MATHEMATICAL AND ECONOMIC THEORY OF ROAD PRICING
Related Elsevier series DRESNER (ed.) Economics Research in Transportation Economics HENSHER & BUTTON (eds.) Handbooks in Transport Related Elsevier journals Transport Policy Preston Editors: Moshe BenAkiva, Yoshitsugu Hayashi and John Preston
Transportation Research Part A: Policy and Practice Practice Editor: Phil Goodwin Editor: Methodological Transportation Research Part B: Methodological Editor: Fred Mannering Mannering Editor:
Elsevier For full details of all transportation titles published under the Elsevier imprint please go to: www.elsevier.com/transportation www.elsevier.com/transportation
MATHEMATICAL AND ECONOMIC THEORY OF ROAD PRICING PRICING
HAI YANG Civil Engineering Department of Civil The The Hong Kong University of Science and Technology Hong Kong, P. R. China
HAIJUN HUANG School of Management Beijing University of Aeronautics and Astronautics Beijing 100083, 100083, P. R. China
2005
ELSEVIER  Heidelberg –  London –  New York –  Oxford – Paris Amsterdam – Boston –  Singapore –  Sydney –  Tokyo San Diego – San Francisco –
ELSEVIER B.V. Radarweg 29 211, 1000 1000 AE Amsterdam P.O. Box 211, The Netherlands
ELSEVIER Inc. 1900 525 B Street, Suite 1900 San Diego, CA 921014495 USA
ELSEVIER Ltd. The Boulevard, Langford Lane Kidlington, Oxford OX5 1GB UK
ELSEVIER Ltd. 84 Theobalds Road London WC1X 8RR UK
© 2005 Elsevier Ltd. All rights reserved. This work is protected under copyright by Elsevier Ltd., and the following terms and conditions apply to its use: Photocopying Single photocopies of single chapters may be made for personal use as allowed by national copyright laws. Permission of the Publisher and payment of a fee is required for all other photocopying, including multiple or systematic copying, copying for advertising or promotional purposes, resale, and all forms of document delivery. Special rates are available for educational institutions that wish to make photocopies for nonprofit nonprofit educational classroom use.
Permissions may be sought directly from Elsevier’s Elsevier's Rights Department in Oxford, UK: phone (+44) 1865 1865 843830, fax (+44) 1865 1865 853333, email:
[email protected] Requests may also be completed online via the Elsevier homepage (http:// www.elsevier.com/locate/permissions). www.elsevier.com/locate/permissions). In the USA, users may clear permissions and make payments through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA; phone: (+1) (978) 7508400, fax: (+1) (978) 7504744, and in the UK through the Copyright Licensing Agency Rapid Clearance Service (CLARCS), 90 Tottenham Court Road, London W1P 0LP, UK; phone: (+44) 20 7631 5555; fax: (+44) 20 7631 5500. Other countries may have a local reprographic rights agency for payments. Derivative Works Tables of contents may be reproduced for internal circulation, but permission of the Publisher is required for external resale or distribution of such material. Permission of the Publisher is required for all other derivative works, including compilations and translations. Electronic Storage or Usage Permission of the Publisher is required to store or use electronically any material contained in this work, including any chapter or part of a chapter. Except as outlined above, no part of this work may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the Publisher. Elsevier's Rights Department, at the fax and email addresses noted above. Address permissions requests to: Elsevier’s Notice No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from from any use or operation of any methods, products, instructions or ideas contained in the material verification of diagnoses and drug dosages herein. Because of rapid advances in the medical sciences, in particular, independent verification should be made. First edition 2005 Library of Congress Cataloging in Publication Data A catalog record is available from the Library of Congress.
British Library Cataloguing in Publication Data A catalogue record is available from the British Library.
ISBN:
0 08 044487 3
@ The paper used in this publication meets the requirements of ANSI/NISO Z39.481992 (Permanence (Permanence of Paper). Printed in The Netherlands.
v
CONTENTS Acronyms Glossary of Notation Preface
xiii xv xix
1
Introduction 1.1 Background 1.2 Theoretical Developments 1.2.1 The firstbest pricing problem 1.2.2 The secondbest pricing problems 1.2.3 Value of time and road pricing 1.2.4 Equity issues and revenue redistribution in road pricing 1.2.5 The road pricing and capacity choice problem 1.2.6 Mathematical programming approach to road pricing 1.2.7 Dynamic road pricing problem 1.3 Outline of the Book
1 1 1 2 3 4 5 6 8 9 10
2
Fundamentals of UserEquilibrium Problems 2.1 Introduction 2.2 Formulations of the Standard UserEquilibrium (UE) Problems 2.2.1 The UE model with fixed demand 2.2.2 Solution method for the fixed demand UE problem 2.2.3 The UE model with elastic demand 2.2.4 Solution method for the elastic demand UE problem 2.3 Traffic Assignment with Link Capacity Constraints 2.3.1 Model formulation 2.3.2 Solution algorithm 2.4 Traffic Assignment with NonSeparable Link Travel Time Functions 2.4.1 Asymmetric link flow interactions 2.4.2 Symmetric link flow interactions 2.4.3 Flow interactions between multiple types of vehicles 2.4.4 Positive definiteness of link cost functions with flow interactions 2.4.5 Practical asymmetric link cost functions 2.4.6 Diagonalization solution algorithm 2.5 Traffic Assignment of MultiClass Users with Different Values of Time 2.5.1 Model formulation
13 13 15 15 17 19 20 22 22 24 27 27 30 31 31 32 34 34 35
vi
Mathematical and Economic Theory of Road Pricing 2.5.2 Equivalence of the timebased and costbased multiclass traffic equilibrium Traffic Assignment with NonAdditive Path Costs 2.6.1 Model formulation 2.6.2 Solution algorithm 2.6.3 An analytical example Traffic Assignment with Stochastic Route Choice 2.7.1 Utility function and discrete route choice model 2.7.2 Model formulation with fixed demand 2.7.3 Model formulation with elastic demand System Optimum Traffic Assignment Sources and Notes
36 38 38 39 40 41 41 42 43 43 45
3 The FirstBest Road Pricing Problems 3.1 Introduction 3.2 MarginalCost Pricing under Standard UserEquilibrium 3.2.1 Graphical interpretation 3.2.2 Fixed demand case 3.2.3 Elastic demand case 3.3 MarginalCost Pricing with Link Capacity Constraints 3.3.1 Pricing model with capacity constraints 3.3.2 Speedflow relationship for congestion pricing 3.4 MarginalCost Pricing with a Total Environmental Quality Constraint 3.5 MarginalCost Pricing with Link Flow Interactions and Multiple Vehicle Types 3.5.1 The case with link flow interactions 3.5.2 The case with multiple vehicle types 3.6 MarginalCost Pricing and Uniqueness of Social Optimum 3.6.1 Uniqueness of OD demands 3.6.2 Uniqueness of link flows with separable link cost functions 3.6.3 Uniqueness of link flows with nonseparable link cost functions 3.6.4 Analytical examples 3.7 MarginalCost Pricing under Stochastic User Equilibrium 3.7.1 Economic benefit measure and maximization 3.7.2 Equivalence to the minimization of expected perceived travel cost 3.7.3 An analytical example 3.8 Summary 3.9 Sources and Notes
47 47 48 48 50 52 54 54 56 60 62 62 63 64 64 65 66 71 74 74 76 77 79 79
2.6
2.7
2.8 2.9
4 The SecondBest Road Pricing Problems: A Sensitivity Analysis Based Approach 81 4.1 Introduction 81
Contents
vii
4.2 Formulation as a Mathematical Program with Equilibrium Constraints 4.3 Sensitivity Analysis of Traffic Equilibria with Pricing 4.3.1 Sensitivity analysis of the restricted network equilibrium problem 4.3.2 Relations to other gradientbased sensitivity analysis methods 4.3.3 Dissection of sensitivity analysis with analytical examples 4.3.4 A sensitivity analysis based algorithm 4.4 Applications of the Sensitivity Analysis Method 4.5 Traffic Restraint and Road Pricing Problem 4.5.1 A conceptual framework 4.5.2 Direct solution with the queuing network equilibrium model 4.5.3 Nonuniqueness of link tolls for environmental capacity constraint 4.5.4 Selection of optimal tolls by bilevel programming 4.6 Summary 4.7 Sources and Notes
82 84 84 91 93 108 109 113 113 114 116 117 120 121
The SecondBest Road Pricing Problems: A Gap Function Based Approach 5.1 Introduction 5.2 Marginal Function Based Approach 5.2.1 Gap function for the VI based traffic equilibrium models 5.2.2 Gap function for the optimization based traffic equilibrium models 5.2.3 Reformulation as continuously differentiable optimization problems 5.2.4 A scheme of the augmented Lagrangian algorithm 5.3 Applications of the Gap Function Approach 5.4 EntryExit Based Toll Design Problems 5.4.1 A UE model with entryexit specific tolls 5.4.2 Solving the UE model with entryexit specific tolls by network decomposition 5.4.3 Alternative tolling schemes 5.4.4 BLPP formulation and gap function approach 5.4.5 Solution by the augmented Lagrangian algorithm 5.4.6 Numerical example and discussion of results 5.5 Summary 5.6 Sources and Notes
123 123 124 124 127 129 130 131 139 140
Discriminatory and Anonymous Road Pricing 6.1 Introduction 6.2 Toll Properties for the Single Class UE and SO Problems 6.2.1 Characterization of link tolls for SO 6.2.2 A simple example for the valid link toll set to decentralize SO 6.3 TimeBased and CostBased MultiClass UE and SO Problems
161 161 163 163 166 168
144 146 148 150 153 159 160
viii
Mathematical and Economic Theory of Road Pricing
6.4
6.5
6.6 6.7
6.3.1 System optimum in time units and pricing for equilibrium 6.3.2 System optimum in cost units and pricing for equilibrium 6.3.3 Analytical examples Multiple Equilibrium Behaviors and SO Problems 6.4.1 UECN multiple behavior equilibrium and system optimum 6.4.2 Anonymous link tolls to decentralize a system optimum 6.4.3 An analytical example MultiUEClass and MultiCNClass Behavior Equilibrium and SO Problem 6.5.1 Multicriteria and multiple behavior equilibrium and system optimum 6.5.2 Anonymous link tolls to decentralize a system optimum 6.5.3 An analytical example Summary Sources and Notes
169 174 176 181 182 186 189 192 193 195 197 200 201
7
Social and Spatial Equities and Revenue Redistribution 7.1 Introduction 7.2 The Social and Spatial Inequity: An Example 7.3 Redistribution of Congestion Pricing Revenue 7.3.1 Existence of a Pareto refunding scheme 7.3.2 Design of Pareto refunding schemes 7.3.3 Remarks on the Pareto refunding schemes 7.4 Network Toll Design Models with Equity Constraints 7.4.1 Traffic equilibrium and social welfare with pricing 7.4.2 Specification of equity constraints 7.4.3 Network toll design with equity constraints 7.5 Alternative Solution Approaches 7.5.1 Solution for bilevel model 7.5.2 Solution for trilevel model 7.6 A Numerical Example 7.7 Summary 7.8 Sources and Notes
203 203 204 210 210 218 224 225 225 226 228 230 230 232 232 237 238
8
Pricing, Capacity Choice and Financing 8.1 Introduction 8.2 User Equilibrium and Performance Evaluation for Private Toll Roads 8.2.1 Multiclass UE model with elastic demand 8.2.2 Alternative performance measures 8.3 Profitability, Social Welfare Gain and Optimal PriceCapacity Adjustments 8.3.1 Experiment setting 8.3.2 Characterization and analysis of numerical results
239 239 240 240 241 243 243 245
Contents
8.4
8.5
8.6
8.7 8.8 9
Profitability and Welfare Gain of Private Toll Roads with Heterogeneous Users 8.4.1 Experiment setting 8.4.2 Characterization and analysis of numerical results SelfFinancial in General Networks 8.5.1 Selffinancing rules with a single type of vehicles 8.5.2 Selffinancing rules with multiple types of vehicles Competition and Equilibria of Private Toll Roads 8.6.1 Monopolistic competition 8.6.2 Alternative formulation 8.6.3 Solution methods 8.6.4 Two analytical examples 8.6.5 Graphical characterization of toll competition Summary Sources and Notes
Simultaneous Determination of Optimal Toll Levels and Locations 9.1 Introduction 9.2 The LinkBased Pricing: Methodology 9.2.1 Genetic algorithm for determination of optimal tolling links 9.2.2 Selection of minimal number and locations of tolling links 9.3 The LinkBased Pricing: Example 9.4 The CordonBased Pricing: Methodology 9.4.1 Mathematical properties of toll cordons 9.4.2 Types of cordon 9.4.3 Determination of tolling cordons 9.5 The CordonBased Pricing: Example 9.5.1 Determination of location of a singlelayered tolling cordon 9.5.2 Determination of location of doublelayered tolling cordons 9.5.3 Impact of the cordonbased pricing on trip length distribution 9.5.4 Sensitivity analysis of demand elasticity 9.6 Summary 9.7 Sources and Notes
10 Sequential Pricing Experiments with Limited Information 10.1 Introduction 10.2 Implementation of the FirstBest Pricing Scheme 10.2.1 An iterative toll adjustment procedure for a single road 10.2.2 An iterative toll adjustment procedure for a general network 10.2.3 Proof of convergence 10.2.4 A numerical example
ix
247 248 251 255 255 263 267 267 268 270 272 274 281 282 283 283 284 284 287 287 290 291 294 296 301 303 303 304 306 306 307 309 309 310 310 313 316 320
x
Mathematical and Economic Theory of Road Pricing
10.3 Implementation of the SecondBest Pricing Scheme 10.3.1 A sequential implementation procedure 10.3.2 Sequential linear approximation 10.3.3 Two numerical examples 10.4 Implementation of the Traffic Restraint and Road Pricing Scheme 10.4.1 The primal and dual formulation 10.4.2 Trialanderror implementation procedure 10.4.3 A numerical example 10.5 Summary 10.6 Sources and Notes
325 325 328 329 335 335 337 339 341 343
11 Bounding the Efficiency Gain or Loss of Road Pricing 11.1 Introduction 11.2 Maximum Efficiency Gain of FirstBest Pricing Schemes 11.2.1 Efficiency gain with fixed demand 11.2.2 Efficiency gain for cost functions with limited congestion effects 11.2.3 Efficiency gain with elastic demand 11.3 Inefficiency of Stochastic User Equilibria 11.3.1 Determination of the cost inefficiency bound 11.3.2 Interpretation of the cost inefficiency bound 11.4 Inefficiency of CournotNash Equilibria 11.4.1 CN equilibrium with a finite number of players 11.4.2 Upperbound of the inefficiency of CN equilibria 11.4.3 Upperbound of special cases and a numerical example 11.5 Maximum Efficiency Loss of SecondBest Pricing Schemes 11.5.1 Measure of efficiency loss 11.5.2 Bound for traffic equilibria with fixed demands 11.5.3 Bound with polynomial cost functions 11.5.4 Bound for traffic equilibria with elastic demands 11.5.5 Alternative approach for bounding efficiency loss 11.5.6 Determination of the inefficiency bound for actual pricing schemes 11.6 Summary 11.7 Sources and Notes
345 345 346 346 352 354 358 358 362 363 363 365 367 369 369 370 375 378 383 384 387 387
12 Dynamic Road Pricing: Single and Parallel Bottleneck Models 12.1 Introduction 12.2 Single Bottleneck Models 12.2.1 The case with homogeneous commuters 12.2.2 The case with heterogeneous commuters 12.3 Parallel Bottleneck Models
389 389 389 390 395 400
Contents
12.3.1 Dynamic tolls on all routes 12.3.2 Dynamic tolls on partial routes 12.4 Bottleneck Pricing and Modal Split 12.4.1 Competition between mass transit and highway 12.4.2 Pricing and logitbased mode choice with elastic demand 12.5 Summary 12.6 Sources and Notes
xi
400 402 403 404 414 420 421
13 Dynamic Road Pricing: General Network Models 13.1 Introduction 13.2 Problem Description 13.3 SpaceTime Expanded Network 13.3.1 Basic assumptions 13.3.2 Static temporal expanded network 13.3.3 Traffic on the STEN 13.3.4 Link exit capacity and travel cost on the STEN 13.4 System Optimum, Externality and Congestion Toll 13.4.1 Model formulation 13.4.2 Optimality conditions 13.4.3 Externality and congestion toll 13.5 Numerical Examples 13.5.1 A single bottleneck network 13.5.2 A general bottleneck network 13.6 Summary 13.7 Sources and Notes
423 423 424 426 426 427 429 430 431 431 432 434 435 435 438 441 441
References Subject Index
443 461
This Page is Intentionally Left Blank
xiii Xlll
ACRONYMS AC BLPP BOT BPR BLVI CN DUE EB EPEC KKT LP MC MPEC OD SC SO SW STEN SUE UB UE VI VOT
Average Cost BiLevel Programming Problem BuildOperateTransfer Bureau of Public Roads BiLevel Variational Inequality CournotNash Deterministic User Equilibrium Economic Benefit Equilibrium Problem with Equilibrium Constraints KarushKuhnTucker Linear Program Marginal Cost Mathematical Program with Equilibrium Constraints Origin Destination Social Cost System Optimum Social Welfare SpaceTime Expanded Network Stochastic User Equilibrium User Benefit User Equilibrium Variational Inequality Value Of Time
This Page is Intentionally Left Blank
XV xv
GLOSSARY OF NOTATION a A
a link a e A the set of all links
A
the subset of links subject to a toll charge
A
the subset of links with observed flows
Bw(dw)
the benefit function between OD pair we W
$Z [d")
the benefit function of user class me M between OD pair weW
cm
the travel cost on path r connecting OD pair weW
c
the vector of path travel costs
c^,
the travel cost on path r between OD pair w by users of class me M
Ca
the capacity of link ae A
dw
the travel demand between OD pair weW
d
the vector of OD demands
Dw
the potential demand or the upperbound of demand between OD pair
weW d™
the travel demand of user class m between OD pair weW
D™
the potential demand of user class m between OD pair w&W
Dw (\xw)
the demand function between OD pair w e W
EC {d*>)
the inversed demand function between OD pair weW
E"
Elasticity of v to u, E" = (u/v)dv(u)/du
fm
the flow on path re R^
f
the vector of path flows
fZ
the flow of user class me M on path r £ R^ between OD pair weW
G
(N,A) be a network with node set A^ and link set A
K
the set of CournotNash (CN) players or the set of vehicle types
Ia{ya)
the construction cost of link a as a function of link capacity ya
m M N r R^,
a user class me M the set of all user classes with different values of time the set of all nodes a path or a route re R the set of all paths connecting OD pair weW
xvi
Mathematical and Economic Theory of Road Pricing
t°a
the freeflow travel time (moving time on link a e A
ta (va)
the travel time on link ae A as a function of link flow va
ia(va)
the marginal social travel time on link ae A , including the congestion externality, ta (va) = ta (v a ) + va dta (va)/dva
K (va>X.)
t n e link travel time as a function of both flow and capacity variables
t(v)
the vector of link travel times
ua
toll charge on link ae A
u va
the vector of link tolls the flow on link ae A
v
the vector of link flows
v™
the flow of user class m on link ae A
v*
the link flow of vehicle type ke K or CN player ke K
w W
an OD pair weW the set of all OD pairs
Wk
the set of OD pairs of which users are controlled by CN player ke K
w
K
Wm M
=Ute^ the set of OD pairs of which users are controlled by UE class me M
w
=UmsA/^m
ya
the capacity of a new link a or capacity increase of an existing link a
y
the vector of link capacity variables
p
the users' value of time
pm
the average value of time of user class me M
bar
1, if link ae A is on path r e R and 0, otherwise
A
the link/path incidence matrix, A = [8 or ]
e
a predetermined tolerance for stopping iterative process
Km
1 if path re 1^ and 0, otherwise
A
the OD/path incidence matrix, A = [ A m ]
[iw
the Lagrange multiplier associated with the demand conservation constraint of OD pair weW, or the generalized travel cost between OD pair we W at equilibrium
(0.™
the generalized travel cost of user class me M between OD pair we W
\i
the vector of generalized OD travel costs
Glossary of Notation Q.
the feasible region of link flows and OD flows
Q7
the feasible region of path flows
£2 V
the feasible region of link flows
Q* Qm
the feasible region of links flows of CN player ke K the feasible region of link flows of a UE class me M
Q.u Vvt(v)
the feasible region of link flows of UE player the Jacobian matrix of link travel time function t with respect to v
e c u °°  •
element membership proper set inclusion union of sets infinity the cardinality of a finite set

the Euclidean norm
argmin^ F(x)
the set of x attaining the minimum of the function F(x)
xvii
This Page is Intentionally Left Blank
xix XIX
PREFACE Road pricing as an effective means of both managing road traffic demand and raising additional revenue for road construction has been studied extensively by both transportation researchers and economists. Practical implementation has been progressing rapidly and electronic road pricing schemes have been proposed and tested worldwide. It is likely that over the next few years, with increasing public acceptability, there will be greater use of road pricing. The incontestable fact is that there is a great need for the development of efficient roaduse pricing models. Despite a few monographs and journal special issues in the literature that have been devoted to the topic in recent years, most studies have focused on empirical studies, policy experience, environmental issue of road congestion and road pricing. There is still scope for methodological development of advanced road pricing systems, such as dynamic pricing, integrated road and transit pricing, as well as practical toll charging schemes in general road networks. This book is intended to deal with a number of timely topics including: fundamentals of userequilibrium problems; principle of marginalcost pricing applied to road pricing; existence of optimal link tolls for system optimum under multiclass, multicriteria, multiple equilibrium behaviors; social and spatial equity issues in road pricing; optimal design of private toll roads; simultaneous determination of optimal toll levels and locations; trialanderror implementation of marginalcost pricing on networks with unknown demand functions; dynamic road pricing. This book would appear to be the first book devoted exclusively to the mathematical and economic investigation of roaduse pricing in general congested networks, which aims at alleviating traffic congestion, improving transport conditions and enhancing social welfare. It constitutes an update of the state of the art of the latest research, mainly by the authors and their colleagues. The book is targeted at students, professionals and scientists who are studying and working in relevant transportation fields. We are most thankful to a number of individuals for their help during our work in this field. First and foremost we would like to acknowledge the contribution and continuing encouragement of Prof. M.G.H. Bell of Imperial College London (ICL). In fact, the manuscript was finalized during the first author's sabbatical leave at ICL in early 2005, and he is particularly grateful to Prof. Bell for his hospitality during a very fruitful and enjoyable stay in ICL. Thanks are also due to Prof. R. Lindsey of the University of Alberta, Prof. E.
xx
Mathematical and Economic Theory of Road Pricing
Verhoef of the Free University of Amsterdam, Prof. W.H.K. Lam of the Hong Kong Polytechnic University and Prof. S.C. Wong of the University of Hong Kong throughout our study of the various road pricing problems and in the preparation of this book. We highly appreciate our research collaboration that produced some of the material included in the book. We also wish to greatly thank three former Ph.D students at the Hong Kong University of Science and Technology, Dr. Qiang Meng (currently at the National University of Singapore), Dr. Xiaoning Zhang (currently at the Tongji University of Shanghai) and Dr. Judith Wang (currently at the University of Auckland) for their contributions to various topics in the book. Thanks also go to Miss Wei Xu and Mr Xiaolei Guo (currently PhD students at the Hong Kong University of Science and Technology) who helped in discussions in earlier drafts of the book. Funding for the research was provided by the Research Grants Council of the Hong Kong Special Administrative Region (HKSAR) and the National Natural Science Foundation, P.R. China.
Hai Yang HaiJun Huang 31 May 2005
INTRODUCTION 1.1
BACKGROUND
Roadway congestion is a source of enormous economic costs. In principle, many of these costs can be prevented, as they result from socially inefficient choices by individual drivers. A number of regions have considered alleviating roadway congestion by introducing congestion pricing. Indeed, road pricing has become one of the priorities on transport policy agendas throughout the world. It is increasingly believed that road pricing may offer an effective instrument to manage travel demand, and to raise revenue that may, for instance, be used for transport improvements. An increasing number of congestion pricing schemes have been proposed, tested or implemented worldwide. Examples include the US's value pricing scheme, recent EU green and white papers, Dutch initiatives, electronic road pricing schemes in Singapore and Hong Kong, and the London congestion charging scheme that was introduced in February 2003. It is conceivable that a new generation of roaduse pricing technologies will be widely considered for introduction on many congested road networks. All these emerging technologies offer efficient tools and challenges for implementation of road pricing for traffic management, and also generate a great need for the development of efficient pricing models that allow for the evaluation and perhaps design of innovative road pricing schemes. Some monographs and journal special issues have been devoted to these topics in recent years. However, the majority of studies have focused on the empirical aspects, policy experiences, and environmental issues of road congestion and road pricing. There is, however, still ample scope for methodological developments for the analysis of advanced road pricing systems, such as dynamic pricing, integrated road and transit pricing, as well as practical toll charging schemes in general road networks.
1.2
THEORETICAL DEVELOPMENTS
The initial idea of road pricing was put forward by Pigou (1920), who used the example of a congested road to make points on externalities and optimal congestion charges (see, also, Knight, 1924, for early interpretation of social cost and toll charge). The fundamental concept
2
Mathematical and Economic Theory of Road Pricing
is easy: to apply a price mechanism in the same way as it applies elsewhere in a market economy. Prices should be higher under congestion conditions and lower at less congested times and locations in order to deter excessive uses. The key question is how to choose the appropriate price in a simple yet practical manner, under what is often a complex set of economic and technical circumstances. The key issue from a practical point of view is how to implement the concept, not only in terms of developing technically efficient charging mechanisms, but also in gaining political acceptance as a valid policy instrument (Button and Verhoef, 1998). There have been both intellectual and practical developments after Pigou's idea of using road pricing measures to regulate road traffic congestion. Seminal works on road pricing that need to be mentioned include Wardrop (1952), Walters (1961), Beckmann (1965) and Vickrey (1969). The subject of road pricing has particularly rejuvenated interest extensively from both economists and transportation researchers in recent years, because of the growing prominence and changing nature of urban transportation problems facing a modern city. Here we do not intend to fully review the large body of literature (which will be done in individual chapters) but merely outline the various types of road pricing problems of interest and some relevant analytical studies. Readers are invited to refer to McDoland (1995), Lo and Hickman (1997), Button and Verhoef (1998), Verhoef (1996), McDonald et al. (1999), Levinson (2002), Yang and Verhoef (2004), and Santos (2004) for excellent reviews or monographs on the subject, from the economic perspective. 1.2.1
The Firstbest Pricing Problem
The theoretical background of roaduse pricing has relied upon the fundamental economic principle of marginalcost pricing, which states that road users using congested roads should pay a toll equal to the difference between the marginalsocial cost and the marginalprivate cost in order to maximize the social surplus. By so doing, each user will face the marginal social cost of road use other than the marginal private costs. For a congested road, this includes the value of time losses imposed on other road users, as well as the value of emissions, noise and accident risks created. The social surplus, defined as the difference between the total benefits and total costs, is often considered as an appropriate indicator for social welfare, and its maximization is a condition for economic efficiency to prevail (Verhoef, 1996). The theory of the marginalcost pricing or the firstbest pricing in the literature by Pigou and followers (for example, Walters, 1961; Vickrey, 1963; Evans, 1992; Hills, 1993) was developed based on the demandsupply (or performance) curves for the standard case of a homogenous traffic stream moving along a given uniform stretch of road, such as an expressway, connecting given entry and exit points. There would be some decision on what
Introduction
3
the optimal traffic flow should be, essentially determined by estimating the traffic speedflow relationship underlying the performance curve. The congestion charge would then be an efficient way of attaining the target flow of traffic. In the case of homogeneous users, the firstbest congestion pricing theory is established in general traffic networks. In line with this theory, a toll that is equal to the difference between the marginal social cost and the marginal private cost is charged on each link, so as to internalize the user externalities and thus achieve a system optimum flow pattern in the network (Beckmann, 1965; Dafermos and Sparrow, 1969, 1971). Investigations have been conducted on how this classical economic principle would work in a general congested road network with multiple vehicle types, such as trucks versus passenger cars (Dafermos, 1972, 1973), with both multiple vehicle types and link flow interactions (Smith, 1979b), with queuing (Yang and Huang, 1998), and in a congested network in a stochastic equilibrium (Yang, 1999). Moreover, Bellei et al. (2002) developed a variational inequality model for network pricing optimization in a multiuser and multimodal context; Liu and Boyce (2002) presented a variational inequality formulation of the marginalcost pricing problem for a general transportation network with multiple time periods. Recently, Yang, Zhang and Meng (2004) developed a trialanderror implementation method for marginalcost pricing on networks in the absence of demand functions. 1.2.2
The Secondbest pricing problems
In spite of its perfect theoretical basis, the firstbest or marginalcost pricing scheme is of little practical interest. The problem stems from the fact that it is impractical to charge users on each network link in view of the operating cost and public acceptance. Note that a simple introduction of the "marginalcost pricing tolls" over a subset of the links of a network based on a local link speedflow relation may distort the allocation of the traffic over the entire network and may cause degradation instead of improvement in social welfare. Thus, setting the toll equal to the marginal cost of the trips may not be valid for the design of an optimal road pricing scheme in a more complex environment. Due to the imperfection of the firstbest pricing, the secondbest pricing is, in fact, often the most relevant case from a practical perspective and has received ample attention in the recent literature (Lindsey and Verhoef, 2001). The key questions to be answered in a secondbest pricing scheme in general include: where to levy the toll and how much? What different impacts would the pricing schemes bring to different users? Most studies in the context of the secondbest pricing problems are generally conducted with respect to the determination of toll levels for given charge locations. The simplest version of the problem concerns the two route problem, where an untolled alternative road is available
4
Mathematical and Economic Theory of Road Pricing
parallel to a toll road. This problem was probably first analyzed by LevyLambert (1968) and Marchand (1968), and more recently by Braid (1996), Verhoef et al. (1996), and Liu and McDonald (1999). There are in general four alternative types of toll charging schemes in road networks that are of practical interest (May and Milne, 2000): traveldistance based charging, travel time or traveldelay based charging, linkbased charging and cordonbased charging. In a travel distance or travel time based charging scheme, a toll is charged in proportion to the distance traveled or the time spent in the network and the optimal toll charge per unit distance or travel time for minimization of total travel time or maximization of social welfare can be easily determined using simple optimization search methods. In a linkbased pricing scheme, tolls are charged only on a subset of selected links, especially where there is a bottleneck. In many cities, congestion tolls are collected on some urban bridges; tunnels or expressways, and tolls at different locations are coordinated for congestion mitigation (Yang and Lam, 1996). Furthermore, the link tolls are considered to act as some additional costs that prevent traffic flow becoming unstable and/or producing unacceptable environmental damage. This consideration is intimately related to the emerging issue of sustainable development and transportation facing most major cities. Road pricing and travel demand management are being considered to be the key instruments to deal with the environmental impacts of transportation. With this pivotal issue in mind, Ferrari (1995, 1997, 1999) and Yang and Bell (1997) developed various mathematical models for the traffic restraint and road pricing problem that have important implications for enhancement of environmental quality and traffic management. Instead of charging on separate individual links, a cordonbased pricing scheme emerged in recent years in some countries or regions for reducing traffic demand in congested urban central areas (May et al., 2002; Santos et al., 2001; Santos, 2002; Mun et al., 2003). There are many rational aims of secondbest congestion pricing. They could be minimizing the system cost, including operating cost, maximizing social welfare gain and maximizing revenue. In addition, simultaneous determination of toll locations and toll levels on a network is of practical importance (Hearn and Ramana, 1998; Verhoef, 2002; Yang and Zhang, 2002a). 1.2.3
Value of Time and Road Pricing
The concept of value of time (VOT) plays a pivotal role in road pricing analysis as it describes how users make tradeoffs between cost and time. In the presence of heterogeneous users with different VOTs, various network equilibrium models have been developed by assuming either a discrete set of VOTs for several distinct user classes or by a continuously distributed VOT across the whole population (Leurent, 1993, 1996 and 1998; Marcotte and Zhu, 2000; Mayet and Hansen, 2000; Nagurney, 2000).
Introduction
5
The optimal pricing problems for users with discrete or continuous VOT distributions had been investigated by Dial (1999a, 1999b) and Florian (1998). Dial (1999a, 1999b) established a model and algorithm for determining a systemwide optimal set of tolls that induce traffic that is simultaneously user and systemoptimal. The model permits the VOT to be continuously distributed for each OriginDestination (OD) pair. Assuming that each trip uses a path that minimizes its own particular perception of generalized cost, Dial demonstrates what economists have always known, that is to minimize the total perceived cost of time and the best toll for a link is its expected value of the social component of its marginal cost. Arnott and Kraus (also refer to Lindsey and Verhoef, 2001) have shown that firstbest pricing remains possible using anonymous tolls with heterogeneous user groups, provided that the tolls can be varied freely over time and the system objective function is measured in conventional unit of dollars. The optimality of anonymous tolling is derived from the fact that the appropriate tolls depend only on the marginal externality costs that drivers impose and that the congestion externality is anonymous at any departure time. Recently, Mayet and Hansen (2000) investigated secondbest congestion pricing with continuously distributed VOT for a highway with an unpriced substitute, where alterenative optimal tolls are found depending on whether the social welfare function is measured in money or in time, and whether toll revenue is or is not included as part of the benefit. Yang and Huang (2004) further investigated the relationship between the multiclass, multicriteria traffic network equilibrium and system optimum problems. Using a product differentiation approach, Verhoef and Small (2004) examined the impacts of user heterogeneity (with a continuously distributed VOT) on the results of the firstbest, secondbest and the revenuemaximizing pricing policies. De Palma and Lindsey (2004) examined the effect of congestion pricing with heterogeneous travelers using a generalequilibrium welfare analysis. 1.2.4
Equity Issues and Revenue Redistribution in Road Pricing
In congestion pricing, the social equity issue is concerned with the unequivocally distributional impacts of road pricing between the poor and the rich users, because fixed charges represent a larger portion of income for a lowincome drivers than for a highincome driver. This socially inequitable issue has often been used as an argument to justify the political unacceptability of road pricing and has been debated extensively in the literature (Foster, 1975; Small, 1983; Starkie, 1986; Hau, 1992; Johansson and Mattsson, 1995; Litman, 1996; Button and Verhoef, 1998; Richardson and Bae, 1998; Giuliano, 1992; Viegas, 2001; Paulley, 2002; Raux and Souche, 2004). Apart from the conventional social equity issue between poor and rich drivers by paying the same toll charge, there exists a spatial equity issue in the sense that the changes of the generalized travel costs of drivers traveling between different OD pairs may be significantly different when tolls are charged at some selected links. Santos and Rojey (2004) assessed the
6
Mathematical and Economic Theory of Road Pricing
potential distributional impacts of a road pricing scheme in three English towns and indeed found that impacts are town specific and depend on where people live, where people work and what mode of transport they use to go to work. Recently, Yang and Zhang (2002b) incorporated both the social and spatial equity issues in the analysis of the secondbest network toll design problems. Road pricing can raise large amounts of revenue and the alternative allocations of the revenue vary from investment in infrastructure, to compensation to losers and improvement of public transport. The way in which the government allocates revenues will determine both the equity and the political acceptability of a road pricing scheme. Several researchers have looked into various revenue distribution strategies. For examples, Small (1992c) proposed a travel allowance for all commuters. Parry and Bento (2001) recommended that income taxes be reduced to offset congestion pricingrelated labor supply restrictions. Goodwin (1989) suggested a combination of revenue uses in order to offset several congestion pricing impacts. Poole (1992) added that it may be possible to introduce offpeak discounts and peakhour surcharges on a toll road. DeCorlaSouza (1994) proposed a cashing out strategy to induce a shift of peakperiod travelers to other modes, thus reducing the need for additional infrastructure. Recently, Kalmanje and Kockelman (2004) proposed a creditbased congestion pricing strategy where the term "credit" refers to a monetary cashout. The method can help tackle the problem of congestion in an equitable and efficient fashion. From a theoretical perspective, Bernstein (1993) examined the possibility of userneutral congestion pricing with both positive and negative tolls (tolls and subsidies) in the bottleneck congestion model; Adler and Cetin, (2001) discussed a direct distribution approach to congestion pricing in which monies collected from users on a more desirable route are directly transferred to users on a less desirable route using a two parallel route example with bottleneck congestion. For a OD pair connected by a number of parallel routes, Elliasson (2001) showed that a tolling and refunding system that reduced aggregate travel time and refunded the toll revenues equally to all users will make everyone better off than before the toll reform; Yang, Meng and Hau (2004) developed an optimal integrated pricing model in a bimodal transportation network with explicit consideration of subsidy to the transit mode from road congestion pricing revenue. Apart from economic efficiency, they also showed that transport pricing and subsidy policy can also be justified on the grounds of the equity issues. 1.2.5
The Road Pricing and Capacity Choice Problem
Even if road capacity can be adjusted to cope with congestion, capacity policies alone will typically not be sufficient to reach a socially optimal outcome, and the use of congestion pricing remains warranted. Given the reality of public ownership of most roads and the possibility of internalizing the congestion externality with corrective congestion charge, there
Introduction
7
have been a number of studies looking at the link between the optimal congestion charge and the road investment strategy (Mohring, 1970). An important result in transport economics, the selffinancing result, was first established by Morning and Harwitz (1962) for a single road: Under certain technical conditions, the revenues from optimal congestion pricing, equal to the difference between the marginal social cost and marginal private cost at any level of traffic flow, will be just sufficient for financing the fixed costs associated with the optimal capacity supply. The conditions are satisfied when there are constant returns to scale in road construction and maintenance, and the capacity can be increased in continuous increments (Morning and Harwitz, 1962). This selffinancing result was later expanded to include damage. In this sense, a road charge equal to the sum of the optimal congestion charge and the road damage charge, covers the fixed operating cost and the variable maintenance cost, provided there are constant returns to scale in construction and, maintenance and use of road capacity (Newbery, 1988; 1989). The theorem has been shown to extend to each road individually in a full network and consequently to the network in aggregate, provided each link is optimally priced and all capacities are optimized (Yang and Meng, 2002), to dynamic models (Arnott and Kraus, 1995). The growing participation of the private sector in constructing, maintaining and operating road infrastructures has been justified for its positive effects on public budgets and cost efficiency in the provision of the required road capacity. Under the socalled BuildOperateTransfer (BOT) schemes, the private sector builds and operates roads at its own expense and receives the revenue from road toll charges over a number of years, after which these roads are transferred to the government. Such commercial and private provision of public roads has attracted fastgrowing interest in recent years and is being used to finance modern road systems worldwide (Tam, 1998). Once road provision moves into the market economy, there are many intriguing issues to be addressed. From the viewpoint of private investors, the profitability of a road project is of a great concern; while from the government side, social welfare should be the main goal. It is thus imperative to assess whether or not the construction of a toll road will give a positive welfare increment if it is profitable, compared with the donothing alternative, and vice versa whether or not a toll road, which adds to welfare, will be profitable and hence can be provided privately. Obviously, answers to these issues crucially depend upon the supply decisions made by individual firms in terms of capacity choice and pricing, which directly affect the cost of the road project and its attractiveness to potential users. The choice of capacity and prices is more complex when two or more competing firms operate multiple toll roads, because their profits are interrelated due to demand interdependence in the network. Apart from the consideration of road user responses, each firm must consider what its competitors' choices are likely to be in making its capacity design and pricing decisions.
8
Mathematical and Economic Theory of Road Pricing
There has been mounting interest in how price and capacity should be chosen by private road operators. General arguments over the private provision of roads are given by, for example, Geltner and Moavenzadeh (1987); GomezIbanez et al. (1991); Banister et al. (1993); David and Fernando (1995); Nijkamp and Rienstra (1995); and Roth (1996). Theoretical studies range from the use of an analytical approach for a single road or bottleneck or parallel roads (Viton, 1995; Mills, 1995, Verhoef, 1996, Braid, 1996; De Palma and Lindsey, 2000, 2002; Verhoef and Small, 2004; Tsai and Chu, 2003; De Rus and Romero, 2004) to the use of general traffic equilibrium approach for general networks (Yang and Meng, 2000, 2002; Verhoef and Rouwendal, 2004). Yang and Tang et al. (2002) developed multiclass network equilibrium models and studied the effect of users' value of time distributions on profitability and social welfare gain from private toll roads. Yang and Woo (2000) examined graphically the competitive Nash equilibrium by considering two toll roads provided and operated by two profitmaximizing private firms. Wang et al. (2004) examined a situation where both highway infrastructure and transportation (bus) services on the highway are provided commercially and privately, and examined the strategic interactions of such a bilateral monopoly. 1.2.6
Mathematical Programming Approach to Road Pricing
By explicitly incorporating the users' responses to the toll pricing into the analysis, the general network toll design problem can be studied in the context of the TNOUEC (Transportation Network Optimization with User Equilibrium Constraints), which has received a longstanding interest in the transportation science community. The TNOUEC problem is to determine the optimal values of a set of decision variables (toll charge in the case of congestion pricing) so that an appropriate measure of transportation system performance is optimized, while taking into account the route choice behavior of network users (for a recent review, see, Yang and Bell , 2001; Clegg et al., 2001). It is well known that the TNOUEC problems can be well described as BLPP (Bilevel Programming Problem) (Shimizu et al., 1997; Dempe, 2002) and more generally as MPEC (Mathematical Program with Equilibrium Constraints) (Luo et al., 1996). There has been a growing number of studies recently on the various network toll design problems, in particular in the case where only a subset of links in the network is subjected to toll charges. Optimal determination of tolls for a subset of links in a network by Yang and Lam (1996) for the case with fixed demand, by Zhang and Ge (2004) for the case with elastic demand and Yang and Bell (1997) when there are link flow restrictions. Labbe et al. (1998) and Brotcome et al. (2001) developed a linear bilevel model for toll optimization in uncongested networks. Yang and Zhang (2002b) dealt with the equityconstrained network toll design problem with multiple user classes. Bergendorff et al. (1997) and Hearn and Ramana (1998) proposed various methodologies to extract congestion toll sets such that the
Introduction
9
tolled user equilibrium is system optimal. Dial (1999c, 2000) proposed efficient algorithms to find the minimalrevenue congestion tolls that can sustain a useroptimal flow pattern as a system optimum; Bai et al. (2004) proposed a decomposition technique for the minimum toll revenue problem. Lawphongpanich and Hearn (2004) recently proposed a cutting plane method for solving the bilevel secondbest toll pricing problem. Chen and Bernstein (2004) and Chen et al. (2004) simplified formulations and solution methods of the network toll design problem with multiple user classes using restricted path sets; Yang, Zhang and Meng (2004) dealt with the network toll design problem with entryexit based toll charges using the network equilibrium model with pathspecific cost. 1.2.7
Dynamic Road Pricing Problem
There has been a vast accumulation of the literature that deals with the dynamic road pricing problem, ranging from a single highway bottleneck to a general transportation network following the bottleneck model developed by Vickrey (1969). In the standard bottleneck model, congestion is assumed to arise when vehicles queue behind a bottleneck. All commuters wish to arrive at work at a certain time but there is a bottleneck with finite capacity that will not allow all of them to arrive at their preferred time. There are costs associated with early and late arrival, which together with the toll, are added to the cost of the trip. Commuters try to minimize their generalized trip costs by choosing their departure time. Departure time evolves during the rush hour and hence forms a time pattern of departure. The evolution of congestion in the rush hour is thus endogenously determined. Arnott, de Palma and Lindsey (1988; 1990a,b; 1992; 1993a,b; 1994; 1998) have further extended the bottleneck model along the lines traced by Vickrey (1969) to include heterogeneous commuters, route choice, stochastic capacity, elastic demand, and timevarying tolls. A second form of the dynamic congestion model, initially developed by Henserson (1974), considers hypercongestion. The approach also involves the distribution of travel delays and scheduling of costs of the peak, and the duration of the peak in the untolled equilibrium and the social optimum, being determined endogenously. This type of generic model was further extended by Mun (1994), Chu (1995), Huang and Yang (1996), Yang and Huang (1997). It is found that the social optimum does not require complete elimination of travel delays. Verhoef (2003) further examined the occurrence of hypercongestion and the optimal road pricing in a continuoustimecontinuousplace dynamic economic model of traffic congestion. A third form of the dynamic congestion model, initially developed by Tabuchi (1993), involves two travel modes, consisting of a road with a bottleneck and a railroad exhibiting economies of scale, in order to investigate the efficiency of several railroad fare and road toll regimes. Huang (2000, 2002) derived and compared the socially optimal combination of transit fare and road toll for the twomode problem in the case of the logitbased mode choice
10
Mathematical and Economic Theory of Road Pricing
and two groups of commuters. Danielis and Marcucci (2002) explored the consequences of price distortion in the railroad mode on road congestion pricing. Such distortion may take place when the railroad operator seeks to make a zero loss, when railroads are privatized or when the market for railroad services is assigned through franchise bidding.
1.3
OUTLINE OF THE BOOK
The theoretical background of roaduse pricing has relied upon the fundamental economic principle of marginalcost pricing, which states that road users using congested roads should pay a toll equal to the difference between the marginalsocial cost and the marginalprivate cost in order to maximize the social net benefit. Up to date, however, most theoretical arguments about roaduse pricing are entirely concerned with abstract demandsupply models with simplifying assumptions. There exists in the literature considerable confusion about the analysis of road congestion that needs to be clarified. There are also many interesting, yet important issues, to be explored when the detailed modeling of a network is involved, such as bounding the efficiency gain achieved from a marginalcost pricing scheme and the selffinancing in a general road network with general traffic equilibrium models. In addition, various difficulties raised from the practice of road pricing, for example how to ensure the social/spatial equity, how to handle road pricing revenue, and how to determine the optimal location of toll links, are in need of efficient solutions. Moreover, there has been a great need for the development of practical and efficient methods for the actual design and implementation. A fundamental question is how to choose the optimal congestion tolls in a simple yet practical manner. This book addresses the important, essential and practical problems of road pricing by making use of advanced modeling techniques with rigorous mathematical and economic theories. Most studies presented in the book are carried out on the economic equilibrium analysis of general transportation networks and hence are quite different from traditional analytical methods of road pricing. This book consists of thirteen chapters that provide indepth accounts of the state of the art, either by dealing with new problems or developing new solution methods. Chapter 2 provides the fundamentals of the concepts, models and algorithms of the traffic equilibrium and system optimum problems on networks and their various extensions, which are essential to the road pricing analysis. This chapter also contains review materials of wellknown solution methods for the various traffic equilibrium models. In Chapter 3, we expound the fundamental principle of marginalcost pricing that is used to consider road congestion pricing. We explain how this classical principle would work in a
Introduction
11
general congested network and how the principle holds when link flow interactions and link capacity constraints are built into the traffic equilibrium model. We examine the conditions for unique user equilibrium and system optimum solutions. In Chapter 4 we illustrate the application of the bilevel programming approach to the various types of secondbest alternative pricing problems when the stringent assumptions of the firstbest or marginalcost pricing conditions are not satisfied as indeed in the case in reality. We develop a simple and general gradientbased sensitivity analysis of network traffic equilibrium, and present its meaningful application to the general traffic restraint and road pricing problem. Chapter 5 continues to investigate the general secondbest pricing problems in networks. We introduce equivalent, continuously differentiable singlelevel reformulation of the general bilevel network toll design problems using differentiable gap functions. Both the gap function, by virtue of the marginal or value function of the convex UE program, and the general gap function of the variational inequality formulation of the generalized UE problem are described. An application of the gap function based approach is given to the network toll design problem with entryexit based toll charges. Chapter 6 examines the question of the existence of optimal link tolls for system optima under multiclass, multicriteria and multiple equilibrium behaviors. We investigate the following questions: whether the useroptimal flows depends on the unit (time or money) used in measuring the travel disutility in the presence of a toll charge; whether there are any uniform link tolls across all individuals with different values of time that can support a multiclass useroptimal flow pattern as a system optimum, when the system objective function is measured by either money or time units. In the presence of multiple equilibrium behaviors, whether a systemoptimum flow pattern remains attainable by meaningful link tolls. Chapter 7 is concerned with the social and spatial equity issues in road pricing. The social equity refers to the different financial burden imposed between lowerincome and higherincome drivers and the spatial equity represents the different changes in the generalized travel costs of drivers traveling between different OD pairs, when tolls are charged at some selected locations in the network. We develop interesting and meaningful pricing and revenue redistribution schemes that can make all users better off. We also incorporate both the social and spatial inequities as constraints in the network toll design models. Chapter 8 analyzes the pricing problems in the context of private roads. We first consider the profitability and social welfare gain of toll roads in a network with either homogeneous or heterogeneous users. We establish the conditions for the selffinancing theorem to hold in a firstbest road pricing and capacity choice scheme in a general network and examine its
12
Mathematical and Economic Theory of Road Pricing
consequence in a secondbest environment. We move to the investigation of the competition and equilibria, based on a situation where there are two or more profitmaximizing private firms that operate multiple interrelated toll roads in a road network. Chapter 9 investigates the problem of the simultaneous determination of optimal toll levels and locations on general networks encountered in the various secondbest pricing problems. Specifically, we deal with the secondbest linkbased and cordonbased pricing schemes that involve simultaneous selection of both toll levels and toll locations. Optimization models with mixed (discrete and continuous) variables are developed and solved with the metaheuristic methods. In Chapter 10, we examine the technical difficulty in determining the toll charge and develop novel new methods for the actual implementation of congestion pricing. Specifically we propose trialanderror implementation of the firstbest and the secondbest pricing problems on networks when the demand functions are unknown. We establish the convergence of the trialanderror procedure for the firstbest pricing problem and propose a heuristic sequential experimenting approach to the secondbest pricing problems. The proposed procedure allows a traffic planner to easily estimate the socially optimal congestion tolls in networks without resort to the demand functions. In Chapter 11, we attempt to find the maximum efficiency gain that can be achieved through a marginalcost pricing scheme in a general network. An upper bound of the efficiency gain of a marginalcost pricing scheme is established, using the latest concept of price of anarchy of atomic and nonatomic congestion games. We also try to bound the efficiency loss of a secondbest pricing scheme in relation to the firstbest one. Chapter 12 is devoted to the dynamic road pricing problems with single or parallel bottlenecks. We present the basic single bottleneck model and then its extension to a network of parallel routes. The two mode problem is also discussed in which an alternative mass commuting mode, such as a railway or subway, exists in parallel to the road with a bottleneck. In all cases, we obtain analytical results and provide sensible economic or behavioral interpretations. Finally, in Chapter 13, a general model simultaneously dealing with departure time, route choice and optimal congestion toll in a general network is developed, using the spacetime extended approach. The model can deal with a general queuing network with elastic demand, and allow for treatment of commuter heterogeneity in their work start time and schedule delay cost, and hence make a significant advance over the previous simple bottleneck models of peakperiod congestion.
FUNDAMENTALS OF USEREQUILIBRIUM PROBLEMS
2.1
INTRODUCTION
Two central concepts pertaining to road pricing are UserEquilibrium (UE) and System Optimum (SO) of traffic flow distribution in networks. In this chapter, we provide a tutorial in the various formulations of the standard UE and SO problems and their extensions, which are frequently used in subsequent analysis of various road pricing issues. A brief outline of the well established solution concepts for the various UE and SO problems is provided as well. The background materials would be useful in understanding the developments in the main body of the book. The fundamental aim of traffic assignment is to find link or path flow patterns, given the OriginDestination (OD) trip rates, network topology, and link performance functions. The result of the assignment is an estimate of traffic volume on each link in the network, and the associated measures of system performance. To do so, we have to make some assumptions that describe how commuters make their travel decisions. A well received assumption in the literature is that all users have perfect information about traffic conditions over the entire network and make consistently correct decisions in route choice, leading to the deterministic (UE or SO) assignment models. The UE model assumes that all users minimize their own individual travel costs on transportation networks, while the SO model hypothesizes that all users cooperate to minimize total network cost. Incorporating uncertainty in travel choice analysis, as usually done in stochastic traffic assignment models, is important, but is not covered in this book. The following terms and concepts are frequently used throughout this book. An urban road transportation system is described as a strongly connected, directed network (N,A) where N and A denote the sets of nodes and links, respectively. The assumption of a strongly connected network means that there exists at least one route from each origin to each
Mathematical and Economic Theory of Road Pricing
14
destination. A link a e A represents a road segment and nodes may be origins, destinations or road intersections. Trips are generated from various origins and absorbed to all destinations. Let W be the set of such OD pairs and dK be the number of trips made between OD pair weW. Users select paths (or routes) for traveling from their origins to destinations in the network. A path is defined as a connected sequence of links and R^ denotes the set of all simple paths connecting OD pair weW. Traffic volume or flow on a link is the sum of flows along the paths that go through this link. Let fm denote the flow on path reRw between OD pair weW
and va the flow on link a e A. The relationship
between link flows and path flows can be expressed by HEW reRw
where 8O, is equal to 1 if link a is on path r and 0 otherwise. Due to traffic congestion, travel time on a link a e A is assumed to be an increasing function of flow on that link. Let ta{va) be the (average) link travel time function for link a e A. It is assumed to be a differentiable, monotonically increasing function of link flow. A typical link travel time function is shown in Figure 2.1, where freeflow travel time corresponds to zero flow. Theoretically, the traffic flow on a link can not exceed its physical capacity, but an approximate form of the function is adopted in most models for facilitating the solution techniques. Traffic assignment models with explicit link capacity constraints are given later. The travel time along a path, crw, re R^, we W, is given as the sum of travel times on all links that constitute the path. We thus have
P
Theoretical curve/ \ Practical approximation
Freeflow travel time 0
Capacity Link Flow
Figure 2.1 Typical form of a link travel time function
Fundamentals of UserEquilibrium Problems 2.2
15
FORMULATION OF THE STANDARD USEREQUILIBRIUM
PROBLEMS 2.2.1
The UE Model with Fixed Demand
As mentioned above, in the UE assignment problems, all trips are assumed to select paths or routes independently that minimize their own individual travel costs. Under the resulting network flow pattern, all routes actually used for a specific OD pair have equal travel cost which is less than that on any of the unused paths. We call such a network being in a Wardrop UE state where no users can find a shorter path by unilaterally changing their routes. Here we provide formulation of the UE problem in the forms of mathematical programs. It is well known that the linkflow pattern satisfying the UE conditions can be obtained through solving the following equivalent mathematical program (23>
_. subject to
fm>0,reRw,weW
(2.5)
where link flows, x = {va,ae A) =(•••, v0,•••) (arranged as a column vector), are given by (2.1). The objective function does not have any intuitive economical or behavioral interpretation. It can only be viewed as a mathematical construction for implementing the UE condition. Constraints (2.4) state that the sum of flows on all paths connecting each OD pair has to equal the OD demand. Constraints (2.5) are nonnegativity requirements on all path flows. The relationships between link flows and path flows are expressed by (2.1). Formulation (2.3)(2.5) is called the Beckmann transformation. It is easy to prove that any flow pattern that solves (2.3)(2.5) satisfies the Wardrop's UE conditions. View link flow v as a function of path flow vector f as defined by (2.1) and construct a Lagrangean function:
where \aw is the Lagrange multiplier associated with (2.4) for OD pair we W and \i = (\iw,weW) f
. The firstorder optimality conditions for an optimum solution
* = {/L'r G R»>wG ^ )
t0tne
minimization problem (2.3)(2.5) are
Mathematical and Economic Theory of Road Pricing
16 3Z,(f»
= 0 and
31 (f*.
>0,
(2.7)
(2.8)
together with the nonnegativity condition (2.5) and the linkpath incidence relationship (2.1). With use of eqn. (2.1), (2.7) and (2.8) can be expressed explicitly as C{crw^) = 0,reRw,weW
(2.9)
cm[iw>0, reRw,weW
(2.10)
£ / ™  J w = ° > weW
(2.11)
where cm = ^ f
S^, r&Rw, w&W. It is readily verified that conditions (2.9) and
(2.10) together yield the following results cm =VW if /™ >0; c w >n.w if/r*w = 0, r e Rw,we W
(2.12)
These conditions imply that, at the optimum, the cost on any path with positive flow, i.e., f"w > 0 , must equal a constant, \iw, the OD specific Lagrange multiplier; otherwise it is greater than or equal to \iw. This constant is the minimum among all possible path costs as stated by (2.10). Hence, the Lagrange multiplier associated with a specific OD pair can be regarded as the minimum path travel time between that OD pair. It is thus clear that the firstorder conditions are equivalent to the Wardrop UE conditions: for each given OD pair, any actually used path has the same, minimum travel time. Therefore, with the assumptions on link travel time functions, solving mathematical program (2.3)(2.5) is equivalent to finding the UE linkflow pattern. It can be shown that (2.3)(2.5) has only one stationary point in link flow, i.e., the optimal solution expressed by link flow. In fact, the objective function (2.3) is strictly convex with respect to v, since its Hessian matrix, 0 0 t'2(v2) (2.13) V2Z(v) = 0
is positive definite. In eqn. (2.13) the prime superscript denotes the derivative of the travel time function to link flow, i.e., t'a (ya) = d/o (vo )/dv o , ae A, and \A\ is the cardinality of the set A. Clearly, all entries in the diagonal matrix (2.13) are positive because the travel time of each link is assumed to be a monotonically increasing function of its own link flow. Note that in deriving (2.13) we employ
Fundamentals of UserEquilibrium Problems
= a ; 0,othenrise)
17
(2.14)
vb This means the interactions of flows and travel times across links are temporarily not considered in UE problem (2.3)(2.5). The right hand side of (2.13) is also called the Jacobian of linktraveltime vector, t, with respect to link flow vector, v. The region of feasible solutions to mathematical program (2.3)(2.5) is made by linear equalities and inequalities. This region is hence a convex set. So, problem (2.3)(2.5) is a strictly convex mathematical programming problem which has only one linkflow pattern solution satisfying the UE requirements. However, it should be noted that this program is not strictly convex with respect to the path flow, hence there may exist more than one pathflow pattern solution, which generates the unique, optimal linkflow pattern solution. 2.2.2
Solution Method for the fixed Demand UE Problem
In the optimization theory, there are many algorithms suitable for solving the minimization problem (2.3)(2.5). Among these, the convex combination method, which is based on feasible directionsearching, has been most widely employed. In this method, the directionsearching step which originally corresponds to solving a linear program (LP) is reduced to a shortestpath problem. The reason is that the gradient of the objective function, Z(v), with respect to link flow at each iteration, e.g., the n'h iteration, equals the link travel time, i.e., 3z(v w )/3v o = t(vl"A. The LP problem at the «th iteration thus becomes
dz(v{)] min Z«(v) = Vz(vW) v = g  i  ^ v a = g ^ V .
(2.15)
subject to
Jm,>0,reRw,weW where Va = ^T
^
(2.17) sm e
fnfia *
path flow for path reRw
auxiliary link flow on link a, and fm is the auxiliary
connecting OD pair we W . The LP problem (2.15)(2.17)
minimizes the total network travel time with temporarily fixed link travel times, t^n) = ta(v[n)), a e A . Clearly, performing an allornothing assignment just reaches this target, which assigns all trips on to the corresponding shortest paths. Consequently, solving the LP problem (2.15)(2.17) is replaced by executing an allornothing assignment. The stepbystep procedure of the convex combination algorithm for solving the problem (2.3)(2.5) is given below.
18
Mathematical and Economic Theory of Road Pricing
Step 0. (Initialization) Perform allornothing assignment based on link travel time ta = ta (0), as A. This yields v(1). Let iteration counter n := 1. Step 1. (Updating) Update link travel times, ^"' = ta (v ( "), as A. Step 2. (Direction finding) Execute allornothing assignment to yield auxiliary link flows v. Step 3. (Line search) Find an optimal step size a(n) by solving mm > 0 4^
0
Step 4. (Updating) Update link flows, v) at v. The number of iterations required for convergence is significantly affected by the initial solution, the congestion level and the stopping criterion. In relatively uncongested networks, the flow pattern after a few iterations approaches the equilibrium state, since the updated link
Fundamentals of UserEquilibrium Problems
19
flows can not heavily change link travel times and then the set of shortest paths basically remains unchanged in the iterative process. As congestion builds up, more iterative works are required to equalize the costs on potentially possibleused paths. However, in most actual largesize applications with typical congestion levels, only six to ten iterations are usually sufficient to find the equilibrium linkflow pattern in terms of the tradeoffs among analytical accuracy, data limitations, and computer budget. All nonshortest paths are not assigned any flow. 2.2.3 The UE Model with Elastic Demand The traffic assignment models presented so far have assumed that the traffic demand for each OD pair is given or fixed. Road pricing has in general longterm impact on traffic demand and thus a sound assessment of a pricing scheme requires consideration of price elasticity of travel demand. Here we briefly introduce the traffic assignment problem with elastic demand that is pertinent to road pricing analysis. Suppose the OD demand in an OD pair is a function of the equilibrium OD travel time or cost between that OD pair (for simplicity and relevance in our subsequent studies, we consider separable demand functions unless mentioned specifically), namely dw = Dw(\iw) where \iw is the shortest path time or cost between OD pair weW. We further assume that the demand function is a strictly monotone, invertible and decreasing function of travel time. Let \iw = D~'(dw) = Bw(dw), weW denote the inverse or benefit function. The inverse or benefit function can be regarded as an amount that a user is willing to pay for his or her travel, or a benefit that he or she can obtain from this travel. The elastic demand traffic assignment problem is to determine a set of selfconsistent link flows and an OD demand solution satisfying demandperformance relations. The convex optimization problem is expressed below
min £ ^(ro)dco£ 15w(co)dco QEA
o
MeH
(2.19)
' 0
subject to %/„=0,reRw,wGW
(2.21)
where link flow va, ae A is given by (2.1).
20
Mathematical and Economic Theory of Road Pricing
Proceeding along the lines for obtaining (2.9)(2.12) yields, for any optimum solution (f *,v*,d*), the following relations cm = H w if /™ > 0; cm > n „ if fr'w = 0, r € Rw, we W Bw(d'w) = nw, if d'w>0; Bw(d'w)(f) can be minimized by each OD pair individually. The problem for OD pair we W is given by
Fundamentals of UserEquilibrium Problems
min Z (n) ({„) = X [0
reRw
(2.28)
XZ^
(2.29)
By inspection, the solution requires assignment of as much flow as possible (i.e.,Dw) for which \c("J B^yd^jl
is the smallest (most negative). One thus needs to determine the
shortest path c^ = min re ^ [c^ 1 Bw(^"))]
and the value of [c'fj  Bw(fi£n))] for each OD
pair. If this term is positive, no flow should be assigned to any path in this OD pair. This procedure will result in the lowest value of objective function (2.27). In summary, for a given OD pair, the solution to the auxiliary linear program (2.27)(2.29) can be simply obtained as \ ^ d  ) , r , [O, otherwise ™>5«r> ae
K
, ^ W
A
> d™ = 2 7%\
reRw
re
w& w
( 2  31 )
"»
The FrankWolfe algorithm for solving the Elastic Demand UE problem is summarized below: StepO: {Initialization)
Let \d{®\ wefFJ be a set of initial OD demands and
jv^0), a e A\ be the corresponding set of feasible link flows; Let e > 0 be a stop tolerance and set n := 0. Step 1: {Finding descent direction) For each OD pair w, weW, compute the shortest path r and the corresponding path cost c{"l based on the link travel costs f
a( v i" > )'
aS
A' obtain a set of auxiliary link flow and demand pattern
{vo 0 , r e R w , weW
(2.39)
va 0 only if va = Ca and Xa = 0 if va < Ca. We can naturally regard this multiplier, Xa , as a equilibrium queuing delay time, i.e., an additional time penalty (besides the moving time) that users traveling on this saturated link are willing to wait for continuously using this link. Introduction of link queuing delays explore a way to create equal path travel times among all
24
Mathematical and Economic Theory of Road Pricing
used paths (saturated or nonsaturated ones). The queuing delay on a link exists only when the link flow reaches its capacity, see Figure 2.2. It should be pointed out that although the equilibrium generalized path travel times, u,^, weW , are unique, this is generally not true for the optimal multipliers, Xa, aeA . A necessary and sufficient condition for obtaining a unique equilibrium queuing delay is that all binding link capacity constraints (i.e., holding with strict equality) are linearly independent. The multiplier ka, can be interpreted as a toll charge imposed on all commuters who use link a € A for commuting, then an unqueued equilibrium flow pattern can be obtained. This toll charge produces no losses for network users if the toll is not beyond the queuing delay, because it simply substitutes the wasted queuing time. 2.3.2
Solution Algorithm
We now turn to study the solution methods for program (2.32)(2.35). When applying the FrankWolfe algorithm to program (2.32)(2.35), the direction searching step is required to solve a linear multicommodity flow problem with inequality constraints, rather than the shortest path problem as done for problem (2.3)(2.5). This is prohibitively expensive to solve repeatedly. Most algorithms employ a strategy that converts the capacitated problem into a sequence of uncapacitated problems through a penalization/dualization of the capacity constraints (2.34), so that various existing efficient approaches for program (2.3)(2.5) can be applied for the solution of (2.32)(2.35). Here, we present an augmented Lagrangean dual algorithm. The penalty function methods, either in the form of an exterior penalty or an interior penalty, are often used to transform a constrained problem into an unconstrained or partially constrained problem. We set a penalty function for the link capacity constraint (2.34) as simple as P ( v , ) = I m a x 2 { 0 , vaCa},asA
(2.45)
Then, the new objective function which contains all penalty components becomes
where p is a penalty parameter (p > 0). For every specific p value, an uncapacitated traffic assignment problem will be solved, i.e., min Z(v) ve£2,
where Qv is defined by
(2.46)
Fundamentals of UserEquilibrium Problems Q v = { v  v = Af,Af = d,f>0}
25 (2.47)
where the link/path incidence matrix A = [8 ar ], the OD pair/path incidence matrix A = [A^,] where Am =1 if re R^ and 0, otherwise, and d = (dw,we W) is a column vector of all OD demands. Denote v(p) as the solution of problem (2.46) for a specific pvalue, v* as the solution of the original problem (2.32)(2.35), one can show that Z(v(p)) 0
(2.48)
limv(p) = v*
(2.49)
The optimal multipliers associated with capacity constraints, or the link queuing delay at equilibrium, A*, ae A , can be estimated below lim p
= X'a,aeA
(2.50)
It seems to us that now the original problem (2.32)(2.35) can be efficiently solved by applying the FrankWolfe algorithm to a uncapacitated traffic assignment problem (2.46) with a suitably large p value and the revised link travel time function ta (vo) + pdPa (v o )/dv a , a £ A. However, such a suitably large p value is not easy to find because of the numerical illconditioning inherent in penalty methods. In order to avoid this illconditioning, one may construct an augmented Lagrangean function as follows Lp(v,A) = Z(v) + £ — [max2{0, Xa +p(va  C J }  A a ]
(2.51)
where A is the vector of dual variables associated with capacity constraints. For every specific pvalue and A vector, denote v(p,A) as the solution of the problem: minZ,p (v,A) subject to v£ Q v , and Lp (X) = miny(nv Lp (v,A) as its objective function value. It is true that, for any p > 0, there exists A* that L p (A.)0/orfl//v6Q,
(2.56)
where Qv is defined by (2.47). Proof. First we note that the traffic equilibrium conditions imply that
Mf*)nJ[/™/:>0 For any fm>0.
Indeed, if fm>Q
(2.57) then [c w (f*)n w ]=0 by letting fm=0
f*m = 0 then r c r v v ( f * )  n l > 0 since fm>0.
in (2.57). If
Thus (2.57) holds. Furthermore, observe that
(2.57) holds for any path re R^, and any OD pair we W, hence summarizing inequality over all paths and OD pairs yields:
or
whereas
which, in vector form, yields (2.56).

We now show that any solution of (2.56) is an UE flowpattern. From inequality (2.56) and the equivalence relation (2.58), we have c(f*) T f>c(f*)Y f o r a l l f e Q /
(2.59)
Q /= {fAf = d,f>0}
(2.60)
where
This implies minc(f*) T f>c(f*) T f*
(2.61)
The lefthand side is a linear program, while the righthand side is a constant. From the inequality above it is clear that f * solves this lefthand side linear program. Let a. and r\ be
Fundamentals of UserEquilibrium Problems
29
the vectors of multipliers associated with constraints Af = d and f >0 in (2.60), respectively, for the linear program. If vector f * e £lf is a solution to the linear program or equivalently VI (2.66), then there exist (X and r\ satisfying c(f*)TiA T n = 0
(2.62)
riTf * = 0
(2.63)
Af"d =0
(2.64)
Tl > 0
(2.65)
From (2.62) and (2.63), if / r * w >0, then cm(f') = \iw for r&Rw. Also, if f'w = Q, then c
™ (f * ) = V^w + 'I™ w i m H ' ™  0> f° r r e Rw Hence, f * is an equilibrium flow pattern. ]
As seen from the above proof, the VI formulation can also be written with the pathflow and path travel time variables, i.e., to find f * e Q.f such that c(f*)T(ff*)>0 for all fe£i.f
given in (2.60) and c = (cm,rsRw,we
(2.66) W) with the following relation
between link and path travel time: c(f) = A T t(v)=A T t(Af) It is easy to determine that solutions to VI problem (2.56) or (2.66) exist, because the function t(v) is continuous and the feasible solution set Qv in VI (2.56) is compact (closed and bounded). Theorem 2.2 If the link travel time function vector, t(v), is strictly monotone, namely, [t(\l)t(\2)f
(y1 \2)>0, for any \\ v 2 eQ v andvl*v2
(2.67)
then problem (2.56) has a unique solution. Proof. Consider any distinct v1, v2 e Q v , we have (t(v1)t(v2))T(v'v2)>0
(2.68)
Suppose v1 and v2 are two distinct equilibrium solutions. Then, from VI (2.56), we have t(v')T(v2v')>0 t(v2)T(v'v2)>0 Adding the two inequalities yields
30
Mathematical and Economic Theory of Road Pricing
Since t(v) is strictly monotone, we have v1 = v2. ] Note that Ihe strictly monotone condition is equivalent to the positive definiteness of the Jacobian, V v t(v), of link cost functions. To see this, let v1, v2 e Q,v be two distinct link flow patterns, then there exists 0 < a < 1.0 such that t( v 2 ) = t(v') + Vt(vI + a ( v 2  v ' ) ) T ( v 2  v 1 )
(2.69)
and we have (t(v2)t(v1))T(v2v')=(v2v1)TVt(v1+a(v2v1))(v2v1)>0
(2.70)
and vice versa. 2.4.2
Symmetric Link Flow Interactions
Mathematically, symmetric link flow interaction is stated as d
^
d
^
*
b
(2.71)
The above equation means that the marginal effect on one link's travel time inflicted by another link's flow is equal to the marginal effect on that link's travel time given by this link's flow. In other words, the Jacobian of the vector of link travel time functions is symmetric. In this case, an equivalent mathematical program that can generate a UE flow pattern can be constructed as follows min Z(v)=J t(ra)dco
(2.72)
where £lv is defined by (2.47). The objective function (2.72) is defined by a line integral, which can be converted to the sum of a series of standard integrals, each with a single integrated parameter. To show this, we consider a wellused example with twoway streets where the travel time experienced by a user depends not only on the traffic volume in his direction (denoted by a) but also on the traffic volume in the opposite direction (denoted by a'). So, the travel times on link a and link a', respectively, can be written as Uv) = '0(va,v,), tA^) = tAva;va),aeA,a'eA
(2.73)
After conversion, objective function (2.72) becomes 1 [ 1 min Z(v) =  £ j\(co,v,,)dco+jr,,(a>,0)dco
(2.74)
Fundamentals of UserEquilibrium Problems with
condition
dta(\)/dva.=dta(\)/dva,
aeA,a'eA.
It
is
31 easy
to
show that
dZ(y)/dva =ta{va,va.), which is important in deriving the firstorder optimality conditions equivalent to the UE conditions. Furthermore, the fact that the Hessian of objective function Z(v) is the Jacobian of link travel time functions, implies that program (2.72) will have a unique solution if the Jacobian is positive definite. 2.4.3
Flow Interactions between Multiple Types of Vehicles
The interaction between different types of flow (e.g, cars, trucks, buses) when sharing the same roadway induces nonseparable and, in general, asymmetric link cost functions. Let K denote the set of flow types. Let v* be the flow on link a associated with flow or vehicle type k and t'a (v) be vehicle typespecific link travel cost function on link a, where v = (v*,k& KJ and v* = (v*,ae A) ,ksK.
Note that the link cost function here includes
interactions between different types of vehicles sharing the same links and between flows on different links examined in the previous subsection. It is reasonable to assume that the demand by users of one type of vehicles depends only on its own price. Let 5* (d) denote the OD benefit (inverse demand) functions where d = (d*,&e K^j and dk = {dkn,we.W} . Likewise, the elastic demand UE problem with multiclass users in vehicle types can be formulated as the following VI: determine (v*,d* j that satisfies (v,d)EQ
(2.75)
where Q=[(\k,dk),keK\
v*=Af*, d*=Af\ f*> 0,d*> o}
with f* =(/*„,re Rw,wz wf,
(2.76)
ksK.
Note that the multiclass traffic equilibrium model with several vehicle types can be reduced to a single user class model with link flow interactions by associating each class with a complete network copy. Then, the problem reduces to what we have discussed in the previous subsection. 2.4.4
Positive Definiteness of Link Cost Functions with Flow Interactions
As shown above, positive definite Jacobian of link cost functions is sufficient to guarantee a unique link flow solution for both symmetric and asymmetric traffic equilibrium problems. A
32
Mathematical and Economic Theory of Road Pricing
typical situation that gives rise to a positive definite Jacobian matrix Jy = V v t(v) is that both J v and JVT are diagonally dominant and each diagonal element is positive. Specifically, the positive definiteness of the Jacobian of link cost functions with all nonnegative elements is guaranteed if the following conditions hold: (i)
(Positive diagonal elements of the Jacobian) The travel time on each link is an increasing function of its own flow: dt (v) i U >0 ) ae^ (2.77)
(ii)
(Rowdiagonally dominant)
^M>^djM,aeA (iii)
(2.78)
(Columndiagonally dominant) ^
>^
*
, a€ A
(2.79)
Conditions (2.78) and (2.79) implied that the main dependence of a link's travel time is on its own flow, and the direct effect of flow of a link on its own travel time is much larger than its effect on any other link travel time: dt (v) dt (v) 3/ (v) dt.(\) l U aK » ', aK ' » bK ; , for any ft* a (2.80) dva dvb dva ava In the case of symmetric link flow interaction, conditions (2.78) and (2.79) are identical since )da>+£ £  U > a
(2.87)
subject to JJfZ=d:,weW,meM
(2.88)
fZ>O,rBRw,w€W,meM
(2.89)
where / ^ is the flow of user class m on path reRw,w eW, and v™ and va are defined as
v:=EE/X,^^M
(2.90)
va = ^v:,asA
(2.91)
From the firstorder optimality conditions of the convex optimization problem (2.87)(2.91), we can easily establish the following equilibrium relations in terms of the minimization of travel disutility in time units:
36
Mathematical and Economic Theory of Road Pricing
IX( V J 5 « +Y—8 =V""™,tffm>O,reRw,weW,meM _^j aeA
\
a/
ar
^^ o aeA tJm
ar
(2.92)
w
S ' . W 8 " + S^5 ar >^: >ttae , if / : = 0 , re ^weW,
me M
(2.93)
where u™1"1™ is the Lagrange multiplier associated with (2.88) and equals the minimum travel disutility in time units (inclusive of equivalent time of toll charge) between OD pair w by users of class m , i.e., n™>time = m i n ^ { C ™ } . Finally, we note that the objective function (2.87) of the UE program is strictly convex in aggregate link flow if the link travel time function ta(va) is monotonically increasing, but linear in the classspecific link flow. Thus, the aggregate equilibrium link flows are unique, but the equilibrium link flows by user class are not unique. 2.S.2
Equivalence of the Timebased and Costbased MultiClass Traffic Equilibrium
The equivalence of the timebased and costbased multiclass traffic equilibrium depends on users' tradeoffs between time and money. The travel disutility of a path re R^,, weW consists of two terms: travel time denoted by tm and travel cost (money) crw. We now define a general value of time function or simply VOT function, gm(t), that transforms a time of t units into equivalent amount of monetary cost for user class me M. Then the following assumption is generally true. Assumption 2.1 The VOT function gm(t) for any me M is a continuous and strictly increasing with respect to t for t > 0 and gm (0) = 0. Assumption 2.1 implies that t = g"1 (x) exists for x >0. With this general VOT function, the travel disutility of path re Rw is given by c^cost =um+gm(tm) cm,t,me _
^ + g^{um)
in cost unit, and is given by
in time unit. Now we have the following equivalence theorem.
Theorem 2.3 The timebased and the costbased equilibrium solutions are identical in any general network if and only if gm{t) is a linear function of t for all me M. Proof: Note that, from the assumption that gm(f) is linear and strictly increasing with 8m (0) = 0, the VOT of users of each class m must be a positive constant. Let Pm > 0 denote this constant VOT for user class m as before. Each user in each class is assumed to choose a
Fundamentals of UserEquilibrium Problems
37
route that minimizes his or her generalized travel cost c^;""" defined below (time is transferred into money according to VOT) ,K)
+
"0
(3.58)
The symmetry assumption is not necessarily needed; it makes the conventional optimization formulation of the UE problem possible. 3.6.2 Uniqueness of Link Flows with Separable Link Cost Functions Now we move on to look into the terms relating to the link flows and link cost functions in the SO problem. In the case of separable link cost functions, continuity and monotonicity are sufficient to ensure a unique link flow solution to the UE problem. Nevertheless, the SO problem requires one additional condition that each link travel time function is convex in order to ensure a unique link flow solution. To see this, we examine the Hessian matrix of the objective function. The Hessian matrix is a diagonal one for the separable link cost functions with entries as follows 2 dOO + v,
dvo
i t o
(3.59)
aBA
dvj
Clearly, the first term is positive since the link travel time function is assumed to be continuous and monotonically increasing. The second term is positive too if the link travel time function is convex. Then, the Hessian matrix is positive definitive and the SO programs presented in Section 3.23.5, involving separable link cost functions, have a unique link flow solution. Here one should particularly pay attention to the fact that, even in the simple case of separable link cost function, the term, ta (ya) va or "£ ta (va) va of the SO problems may not be convex if the link cost function is monotonically increasing but not convex, and hence the SO solutions may not be unique, as shown in the following example. Example 3.1 (Unique UE but multiple SO solutions) Set the separable link cost functions
where a>4
is a constant. Note that both A( V I) a n d h^i)
are
functions. Let the demand d = 1. The UE objective function is given by
Z UE (v 1 ,v 2 )=Z UE (v 1 ,lv 1 )=J(l^)d«+?(l^ t0 )d»=l +
strictly increasing
Mathematical and Economic Theory of Road Pricing
66
Since the model is symmetric, we conclude that the equilibrium flow is given by v* = v* = 1/2. Now look at the derivative of the total social cost 2s0 (v,, v2) at this point. Since v, + v2 = 1, we can consider 2 s0 (v,,v 2 ) as a function of one variable:
It is easy to see that dZ s o (v,,lv,) dv,
1.10 

1
= (4aa 2 )e" o/2 4
= 0,
\
UE
/
—_
J
1.05 
1.00 O o u 0.95 Vali
—so
0.90 •
0.85 0.00
0.20
0.40 0.60 Flow on Link 1
0.80
1.00
Figure 3.7 Curves of the UE and SO objective values versus link flow v, (0 < v, < 1.0) Thus, we can see the UE in this network coincides with a local maximum of the total social cost. Moreover, it is easy to observe that the social optimum is not unique: there are two points with the same minimum value of the social cost. These points are symmetric with respect to 1/2. Figure 3.7 shows the values of UE and SO objective functions when a = 8, where the SO objective value reaches its minimum at two points, i.e., v, = 0.1265 and v, =0.8735. Clearly, the reason for the nonuniqueness of the SO result in this example is the nonconvexity of the social cost, as shown in Figure 3.7. The nonconvexity of the social cost is due to the fact that the link cost function is not convex. This example confirms that even if the link travel cost functions are strictly monotone, the social cost, 7C(v), may not be convex. 3.6.3
Uniqueness of Link Flows with Nonseparable Link Cost Functions
The General FirstBest Pricing Problem
67
In the presence of nonseparable link cost functions, ta (v), ae A, the vector of link cost functions, t(v), is said to be convex if each ta(v), as A is a convex function for v>0; t(v) is said to be strictly monotone if for any distinct v1, v 2 eQ v ,we have ( t ( v 2 )  t ( v ' ) ) ( v 2  v ' ) > 0 , where Qv ={v v = Af,Af = d,f>0}. Alternatively,
t(v)
is strictly monotone if and only if its Jacobian V v t(v) is positive definite. As mentioned in Chapter 2, the strictly monotone condition or positive definite Jacobian of link cost functions is sufficient to ensure a unique link flow solution to the UE problem. In contrast, the social cost for the SO problem is given as TC(y) = ^
ta (v)va = t(v) T v
and its convexity, to ensure a unique SO link flow solution, is somewhat complicated. First, differing from the case of separable link cost functions, the convexity of the vector functions t(v), in the case of nonseparable link cost functions, does not necessarily imply the convexity of rC(v) = t(v) T v for v>0. For example, linear (and hence convex, but not strictly convex) and nonseparable link cost functions do not guarantee the convexity of 7C(v), as seen from Example 3.2 in the next subsection. Second, strict monotonicity of the vector functions t(v) does not imply the convexity of TC(y) either. This is true even if the link cost functions are separable, as shown in the above Example 3.1. Nonetheless, let us consider the Hessian of TC (v):
dvadvb
dvb
dva
£}
dvadvb
Suppose the secondorder cross partial derivatives can be omitted in relation to the rest of the terms, we have ^ dva2 dvadvb
+
v
dvb
( 3. 6 3)
dv2
dva dva
b
(3.64)
The conditions for having a unique SO link flow solution in the special case of symmetric link flow interactions can then be stated below.
68 (i)
Mathematical and Economic Theory of Road Pricing The travel time on each link is an increasing and convex function of the flow on that link, that is, dva
(ii)
'
dva2
0, aeA
(3.65)
The main dependence of a link's travel time is on its own flow.
Condition (3.66) together with (3.65) ensures d2TC(\)/dva2 > £ ^
^d2TC(\)/dvadvb.
By comparing the above conditions for the SO problem with those for the UE problem in subsection 2.3.2 of Chapter 2, one can see that the additional condition for a unique SO link flow solution in the case of symmetric link interaction is that the travel time on each link is a convex function of its own flow. This observation is similar to that in the case of separable link cost functions. As discussed in Chapter 2, one may expect that the symmetric situation may appear in the flow interactions between different links. Nonetheless, in the case of interaction between multiple types of vehicles sharing the roadway, as mentioned already, the Jacobian of link cost functions becomes asymmetric; their strict monotonicity or their positive definite Jacobian does not guarantee the positive definiteness of fee Hessian or the convexity of the SO objective function TC(v), and as a result, the SO link flow solutions may not be unique. It seems there is little hope to establish the general sufficient conditions for the convexity of TC(\) in terms of the properties of asymmetric link cost functions t(v). Nonetheless, we now examine the convexity TC(\)
for the specific asymmetric link cost functions
introduced in Section 2.4 of Chapter 2:
taT = t:TL + a^^±^j\
^=c{l.0+a^^±^Ji
where all variables and parameters are defined before. The Hessian of TC(\) consists of 2x2 diagonal subblocks with each corresponding to one link. Hence it suffices to check the positive definiteness of each subblock. In view of Ta = tacvac + ( a r v a n we have:
(3.67)
The General FirstBest Pricing Problem
69
dv dvT ac
dTa2
at
dT[_ d\.
dv
 + v_.
2
V
aT
dV
ac
"Cdvacdvar
"^vjv^
. dt _ dt d2t T 2 —+v — + v ——
"*dj
'T*
2
where the specific partial derivatives can be obtained from (3.67) and are given below:
ac
ac
\cl
3v,
CaT
CaT
CaT dvj
c ac
'2)
is of course strictly monotonic since for any distinct ,
where
xl+x2=5
and
yt+y2 = 5,
we
have
( t ( v 2 )  t ( v ' ) ) ( v 2  v ' ) = 4 ( > ' 2  x 2 ) 2 > 0 . The unique UE solution is: v*=3, v*=2, t' = t2 = 21. The total travel cost is given by rC(v 1 ,v 2 ) = (2 + 5v 1 +2v 2 )v 1 +(4+3v 1 +4v 2 )v 2 =5(v,) 2 +4(v 2 ) 2 +5v 1 v 2 +2v,+4v 2 Since v , = 5  v 2 , we have TC = 4(v2) 23v 2 +135 which is a convex function. The unique SO solution is: v* = 2.125, v* =2.875, TC = 101.94.
3.7
MARGINALCOST PRICING UNDER STOCHASTIC USER EQUILIBRIUM
3.7.1 Economic Benefit Measure and Maximization As already mentioned in Chapter 2, with the logitbased route choice model, the expected indirect utility received by a randomly sampled individual can be expressed as Sw = E\max(Qcm + & j l = In £ e x p (  9 c J
(3.69)
This measure has an economic interpretation related to consumer surplus. For example, in the absence of income effects, a change in price or any other characteristics of the travel environment results in an expected change in consumer's surplus which is equal to the change in the above welfare measure. According to the representative consumer theory of the logit model (Oppenheim, 1995), the behavior of consumers with different tastes can be described by the choices made by a single individual who has a preference for diversity. The direct utility function of the representative consumer or user (i.e., the utility corresponding to the aggregate demand) can be expressed as follows:
The General FirstBest Pricing Problem
75
The utility function (3.70) consistent with the logit model is an entropytype function which has also been used as a benefit measure in terms of interactivity in trip distribution (Erlander and Stewart, 1990). On the other hand, the total social cost incurred by all users in the network is given by Y t (v )v . The net economic benefit could be measured as the traveler's benefit minus the total travel cost, and the socially optimal pricing requires that the net economic benefit is maximized. Observing that the problem can be transferred into a minimization problem by changing the sign of the objective function, thus, after deletion of the constant term, we have the following minimization program:
%fr=dw,weW
(3.72)
reRw
fr>0,reRw,weW
(3.73)
where the demand dw between OD pair we W is fixed and given and the last term is constant (included for later use). Note that the objective function (3.71) is strictly convex, thus the following KuhnTucker conditions for any / ^ > 0 are also sufficient to obtain the unique optimal solution:
where Xw is the Lagrange multiplier associated with constraint (3.72), and
t W) = t (O + v ' ^ J , ^ '
^'
aeA
(3.75)
dv a
Evidently, the first term of the righthand side of (3.75) is the actual link travel time incurred by a traveler and the second term is the additional travel time that a traveler imposes on all other travelers in the link. It can thus be seen that user externality is present due to the effect of congestion on the network links. Define
C = IS.[e(^C)l
(3.77)
Solving this equation and constraint (3.72) yields: I
f*
exp(0c* )
reRweW
(3.78)
76
Mathematical and Economic Theory of Road Pricing
From the above derivation, we can observe that the optimality condition (3.74) for an optimum of the economic benefit maximization problem (3.71)(3.73) is also the condition that governs the logitbased stochastic user equilibrium (SUE) model (3.78). The only requirement for this equivalence is that the link cost and the route cost should be given by (3.75) and (3.76) where the additional cost or user externality specified by the second term of eqn. (3.75) is included. This means that from the viewpoint of social welfare maximization, user externality exists in the system optimality conditions. Therefore, by imposing a toll at each link exactly equal to the externality, we can ensure that the users' optimal private choices will also be optimal social choices in terms of the maximization of net economic benefit. Consequently, the classical principle of traditional marginalcost pricing is still applicable in the logitbased SUE situation. 3.7.2
Equivalence to the Minimization of Expected Perceived Travel Cost
In fact, maximization of the net economic benefit (3.71) is equivalent to the minimization of the total expected perceived user cost in the logitbased stochastic route choice model. To see this, we now change the objective function (3.71) as follows:
(3.79)
From the logitmodel in Chapter 2: fm/dw = exp (6c TO )/^ fe/i exp(6ciw), re Rw,weW we have  l n ^  l n ^ e x p t  e c J , reRw,weW
(3.80)
Substituting (3.80) into (3.79), we have
6
"*"**•
L
*•0
(4.7)
for all (v,d)eQ where £2 is defined in (4.4). It is assumed that t(v*,e) and B(d*,e) are once continuously differentiable in e . Typically, the development of sensitivity analysis for a traffic equilibrium problem involves several technical difficulties. Due to the special structure of the problem, its VI formulation usually includes path flow variables. However, the fact that the path flow pattern is usually not unique at equilibrium prohibits the direct application of the VI sensitivity analysis approach to the traffic equilibrium problem. Here we introduce a restriction approach in which an equivalent restricted problem for the network equilibrium problem is developed which has the desired properties required by the general implicit function theorem. The approach presented here is a rectified version of the original restrictive approach proposed by Tobin and Friesz (1988) that contains a few serious flaws. The approach is easily accessible by using a rather intuitive implicit function theorem and can be used as a popular tool for producing gradient information needed in traffic optimization analysis. 4.3.1 Sensitivity Analysis of the Restricted Network Equilibrium Problem The restricted network equilibrium approach is intended to derive the derivative expressions and has the following advantages: 1) the information on traffic assignment results can be effectively utilized for calculation of derivative values; 2) the derivatives of link flows, OD demands, OD costs and other solution variables can be obtained and expressed explicitly in terms of the equilibrium flow solutions. Now we make the following assumption: Assumptions 4.1
The functions B(d,e) and t(v,e) are positive and strictly monotone
The General SecondBest Pricing Problem
85
in d and v (decreasing and increasing) for d > 0 and v > 0, respectively, and once continuously differentiable in (d,e) and (v,e), respectively. We assume that we already have a solution d*(0), v*(0) and f*(0) to the above perturbed problem (4.7) for e = 0. This solution is unique under assumption 4.1. In addition, under the strong monotone assumption, the equilibrium flows and OD costs vary continuously with perturbations of the link cost and OD demand functions (Hall, 1978; Dafermos and Nagurney, 1984). We also assume that at equilibrium the demand between each OD pair is strictly positive, d'w = Dw \\i*w) > 0, we W. To overcome the difficulty of the nonuniqueness of the path flow in the network equilibrium problem, the equivalent restriction approach is to select a maximal set of independent paths in the feasible region of equilibrium path flows. Here we focus on the equilibrated paths or the paths with the minimum travel cost in relevant OD pairs only in the network. Let A and A be the link/path and OD/path incidence matrices associated with the equilibrated paths only. Then any nonunique equilibrium path flows at e = 0 must be contained in the following set: Q(0) = Q(e = 0) = {7 Af=v*(0), Af = d*(0),f >0J
(4.8)
where v*(0) and d*(0) are unique link flow and OD demand solutions at e = 0. Then the linearly independent paths selected correspond to the linearly independent columns of
rsi the matrix
 .
The linearly independent equilibrium paths can be obtained easily if the convex combination method (diagonalization algorithm in conjunction with the FrankWolfe method for the general asymmetric traffic equilibrium problems) is used to solve the network equilibrium problem. The FrankWolfe algorithm generates a unique set of minimum time paths between each OD pair at each iteration. If the paths generated are saved from iteration to iteration, upon termination the algorithm provides an equilibrium path flow pattern and a link/path incidence matrix for the paths used, from which a maximal set of linearly independent paths can be identified (note that this is not always feasible, and a postselection of equilibrated paths might be needed after executing an equilibrium traffic assignment). At this point, we can assume that at equilibrium every link carries a positive flow, v* > 0, a e A in the righthand side of system (4.17). Otherwise, we can eliminate such links
86
Mathematical and Economic Theory of Road Pricing
with zero flows in the sensitivity analysis, because, by Assumption 4.2 and uniqueness of link flow, any link with zero flow must be on the nonequilibrated paths only, which are already eliminated from our further analysis without effect on the derivation results. In this case, under Assumption 4.2, links with zero flows at e = 0 will continue to have zero flows for £ near zero and the relevant derivatives in perturbations are equal to zero. To proceed, we now introduce the following assumption. Assumptions 4.2 +
f,
+
There exists a set of strictly positive equilibrated path flows denoted as
+
7 = 7 ( 0 ) e Q ( 0 ) at 6 = 0.
Assumption 4.2 implies that, the path flows on the minimum cost paths are not unique, but we can at least find a set of positive flows from the set of equilibrium path flows on the equilibrated paths. Now we look at the general necessary conditions for the perturbed network equilibrium problem (4.7) at e = 0. The existence of equilibrium is that there exists a solution f * of the full path flow vector to the following system equations: c(f*,0)7iA T n = 0
(4.9)
7iTf*=0
(4.10)
Af*D(n,0) = 0
(4.11)
f*>0 7t>0
(4.12) (4.13)
where c in (4.9) is a full vector of all path cost functions, and n is a vector of multipliers associated with the nonnegativity condition of path flows. Under Assumption 4.1, the continuity of the equilibrium link flows and OD costs in perturbations of the link cost and OD demand functions is preserved. Thus, the nonbinding nonnegative constraints in (4.13) remain nonbinding near e = 0, and can be eliminated without changing the solution near e = 0, or tie Lagrange multipliers, n, associated with the nonbinding constraints (nonequilibrated paths) at e = 0 are strictly positive and remain strictly positive for e near zero, and hence all nonequilibrated paths (nonshortest paths) are out of the question. Therefore, we have the following equivalent system of equations: c(f*,0)AV = 0
(4.14)
A7*D(n,0) = 0
(4.15)
f*>0
(4.16)
The General SecondBest Pricing Problem
87
By our assumption 4.2 we have a strictly positive flow f e Q, or we have a strictly positive solution f > 0 to the following linear system of equations:
V(o)
A f=
(4.17)
d*(0)
A
We now choose a maximum set of linearly independent columns or (equilibrated) paths in 
. Denote the set of equilibrated, linearly independent paths as R and the
corresponding path variables as f , and the further reduced link/path and OD/path incidence matrices as A and A . Equation (4.17) can be rewritten as:
"A A T f U V * ( 0 ) l c
(4 18)
d
A A J[fJ L *(°)J where ' c' denotes the corresponding 'complementary' matrices and vector associated with the equilibrated, but dependent paths. For sufficiently small E near zero, we can always fix the complementary or nonbasic path flow variables as fc = fc+ and solve the following linear systems of equations for f, for any £ near zero: (4.19) Note that the above system may generally contain linearly dependent equations (not of full row rank) in view of the fact that the network link flow vector must satisfy the flow conservation conditions at the nodes. Notwithstanding, because

A
is of full column rank,
LJ for fixed fc = fc+, the solution f(s) is uniquely determined and varies continuously with e in view of the fact that both V*(E) and d*(£) are continuous functions of E.Byour Assumption (4.2), f (0) = f+ > 0 , we thus conclude that f(e)>0 for E near zero. As we already show that, by fixing the flow of each equilibrated and dependent path (equal to a positive value) f c =f c+ , flow variables f are positive in the reduced, linearly independent system (4.14)(4.16) at e = 0 and will remain so for perturbations of £ near zero. This means that the nonnegative constraints (4.16) on f are not binding and can be eliminated without changing the solution in a neighborhood of E = 0. Consequently, it is sufficient to consider the equilibrated and linearly independent working paths only, and the system (4.14)(4.16) thus reduces to: c(f*,0)A T n = 0
(4.20)
Mathematical and Economic Theory of Road Pricing Af % D(u,0) = 0
(4.21)
where c(f*,OJ represents the corresponding reduced vectors (cost vector of the chosen linearly independent, equilibrated paths). By our Assumption 4.1 of the differentiability of link cost and OD demand functions, differentiating both sides of the system of equations (4.20) and (4.21) with respect to perturbations e yields: A T
L
Vec(f,o)"
v
A
(4.22)
~^
where the Jacobian
V f c(f,0) A
AT VHE
(4.23)
is well defined and Vfc(f*,0) = ATVvt(v*,0)A Theorem 4.1
(4.24)
Under Assumptions 4.1 and 4.2 and the linear independence of the working
path set R, the Jacobian, J ? (0) in (4.23) is nonsingular. Proof. To prove Theorem 4.1, it suffices to prove that all the columns of the Jacobian matrix are linearly independent. Consider a nonzero vector A,= (A/,A.1) 5*0, where A, is a column vector with the same number of elements as that in f and A. is a column vector with the number of elements equal to the number of OD pairs (number of rows in A , A or
A). Let [ J f lA, = 0, and we have Vfc(f*,0)A"ATA. = 0
(4.25)
AA.V11D(n,0)A. = 0
(4.26)
Multiplying both sides of eqn. (4.25) by AT (from the left side) and using eqn. (4.24) yields: (A£)TVvt(v*,0)AA~(A£)TA. = 0
(4.27)
Substituting eqn. (4.26) into (4.27) yields: (AX)TVvt(v*,0)AX + AT[V(1D(^0)]A. = 0
(4.28)
From Assumption 4.1, the link cost function vector t(v,e) and the demand function vector
The General SecondBest Pricing Problem
89
D(n,e) are strictly monotone (increasing and decreasing) in v and \x respectively, eqn. (4.28) thus implies that AA = 0 and A = 0. Here we still need to show that A must equal A
zero. Since A. = 0, we have AA = 0 from eqn, (4.26). By our early assumption that
consists of only independent columns, we thus conclude that A = 0 from AA = 0 and AA = 0. These results contradict our assumption of the nonzero vector: A = (AT, AT J * 0. Therefore, the Jacobin matrix \Jf
1 in (4.23) is invertible. This is also true in the case of
fixed OD demand where V^D (n, 0) = 0.
]
From the early assumptions and Theorem 4.1, the conditions for the general
implicit
function theorem are met, and from eqn. (4.22) we arrive at:
r  ^ l
(429)
VED((i,0)J Let (4.30) Then the derivatives of the working path flows with respect to E at 6 = 0 are Vef = B n V E c ( f \ 0 ) + B12 VED (n, 0) In view of VEv = A V / + A c V E f
(4.31)
and V E f c = 0 as well as V E c(f*,o) = A T V e t(v',0),
tne
derivatives of link flows with respect to e at e = 0 are obtained as: VEv = AB n A T V e t(v\0)+AB 1 2 V 6 D(u,0)
Theorem 4.2
The values of VEv as calculated in (4.32) are independent of the choice
of the equilibrated and linearly independent paths pathflows
Proof.
(4.32)
R and the specific values of positive
f + = f ~f
c+
The former independence of the choice of R is implicitly due to the fact that the
basic linear system of equations (4.19) holds at equilibrium for the values and small variations of link flows and OD demands, and path flows for any maximal number of linearly independent equilibrated paths R. The latter independence of the specific values of equilibrated positive path flows is selfevident from all the sensitivity analysis equations, the only term that involves the selected path flow values is the Jacobian of the path cost vector in
90
Mathematical and Economic Theory of Road Pricing
(4.24), which can be uniquely evaluated by the Jacobian of the link cost functions and the reduced link/path incidence matrix associated with the selected path set. To see this explicitly, like Tobin and Friesz (1988) we consider £=(••,£,,•••)
and a unit vector ei of
the same dimension of £, with unity in the i th position and zeros elsewhere, and 8 > 0 is a scalar. By definitions and assumptions, for any link a& A we have:
= X X f™ ( e ) ^
(Because nonequilibrated path has zero flow)
= X X L (6)8, + X X/™ ( E ) 5  ^ definition in (4.18)) tion = X X /™ ( e ) 5  + X X /™8  (by assumptic The secondterm is constant. Thus,
This completes the proof. ] Theorem 4.2 states that the value of Vev is uniquely defined irrespective of the choice of the equilibrated, linearly independent paths and their specific positive values without perturbation. This means that link flows are differentiable with respect to a variety of perturbations of the link cost and OD demand functions and their welldefined derivatives are obtained in eqn. (4.32). Other derivatives, such as that of OD demand in perturbation parameters, can be obtained below. From eqn. (4.22) the derivatives of OD travel cost with respect to e at e = 0 are: Ven = B21ATVEt(v*,0) + B22VeD(u,0)
(4.33)
We thus have Ved = VeD(j.,0) + VtlD(^,0)Ven
(4.34)
where Vtj. is given by eqn. (4.33). Note that, in the case of specific applications, the derivative expressions (4.32) and (4.34) can be considerably simplified. In particular, when the perturbation parameters are specified,
The General SecondBest Pricing Problem
91
the derivatives VEt(v*,0) and VeD(u,,0) can be obtained in a straightforward manner. Of course, the sensitivity analysis formula presented so far does apply for the special case with fixed travel demand, but the perturbation parameter can then appear in the demand vector d(e).
In this case VeD(n,0) = VEd(0) in (4.22), (4.29), (4.32) and (4.33), and
( ) = 0 in (4.22), (4.36) and (4.34). 4.3.2 Relations to Other Gradientbased Sensitivity Analysis Methods In the sensitivity analysis method by Tobin and Friesz (1988), an equilibrium path flow vector f * is chosen such that it is an extreme point of the polyhedron £2(0) in (4.8), which has exactly as many paths with a strictly positive flow as the rank of the matrix

(such a path flow vector is called a nondegenerate extreme point of Q(0)). An extreme point of the path flows at e = 0 in this case correspond to the solution of our system (4.18) by setting fc = 0. Namely,
Unfortunately, there is no guarantee that the solution to the above system is strictly positive (some are zeros) no matter which linearly independent columns in

are selected from
 . This exactly corresponds to the case where the number of paths with strictly positive flows at any extreme point of Q(0) cannot equal the rank of
— • The latter observation
was first shown by Josefsson and Patriksson (2003) through a simple example, which will be further presented later. The Jacobian of the working path cost vector, Vfc(f*,o) = ATVvt(v*,o)A, is in general not invertible, and we have to work directly with the inverting of the Jacobian in (4.23). The reason is simple. Let the number of (linearly independent) paths included in A be m and the number of links in the network with positive flow be n. Then ATVvt(v*,o)A is an nxn matrix, but its rank is m if m0
(4.43)
for all r<ERw,weW. Condition (4.43) implies that f'w>0
if cnv(f*)e)ji'w = O, and if
f'w = 0 then crw(f*,e)p.^ >0. This strictly complementarity condition is used together with the strict monotonicity condition as sufficient conditions for the differentiability of the traffic equilibrium link flows. Nevertheless, the above strict complementarity condition is not well defined for the traffic equilibrium problem, because of the nonuniqueness of the equilibrium path flows. Let us consider the following example network that often appears in the literature. Example 4.1
Consider the following network shown in Figure 4.1, with a single OD pair
from node 1 to node 3, with a fixed demand of rf,_)3 = 2 (flow units). The link cost functions are given by: tl = l + vl+ui, with link 1 subjected to a toll charge. 1
Figure 4.1
t2=l + v2, /3 = l + v 3 , t4=\ + v4
(4.44)
3
The network used in Example 4.1
There are four paths denoted by: rx ={l,3}, r 2 ={l,4}, r 3 ={2,3}, r4 ={2,4}. Without toll charge u\ = 0, we have an obvious unique (unperturbed) equilibrium link flow solution: v* = (1,1,1,1) . All four paths are the minimum cost path between the single OD pair and the equilibrium path flows are not unique but can be expressed as follows:
j;=p, 7 2 '=lp, 7 / = l  P , 7;=P, 0 +°
Likewise, when only link 1 with positive flow is included, applying the sensitivity analysis formula yields: Vu v = (0,0) . These results are consistent with our earlier observation that the link flow is differentiable at u, = 1 (with identical left and right derivatives). At this point it is still too early to say anything definite and the following example perhaps offers a further clue to the above observation.
Link 1
V_y
Link 2
Figure 4.5 A simple network of three links connecting three OD pairs Example 4.5 Consider a simple network again depicted in Figure 4.5 with 3 nodes, 3 links and 3 OD pairs: 1—>2, 2—»3 and 1—>3. The link cost functions are shown in the
100
Mathematical and Economic Theory of Road Pricing
network, with link 1 subjected to a toll charge w,. The OD demands are fixed and given by rf,_,2 = ^ 3 = c?1_>3 =2. There are four paths denoted by: rx ={l}, r2={2],
r 3 ={l,2},
The equilibrium link flows as a function of link toll «, is given below: Mi 3—L,
[ 0<M, 6
U, 1 +—,
0.,]'=
1 0 0 1 0 0
0 1 0 0 1 0
01 0 0 0 0  1 0 01 4 0 0 0 0 0 0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 1
1 0 0 1 0 0
0 1 1 0 1 0 0 1 1 0 0 4
"1 0 0" "0 0 0" "1" "0" = AVcf = 0 1 0 0 0 0 0 = 0 0 0 1 0 0 0 0 0 which once again exactly agrees with the right derivatives of equilibrium link flows in (4.51). From the above simple example, we do find that application of the gradient formula can still provide meaningful information. It provides the left and right derivatives of link flows (if not differentiable) and the desired derivatives (if differentiable) when including or not including the degenerate path. This is indeed true when there exist one degenerate path and one perturbation only, as in the above examples. An intuitive interpretation is given here. The equilibrium link flow is piecewise differentiable, and the left (or right) derivative with respect to a perturbation corresponds to the case that the perturbation just causes the degenerate path to be abandoned or to come into use. Therefore, the inclusion or exclusion of the degenerate path provides either the right or left derivative of equilibrium link flows.
102
Mathematical and Economic Theory of Road Pricing
Note that the above observation cannot be generalized to the cases where a degenerate equilibrium point contains two or more degenerate paths (this is very unlikely in practice). The reasons are as follows. When there are two or more degenerate paths, the left (or right) derivatives of equilibrium link flows correspond to a particular inclusion (or exclusion) of a subset of the degenerate paths, but not inclusion (or exclusion) of any subset of the degenerate paths can provide the desired results. It is thus necessary to identify appropriate or relevant degenerate paths for the analysis of the left and right derivatives. vi)
Non Invertibility of the Jacobian of the cost vector of independent working paths
Here it is worthwhile to emphasize that the Jacobian of the cost vector of the equilibrated, linearly independent working paths: Vfc(f*,cA = ATVvt(v*,0)A that appears in the sensitivity analysis formulas is in general not invertible (and, of course, not invertible at the extreme point of the feasible region of the equilibrium path flows considered in Tobin and Friesz (1988)). The reason is clear. Let the number of (linearly independent) paths included in A be m and the number of links in the network with positive flow be n . Then ATVvt(v*,0jA is an nxn matrix, but its rank is m if m 2, 2 —» 3 and 1 —»3, all having the same (fixed)
demand d^2 = dM=dM
= 5.
The General SecondBest Pricing Problem
Linkl
103
Link 2
Figure 4.6 A simple network of two links in series connecting three OD pairs Clearly, there are three paths denoted by rx ={l], r2 ={2], and r2 ={l,2}, one for each OD pair. The path costs in terms of path flows are given by: Since each OD pair is connected by a single path, all the three paths in the network must be included in the working path set. Let f = f = (fl,f2,f3)
and c = c = (c1>c2,c3) . In this
case, the number of paths included for the sensitivity analysis is more than the number of links in the network and indeed Vfc(f*,0) or ATVvt(v*,0)A (A = A and A = A)isnot invertible as we can see below: "l 0 Vfc(f*,0) = 0 1 i 1 1 2
r
and det[vfc(f*,0)] = 0
Nevertheless, from (4.23) where V(iD(n,0) = 0 we have 1 0 1 1 0 0
0 11 0 0] 1 1 01 0 1 2 0 01 0 0 0 0 0 10 0 0 0 0 1 0 0 0
and its inverse exists and is given as follows 0 0 0 0 0 0 1 0 01 0 0
0 0 0 0 0 
1 0 0^ 0 10 0 0 1 10 1 0 11 1 1 1 2
If we assume a toll charge is introduced on link 1 and let e = w,,then V u t(v*,o) = [l 0]T and Vu D (jx, 0) = [0 0 0] . All the derivatives of link flows in toll u, are intuitively zero because there is no route choice in this simple network with fixed demands. The derivatives at e = «, = 0 can be readily found to be
Mathematical and Economic Theory of Road Pricing
104
1 0 0 1 Substituting the above results of ~Jf 1
l]Tl 1 0 and V E c(f\o) into (4.29) yields
, Vev = AVef =
Vf =
The result tells us that there is no change in path and link flows in all three OD pairs, as expected. Next we assume that there is change h the fixed OD demand in OD pair 1 —> 3 and let e = &/1_>3. Then
in
a
similar
manner,
we
have
VEt(v*,0) = [0 0]T
and
VED(n,0) = [0 0 l] T . The derivatives at e = 8J13 = 0 are given by
11' Once again we obtain the expected derivative results. Example 4.7
We consider the Network in Figure 4.1 with link cost functions (4.44) and
OD demand d{_J>i = 2 used in Example 4.1. Without toll charge u\ = 0, we have the unique equilibrium link flow solution: v* = (1,1,1,1) . The derivatives of equilibrium link flows with respect to «, at u* = 0 can be easily found from (4.46) in Example 4.2: (4.53)
Now we consider application of the sensitivity analysis formula: Because all four paths are equilibrated paths, we have "10 1
0
10 0
1
1" 1
(4.54) y 0 1 1 0 1 ' 0 10 11 This matrix, with each row corresponding to a path, has a rank 3, so we can choose three independent paths. It can be easily checked that paths /;, r2 and r3 are linearly independent and are chosen here for further consideration. With this reduced set of paths we have
The General SecondBest Pricing Problem
105
Vfc(f*,O) = ATVvt(v*,O)A 1 1 0 1 0] 0 1 0 0 1 0 0 1 1 ojJ _0
0 0' M 10 0 0 0 10 1 0 0 1 0
0
1 0 0 1
0 1 1 0
=
2 1 1 1 2 0 1 0 2
This matrix is invertible and hence the formulas for the inverse of a partitioned matrix can be applied. Here we use the following inverse of the Jacobian directly for our calculation: 2 1 1 IT 1 [ 1 1/2 1 2 0 1 1/2 1/2 1 0 2  1 " 1/2 0 1 1 1 oj _ 0 1/2
1/2 0 1/2 1/2
0" 1/2 1/2 1
Thus we obtain the following link flow derivatives: V a v(u,). "1
V
l
'\u'=0.0
1 0 =AVEf = 1 0 E
1 0] • 1 1/2 1/2 0 1 0 1 1 1/2 1/2 0 1/2 0 1 0 1/2 0 1/2 1/2 0 1 0
1/2 1/2 0 0
(4.55)
which is identical with the early solution in (4.53). Example 4.8 We consider again the same network in Figure 4.1, but with the following link cost functions: f,=l + v, + M,, t2=\ + v2, ? 3 =l + v3 + «3, ? 4 =l + v4 where both link 1 and link 3 are subjected to toll charge. In addition to the OD pair 1 —» 3 considered earlier, we further add two OD pairs 1 —»2 and 2 —> 3 , and assume perturbations exist in all three OD demands given by d^ = 2 + bd^ , d^2 = 2 + 8d^2, d2^, = 2 + 5e?2_)3. In this case we have five perturbation parameters denoted by £ = («„ u2, MM,
&/lH2, oW2^3)T
There are eight paths in the network denoted by: r{ ={l,3},
r 2 ={l,4},
r 3 ={2,3},
r4 ={2,4}, r5 ={l), r6 ={2}, r, ={3}, rB ={4J. The number of paths is larger than the sum of the number of links and number of OD pairs. Without toll charge and demand perturbation: £ = ( « , , « 3 , & / M , &/ lHe , 5J2_3)T =(0,0,0,0,0) T
We have an obvious unique (unperturbed) equilibrium link flow solution: v* = (2,2,2,2) . Without difficulty, the derivatives of link flows with respect to e at e = 0 can be found analytically as:
Mathematical and Economic Theory of Road Pricing
106
Vev(e) e
=fv v(e)
^ ' e=0.0
_
"
\
/
, V5.v(e u = 0 .0
5d
"1/2 0 1/2 0 0 1/2 0 1/2
8d=o.oj
V
1/2 1/2 0" 1/2 1/2 0 1/2 0 1/2 1/2 0 1/2
(4.56)
Now we consider application of the sensitivity analysis formula. Again all four paths are equilibrated, and we have
A1
" 1 0 1 0 1 0 0" 10 0 1 1 0 0 0 1 1 0 10 0 0 10 1 1 0 0 = 10 0 0 0 10 0 10 0 0 10 0 0 10 0 0 1 0 0 0 10 0 1
(4.57)
This matrix, with each row corresponding to each equilibrated path, has a rank 5, so we can choose five independent paths out of the 8 equilibrated paths. It can be easily checked that paths ru r2,r,, r5, rn are linearly independent and are chosen here for further consideration. With this reduced set of working paths, we have Vfc(f*,0) = A T V v t(v\0)A 1 1 0 1 0
0 0 1 0 0
1 0 1 0 1
0 1 0 1 0 1 0 0 0 0 0 0 0
0 0] 0 0 1 0 0 lj
1 1 0 0 1 0 0 1
0 1 1 0 1 0 0 0
0" 0 1 0
2 1111' 12 0 10 10 2 0 1 1 1 0 10 10 10 1
The above Jacobian of the reduced path cost vector is singular (with a rank equal to 4). Nevertheless, the Jacobian J f is invertible and its inverse is given by: 2 1 1 1 1  1 0 0" 1 1 2 0 1 0  11 0 0 1 0 2 0 1  11 0 0 1 1 0 1 0 0 1 0 1 0 1 0 1 0 0 1 1 1 1 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 1 0 0 0 Then the link flow derivatives are given by:
['..I
1 1/2 1/2 0 0 1/2 0 0 0 0 1/2 0 0 0 0 0 0 0 0 0 0 0 0 0 1/2 1/2 0 0 1/2 0 1/2 1 0 1/2 1/2 0 0 1
1/2 1/2
0 1/2 1/2 1/2 1/2 0 0 1/2 1/2 0 0 1 1 0 0 1 1/2 1/2 0 1/2 1/2 1/2 1/2 0
The General SecondBest Pricing Problem
v v(e)'le=0.0== AV
107
v
"1
=
1
o
•1
1 o" 1/2 o o
r1/2
. 1/2 0 0
i
1
0 0 1 0 i o 1 1/2 0 0 1 o 0 0 0
1 0 1/2 1/2" o
c) () ()
0 1/2 0 0
0 0 0 0 0 0 0 0
1/2 1/2 0 0
0 1/2 1 0
1 0 0 0 0 0 0 0 _1
o o o
1 1/2 o o o 0 o 1 o o 0 0 0 1 0 1 0 0 0 1 0 0 0 0
o 0 0 0
1
0 1/2 1/2 0 1/2 1/2
1/2 0 1/2 1/2 0 1/2 1/2
0" 0 0 1/2 0 1/2
The results are identical with those identified analytically. vii)
Infeasibility with an extreme point
As mentioned before, the sensitivity analysis formula developed by Tobin and Friesz (1988) is based on the choice of a nondegenerate extreme point f* of the polyhedron Q(0) which should have exactly as many paths with a strictly positive flow as the rank of the matrix
 • This condition cannot always be fulfilled as already demonstrated by an
example in Josefsson and Patriksson (2003). To see this, we revisit Example 4.1 with the equilibrium path flows given by (4.45). 7T=P, 72* = I  P , 7/ = I  P , 7 ; = P ,
o0),
S"i( a< " > )
i
^
simple
sequence,
can be used.
Step 5. (Convergence check) If u ( n + 1 )  u ^ ' l ^ e f o r a l l a e ^ then stop where e is a predetermined tolerance. Otherwise, let n:=n+l
and return to Step 1.
The General SecondBest Pricing Problem
4.4
109
APPLICATIONS OF THE SENSITIVITY ANALYSIS METHOD
We now present a numerical example to illustrate how to use the sensitivity analysis method for obtaining the derivative information of interest. The road network, as shown in Figure 4.7, has 11 links and 7 nodes. The following link cost function, including a link toll, is the input, together with link freeflow travel time, t° , link physical capacity, Ca , and link environmental capacity (to be used later), Ca,
given in Table 4.1. (4.59)
10
Figure 4.7 The network used for sensitivity analysis and application Table 4.1 Link a
tl Ca Ca
1 6 200 160
2 5 200 150
3 6 200 200
Input data for the example network 4 7 200 150
5 6 100 100
6 1 100 100
7 5 150 150
8 10 150 100
9 11 200 160
10 11 200 160
11 15 200 150
It is assumed that there are 4 OD pairs (1—>7, 2—»7, 3—>7 and 6—»7) and the demand functions are given below: ATO^T)
iW^
7
= 600exp(0.04^1_7); D^{\i2^)
= 500exp(0.03n2^7)
) =500exp(0.05u^ 7 ); D^ 7 (i 6 ^ 7 ) = 400exp(0.05u^ 7 )
where ji is the generalized travel costs between OD pairs. Table 4.2 presents the derivatives of link flows v and OD demands d with respect to all
110
Mathematical and Economic Theory of Road Pricing
link tolls at ua =1.0 for all aeA
(note that link toll is measured in equivalent travel time
units). From Table 4.2 it is interesting to note that d^Xu) = *a, beA, aeA aua oub
(4.60)
This means that the marginal impacts of link tolls on link flows are symmetric. In fact this symmetric effect can be easily confirmed from the theoretical sensitivity analysis formula (4.32) derived earlier. The derivative information obtained is very useful in the design and operation of toll road systems. It produces the optimal 'direction' in which the flow pattern of the elasticdemand network equilibrium problem can move if the road tolls are changed, and thus the pricing strategies that should be sought for optimization of a system objective function. An important application of the sensitivity analysis results shown in Table 4.2 is the estimation of nearby solutions for any combination of link toll changes, once an equilibrium solution has been calculated. Table 4.3 presents the unperturbed solutions (link flows and OD demands) to the network equilibrium problem at link tolls ua = 1.0, aeA. Numerical comparisons of the estimated perturbed solutions using the derivative information with the actual perturbed solutions are also presented, when 3 of the
11 link tolls are changed by
8w9 = 8u10 =5M,, =0.10. The estimation is made using a linear approximation based on the derivative in Table 4.2. One more important application of the sensitivity analysis results is the calculation of the elasticity of OD demand or link flows with respect to the toll charge levied on a particular link. This kind of elasticity is quite useful in road pricing analysis, but cannot be simply obtained with the conventional simplified technique that is based on explicit, abstract demand models. Given the values of the derivative of link flows in Table 4.2, the elasticity, E""b of link flow, vb, be A with respect to link toll, ua, ae A can be calculated as E"v° =
v u )
^  ^g) b e A , a e A
(4.61)
where we have the direct toll elasticity of link flow when a = b and cross toll elasticity of link flow when a *• b. For example, from Table 4.2 and Table 4.3 we have Eu'=  2 6 . 6 l x  M _ = 0.14, E"> = 8.34x—!^ = 0.04, £"' =  1 3 . 1 6 x — ^  = 0.07 V 196.30 196.30 "' 196.30 It is clear that the cross elasticity (or derivative) can be either positive or negative, indicating two substitutable or two complementary toll roads. Formally, toll road a and b are substitutes if dvb(u)/dua > 0 (or equivalently E"' > 0) and complements if dvb(u)/dua < 0
The General SecondBest Pricing Problem
111
(or equivalent^/ E°°t < 0 ) . The cross toll elasticity of the link flow given by (4.61) can be used to measure the degree of flow interdependence among toll roads, which depends on, among many factors, the network topology and OD demand pattern. Clearly, If dvb(\i)/dua = 0 , a*b, then the two toll roads are said to be independent.
Likewise, given the values of the derivative of OD demands in Table 4.2, the elasticity, £ £ of demand between a specific OD pair weW
with respect to toll ua levied on a particular
link a e A can be calculated as E"d = ^ i H ) i ^ ,
WE
W, ae A
(4.62)
Once again, as examples, we have E"> =4.43x M
1 0
246.86
=0.018 and E? =  3 . 8 6 x L ° =0.016 d " 246.86
The toll elasticity or derivatives of OD demands can be either positive or negative.
Table 4.2 Derivatives of link flows, OD demands and objective functions with respect to link tolls at u,, =1.0 for all links Variable
a(.)/a»,
d(.)/du2 8.34
3()/3u3
26.61 8.34
38.26
13.16 4.12
13.16 4.41
4.12 17.54
24.89 2.18
1.50
6.80
0.74
v6
3.92 13.45
20.71
Vg
0.72
v, v10 vn
3.69 9.02
3(0/3«4 4.41 17.54
d()/du5
d()/du6
1.50 6.80
3.92 20.71
0.74
1.94
21.51
0.89
3.97
0.89
12.44
5.91
3.97
5.91
4.21
1.94 11.73
2.23
0.76
24.68 1.98
7.45 10.53
0.36 1.82
10.17 3.38
2.83
7.87
0.25 3.35 1.50
0.51
13.88 1.33
14.47
19.94
7.16
3.58
9.39
d^n
2.31
3.18
1.14
10.56 1.68
0.57
1.50
di,1
0.78
0.65
0.64
2.27
1.29
4.43 0.96
1.39
0.38 3.86
0.25
0.65
2.73
0.47
0.73 0.87
0.88
3.60
v, V
2
v4 v5
d}_>1 d^7
2.18
7.20
Mathematical and Economic Theory of Road Pricing
112
Table 4.2 Continued Variables v
i
V
2
d{)/du,
d(.)/du9
d{)/dult
d()/duu
3.69
9.02
14.47
13.45
3()/3«. 0.72
4.21
7.45
10.53
2.83
19.94
11.73
0.36
1.82
7.87
7.16
v4
2.23
0.25
3.35
1.50
10.56
v5 v6
0.76
10.17
3.38
0.51
3.58
1.98
7.20
13.88
1.33
9.39
v7
25.18
0.36
1.86
16.89
7.31
Vg
0.36
14.02
5.42
0.24
1.72
v9
1.86
5.42
23.56
1.25
8.82
V
18.69
0.24
1.25
19.61
4.91
10
V]1
dx^ d
^ i
d^i
7.31
1.72
8.82
4.91
43.50
1.17
0.27
1.41
0.78
5.51
0.39
3.85
2.03
0.26
1.85
8.29
0.12
0.61
2.72
2.41
0.48
1.41
4.26
0.32
2.29
Table 4.3 Estimated and exact solutions of link flows and OD demands for perturbed link tolls by 5M9 = 8M10 = 8w,, =0.1 with respect to link tolls at ua = 1.0 for all links Solution Variable
Unperturbed Solution
v, v2 v3
196.30 95.67 230.34
Exact Solution 197.32 96.97 231.83
Estimated Solution 197.22 96.90 232.03
v4 v5
245.55 112.34
247.18 112.30
247.09 112.27
v6
149.88
150.21
150.20
v7
34.04
34.50
34.81
v8
124.87
124.48
124.52
v9
182.37
180.94
181.02
v10
212.81
211.63
211.47
v,, dx^
131.54 311.17
128.31 310.30
128.56 310.40
d2^
237.21
236.78
236.79
d^
246.86
246.13
246.28
206.67
206.69
C and v0 > C). In reality, as traffic flow reaches network physical capacity (v = C), vehicle queues will occur and drivers experience corresponding queuing delay. The actual supply performance curve will become abe'Q where the vertical segment represents the queuing delay. The actual queuing delay will grow sufficiently to match the corresponding demand at network capacity, and thus the equilibrium point e'o is reached, and the corresponding queuing delay is be'o. Because queuing delay constitutes wasted time, it may be desirable to remove the queues by toll charges from the viewpoint of society. As stated earlier, this could be done by simply substituting the queuing delay be'o with an equivalent amount of toll. However, in reality, it might be required to bring down the network demand to a level lower than its physical capacity. A typical situation is that the traffic volume through each street of the network or total amount of vehiclekilometers over the network should not exceed a certain threshold for environmental damage, as discussed briefly in Chapter 3. From Figure 4.8, we can see that under road pricing scheme P,, the equilibrium demand, v1, corresponding to point e, is less than the network physical capacity, but greater than its environmental capacity (C < v, < C). Under road pricing scheme P2, the equilibrium demand, v2, corresponding to point e2 satisfies both physical and environmental capacity constraints. Therefore the problem of interest here is to find a road pricing scheme that holds the traffic flow within a given level such as below the network environment or physical capacity. 4.5.2
Direct Solution with the Queuing Network Equilibrium Model
Consider an elastic demand UE problem with link capacity constraints introduced in Chapter 2 as follows: v,
dw
rnin £ j'.(«>)da> X J*»(
(463)
subject to va 0; crw >\iw if f'w = 0, re Rw,we W
K«)
= K>
if
0; KK)
^ ^,
if
115 (4.65)
«C =0,weW
(4.66)
X o =0, if v[0, if v > C a , aeA
(4.67)
where u^ is the minimum travel cost between OD pair weW,
Xa,ae A is the Lagrange
multiplier associated with capacity constraint (4.64) and
^ = S^(v«')8«r' l{vl) = ta{vl) + K
( 4  68 )
aeA
and Xa is the Lagrange multiplier associated with constraint (4.64) for link as A. The constraint set of problem (4.63)(4.64) is linear and hence convex, and the objective function is strictly convex with respect to link flow and OD demand variables, and a feasible solution is guaranteed in which the equilibrium OD demands and link flows are determined uniquely. Multiplying both sides of the equality cm = u.w for f^, >0 and inequality cm>\x.w for f'w
=
0 by f'm for each path reR^
lUX
and for each OD pair we W in (4.65), we have
+ E V X ^ X , rsR^weW
(4.69)
Clearly the equation holds for both frw > 0 and /*„, = 0. Taking the summation of both sides of (4.69) over all paths and all OD pair and using the flow conservation relations v
«=S^S
r e
^O0r.
a&A
>
and
4* = £ re/ t,/™ '
I> o vl = 5> a C 0 = X M»  £'» { 0 ()(u,v,d,v,d) = G a (u,v,d), where v = Af and d = Af]
(5.14)
We then have the following Lemma immediately. Lemma 5.1
Forgiven ueU and (v,d)s Q, let f'eQj(u,v,d) be an optimal path
flow solution associated with (u,v,d) to the problem (5.12)(5.13), then f* satisfies the Linear Independence Constraint Qualification (LICQ) for the problem (5.12)(5.13). Lemma 5.1 is selfevident For any active nonnegative path flow constraints in (5.13), the corresponding gradient is an unit vector with the element corresponding to the binding zero path flow equal to 1 and the rest are zeros. Because all path flows (within each OD pair or across different OD pairs) are independent variables for the elastic demand traffic equilibrium problems (5.12)(5.13), all the gradients associated with the active nonnegative path flow constraints are thus linearly independent. The LICQ property for any path flow solution of the problem (5.12)(5.13) allows for the direct application of the sensitivty analysis results for parameterized nonlinear programming problems developed in Gauvin and Dubeau (1982). From Lemma 5.1 and the differentiability results of marginal function in nonlinear programming, the following theorem on the gap function G a (u,v,d) follows immediately. Theorem5.2
Ga(u,v,d)>0 is continuously differentiate in ueU and (v,d)e£2
and its gradient is given by
The General SecondBest Pricing Problem
Vut(v,u)
0
127
¥v*v"l
V' ) \ Vvt(v,u) 0
) 0 V d B(d,u) ' "" " '
"'
(5 16)

where v*=v*(u,v,d) and d*=d*(u,v,d) is the unique solution of the positive definite quadratic programming problem (5.8), or equivalently, the unique equilibrium link flow and OD demand solution of the traffic assignment problem (5.9) for given (u, v, d). In their detailed components, the expressions of the above derivatives are in fact simple, as given below (consider link based toll charge only and unit value of time or toll charge measured in equivalent time unit): aG a (u,v,d)_,,,. (
dua
v )
\"
''[
3G a (u,v,d)_ v ^. 9G a (u,v,d) = _
(
dua
0,
^dtj±n)(t ^^.(d.u) > dd
w
/V)U) + ^ a f (du ( "K
where v* = v*(u,v,d), n e A and d"w = J^(u,v,d), we. W are again the unique link flow and OD demand solution of the traffic equilibrium problem (5.9) for given (u,v, d), 9Svv(d,u)/3«lv can be calculated once the role of u in the demand function is specified (e.g., potential demand or cost sensitivity parameter in the demand function). 5.2.2
Gap Function for the Optimization based Traffic Equilibrium Models
We first define the marginal or value function for the lowerlevel convex programming problem of the traffic equilibrium model (5.4)(5.5) as follows
(M1,V1,V2,VPV2) where + (v 2 v 2 )(v 2 + 4) + (v1 Substituting v2 = 5  v, and v2 = 5  v, into the above expression yields
(5.38)
132
Mathematical and Economic Theory of Road Pricing x/ _x r_ 1 1 7V (3 1 vi>vi) = vi + ^ v i +^"i  J  ^ v i +  " i  J
One can thus readily obtain that , , , — (3v, +M.7) 2 if «, +v.7
Let Ga(ul,vl) = O in (5.39), we obtain the same equilibrium link flows, v*(«,) (and v* («,) as well) as a function of w, as given in (5.37). Next we consider the following penalized problem for finding a solution of minimizing total travel time cost (5.40) where p > 0 is a positive penalty parameter. By eqn. (5.39), we can separate the above problem into two cases. Case (a) When «, +v, 0 is a positive penalty parameter.
From the definition of the gap function
(5.26), H(ux, Vj) is given by H («,,v,) = }(2? +2+Ml) d? + j (4? +8) d?  q> (u,) 0
0
where the value function
(P(M,)
is given by (5.43). In view of the fact that the current
objective function has different explicit expressions in different intervals of variable w,, we can decompose the revenue maximization problem (5.45) into two subproblems defined within the intervals of 0 < u, < 18 and u, > 18 respectively. Case (a) When 0 < u, < 18, the partially penalized objective function is quadratic and the problem can be written as:
r
1
where 4 = [ l 8 p , 3p], 5, =
f6?
_ ?'
1? 1
' \, C,=27?. In this case, S, is negative
definite for any ? > 1/2, which means the problem is a strictly concave maximization and the stationary point of the partially penalized problem is
which is feasible for the problem. Hence, the above (v,(?),u,(?)) is the optimal solution and we have
136
Mathematical and Economic Theory of Road Pricing Um v,(?) = 1.5,
Hm ux{l) = 9, lim ^(wpV,) = 13.5
which is exactly the optimal solution identified earlier. Note that for any limited value of p > 1/2, the solution of the penalized problem will not be exactly equal to the solution of the example. Case (b) When M, > 18, the problem can be written as
where y42 = [18p, 0] B2 =
f—6D
1— Q~\
'
, C2 = 0. The optimal solution of the problem here
is not obvious for a limited positive parameter p . Nevertheless, it is straightforward to check the optimal solution as p> +°°. The problem can be rewritten below max 0 . ) 5  > reRw,weW
P
(5.57)
P ^ ±J
where P is the user's value of time. Thus, the path cost function (5.57) is a special case of the generalized nonadditive path cost function introduced by Section 2.3 of Chapter 2, where a nonlinear (instead of constant) value of time is considered. TOLLING SUBNETWORK • • • I n
BASIC SUBNETWORK
Figure 5.3 A schematic representation of the network system (n'b and n\ represent node i in the basic and tolling subnetworks, respectively) Since a toll scheme influences the demand for travel, it is desirable to consider the elasticity of travel demand. We thus introduce the demand function {
dK =Dlv(u,lv)
and
Bw=D~ (dw)
(suppose separable OD demand function), which is a strictly decreasing
function
the
of
minimum
generalized
OD
travel
cost
(in
time
unit)
\iw
(\xw = minre/, c^, we W\. Furthermore, we assume that users have a free choice of whether to use the roads with or without toll charge and the route choice behavior of users satisfies the user equilibrium principle. In this case, a pathflow vector f * with element f*w, is said to be a user equilibrium if and only if /^>O,thenc ra (f*) = niv(f*) for all r e i^, we W
(5.58)
/^,=O,thenc TO (f*)>n w (f*) f o r a l l r e ^ , we W
(5.59)
142
Mathematical and Economic Theory of Road Pricing
Together with the link flow vector v and OD demand vector d, we also introduce vector q = (•••, ue)
(5.91)
dd.. (u,xWu"",x3 10
2 3 l»10 3»l 10
9
4 5 6 3>8 8^3 8>10 11
10
10
7 8 10>l 10>8 12
8
As discussed previously, a regularization approach is utilized to determine a unique solution of the entryexit flows. We choose an initial value of the regularization parameter r(1) = 0.04 and reduce it by half after each iteration, ri("+1) = 0.5r)(n>. In the augmented Lagrangian
The General SecondBest Pricing Problem
155
algorithm used in the example, we set the initial Lagrangian multiplier (])(1) = 0.0 and the initial penalty parameter p < ! ) =10.0, and employ the following updating procedure: If #(u ( " + 1 ) ,x ( n + 1 ) )
; Otherwise, set p("+1) =10.0p12
15.42
15.61
12.42
13.37
"2
11—>13
37.82
37.78
35.74
37.30
"3
11>14
51.21
51.91
50.16
49.67
«4
12>11
14.52
14.26
13.03
13.37
«5
12>13
25.13
24.31
23.32
22.92
«6
12>14
38.71
39.39
37.74
37.30
"7
13>11
37.56
37.33
37.44
37.30
«8
13—>12
23.59
23.31
23.41
22.92
u9
13>14
15.71
15.82
14.42
13.37
14>11
51.47
51.82
50.75
49.67
14>12
38.54
37.72
37.72
37.30
14>13
15.08
15.61
14.31
13.37
"11 M
12
E 48.66 48.66 48.66 48.66 48.66 48.66 48.66 48.66 48.66 48.66 48.66 48.66
Table 5.10 The optimal tolls (HK$) for social welfare maximization Pricing Scheme
A
B
C
12.13
5.62
D 8.72 23.68 32.40
E 28.12 28.12 28.12 28.12 28.12 28.12 28.12
"1
11>12
12.01
u2
11—>13
24.82
24.75
22.01
"3
11—>14
31.89
31.93
32.03
«4
12>11
9.38
9.41
5.89
"5
12>13
14.51
14.25
17.39
"6
12>14
23.88
24.17
27.41
8.72 14.95 23.68
«7
13>11
22.02
21.92
21.48
23.68
«8
13>12
15.97
17.14
15.59
«9
13—>14
10.31
10.27
10.02
14.95 8.72
«10
14^11
30.91
30.85
30.67
32.40
«11
14^12
28.41
28.76
24.78
"12
14—513
13.01
13.11
9.19
23.68 8.72
28.12
28.12 28.12 28.12 28.12
Table 5.10 and Figure 5.13 present the solutions for the case of social welfare maximization. Similar phenomena are observed, with declining maximal social welfare, from Scheme A to Scheme E. Again, Scheme A and Scheme B gives rise to nearly the same solutions. Moreover, Scheme E for uniform toll charge gives rise to the lowest social welfare and the lowest revenue. Scheme D for the distance based toll pricing scheme gives the highest revenue among the five cases of social welfare maximization.
The General SecondBest Pricing Problem
1.10 •
•
Social Welfare 107HK$ •
159
Revunue 106HK$
1.051.000.95 •
0.90 0.85 0.80 0.75 A
B
C
D
E
Pricing Scheme
Figure 5.13 Maximum social welfares and the corresponding revenues
5.5
SUMMARY
In this chapter we developed a gap function based approach for the second best road pricing problems which are formulated as MPEC or BLPP. We proved that the gap function associated with the VIbased or optimization based lowerlevel UE problem is a continuously differentiable function and its value and gradient in flow and toll variables can be readily calculated via implementing an equilibrium traffic assignment only. The MPEC and BLPP are then transformed into a singlelevel continuously differentiable optimization program by treating the UE conditions as a nonlinear, implicit gap function constraint. By incorporating the gap function constraint into the objective function as a penalty term, the singlelevel differentiable problem was solved by the augmented Lagrangian algorithm, in which the critical suboptimization problem is favorably solved by the conventional convex combination method. The gap function based approach was well illustrated with both simple analytical and larger numerical pricing examples with known solutions. We have specifically addressed the application of the proposed gap function based approach for solving the optimal secondbest pricing problem in a network with entryexit based toll charges. Although the solution of equilibrium entryexit flows using the toll roads is not uniquely determined, the resulting total toll revenue is found to be always unique. The FrankWolfe method is adapted to solve the traffic equilibrium model, where the de scentdirectionfinding subproblem is solved by a conventional shortest path algorithm through an appropriate network transformation, and the network transformation avoids the difficulty of
160
Mathematical and Economic Theory of Road Pricing
path information storage and enumeration. A regularization approach was used to address the nonuniqueness issue of the entryexit flow on the tolling subnetwork, thereby allowing for the application of the gap function approach to solve the optimal entry exit based toll design problem. The performance of the solution method is demonstrated and the solution properties of alternative toll pricing schemes are explored with a numerical example. The proposed gap function method proved to be indeed promising and can be further extended to address the following issues: 1) It is meaningful to investigate the exact and inexact nature of the employed augmented Lagrangian penalty function to recover the local optimum for finite penalty parameter values; 2) The step size search procedure is computationally demanding in solving the singlelevel continuously differentiable optimization problem, and alternative algorithms, such as a trust region algorithm, can be explored to circumvent this difficulty; 3) A rigorous attempt for computing globally optimal solutions is presently out of the reach of existing methods, nevertheless, it is possible to further transform the singlelevel continuously differentiable program into a concave minimization problem subject to convex constraints, and then apply the wellestablished cutting plane algorithm to find the global optimum.
5.6
SOURCES AND NOTES
Value (or marginal) function is an important concept in nonlinear programming, and its differentiability property is well expounded in Gauvin and Dubeau (1982). In transportation literature, the value function for the convex UE program and its differentiability property was explored and applied for the traffic network design problem in Meng et al. (2001). Several gap functions are available for Vis, but not all of them are differentiable. The quadratic gap function used in this chapter for VI was proposed by Fukushima (1992) and has a desirable differentiable (in argument of VI itself) property. Further differentiability in upperlevel decision variables in the context of MPEC for the secondbest pricing problem was shown here. Readers may refer to Ye et al. (1997), Chen and Florian (1995) and Marcotte and Zhu (1996) for use of the gap function in the analysis of the general BLPP. Detailed description and discussion of the augmented Lagrangian algorithm can be found in the excellent textbook by Bazaraa et al. (1993) and Bertsekas (1982; 1999). Readers are recommended to consult the references: Meng et al. (2001) and Yang Zhang and Meng (2004), from which part of the materials presented in this chapter are drawn.
6 DISCRIMINATORY AND ANONYMOUS ROAD PRICING
6.1
INTRODUCTION
In this chapter, we move on to investigate the road pricing problem with extended multiclass equilibrium behaviors in the transportation networks introduced in Chapter 2. We begin our investigation by first recognizing that the terms of multiclass traffic equilibrium and pricing problems may refer to three distinct situations of observable and unobservable user heterogeneity, as summarized in Table 6.1. The first situation is that the flows in a transportation network are divided into different classes of vehicles or modes, each of which has an individual costflow function, and at the same time contributes to its own and other class's cost functions in an individual way. As examined in Section 2.3 in Chapter 3, a classification of vehicle types could distinguish trucks and buses from cars; heavy vehicles from light ones; private transport from public transit. In this situation, applying the marginal cost pricing principle would burden each class with a different toll for social optimum, as shown in Section 3.5 in Chapter 3. This is economically meaningful and practically feasible. It is economically meaningful because the firstbest discriminatory toll charges reflect the different congestion externalities brought about by different types of vehicles. It is practically feasible because toll charges can be differentiated across user classes on the basis of their observable differences. Differentiated toll pricing across vehicle types is indeed carried out in most pricing schemes. The second and third situations are that all users or drivers are assumed to use the same type of vehicles and thus have identical congestion effects when using the same roadway, but users differ from one another in other, unobservable ways. In the second case, users differ from one another in the value they place on time; in the third case users differ in their routing behavior. In the latter two situations, users are observationally indistinguishable and then the congestion charge must be anonymous or uniform when they use the same road link.
162
Mathematical and Economic Theory of Road Pricing
In this chapter we deal with the multiclass, multicriteria and multiple behavior equilibrium and system optimum problem. Here we refer the multiclass and multicriteria to the situation that users have distinct values of time (VOT) and thus place different subjective weightings of two criteria, travel time versus travel cost, in route choice; we refer multiple behaviors to the situation that users make their route choices in different behavioral manners, such as the case when deterministic User equilibrium (UE) users and CournotNash (CN) players share the same network resources. We establish the existence of meaningful anonymous link tolls in this case to decentralize a system optimum as a multiclass, multicriteria and multiple behavior equilibrium, with resort to rigorous mathematical programming approach. Table 6.1 Classification of multiclass traffic equilibrium problems Singleclass Traffic Equilibrium a) Traffic equilibrium without link flow interactions (standard traffic equilibrium formulated as a convex mathematical program) b)Traffic equilibrium with asymmetric link flow interactions (generalized traffic equilibrium formulated as a VI)
Multiclass Traffic Equilibrium By vehicle types a) Each type of vehicle has an individual costflow function, and meanwhile contributes to its own and other class's cost functions in an individual way (e.g., truck vs car).
By value of time a) Discrete value of time distribution (finite number of user classes)
b)The resulting multiclass equilibrium problem can be reduced to a single class problem with asymmetric link flow interactions through multiple network copies (each class is associated with a network copy).
b) Continuous value of time distribution (infinite number of user classes)
By routing behavior a)UEvjSUE b) UE vs CN controller or player (CN altruistic controller or CN private controller) c) UE vs SO player (SO player acts a leader in a Stackelberg game or SO player is capable of predicting the reaction of UE users)
Remark: UE: User Equilibrium; SUE: Stochastic User Equilibrium; CN: CournotNash; SO: System Optimum
This chapter is organized as follows. After revisiting the single class traffic equilibrium and system optimum problem, we deal with the two criteria of cost versus time traffic equilibrium and system optimum problem, where system performance can be measured in either time or cost units. We then continue to examine the multiple behavior equilibrium and system optimum problem, followed by the combined multiclass, multicriteria and multiple behavior equilibrium and system optimum problem.
Discriminatory and Anonymous Road Pricing 6.2
163
TOLL PROPERTIES FOR THE SINGLE CLASS UE AND SO PROBLEMS
6.2.1
Characterization of Link Tolls for SO
As shown in Chapter 3, in the standard traffic network equilibrium model with homogeneous users, a socalled marginalcost pricing toll can drive a user equilibrium flow pattern to a system optimum in the sense of total travel time minimization for the fixed demand case and social welfare maximization for the elastic demand case. The link toll according to the marginalcost pricing principle equals the congestion externality given as the difference between marginal social cost and marginal private cost on each link. As shown in Chapter 3, the SO problem with elastic demand can be formulated as:
where
\
(
)
\
J
J
The same notation is used as before. Let d~l°', weW
^
\
and vaso, ae A be the unique OD
demand and link flow solution of the above SO problem (6.1)(6.2). It is already shown in Chapter 3 that the marginalcost pricing toll given by ua =Pvo •dta(va)/dva\
_M , as A,
where P is the VOT of homogeneous users considered here, will support the SO flow pattern (v s o ,d s o ), where v so = ( v f , ae A)1 and d so = ( 5 f , we W ) \ as an UE flow pattern that satisfies the following variational inequality (VI)
(y.d)eQ
(6.3)
It is known that the link tolls that can decentralize a given SO flow pattern as a UE is not unique; this observation is particularly prevalent in the traffic equilibrium problem with fixed demand. Our task here is how to identify all such valid link tolls including the economically meaningful marginalcost pricing link toll. It is interesting to note that the desirable valid link tolls for a given SO flow pattern (v s o ,d s o ) obtained from (6.1)(6.2) is a polyhedron as shown in the following theorem.
(6.2)
164
Mathematical and Economic Theory of Road Pricing
Theorem 6.1
Any link tolls u = (ua,aeA)
contained in the following polyhedron
consisting of equalities and inequalities can support a SO flow pattern (v s o ,d s o ) obtainedfrom problem (6.1)(6.2) as an user equilibrium: ^ ,
re^weW
(6.4)
avaso =fiY,Bw(d?)d? we.W
aeA
P2>a(voso)voso = constant aeA
In the case with fixed OD demand dw, weW, the SO problem is to minimize total travel time, max ^
v eA
Ja(va)
subject to flow conservation conditions. For a given unique SO
link flow solution v s o , the valid link tolls to decentralize the SO as an UE is given by the following linear system: \ ^ .
re^wsW
(6.15)
166
Mathematical and Economic Theory of Road Pricing
Clearly in the fixed demand case, the total toll revenue is not constant. However, if u.w is given (for example, as the unique minimum marginal route travel time at the SO solution) MVS0=BY
then Y
U,
d BY
t (vs°)vso is unique.

In view of the nonuniqueness of the valid link tolls for decentralizing the SO flow pattern, one may introduce various secondary objectives such as toll revenue (for the fixed demandcase) or number of toll booths are optimized with respect to the feasible link tolls contained in the polyhedron developed above. 6.2.2
A Simple Example for the Valid Link Toll Set to Decentralize SO
Example 6.1 (Valid link toll set for the singleclass UE and SO problem) Now we provide a simple example to explain the toll characterization results for the single class UE and SO problem discussed above. The network, shown in Figure 6.1, consists of 3 nodes and 4 directed links with one OD pair (origin node 1 and destination node 3). Suppose the VOT for all users is B = 1.0 and the link cost functions are given by
Figure 6.1 An example to explain the toll pricing for the single class UE and SO problem 1) The elastic demand case. Suppose the demand function between origin 1 and destination 3 is given by d  13  (X (omit the OD index for simplicity). The system optimal solution for the social welfare maximization is unique and is given by
vso = (v.vf.vf.vf) T = [ y ,  A f J and d™ = i± The corresponding link travel times are
The set of valid tolls to decentralize the SO flow pattern is given below 29 — + 3 + M. +u, > u 8
29
1
7
3
"
Discriminatory and Anonymous Road Pricing
167
33 33
7
+ — +M, +U, >\ 2 4 8 2
41 29
("33
"\ 13
^ 9
,,
N
„
("7
— + w, x — + — + M, X  + ( 3 + M , ) X 2 +
8 'J 8 [ 8 which can be simplified into
2
J8
V
3/
" 3
41
— +M. x—= — x
[2
4
11 —
J4 4 4
u, + u > — 1 3 8
25
13 9 » 3 289 3 8 ' 8 2 4 4 32 The above system is simply equivalent to the linear equality systems: 1 1 21 H , = H 2 +  , U 3 = M 4 +  , u2+uA= — L
Z
o
It can be easily checked that the marginalcost pricing toll: u = (13/8,9/8,2,3/2) , is included in the above system of equalities and inequalities. Clearly, the valid toll set is a convex polyhedron. The total toll revenue is 289/32, and constant. 2) The fixed demand case. Suppose the demand between origin 1 and destination 3 is fixed to be d = 5. The system optimal solution for total travel time minimization is unique and is given by V
vl
r=
11 _ s0
9
4
4
> "2
_so
' "3
7 2
_so
' "4
3 2
and 10 91 0 rso _ £ ^ rso _ £j_ 7so _ ^_ rso _ c h ~ 4 ' ( 2 ~ 4 ' ^ ~ 2 4
The set of valid tolls decentralizing the system optimum is given below 19 9 4 2 ' 3
168
Mathematical and Economic Theory of Road Pricing 19 4 21 9 4 +2 21 4 19 "\ 11 (21 ^ 9 (9 ) 7 ,t . x 3 — + u, x—+ Yu, x—+ —+M, x—+(5 + u, )x—= 5 x u 4
'
4
4
2
4
2
3
2
V
"
;
2
^
The valid link toll set satisfying the above linear system can be further simplified into 1 1 u}=ui + W,=u2+, The above valid toll set is a unbounded convex polyhedron which includes the marginalcost pricing toll: u = (11/4,9/4,7/2,3) , the total toll revenue with the above valid link toll set is unbounded. Notwithstanding, if we set u, to be the unique minimal marginal OD travel cost cost of 31/2, then the valid link toll set becomes:
_25
_21
J_
_
J_
which gives rise to a constant revenue of 238/3.
6.3
TIMEBASED AND COSTBASED MULTICLASS UE AND SO PROBLEMS
In Subsection 2.3.4 of Chapter 2, we introduced the multiclass bicriteria UE models in which users select their routes to minimize total disutility measured either in time unit or cost unit, respectively, and concluded that both timebased and costbased equilibria have identical link flows pattern that can be obtained from a convex minimization program. In this section, we examine the multiclass, bicriteria or cost versus time network equilibrium and system optimum problem for several user classes in a network with a discrete set of VOTs, but with identical types of vehicles. We already know from Chapter 2 that the useroptimal link flow pattern is independent of the unit (time or money) used in measuring the travel disutility in the presence of road pricing when users value their travel times linearly. Here we are specifically interested in the following questions: are there any anonymous or uniform link tolls (link tolls that are identical for all user classes) that can drive a multiclass traffic network equilibrium into a system optimum? If yes, does such link toll pattern depend on the criterion or the unit (time or money) that defines the total travel disutility for the system optimum? What are the general properties of the valid toll set?
Discriminatory and Anonymous Road Pricing
169
Recalling the following minimization problem with separable link cost functions and fixed classspecific OD demand introduced in Subsection 2.3.4
^yX
(6.17)
aeA mf=M P m
subject to ^fZ
= d",weW, meM
(6.18)
fZ>O,reRw,weW,meM
(6.19)
, me M
(6.20)
This minimization problem leads to a multiclass network equilibrium in terms of either generalized path travel time given by
vJ + ^  i s , , re i C we fF, me M Pm J
(6.22)
or generalized path travel cost given by (va) + K.}5ar, r e Rw, weW, me M
(6.23)
In problem (6.17)(6.21) and eqns. (6.22)(6.23), M is the set of all user classes, (3m (P m >0) is the average VOT for users of class m, /r™ is the flow of user class m on route reRw
between OD pair weW, vj is the flow of user class /M on link aeA,
d™ is the demand for users of class m between OD pair weW, and ua is the exogenously given toll levied on link a e A. As assumed previously, all link travel time functions, ta(va), as A, axe continuous, strictly increasing and convex. For users with a single VOT, problem (6.17)(6.21) collapses to a standard UE traffic assignment problem with fixed demand. It is well known that the marginalcost pricing link toll, ua = fivj'a(ya), evaluated at the system optimum link flow, ~va, ae A, can support a SO flow pattern as a UE. The SO here refers to the minimization of total travel disutility that can be measured in terms of either time or monetary units. 6.3.1
System Optimum in Time Units and Pricing for Equilibrium
We next consider the following system optimum in time units and see how it leads to a different result:
170
Mathematical and Economic Theory of Road Pricing min ^vja(va)
(6.24)
aeA
subject to (6.18)(6.21). Problem (6.24) is a strictly convex program in aggregate link flow v if the separable link travel time function is monotonically increasing and convex (see Chapter 2), and thus, the solution of aggregative link flow is unique. The link flow by user class is indeterminate, and any
v™, a e A, m e M
that leads to the unique aggregate link flow solution,
2ma e 4 i s optimal to the timebased SO program (6.24). From the firstorder necessary optimality conditions, for an optimum ~va, as A to the timebased SO program (6.24) we have r =jC'
ime
if fZ>0,re
Rw,weW, me M
if fZ=0,reRw,weW,meM
(6.25) (6.26)
Similarly, the optimality conditions (6.25)(6.26) for the timebased SO program (6.24) can be regarded as the network equilibrium conditions in time units, with ^™'"me being the corresponding minimum OD travel time. As seen from (6.25)(6.26), u™"'1™ is identical for all user classes me M traveling between the same OD pair weW. The requirement for such equilibrium is that each user should face a marginal social travel time consisting of a private or experienced travel time and a travel time congestion externality common to all users, when using each link in the network. As we can see, the above link travel time externality is anonymous, but can be internalized only when the toll charge for each link is differentiated according to each user's VOT: ,aeA,meM
(6.27)
This discriminatory link toll pattern is of course unrealistic and difficult, if not impossible, to implement in reality because, as mentioned earlier, users differ from one another in an observationally indistinguishable way. Given the impossibility of introducing a discriminatory link toll pattern (6.27) for various user classes for achieving an exact timebased SO, we naturally wonder whether there exists an alternative uniform link toll pattern to decentralize a timebased SO flow pattern into a multiclass network equilibrium. In contrast to the single class case examined in Section 6.2,
Discriminatory and Anonymous Road Pricing
171
where the set of systemoptimal link tolls can be characterized by a nonempty polyhedron, the answer to the existence of anonymous systemoptimal link tolls to the present multiclass user equilibrium problem is not evident and straightforward. Let v"as°, a e A be the link flow solution of the timebased SO program (6.24). Under our aforementioned assumption of increasing convex link travel time functions, the timebased SO link flows v^0 and hence link travel time ^(v a s o ), as A are uniquely determined. We now begin to seek an anonymous link toll pattern to decentralize this unique timebased SO flow pattern as a multiclass bicriteria UE. Consider the following linear programming (LP) problem:
fcsok
(628)
subject to
IXX/x=ffl^
( 6  29 >
ixW neM e ( ,
£ f™ = dl, w e W, m G M
(6.30)
fZ>0,reRtt,wsW,msM
(6.31)
where v ^ X ^ X ^
fZ&ar>
a&
^, meM. The dual formulation of the above LP is
given by
£ £ C ^
aeA
(632)
KEW neM
subject to K + 2 ^ A r ^ 2 $Ja(vaS0)Sar,
r e Rw, weW,m^:,
reRw,we W, m€ M
(6.34)
172
Mathematical and Economic Theory of Road Pricing
Let / ^ and v"™ be an optimal solution to the LP program (6.28)(6.31). According to the duality theory, at the optimal points of the original and dual problems the following complementary slackness conditions hold: Z=0,reRK,weW,meM
(6.35)
( ue.l
The above conditions can be regarded as the multiclass user equilibrium conditions in cost units, with \i™ being the corresponding minimum OD travel cost. The dual variable, Xa, can be interpreted as a shadow price, i.e., an additional cost implied by the need for consistency between the link flows generated by the timebased SO program (6.24) and used in the LP (6.28)(6.31). If ua is positive (or Xa is negative), users traversing link ae A are charged by a toll equal to ua, and otherwise subsidized with ua. Therefore, we conclude that if the link travel cost perceived by an individual user is modified by charging or subsidizing an amount of money, ua, ae A, given by the solution of the dual problem (6.32)(6.33), then a new network equilibrium is established. This new equilibrium is precisely a timebased SO in the sense that the same system optimal link flows, v"aso, ae A, (and hence the same link travel times, ta (vf 0 ), ae A) are reproduced. In summary, we have the following finding: Theorem 6.2 A timebased SO for minimizing the system travel disutility measured in time unit can be supported as a multiclass network equilibrium by a uniform link toll pattern across user classes. The toll is given as the solution of the dual LP (6.32)(6.33) and can be positive (charge) or negative (subsidy) to link users. Note that the Lagrange multipliers ua (or A.fl) and u™ satisfying (6.35)(6.36) may not be unique. The valid feasible set of uo, a e A and the corresponding OD cost u™, weW, me M can be alternatively written as the following polyhedron consisting of a system of linear equalities and inequalities:
x«
(637) e W, me M
(6.38)
where v™, a&A, meM solves the primal LP (6.28)(6.31). Equation (6.37) follows directly from the duality theory that the dual LP has the same optimal objective value as the primal one, and (6.38) is simply an altered form of the dual feasibility constraint. Thus,
Discriminatory and Anonymous Road Pricing
173
(6.37)(6.38) are both necessary and sufficient to characterize the dual optimal solutions. Note that the firstterm in the righthand side of eqn. (6.37) is in fact the total revenue collected from the toll charge. In reality, negative tolls or subsidies on road networks are difficult and seldom implemented by policy makers, and thus an optimal toll pattern with all link tolls being positive is more meaningful and acceptable to the policy makers. We note that the solution of a LP problem (including the dual problem) is not necessarily unique; this means that in certain cases, we are likely to obtain an anonymous positive link toll pattern from the set of feasible solutions of the dual LP (6.32)(6.33). This is indeed possible by modifying the primal LP (6.28)(6.31) slightly. Consider a slightly modified version of the primal LP (6.28)(6.31) with the equality constraints (6.29) replaced by the following inequality constraints:
I I S / : S o r = va J If,' J\,b'
The optimal solution is
/,mo
™
/•» > Q
Jl,b> J2,b ~
"
= 10, ft = 0, ft = 0, ft = 10, / , ; = 10, ft = 0, ft = 10,
// 4 =10, with a corresponding minimal value of objective function equal to 2700. The dual problem of the above LP is formulated as follows: max 10?, + 10?2 +20?3 +10?4 +20?5 +10uJ +10ua" subject to u:+?,i w , if /„=0,
aeA.
In addition, at the optimum of the revised quadratic
program, the inequality constraints are always binding and hence the system optimum link flows are always reproduced. 6.4.3 An Analytical Example
Figure 6.5 The simple network used in the example Example 6.4 (Valid link toll set for the multiple behavior equilibrium and SO problem) Consider a simple network shown in Figure 6.5, consisting of 3 nodes and 4 directed links.
190
Mathematical and Economic Theory of Road Pricing
Link cost functions are given by f,=6 + 2v,, t2=20 + v2, f 3 =8 + v3, ?4=v4 There are two OD pairs (1 —> 3 and 2 —> 3) in the network, and the demands are fixed to be dn = 20 and d2i = 20. There are three paths connecting the OD pair 1 —»3: path 1 with single link 1, path 2 with link 2 and link 3, and path 3 with link 2 and link 4. There are two paths connecting the OD pair 2 —> 3: path 4 with single link 3, and path 5 with single link 4. Flows through the five paths are denoted by /,, / 2 , / 3 , / 4 and / 5 . The unique link flows at the system optimum are obtained as: v so =(v1so)v2so,v3so)v4so)T =(14,6,11,15)T The corresponding link travel times are given by:
The total travel time at system optimum is 1066. Now we seek anonymous link tolls to decentralize a system optimum to various multiple behavior equilibria. The following three cases are examined and compared. Case A:
OD pair 1 —»3 is controlled by a CN player (denoted by " C "), and the OD pair 2 —> 3 is controlled by a UE player (denoted by " U ").
Case B:
The two OD pairs are controlled by two distinct CN players. OD pair 1 —» 3 is controlled by player " C\", and the OD pair 2 —> 3 is controlled by player " C2 ".
Case C:
Both OD pairs are controlled by a UE player.
Case A. At the multiple behavior equilibrium without toll charge, the link flows routed by each player are:
The total travel time is 7568/7. Now we apply program (6.62)(6.67) to find the unique link flows routed by each player when anonymous link tolls are charged for system optimum. Substituting the above SO solution of link flows and link costs into the program, we have
+— subject to v 2 a +v 2 c =6, v 3 u +v 3 c =ll, v?+v
/3,
/4.
/5SO
This is a simple quadratic programming problem, to which the optimal solution can be easily identified. The link flows routed by each player at optimum are given by
Substituting the above solution into system (6.72)(6.74) yields the following valid toll set: 14u, + 6w2 +11«3 +15w4 +1512 = 2 0 ^ + 20uc
1>u
c
, 54 + « 2 +w 3 >u c , 50 + u2+u4>\ic
The valid toll set can be further simplified into: M, = M2 + M3  8 , M4 = M3 + 4
The resulting anonymous link tolls are unbounded. To seek positive toll patterns for supporting the system optimum as a UECN multiple behavior equilibrium, the aggregate link flow constraints are modified as: v^+v, c Zjfn, = dw,fm ^ 0 , r s ^ , w € W",ae A >, me M (6.78) wEWm f^Rw
f=Rw
J
where v* = (• • •, V™, • j , aE A and v™ is the flow of link a arising out of the OD flows from the set W", Wm is the set of OD pairs of which users are controlled by UE player me M. For a UE player me M, the route choice behavior of its UEfollowing users can be characterized by a minimization of the Beckmanntype objective function over its feasible set:
P7 J ta (v~am + co) dco + p^ I ca (vm + (o)dco 0 where v ; m = v f + Y .
0
(6.79)
J
. v' and \'m = (• • • ,v'm,• • •) , ae A is taken as fixed, P,m and
P™ are the weighting factor of travel time and travel cost in the travel disutility evaluated by
194
Mathematical and Economic Theory of Road Pricing
UE user class me. M. From the firstorder optimality conditions for the UE program (6.79) of UE class m e M , w e have:
K) + PXK))^=JC if/™>0, re^, w e r
(6.80)
a&A
X(P^(v a ) + p; C^, iffrw = Q,reRw,weW"
(6.81)
aeA
where (i*, we Wm are the Lagrange multipliers associated with OD flow conservation constraints and equal the minimum travel disutility between OD pair we Wm by users in UE class me M. A CN player ke K is to minimize the total travel disutility of all users under its control based on the current routing decisions of other players: m v e
"! Z i P M ^ + v ^ + P'c^+v*))/
(6.82)
aeA
where v~*=v M +Y
vJ a n d vM = Y
v' , y ' k = (    , v ' k ,•••) , a e A i s t a k e n a s
fixed, P* and P* are the weighting factor of travel time and travel cost in the travel disutility evaluated by CN player k e K. From the firstorder optimality conditions for the partial SO program (6.82) of CN player keK we have:
, weWk
tffrw=0,reRw,weWk
(6.83)
(6.84)
where C(yll) = &ta (vo)/dvo and c[(ya) =dca (v o )/dv o , and i.* is the Lagrange multiplier associated with OD flow constraints and equals the minimum perceived travel disutility between OD pair weWk
by users under CN player keK. Again we should distinguish
[iw in (6.80)(6.81) and (6.83)(6.84) according to the owner of the concerned OD pair. Namely, \iw represents the actual or experienced travel disutility if OD pair w is owned by a UE player and the perceived travel disutility if owned by a CN player. Now we consider system optimum in terms of total travel disutility minimization. The SO player controls all the flows on the network and freely chooses individual weights Pf° for total travel time and P f for total travel cost, regardless of the weights taken by individual UE and CN players. The SO problem is formulated below:
Discriminatory and Anonymous Road Pricing
195
min r i X K k + Pf 2ca(va)va aeA
(6.85)
aeA
subject to v
a = Sv«m + Z v « '
aeA
> v"" 6 ^" 1 . meM; v*a, aeA we have:
M
if 7™ > 0, re Rw,weW
(6.87)
if fm = 0,re Rw,weW
(6.88)
K
k
where W = W \JW where fF" = LLMtf"" and fF* = [)keK W . Again let V*°, aeA
denote the unique aggregate link flow solution to SO problem (6.85)
(6.86). The classspecific link flows v™ and v* at the SO solution may not be unique, but must satisfy the feasible conditions: \meQ.m, me M and v ' e Q 1 , keK
together with:
Now we compare the SO optimality conditions (6.87)(6.88) and the multiclass UE conditions (6.80)(6.81), and also the multiclass CN equilibrium conditions (6.83)(6.84). It is easy to observe that those SO optimality conditions and the individual player equilibrium conditions cannot be reconciled through the marginalcost pricing even with toll differentiation. The reason is that the criteria (P7>Pr) by UE user class meM (P*,P*) by CN player keK
and
are not consistent with the SO criteria (Pf°,Pf), even if the
'full' marginal social travel time equal to ta (va) + v/a (ya) and the marginal social cost c
a (va) + Vac'a (Va)
6.5.2
on eacn
link
a€
^ are perceived by all users of each player.
Anonymous Link Tolls to Decentralize a System Optimum
Along the same line to obtain an anonymous link toll pattern for the system optimum, we construct the following quadratic program for given unique SO link flow solution v^0 as
wellas ta{v!°), ca{v?°), /.(yf») and c ^ v f ) ,
aeA.
196
Mathematical and Economic Theory of Road Pricing
VJ2$MK
X{XPr*. fcs°K + X Pkfc50 K + X &c° fcso)v"+XP*C fcso K 1
subject to
Ev«m+Ev*=17«so'fle^
(69°)
VT = X X •/""«,, fl 6 ^, « e M
(6.91)
= X. E / A . « e ^ ™ ^
(692)
we IT
/
>0, re7? , w e f
(6.94)
Again in a similar argument as before, the set of feasible solutions to program (6.89)(6.94) is nonempty, and for a classspecific optimal link flow solution, v7™ and ¥*, m&M, ke K, a€ A, the following relationships come from the firstorder optimality conditions of the convex program (6.89)(6.94):
reRw, WGW", m^M
(6.96)
(6.97)
reRw,weWk,keK
(6.98)
Once again, in the above equations, ua is the Lagrange multiplier associated with the link flow constraint (6.90) and u.^ is the Lagrange multiplier associated with the OD demand constraint (6.93). If the travel disutility for all user classes is measured in monetary units, then ua can be viewed as the anonymous toll levied on link a e A regardless of user classes, and [iw as the minimal perceived travel disutility inclusive of link tolls between OD pair weW controlled by a specific player (class), and the group of equations (6.95)(6.98) together with ttie nonnegativity path flow constraint (6.94) imply the multiUEclass and multiCNclass behavior equilibrium conditions.
Discriminatory and Anonymous Road Pricing
197
Here we should note that it is required to measure the generalized travel disutility by each user class in monetary units in order to interpret ua, ae A in (6.95)(6.98) as the anonymous link tolls sought. In this case, the weighting factors, P", me M and P', ke K, attached to travel time in all the relevant equations, corresponds to the value of time of each UE class (player) and each CN player. The weighting factors, ffi° and Pf, for the SO objective, although unrestricted, can be meaningfully chosen as the average of (P™,P*) and (Pr>$*) of all user classes me M, ke K, respectively, weighted by their OD demands. By simplifying the system of equalities and inequalities (6.95)(6.98) as before, we obtain the following polyhedron of valid links tolls:
fcS°)f X PX" + S P ^ V £P& fcS0)fc*)2 +M^S°1 = X M . (699) ^
J
J
keK
^w^O, reRw,weW,
»€»'
me M
reRw,weWk,keK
(6.100)
(6.101)
0
where v"^ , a e ^4 is the given aggregate link flow at SO or the unique solution to the SO program (6.85)(6.86), ~v™, vka, meM, ke K, ae A are again the unique classspecific link flow solutions of convex program (6.89)(6.94). As before, we can modify the program (6.89)(6.94) with the equality constraints (6.90) replaced by relaxed inequality constraints and show the existence of anonymous positive link tolls to decentralize the system optimum into a multiclass multicriteria behavior equilibrium. 6.5.3 An Analytical Example Example 6.5 (Valid link toll set for the multicriteria and multibehavior equilibrium and SO problem)) Consider again a simple network shown in Figure 6.6, consisting of 3 nodes and 3 directed links. In addition to the invehicle travel time, users have to pay the environmental cost. The link travel time functions are given by fi=4v,, t2=50 + v2, ?3=50 + v3 The environmental cost functions are given by
198
Mathematical and Economic Theory of Road Pricing c,=20 + 2v,, c 2 =30 + 2v2, c 3 =10 + 2v3
Figure 6.6 The small network used in the numerical example
45
Flow on Path 1 •6.5
 o  Flow on Path 2
O
—O— Value of Objective Function
15
6.0 0.0
0.2
0.4
0.6
0.8
1.0
Figure 6.7 The SO path flows and objective function value There are four classes of users (denoted by A, B, C and D) traveling from node 1 to node 3. Users in class A and class B are controlled by two different UE players, and users in class C and class D are controlled by two different CN players. The weighting factors for travel time and environmental costs by the four classes are:
# = 0 . 4 , pf = 0.5 , Pf=0.2, Pf =0.6 #=0.6, #=0.5, #=0.8, #=0.4 The demands for the four classes in the single OD pair are dA=\0,
dB = 20, dc =20 and
dD = 10, respectively. There are two paths connecting the OD pair 1 —> 3: path 1 with single link 1 and path 2 with link 2 and link 3. The classspecific flows on these two paths are
denoted by /,", / / , f°, / / , ff,
f2c,
tf,
f2D.
The classspecific link flow solution of the multicriteria, multiclass behavior equilibrium is:
Discriminatory and Anonymous Road Pricing
199
v" = (10,0,0)T, vB = (20,0,0) T , v c = ( 9 , l l , l l ) T , v D =(0,10,10) T Consider the system optimum case with weighting factor pf°+p^° =1.0. The path flow solutions and the SO objective values for the convex SO program (6.85)(6.86) are plotted in Figure 6.7 for reference as Pf° changes from 0.0 to 1.0 ( P f changes from 1.0 to 0.0). Now we seek anonymous link tolls to decentralize the system optimum corresponding to SO weighting factors: Pf°=P^° = 0.5, with the following aggregate link flow solutions, together with link travel time and environmental costs: v S0 =(35,25,25) T , t(v s o ) = (140,75,75)T, c(v s o ) = (90,80,60)T The details of program (6.89)(6.94) are written as follows: c /l c D A min D 1 lOv^ +115vf + 100v, + 120vf + 78v2 + 77.5v* + 79v2 + 77v2 + 67.5vf + 63v3c +69v3D
subject to +v C
i +v.° = 3 5 . V2 + v' +v2c +v2D =25, v3" + v3* +v3c +vf =25
vt+v? v*
 f*
VB
_ ft
vf=/ 2 D , ()
Solving the above simple quadratic program, we obtain the following optimal solution: v s =(7,13,13) T , v c = (l4,6,6) T , vD = (4,6,6)T
V^=(10,0,0)T,
Substituting the above results into (6.99)(6.101) yields the following set of valid link tolls: 8 1 5 0 + 35M, +25w2 +25M, = 1 0 ^
s
^
°
C
133.6 + M, >M
, 16
D
132.8+w,>n , 1 This set can be further simplified into M, = U2 + M3 + 3 0
One can, of course, further add nonnegative constraints to the above link toll set and the resulting nonnegative toll set remains nonempty and valid, as discussed before.
200
6.6
Mathematical and Economic Theory of Road Pricing
SUMMARY
In this chapter we examined the existence and properties of the valid link tolls that can decentralize a system optimal flow pattern into a standard user equilibrium or various multiclass, multicriteria behavior equilibria. Our general point is that the toll charge must be anonymous or uniform in the case of unobservable user heterogeneity, such as users with different values of time or different routing behaviors but with the same type of vehicles. In the case of single class or homogeneous users with identical VOTs, the relationship between UE and SO and the properties of valid tolls holds regardless of the cost (money) or time units that are used in expressing the objective function of the system optimum and the criterion for user equilibrium. The valid link toll set for SO is in fact a convex polyhedron in terms of a system of linear equalities and inequalities. All valid link tolls give rise to the same constant total toll revenue in the elasticdemand case. In the case of multiclass users with a discrete set of VOTs, we showed that there exists an anonymous link toll pattern that supports a SO as a multiclass user equilibrium, whenever the system objective function is measured by either money (cost) or time units. For the costbased SO, the anonymous link toll is equal to the user externality of travel time multiplied by the arithmetic mean of the values of time of all users traversing that link. For the timebased SO, the anonymous link toll is not based on the user externality, but can be determined from the solution to an artificially constructed linear dual problem. With multiple equilibrium behaviors, such as the presence of both UE and CN users on a common network, we showed that, applying the traditional marginalcost pricing for SO requires link tolls to be differentiated across user classes, because users of different classes have different perceived costs on each link in their route choice. By artificially constructing a convex quadratic program, we then showed that alternative meaningful, nonnegative, and anonymous link tolls identical to all user classes do exist to decentralize a SO flow pattern as a UECN multiple behavior equilibrium. The desirable link tolls are generally nonunique, but constitute a nonempty polyhedron, again expressed in terms of a linear system. We further generalized our results to the multiclass, multicriteria behavior equilibrium and system optimum problems. In fact, our findings from this chapter can be generalized as follows: Any feasible link flow pattern can be supported by anonymous link tolls as a deterministic traffic equilibrium. Here, feasibility of link flow patterns means that there exist path flows that satisfy the OD demand constraints and induce the given target link flows; the
Discriminatory and Anonymous Road Pricing
201
deterministic traffic equilibria herein comprise the standard user equilibrium, the equilibrium with multiple VOT user classes, the equilibrium with multiple UECN routing behaviors, and their combinations examined in this chapter. One should realize that the anonymous valid link tolls are not necessarily positive, depending on the nature of the chosen target link flow pattern. Finally, we point out that the explicit characterization of the valid link tolls set for decentralizing the SO flow patterns allows for further extraction of optimal link tolls with secondary objectives such as minimization of total revenue collected or minimization of number of toll booths.
6.7
SOURCES AND NOTES
The toll pricing framework in Section 6.2 for the single class traffic equilibrium and system optimum have been extensively developed by Bergendorf et al. (1997), Hearn and Ramana (1998), Hearn and Yildrim (2002), and recently by Yildirim and Hearn (2005). Selection of optimal link tolls with secondary objectives are also considered by Dial (1999c, 2000), Bai et al. (2004), Penchina (2004) and Lawphongpanich and Hearn (2004). The multiclass UE and SO problem in Section 6.3 is based on the recent work of Yang and Huang (2004), and further extension was made by Yin and Yang (2004). The SO and multiple equilibrium behaviors presented in Sections 6.4 and Section 6.5 was recently developed by Yang and Zhang (2003) and Yang, Zhang and Meng (2003), and details of the materials in the two sections are also available in the thesis of Zhang (2003). Reader may further refer to Haurie and Marcotte (1985) about the relation between Wardropian equilibrium and the CournotNash game, to Harker (1988) about the multiple behavior equilibrium, and to Nagurney (2000) and Nagurney and Dong (2002) about multiclass, multicriteria traffic equilibrium. Additional background on anonymous pricing can be found in Arnott and Kraus (1998) (also, Lindsey and Verhoef, 2001) in the case of bottleneck congestion models.
This Page is Intentionally Left Blank
SOCIAL AND SPATIAL EQUITIES AND REVENUE REDISTRIBUTION
7.1
INTRODUCTION
Although congestion pricing is theoretically and technologically easy to implement, it has long been viewed as a political issue. A common criticism is that road use charges make unequivocally distributional impacts on users with different incomes. Generally speaking, the equity implications of congestion pricing are complex because of all the different options facing users under a congestion pricing scheme. People who continue to use the highway after a toll is imposed, pay the toll, but also have a lower travel time: the toll decreases traffic volume, which decreases travel time. Some users with very high values of time (VOT) would find that they become better off (the reduced congestion can more than compensate the users for the extra cost of toll charges). However, those with low values of time but still using the roads are generally made worse off than before. People who stop using the highway avoid the toll, but forgo the benefits associated with using the highway and experience the inconvenience of switching to another mode of transport. This inequitable issue among different social classes of users is termed the social equity issue and becomes the primary focus of road pricing arguments. Apart from the social equity that is often measured with respect to income (and hence VOT), the spatial equity issue is directly applicable in the current context of road pricing. It is about accessibility to different points in space, or specifically about users traveling between different locations. It is evident that after introducing congestion pricing in a road network, the changes in travel costs (inclusive of toll charges) between different origindestination (OD) pairs can be substantially different, depending on the amounts and locations of toll charges. Thus people traveling between different OD pairs will receive different effects from congestion pricing scheme s, and this could result in another kind of unfairness on users and hence become a new obstruction on the implementation of pricing policy due to the public rejection. It is very often the case that social and spatial segregation go hand in hand, their treatments thus requires explicit considerations of the various social groups (particularly the poorest
204
Mathematical and Economic Theory of Road Pricing
social group on the basis of income) and the various (particularly the least wellserved) geographical zones. One possible way to allay these equity concerns raised is to adopt a revenue neutrality policy, through a direct even or uneven redistribution of the revenue generated by the congestion charge, among users in different social and spatial groups. It is also desired to take explicit account of social and spatial equities in designing a road pricing scheme, even without direct revenue redistribution. In this chapter, we show that a congestion pricing scheme that reduces the total travel time and redistributes the toll revenue to all users can make everyone better off, and then we propose various Pareto refunding schemes. Without revenue redistribution, we explicitly incorporate the social and spatial equity constraints into the bi level programming models for the secondbest network toll design problem examined in Chapter 5. Social and spatial inequities are simply measured by the maximum relative increase of the generalized equilibrium OD travel costs between all OD pairs for various classes of drivers with different VOTs.
7.2
THE SOCIAL AND SPATIAL INEQUITY: AN EXAMPLE
Example 7.1 (Demonstration of the social inequity in road pricing) We begin with a simple example to demonstrate the aforementioned social and spatial inequity problem. The network depicted in Figure 7.1 consists of 2 parallel links connecting a single OD pair. There are two classes of users, denoted with equal travel demand dx = d2 = 5 (flow unit, say in 103 veh/h) traveling from node 1 to node 2, and their VOTs are assumed to be P, = 1.0 and (32 = 2.0 (HK$/min), respectively. Link travel time functions (in minutes) are shown in the figure: Linkl
/ 2 (v 2 )=v 2 +10
Figure 7.1 Network used for Example 7.1 Traffic equilibrium in the absence of toll pricing In the absence of a toll charge, it is straightforward to obtain the unique aggregate equilibrium link flow, v, corresponding equilibrium link travel time, t, total system travel disutility, C in money units (vehHK$), and T in time units (vehmin), as follows:
Social and Spatial Equities and Revenue Redistribution
205
The OD travel disutility by each class in time and money units is W = — (HK$) = — (min); ^ 2 = — (HK$) = — (min)
(7.1)
As stated in Section 2.3 of Chapter 2, the timebased and the moneybased equilibrium problems for multiclass users with distinct VOTs are equivalent to each other in the presence or absence of toll pricing given exogenously. The above aggregate equilibrium link flow is unique and independent of the unit (time or money) to measure OD travel disutility, but the classspecific link flow is not unique. The costbased system optimum and anonymous link tolls As examined in Chapter 6, we now consider the system optimum in terms of travel cost minimization (system disutility is measured in monetary units and is denoted by C): min
C = X E Pmvmja(v.) = (vj + 2v2){lv\ + 2vf + 6) + (v2 + 2v\)(v2' + v2 +10)
vJO.v SO
3=1,2 m=l,2
subject to v,'+v 2 =5; v,2 + v 2 = 5 Like Example 6.2 in Chapter 6, the objective function is nonconvex and there are minimizing points and one saddle point. The local mimimum is given by 6 6
29 2iY t_/
f _ ( i l _ _ Y c=^
r =1849
where T is the corresponding system disutility measured in time unit). The global minimum is given by '43 12'12 , V 1*
XT 2/
[ 4 3 77Y I '  t ' i ' t *
.
.
,XT
V P 2l 2/ Vl
( 7 9 197Y r
'
n
~ '
5351
7321
>yi
12 12J { 6 12 ) 24 48 The valid link tolls to decentralize the global costminimizing system optimum is given by 13 which includes anonymous firstbest link tolls
U = (M,,U2)
=(43/3,47/6) for the costbased
system optimum. In this case, the OD travel disutility inclusive of equivalent toll charge by each class is
206
Mathematical and Economic Theory of Road Pricing
(7.2)
)
(7.3)
The total revenue IT is given by IT = 559/24 + 10u2 , leading a revenueneutral ( n = 0) pricing u = (M P M 2 ) T = (l001/240,559/240)T. The timebased system optimum and anonymous link tolls Next we look at the system optimum in terms of travel time minimization (system disutility is measured in time unit): mi? X V A ( v j = v, (2v, + 6) + v2 (v2 +10) =1,2
subject to v,+v 2 =10 The timebased aggregate link flow SO solution is given by: v = (v,,v 2 ) T =(4,6) T , t = ( ^ 2 ) T = ( 1 4 , 1 6 ) \ C = indeterminate, 7 = 152 Note that the system disutility in monetary units is indeterminate, because the classspecific link flow is not unique at SO and travel times on the two links are not identical (different from the untolled UE case where both timebased and costbased system disutilities are well defined). Again from Chapter 6, anonymous link tolls to decentralize the above timebased SO flow pattern can be obtained as M, — u2 — 4 = 0
With the above anonymous link toll charge, the classspecific equilibrium link flows together with total system disutility in monetary units are uniquely determined as v'
= ( v ;,v 2 ) T =(0,5) T , v 2 =(v 1 2 ,v 2 2 ) T =(4,l) T , C = 224, T = 152
The OD travel disutility inclusive of equivalent toll charge by each class is (7.4) H 2 =(32 + w 2 )(HK$)= 16 + M 2 ](min)
(7.5)
The total revenue n is given by n = 16+10w2, leading a revenueneutral (11 = 0) toll charge u = («1,«2)T =(12/5,8/5) T . System efficiency and inequity measure for alternative pricing schemes
Social and Spatial Equities and Revenue Redistribution
207
Table 7.1 compares the system efficiency in terms of total system travel disutility in time and monetary units respectively for the alternative pricing schemes discussed above. Two points are worthwhile mentioning. First, the timebased and (global) costbased SO enhances the system efficiency in terms of total system disutility, in both time and monetary units in comparison with untolled traffic equilibrium. The local costbased SO actually worsens system performance in terms of total travel time. Therefore, it is critical to identify the global optimum solution for the costbased SO program with a nonconvex objective function. Table 7.1 Comparison of system efficiency for alternative pricing schemes
Untolled Equilibrium
Total Costbased System Disutility (vehHK$) 230.00
Total Timebased System Disutility (vehmin) 460/3 = 153.33
Local Costbased SO
2759/12 = 229.92
1849/12 = 154.08
Global Costbased SO Timebased SO
5351/24 = 222.96 224.00
7321/48 = 152.52 152.00
Our attempt here is to look into the inequity impacts on the two user classes brought about by the above realistic nonnegative, anonymous link tolls to decentralize the timebased and costbased system optimum into multiclass traffic equilibria. To do this, we now look at the ratios of the OD travel disutility or generalized OD travel costs for each user class after and before introducing anonymous pricing. As noted previously, the timebased and the costbased equilibrium problems for multiclass users with distinct VOTs are equivalent to each other, and the said ratio of OD travel disutility for a specific user class is independent of the unit (time or money) used to measure the travel disutility. Now we provide the ratios of OD travel disutility of the two user classes in the case of costbased SO with travel disutility given by (72)(7.3) and timebased SO by (7.4)(7.5) to the case of untolled equilibrium given by (7.1). 1
[l84 1
23
46 2 J 46
2
2
2
^184
92
23 92
2
2
Figure 7.2 plots the change of the above ratios for the costbased and timebased SO programs, respectively, against equilibrium link toll charge, u2, starting from their revenueneutral toll charges determined above: u 0 0 " 0 =(w,,M2)T =(1001/240,559/240)1 and =
made.
( M i j M 2 y r = (12/5 ) 8/5) T . From the two figures, the following observations are
208
Mathematical and Economic Theory of Road Pricing
1.80 1.60 •2
1.40
I 10 ° 
0.80  User Class 1
£ 0.60
User Class 2
Q
O 0.40
2.33: Revenueneutral pricing
0.200.00

3

2

1
0
1
2
3
Equilibrium Toll Charge on Link 2
a) The case of costbased SO pricing policy 1.80 1.60.2 1.40% 1.20
I
10
°
s •3 0.80H 0.60
• User Class 1
User Class 2
Q
° 0.40 
1.60: Revenueneutral pricing
0.200.00

3

2

1
0
1
2
3
4
5
Equilibrium Toll Charge on Link 2
b) The case of time based SO pricing policy Figure 7.2 Change of travel disutility ratios between the SObased pricing policy and the untolled traffic equilibrium against toll charge on link 2 First, the OD travel disutility ratios of course increase with toll charge, but the increasing rate for user class 1 is higher than that for user class 2 in both costbased and timebased SO pricing policies. This indicates an inequitable impact of toll charge to the users with different VOTs, or lowerincome users are more likely to be made worse off than before (social inequity). Second, t is interesting to note that revenueneutral toll pricing policy can bring
Social and Spatial Equities and Revenue Redistribution
209
good news to both classes of users and of course to the whole society. As observed in Figure 7.2, the revenueneutral pricing policies for both costbased and timebased SO programs are indeed Paretoimproving; the travel disutility change ratios are strictly less than unity for both user classes. This observation raised the possibility that an appropriate pricing and revenue refunding scheme may make all users better off. Example 7.2 (Demonstration of the spatial inequity in road pricing) Consider a simple network example depicted in Figure 7.3, consisting of 3 nodes and 3 links. Link travel time functions are given below:
2 Figure 7.3 Network for Example 7.2 There are two OD pairs from node 1 to node 3 and from node 2 to node 3, with demand rf,^3 = 5 and d2^ = 15, respectively. For OD pair 1 —> 3, there are two classes of users, with travel demand d\^ = 1 and d\^ = 4 , VOT P, = 2 and P2 = 1, respectively. For OD pair 2 —> 3, there is only one single class of users with VOT p = 1. In the absence of toll charge, traffic equilibrium solution is given by: v,=0, v 2 = 5 , v3 = 20; F, =50, F2 =18, F3=26 ^ = 4 4 , (la™ =26, f = 610, C = 654 The timebased system optimum solution is given by: _
=
110W39I
3J h

=
260
5371

T=
9226039 10A/39T :!
_ , ... ,.
5^95 ^
^ ^
. ,
1 2 9
9
_
)
5^50
=1629
06
. „
t = 553.51, C = indeterminate..
27 One of the simplest positive link toll schemes to decentralize this timebased SO is given by
which gives rise to:
210
Mathematical and Economic Theory of Road Pricing
v? = 0 , 3 = 1 ,
3 2605V391
_time
58120^91
For OD pair 1 > 3,
which shows that the the relative changes of travel time costs are different for different user classes with different VOT, even if they are traveling between the same OD pair, and thus we encounter the social inequity issue associated with the road pricing considered here. In addition, we also have the following spatial inequity issues by furthering looking at the relative change of travel time cost between OD pair 2 —> 3 : F%
7.3
20.61 " = ° 79 ( s P a t i a l inequity) 26
REDISTRIBUTION OF CONGESTION PRICING REVENUE
The possible way to allay the equity concerns raised in the earlier section is to adopt a revenue neutrality policy through a direct redistribution of the revenue generated by the congestion charge among users in equal or unequal shares. In this section, we show that a congestion pricing scheme that reduces the total travel time for given OD travel demand and redistributes the toll revenue to all users can make everyone better off. We investigate the conditions under which this observation still holds with an even redistribution across users traveling in the same OD pair, regardless of their values of time and how this ODspecific even redistribution can be undertaken without crosssubsidization or wealth transfer among users traveling in different OD pairs. 7.3.1 Existence of a Pareto Refunding Scheme Like Section 2.3 in Chapter 2 and Section 6.3 in Chapter 6, we consider a multiclass traffic equilibrium problem with a set of discrete VOTs. Let M denote the set of all user classes, Pm (Pm > 0) be the average VOT for users of class me M and d™ be tie fixed demand for travel of class meM
between OD pair we W. Consider a network under a toll system
u = (...,«„,...) ,aeA
where A is a subset of links subject to toll charge. Let
Social and Spatial Equities and Revenue Redistribution
211
v™, ae A, me M be the equilibrium flow distributions on the network in the absence of congestion charge (u = 0) and the corresponding travel times be ta (v), a e A . Furthermore, the equilibrium flow pattern after the toll system is introduced is denoted by V™, a e A, me M, with corresponding travel times ta (v), a e A. Here we consider a general travel
time
function
ta (v)
that
allows
interdependence
among
link
flows
We consider a general secondbest pricing scheme, which means Ac A or A = A. The congestion pricing scheme is meaningful only if it reduces the total system disutility. As pointed out in Chapter 5, in the presence of multiple user classes with different VOTs, the total system disutility can be measured in either cost (monetary) or time units. This means that a congestion pricing scheme is desirable if either one (or both) of the following two conditions is met
T=YJta{^K™, weW', meM], such that *=n
( 7  22 )
v^W neM
and P C r f " " ®" < VTmdZfarany weW, meM
(7.23)
Proof. Let
'meM
where a™ is a positive number (a™ > 0 for all we W, m e M) and satisfies
a>O,weff,meMand £ £c£ = 1 meM weW
Then,
y y om = y y (rrsi W ^  «
Equation (7.34) simply means that w s w meM
J C 
be
t n e tota
l travel time of
users
between OD pair we W and
ATw = Twfn, weW be the corresponding change after introducing the pricing scheme; similarly, let AT = ^
wATw=T
assumption, A 7 < 0 ; ATw,weW
f
be the change in total system travel time. By our
can be either negative or positive.
Let w be the total amount of refund to be evenly returned to all classes of users in OD pair weW and consider the following refunding scheme: OW = UW+AUIV,WBW
(7.36)
where A]\w denotes the crosssubsidizations among different OD pairs, and is determined by: $ATW, if A7; > 0 AEL = PA7;, if ATW < 0 0, if &TW = 0
(7.37)
where P = max{Pm, me M) and P = min{Pm, me M}. We say a refunding scheme O = {w, we W] is feasible if the total amount of refund needed is less than or equal to the total revenue generated from the pricing scheme concerned, or if it satisfies:
5>^]1
(7.38)
We also define the subsets of OD pairs in the network in which total OD travel time is increased or decreased after introducing the pricing scheme respectively as W* ={w\ATw>0, weW); W ={w\ATw = {•&„,, we W}, defined by eqns. (7.36) and (7.37), to be feasible is
Social and Spatial Equities and Revenue Redistribution
fT
221
~ \AT\ "p^p"
Proof.
XAFL = P X AT; + P E AT; = (pp) £ AT n€)f
\xw*
«sn' +
tx\r
«€W
=(pp)5>rw+[kr Clearly, if eqn. (7.40) holds, then we have ^^w
(7.41)
AU.W ^ 0 and hence the proposed refunding
scheme given by (7.36) and (7.37) is feasible. Conversely, if E^^AII*, ^ 0 , then eqn. (7.41) leads to eqn (7.40) immediately. ] From (7.40) and (7.41) it is trivial to note that if
ItAT" P TT
(7.42)
PP
then we have = 0; Y O
=FI
(7.43)
and if
ff
\ such that $>"w >O w , we W and ^ ^ ^ ^ C =YlLemma 7.3 The refunding scheme given by eqns. (7.36) and (7.37), if feasible, satisfies the following equations:
PC  —  f O
we w m
'
e M
(7 46)

where "= " may holdfor only three cases a) ATw=0 b) ATw>0, a n ^ P m = P c) ATw
Ow, we W and ^
„,
(7.53)
we obtain
C = JCt  4C1 ,weW,me M
(7.54)
From O™ = §1d™ and eqn. (7.22), we have
e = \ ?. >. U;.J.:j(™ is readily given by (7.54). Given that e is a measure of a uniform betteroff degree across all users in the network, eqn. (7.56) simply implies that minimizing the costbased system disutility C is equivalent to maximizing the average betteroff degree of all users for a road pricing and revenue refunding scheme.
224
Mathematical and Economic Theory of Road Pricing
Note that the above refunding formulas (7.53){7.56) apply only if the various user classes can be identified and the crosssubsidizations among users in different classes and OD pairs are possible. Otherwise, one has to check whether the conditions for Theorems 7.3~7.5 are met and then apply the more plausible refunding schemes under the required conditions. 7.3.3
Remarks on the Pareto Refunding Schemes
Needless to say, the efficiency of a tolling system can be evaluated by the amount of total travel time savings, and the firstbest pricing scheme induces a UE flow pattern to a system optimum and thus achieves the maximum efficiency gain which can be shown to be bounded in terms of the percentage reduction of total travel time for fixed OD demands. While everyone can be made better off, one intriguing question remains here as to how the system efficiency improvement and individual benefit gain are related under a pricing and revenue redistribution scheme. Specifically, it is not clear whether a pricing and refunding scheme that can bring a greater decrease in total travel time can nevertheless give rise to more positive individual benefit gain. In the Pareto refunding scheme developed above, we assumed that the OD demand is fixed before and after introducing the tolling system. This is in some situations a restrictive assumption, because road pricing certainly has longterm impacts on travel demand. Furthermore, congestion pricing is in general introduced to ensure that users recognize the true social traveltime costs of their tripmaking. It is thus immediately clear that refunding will indeed create distortions and the firstbest conditions for the socialwelfare maximization with elastic demand cannot be optimal, because users will not be able to perceive their real marginal social cost of tripmaking. Nevertheless, the refunding scheme does report good news in that a socially optimal flow pattern for given demands can be accomplished by offering users compensation incentives for selecting socially optimal routes on a general network, while maintaining revenue neutrality. Incidentally, the refunding scheme does not discourage people from driving; this observation itself in turn justifies the assumption of fixed OD demands in the refunding model. One may find that the constant travel demand is in fact a rather innocuous assumption and more good news can be shown if applying the above method and results to a multimodal transportation network. One may reasonably assume that commuting is a necessity to all users and users divide themselves among the various modes and routes based on their preferences for trading money for time saving, then the pricing and refunding scheme developed for fixed travel demand does make sense in developing a transmodal integrated transport pricing and subsidy strategy that is both equitable and efficient. Under such a strategy, people tolled off the road often shift to transit, and road pricing revenues can be used to benefit users of more socially desirable modes, with lower social marginal cost like
Social and Spatial Equities and Revenue Redistribution
225
the transit mode. Indeed, the proposed refunding scheme does allow for both positive (subsidy) and negative (further charge) redistribution of congestion pricing revenues among socially and geographically diverse groups of users.
7.4
NETWORK TOLL DESIGN MODELS WITH EQUITY CONSTRAINTS
The refunding strategy proposed in the previous section to address the social and spatial equity issues is to maintain " revenue neutrality" by returning congestion to each class users. if, reasons, a direct redistribution users cannot be adopted, would then necessary address impacts on equities designing an equitable efficient network tolling scheme. this section, develop toll design model with explicit consideration social spatial equities. 7.4.1
Traffic Equilibrium and Social Welfare with Pricing
Differing from the previous sections, we consider demand elasticity in the multiclass network equilibrium model. This model adopts the discrete approach for VOT distribution, which is considered to be particularly suitable for congestion pricing modeling with elastic demand. The whole society is divided into a number of classes and each class is assumed to have an average VOT belonging to a certain interval, and can be characterized by classspecific demand functions. This treatment in certain cases can be regarded as a discrete approximation of the model using continuously distributed VOT and has some practical advantages. It is consistent with the conventional market segmentation approach for travel demand modeling and thus greatly facilitates model calibration with actual data. Following the previous notation, let A be the subset of links with toll charge and B™ (d™) be the classspecific benefit function (inversed demand function) of user class me M between OD pair weW. User classes are ordered according to their increasing VOTs, namely, P = Pi < P2 <    < P^ = P. Let v " = ( v ; , a e i ) T and dm =(d",we wf, and the equivalent convex program formulation for the elasticdemand multiclass network equilibrium problem is given below (Section 2.3, Chapter 2):
where
fZ^0,reRw, r£Rw
weW\, meM
(7.58)
226
Mathematical and Economic Theory of Road Pricing
Note that the travel disutility used here in characterizing the route choice behavior is measured in time units where the toll is converted into corresponding travel time by a given VOT of each class. All standard assumptions hold for the link travel time and OD benefit functions for the unique solution of aggregate link flows and classspecific OD demand. For a given traffic equilibrium solution to program (7.57)(7.58) associated with a pricing scheme u, we can define the corresponding social welfare, S (u), in time units as follows:
f K{ (7.59)
Here \m (u) and dm (u) are solutions to the UE program (7.57) for a given link toll u . Although the classspecific equilibrium link flow, v m (u), is not unique, the aggregate equilibrium link flow is unique and hence the social welfare, S, in (7.59) is well defined for given u. 7.4.2
Specification of Equity Constraints
In reality, it is very difficult, if not impossible, to design a toll pattern that brings a completely equitable impact on all users traveling between different OD pairs. A simple yet practicable method is to prevent the percentage increase of travel cost of all network users from exceeding a certain threshold in designing a toll scheme. In this case, we can say that the relative inequity impact from a toll scheme is limited to a certain level. With this in mind, we deal with the equity issue by incorporating an equity constraint in the upperlevel problem. The equity is measured as the relative change of the OD travel disutility (inclusive of toll charge) and thus the equity constraint can be specified as follows: ^"^
<X , weW, meM
(7.60)
The term (iw is the original equilibrium OD travel disutility in time units without pricing (which is commum to all user classes), and u™ , (u) is the equilibrium OD travel disutility in time units (inclusive of toll charge) for class m after introducing pricing scheme u; %w is designated here as an equity index that dictates the degree of tolerance of the inequity associated with a pricing scheme. A sensible selection of the value of this index is given by: (7.61) 1, otherwise
Social and Spatial Equities and Revenue Redistribution
227
Here the term \xw is the OD travel disutility after the firstbest or marginalcost pricing scheme is implemented; f \^'(u), we W, for any anonymous link pricing scheme with nonnegative tolls, where user classes are ordered according to increasing VOT. This is evident. Consider any two classes of users, class m and class n with Pm < P n , traveling between a given OD pair we W. Let tm and um (um > 0) be the travel time and total toll charge along any route reRw. c^ = trw + um/$m>c?M,= trw+urw/$rt for any reRw,w€W,
Clearly, we always have
and thus
JC>M£,
we W as
long as Pm < P n . In view of the fact that I™ = \inw = £„, in the absence of pricing, the set of constraints (7.60) can now be replaced by
! i ^ M < x (cp), weW
(7.62)
Note that, as mentioned in Section 7.1, the equilibrium travel disutility ratio in (7.60) for a specific class is independent of its measurement unit (time or money) and hence eqn. (7.62) is always sufficient to ensure all equity constraints are met, despite the fact that ^(u^i^u^^^'^),
weW,
for (3, < (32 < ••< PA/, if the travel disutility is
measured in monetary units (c^ = $mtrw + um j . 7.4.3
Network Toll Design with Equity Constraints
With the aforementioned social and spatial equity considerations, the network toll design problem with an equity constraint can now be formulated as the following bilevel programming problem:
mEM
\*=W
0
subject to (7.64)
Social and Spatial Equities and Revenue Redistribution
229
where v m (u) and d"1 (u) solves the equilibrium problem (7.57) with the timebased OD travel disutility ^^(u), weW associated with the lowest VOT user class. In the model, Xw((p), we W is specified by eqn. (7.61) and U is the feasible set of link tolls defined by
U = { u uf < ua < « r , a € 1}
(7.65)
where u™ and «™ax are the prespecified lower and upper bounds of link tolls ua, ae A. As pointed out earlier, parameter cp (0 < cp < 1.0) reflects the allowable degree of inequity in terms of the OD travel disutility ratios after and before implementing a pricing scheme. This parameter is selected by the decisionmaker and can be, in fact, treated as a decision variable in the programming model. For each given (p, let S'(q>) be the maximum social welfare obtained by the bilevel model (7.63)(7.64), ^'(cp) can be regarded as an implicit function of parameter cp. By incorporating this equity decision parameter (p, we have the following trilevel programming model with dual toplevel objective functions: (7.66)
where, according to the specification and analysis of (p in Section 7.4.2, S' (0) with (p = 0 generally corresponds to the 'minimum' social welfare in the donothing or unto lied situation; while S' (l) with
l'tnen Step5.
uniformly generate, at random, a point denoted by x over U.
(Metropolis' rule) If F(\)< F(uM) random [0,1] when
F(X)>F(U4^5=2000 (veh/h) for the four OD pairs ( l  > 5 , 2  > 5 , 3  > 5 , 4 > 5). Parameter y is set to be 0.01 for all user classes.
Figure 7.5 The network used in the numerical example Table 12 Input data for test network in Figure 7.5 Link C (min) C (veh/h)
1 10
23
42
10
15
10
23
42
1000
2000
3000
4000
6000
1000
2000
3000
Suppose there are 3 user classes with m = l,2 and 3 representing low, medium and high income user class, respectively. The potential market share of the three classes are assumed to be identical for all OD pairs and of values of 20%, 60% and 20%, and their VOTs are 75, 100 and 150 (HK$/h), respectively. Without implementation of a roadpricing scheme, the equilibrium OD travel disutility is presented in Table 7.3. The social welfare gain is 1.7701xl04 (vehh). When differentiated
234
Mathematical and Economic Theory of Road Pricing
marginalcost pricing link tolls are charged on each link, and the resulting generalized OD travel disutilities (in time units) are also listed in the same table, and the social welfare gain increases to a maximum value 1.7940x10" (vehh). Table 7.3 also shows the ratios of the OD travel disutility after and before marginalcost pricing is introduced, which are identical among different user classes in the same OD pair. It can be seen from the table that travel disutilities between all four OD pairs increase after the marginalcost pricing link tolls are charged. In particular, the travel disutility between OD pair 4 —> 5 increases most. Table 7.3 Change in equilibrium OD travel disutility before and afterroad pricing ODpair Travel disutility without pricing (min) Travel disutility with marginalcost pricing (min) Ratio of equilibrium OD travel disutility
l>5
2>5
3>5
4>5
42.631
30.587
42.779
19.544
48.651
37.120
49.084
25.408
1.141
1.214
1.147
1.300
Table 7.4 Numerical results for VOTs = 75, 100, 150 (HK$/h) 9 «, (HK$)
0.0 0.00
0.2 0.30
0.4 1.35
0.6 2.91
0.8 4.22
1.0 5.21
u2 (HK$) Welfare (104 vehh) Revenue (104 HK$) Binding OD pair
0.00
2.90
4.35
5.01
5.68
5.81
1.7701 0.0000 All
1.7815 2.0317 2>5
1.7866 3.2232 2>5
1.7898 4.0264 2>5
1.7908 4.6898 2»5
1.7910 5.0274 no
Now we consider a secondbest pricing scheme with link 4 and link 5 subjected to toll charge for maximizing social welfare, while considering equity constraints with model (7.63)(7.64). Table 7.4 shows the numerical results for 10 different values of the inequity threshold (p. The last row of the table lists the OD pairs, for which the constraint, ]l[,(u)/fiw < %„ (5 for cp = 0.1~0.8, this is clearly due to the fact that a free alternative route is not available for traveling from node 2 to 5 and lowincome users from 2 to 5 suffer most from the pricing scheme. As cp > 0.9 relaxing the equity constraint does not lead to further increase in social welfare (objective function), and all equity constraints become inactive.
Social and Spatial Equities and Revenue Redistribution
235
In fact, the dispersion of VOT distribution across user classes has significant impacts on the analysis result. Namely, as the VOT distribution is more dispersed, the inequity effect among user classe s becomes more profound, and thus the equity constraints act more tightly. To verify this effect, Table 7.5 presents numerical results for another set of more dispersed VOT distribution given to be 50, 100 and 200 (HK$/h)), for the three classes. In comparison with the former case, optimal tolls for both links are smaller for the same cp value. Even when cp increases up to its maximal value of 1.0, the equity constraint is still binding. Thus the social welfare for 0 < cp < 1 cannot reach its maximum value achieved in the case without equity constraints. This highlights the fact that both social and spatial equity issues deserve more attention as VOT differential across users becomes more remarkable. Table 7.6 provides the numerical results for 5 different values of weighting factor to. As to increases, or as more emphasis is placed on maximization of social welfare than on the equity constraint, the value of 5
1.7839 2.3587 2>5
1.7871 3.1705 2>5
1.7895 3.8562 2>5
4.66
Table 7.6 Numerical results for VOT = 75,100,150 (HK$/h) CO
5
1.7866 3.2232 2>5
1.7901 4.2024 2>5
1.7908 4.5504 2>5
We now demonstrate how the two kinds of inequities examined here occur on the network and how the proposed models work. Figure 7.6 plots the percentage change in the average OD travel disutility (average over 3 classes weighted by their shares) for the 4 OD pairs.
Mathematical and Economic Theory of Road Pricing
236
From this figure, we can see that when cp is less than 0.4, the average OD travel cost of 4 —> 5 increases most. As cp grows up, travel costs of 2 —> 5 increases most.
 O  OD from 1 to 5  D  OD from 2 to 5 OD from 3 to 5  *  OD from 4 to 5
0 0.0
0.1
0.2
0.3
0.4 0.5 0.6 Equity Parameter
0.7
0.8
0.9
1.0
Figure 7.6 Percentage change in weighted average OD travel disutility
—O— highincome class from 2 to 5 —D— mediumincome class from 2 to 5 owincome class from 2 to 5 X highincome class from 3 to 5 mediumincome class from 3 to lowincome class from
5 0.0
0.1
0.2
0.3
0.4 0.5 0.6 Equity Parameter
0.7
0.8
0.9
1.0
Figure 7.7 Percentage change in OD travel disutility by user class Figure 7.7 depicts the percentage change in OD travel disutility by user class for two selected OD pairs 2 —»5 and 3 —> 5. It is observed that for the OD pair 3 —> 5, highincome class users actually benefit from the toll charge. As cp increases up to about 0.70, they enjoy more benefits, and as cp increases further beyond 0.70 they enjoy less benefits. In contrast, all other
Social and Spatial Equities and Revenue Redistribution
237
users suffer from the toll charge, and suffer more losses as (p increases. Note that for OD pair 3 —> 5, the medium and lowincome class users have equal travel cost because they are both exclusively concentrated on the nontoll path (link 8). These results reveal that parameter (p plays an essential role in controlling the spatial and social equity impacts of a networkpricing scheme on heterogeneous users.
7.7
SUMMARY
After a general discussion of the social and spatial inequity issues in congestion pricing and a demonstration with a simple analytical example, this chapter first proved the existence of a Pareto refunding scheme that returns the congestion pricing revenues to all users to make everyone better off. More plausible refunding schemes are sought under specific conditions, such as whether aggregate travel times or costs by individual OD pairs are reduced, after introducing a tolling system. The Pareto refunding schemes developed in this chapter could be used as a basis for an acceptable implementation of congestion pricing in terms of both economic efficiency and equity. It holds promise as a way for invoking congestion pricing and using pricing revenue in a constructive manner that is viewed favourably by the public. With the advent of intelligent transportation systems, it is possible today to create complex systems capable of operating direct redistribution methods of congestion pricing Without direct redistribution of congestion pricing revenues to users, we then proposed and solved mathematical programming models that explicitly address the social and spatial equity concerns in designing a networkwide efficient pricing scheme. The impact of congestion charge on spatial equity is measured by the relative changes in Ihe generalized OD travel disutility in each OD pair, whereas the impact on social equity is considered by examing the inequitable relative changes in travel disutilities of users of different social classes. Both equities are incorporated into the bilevel network toll design model as constraints. A penalty function approach, combined with a simulated annealing method, was applied for solving the equityconstrained network toll design problem. As demonstrated with numerical examples, the proposed models and algorithms allow for selection of a fair and reasonable link toll pattern that maximizes social welfare gain, and meanwhile attempts not to bring excessive negative impact on certain groups of users. Finally, we should recognize the fact that, although the various equity issues and revenue allocations of congestion pricing have been argued extensively in the literature for more than a decade, only until recently, did their rigorous quantative analysis receive attention. There are many worthwhile topics for further studies. It remains an open question as to what extent the Pareto refunding scheme applies in the more realistic case of elastic demand. It is challenging to develop combined congestion pricing and revenue refunding models to create
238
Mathematical and Economic Theory of Road Pricing
the best winwin situation, particularly under various social and economic constraints. Furthermore, instead of the change in travel disutility between individual OD pair used in the bilevel network toll design model, one may measure the impact of congestion pricing on social and spatial equities by the accessibility of different user classes from different zones. Accessibility aggregates the potential access from a zone to opportunities in different zones, each of which is weighted by a deterrence function of travel cost.
7.8
SOURCES AND NOTES
The Pareto refunding scheme for congestion pricing revenue presented in Sections 7.2 and 7.3 is a recent (unpublished) development by the authors. The network toll design models with equity constraints presented in Sections 7.47.6 are drawn from the work of Yang and Zhang (2002). There is substantial and extensive literature 4>out social and spatial equity issues in road pricing, and readers are suggested to further refer to Foster (1975), Small (1983), Hau (1992), Giuliano (1992), Johansson and Mattsson (1995), Litman (1996), Button and Verhoef (1998), Richardson and Bae (1998), and, for most recent discussions, to Viegas (2001), Raux and Souche (2004) and Santos and Rojey (2004). A few research studies have looked at the revenue distribution strategies including Small (1992a,b,c), Parry and Bento (2001), Eliasson (2001) and recently Kalmanje and Kockelman (2004). Readers may refer to Bernstein (1993), Adler and Cetin, (2001), Elliasson (2001) and Yang, Meng and Hau (2004) for theoretical developments of congestion pricing revenue redistribution strategies. The simulated annealing method for solving the equityconstrained network toll design problem can be found in many reference books, such as Aarts and Korst (1989), Dekkers and Aarts (1991) and Romeijin and Smith (1994). Readers may also refer to Friesz et al. (1992), Huang and Bell (1998) and Meng and Yang (2002) for the successful applications of this method in solving the continuous equilibrium network design problems in transportation. Finally, the concept of accessibility is covered in many textbooks, such as Erlander and Stewart (1990).
8 PRICING, CAPACITY CHOICE AND FINANCING
8.1
INTRODUCTION
How should society go about expanding its road systems? Who would decide where to provide more road capacity, and how much more? Where would funds for expansion come from? The recent worldwide tendency toward the introduction of commercially and privately provided public roads proves to be an efficient answer to these questions. In a socalled BuildOperateTransfer (BOT) project, the private sector would build and operate roads at its own expense and in turn should receive the revenue from road toll charge for some years, and then these roads will be transferred to the government. BOT projects use the market criterion of profitability for road development and rely on the voluntary participation of private investors, who hope to benefit financially or otherwise from their participation. This commercial and private provision of transport infrastructure has attracted fastgrowing interest in recent years and is being employed to finance modern road systems. Private provision of roads is driven by a number of factors. A primary motivation is a widespread belief that the private sector is inherently more efficient than the public sector, and therefore builds and operates facilities at less cost than the public sector. Also, the public sector, facing taxpayer resistance, may simply be unable to finance facilities that the private sector would be willing and able to undertake for a profit. In addition, if new road space is provided as an "addon" to an existing network system, and if road users find it worthwhile to patronize this new road and pay charges, and if the charges cover all costs (including congestion and environmental costs), all may gain benefit, and there would be no obvious losers. Even those who do not use these new roads would benefit from reduced congestion on the old ones. Once road provision moves to a market economy, there are many intriguing issues to be addressed. From the viewpoint of private investors, profitability of a project is of great concern because their private firms are put at risk. Specially, the paucity of experience with tollcharged roads and the long payback period after the initial investment make it more difficult to predict future revenues and thus makes the uncertainty and risks of investment in
240
Mathematical and Economic Theory of Road Pricing
roads very high compared to alternative investment options. Therefore, careful project selection and a clear identification and assessment of the risks are important in order to guarantee that willing buyers are prepared to pay sufficiently to induce willing suppliers to provide it. From the viewpoint of society, it is meaningful to assess whether the construction of a road will give a positive welfare increment if it is profitable, compared with the donothing alternative, and vice versa whether any road which adds to welfare will be profitable, and hence can be provided privately. More importantly, under the commercial and private provision of roads, what roles can a government play? What government regulations concerning private toll roads should be imposed? For example, should the road charge level be controlled by the government to prevent private firms from abusing their monopoly over the road and for equity reasons (access to services for everyone)? It is obvious that these issues must be addressed carefully, because transport infrastructure has major strategic, economic, social, financial and environmental effects and because the interests of the private sector are different from those of the public sector. In this chapter we look into the interrelations among pricing, capacity choice, financing and competition in simple and general networks. We begin with a description of traffic equilibrium models with heterogeneous users and alternative performance measures for a given pricing scheme. We elucidate the various possibilities of profitability and social welfare gain of a single toll road in a general network and then move to the analysis of the impacts of user heterogeneity on its profitability and social welfare gain. We investigate whether and under what conditions the classical selffinancing principle holds in general transportation networks. Finally, we examine, both theoretically and numerically, the competition and equilibria based on a situation where there are two or more profitmaximizing private firms that operate some toll roads in a road network.
8.2.
USER EQUILIBRIUM AND PERFORMANCE EVALUATION FOR PRIVATE TOLL ROADS
8.2.1 Multiclass UE Model with Elastic Demand Suppose that the highway BOT project consists of a finite sequence of connected links (segments) and the location of the BOT project and the number of years of operations by the private sector are already determined. Let A denote a subset of these links and ya, ae A denote the capacity of a link if this link is a new link to be added to the network, or the capacity increase if this link is an existing link to be expanded, and furthermore let
Pricing, Capacity Choice and Selffinancing
241
y = (ya,ae A) . Without loss of generality, here we consider new toll road links only and denote the toll to be charged on a new link ae A by ua and u = (ua,ae A) . Like previous chapters, subdivide the demand for travel into a set M classes, corresponding to groups of user with different socioeconomic characteristics (for example, income level). Because a BOT project generally involves longterm investment and costrecovery made in a road network, and thus certainly influences the demand for travel, it is therefore desirable to consider the elasticity of travel demand. Let the demand, 1, the revenue will exceed the capital costs and the road will earn a surplus. We thus have the following Selffinancing theorem. Theorem 8.1. In a general network under Assumptions 8.18.3, the revenue from socially optimal pricing on a road just covers the capital cost for constructing the road. This breakeven result holds for each road individually in the full network and consequently to the network in aggregate, provided that each link is optimally priced and all capacities are optimized. Note that the conditions for the above selffinancing theorem to hold are rather ideal and restrictive. In the first place, it is impracticable to optimally design and construct a complete
258
Mathematical and Economic Theory of Road Pricing
new network. A very frequent case is, after all, that one or more new roads are added to an existing network. Nevertheless, it can be found from its derivation that the selffinancing theorem does throw light on the case when only a subset of new links is introduced to an existing network but all links (new and existing links) are optimally priced. Namely, if eqns (8.24) and (8.25) hold for all links, then the network traffic equilibrium conditions are maintained, but eqn. (8.28) is applied to individual links without direct relevance to the equilibrium conditions. Ffence, the breakeven result does hold for the subset of new links individually for given and fixed capacities of existing links in the network. We thus have the following theorem under this relaxed condition. Theorem 8.2. Suppose one or more new roads are introduced to an existing network and Assumptions 8.28.3 are satisfied, then the revenue from socially optimal pricing on a new road just covers the capital cost for constructing this road, as long as all existing and new links are optimally priced and the capacities of new links are optimally selected. Two points deserve our attention here. First, the breakeven result applies for the new links only individually or in aggregate, the revenue from marginalcost pricing on the rest of the network links is intact and hence the whole network makes a profit from the socially optimal pricing. Second, albeit the selffinancing result holds for individual links in both Theorems 8.1 and 8.2, the resulting optimal toll setting and capacity choice are intertwined because the tolls and capacities are calculated on the basis of link flows that are determined from the global network equilibrium conditions (see Section 3.2 in Chapter 3 for a prior discussion). Theorem 6.1 and Theorem 6.2 are with respect to the situation under a firstbest pricing environment, albeit capacity choice can apply for a subset of new links only. We naturally wonder what happens with financing road capacities when toll roads coexist with free roads in a common network. A general answer to this question proves to be difficult, as it depends on the network topology, the locations of toll roads as well as the OD demand and link performance functions. Nonetheless, some meaningful insight can be obtained from the following two styalized examples.
Figure 8.10 A twoparallelroute problem Example 8.1 (The two parallel route problem) Consider a simple network in Figure 8.10,
Pricing, Capacity Choice and Selffinancing
259
with two competing congested roads (or links) (the standard tworoute problem in the literature). Route (or link) 1 already exists and remains free of toll and route 2 will be built and tolled. Omitting the OD index for a single OD pair, we have the following equilibriumconstrained welfaremaximizing problem with respect to the toll and capacity of the new route 2 (suppose both routes are used at optimum, and unit capacity construction cost for route 2 is constant). d
max
[5(co)dco/1(v1)v1f2(v2,72)v2TiK2y2
(8.32)
subject to
B{d) = tl{vl)
(8.33)
B{d) = t1{y2,y2) + u2
(8.34)
where d = v, + v2 and constraints (8.33) and (8.34) characterize the UE conditions. Consider the following Lagrangian: v,+v 2
L{Vl,v2,u2,y2,XvX2)
=
The firstorder conditions are: ^ = B(d)ti(vl)v/l(vl)+X/1(v1)(Xl+X2)B'(d)=0 ^
(8.36) 0
tl{Vi)B(d)
0
(8.37) (8.38)
^
(8.39)
^L = X2=0 au2
(8.40)
^^M^iL^O
(8.41)
where B'() = dB()/d(),
0. This equation shows that the
optimal toll should be equal to the marginal external congestion costs on the new tolled road minus a (positive) term consisting of a fraction (between 1 and 0) of the marginal external congestion costs on the untolled route. This result reflects the fact that one should also take into account the flow "spillover" effect on the untolled route in order to maximize the social welfare. If the link travel time function t2 (v2 ,y2) is homogeneous of degree zero, then eqn. (8.28) holds for link 2. In addition, from eqns. (8.42) and (8.41), we arrive at u2v2 =r]K2y2+a.v/l(vx)v2 •{3
Figure 8.11 A twoserialroute problem The equilibriumconstrained welfaremaximizing problem with respect to the toll and capacity of the new route 2 (suppose again both routes are used at optimum, and unit capacity construction cost for route 2 is constant) is given as: max j £^2(co)dco+ j B^3(co)dco/1(v1)v, t2(v2,y2)v2r\K2y2 Ul yi
'
o
(8.45)
o
subject to S .  * ( ^ 2 ) = ',(v,)
(846)
5 M K , 3 ) = ' 1 (v,) + ' 2 ( w 2 ) + " 2
(8.47)
where di_t2 + rf,^3 = v, and dM = v2 and constraints (8.46) and (8.47) characterize the demandsupply equilibrium.
Pricing, Capacity Choice and Selffinancing
261
Proceeding along lines similar to those which led to eqn. (8.42) yields U2=vA{v2,y2)
+ a.v,t[(vx)
(8.48)
V1
(8.49)
where
I
1
,
\
and 0 < a < 1 . 0
This equation says, simply enough, that users who pay toll are charged in excess of their marginal external cost on the new tolled links. The extra toll charge (the second term) is a fraction of the marginal external congestion cost on the untolled link 1 and the fraction depends on the demand elasticity for the untolled users from node 1 to node 2. This is intuitively obvious because the external costs of users from 1 to 3 on link 1 are partly internalized through the charge on link 2. If we further assume that users make trips from node 2 to node 3 as well, we also obtain the same toll expression given by (8.48), but a is given by and 0 < a < 1 . 0
(8.50)
An immediate consequence of the charge on the new toll road 2 greater than the external congestion cost is «2V2=TlK23'2+av10,reRw,weW
where v = V
V
(8.54)
f 5 , ae A and d = \
f , we
Now we consider the following Lagrangian in terms of path flow andLagrange multipliers:
1
f ' «S.
where \rw,re
}
OEA\ \^W /€«„

]
(8.55)
aeA
Rw,we W are the Lagrange multipliers associated with constraints (8.53). By
our assumption that fm > 0 for any r e Rw, w eW, the firstorder optimality conditions for the above minimization problem become
(8.57) dL
B
=0
(8.58)
Substituting eqn. (8.58) into eqn. (8.56) yields the following expression for the Lagrange multiplier
Substituting eqn. (8.59) into eqn. (8.57) yields, after rearrangement, the following expression of the link toll u, =v/,+ V a v i
(8.60)
where ,aeA,
a *6(8.61)
Pricing, Capacity Choice and Selffinancing
263
In view of 8ar = 0 or 1, we have 5 5,,.. asA
;: 0 < _vy < °°j . In this case, Uj x Y} is a closed and convex subset and each player's strategy is independent of the rival players' strategies. Player j's problem is to determine, for each fixed (u_y,y_;) of other player's strategies, an optimal strategy u". € Up y'. e Yj that solves the profitmaximizing problem (8.77), a set of u* = (u'j, je j \ and y* = (y', je j \ of strategies is said to be in equilibrium if no player is able to enhance its profit by unilaterally modifying its chosen strategy. That is, rc.^y^n.^.y^y:.), jeJ
(8.81)
Let us assume that each one of the profit functions ny. is convex on the set U} x Y} when viewed as a function of \upy^
alone and the other components are fixed. Using the
optimality conditions for convex optimization, we see strategy set u* =(u*,js j \ and y* =(>>*•,./€ 7J is in equilibrium if and only iffor any (upy^e.UjY.Yj
and every jeJ :
Adding these conditions, we conclude that u* =(u'j,je j) and y* =(j"pje
j) must be a
solution of the following VI: (uu*) T 7i u (u*,y') + ( y  y * ) \ ( u * , y ' ) > 0 for any u e U= YlUp y€ Y = UYj
where 7tu(u,y) =(dnj(u,y)/duj, j€ j) and ny(u,y) = (dnJ(u,y)/dyj,jej)
.
As shown in Chapter 2, for given (u,y) there will be a unique network equilibrium flow pattern, a vector of all OD demands, d(u,y) = [dw (u,y), we W) and a vector of all link flows, v(u,y) = (v o (u,y), a e A) . The unique equilibrium solution can be formulated as the VI of finding v and d such that t ( v , u , y ) T ( v  v )  B ( d ) T ( d  d ) > 0 forall (v,d)eQ where, as defined before, Q = {(v,d)v=Af, d=Af, f > O,d>OJ, B(d)=(Bw(dw), and t(v,u,y) = (ta(ya,ua,ya),a e A; ta(va),ae A,ai A^j .
(8.82) weW)T
270
Mathematical and Economic Theory of Road Pricing
To sum up, we thus arrive at the following BVI formulation for the toll road competition problem that belongs to the class of EPEC: Find u=(uj,je
j)eU = Yl^JUJ and y* = (yp je j)e Y = UjsJYJ such that
( u  u ' ) T n u ( u \ y \ v , d ) + ( y  y * ) T 7 i y ( u \ y \ v , d ) > 0 forany ueU, yeY
(8.83)
where for each u € U and yeY, I v, d) 6 O is the unique solution of the VI defined by t ( v , u , y ) T ( v  v )  B ( d ) T ( d  d ) > 0 forall (v,d)eQ
(8.84)
8.6.3 Solution Methods Although the EPEC problem is a sensible mathematical model with a welldefined solution concept, its complex bilevel VI structure makes it a computationally intractable problem. Any rigorous attempt to compute an EPEC solution (if it exists) is presently out of the reach of existing methods. Here we suggest two heuristic methods for the computation of EPEC solutions. Diagonalization Method. Observe that the problem of a single dominant firm is an MPEC (or a Stackelberg leaderfollower game) that chooses a simple pair of strategies {uj,y^e UjXYj,jeJ
subject to the UE constraint
^ w ^ ^ . u . j . y . ^ v ^ M ^ ^ . u ^ . y . ^ M ^  t i / ^ ^ ) , jeJ
where vj{uj,yj,u_J,y_j)
(8.85)
is obtained by solving VI (8.84). This is relatively a simple MPEC
with two upperlevel decision variables only, to which a host of computational methods is applicable such as the sensitivity analysis based or the gap function based methods developed previously in Chapters 4 and 5. One may as well employ a direct search method such as the HookeJeeves method at the expense of iteratively solving the lowerlevel UE problem. The diagonalization method thus solves the individual MPECs of the firms separately and sequentially holds the decision variables of the other firms (or players) fixed in turn until the sequence converges. Synchronous Iterative Method. The firstorder conditions of each firms' MPEC is given by eqns. (8.78) and (8.79), which is in fact amenable to a decoupled simple solution adjustment. For a given current solution (u'"',y'"M at iteration n, one can find the corresponding link flow solution v'"' and the necessary derivative information dvf1 jdiij and 3v'"'/3y; for each
Pricing, Capacity Choice and Selffinancing jeJ
271
by conducting a UE traffic assignment and its sensitivity analysis (see Chapter 3).
Then, by virtue of eqn. (8.78), one can find an auxiliary toll «"]"' as given below: (8.86) By substituting «jn) into eqn. (8.79) and solving the resulting equation, one can obtain an auxiliary capacity y^ as well, if dljiy^/dyj
is not a constant, but includes yp such as the
case of nonlinear construction cost function: I. (y\ = K;. (yjf' (Y7 * l ) ; 7e J I n the case of constant return to scale construction cost function, / ; (y ; ) = K y j y , 76 J, then we have dnj(yi,y) jdyj = Uj dvj (u,y)ldyj tlK^, which becomes a (positive or negative) constant and thus not solvable in terms of y}. In this case, the auxiliary yf1 is instead given by: 0, if
y)' =
A Uj
'y)i\Kj>0
r\Kj =0 ,
jeJ
(8.87)
where ^™ax is the predetermined upper bound of capacity for toll road je J. After obtaining the auxiliary link toll and link capacity, the solution (u'"+1 ,y
I at iteration n + 1 can be
updated by the Method of Successive Average (MSA) as follows: (8.88)
Consequently, the method simply iterates between the synchronous tollcapacity adjustment for all firms and the equilibrium traffic assignment together with its sensitivity analysis, and this process is repeated until an acceptable convergence is achieved. We conclude this section by pointing out that the study of EPEC and/or BVI is still in its infancy, the computation of global solutions even to MPECs remains elusive, if not impossible, and hence efficient convergent methods for computing equilibrium solutions to the general EPEC are to date not available. Though we set forth two appealing computational methods for the specific EPEC of toll competition, the issue of convergence of the heuristic methods is not easily resolved.
272
Mathematical and Economic Theory of Road Pricing
8.6.4 Two Analytical Examples Example 8.3 (Toll competition between two substitutable roads) We consider a simple network of two parallel links shown in Figure 8.10 and used in Example 8.1 to illustrate the synchronous iterative method. For simplicity, we suppose both firms already make their longterm choices of capacities and face shortterm toll competition only. We further assume that both links are used (v, > 0, v2 > 0) throughout the iterative adjustment of tolls by the two firms. Thus, the following equilibrium conditions are always satisfied *(v, + v2) = f1(vI) + K1
(8.89)
5(v,+v 2 ) = /2(v2) + «2
(8.90)
Differentiating both sides of each equation with respect to toll and capacity respectively yields: B'\
3v. du]
3v, ^1 , 3v, 3M, I 3M,
du2
du2 I
2
du2
,
„ / 3v. I du2 '
3v, ^ I
, 3v,
3M2
3M2
I 3M,
Solving the above equations for dvjdu, and dv2/du2 yields: 3vL_t^B' 3M, (t'l + t'2)B't'/2'
3v, t\B' 3M2 \t{ + t'2)B't[t'2
Substituting these derivatives into (8.86) leads to the following auxiliary solution:
B't2 Now let B = (p 
2 2 ' In
view
5
of
i>3 = tPi_>3 — tpi^3(V2)' h=a\
+a v
i i» h = 3). The following form of the OD demand function is adopted: n\>
m orn
2 2
where K is a proportionality parameter with value l.OxlO6 (HK$/(h • veh/h)). The link travel time function (8.11) is used, with capa =ya for ae {«,,«,,z^,/^}. The basic input data of link travel time
function
are presented
in Table
8.4. Other parameters are:
5
TI = 1.14xl(T (l//i), (3 = 100 (HK$/h) (average VOT to transfer toll in HK$ into equivalent time in h).
(a) The base network c
  — I.
— 
(b) The network with two substitutable toll roads, mi and
(c) The network with two complementary toll road, m2 and n2 Figure 8.14 The base network and the proposed private roll roads
276
Mathematical and Economic Theory of Road Pricing
Table 8.4 Input data for the link travel time function Link No.
Freeflow Time (h)
Capacity (veh/h)
a
0.4 0.6 1.0 0.8 0.4
3000 3000 4000 to be determined ditto
0.3 0.4
ditto
b c m \ n \ m1 n2
ditto
Two representative games. In view of the fact that road provision is inflexible in the short run (capacity cannot be changed frequently), while tolls are flexible, we will focus on the following two possible competitive games. Game A. This game is defined as a oneshot game where all decisions are made at the same time by both firms. Namely, two firms choose their capacities and tolls simultaneously when two new private toll roads are introduced at the same time. Game B. In this game we suppose both firms have chosen their capacity and both capacity supplies are known to each other, the two firms simultaneously choose their toll levels (toll adjustment is flexible in the short run). For each of the games under the aforementioned demand conditions (demand substitutes or complements), we determine and compare the solutions of the mixed supply and price competitions, including the total social welfare gain and the private firms' profits. Representative results. For each of the networks, the results of both Game A and Game B are presented here. Table 8.5 shows the results associated with Game A, as well as the social optimum solution (welfare maximization) for both substitutable and complementary cases. Note that the social optimum solution can not be implemented because at least one firm receives negative profit. We note that for Game A, both firms are free to select their capacity and toll level in order to maximize their own profits simultaneously, and we find that a Nash equilibrium can always be achieved by the end of the game. We also find that there may be more than one equilibrium solution, depending on the initial combination of the decision variables. The results in Table 8.5 are just some that we believe to be practically meaningful.
Pricing, Capacity Choice and Selffinancing
277
Table 8.5 Representative results Substitutable Network (b)
_ . Firm m Capacity Firm n Toll
Fiml m
Firm n Firm m
rroiit
Firm n Welfare Gain
Competitive Solution 6401.00 5000.10 17.00 17.90 18907.85 39463.99 229904.30
Social Optimum Solution 6212.50 5509.00 9.91 10.08 12220.16 14628.97 291813.20
Complementary Network (c) Competitive Solution
Social Optimum Solution
3798.10 7301.00 9.20 18.70 12908.96 62711.04 296352.00
9388.16 11947.57 2.360 8.890 16198.34 21623.16 302840.70
Graphical interpretation As the road capacity is more likely to be fixed, and toll is flexible to change in the short term, here we are more interested in examining the competitive characteristics of the toll roads with respect to toll charges only. This is the case of Game B, where the capacities are preselected and fixed, and we find there is only one toll charge solution for each case. By varying the combinations of the tolls by the two firms, the variations of social welfare and profit can be represented by the isowelfare and isoprofit contours in Figure 8.15 to 8.19. In scenario 1, for the network in Figure 8.14(b), the two firms compete with each other to attract traffic flow to their road. When the toll is set too low, even if a large amount of flow is attracted to the toll roads, the revenue received is not enough to cover the cost. On the other hand, if the toll is set too high, no one will use that road which will result in negative profit again. As shown in Figure 8.15, the profit increases from negative to its maximum, and then falls as the toll further increases. It is also interesting to notice that there is no contour at the right side for firm wz,, as does firm n, at the top of the graph, because the toll charge of either firm is too high. Two reaction lines of the firms can also be plotted on the graph. They represent the optimal selection of toll by one firm once the other firm has already selected its toll. Starting from any point on the graph, the firms will select their new tolls on these lines and finally they will converge to the intersection of these lines and reach the Nash equilibrium there. In Figure 8.16 for the substitutable case, we can identify the corresponding social welfare gain in each combination of toll charges by the two competitive firms. As shown in the figure, the point of Nash Equilibrium has a positive welfare gain in this example, and the point of maximum social welfare gain also occurs inside the region of positive profits. Both toll roads
278
Mathematical and Economic Theory of Road Pricing
have benefit for the whole society and the private firms. However, this is not always true; there are also cases where only one firm enjoys positive profit. Figure 8.17 depicts the profits for the complementary case, with the network in Figure 8.14(c). This graph is more or less the same as that in Figure 8.15 except that the maximum profit of one firm occurs at where the other firm has zero toll. This result stems from the fact that in this case the two toll roads are complementary. Figure 8.18 represents the corresponding social welfare gain for different toll combinations for the complementary case. Similar to Figure 8.16, we can find that both Nash Equilibrium and maximum social welfare gain occur in the positive profit and welfare region. Thus both toll roads benefit the private firms and the whole society. In Figure 8.19, we focus on the three curves for the substitutable case (the graph for the complementary case is similar), the zero profit curves and the zero welfare gain curve. These curves divide the space into a number of regimes that have different economic meanings. Regime I is bounded by the two zero profit curves within the positive welfare gain region. Thus in this regime, both firms enjoy positive profits and introduction of the two competitive roads is socially desirable. In Regime II, one firm is profitable and the other is unprofitable. Firm n is unprofitable within Regime Ilb and Regime Ild, while firm m is unprofitable within Regime Ila and Regime IIc. In playing the competitive game, the unprofitable firm will always try to alter its toll charge and move toward profitable Regime I. Note that neither firm will choose regime III since both of them are unprofitable, unless being subsidized by the government. Finally, Regime IV is not preferred to both private firms and the whole society because it generates negative social welfare gain and negative profits.
Profit for Firm n Profit for Firm m
15
20 25 30 Toll Charged by Firm n (dollar)
35
Figure 8.15 Profits for the substitutable case with capacity fixed
Pricing, Capacity Choice and Selffinancing
Social Welfare Curve
Figure 8.16 Social welfare gain for the substitutable case with capacity fixed
Profit for Firm n Profit for Firm m
10
15
20 25 30 Toll Charged by Firm n (dollar)
35
Figure 8.17 Profits for the complementary case with capacity fixed
279
Mathematical and Economic Theory of Road Pricing
280
50  f
10
15
20
25
30
35
40
45
50
Toll Charged by Firm n (dollar)
Figure 8.18 Social welfare gain for the complementary case with capacity fixed
^^—^^— —
 —
Zero WE I tare Cain 
M.\l.im:i: Profll far Firm I
Zero Profit for Firm u Ztro FroHr for Firm m
IV . • ; : . . .  I V
3 S R,
..;•
: j I
ii
. ; •
Regime Uc
•
Illc I /
Regime [ l i b 0 .
I
Wdtaim
WAR fl"
Nuh Equilibrium I
I II.::li:.;l.K x,
^
0,dw>0,reRw,weW
(9.5)
The notation is as defined as before, particularly, A cz A is a subset of candidate subject to toll charge;/a(va,wa) = ta(ya) + ua/fi, ifae^4;and ta(va), ifaeA,
a £ A. Link travel time
ta (va) is a strictly increasing and continuous function of its link flow va, and benefit or inverse demand function Bw(dw) is a strictly decreasing and continuous function of OD demand dw. U is the feasible set of link tolls, which could be defined by
U = {u\ur
1 0 0 1 0 0 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 0 0 _] 0 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
Mathematical and Economic Theory of Road Pricing
294
The rank of this matrix is Rank = 6, then we get the number of components N
Figure 9.6 The cut in Example 9.3
t
= 3.
Figure 9.7 The cut in Example 9.4
Example 9.4 Set Nt = {n6, «8} and N2 ={«,, n2, «,, n4, rc5, «7, «,} as shown in Figure 9.7. The cut is given by the link set {e7, et, eu, en, eu, el4, e15, el6, e,,, e^,} . Intuitively, there appear three components after removal of these links from G, i.e. two disjointed shaded areas and one plain area, hence we say that the node set Nt is multicentered. In another way, since the number of components after cutting is increased by two, the cut is not a cutset. The incidence matrix of the graph after cutting is given by
e5 e6
3o e17
0 0 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0 —1 0 1 1 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1
•1
A = «5
It is easy to calculate that the rank of this matrix is Rank = 6 and then the number of components is 7Vcompl = 3 . With the mathematical definition of cut and cutset in graph theory, we now introduce rigorous definition, classification and interpretation of various cordons for areabased congestion pricing with different geographical features. 9.4.2
Types of Cordon
Simultaneous Determination of Toll Levels and Locations
295
Singlelayered cordon. When an areabased pricing scheme is considered for a congested area such as the central business district, the subject area is cordoned off by a simple imaginary closedloop; the area enclosed within this loop is defined as the cordon area subject to toll charge. The closedloop C, and C2 in Figure 9.8 are examples of such simplest complete cordons. Note that in the same figure, the screen line S divides the study area into two sections and can be considered as a special cordon as well. Differing from the closedloop cordon that divides the whole area into an inner and an outer subarea, screen line S crosses the boundary of the study area. Both types of cordon may be adopted for areabased congestion pricing.
Figure 9.8 Examples of singlelayered cordons and screenlines It is straightforward to see that the set of links crossing the cordon is a cutest, as defined before. Since the cordon divides the whole network into two parts, if we take the nodes in one part as set TV,, and the other nodes as N2, the links crossing the cordon are exactly the set of links with one endpoint in JV, and the other in N2, which is a cutset by definition. Note that the singlelayered cordon divides the whole network into two components; hence the rank of the graph will decrease from (7VJl) to (/V"2), where \N\ is the number of nodes. This was illustrated in Example 9.2. Multilayered cordons. Suppose there is one large city center, traffic becomes more congested on moving to the innermost part. In this case, multilayered cordons may be more preferable in order to mitigate traffic congestion to a desirable level. Similarly, the multilayered cordons are also a disjointed combination of multiple closedloops and a noncutset
296
Mathematical and Economic Theory of Road Pricing
cut. In this case, one cordon contains the other and users have to pay money once crossing a closedloop of the cordon. In general, for a mlayered cordon, totally there are (m + \) components generated after the cutting. Again, the rank of the graph will decrease from (Afl) to (7V»!  l ) after removal of the links crossing the cordons. This was illustrated in Example 9.3. Multicentered cordons. A large city may contain more than one separate business centers, and an areabased toll charge may be desirable for each center to mitigate traffic congestion. Since the multicentered cordons are a disjointed combination of individual simple closedloops, it is a cut but not a cutset by definition. If there are m centers enclosed with closedloops, then there are (m + l) components generated after cutting, namely, the m city centers and the remainder. In other words, if the links crossing the cordons are removed, the rank of the graph will decrease from (7Vj 1) to (^V  m 1). This was illustrated in Example 9.4. There might be more complex cordons consisting of a combination of the various aforementioned forms of cordons. Their properties can be discussed in a similar manner and are omitted here. 9.4.3
Determination of Tolling Cordons
Determination of optimal tolling cordons is one kind of location problem in networks. It is well known that solving the network location problem by traditional discrete optimization program is very hard. Here, we again apply the GA method used in previous sections, in which the populations refer to the feasible cordons and their performances are evaluated by the maximal social welfare that can be achieved with optimal toll charges on them. The general idea of using GA for selection of tolling cordons is given below. General procedure for selection of tolling cordon (i) Determine the type of cordon by examining the characteristics of the city. (ii) Generate a set of feasible cordons of the selected cordon type randomly, and evaluate their performances with the bilevel model (9.1)(9.6). (iii) Discard the cordons with poor performance, and let those with better performance survive, (iv) Pair and cross the surviving cordons, then give birth to new cordons of the next generation. Meanwhile, the mutation will generate other new cordons.
Simultaneous Determination of Toll Levels and Locations
297
In step (ii) the bilevel model (9.1)(9.6) can be simply applied to find the optimal tolls on predetermined tolling cordons. The single layeredcordon has one toll level and the multilayered or multicentered cordons have several toll levels (one toll level for each closed loop). All links crossing the same closed cordon loop are charged with the same level of toll. Links not crossing any cordon loops are free of charge. As discussed before, there are a couple of methods available to solve the bilevel toll optimization program, such as simulated annealing and grid search algorithm. Since in our problem there are very few variables (levels of cordon toll charge) in the upper level, the grid search method is employed for search of a global optimum. In this case, a traffic assignment has to be executed for each trial toll charge. The key problem in the above procedure is how to represent the cordons by chromosomes and how to generate feasible cordons. This problem is discussed in detail for different types of cordons below. Singlelayered cordon The singlelayered cordon divides the whole network into two subareas, namely the tolling area and the nontolling area. Users have to pay when crossing the cordon from nontolling area to the tolling area. Let us term the nodes in tolling area tolling nodes and those in nontolling area nontolling nodes. Of course, the area inside the cordon is the tolling area and the area outside the cordon is the nontolling area.
Figure 9.9 An example network used for illustration of selection of a singlelayered cordon and parameter representation of genetic algorithm Being the most fundamental and commonly used, the singlelayered cordon is discussed in more detail, as it is also helpful to the understanding of other forms of cordons. Suppose a
Mathematical and Economic Theory of Road Pricing
298
road network is given in Figure 9.9 and a singlelayered tolling cordon is to be set around the CBD area. Instead of determining the optimal tolling links directly, we determine the optimal tolling nodes (tolling area), and thus the optimal tolling cordon is determined subsequently. Around the most congested CBD area, some nodes are predetermined as candidate tolling nodes, for example, in Figure 9.9 the nodes within the closedloop C\ are set to be candidates. Note that the tolling area containing the tolling nodes should be large enough to include the future optimal tolling cordon; otherwise an optimal tolling cordon may not be found. Then the GA is employed to search for the optimal tolling nodes (tolling cordon). The procedure of searching for a feasible cordon is as follows. From the candidate set, some nodes are randomly picked up as a subset of tolling nodes, and then a cut is determined automatically, having the links with one endpoint in the tolling area and the other in the nontolling area. Then the rank of the incidence matrix, after removal of the cut, is calculated. If the rank is equal to (1^1  2), where iV is, as mentioned before, the number of nodes in the network, this cut is accepted as a cutset or a feasible cordon; otherwise, it is rejected and another set of nodes is randomly picked up until a new feasible cordon is found. For each given cordon (tolling links), the bilevel model (9.1)(9.6) is applied to obtain the maximal social welfare and the corresponding optimum uniform toll. Through the process of natural selection in the GA, the cordon with best performance will be chosen in the end. The GA parameter representation here is somewhat different from those in other applications. The parameter (gene) denotes the state of each candidate nodes, for each candidate node « e N, where N is the set of preselected candidate nodes, it has only two alternatives, either to be charged or not. We define that fl, if candidate node n is selected as a toll node (9.15) O, otherwise The candidate nodes h with g^ = 1 constitute the set of tolling nods N'. Therefore, we can encode the parameters (genes) with the simplest binary variable, and 1 bit is enough for each gene, i.e. 01 integers. Example 9.5 Suppose that a singlelayered cordon is to be installed in the network of Figure 9.9. In this network, every undirected line represents two opposite directed links. Parameter representation in a binary GA is given below. (a)
1
2
3
4
5
6
7
8
Q
10 11 12
(b)
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
(c)
0
1
1
1
1
1
0
0
0
0
0
0
0
0
0
13 14 15
Simultaneous Determination of Toll Levels and Locations
299
The 15 nodes within the closedloop C\ is set as candidate tolling nodes, whose numbers are given by the above row (a). Row (b) represents a chromosome whose length is 15. Hence the tolling node set represented by this chromosome is N* — {1,2,3,4,5,6}, which is given by the nodes within the closedloop C2 in Figure 9.9. The closedloop C2 is a cut, which is also a cutset and hence it is accepted as a feasible cordon. Row (c) represents another chromosome, where the tolling node set is N* ={2,3,4,5,6}, which is given by the nodes between the closedloops C2 and C3 in Figure 9.9. Since the cut given by the union of C2 and C3 is not a cutset and hence not a feasible singlelayered cordon, and therefore it should be rejected. Actually, it is a twolayered cordon to be examined in next subsection. The procedure of GA used for determining optimal singlelayered cordon is stated as follows: Step 1. (Initial population) Randomly generate the initial population of feasible singlelayered tolling cordons, according to the abovementioned method. Step 2. (Function evaluation) Apply the bilevel model to obtain maximum social welfare for each given cordon. Step 3. (Natural selection) Select those cordons with higher social welfare as survivors and discard the rest. Step 4. (Crossover) Conduct pairing among survivors and exchange tolling nodes between pairs. Step 5. (Mutation) Randomly modify the parameters of some nodes. Step 6. (Next generation) For each new set of tolling nodes, determine the corresponding cut automatically. Maintain those cuts that are cutset (i.e. feasible cordon), and discard the rest. Randomly generate new cordons to complement the population. Step 8. (Verification of stopping criterion) If the stopping criterion has not been reached, go to Step 2; otherwise stop. If the cordon to be determined is a screenline, then the candidate nodes along the predetermined boundary are preselected, and the above approach can be modified slightly to determine the optimal screenline. Multilayered cordons For multilayered cordons, the genetic algorithm can be used in a similar manner; the difference lies in the generation of feasible cordons. Here, for simplicity, we use closedloops to represent the candidate tolling and nontolling areas without adding specific nodelink structure of the network.
300
Mathematical and Economic Theory of Road Pricing
\ \
Figure 9.10 Generation of multilayered cordons In Figure 9.10, there is only one city center, but this center is to be cordoned in two layers. Hence there will be three components after cutting, namely the inner tolling area, the outer tolling area and the nontolling area. The nodes in the region between curve a and curve b are set to be candidate tolling nodes between the outer and inner cordons, furthermore to keep the cordon to be layered, a set of nodes in the ring between loop c and loop d are preselected as compulsory tolling nodes. In this way, the tolling nodes between the outer and inner cordons are randomly generated in the candidate set, to construct feasible tolling cordons subject to the rank of the graph after cutting equal to (N  3). In this way, the outer feasible cordon will be generated between loop a and loop c, such as loop e; and the inner feasible cordon will be generated between loop b and loop d, such as loop / . Then the GA is used to search for the optimal toll cordons. Multicentered cordons As represented in Figure 9.11, the city has two centers, eastern and western, and one may consider areabased pricing with two disjoint cordons. The method is similar to the case of the singlelayered cordon, where two sets of candidate nodes are predetermined, which are demarcated by loops a and b . Within the two loops, two disjoint single cordons are selected, such as loop c and loop d. The multicentered cordons are a cut and can be determined by picking out two sets of nodes from the two candidate sets as tolling nodes. Since two disjointed sets of nodes are picked out, there should be at least three components after cutting, and the rank of the network must be less than or equal to (./V  3). If the rank is exactly
Simultaneous Determination of Toll Levels and Locations
301
  3), then the nodes in each candidate set are connected, implying that there is only one cordon around each center. In this case, one can claim that feasible twocentered cordons are found. Thus we can generate feasible cordons by randomly picking out the tolling nodes and then examining the resulting rank of the incidence matrix. In this manner, the GA can be readily used to search for the optimal multicentered cordons.
Figure 9.11 Generation of multicentered cordons Finally, we point out that in reality there might be various complex cordons consisting of a combination of the aforementioned three types of cordon, and the selection of optimal tolling nodes or tolling areas can be conducted in a similar mariner by presetting appropriate candidate tolling nodes and applying the genetic algorithm and the cut/cutset concepts in graph theory.
9.5
T H E CORDONBASED PRICING: EXAMPLE
Figure 9.12 depicts the arterial road network of Shanghai metropolis, the largest city in China. The network consists of a total of 690 directional links (each line segment in the figure represents two opposite directed links) and 187 nodes, out of which there are 30 origin nodes and 30 destination nodes. Homogeneous users are considered with a mean time value of 60 (FIK$/h). We consider the traffic condition of the morning peak hour during which most trips are generated in the suburban area and attracted to the CBD. The link travel time function (9.13)is used for all road links in the network. Elastic demand function is used here to describe the reaction of travelers to the travel disutility (inclusion of time and monetary costs), and the demand function takes the following form for each OD pair:
pfl.O^j , weW
(9.16)
here dw is the realized OD demand, Dw is the potential OD demand, \xw is the present OD travel disutility, and \xaw is the freeflow OD travel disutility between OD pair weW. The
302
Mathematical and Economic Theory of Road Pricing
demand elasticity with respect to the OD travel disutility is given by p\xw/\x°w where p is regarded as a dimensionless demand elasticity parameter.
Figure 9.12 The optimal location (closedloop B) of a singlelayered cordon
Figure 9.13 The optimal location (outer loop B and inner loop D) of a doublelayered cordon In this example, the elasticity parameter p is taken to be 0.25. The original social welfare in the absence of toll charge is 1.619xlO7 (HK$), and the maximal social welfare achieved under the firstbest pricing is 1.672xlO7 (HK$), which implies a maximum percentage welfare gain of 3.27%. To mitigate network traffic congestion and air pollution problems during the morning peak hour, the cordonbased congestion pricing around the CBD area is considered. In the following, we investigate the various cordonbased pricing strategies for the network.
Simultaneous Determination of Toll Levels and Locations
9.5.1
303
Determination of Location of a Singlelayered Tolling Cordon
The genetic algorithm described in previous section is used to determine the optimal single cordon location. In Figure 9.12, the nodes within loop A are set to be candidate tolling nodes. By executing the GA, loop B is eventually selected to be the optimal tolling cordon. On this cordon the optimal toll level is 10.6 (HK$), and the corresponding maximum social welfare is 1.634xlO7 (HK$), which represents a welfare increase by 0.93% compared with the notolling equilibrium case, and a 2.27% gap from the maximum level of social welfare under the firstbest pricing. Table 9.4 Summary of the results for the single and doublelayered cordon pricing Type of cordon Singlelayered cordon Doublelayered cordons
9.5.2
Cordon toll (HK$)
Percentage social welfare gain over the nontolling equilibrium
Percentage welfare gap from the firstbest social optimum
Percentage welfare gain as a share of the max welfare gain
10.60
0.93%
2.27%
28.30%
Inner: 4.90 Quter: 8.50
1.35%
1.85%
41.51%
Determination of Location of Doublelayered Tolling Cordons
The GA is again used to determine the optimal location of doublelayered cordons. In Figure 9.13, the nodes inbetween curve A and curve E are set to be candidate tolling nodes between the outer and inner cordons. The nodes on the closedloop C (solid dark line) are preselected tolling nodes to keep the cordons being doublelayered (this corresponds to the nodes inbetween loops c and d in Figure 9.10). After the process of natural selection, eventually loop B and loop D are found to be the optimal outer and inner cordons. The optimal toll levels on the outer and inner cordons are found to be 8.50 and 4.90 (HK$) respectively, and the maximal social welfare achieved is 1.641xlO7 (HK$). This represents a 1.35% welfare increase over the nontolling case and a 1.85% welfare gap from the firstbest social optimum. Note that the reported percentage welfare gain is very marginal, which is largely due to the relatively high total consumer surplus realized (the type of demand function and the parameter values used also make a difference). Therefore, we now introduce the following more meaningful efficiency index to compare the performance of different pricing schemes: „
SS°
s's°
(9.17)
Mathematical and Economic Theory of Road Pricing
304
where S represents the social welfare with a given cordon pricing scheme, 5° is the original social welfare without pricing, and S' is, as used before, the maximal social welfare under the firstbest pricing scheme. Thus, this efficiency index measures the welfare gain as a share of the maximum gain under the firstbest pricing. From the above numerical results with parameter p =0.25, the efficiency index 0 = 0.2830 (or 28.30%) in singlelayered cordon pricing, and 9 = 0.4151 (41.51%) in doublelayered cordon pricing. Table 9.4 summarizes the above results for both the singlelayered and doublelayered cordon pricing cases. 4.0
• Margical cost pricing • Double cordon pricing • Single cordon pricing Without toll pricing
05
510 1015 1520 2025 2530 3035 3540 4045 4550 5055 5560 Trip Length (min)
Figure 9.14 The impact of alternative toll pricing schemes on trip length distribution 9.5.3 Impact of the Cordonbased Pricing on Trip Length Distribution Here we further investigate the impact of alternative pricing strategies on trips of different lengths. Figure 9.14 displays the trip length distribution for the following four cases: the marginalcost or firstbest pricing, the single and doublelayered cordon pricing and the no toll pricing. The length of a trip is measured with respect to the shortest freeflow travel time from its origin to destination, which is proportional to the actual physical distance. The discrepancy among the four trip length distributions reflect the impacts of the different pricing schemes on the users traveling between different OD pairs with different trip lengths. As observed in Figure 9.14, for trip lengths less than 10 (min), the demand after introducing cordonpricing increases, compared with the no tolling case, and is higher than that in the marginal cost pricing case. This is due to the simple fact that the users with short distance trips do not cross over the tolling cordon and thus avoid toll charges, but benefit from reduced traffic congestion on the network. For trip length between 10 and 35 (min), the demand after introducing cordonpricing decreases compared with the no tolling case, and is lower than those in the marginalcost pricing case. In addition, the demand under single
305
Simultaneous Determination of Toll Levels and Locations
cordon pricing is slightly higher than that under doublecordon pricing. This implies that the cordonbased secondbest pricing mainly causes mediumlength travel demand to drop. For trip lengths greater than 35 (min), the demand after introducing cordonpricing decreases, compared with the no tolling case, but is higher than that in the marginalcost pricing case. This means that the longdistance travel demand, although dropped after introducing cordon pricing, still exceeds the socially optimal demand level. The above numerical results show that trips are underpriced in locations inside the cordon, overpriced just outside the cordon and underpriced on the fringe of the urban area. 14.0
D— Optimal toll on a singlelayered cordon 12.0 
—O— Optima! outer toll on doublelayered cordons 6— Optima] inner toll on doublelayered cordons
4.0 2.0 0.1
0.2
0.3
0.4
0.5 0.6 0.7 Elasticity Parameter
0.8
0.9
1.0
Figure 9.15 Sensitivity of optimal cordon tolls to demand elasticity parameter 1.0
• Double cordon pricing
—d—Single cordon pricing
0,8
0.2
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Elasticity Parameter
Figure 9.16 Sensitivity of social welfare efficiency index to demand elasticity parameter
306
Mathematical and Economic Theory of Road Pricing
9.5.4 Sensitivity Analysis of Demand Elasticity Now we conduct a sensitivity analysis of the elasticity of demand by varying the elasticity parameter p in the demand function. In the following analysis, the cordon locations are fixed, as given in Figure 9.12 and Figure 9.13 for singlelayered and doublelayered pricing cordons. Figure 9.15 displays the change of optimal tolls as parameter p varies from 0.1 to 1.0. Clearly, as the value of p increases, optimal toll levels on all cordons decrease. This means that as travel demand is more elastic or more sensitive to travel cost, a lower toll will be enough to achieve social welfare maximization. Figure 9.16 shows the change of the social welfare efficiency index defined in eqn. (9.17) as parameter p varies from 0.1 to 1.0, for both the singlelayered and doublelayered cordon pricing schemes. Clearly, the efficiency index increases with p value, which means that the cordonbased pricing scheme is more effective when travel demand is more sensitive to travel disutility. Within the practical range of demand elasticity (0.30 ~ 0.40), the welfare efficiency achieved through the cordon based pricing schemes is between 2060%. There is still a large gap from the firstbest social optimum. This result is different from the theoretical observation made in Mun et al. (2003) for a monocentric city in which he found that a single cordon pricing scheme attains an economic welfare level nearly as good as the firstbest optimum. One possible explanation for this discrepancy is given below. The firstbest pricing in a network influences both trip rate and route choice in a socially optimal manner, and the cordon pricing scheme could effectively restrain the demand for travel but has very limited impact on users' route choice in a real network, as examined here. In contrast, the spatial model of traffic congestion for the monocentric city does not entertain route choice behavior and thus may overestimate the actual social welfare gained through cordon pricing in a road network.
9.6
SUMMARY
In this chapter we have examined how to determine the toll levels and tolling locations simultaneously for the linkbased and cordonbased secondbest pricing schemes. In each scheme, optimal tolls can be determined by the methods developed in previous chapters for given charging locations; while the optimal charging locations can be identified effectively by a genetic algorithm. For a linkbased pricing scheme, the optimal toll levels on given tolling links are determined using a simulated annealing algorithm. For the linkbased pricing scheme, the model was slightly altered to determine the minimum number of tolling links required to achieve a system optimum with and without inclusion of implementation costs of toll collection.
Simultaneous Determination of Toll Levels and Locations
307
For a cordonbased pricing scheme, the concepts of cutset and cut in graph theory were used to describe the mathematical properties of singlelayered, multilayered and multicentered tolling cordons. We showed that each type of cordon corresponds to a certain kind of cut or cutset in a network. The nodelink incidence matrix of a graph is applied to examine the rank of graphs, which can identify whether or not a given subset of candidate nodes forms a feasible cordon of a certain type. The genetic algorithm is employed to naturally select optimal tolling cordons by setting appropriate candidate tolling nodes and examining the resulting nodelink incidence matrix. A case study was conducted with the urban arterial road network of Shanghai City. From our numerical experiment results, the cordonbased pricing schemes achieve about 2060% welfare gain as a share of the maximum welfare gain by appropriate selection of (single or double) cordon locations and toll levels. For both the single and doublelayered cordon pricing schemes, there remains a substantial gap in social welfare from the firstbest social optimum level. In addition, the cordonbased pricing generally causes the short and longdistance trips to be underpriced and the mediumdistance trips overpriced, in comparison with their respective firstbest pricing toll levels.
9.7
SOURCES AND NOTES
A few authors considered the linkbased and cordonbased secondbest pricing problem. Hearn and Ramana (1998) proposed an approach to determine the minimal number of tolling links to achieve a system optimum. May and Milne (2000) tested and compared various kinds of road pricing schemes, including cordonbased charging. Shepherd et al. (2001) investigated the sensitivity of optimal toll charges with respect to prespecified cordons of different sizes. Verhoef (2002) examined the linkbased secondbest pricing problem and proposed heuristic algorithms for finding secondbest optimal toll levels and tolling points in networks. Mun et al. (2003) presented a theoretical analysis of cordon pricing in a monocentric city. Shepherd and Sumalee (2004) and Sumalee (2004) applied genetic algorithm for optimal cordon design. Santos (2004) and Santos and Rojey (2004) simulated cordon tolls for eight English towns and assessed and examined the distributional effects and environmental impacts. This chapter is developed mainly based on the works of Yang and Zhang (2002) and Zhang and Yang (2004). The background on graph theory used in this chapter is from Chen (1997). The genetic algorithm and its application are well documented in Haupt and Haupt (1998).
This Page is Intentionally Left Blank
10 SEQUENTIAL PRICING EXPERIMENTS WITH LIMITED INFORMATION
10.1
INTRODUCTION
In previous chapters, we have introduced the significant advance that has been made towards better modeling and understanding of traffic congestion and possible impacts of congestion pricing. There remain, however, a few questions surrounding the actual design and implementation of road pricing schemes, for which policy makers and traffic engineers require answers. A fundamental question is how to choose the optimal charge level of congestion tolls in a simple yet practical manner. Previous analysis and discussion of optimal congestion pricing centered on three fundamental elements, namely, the speedflow relationship, the demand function, and the generalized disutility (and hence the value of travel time savings). Among all these three elements of primary concern in actual implementation of a congestion pricing scheme, and analytical demand functions tailored for congestion pricing are, however, difficult to establish in practice, even with advanced transport modeling techniques. In this chapter, we propose trialanderror implementation methods for the various congestion pricing schemes examined in previous chapters, when particularly the demand functions (and/or value of time) are unknown. Three typical pricing policies are considered here including: the firstbest pricing scheme, the secondbest pricing scheme and the traffic restraint and road pricing scheme. The fundamental idea behind the methods is to adjust toll charges at frequent intervals as traffic patterns changes brought about by price change. Given a trial of a set of link tolls, the revealed aggregate link flows can be observed with ease, and based on the observed link flows, a new set of link tolls can be determined and used for the next trial. One can thus expect to identify the efficient toll rates through an iterative toll adjustment procedure, which would allow for a traffic planner to easily estimate the socially optimal congestion tolls in a network, without resort to the demand functions.
310
Mathematical and Economic Theory of Road Pricing
This chapter is organized as follows. We begin by proposing an iterative procedure for determining the socially optimal firstbest link tolls based on the method of successive averages and give a rigorous theoretical justification for its convergence. Then, we develop a sequential experimental approach for analyzing the secondbest pricing problem and describe the necessary bilevel programming models and algorithms for the subproblems of OriginDestination (OD) demand estimation and link toll selection. We move on to examine the traffic restraint and road pricing scheme and develop an iterative toll adjustment procedure derived from the dual formulation of the traffic equilibrium problem. The procedure updates link toll charges from observed link flows using a simple gradient projection method and its convergence is established under mild conditions. We conclude this chapter by highlighting a few behavioral and political issues in the implementation of the trialanderror procedure.
10.2
IMPLEMENTATION OF THE FIRSTBEST PRICING
SCHEME
The basic logic behind the iterative implementation procedure for the firstbest or the marginalcost pricing problem is given below. "With information on the underlying speedflow relationship on a highway and the generalized cost, the optimal congestion toll occurs if and only if the intended speed level or traffic volume that is used to calculate the theoretical congestion toll is exactly equal to the observed speed level or traffic volume after the imposition of the toll." As a result, the exact knowledge of the demand function is not really necessary for identification of the optimal toll charge. We begin our exploration of the trialanderror procedure by examining the simple standard case for a single highway, and then moving on to investigate the marginalcost pricing problem on general networks. 10.2.1 An Iterative Toll Adjustment Procedure for a Single Road Consider the demandsupply (performance) curve and the underlying theory of marginal cost pricing illustrated graphically in Figure 10.1. As examined in Section 3.1 in Chapter 3, this corresponds to the simplified but, in the literature, standard case of a homogenous traffic stream moving along a given uniform stretch of road. The optimal demand (or flow) is, as we can see, equal to dK , where marginal cost (MC) and demand are equated, while the actual demand in the absence of toll charge tends to be d"° (intersection between the average cost (AC) and the demand curve), because road users ignore the congestion that they impose on
TrialandError Implementation of Road Pricing
311
others. Therefore, the optimal toll to be charged is equal to wopt. Nevertheless, the demand function is generally unknown and hence «°pt cannot be determined analytically. What we know is that the revealed demand or traffic flow d on the road is observable for a given level of toll charge. We are thus able to observe the responses of traffic flow to alternative toll charges, such collectable reaction data information is in fact sufficient to find the optimal toll for at least the simple demandsupply curve on a single highway. MC(d)j \
Bid) (unknown benefit or \ inverse demand function) AC(d)
N
1
>
0
Traffic Demand
Figure 10.1 Demandperformance equilibrium on a single highway The iterative trial and error procedure works as follows. Starting with an initial targeted flow (or a targeted speed according to the underlying speedflow relationship) to be achieved by imposing a corresponding congestion toll. Let the initial target of flow be dw (of course, dm = da° is an acceptable initial choice; dm
should be larger than ds° ) and the
corresponding toll to achieve this level of flow is estimated by um = MC\dm)
AC(dmj
without resort to the demand function. Once um is imposed, the revealed or observed traffic flow becomes dm, which is an aggregate result of the users' willingnesstopay (dictated by the demand or benefit function unknown to the planner) according to their marginal private cost {AC curve) and the amount of toll in place. From Figure 10.1, dm < dm .
312
Mathematical and Economic Theory of Road Pricing
Now consider an appropriate one dimensional search method, the bisection method, to modify the intended traffic flow: dm = {dm+ dm\ll.
Given dm,
the new theoretical
estimate of congestion toll is determined by um = MC[di2))  AC(dm).
Since dm < dm,
we must have w(2) < um. Once this new toll is imposed, there will be a new response from the market, which leads to a new revealed level of flow dl2\ corresponding to the users' willingnesstopay of [AC{dm) + um\
by the same argument used above in deriving
dm.
The above operations can be expressed by the following three equations:
B(dm) = AC(dm)+MC{dm)AC{dm) dm=[dm+dm) B(dm) = AC(dm) +MC{d^)AC{d(1)) where B() = D~x (•) is the inverse of the true but unknown demand function (or benefit function). Assume that AC is a convex, monotonically increasing function and B() is a convex, monotonically decreasing function. Then, the observed flow must be decreasing with respect to the intended flow since MC(d)AC(d) B\d\AC\d\
is monotonically decreasing. So, we can define a decreasing function,
d = F(d), and have dm =F[dm)
and d(2) = F(d(2)). Furthermore, we have +
Note that F^d^KF^d2)
holds due to d]>d2.
After k iterations, a sequence w
is monotonically increasing and
{ '
k )
) {
w
JJ'^.M**',^'*'}
) ,
~m _
Hence, we must have
is generated with the following properties:
k>\ With
In Figure 10.1 there is only
one point, d*°, that satisfies this condition. So, we have lim d(k) = lim d(k) = ds° and lim u{k) = Mopt =
MC(dso)AC(ds°)
From the above analysis, one can easily see that the estimated congestion toll will converge (to the optimal value) only when the traffic flow is at the socially optimal level. Even without the full knowledge of the demand function, the planner is able to 'find' the correct amount of toll as long as there is an adjustment process in the toll charge. 10.2.2 An Iterative Toll Adjustment Procedure for a General Network The above implementation mechanism applies for the standard demandsupply model, with which most existing theoretical arguments on road pricing are concerned. The analysis of convergence does hold for the convex, monotonic demand and supply curves for a single highway or for the aggregate demandsupply relationship for a congested area. The convergence is largely due to the fact that the socially optimal level of demand always lies in between the intended and the revealed flow levels at each trial, as long as the true but unknown demand function is monotonically decreasing and the supply or performance function is monotonically increasing. Furthermore, an implicit assumption actually adopted in the argument is that the revealed or observed flow is always equal to the origin to destination demand since only a single road is involved. Once we move to a general road network, the problem is not so simple and a number of factors compound its complexity. First, given a complete set of observed link flows on the network, arising from a trial of a complete set of link tolls, the corresponding OD demands remain unknown. One may naturally consider identification of the OD matrix from link traffic counts by virtue of many available techniques. Unfortunately, the OD matrix that reproduces a given set of equilibrium link flows when useroptimally assigned to the network, is in general not unique. Second, when the toll on a given link (or a set of links) is altered, the demand between a specific OD pair and flow on a specific link could either increase or decrease or remain unchanged. In other words, the change of OD demands and link flows in response to link tolls is not as simple and obvious as the standard textbook demandsupply case examined above. These facts imply that the bisection trial scheme does not necessarily work on a general road network. This conjuncture is indeed true. As verified
314
Mathematical and Economic Theory of Road Pricing
later by numerical examples, the bisection (or other fixed step size) trial scheme does not always converge on a general network. We begin with restating the following assumptions required throughout the analysis. Assumption 10.1 All users in the network are homogeneous in terms of their value of time and follow the deterministic UE behavior in their route choices. Assumption 10.2 For each link aeA,
the corresponding link travel cost ta(va) is a
strictly increasing, convex and twicecontinuously differentiable function of link flow va. Assumption 10.3 For each OD pair we W, the corresponding true, but unknown OD demand function Dw(\iw) is nonnegative, strictly decreasing, bounded and continuously differentiable in OD travel cost \xw, 0 ( ' ) = 5 » . h  2 f 5>»do) 1V>
^
y d
aeA
(io.2)
v*W 0
where Q, is the feasible set of link flows and OD demands as defined before by \ where fm
is the flow on route re Rw between OD pair weW,
charge expressed in equivalent time unit on link at A.
(10.3)
ua denotes the toll
TrialandError Implementation of Road Pricing Because ta(va)va = V"(ta((£>) +(nt'a(co))dco, where t'a(va) = dta(va)/dva,
315 problem (10.2)
can be solved by the user equilibrium traffic assignment procedure with elastic demand via setting a link travel cost function as the following marginal link cost function: ia{Va)=ta{Va) +Vj'a{Va), "SA
(10.4)
Let v* and d* be the optimal solution of problem (10.2), then the optimal link tolls based on the marginalcost pricing scheme, as shown in Chapter 3, are set as follows: u>vX{a€ A) can be identified by solving the following typical elasticdemand traffic assignment problem:
Note that one can use a more general link travel cost function with flow interactions (see Chapter 3 for more details). With such a generalization, the subsequent discussion still holds as long as the link travel cost and OD demand functions meet appropriate monotonicity and differentiability assumptions. As seen from above, obtaining optimal link tolls from the mathematical model (10.2) requires knowledge of the analytical demand functions. In reality, it is extremely difficult to know exactly users' willingnesstopay for a journey and thus identify analytical demand functions tailored for congestion pricing. To circumvent this difficulty, we now propose the following trialanderror iterative method to find the systemoptimal link flows and link tolls defined in eqn. (10.5), which obviates the need of the OD demand functions. What is required is to observe traffic flow on each link after each trial of link toll pattern. Trialanderror procedure for the firstbest pricing scheme Step 0. (Initialization) Let {v^0), a e ^ J be an initial set of feasible link flows. Set k:=0. Step 1. (Estimate link tolls) For each link ae A, calculate the current link toll charge by u v'a,asA
(10.20)
„* _> u'ay as A
(10.21)
> 0,aeA
Proof
(10.22)
For any (v, d) € Q., we have
(10.23)
KM where 7a (va) is the derivative of marginal link travel time function (10.4) with respect to link flow va, i.e., dia(va)/dva =2dta(va)/dva + va d2ta(va)/d\ derivative of benefit function with respect to OD demand dw. Hence,
d('
V2F,(v,d)
(t)
and B'w(dw)
is the
320
Mathematical and Economic Theory of Road Pricing (10.24)
Because £2 is a compact set, it follows that the right hand side of eqn. (10.24) must be bounded by the assumptions on link travel cost and OD demand functions. In view of Lemma 10.2, all the necessary conditions for the proof of the convergence of the method of successive averages are satisfied. We thus conclude that v(ak) —> v*, a e A and —> d'w, we W, where {v*, ae A} and [d'w, we ff} are the optimal solution of the minimization problem (10.2)(10.3). This also implies that (10.7), it follows that u(k) +u'aeA.
;«_,, 0, a G A. By eqn.
\
10.2.4 A Numerical Example Example 10.1 We now illustrate the proposed iterative procedure with a simple example that was used in Section 4.4 in Chapter 4. Here for convenience, we present the network and input data again. The network, as shown in Figure 10.2, consists of 11 links, 7 nodes and 4 OD pairs (1—>7, 2—>7, 3—»7 and 6—>7). The true, but unknown demand functions are given below: A*(m 7 ) = 600exp(0.04u^ 7 ); D M (  i w ) = 500exp(0.03u^7) ^ , 0 * 3  7 ) = 500exp(0.05u^ 7 ); £>^7(n6^7) = 400exp(0.05u^ 7 )
Figure 10.2 The network used for numerical example The link travel time function:
* > . ) = /» 1 + 0.151^?
TrialandError Implementation of Road Pricing
321
is used with link freeflow travel time, t"a, and link physical capacity, Ca, given in Table 10.1 (C a in the table is the link environmental capacity to be used later). Here link toll is given in equivalent time unit. Table 10.1 Input data for the example network Link a
£
ca
1
2
3
4
6
5
6
7
5 6
6
7
1
5
200 200 200 200
100
160 150 200
100 100 150
150
100 150
8
9
10
11
10
11
11
15
150 200 200 200 100
160 160
150
With these demand functions, we have the true optimal link flows and link tolls shown in Table 10.2 by accurately solving the systemoptimal marginalcost pricing problem (10.2)(10.3), using the FrankWolfe convex combination method (2000 FrankWolfe iterations are employed for a sufficiently accurate solution). The initial link flows are not restrictive and are generated via an allornothing traffic assignment with an initial predetermined OD matrix having an identical value for all elements. One can of course use an equilibrium traffic assignment procedure or use the observed, untolled link flows to obtain an initial set of link flows. The observed link flows at each trial are simulated by using an elasticdemand UE traffic assignment algorithm. Namely, for given toll pattern in eqn. (10.7) at iteration k, the observed link flows {v^\ aeAj generated by solving the UE traffic assignment problem (10.6).
Table 10.2 Systemoptimal link flows and link tolls Link Number 1 2 3 4 5 6 7 8 9 10 11
Link Flow 152.18 86.74 203.32 208.49 97.86 121.75 51.14 115.25 174.88 175.23 147.33
Link Toll 1.21 0.11 3.85 4.96 3.30 1.32 0.04 2.09 3.86 3.89 2.65
are
322
Mathematical and Economic Theory of Road Pricing
—o— Initial link flows with a uniform OD demand = 1 0 O— Initial link flows with a uniform OD demand = 1 0 0 —A— Initial link flows with a uniform OD demand = 200 O— Initial nontolled UE link flows
p—o—q—a—p—Q—q—•—q 0
2
4
6
8
10
12
14
16
18
20
Number of iterations
Figure 10.3 Change of the Euclidean distance with iterations for various initial flows Table 10.3 Estimated optimal link flows and link tolls Link Number 1 2 3 4 5 6 7 8 9 10 11
Link Flow 153.22 88.13 203.81 208.85 98.27 120.72 50.59 116.06 175.92 175.90 149.04
Link Toll 1.24 0.11 3.88 4.99 3.36 1.27 0.04 2.15 3.95 3.95 2.78
Using the true optimal link flow pattern shown in Table 10.2, we now use the following Euclidean distance from an intermediate estimated link flows to the optimal link flows to assess the convergence of the iterative procedure: \2
\2
(10.25)
For various initial link flows (including the observed, untolled UE link flows), Figure 10.3 plots the change of this Euclidean distance with iterations of the proposed procedure using a
TrialandError Implementation of Road Pricing
323
predetermined step size: ccw = l/(k +1), where k is the number of iterations. As seen for the first 20 iterations, the procedure does converge and the final convergent point is independent of the initial link flows. When a uniform OD demand equal to 100 is used to generate the initial link flows, the proposed method terminates after 8 iterations with a stopping criterion of 8 = 0.05. The estimated systemoptimal link flows and optimal link tolls are given in Table 10.3, which are sufficiently close to the true value in Table 10.2. Now we examine the convergence of a fixed step size in the link flowupdating scheme in (10.8). Here we consider two typical values of a w =0.5 and a w =1.0, leading to the following heuristic procedures. For cc(t) =0.5, k=l,2,, eqn. (10.8) yields: V(M)
=
Va
+Va
^a
e A
(1Q
2 6 )
For a w = 1 . 0 , £ = 1,2,, eqn. (10.8) results in: v,6)d(0
(10.29)
» ^ 0
where Bw (,6) = Z)"1 (,0), being the inverse function of the currently calibrated, approximate demand function
DK(\JLW,Q)
and U is the feasible set of toll vector u=(ua,ae A}.
After imposition of the newly determined link tolls u~f , aeA, a subset of revealed link flows v"(*+1), ae A is observed. The observed link flows are the result of distribution of the
326
Mathematical and Economic Theory of Road Pricing
realized OD demands on the network, and thus provide useful information about the true but unknown demand functions. Thus we can use the BLPP technique to estimate or update the OD matrix and the approximate demand functions from the newly available link traffic counts, thereby obtaining a new set of updated OD demand functions
There are two ways to update the demand functions: one is the updating of the OD matrix obtained in the last iteration from the traffic counts, and then using the new OD matrix to tune the approximate OD demand functions; the other is calibrating and updating the parameters of the assumed OD demand functions directly from the observed new trial results. The OD matrixupdating problem from traffic counts can be formulated as the following BLPP: y
l
( ( ) , )
y/
2
(,
(10.30)
where v(d) solves the following fixed demand UE problem for given de£2 d and ueU: min 5)((r.(a))+«,)dfl)+ aeA o
£_t a (co)dco
(10.31)
a£A,aiA 0
where
{
(
)
\
}
(10.32)
(10.33) In this model, v = (v o ,ae A\ represents the link flows vector observed in the last trial; d = ldv,ws W\
is the OD matrix (rearranged as a vector) obtained in the last trial;
Fj (v,v) and F 2 (d,dJ are functions of the generalized distance or error measurement between v and v, and d and d, respectively. The following Euclidean distance function is frequently used: Y1FA\(d),\) + y2F2(d, Y2)
mav
be preferred.
TrialandError Implementation of Road Pricing
327
In the case that the approximate OD demand functions are tuned by updating a set of parameters 6 e 0 directly, the problem is formulated as the following similar BLPP: mm YlJ F 1 (v(e),v) + Y 2 F 2 (d(e),d)
(10.35)
where (v(0),d(9)) solves the following elasticdemand UE problem for given 0 G B and ueC/
J
 H X j ( )
(10.36)
The same quadratic function in (10.34) can be used for the objective function (10.35). The above iterative bilevel OD demand function estimation and bilevel link toll optimization process is repeated until an appropriate convergence criterion is reached. To sum up, the general iterative trialanderror procedure through sequential price experimentation for estimating the secondbest link tolls is given below. Trialanderror procedure for the secondbest pricing scheme Step 0. (Initialization) Let {Dw(nw,0(O)), WE W] be an initial set of approximate demand functions. Set k := 0. Step 1. (Estimate link tolls) Solve the BLPP (10.28)(10.29) to obtain a new set of link tolls {«, a
el).
Step 2. (Observe link flows) Observe the revealed link flows denoted by (vj;*1, a e A\ after the imposition of the new link tolls
{M£A),
a e Aj.
Step 3. (Check convergence) If u(*+1)  u ( *'/W k) l < e, then stop. Otherwise, go to Step 4. Step 4. (Update OD demand function) Solve the BLPP (10.30)(10.31) or (10.35)(10.36) to obtain a new set of updated OD demand functions (Dw{y.w,Q{k+l)),wew).Let
k:=k + l and go to Step 1.
In the algorithm, E is a positive number for convergence tolerance. The procedure will be terminated after the difference between two successive trial results is within a given tolerance. Two remarks on the proposed procedures are appropriate here. First, to execute the procedure, one needs to assume a specific functional form as an approximation of the true
328
Mathematical and Economic Theory of Road Pricing
but unknown demand functions, and an existing reference OD matrix, if available, could be helpful for estimating/updating the OD matrix and the approximate OD demand functions initially. Given a set of limited trial experiments around the optimum solution, the procedure is expected to derive the demand curve for a limited range of traffic levels that is relevant to road pricing together with the estimation of optimal link tolls. Second, as a special case of the general secondbest pricing problem envisioned here, the proposed sequential linear demand function approximation (presented below) can of course be used to determine socially optimal tolls in a firstbest pricing environment, without the demand functions considered in Section 10.2. We may reasonably expect that the secantlike algorithm based on successive approximations would zone in on the optimal price much faster than the painfully slow process of the method of successive averages. This is important in practice, since the price experiments are generally limited to a few trials. Thus the proposed approach has an added bonus for the firstbest pricing problem studied in previous section, it involves fewer price changes and thus reduces the cost of implementing the pricing scheme, which is itself a social benefit. 10.3.2 Sequential Linear Approximation As assumed throughout this study, we have to specify an approximate (but easily invertible and computable) function Dw(\iw,&), we W, 6 e O , to use in place of the true, but unknown, demand function in the neighborhood of the optimum point of the secondbest pricing problem. One favorable choice is an interpolation and extrapolation scheme by a polynomial, including the simple linear and quadratic approximations. Here we adopt the following simplest sequential linear approximation scheme based on successively observed and estimated experimental data: { M " ) )
^) ( )
^
)
)
where 0 ^ = lz ^ ,d^\\i^ ),we
(
^
)
)
,
k=l,2,.
W are the function parameters,
(10.37) {ZJ^WEJF}
are the
slopes of the linear demand function approximations, id^\wG w] are the OD demands estimated from observed link flows at iteration k using the OD matrix updating model (10.30)(10.31), and jji^^ws fFj are the corresponding equilibrium OD travel cost inclusive of toll charges if any. By considering the straight line through the latest two experiment points, we have: =l,2,...
(1038)
TrialandError Implementation of Road Pricing
329
The line slopes, {zf\ w e WJ, are updated from iteration to iteration, based on the observed results in each trial. This successive linear approximation, if convergent, is expected to become a tangential approximation of the true demand functions and the tangent point is the resulting equilibrium point associated with the secondbest optimum toll charge. Note that, when estimating the current OD matrix d^\weW l
, the OD matrix d^' \we
by the model (10.30)(10.31)
W obtained in the previous iteration is used as the reference
matrix to ensure that the sequence of the OD matrices generated converges to the final optimum. In this regard an initial seed or reference OD matrix is used to estimate the initial OD matrix for the linear approximation and an old matrix from a previous study could be used for this purpose. The quality of the initial seed OD matrix is immaterial because its impact on the resulting OD matrix estimates will diminish as the OD matrix is successively converged. In addition, a rational linear approximation of the true demand curve requires that the slope \z^\ we W] calculated by (10.38) from the latest two experiment points be negative, a property of a demand function as mentioned in Assumption 103. Nonetheless, this requirement is not always ensured as the iteration continues, particularly when the trial solutions approach the real optimum. Because the OD demands, d!f and d^~^, are merely estimates from traffic counts, the slope z * may become unstable as d^ x)
d^~
and
move closer to each other due to estimation inaccuracy. Thus, a correction has to be
made whenever a positive slope emerges by, for example, resetting it to a small negative value. 10.3.3 Two Numerical Examples We now present two simple numerical examples to elucidate the applications of the proposed sequential price experimental approach to the firstbest and the secondbest road pricing problems with unknown demand functions, respectively. Example 10.2 In this example, we show that the algorithm based on successive linear approximations may involve fewer price changes than the method of successive averages does for the firstbest pricing scheme. For simplicity, we consider the standard textbook case; a single OD pair connected by a single road, with the following true demand and link travel time functions: = 600exp(0.05n)
(10.39)
Mathematical and Economic Theory of Road Pricing
330
f
(
(10.40)
f(v) = 10 1 + 0.15 —
Since a single road is concerned, the revealed or observed flow, v, is always equal to the OD demand, d, and hence the OD demand estimation process is bypassed here. 25.0 A—Bisection Method XMethod of Succesive Averages
20.0
^ — Linear Approximation
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18
Iteration Number
Figure 10.6 Comparison of the convergence rates by the bisection method, the method of successive averages and the sequential linear approximation method in Example 10.2 By equalizing the inverse of the true demand function and the following marginalconfunction denoted as t (v)
dv
(10.41)
the socially optimal level of flow, full trip cost and toll charge can be identified to be: v'=d' = 216.898; \i'= 20.350; u= 8.300 The equilibrium flow and full trip cost without tolling are equal to v(0> = dm = 276.581 and (x(0) =15.489. Starting with this observed flow with the corresponding initial toll charge um = 21.933 calculated by (10.7), the results by three methods (the bisection method with a w = l / 2 in (10.8); the method of successive averages with a(k) = l/(& + l) in (10.8), and the linear approximation method proposed in this section) are shown below.
TrialandError Implementation of Road Pricing 40 T
"
"
331
Unknown Demand Curve
— 0 — Search Path
35 
Marginal Cost Curve Average Cost Curve
30o.
£25 0*
a.
I 20
u 1510
90
130
170
210 250 Number of Trips
290
330
370
Figure 10.7 Convergence pattern of the sequential linear demand function approximation in Example 10.2 As already mentioned, the bisection method does converge for the simple single road case considered, and hence is applied here for a comparison. For a given tight convergence criterion e =0.001, the number of trials required by the bisection method, the method of successive averages and the linear approximation method are respectively 5, 18 and 7 (also refer to Figure 10.6). For this special example, the bisection method has the quickest convergence (but it does not work for a general network), the proposed sequential linear demand function approximation method outperforms the method of successive averages. Note that the relative convergence of the three methods depends on the curvature of the demand function. Clearly, if the true but unknown demand function is linear, the linear approximation method will arrive at the optimum with two iterations only. The average private cost and the marginal social cost curves are plotted in Figure 10.7 together with the true but unknown demand curves (dashed). The linear extrapolation and interpolation lines are drawn through the two most recent experimental points (except the initial point which is the observed point without toll charge or in the donothing case). The points are numbered in the order that they are identified in the sequential price experimentation. For this simple single road example, the true OD demand and OD cost in each trial are observed directly without an estimation process and thus all trial points are located on the true demand curves. As observed from the figure, the sequence of
332
Mathematical and Economic Theory of Road Pricing
experimental points converges to the social optimum (the intersection between the true demand curve and the marginal social cost curve). Example 10.3 In this example we use a small but general network to illustrate the application of the trialanderror procedure for the secondbest pricing problem. The network, as shown in Figure 10.8, consists of 4 links, 3 nodes and 2 OD pairs (1—>3, 2»3). The true demand functions are given below: A^ 3 (Hi^j) = SOOexp(0.03u^3), D^ ( u ^ 3 ) = 400exp(0.05u^ 3 ) The following link travel time functions are used for the 4 links respectively: I
^200j I
r,(v,) = 9 l + 0 . 1 5   ^  
n 3J
y
15O
,
1 JJ
v
'
I
^ 150J
?4(v4)= 11 l + 0 . 1 5   ^  
[
1 2 O O JJ
The set of tolled links and observed links are assumed to be identical with two links: ^ = ^={3,4}.
Figure 10.8
The network used for Example 10.3
For given toll charges in each trial experiment, the observed link flows are simulated (generated) by using the elasticdemand traffic assignment method with the true nonlinear demand functions. In view of the small size of the example network, we used an extensive direct search method without derivatives (HookeJeeves pattern search method) for finding the global optimal solutions of both the bilevel OD estimation and toll optimization problems in each trial. For the purpose of validation, we identified the true optimum (never known in reality without demand functions) by extensive direct search using the true OD demand functions. With the known optimum toll solution, we can test whether or not the sequential alternate bilevel OD estimation and toll optimization algorithm converges to the real optimum. To start the iterative procedure, the initial observed link flows are given as the equilibrium link flows without toll charge (donothing case), and the initial OD demand and OD cost for the
TrialandError Implementation of Road Pricing
333
first linear approximation of the demand function are estimated from the initial observed flows with an initial seed OD matrix: d,^ = d2_ti = 200. The starting point of the toll charge is (« a e Ay
and, according to the duality theory in nonlinear programming, the
objective function ). Namely, v^ = va (u (i) ), ae A. From eqn. (10.49) in Lemma 10.3, it can be readily verified that the righthandside of eqn. (10.51) amounts to a projection operation u = Proj^ (u(lr) + a w Vcp(u ( "))
(10.52)
where 9T ={u u>0}, V(p(u(*)) = (9(p(u(*r))/3wo, ae A \ where the projection operator is defined by Prcv (y) = argmm y  x2
(10.53)
Hence, the iterative formula (10.52) is a gradient projection method with a step size a w satisfying condition (10.9) for the dual problem (10.47). The following theorem shows the convergence of the iterative scheme (10.52) in finding an effective toll pattern for the traffic restraint and road pricing scheme.
TrialandError Implementation of Road Pricing
Theorem 10.3
If u(t+1) = u (t) at iteration k, then uik)el/',
339
being the set of optimal
solutions for the dual problem (10.47). Otherwise, l i m u ^ ^ u * and
u'ell'.
As the gradient Vcp(u) defined in (10.49) is a bounded function, then the sequence of vectors, u
j , generated by iterative scheme (10.52), will be convergent to an optimal
solution for the dual problem (10.47). The theorem comes from the fundamental property of projection operators and the (sub)gradient algorithm for the Lagrangian dual formulation of a convex program. We conclude this section with two remarks. First, unlike the trialanderror method for the firstbest and secondbest pricing problems examined in Sections 10.2 and 10.3, the procedure proposed here for the traffic restraint and road pricing scheme updates the link tolls in a straightforward manner, based on the difference between the given flow restraint levels and the observed link flows only after each trial; neither link travel time nor demand functions and nor users' value of time are needed explicitly. Second, the trialanderror procedure introduced in this section deals purely with the traffic restraint policy through pricing, and it can be simply altered to handle the marginalcost pricing problem with link (physical or environmental) capacity constraints described in Section 3.3 in Chapter 3. In this case, the procedure given in this section and that developed in Section 10.2 should be combined or the primal and dual method for a convex program should be used to identify the optimal link tolls. Namely, the new trial tolls should be given as the sum of the tolls determined by eqns. (10.51) and (10.7). Clearly, in the joint implementation, step sizes used in eqns. (10.51) and (10.8) in the two procedures should be different, and the users' value of time is needed for determining the marginalcost pricing link tolls. 10.4.3 A Numerical Example Example 10.4 We consider the same network used in Example 10.1. The traffic restraint and road pricing scheme here is to find an efficient set of link tolls to hold the flow on each link in the network (A = A) below or equal to its environmental capacity, that is already given in the bottom row of Table 10.1. To test the proposed trialanderror procedure, we apply two step size sequences satisfying condition (10.9): a{k) = \/{k + l) and a w e [ 0 . 5 / ( l + A:), 2.5/(l + A;)]. For the second sequence, an appropriate step size at each iteration is given by randomly generating a number
340
Mathematical and Economic Theory of Road Pricing
in the relevant interval. In both cases, the initial toll charge is zero and the observed, untolled UE link flows are used to obtain a first set of link tolls for the toll experiments After implementing the iterative trialanderror scheme, the convergent link toll pattern (in equivalent time units) and the ratios of equilibrium link flow to link environmental capacity are together summarized in Table 10.4. It can be seen from Table 10.4 that the ratio of flow to environmental capacity for each link is less than or equal to one; a link is subjected to a toll charge only when its flow reaches its environmental capacity threshold; otherwise it is free of charge. This means that the objective of environmental capacity constraints is achieved perfectly through the effective link toll charge identified from the iterative trialanderror implementation procedure. Figure 10.11 shows the convergence of the iterative trialanderror scheme in terms of the absolute difference of link tolls between two successive iterations:
u ( * +1) u )d(O
(11.1)
0
where Qv is given by Q v ={v v = Af, Af = d, f>0] with v = (va,aeA)
, d = (dw,weW)
(11.2) and f = (fm,reRw,weW)
being the vectors of
link flows, OD demands and path flows respectively, as defined before. Alternatively, the UE problem can be formulated as equivalent Variational Inequalities (VI) in terms of link flow variables. Find vue e Qv such that t(v u e ) T (vv u e )>0, foranyven v
(11.3)
ue
With the UE link flow solution v , then the total system travel time is given by Tm = 7\vue) = ^
/„ (vf) v". On the other hand, the standard SO model that minimizes the
total system travel time is given by min 7\v) = ] > > » „
(11.4)
where, as shown in Chapter 2, if ta(va) is monotonically increasing and convex, then ta(va)va
is convex and hence the solution to the SO problem (11.4) is unique. Let vso
denote the link flow solution of the SO problem and define the following ratio
p=iHL = _ L 4 Tso
(11.5)
r(v)
where ' fx' denotes the case of fixed demand. Clearly, pj^ > 1, this ratio is called the inefficiency, or price of anarchy, of the selfish user equilibria (Papadimitriou, 2002; Roughgarden, 2003). The ratio, p£, can also be regarded as the efficiency gain of a marginalcost pricing scheme, because it can drive a UE flow pattern to a system optimum or completely remove the inefficiency of selfish user equilibria. Here we introduce the following relative efficiency gain of a marginalcost pricing scheme in terms of the ratio p"':
r(v ue )r(v so )
Efficiency Gain and Loss of Road Pricing
347
Our purpose here is to find an upper bound of p"xe, thereby quantifying the maximum efficiency gain of marginalcost pricing schemes. The case with linear cost functions The following theorem quantifies the maximum inefficiency of the UE link flow pattern with linear link cost functions. Theorem 11.1 Let vue be the UE link flows with separable linear link cost functions, and let vs° be the SO link flows. Then p£ Proof.
=T(Y"°)/T(VS°) 0, ba > 0, a e A. From VI (11.3) of the UE problem, we readily have
(r) where the second inequality holds because (va0.5v™) >0, and the third inequality holds because aa(v™f 0, we thus arrive at
This completes the proof. Travel Time
Freeflow Travel Time 0
»"•
V
a
»Flow
a
Figure 11.1 Geometric illustration of the proof of Theorem 11.1 From the above proof, we have the following immediate result of inequality for any nonnegative link flow vector v > 0: asA
4
The following theorem shows a pseudoapproximation result that upper bounds the social cost of a UE link flow pattern by that of an SO flow distribution with uniformly upscaled OD demands. Theorem 11.2 Let vue be the UE link flows with separable linear link cost functions, and let V(°+1/4) be an SO link flow solution with the uniformly scaled OD demand (1 +1/4) d or 5/4d, then r(v u e ) o (vr K" * f E ' . ( c ^
(]+1/4) =£'„
(vrK,(i+V4,
Whereas the second inequality follows from (11.9). The case with nonlinear link cost functions We now consider the case with general nonlinear separable link cost function. Like Figure 11.1 for the linear cost functions, Figure 11.2 depicts the situation for arbitrary, but still separable, link cost functions. Travel Time
'.(O
Link Cost Function Freeflow Travel Time
•Flow
Figure 11.2 Geometric illustration of the definition of j(C) We now consider how to upper bound the area of the shaded rectangle, or the term (^( v r)~ ? o ( v a)) v a >
m
terms of the area of the large rectangle, which is of size fo(v™)v^e.
For each link cost function ta=ta{za)
and nonnegative link flow za>0,
we define the
following parameter (11.10) Here, 0/0 = 0 by convention. Since (ta(za)ta(va))vaza>
we
have
if 0£ = 1.1505. In addition, p" x ^l and
C ^ °
as p > 0 (without traffic congestion), but p"xe > °° and (t)fXe > °o as p H> +=» (with severe congestion). Example 11.1 (Network example of furnishing the maximum efficiency gain) Consider a network having one OD pair connected by two links. Link travel time functions are
352
Mathematical and Economic Theory of Road Pricing
tt(vt) = \ and t2(v2) = (v2)p,p>0,
respectively. Let the O D demand b e d =
vl+v2=\.
For any p , the UE solution is v,ue=0 and v 2 e =l with r(v u e ) = l. The SO solution is v," =!(/> +1)""'and vs2°={p + \yllp
with r(v s °) = l  p ( p + l)~(p+1)/* 0, represents the worstcase example of the inefficiency of selfish user equilibria, or the maximum efficiency gain of a marginalcost pricing scheme for a class of polynomial link cost functions. 11.2.2. Efficiency Gain for Cost Functions with Limited Congestion Effects We note that the inefficiency bound given in Theorem 11.3 for the standard traffic equilibrium problem is a worstcase measure, taken over all possible instances. The actual ratio of the UE total cost to the SO total cost in realistic instances could be substantially smaller. Indeed, in a traffic network, the freeflow travel time is usually not a negligible fraction, which is encountered in both UE and SO flow states. Here, we present a parameterized, improved bound on the inefficiency of equilibria introduced by Correa et al. (2005), where the ratio of fixed to total travel cost at equilibria come into play. Lemma 11.2 Let vue be a UE link flow pattern, with separable link cost functions drawn from a given class C, such that toa=ta(Q)>r\{\a°^ta{y^
for all a e A, for some positive
constant 0 « + 1 J ^  U ^ H 
If
v a >v^ e for all aeA, eqn. (11.24) clearly holds. Now, we assume vaT(v ue )? a (vf),
the area
of the smaller shaded rectangle is at most y(C) times that of the
Efficiency Gain and Loss of Road Pricing
353
rectangle with upperleft corner point (0,f a (v"n and lowerright corner point which is of size at most (lri(v u e ))/ a (v^)vf. The result follows.
•
Travel Time
Link Cost Function
0
vu*
vu
•Flow
Figure 11.3 Geometric illustration of the proof of Lemma 11.2 The following theorem is a generalization of Theorem 11.3 as a direct result of Lemma 11.2. Theorem 11.4 Let vue be a UE link flow pattern with separable cost function drawn from a given class C, such that t° =ta(0)>r\{\"°^ta{v^
for all as A, for some positive
constant 0 0 (the demand is perfectly elastic). In what follows, we attempt to derive a pseudoapproximation bound of pec in terms of the social welfare and user benefit at a given user optimum solution as well as the parameter y(C) established previously. We first introduce the following lemma. Lemma 11.3 If Bw(dtr) is a nonincreasing function of dw for dw>Q, then
J Proof.
J w (co)dco + X^(C)KC)
Since Bw(dw)
(H34)
is nonincreasing, we then have
(dwdl°)Bw{dle)>
)da + (dwd:°)Bw(d:°),weW 0
(11.36)
0
Summing up over all OD pairs weW
leads to (11.34).
•
Theorem 11.5 Given general nondecreasing link cost functions, ta(va),aeA, increasing demand functions, Bw{dw),wsW,
and non
then (11.37)
where y(C) is defined by eqn. (11.11) and co(vue,due) is the ratio of user benefit (denoted by U) to social welfare at user optimum:
(11.39)
Proof.
Since (vue,due) is an equilibrium solution, from the VI formulation of the elastic
Efficiency Gain and Loss of Road Pricing
357
demand traffic equilibrium problem we have
I' o (vr)(v a vr)XMC)KC)^0, for any (v,d)eQ aeA
(11.40)
weW
Substituting (11.34) into (11.40) yields
J
J
It can be rewritten as
Using the definition (11.29) and Lemma (11.1), it follows that (11.41) Using the definition (11.32) and let (v,d) = (vso,dso) in (11.41), we have
;
Using
5(v u e ,d u e )
= 5'(v'IC,du
O6/4 o
subject to =d' fm>0,
W
*W
(11.47)
reR^, weW
(11.48)
With the assumption of monotonically increasing link cost function, the above problem of minimizing a strictly convex function over a compact (closed and bounded) set guarantees the existence and uniqueness of a path flow solution fsue e Q / . In addition, the entropytype function V =
» Z
ensures (cZ
that the optimum a
/w^or»
is achieved
at an interior
point.
Using
necessarily and sufficient condition for a unique optimum
Vue&D.f is that [V f Z(f S M )] T (ff")>0
for any f e ^
(11.49)
Substituting
into (11.49) and in view of
we have VI (11.44).
•
Let f e F be a systemoptimal path flow vector. From eqn. (11.44) we have
I lL(r')+iin/rl(/;°/r)^o This leads to
where C = c m ( f s u e ) , C = c w ( f s t ) ) and
Note that
Thus eqn. (11.51) becomes
01.50)
360
Mathematical and Economic Theory of Road Pricing
01.52) As discussed before, the second term of the righthandside of (11.52) is bounded by 01.53) as A
Now we seek a bound of the third term of the righthandside of (11.52) and thereby bound the overall cost inefficiency of the SUE. Lemma 11.5 Consider the following maximization problem maxZ(x,y)=>,x,)lnx,
(11.54)
subject to
Jx,=rf
(11.55)
1=1
t,y,=d xi,yi>0,i
(11.56) = \,2,,n
(11.57)
where d>0 is a constant. The optimal value of this problem is Zmax = kd, where k solves equation, kek ={n — l)/e, with e being the base of natural logarithm, or if it is defined that g{x) = xex, then k =
g~l((nl)/e).
Proof: The KKT necessary conditions for optimality are dx
{dxt
^ + X20, (f + where X, and X2 are the Lagrange multipliers associated with the equality constraints (11.55) and (11.56), respectively. Note that the entropy type of function ensures an interior point of positive solution:
xf >0, / = 1,2, •••,«. Using
dZ/dxj = y>/xi  Inx;  1 and
dZ/dy, = In x, for / = 1,2,.. .n, the KKT conditions reduce to ^   l n x ,  1 + ^ = 0 , x ; > 0 , i = l,2,",w
(11.58)
lnx,+^. 2 =0,if yf>0,
(11.59)
\nXi+X2>,) = (e~' \o) or
Efficiency Gain and Loss of Road Pricing 2)e
Xl
),
i = \,2,,n,
\XlX2>0.
361 Let
k,=\ + \
and
k2 = X2, the solutions become ( W , ) = ( e \ 0 ) or ( W , ) = ( e \(fc 2 £>*>), i = U2,,n,
k2k,>0
Let m be the number of (x^y,)
taking value (e*2,(&2—fc^e*2) at optimum, then
(n — m) is the number of (x^y,)
taking value ( e \ o ) . Substituting these values of
(X^JA)
into the objective function (11.54), we have Zopx={n~m)ek'kl+m{k2kll)e^k2
(11.61)
Also, from constraints (11.55) and (11.56), we have (nm)ekae
A)
be the systemoptimal link flow solution
to the SO problem (11.4). From VI (11.77) and the fact that Vs0'* e ii*, ke K, we have
Since v
«=Z
t
^
' . K ) = '°+ .)' and t'M =
max ell ,k
(11.79) k
The last inequality is due to the fact that the constraint v* e £l
is relaxed into
v* > 0, ke K. Now we introduce the following theorem to establish the upperbound of the inefficiency of the CN equilibria. Theorem 11.7 let v
so
Let vQnt,keK
be a solution of the CN equilibrium problem (11.77) and
be a solution of the SO problem (11.4), with polynomial link cost function
(11.18).
Then, (UM>
for \K\>2 and p£ =r(v c ")/r(v s o ) = l for \K.\ = \, where £, is defined by (11.90) and (11.89) given below.
Mathematical and Economic Theory of Road Pricing
366
Proof. The task here is to find a solution of the second term of the righthand side of (11.79), i.e., v*>O,teK
J
l
r
^
j
R r l l
(11.81)
Introducing the Lagrange multiplier for the constraint v* >0, ke K, ae A, we can write down the firstorder optimality conditions for (11.81) as:
aapv:nA(v:n)p~l+aa[(v:«y {vayyaaP{vay+\ka=o,
kex
(n.82)
Xka > 0 , v* > 0 , Xkavka = 0 , ke K
(11.83) 1
2
2
From (11.82) and (11.83), we conclude that if v™ > v^" , then v a > v = 0 . Thus, let v™'* =argmaxv a cn '*, ae A
(11.84)
kzK
and (11.85) Clearly, 0 < £a < 1, ae A. Then, from (11.82), we obtain (11.86) and
\ +p
vr,yj=o,
Therefore,
aa\(v:y (v.y]v.
max
keK.ktk
(11.87) (11.88) where PL+1 \+P
\K\l
(11.89)
Efficiency Gain and Loss of Road Pricing
367
Inequality (11.87) follows from the fact that >
/
jLmJ
^
\
u\ 2 Q
1
}
I
I E^"
i
»
en A I
T
V "~ ^/7 I
a
£—4
I jy~\
i
( \
\^ "
/
and (11.88) follows from the fact that f° >0 for any ae A. Define § = max^
(11.90)
aeA
Then it follows from (11.79) and (11.88) that r ( v c n ) < r ( v s o ) +^r(v c n ) Hence, the assertion (11.80) follows immediately.
•
Note that, in the proof of the above theorem, we assume that there is a unique k E K, such that
v™J>v™*
for
all
v cn.*i = v cn^ = ... = vcM»>von.t
keK,k±k. for
au
If
there
are
kx, k2, •••, km,
an
0, keK Z"ivo'=va
such
y k^kn
i: = l,2,,m,
such that
and the assertion (11.80) also
holds. 11.4.3 Upperbound of Special Cases and a Numerical Example If there is only one CN player, i.e., KJ = 1, then the CN equilibrium problem of interest reduces to the standard system optimum problem. Indeed, in this case, C,a = 1 in eqn. (11.85) and ^,,=0 in (11.89) for all aeA,
and as a result, the inefficiency bound is
pj" = 1, which simply implies that there is no efficiency loss. For the linear cost function with p = l, from (11.89) we have ^a=
4(\K\l\
~>aeA
(11.91)
It is obvious that when \K\ —* °° or when there are an infinite number of infinitesimal players, then Co—»0 and ^—>l/4
in eqn. (11.91) for all aeA,
we thus have
l/(l — £) = 4/3, a bound obtained and given previously in Theorem 11.1 for the Wardropian user equilibrium. Note that one can interpret this extreme case in the following manner. In a network with a certain number of OD pairs weW
and demand dw, suppose there are mw
players in OD pair w, each controlling a positive amount of demand dJw,j = \,2,,mw and ^m"=ldJw=dw, weW.
One may assume mw identical players and thus
dJw=dwjmw.
368
Mathematical and Economic Theory of Road Pricing
Then, as mw —> °° (hence \K —> °°), the behavior of the CN equilibrium yields a total link flow vector corresponding to a Wardrop equilibrium (Haurie and Marcotte, 1985). For the general case where p > 1, if there are infinite number of players competing on the network and t,a —> 0, ae A (H.92) and the bound of the CN equilibria becomes
which is identical with the upper bound result (11.21) obtained previously for the deterministic Wardropian user equilibrium problem. Example 11.4
Consider a simple network shown in Figure 11.5, consisting of 3 nodes and
4 directed links, used in Section 6.4 of Chapter 6. Link cost functions are given by
Figure 11.5 The simple network used in Example 11.4 There are two OD pairs (1 —»3 and 2 —> 3) in the network, and the demands are assumed to be fixed with d^3 = 20 and d2^3 = 20. There are three paths connecting OD pair 1 —> 3 : path 1 with a single link 1, path 2 with link 2 and link 3, and path 3 with link 2 and link 4. There are two paths connecting OD pair 2 —»3: path 4 with a single link 3, and path 5 with a single link 4. The unique link flows at the system optimum are v,so =14, vf = 6 , v° = 11 and vf =15, with the corresponding link travel times r,(v1S0) = 34, U (VT)=
? 2 (v") = 26,
15. The total travel time at system optimum is 1066.
r3(v3so) = 19 and
Efficiency Gain and Loss of Road Pricing
369
Suppose that the two OD pairs are controlled by two distinct CN players, with OD pair 1 > 3 by player 1 and OD pair 2 > 3 by player 2. At the CN equilibrium, the playerspecific link flows are unique and are given by v,"1'1 =88/7, v2n'' =52/7, v3c"' =50/21 and v4cnJ =106/21; v,cn2=0, v2cn2=0, v3cn2=26/3 and v4cM =34/3. The total travel time is 67664/63. The ratio of inefficiency is p™=67664/63/1066=1.007534. In our notation, £,=£ 2 =1, ^3 =0.7845 and £4 =0.6919. ^ = £ 2 = 0 , £3 =0.1342 and £ 4 =0.1420.
Thus, ^ = max{^,,^2,^3,^4} = 0.1420
and our bound is p™=(l^)~'
= (10.1420)"'=1.1655.
11.5
MAXIMUM EFFICIENCY LOSS OF SECONDBEST PRICING SCHEMES
As pointed out before, due to the imperfection of the firstbest pricing from a political and technical implementation perspective, the secondbest charging schemes are more practically relevant, and indeed have received ample attention recently. A second best pricing scheme, if introduced, is expected to lead to system efficiency in between those of the untolled UE and SO flow patterns. Therefore, we are interested in establishing the efficiency loss of a secondbest pricing scheme in comparison with the firstbest one, and determining the room or potential for further improvement of the secondbest pricing scheme per se. 11.5.1 Measure of Efficiency Loss We now consider measuring the efficiency loss of a secondbest pricing scheme by the (absolute and relative) gap between the performance measures under the current secondbest pricing scheme and the systemoptimal one. As before, the total system travel time for the fixed demand case or the total social welfare for the elastic demand case can be used as the system performance measure. Let r(v ue (u)) be the total system travel time under a secondbest pricing scheme u, where v ue (u) is the corresponding UE link flows with pricing scheme u. Again let 2"(vso) be the minimum system travel time associated with SO link flows vso. Again we define
Again, p™(u)>l, by definition of u, being a secondbest pricing scheme. We can thus
370
Mathematical and Economic Theory of Road Pricing
have the following relative efficiency loss associated with a secondbest pricing scheme in terms of the ratio PfXe (u): /
^ r(vue(u))r(vso)
01.95)
Similarly, in the case of elastic demand, we have the following counterparts: S(v,d) «e u) = n±——I— Y S[y ue (u),d ue (u))
C u=
y
(11.96)
' ,y—'
v
(H97)
"=\—VT
Therefore, determining the maximum efficiency loss of a secondbest pricing scheme is equivalent to establishing an upper bound of the inefficiency index PfX(u) in (11.94) or pJ°(u) in (11.96), associated with a given secondbest pricing scheme, u, with fixed or elastic demand. 11.5.2 Bound for Traffic Equilibria with Fixed Demands Consider a secondbest pricing scheme n = (ua,ae A) >0, with the modified link cost function ta(va) + ua, the VI formulation (11.3) oftheUE problem can then be rewritten as: (t(v ue ) + u) T (vv u e )>0, foranyveQ v
(11.98)
which leads to
For each link cost function ta = ta (va) and nonnegative user equilibrium link flow v"a° > 0, we define the following parameter for each link a e A, associated with link toll charge ua:
^ S « ( C ) (".I") y (t ,v,« )= m B [ ( f ' W t  M i ;  + ( v
;
vfl>o
M v )v l
a\
a/ a
v
'^,
i f
u S v v(v)
(11.101)
Efficiency Gain and Loss of Road Pricing
371
Travel Cost
>Flow Figure 11.6 Geometric illustration of the numerator of eqn.(l 1.100)
for 0voe'Ma) becomes much greater than 1 as ua becomes large. Also, from (11.109), y(C,u) should not always go to positive infinity for meaningful u, otherwise, the presented
bound
becomes
ineffective.
If we still
0<wo < v ^ ( v " ' ) , then the value of la{ta,v^,ua^
use definition
(11.101) for
may become huge if ua is sufficiently
small. This can be seen in the extreme case of ua = 0 and zero freeflow travel time because Yo (?a>vr>"a) ~~>°°
as v
a~*0 in this case.
Note that the bounding equations (11.108) and (11.109) assume that the toll charge is either uniformly less (or greater) than the congestion externality for all links in the network. It is obviously unrealistic in practice. In particular, for the trialanderror pricing scheme developed in Chapter 10, the toll charge in each trial for each link can be either less or greater than the congestion externality. It is therefore necessary to establish an inefficiency bound for general secondbest pricing schemes that can accommodate both possibilities. This can be done by the following relaxation. Let set A\ consist of those links whose toll charges are less than their congestion externalities, i.e., 0>(I aeA
as. A
As before, using the definition (11.94) and let vfl=v^°, we have
... . . ..
Efficiency Gain and Loss of Road Pricing
375
that is p
*(u)
\_
(C^)
(11112)
To sum up, we state the following Theorem. Theorem 11.9 Let vue(u) be a UE link flow pattern associated with a pricing scheme u, with separable cost functions drawn from a given class C, and let vso be an SO link flow solution, then
11.5.3 Bound with Polynomial Cost Functions Consider the convex polynomial cost function (11.18), ta (va) = t° + aa (va ) p , a e A,
t°>0,
aa > 0 and p > 1. Let ua be the toll charge on link a e A. For the case where 0 0 for all
a e l
Next we consider the case where ua > v™t'a (v^e) or ~Ka > 1, a e ^ . By definition,
(11.119)
<max
where the inequality holds for positive value of numerator and t° > 0 for all a e A. This is true because, for the maximization of Ja(ta,v™,ua^
and the bounding inequality (11.99),
we only need to consider va, for which a a ( ( v ^ e ) ' '  ( v a ) ' ' ) v 0 + ( v a  v " e ) « a > 0 . This means that the maximum value of la(ta,v™,ua} polynomial cost function.
can be obtained by taking t°=0
for the specific
Efficiency Gain and Loss of Road Pricing
377
Now we are at the stage to find the optimal v* for the maximization problem (11.119). Define
(1L120)
Let dF(v a )/dv o =0, we have (11.121) Where KO is defined in (11.116). Solving the equation yields
Then from eqn. (11.119), we have
and
If we further assume that the secondbest pricing scheme, u, is chosen such that Ka = K for all ae A, then finally we arrive at i
 Kp, if 0 < K < 1
Y(C,u) =
(11.123)
Clearly, if K = 0 (U = 0), we have y(C,u = 0) = p(l +py(p^)/p.
As expected, we have the
same result given in (11.20), in the absence of pricing. In contrast, if KO = K = 1 for all ae A or in the case of a perfect marginalcost pricing scheme, we obtain y(C,u) = 0. This result simply implies that "there is no efficiency loss" or "the UE flow pattern is also systemoptimal" under the marginalcost pricing scheme. Example 11.5
Consider a traffic network with the polynomial link cost function,
ta(va) = t° + aa(va)p,
ae A. Let v ue (u) be a UE link flow pattern on the network
378
Mathematical and Economic Theory of Road Pricing ay°t
(v"e
= Ka p I v"e
for each link as A, then, by definition, (11.94) and (11.95), the relative efficiency loss in relation to the system optimum, associated with a pricing scheme, u, is given by )
K1
(11.124)
[y(C,u), if K > 1 where y(C,u) is given by (11.123). Figure 11.9 plots the relationships between parameter K in the interval 0 < K < 3.0 and the maximum efficiency loss (Pf*(u) for p = \ (linear), p = 2 (quadratic), p = 3 (cubic) and p = 4 (BPR functions of degree 4). In this figure, the values of efficiency loss at u = 0 corresponds to the early bounding results without pricing (refer to Figure 11.4 with i\ = 0).
Quadratic Cubic B.P.R.
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Parameter Kappa Figure 11.9 The maximum inefficiency loss associated with a secondbest pricing scheme for polynomial cost functions 11.5.4 Bound for Traffic Equilibria with Elastic Demands Now we move on to deal with the bounds of efficiency loss of a pricing scheme for the case with elastic demand. We begin our analysis with the secondbest toll charge being uniformly less (or greater) than the congestion externality for all network links, and then present our results for the general mixed situation. The following theorem follows immediately. Theorem 11.10 Let vue(u) be an elasticdemand UE link flow pattern associated with a pricing scheme u, with separable cost functions drawn from a given class C, and let vso
Efficiency Gain and Loss of Road Pricing
379
be an SO link flow solution, then (co(u)l) Y (C,u) if
0vl%(v™) in a single bound. Using the same definitions of 4> A^, Y,(C,u) and Y2(C,u) as in the fixed demand case, we now have the following Theorem.
380
Mathematical and Economic Theory of Road Pricing
Theoremll.il
Let vue(u) be an elasticdemand UE link flow pattern associated with a
pricing scheme u, with separable cost functions drawn from a given class C, and let vso be an SO link flow solution, then
where ro(u)
=E l M
ffl
(U\= J £ L
(11.129)
C/ue(u) = C/(doe(u)) awJ 5lue(u) = 5'(v"e(u),dlie(u))
are the user benefit and social
welfare at user equilibrium associated with pricing scheme u, and t/ so =C/(d so j is the user benefit at the social optimum. Proof. Like the proof of Theorem 11.5, the counterpart of eqn. (11.41) becomes 5(v u e (u),d u e (u))5(v ) d) + Y1(C,u)r(vue(u)) + Y 2 (C,u)r(v)>0 Substituting
r(v u e (u)) = 5 r (v lie (u),d ue (u)) + f/(d ue (u))
and
(11.130)
T(\) = S(y,d)
into (11.130) yields
0 Using the definition (11.96) and let (v,d) = (vso,dso) in (11.131), we have
By the definitions of co, (u) and co2 (u), we obtain eqn. (11.128).
•
Theorem 11.11 is particularly useful to establish the inefficiency bound or efficiency loss for practical secondbest pricing schemes with elastic demands. Once p^(u) is determined, we can calculate the efficiency
loss,
0 for any limited value of co(u). Moreover, l + y2(C,u)
l + Y2(C,u)
)
as Yi (C,u) —> 0 and y2 (C,u) —> 0 for any limited values of co, (u) and co2 (u). To put it differently, with an approximate value of co(u) , the convergence of a sequential experimental pricing scheme or the nearness of a given current secondbest pricing scheme to the firstbest one in its neighborhood, can still be well understood. The following lemma is useful for gauging a practical value of GO . Lemma 11.6 If the demand function d = D(\i), where \i is the generalized travel cost, inclusive of toll charge, if any, is monotonically decreasing and convex, then ©= — < 1  2 £ J
(11.132)
where U and S denotes the user benefit and social welfare for any realized (\l,d) and Ej is the price elasticity of demand at (\i,d), defined by d
d
(< 0 )
(11.133)
Travel Cost Demand function
t Demand
Figure 11.10 Geometric illustration of the proof of Lemma 11.6 Proof. Consider Figure 11.10, where T = td denotes the total travel time cost with t being the travel time, R = ud denotes the toll revenue with u being the toll charge, and CS denotes the consumer surplus (or net user benefit) given by the area under the demand
382
Mathematical and Economic Theory of Road Pricing
curve and above the line segment ab. By definition U U CS + R + T , T T oo= — = = = 1 + < 1 + S UT CS + R CS + R CS S UT CS R since R>0. Because the demand function, d = D(\i), is monotonically decreasing and convex, clearly we have CS > Aabc, where Aabc is the area of the triangle abc with be being the tangent line of the demand curve at point (d,\i). One can easily find that Ah =
—2 D (p.)
where D'(ii) < 0
Hence, i
= 12 since [i = t + u>t with u>0. Thus, eqn. (11.132) is true.
•
Lemma 11.6 shows that co(u) = C/ ue (u)/5' ue (u)>l is upperbounded by (l2£^) for any demand function and any pricing scheme u. The equality of the bound holds for linear demand functions. Since users have to pay their full social cost (private cost plus congestion externality) at the system optimum, it is generally true that (7SO 5 ue (0). These observations together imply that V ;
77SO ^(u)
t/"ef0i 5"e(0)
y
'
For the case of multiple OD pairs, because the user benefit and social welfare are OD pair additive, we have immediately that
If the realized demand at equilibrium is in the inelastic range, then 1Q
(12.1)
Then, the travel time by a commuter from origin to destination is tf + q(t)/s, where q(t)/s is the waiting time in the queue, constant tf is the uncongested travel time from origin to destination (it is referred to as moving time hereinafter).
Dynamic Road Pricing: Bottleneck Models
391
A commuter's total travel cost depends on his/her travel time, schedule delay (timeearly or timelate in arriving at the destination) and road toll. For simplicity, we assume a linear travel cost function a C(0=
+p t   t+tf r
a
_f
q{t)~ +Y s
+ w(0, forte [ts,to] )\
(12.2) + u(t), for te[to,te]
where a is the unit cost of travel time, (3 is the unit cost of schedule delay timeearly, y is the unit cost of schedule delay timelate, t is the official work start time (assumed to be identical to all commuters here), u(t) is the road toll at time t, ts is the departure time at which the queue begins, te is the departure time at which the queue ends, and to is the departure time at which an individual arrives at the destination on time /*, i.e., (12.3)
= t
to+tf
In accordance with empirical results, the relation y > ct > P must hold. In fact, a>P is the necessary condition for the existence and uniqueness of the equilibrium considered. In place of the linear function (122), one may adopt a nonlinear or any other more flexible form to study the bottleneck models. In equilibrium, any commuter is unable to find a departure time to reduce his/her travel cost. In other words, dC(t)/dt = 0. Hence, for the notoll equilibrium (i.e., u(t) = 0) we have
dt
$, for t&[ts,to] ap Y s, for te[t ,t.] o a +y
(12.4)
and a s, for fe [*„*.]
ap a Integrating
(12.5) s, for te[to,te] = 0 leads to:
{tt,)s,
for te[t,,to] (12.6)
The first and last persons face no queue, and have the same travel cost. Note that the exiting rate from the bottleneck during [ts ,te ] is s, hence we have
392
Mathematical and Economic Theory of Road Pricing (12.7) N = s(t.t,)
(12.8)
Combining these two equations and the definition of to in (12.3), we can determine the three unknowns related to time, i.e., , 5 iV . 8 N . 5 JV ts=t x tt; te=t +  x tf\ to=t — x tf Ps
y
s
(x
..... (12.9)
s
where 8 = PY/(P + Y) AS a result, the total travel cost perceived by each auto commuter without tolling during [ts,te] is C = 8— + atf s
(12.10)
which is independent of their departure time. The maximum queue length occurs at time ta: q' = q(to)=N
(12.11)
The total social cost of the system comprises three parts, i.e., total schedule delay cost (t.s.d.c), total queuing time cost (t.q.t.c.) and total moving time cost (t.m.t.c). The sum of these three parts is given below: 5N2 8N2 .. .J6N } 2s 2s ' { s ') ( 12  12 ) (t.s.d.c.) (t.q.t.c.) (t.m.t.c.) It is interesting to observe that both (t.s.d.c.) and (t.q.t.c.) are all independent of a. An explanation of this result is given below: The length of the rush hour is independent of a. Since the first and last individuals to depart experience only schedule delay, identical moving time and no queuing time, and since their trip costs are the same in equilibrium, their schedule delay costs must be identical (see (12.7)). It follows that the start and end of the rush hour are independent of a (see (12.9)). With knowledge of the start and end of the rush hour, (t.s.d.c.) and (t.q.t.c.) can be calculated independently of a. The timeunvarying roaduse toll, i.e., u(t) = constant, will not change the distribution of commuters' departure time choices and their costs relating to moving and waiting times, so from (12.10), the individual cost in equilibrium becomes: C = 8N/s + cttf + constant toll. Dynamic Pricing It is well known that there exists a timevarying roaduse toll (beginning and ending at zero) which can eliminate any queue and then reduce the total social cost. To eliminate queuing while minimizing schedule delay costs, the departure rate from the origin and the exiting rate from the bottleneck must equal the bottleneck capacity s throughout the rush hour [ts,te]
Dynamic Road Pricing: Bottleneck Models
393
where ts and /„ are the first and last departure times, respectively. The first and last commuters face no toll and have the same travel cost, so the timing of the rush hour is the same as in the notolling case, see (12.9); and the individual cost in equilibrium is also the same as in the notolling case, i.e., C = 6N/s + atf. The individual cost remains unchanged because the waiting time cost incurred by a commuter departing at time / is just replaced by an equivalent amount of toll that he/she has to pay at that time. So, for holding the equilibrium of individual cost over the whole rush hour, the toll u(t) at any time should be given bt: 0, for tft
fotte[t',f.]
(12.43)
In the above analyses, the moving time on each route is assumed to be zero, so that all routes are actually used at equilibrium state, i.e., Nr>0, re R. However, if some routes have relatively large moving times, i.e., the equivalent cost to moving time is larger than B[N") where N* is the sum of numbers of commuters on the used routes in equilibrium, they may not be chosen by commuters, even without toll charges on them. We can make an ascending
402
Mathematical and Economic Theory of Road Pricing
order for all routes according to their values of moving times, and try to find such a route that the routes with lower moving times will be more likely used at equilibrium. 123.2 Dynamic Tolls on Partial Routes Now we consider a typical example of the secondbest pricing in a network of parallel routes where some untolled alternatives exist. A simple tworoute network is used to highlight this kind of pricing problems. We assume that an optimal dynamic toll is designed to eliminate any queuing on one route and the alternative route is free of charge. Let t] and t\ be the first and last departure times on the tolled route, respectively; and let t) and t] be the beginning and ending times of queues on the untolled route, respectively. In equilibrium, if there are Nl and N2 commuters choosing the tolled and untolled routes, respectively, we then have,/V, = s, (t]e t]\ and N2 = s2 [t] t]\. We know that for the tolled route, the equilibrium travel cost (including toll) is SNjs,; for the nontolled route, the equilibrium travel cost is 6N2/s2. From the following equilibrium relationships between individual costs and marginal trip benefit, we can solve for 7^, and N2: B(N] + N2) = B(N) = —N^=—N2 s, s2
(12.44)
Then, the four departure times can be solved from equations: Nr=s(trefs) P(f* ~ C )
=
Y(C~O>
and
'" = 1.2. The dynamic toll for the tolled road is
0, for tt\
for ;
te[t\t'1 L
J
(12.45)
for t e\_t',t]\
Note that the total social cost of the tolled route consists of schedule delay costs only, and is given by 8(7V,) /2s, with reference to eqn. (12.14); the total social cost of the untolled route that includes schedule delay and queuing costs is valued as &(N2) js2 with reference to (12.12). Hence, the total social cost of the tworoute network is 5(JV, ) /2sl +5{N2)
/s2.
However, the route use given by (12.44) is not the optimum. The optimal use of the network is achieved when the social welfare is maximized. The social welfare is defined as total trip benefit minus the total social cost given by:
Dynamic Road Pricing: Bottleneck Models
403
N
(12.46) 0
where 8 = PY/(P + Y), NI =SI (tl ~ll),
N 2
= s 2 ( ^ ~1])» 2
partial derivatives of S with respect to t], t\, t] and t
e
and
N=Nl + N2. By setting the
equal to zero, we obtain:
Sl
(12.47)
2S(^)A 2 From (12.47), we can obtain the optimal road use. Besides a dynamic toll is charged on route 1, however, the second equation in (12.47) requires that route 2 should be charged a timeinvarying toll equal to the congestion externality &N2/s2. So far, we have studied various bottleneck pricing schemes using the deterministic queuing theory. A particularly important feature of the timevarying tolls is that queuing is eliminated without increasing user costs. Considering the difficulty of implementing a fine timevarying toll in reality, one may further consider an optimal coarse toll which is a fixed fee paid at the front of the queue over a time interval and is expected to significantly reduce queuing congestion as well.
12.4
BOTTLENECK PRICING AND MODAL SPLIT
In many cities, there usually exists an alternative mass commuting mode such as a railway or subway parallel to the road with a bottleneck. For example, in Hong Kong, there are three crossharbor road tunnels and subways connecting Hong Kong Island and the Kowloon Peninsula. In such a competitive system with parallel transit and highway modes, road pricing can be regarded as a measure for restraining auto use and providing revenue for mass transport improvement. Unlike the road, the generalized cost by a mass transit mode mainly depends on its fare level and service quality, such as station and invehicle crowding, and its generally constant journey time. Obviously, analysis of this bimodal system differs from dealing with the abovementioned parallel road bottleneck system in the sense that the mode heterogeneity should be considered in attracting commuters, but has commonality in the sense that both systems provide commuters two or more substitutable choices. In this section we extend our early analysis by dealing with bottleneck congestion pricing and mode split jointly. In particular, we pay attention to the crowding/discomfort effects of mass transit and travel time difference between the transit and road modes. The logitbased modal split is introduced for the elastic total commuting demand.
404
Mathematical and Economic Theory of Road Pricing
124.1 Competition between Mass Transit and Highway Consider a simplified corridor network which comprises two types of modes to provide transportation service between H (a residential area or home) and W (a workplace). Mode R represents a mass transit (e.g., railway) with an assumed infinite capacity and mode A represents a highway with a single bottleneck which is located at the entering point of highway and has a deterministic capacity of s commuters per unit time. There are N identical commuters who can either travel on the highway by car (auto mode, one person per car) or on a railway by transit from H to W daily each morning. Assume that in equilibrium, there are NA commuters who choose auto mode and NR commuters choosing the mass transit mode, N = NA+NR. Per bottleneck model study presented earlier, the individual travel cost of auto commuters in a departure time choice equilibrium in the absence of toll charge is given by: CA = atA+8^
(12.48) s where tA is the constant moving time by auto mode from H to W. The individual cost of transit commuters is defined as: CR = 6g (NR) +uR + atR
(12.49)
where uR is the transit fare which is assumed to be timeinvariant, 0 is the unit cost of discomfort, g(NR) describes the discomfort experienced by a transit commuter, and tR is the constant travel time by railway. The travel time by railway includes incarriage time and access (egress) time from H to the railway station and from the railway station to W. The waiting time for a train at the station, which depends on the transit service frequency, is assumed to be constant and incorporated into tR. The body congestion function g (NR) in (12.49) takes explicit account of the crowding effects at a railway station and inside a train on the commuters' mode and departure time choice (Lam et al., 1999). To facilitate our analytical investigation, we use a linear crowding function g(NR) = NR . In the case of notolling on the highway bottleneck, equilibrium modal split is characterized by: CR = CA ifNR >0 and NA > 0; CR < CA \fNA = 0; CR> CA if NR = 0
(12.50)
subject to NR + NA = N. The total social cost of the system is TSC = NACA + NR(atR+QNR+c) + F
(12.51)
where c represents the marginal (variable) cost of transit service, F is the fixed cost of the transit mode, consisting of facility costs and fixed operating costs. The marginal cost mainly
Dynamic Road Pricing: Bottleneck Models
405
comprises the daily expenses on labor, fuel and electricity per commuter being burdened by the transit operator. The fare and toll revenues are excluded from the total social cost. Pricing on Transit versus Notolling on Road Assume that the public authority sets the railway fare uR equal to the marginal (variable) cost, average cost and an optimum of minimizing the total social cost, respectively, while the road toll charge is not implemented for certain technical or political reasons. In this case, the railway fare becomes the unique means to adjust traffic flow distribution between the two modes. Our objective is to derive the modal split, evaluate the individual travel cost and the total social cost at an equilibrium state of the system. In the following analyses we further assume that the total number of commuters, N, is sufficiently large to lead to both NA>0 and NR>0. It is easy to show that, when N is less than a certain number, there exists the case where only the auto mode is preferable and the railway would not be worth building. (i)
Setting uR=c
This means that the mass transit operates at a deficit since it can not cover its fixed cost F. We can readily obtain the modal split in equilibrium by using eqn. (11.48) as follows:
(1252)
where superscript ' m' represents solution of 'marginal' cost pricing on the transit mode. From (12.52), we can see that the number of transit commuters is proportional to (tA tR) if tR 0 and then N°A > NA since NR + NA= N. This simply stems from the fact that the transit patronage decreases with increase in fare. It is worthwhile to mention here that the transit patronage is still inversely proportional to the 9 value under average costbased fare policy, but it is not reflected in (12.55) explicitly. To see this, we consider two distinct 0values: O(N"f
407
(& + Qs) in view of (12.52). Combining (12.56)
(12.57), we conclude that N"R > NR. NR remains larger than N"R if the fixed cost F rises to F' as long as F'/F NA . From (12.51) and(12.58), the total social cost is: TSC' = N'A(atA+^]
+ N'R(atll+eN'R+ c)+ F= JatA 1 a
= N\ a.tR+QN
R
K
+ ^ '
(12.59)
+ c + —^ N"
Subtracting (12.54) from (12.59) yields TSCaTSCm=F\—\\NQ{NgN°R)
(12.60)
Since NR < N, the first term of the right hand side is positive, which indicates an increase in total social cost associated with the average costbased fare policy, but the second term represents an reduction in commuter discomfort cost due to N"RLJ*L.\ 2 2 ^ 2 ' 2
Ml 2 
F8 +
*
(12.66)
5 + 05
usA=~LT
(12.67)
where NR and NA are given by (12.52). The differences between (12.66) and (12.55) should be noticed carefully. From (12.66), we know that NA < NA , hence the maximum queue length on the highway is less than that in the case of using marginal costbased fare versus notolling on road. However, the equilibrium individual cost rises as NA < NA atR + QNSR
S
+ c>atR + QNR + c. The comparison between (NR,N A}
yields N'R > NR and and (NR,N°A}
is case
dependent. The equilibrium individual cost under a transit subsidy policy is higher than that under optimal transit fare policy if a(tA tR)\n(N/NmiX) where A^max =10 3 . A larger value of co in the function implies a less sensitivity of demand to marginal trip benefit and thus the final realized demand will be higher. The crowding discomfort function takes the form of g(NR) = 0.05(NR ) + 0.25NR . We focus on the solution sensitivity with respect to parameter ffl, it is of course straightforward to do so with other parameters such as c, 0 and ji. We consider three scenario, 1) we simply set uA = 0 and uR = c in model (12.80) for the nonoptimized pricing case, or the road toll is null and the transit fare equals the variable cost, 2) the firstbest pricing, and 3) the secondbest pricing with road toll equal to zero. The optimal toll (case 2) and fares (cases 2 and 3) are plotted in Figure 12.1.
Dynamic Road Pricing: Bottleneck Models
419
90 —X—Transit fare for the firstbest pricing —O— Transit fare for the secondbest pricing with zero road toll
75 •
—&— Road toll for the firstbest pricing H •a 60 
.X
I §
45
30 •
15 •
0
30
60
90
120
150
180
210
240
270
300
Value of Omega
Figure 12.1 Toll and fare levels of the firstbest and secondbest pricing schemes
1000 00 W C
g
800
A— Fare=15andtoll=0
—X—Firstbest pricing
—O— Secondbest pricing
 AFare=15andtoll=0
• X  The firstbest pricing
 • O  Secondbest pricing
I
Or830
60
90
120 150 180 Value of Omega
210
240
270
300
Figure 12.2 Impact of pricing on total demand and modal split
Figure 12.2 shows the total demand and number of transit commuters generated by the three pricing policies. The first policy with null toll and fare equal to the variable cost generates the largest number of commuters, then followed by the secondbest pricing and the firstbest pricing policies. It is found that the difference in transit patronage among the three cases is
Mathematical and Economic Theory of Road Pricing
420
insignificant, but the total demand exhibits much big disparity and hence is more sensitive to the pricing policies adopted. 0.16
—O— Firstbest pricing —A— pricing with zero road toll
0.00 0
30
60
90
120
150
180
210
240
270
300
Value of Omega
Figure 12.3 Impact of pricing on relative change in total social welfare Figure 12.3 displays the relative change in total social welfare. The relative change is defined as the ratio of the overall welfare gain in comparison with the first case with uA=0 and uR =c in model (12.80). As expected, the firstbest pricing yields higher social welfare improvement than the secondbest pricing.
12.5
SUMMARY
In this chapter, we provided a thorough analysis and review of dynamic road pricing which is of great importance due to everincreasing traffic congestion. Albeit limited to a simple spatial setting with either a single or parallel bottleneck routes, the analysis presented in this chapter does offer interesting insights into the dynamic nature of traffic congestion and congestion pricing. Our analysis added realism by including user heterogeneity, demand elasticity and mode split into the dynamic bottleneck pricing models. In the case of a single bottleneck, analytical solutions of timevarying tolls and departure rates can be obtained. Demand elasticity can be easily incorporated in both the single and parallel bottleneck models. Further meaningful results are obtained from a competitive twomode transportation system, in which an alternative mass commuting mode, such as railway or subway, exists in parallel to the road with a bottleneck.
Dynamic Road Pricing: Bottleneck Models
421
The models and analyses in this chapter explored some essential elements of dynamic or timevarying pricing and provided important references for further investigations, but they are limited in the spatial context. In the subsequent and also the last chapter, we make one more significant step forward by developing a numerically tractable dynamic road pricing model with bottlenecks in a general road network.
12.6
SOURCES AND NOTES
There is a vast body of literature geared toward bottleneck congestion pricing mainly by wellknown economists. Vickrey (1969), Braid (1989) and Araott et al. (1993a,b) and others presented analysis of a pure road bottleneck. Henderson (1974), Agnew (1977) and Chu (1995) determined timevarying tolls by applying a speedflow function for a c ongested road. BenAkiva et al. (1986), Arnott et al. (1993a) and Braid (1989) and Mun (1994) conducted equilibrium analyses of bottlenecks with elastic demand. Laih (1994) examined queuing at a bottleneck with single and multistep tolls. Henderson (1974, 1981) incorporated two groups of commuters differing in value of time and schedule delay penalty into a dynamic model. Cohen (1987), Newbery (1988) and Evans (1992) investigated the welfare effects of tolls on two groups of commuters. Arnott et al. (1987, 1988, 1992, 1994) studied various aspects of the bottleneck problem with general groupspecific heterogeneous commuters. Newell (1987) dealt with a general case where commuters are continuously distributed in their work start times, costs of travel time and costs of schedule delay. Yang and Huang (1997) and Huang and Yang (1996, 1999) applied the optimal control theory to develop a general timevarying roaduse pricing model for a statevarying exit capacity bottleneck or Hypercongesion (Herman, 1982) with elastic demand. Bottleneck road congestion pricing with a competing railroad service were investigated by Tabuchi (1993), Huang (2000) and Huang (2002) and Danielis and Marcucci (2002). An excellent early review was given by Arnott et al. (1998). For mo st recent developments of the bottleneck models, readers may refer to Van der Zijpp and Koolstra (2002), de Palma and Lindsey (2002), Marvin (2003), Verhoef (2003), Konishi (2004), Lindsey (2004) and Zhang et al. (2005). For more details about Section 12.2, readers may consult works by Arnott et al. (1990a, 1990b, 1993, 1994, 1995), Braid (1986), Yang and Huang (1997), Lindsey and Verhoef (2001). Relevant studies in Section 12.3 can be found in Arnott et al. (1990a, 1992), Braid (1996) and Huang and Yang (1996). Section 12.4 is based on Tabuchi (1993), Huang and Yang (1999), Huang (2000) and Huang (2002). Readers can also consult Huang et al. (2000) for extension to the logitbased models in carpooling and pricing problems. Empirical results about Y>oc>P were first reported in Small (1982). In fact, cc>P is also the condition required for existence and uniqueness of notoll equilibrium (Smith, 1984 and Daganzo,
This Page is Intentionally Left Blank
13 DYNAMIC ROAD PRICING: GENERAL NETWORK MODELS* 13.1
INTRODUCTION
In Chapter 11, we examined dynamic pricing models with a single bottleneck by seeking timevarying tolls to remove queues with or without elastic demand, and investigating the welfare effects of tolls on different groups of commuters. Extensions to parallel bottlenecks or a bottleneck parallel to a mass transit line were made as well. In this chapter, we make one major step ahead by dealing jointly with the time and space dimensions of road pricing in a general network with bottleneck congestion. The problem becomes much more difficult because of the complexions of dynamic traffic assignment and the network structure, and indeed only very limited studies of dynamic pricing in general networks are available in the literature. This Chapter deals with the modeling of peakperiod congestion and optimal pricing in a queuing network with elastic travel demand. The approach employed here is a combined application of the spacetime expanded network (STEN) representation of timevarying traffic flow and the conventional network equilibrium modeling techniques. Given the elastic demand function for trips between each OriginDestination (OD) pair and the schedule delay cost associated with each destination (workplace), the departure time and route choice of commuters and the optimal variable tolls of bottlenecks will be determined jointly by solving a system optimization problem over the STEN. The STEN approach can deal with a general queuing network with elastic demand, and allow for treatment of commuter heterogeneity in their work start time and schedule delay cost, and hence make a significant advance over the previous simple bottleneck models of peakperiod congestion. This chapter is organized as follows. We give a general description of the departure time and route choice problem with schedule delay costs, and then present a spacetime expanded network where commuter's behavioral decision on departure time and route choice can be described explicitly. We develop a dynamic traffic flow model, followed by the derivation of * Notation is used in this chapter independently of the previous Chapters 110
424
Mathematical and Economic Theory of Road Pricing
the optimal congestion tolls for each link over the spacetime expanded network, which aims to maximize the social benefit over the whole study horizon. Simple and general numerical examples are provided to evidence the potential application of the dynamic traffic congestion and pricing network model in practice.
13.2
PROBLEM DESCRIPTION
The notation used in this chapter is independent of previous chapters. Let G = (N,A) be a network with P being the set of origin nodes and Q the set of destination nodes, a typical origin and destination node is denoted by p and q respectively. Let Rpq denote the set of routes between OriginDestination (OD) pair (p,q). Let the overall time horizon of study [0,T] be divided into equally spaced discrete time periods sequentially numbered t = l,2,,T. Suppose the time horizon [0,T] is chosen to be large enough such that it does not affect the results of analysis (i.e., it covers the whole time period within which commuter's departure time may be adjusted and all commuters arrive at their destinations). Let dpq (t) be the demand for travel from origin p to destination q in time period t, and dpq be the total demand over the whole study horizon between OD pair (p,q). Obviously, T
PI = 2J"H(0>
PE °> # e Q
(i3.i)
(=0
Here we suppose commuters have a free choice of their departure time from home, and thus, dn (?) are endogenous variables in our model. Furthermore, in the medium to long run, demand may be reasonably elastic, since commuters can change residences, jobs, trip frequency, transport mode and so on to avoid bottleneck congestion and tolls. We thus suppose the total demand between each OD pair over the whole time horizon is determined by a demand function: dpq = Dpqhx.\, pe P, qeQ
(13.2)
where ]i.pq is the generalized cost for each commuter which includes travel and waiting time cost, schedule delay cost and toll at the equilibrium state of the system. The assumption that the total demand over the whole study horizon [0, T] is elastic with respect to the equilibrium generalized cost, is in accordance with the conventional treatment of bottleneck analysis. This assumption of demand function could allow adjustment of trip departure time over the entire time study horizon, which is a significant characteristic different from the instantaneous demand function adopted by previous researchers. Instantaneous demand function is simply
Dynamic Road Pricing: Network Models
425
defined as a decreasing function of the travel cost associated with each instant of departure. This type of demand function is difficult to establish in practice and is difficult to capture the adjustment of departure time. Because travel demand is time dependent, traffic flow and travel time along a route will also be timedependent. Let Trpq (t) be the travel time incurred by a commuter who departs from origin p , in time period t, and travels on route r e Rpq, to destination q . Trpq(t) also depends on the demand and supply characteristics of the queuing network. Suppose the destination q, represents the workplace of a group of homogeneous autocommuters who have the same preferred arrival time or common work start time. This is specified by an interval \tq  A ? , f * + A ? l c [ 0 , T] where (O
Link 2
^
Base Network
Expansion of Base Network Figure 13.1 A simple network and its spacetime expansion 133.2 Static Temporal Expanded Network According to the above assumptions, a vehicle that enters a link a e A at period t, must travel during t\ periods of time, and then it may either exit the link or join the exit queue of the link at period t +1\; in the later case, after one period of time, it will have again a choice of exiting the link or joining the exit queue of the link corresponding to the time period t + t°+l, and the process is repeated until the vehicle exits the link. For a given base traffic network G = (N,A),
this process may be described as path following in the spacetime
network to be constructed in the following ways. Nodes:
Each transshipment node nk (ksN)
of the base network is expanded
to T + 1 nodes n'k, t=Q,\,,T , of the expanded network.
428 Links:
Mathematical and Economic Theory of Road Pricing For each link a = (n,, n y ) e A , we construct, for each t = 0,l,,T  one node l'a one link [n'^,l'\
if tt°a>0
one link (l'a,rij)  one link (l'a,C) if t + lsRpq,pBP,qsQ (13.9) XA
where 5or = 1 if link a is in path r and 0 otherwise, Rpq is the set of paths between OD pair (p,q) in the STEN. The flow on a given link (including queuing links) is the sum of all those assigned to the paths passing on this link; this is expressed mathematically as
v, = S E U , " ^
(1310)
The total social cost incurred by all commuters is thus given by ^Vaca
(13.11)
It is assumed that the optimal use of the network capacity is achieved when the net economic benefit (total commuters benefit minus total social cost) is maximized. Therefore, the problem can be formulated as:
432
Mathematical and Economic Theory of Road Pricing
J
5> 0 v a
(1312)
subject to i,dM(t)=dp,,peP,qeQ
(13.13)
E / , =dPq(t)'PeP'
(13.14)
? e 6, / = 0,l,,7*
v 6 < 5 a ,6€4 e ,flE/(
(13.15)
fr>O,reRpq,peP,qeQ
(13.16)
where ^
is the set of paths available between OD pair (p, q) during time period t in the
STEN, and A^ is the set of exit links in the STEN corresponding to link a e A of the base network. 13.4.2 Optimality Conditions To derive the necessary conditions for an optimal solution of the above optimization problem, we construct the Lagrangian function:
(p,q)
where cpw and \xpq (?) are the Lagrange multipliers associated with the flow conservation constraints (13.13) and (13.14), respectively, ub is the Lagrange multiplier associated with the link exit capacity constraints (13.15), and va is again defined by (13.10). The Lagrangian should be minimized with respect to the flow variables (and maximized with respect to the dual variables) and subject to the following constraints: fr>0,rsRpq,peP,qeQ
(13.18)
dpq(t)>0,peP,qeQ,t=0,l,,T
(13.19)
dn>0,peP,qeQ (13.20) The firstorder conditions for a stationary point of the program (13.12)(13.16) with constraints (13.18)( 13.20) are given below:
crJt)yLpq(t)>O,reR'pq,peP,qeQ,t=Q,\,,T
(13.22)
Dynamic Road Pricing: Network Models
433
ut(vbsa)=O,be£,aeA
(13.23)
dpq(t)[\ip,{t)