “A tractable framework for coverage and outage

in heterogeneous cellular networks,” in Information Theory and Applications Workshop, San Diego,

CA, Feb. 2011.

[43] H. Keeler, B. Blaszczyszyn, and M. Karray, “SINR-based k-coverage probability in cellular net-

works with arbitrary shadowing,” in Proc. of IEEE ISIT, Istanbul, Turkey, Jul. 2013, pp. 1167–

1171.

[44] Y. Kim, S. Lee, and D. Hong, “Performance analysis of two-tier femtocell networks with outage

constraints,” IEEE Trans. on Wireless Communications, vol. 9, no. 9, pp. 2695–2700, Sep. 2010.

[45] H. S. Dhillon, R. K. Ganti, and J. G. Andrews, “Load-aware heterogeneous cellular networks:

Modeling and SIR distribution,” in Proc. of IEEE GLOBECOM, Anaheim, CA, Dec. 2012, pp.

4314–4319.

[46] H.-S. Jo, P. Xia, and J. G. Andrews, “Open, closed, and shared access femtocells in the downlink,”

EURASIP Journal on Wireless Communications and Networking, vol. 2012, no. 363, Dec. 2012.

[47] S. Kishore, L. Greenstein, H. Poor, and S. Schwartz, “Uplink user capacity in a CDMA macrocell

with a hotspot microcell: exact and approximate analyses,” IEEE Trans. on Wireless Communi-

cations, vol. 2, no. 2, pp. 364–374, Mar. 2003.

[48] ——, “Uplink user capacity in a multicell CDMA system with hotspot microcells,” IEEE Trans.

on Wireless Communications, vol. 5, no. 6, pp. 1333–1342, June 2006.

[49] X. An and F. Pianese, “Understanding co-channel interference in LTE-based multi-tier cellular

networks,” in Proc. of ACM PE-WASUN, Paphos, Cyprus, Oct. 2012, pp. 107–112.

[50] T. D. Novlan, H. S. Dhillon, and J. G. Andrews, “Analytical modeling of uplink cellular networks,”

IEEE Trans. on Wireless Communications, vol. 12, no. 6, pp. 1536–1276, Jun. 2013.

[51] T. Novlan and J. Andrews, “Analytical evaluation of uplink fractional frequency reuse,” IEEE

Trans. on Wireless Communications, vol. 61, no. 5, pp. 2098–2108, May 2013.

[52] N. Chakchouk and B. Hamdaoui, “Uplink performance characterization and analysis of two-tier

femtocell networks,” IEEE Trans. on Vehicular Technology, vol. 61, no. 9, pp. 4057–4068, Nov.

2012.

[53] V. Chandrasekhar and J. Andrews, “Uplink capacity and interference avoidance for two-tier fem-

tocell networks,” IEEE Trans. on Wireless Communications, vol. 8, no. 7, pp. 3498–3509, Jul.

2009.

[54] W. C. Cheung, T. Q. S. Quek, and M. Kountouris, “Stochastic analysis of two-tier networks:

Effect of spectrum allocation,” in Proc. of IEEE ICASSP, Prague, Czech Republic, May 2011, pp.

2964–2967.

[55] W. Bao and B. Liang, “Uplink interference analysis for two-tier cellular networks with diverse

users under random spatial patterns,” IEEE Trans. on Wireless Communications, vol. 14, no. 3,

pp. 1252–1265, Mar. 2015.

Bibliography 120

[56] P. Xia, V. Chandrasekhar, and J. Andrews, “Open vs. closed access femtocells in the uplink,”

IEEE Trans. on Wireless Communications, vol. 9, no. 12, pp. 3798–3809, Dec. 2010.

[57] P. Tarasak, T. Q. S. Quek, and F. P. S. Chin, “Uplink timing misalignment in open and closed

access OFDMA femtocell networks,” IEEE Communications Letters, vol. 15, no. 9, pp. 926–928,

Sep. 2011.

[58] Z. Zeinalpour-Yazdi and S. Jalali, “Outage analysis of uplink two-tier networks,” IEEE Trans. on

Communications, vol. 62, no. 9, pp. 3351–3362, Sep. 2014.

[59] H. ElSawy and E. Hossain, “On stochastic geometry modeling of cellular uplink transmission with

truncated channel inversion power control,” IEEE Trans. on Wireless Communications, vol. 13,

no. 8, pp. 4454–4469, Aug. 2014.

[60] F. Tong, M. Ahmadi, and J. Pan, “Random distances associated with arbitrary triangles: A

systematic approach between two random points,” arXiv:1312.2498 [math.GM], 2013.

[61] Y. Zhuang and J. Pan, “Random distances associated with rhombuses,” arXiv:1106.1257

[math.GM], 2011.

[62] ——, “Random distances associated with hexagons,” arXiv:1106.2200 [math.GM], 2011.

[63] M. Ahmadi and J. Pan, “Random distances associated with trapezoids,” arXiv:1307.1444

[math.GM], 2013.

[64] M. Ahmadi, F. Tong, L. Zheng, and J. Pan, “Performance analysis for two-tier cellular systems

based on probabilistic distance models,” in Proc. of IEEE INFOCOM, Hong Kong, China, Apr.

2015.

[65] J.-W. Lee, R. Mazumdar, and N. Shroff, “Joint resource allocation and base-station assignment

for the downlink in CDMA networks,” IEEE/ACM Trans. on Networking, vol. 14, no. 1, pp. 1–14,

Feb. 2006.

[66] Q. Ye, B. Rong, Y. Chen, M. Al-Shalash, C. Caramanis, and J. G. Andrews, “User association for

load balancing in heterogeneous cellular networks,” IEEE Trans. on Wireless Communications,

vol. 12, no. 6, pp. 2706–2716, Jun. 2013.

[67] M. Hong and A. Garcia, “Mechanism design for base station association and resource allocation in

downlink OFDMA network,” IEEE Journal on Selected Areas in Communications, vol. 30, no. 11,

pp. 2238–2250, Dec. 2012.

[68] C. S. Chen, F. Baccelli, and L. Roullet, “Joint optimization of radio resources in small and macro

cell networks,” in Proc. of VTC (Spring), Budapest, Hungary, May 2011.

[69] K. Shen and W. Yu, “Downlink cell association optimization for heteregeneous networks via dual

coordinate descent,” in Proc. of IEEE ICASSP, Vancouver, Canada, May 2013, pp. 4779–4783.

[70] D. Fooladivanda, A. A. Daoud, and C. Rosenberg, “Joint channel allocation and user association

for heterogeneous wireless cellular networks,” in Proc. of IEEE PIMRC, Toronto, Canada, Sep.

2011, pp. 384–390.

Bibliography 121

[71] W. Bao and B. Liang, “Near-optimal spectrum allocation in multi-tier cellular networks with

random inelastic traffic,” in Proc. of IEEE ICASSP, Florence, Italy, May 2014.

[72] ——, “Radio resource allocation in heterogeneous wireless networks: a spatial-temporal perspec-

tive,” in Proc. of IEEE INFOCOM, Hong Kong, China, Apr. 2015.

[73] S. Singh and J. G. Andrews, “Joint resource partitioning and offloading in heterogeneous cellular

networks,” IEEE Trans. on Wireless Communications, vol. 13, no. 2, pp. 888 – 901, Feb. 2014.

[74] Y. Lin and W. Yu, “Optimizing user association and frequency reuse for heterogeneous network

under stochastic model,” in Proc. of IEEE Globecom, Atlanta, GA, Dec. 2013, pp. 2045–2050.

[75] ——, “Joint spectrum partition and user association in multi-tier heterogeneous networks,” in

Proc. of CISS, Princeton, NJ, Mar. 2014.

[76] Y. Lin, W. Bao, W. Yu, and B. Liang, “Optimizing user association and spectrum allocation in

HetNets: a utility perspective,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 6,

pp. 1025–1039, Jun. 2015.

[77] R. Serfozo, Introduction to Stochastic Networks. Springer, 1999.

[78] W. Bao and B. Liang, “On the insensitivity of user distribution in multicell networks under general

mobility and session patterns,” in Proc. of IEEE INFOCOM, Turin, Italy, Apr. 2013.

[79] ——, “Insensitivity of user distribution in multicell networks under general mobility and session

patterns,” IEEE Trans. on Wireless Communications, vol. 12, no. 12, pp. 6244–6254, Dec. 2013.

[80] ——, “Understanding the benefits of open access in femtocell networks: Stochastic geometric

analysis in the uplink,” in Proc. of ACM MSWiM, Barcelona, Spain, Nov. 2013.

[81] ——, “Stochastic analysis of uplink interference in two-tier femtocell networks: Open versus closed

access,” to appear in IEEE Trans. on Wireless Communications, 2015.

[82] ——, “Structured spectrum allocation and user association in heterogeneous cellular networks,”

in Proc. of IEEE INFOCOM, Toronto, Canada, Apr.-May 2014.

[83] ——, “Rate maximization through structured spectrum allocation and user association in hetero-

geneous cellular networks,” to appear in IEEE Trans. on Communications, 2015.

[84] ——, “Handoff rate analysis in heterogeneous cellular networks: a stochastic geometric approach,”

in Proc. of ACM MSWiM, Montreal, Canada, Sep. 2014.

[85] ——, “Stochastic geometric analysis of user mobility in heterogeneous wireless networks,” IEEE

Journal on Selected Areas in Communications, vol. 33, no. 10, pp. 2212–2225, Oct. 2015.

[86] ——, “Handoff rate analysis in heterogeneous wireless networks with poisson and poisson cluster

patterns,” in Proc. of ACM MobiHoc, Hangzhou, China, Jun. 2015.

[87] V. Paxson and S. Floyd, “Wide area traffic: the failure of poisson modeling,” IEEE/ACM Trans.

on Networking, vol. 3, no. 3, pp. 226–244, Jun. 1995.

Bibliography 122

[88] D. Kotz, T. Henderson, I. Abyzov, and J. Yeo, “CRAWDAD

trace dartmouth/campus/snmp/fall01 (v. 2004-11-09),” Downloaded from

http://crawdad.cs.dartmouth.edu/dartmouth/campus/snmp/fall01, Nov. 2004.

[89] ——, “CRAWDAD trace dartmouth/campus/snmp/spring02 (v. 2004-11-09),” Downloaded from

http://crawdad.cs.dartmouth.edu/dartmouth/campus/snmp/spring02, Nov. 2004.

[90] ——, “CRAWDAD trace dartmouth/campus/snmp/fall0304 (v. 2004-11-09),” Downloaded from

http://crawdad.cs.dartmouth.edu/dartmouth/campus/snmp/fall0304, Nov. 2004.

[91] [Online]. Available: http://sysnet.ucsd.edu/wtd/

[92] M. Balazinska and P. Castro, “CRAWDAD data set ibm/watson (v. 2003-02-19),” Downloaded

from http://crawdad.cs.dartmouth.edu/ibm/watson, Feb. 2003.

[93] M. Lenczner, B. Gregoire, and F. Proulx, “CRAWDAD data set ilesansfil/wifidog (v. 2007-08-27),”

Downloaded from http://crawdad.cs.dartmouth.edu/ilesansfil/wifidog, Aug. 2007.

[94] D. Kotz and K. Essien, “Analysis of a campus-wide wireless network,” in Proc. of ACM Mobicom,

Atlanta, GA, Sept. 2002.

[95] T. Henderson, D. Kotz, and I. Abyzov, “The changing usage of a mature campus-wide wireless

network,” in Proc. of ACM MobiCom, Philadelphia, PA, Sept. 2004.

[96] M. Kim and D. Kotz, “Periodic properties of user mobility and access-point popularity,” Personal

and Ubiquitous Computing, vol. 11, no. 6, pp. 465 – 479, Aug. 2007.

[97] D. Knisely, T. Yoshizawa, and F. Favichia, “Standardization of femtocells in 3GPP,” IEEE Com-

munications Magazine, vol. 47, no. 9, pp. 68–75, Sep. 2009.

[98] C. C. Chan and S. Hanly, “Calculating the outage probability in a CDMA network with spatial

Poisson traffic,” IEEE Trans. on Vehicular Technology, vol. 50, no. 1, pp. 183–204, Jan. 2001.

[99] Q. Ye, M. Al-Shalash, C. Caramanis, and J. Andrews, “Resource optimization in device-to-device

cellular systems using time-frequency hopping,” IEEE Trans. on Wireless Communications, vol. 13,

no. 10, pp. 5467–5480, Oct. 2014.

[100] S. Singh, X. Zhang, and J. G. Andrews, “Joint rate and SINR coverage analysis for decoupled

uplink-downlink biased cell associations in HetNets,” arXiv:1412.1898 [cs.IT], 2014.

[101] K. Gilhousen, I. Jacobs, R. Padovani, A. Viterbi, L. Weaver, and C. Wheatley, “On the capacity

of a cellular CDMA system,” IEEE Trans. on Vehicular Technology, vol. 40, no. 2, pp. 303–312,

May 1991.

[102] A. J. Viterbi, A. M. Viterbi, K. Gilhousen, and E. Zehavi, “Soft handoff extends CDMA cell

coverage and increases reverse link capacity,” IEEE Journal on Selected Areas in Communications,

vol. 12, no. 8, pp. 1281–1288, Oct. 1994.

[103] 3GPP-TS-36.104, “Evolved universal terrestrial radio access (E-UTRA); base station (BS) radio

transmission and reception (release 12),” Dec. 2014.

Bibliography 123

[104] 3GPP-TR-36.814, “Evolved universal terrestrial radio access (E-UTRA); further advancements for

E-UTRA physical layer aspects (release 9),” Mar. 2010.

[105] 3GPP-TR-25.967, “Universal mobile telecommunications system (UMTS); home node B (HNB)

radio frequency (RF),” Oct. 2014.

[106] P. F. Ash and E. D. Bolker, “Generalized Dirichlet tessellations,” Geometriae Dedicata, vol. 20,

no. 2, pp. 209–243, Apr. 1986.

[107] J.-S. Ferenc and Z. Neda, “On the size distribution of Poisson Voronoi cells,” Physica A: Statistical

Mechanics and its Applications, vol. 385, no. 2, pp. 518–526, Nov. 2007.

[108] M. Udell and S. Boyd, “Maximizing a sum of sigmoids.” [Online]. Available: http:

//www.stanford.edu/∼udell/doc/max sum sigmoids.pdf

[109] M. Chiang, “Nonconvex optimization of communication systems,” Advances in Mechanics and

Mathematics, Special Volumn on Strang’s 70th Birthday, pp. 137–196, Oct. 2007.

[110] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.

[111] A. M. Ibrahim, T. ElBatt, and A. El-Keyi, “Coverage probability analysis for wireless networks

using repulsive point processes,” in Proc. of IEEE PIMRC, London, UK, Sep. 2013, pp. 1002–1007.

[112] A. Guo and M. Haenggi, “Spatial stochastic models and metrics for the structure of base stations

in cellular networks,” IEEE Trans. on Wireless Communications, vol. 12, no. 11, pp. 5800–5812,

Nov. 2013.

[113] A. Chambolle, S. Lisini, and L. Lussardi, “A remark on the anisotropic outer Minkowski content,”

Advances in Calculus of Variations, vol. 7, no. 2, pp. 241–266, Apr. 2014.

[114] H. Federer, Geometric Measure Theory. Springer, 1969.

[115] V. Chandrasekhar and J. G. Andrews, “Spectrum allocation in tiered cellular networks,” IEEE

Trans. on Communications, vol. 57, no. 10, pp. 3059–3068, Oct. 2009.

[116] Y. Chon, E. Talipov, H. Shin, and H. Cha, “Mobility data collected by

Lifemap monitoring system at Yonsei University in Seoul,” Downloaded from

http://crawdad.cs.dartmouth.edu/yonsei/lifemap/, May 2012.

Stochastic Analysis and Optimization of Heterogeneous ... · Stochastic Analysis and Optimization...

Documents

Transcript of Stochastic Analysis and Optimization of Heterogeneous ... · Stochastic Analysis and Optimization...