Optimal Space Flight Navigation

Control Engineering
Ashish Tewari
Optimal
Space Flight
Navigation
An Analytical Approach
Control Engineering
Series Editor
William S. Levine
Department of Electrical and Computer Engineering
University of Maryland
College Park, MD
USA
Editorial Advisory Board

Richard Braatz Mark Spong
Massachusetts Institute of Technology University of Texas at Dallas
Cambridge, MA Dallas, TX
USA USA
Graham Goodwin Maarten Steinbuch
University of Newcastle Technische Universiteit Eindhoven
Australia Eindhoven, The Netherlands
Davor Hrovat Mathukumalli Vidyasagar
Ford Motor Company University of Texas at Dallas
Dearborn, MI Dallas, TX
USA USA
Zongli Lin Yutaka Yamamoto
University of Virginia Kyoto University
Charlottesville, VA Kyoto, Japan
USA
More information about this series at http://www.springer.com/series/4988

Ashish Tewari
Optimal Space Flight

Navigation
An Analytical Approach
Ashish Tewari
Department of Aerospace Engineering
Indian Institute of Technology, Kanpur
IIT-Kanpur, India
ISSN 2373-7719 ISSN 2373-7727 (electronic)

Control Engineering
ISBN 978-3-030-03788-8 ISBN 978-3-030-03789-5 (eBook)
https://doi.org/10.1007/978-3-030-03789-5
Library of Congress Control Number: 2018961705
Mathematics Subject Classification: 70F05, 70F07, 70F10, 49N05, 49N25, 49J15, 49K15, 49J30, 37N05,
34B10
© Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This book is published under the imprint Birkhäuser, www.birkhauser-science.com by the registered
company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
This book is primarily written to consolidate optimal space flight navigation

theory as a discipline. There are several excellent textbooks on optimal control
theory and equally good books on orbital space mechanics, but the application
of optimal control theory to orbital mechanics is mainly confined to research
articles. Over the last several decades, the research literature on optimal space
flight navigation has become voluminous and has fragmented into many esoteric
subdisciplines. However, a monograph which organizes this content and vitalizes it
for contemporary application is hitherto lacking.
Optimal space flight navigation is a problem of practical interest and has been
so ever since humans imagined journeying into space. While visionaries such as
Tsiolkovsky and Goddard helped in taking the first steps by developing rocketry, it
required enthusiasts to devise fuel-optimal trajectories for practical space missions.
The vision of Walter Hohmann, a civil engineer, revealed in his 1925 book Die
Erreichbarkeit der Himmelskörper (The Accessibility of Celestial Bodies) is one
of the most useful and commonly applied space flight maneuvers—the Hohmann
transfer. As recently as in 2014, the Hohmann transfer was successfully applied
to efficiently send India’s first mission to Mars. While arising as an elegant
idea in Hohmann’s mind, the optimality of the Hohmann transfer can now be
rigorously proved in various ways, including by optimal control theory. Similarly,
the straightforward and insightful application by D.F. Lawden of the erstwhile
nascent optimal control theory to the space navigation problem produced in 1963 in
the monograph entitled Optimal Trajectories for Space Navigation has guided many
researchers since then. The first footsteps on the Moon would not have appeared
if NASA’s engineers (including Richard Battin who passed away in 2014) had
not devised simple but efficient techniques to guide the lunar module on a flat
thrusting trajectory to a safe landing. Battin in his treatise An Introduction to the
Mathematics and Methods of Astrodynamics reminisces about how he came up
with the velocity-to-be-gained and cross-product steering as simple but practical
guidance strategies for space flight in the era of primitive computer technology. He
writes of the enormous buildings required to house the computer mainframes of
the day with only very modest computing power. Perhaps the total computational
v
vi Preface
resources present on board Apollo 11 could be surpassed by those of today’s pocket

calculator. The feat performed by those pioneering guidance engineers can only
be appreciated by the complexity of their task—to send people to the Moon and
to bring them successfully back to the Earth. The design and analysis of practical
lunar and interplanetary trajectories could be conducted in the analytical framework
established by Victor Szebehely in his 1967 masterpiece, Theory of Orbits: The
Restricted Problem of Three Bodies.
Rather than highlighting the computational (and often unrealizable) solutions of
mathematically complex problems, this book emphasizes the analytical approach
to optimal space navigation. The need for simplicity and practicability of guidance
methods is to be contrasted with the unfortunate emphasis which has been placed
these days on increasing the complexity of control schemes. Multilevel iterative
algorithms must be solved in real-time computations and therefore have unresolved
convergence issues which prevent applications to actual missions. It should thus be
asked: Is more sophistication always better, and does it lead to any advantage in
the practical sense? I think a simpler guidance solution must be sought, wherever
possible, due to its reliable implementation in a realistic situation, which has been
amply demonstrated by the analytical guidance methods. That is why even the latest
space flight missions are based upon analytical techniques.
The complex interplanetary and asteroid (or comet) flyby and rendezvous
missions are analytically derived from the application of optimal control theory
to multi-body dynamics, such as the optimal transfers between halo and quasi-
halo orbits and Lissajous trajectories using the stable and unstable manifolds of
the restricted three-body problem. Studying such optimal paths often imparts new
insights into the problem of multi-body transfers, which would be lacking in a purely
numerical search for suboptimal solutions. In a similar vein, the problem of orbiting
and landing on an irregularly shaped body with an uncertain gravity field (such as an
asteroid or a comet) can be solved out by time-optimal and fuel-optimal methods,
as has been demonstrated in many practical missions.
The thought of writing this book arose from the course AE-777 (Optimal Space
Flight Control), which I have taught for the past several years and for which a single
textbook containing all the relevant material was unavailable. Therefore, the book is
designed to be used as a textbook in a course on optimal space flight at the graduate
and senior undergraduate levels. The first three chapters give a basic introduction
to the topic, namely, optimal control theory (Chaps. 1 and 2) and orbital mechanics
with impulsive transfers (Chap. 3). Chapter 4 covers optimal transfers in a spherical
gravity field with continuous, unconstrained acceleration inputs, whereas Chap. 5
extends the treatment to trajectory optimization with bounded acceleration inputs.
Finally, Chap. 6 introduces the reader to the advanced topics of optimal transfer
in time-varying gravity fields, including multi-body transfers. Unfortunately, due
to the vast scope of these topics, it is not possible to do them full justice in
an introductory text. Hence, references to research articles are provided for the
advanced topics. The end-of-chapter exercises test the understanding of the basic
concepts, whereas several references are provided for undertaking research and
for advanced applications. A basic knowledge of control systems and dynamics is
Preface vii
assumed of the reader, which can be supplemented by textbooks on these topics,

such as those previously published by me (e.g., Modern Control Design, and
Atmospheric and Space Flight Dynamics, Automatic Control of Atmospheric and
Space Flight Vehicles).
It is a pleasure to offer this manuscript for publication with Birkhäuser. I would
like to thank the editorial and production staff at Birkhäuser for their invaluable
assistance, especially Bill Levine, Benjamin Levitt, Samuel DiBella, and Christo-
pher Tominich. I would finally like to thank my wife Prachi and daughter Manya for
their patience during the preparation of this manuscript.
Kanpur, India Ashish Tewari

September 2018
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Space Vehicle Guidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Analytical Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Optimization of Static Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Static Equality Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Inequality Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Dynamic Equality Constraints and Unbounded Inputs . . . . . . . . . . . . . . 20
2.4 Special Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.1 Fixed Terminal Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.2 Free Terminal Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Sufficient Condition for Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Pontryagin’s Minimum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.7 Hamilton–Jacobi–Bellman Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.8 Time-Invariant Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.9 Linear Systems with Quadratic Cost Functions . . . . . . . . . . . . . . . . . . . . . . 35
2.10 Illustrative Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.11 Singular Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.11.1 Generalized Legendre–Clebsch Necessary Condition . . . . . 58
2.11.2 Jacobson’s Necessary Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.12 Numerical Solution Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.12.1 Shooting Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.12.2 Collocation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3 Orbital Mechanics and Impulsive Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2 Keplerian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2.1 Reference Frames of Keplerian Motion. . . . . . . . . . . . . . . . . . . . . 80
ix
x Contents
3.2.2 Time Equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.2.3 Lagrange’s Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.3 Impulsive Orbital Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.3.1 Minimum Energy Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.4 Lambert’s Transfer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.4.1 Stumpff Function Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.4.2 Hypergeometric Function Method . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.5 Optimal Impulsive Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.5.1 Coasting Arc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.5.2 Hohmann Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.5.3 Outer Bi-elliptical Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4 Two-Body Maneuvers with Unbounded Continuous Inputs . . . . . . . . . . . 115
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.2 A Motivating Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.3 Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.4 Optimal Low-Thrust Orbital Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.4.1 Coplanar Orbital Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.4.2 Plane Change Maneuver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.4.3 General Orbital Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.5 Variational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.6 Optimal Regulation of Circular Orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.6.1 Coplanar Regulation with Radial Thrust. . . . . . . . . . . . . . . . . . . . 128
4.6.2 Coplanar Regulation with Tangential Thrust . . . . . . . . . . . . . . . 131
4.7 General Orbital Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.8 Basic Guidance with Continuous Inputs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.9 Line-of-Sight Guidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.10 Cross-Product Steering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.11 Energy-Optimal Guidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.12 Hill–Clohessy–Wiltshire Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5 Optimal Maneuvers with Bounded Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.2 Optimal Thrust Direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.2.1 Constant Acceleration Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
5.2.2 Bounded Exhaust Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
5.3 Time-Invariant Gravity Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.4 Null-Thrust Arc in Central Gravity Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.4.1 Inverse-Square Gravity Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
5.5 Intermediate-Thrust Arc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
5.6 Lawden’s Spiral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
5.7 Powered Arcs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Contents xi
5.8 Linearization Relative to a Circular Orbit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

5.8.1 Out-of-Plane Rendezvous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
5.8.2 Coplanar Rendezvous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
6 Flight in Non-spherical Gravity Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
6.2 Gravity Field of a Non-spherical Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
6.2.1 Gravity of an Axisymmetric Planet . . . . . . . . . . . . . . . . . . . . . . . . . 201
6.2.2 Gravity Field of an Oblate Planet . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6.2.3 Gravity of an Irregular Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.3 Circular Restricted Three-Body Problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.3.1 Lagrangian Points and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
6.3.2 Hamiltonian Formulation and Jacobi’s Integral . . . . . . . . . . . . 216
6.4 Special Three-Body Trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
6.4.1 Perturbed Orbits About a Primary. . . . . . . . . . . . . . . . . . . . . . . . . . . 219
6.4.2 Free-Return Trajectories Between the Primaries . . . . . . . . . . . 220
6.5 Optimal Low-Thrust Three-Body Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
6.6 Missions Around Collinear Lagrangian Points . . . . . . . . . . . . . . . . . . . . . . . 232
6.6.1 Motion About the Collinear Points. . . . . . . . . . . . . . . . . . . . . . . . . . 234
6.6.2 Lindstedt–Poincaré Method for Halo Orbits. . . . . . . . . . . . . . . . 237
6.7 Numerical Computation of Halo Orbits and Manifolds . . . . . . . . . . . . . 242
6.7.1 Monodromy Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
6.7.2 Stable and Unstable Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
6.8 Optimal Station Keeping around Collinear Points . . . . . . . . . . . . . . . . . . . 249
6.8.1 Impulsive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.8.2 Continuous Thrust Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Answers to Selected Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Chapter 1
Introduction
Navigating a space vehicle consists of first determining and then steering a desired
path under the gravitational influence of several large bodies. This results in two
problems to be solved: (a) determination of a suitable flight path which takes the
vehicle from an initial to a final state, and (b) guiding the spacecraft such that it
always remains close to the desired optimal trajectory. Either of these problems
could lead to infinitely many solutions, each of which can have a different energy
expenditure and a time of flight. However, a unique solution is obtained when either
the energy or the time (or any other performance measure) of flight is minimized for
a given set of initial and final conditions, and with realistic physical constraints.
Such an approach—called the optimal control—thus provides a closure to the
spacecraft navigation problem, which would otherwise remain both arbitrary and
intractable.
The application of the systematic framework supplied by the optimal control
theory applied to practical problems in space flight navigation is the objective of
this monograph.
1.1 Optimal Control
A general control problem is solved by determining the control history for changing
the state of a given system as required. Such a determination can be carried out
only if a mathematical model of the actual system is available. Deriving a suitable
math model in terms of differential, integro-differential, and algebraic equations
requires various idealizations and simplifying assumptions to be applied to the
actual system. The solution to such equations subject to given initial conditions and
prescribed control history must be unique if it has to be applied for controlling the
actual system. The cases where either the governing equations are uncertain or have
non-unique solutions to a prescribed control history and a set of initial conditions
© Springer Nature Switzerland AG 2019 1

A. Tewari, Optimal Space Flight Navigation, Control Engineering,
https://doi.org/10.1007/978-3-030-03789-5_1
2 1 Introduction
are termed non-deterministic and must be handled by stochastic (probabilistic)

techniques. For our present purposes, we shall assume that the mathematical model
is a deterministic one. Furthermore, the system of interest in spacecraft navigation is
completely described at any instant by a finite number of variables x(t) : R → Rn×1
called the state variables, where n is the finite order of the system. The set of
state variables is said to uniquely determine its state. An additional set of variables
u(t) : R → Rm×1 called the control variables is applied externally to the system
in order to manipulate its state. The state vector x(t) and the control input vector
u(t) are mutually related by a set of ordinary differential equations governing the
system.
Consider a system governed by the following set of first-order ordinary differen-
tial equations called the state equation:
ẋ = f (x, u, t) (1.1)
where x(t) ∈ Rn×1 is the state vector and u(t) ∈ Rm×1 is the control vector, subject
to the initial condition
x(t0 ) = x0 (1.2)
The optimal control problem refers to finding a set of control variables, u(t), which
minimizes a performance index,
tf
J = φ[x(tf ), tf ] + L(x, u, t)dt (1.3)
t0
such that the dynamic equality constraint given by the state equation is satisfied,
while taking the terminal state, x(tf ), to a specified set, x(tf ) ∈ Xf ⊂ Rn×1 .
Both the Lagrangian, L(x, u), representing a cost on the transient behavior, and the
terminal cost function, φ[x(tf ), tf ], denoting the penalty at the terminal time, tf , are
continuously differentiable functions up to an indefinite order. The control interval,
t ∈ (t0 , tf ), could be either open or closed, depending upon whether the terminal
time, tf , is specified a priori. There could be additional inequality constraints to be
specified for the state and control variables, given by
g(x, u, t) ≤ 0 (1.4)
It is clear from the Weierstrass’ extreme value theorem that a continuous, real-valued
functional, J (x, u), would possess both a maximum and a minimum with respect
to u either inside or on the boundary of any bounded region of space, u ∈ U ⊂
Rm×1 . Thus by imposing constraints on the state and control variables in the form
of equality and inequality constraints, as well as the initial and terminal boundary
conditions, the existence of a minimum (or maximum) point of J relative to u is
guaranteed. The optimal control theory is a mathematical framework for deriving
the necessary and sufficient conditions of the existence of a minimum (or maximum)
point, from which the optimizing control variables could be determined uniquely.
1.2 Space Vehicle Guidance 3
The solution to the optimal control problem can be derived either by analytically
through the framework of calculus of variations or by a direct transcription method
where the solution is directly sought by nonlinear programming methods. The direct
solution methods involve a “blind” search over a feasible space by computationally
intensive schemes, which either may or may not produce a solution. The latter
invariably focus on sophisticated algorithms, without offering any insight into the
physical problem being solved. This book focuses on the analytical methods of
investigating the extremal trajectories, which drive the state from x0 to x(tf ) ∈
Xf while minimizing (or maximizing) J relative to u, subject to the constraints,
Eqs. (1.1) and (1.4). This optimal control theoretic method has been developed
over the last century [15, 19, 68]. The formulation of the analytical method results
in a two-point boundary value problem which must be solved iteratively, and
has a guaranteed solution when the sufficient conditions of optimality are met.
The analytical derivation of the necessary and sufficient conditions of optimality
produces an insight into the actual problem being solved, which can be employed
to guide the direct search methods. Hence, the two approaches can be combined to
yield an efficient procedure. For the sake of brevity, we shall not discuss the direct
search methods any further, and instead refer the reader to the excellent survey paper
by Betts [14].
1.2 Space Vehicle Guidance
All spacecraft require a precise control of their position for a successful mission.
The flight control tasks can be broadly classified into two distinct categories:
1. Navigation (or guidance) consisting of a control of the spatial position, r(t) ∈
R3×1 , and linear velocity, v(t) ∈ R3×1 (i.e., translation) of the craft’s center of
mass relative to a fixed reference frame.
2. Attitude control, which involves a control of the rigid vehicle’s angular orienta-
tion (or attitude) represented by the kinematic parameters, q(t) ∈ R4×1 , called
the quaternion, and angular velocity, ω(t) ∈ R3×1 , about the center of mass,
relative to a non-rotating frame of reference.
The reference frame for the navigation problem is regarded as being non-
accelerating (or inertial), with its origin conveniently chosen to be at the center of
a large body (such as a planet, the Moon, or the Sun). During the time of flight, the
motion of the spacecraft is resolved relative to that of the body using the translational
dynamic variables, r(t), v(t). The space vehicle’s translational dynamics relative
to the large body is referred to as its orbital motion, and can be controlled by
applying external thrust via rocket engines. In contrast, the reference frame for
attitude control has its origin at the spacecraft’s center of mass (thus translates
with it), and has its axes fixed with respect to the distant stars. The angular velocity,
ω(t), and the attitude, q(t), of the rigid spacecraft are measured relative to this
reference frame, and are resolved in a frame called the body axes, which is rigidly
4 1 Introduction
attached to the spacecraft. The attitude is controlled by externally applied torque

inputs and/or internally variable angular momentum via rotors.
The desired attitude of the spacecraft is dependent on the vehicle’s spatial
position in most space missions, which invariably involve pointing one of the body
axes in a particular direction. Furthermore, the vehicle’s rotation about its own
center of mass causes the thrust of its engines to produce time-varying acceleration
components in the inertial space, thereby affecting the translational motion. Thus
the two motions are inherently coupled. However, since the characteristic time scale
of translational dynamics is typically much larger than that of attitude dynamics,
the two control problems are addressed separately by dedicated systems. Due to its
much shorter time scale, the attitude controller acts as a slave (or actuator) for the
navigation system. Another control subsystem that is subservient to the navigational
system is the spacecraft’s propulsion system, whose time scale is much shorter
than even the attitude subsystem. The propulsion system generates a thrust vector,
T (t) ∈ R3×1 , from the engines, as demanded by the navigational system.
A spacecraft of mass, m(t), has the following equations of translational motion:
ṙ = v (1.5)
v̇ = g(r, t) + T /m (1.6)
where g(r, t) represents the acceleration due to gravity, and the mass variation is
related to the net mass exhaust rate, β(t), of all rocket engines as follows:
ṁ = −β (1.7)
If the spacecraft is assumed to be rigid with a constant inertia tensor, J ∈ R3×3 ,

its rotational (or attitude) dynamics equations resolved in the body axes are the
following:
q̇ = A(ω) (1.8)
ω̇ = J−1 τ − J−1 (ω × Jω) (1.9)
where A(.) : R3×1 → R4×1 is a nonlinear attitude kinematics functional and τ (t) ∈
R3×1 is the applied torque input. The thrust vector, T (t), in the inertial space is
related to a changing attitude as follows:
T = C(q)T̄ (1.10)
where T̄ is the thrust vector resolved in the body axes, and C(.) : R4×1 → R3×3
is the orthogonal rotation matrix representing the coordinate transformation from
the body axes to the inertial space, with the property det (C) = 1. Therefore,
| T (t) |=| T̄ (t) |. The spacecraft’s engines can be rotated relative to the body axes
1.2 Space Vehicle Guidance 5
r(t0)=r0 , v(t0)=v0
F[r(tf), v(tf), tf]=0
Navigational
Feedforward Attitude
Controller Feedforward
Controller q(t),
rd(t), vd(t) qd(t),
Navigational ωd(t) Attitude τ (t) Spacecraft ω (t) Spacecraft r(t), v(t)
Feedback Feedback Attitude Orbital
Controller Controller Dynamics Dynamics
− −
Td(t) Spacecraft (t)

Propulsion
System
Fig. 1.1 Spacecraft control system comprising the navigational, propulsion, and attitude
subsystems
using two gimbal angles per engine, and can also be throttled by varying the mass
exhaust rate, β, thereby producing a variable, T̄ , as required.
The inter-relationship of the various spacecraft control systems is schematically
depicted in Fig. 1.1, wherein the outer navigation (or guidance) loop takes reference
inputs, rd (t), vd (t), from either a ground station or an on-board flight computer
(called the navigational feedforward controller). The navigational feedforward
controller generates the trajectory waypoints, rd (t), vd (t), to meet the boundary
conditions specified by the initial conditions, r(t0 ) = r0 , v(t0 ) = v0 , and a terminal
condition, F [r(tf ), v(tf ), tf ] = 0, where F (.) : R3×1 ×R3×1 → R. A navigational
feedback controller compares the actual trajectory, r(t), v(t), with the specific
waypoints to issue corrective control inputs. The navigational control inputs are the
commands issued to the propulsion subsystem (its internal structure is not shown
here) for producing a desired thrust vector, Td (t), and to the attitude control system
to rotate the vehicle for achieving the desired rotational state, qd (t), ωd (t), at any
time, t. A fully automated (autonomous) spacecraft has navigational, propulsion,
and attitude control subsystems functioning without any manual intervention.
Although the three control systems work in tandem, they can be designed and
studied independently of each other due to their much different time scales. For this
reason, the present text considers only the design and analysis of optimal navigation
systems, and assumes the tasks requiring the generation of the required thrust as well
as pointing of the vehicle in desired directions are instantly achieved by separate
propulsion and attitude control systems, respectively. Such an assumption involves
neglecting the time scales of engine control system dynamics and attitude motion
in comparison with that of the translational dynamics. Figure 1.2 is the schematic
diagram of the navigational system assuming instantaneous vehiclerotation and
6 1 Introduction
r(t0)=r0 , v(t0)=v0
F[r(tf), v(tf), tf]=0
Navigational rd(t), vd(t) Navigational a(t) Spacecraft r(t), v(t)

Feedforward Feedback Orbital
Controller Controller Dynamics
-
Fig. 1.2 Spacecraft navigational system with propulsion and attitude dynamics neglected
thrust vector generation, such that T̄ (t) = Td (t), q(t) = qd (t), and ω(t) = ωd (t) at
any given time, t. The acceleration input, a(t) = T (t)/m(t) ∈ R3×1 , which actually
arises out of T̄ (t), q(t), ω(t), β(t), is thus assumed to be instantly variable, and
hence regarded as the sole control input vector for the navigation problem. Therefore
we have no further need to model the propulsion and the attitude control subsystems
in order to study the navigational system.
Chapter 2
Analytical Optimal Control
2.1 Introduction
Optimization refers to the process of achieving the best possible result (objective),
given the circumstances (constraints). When applied to determine a control strategy
for fulfilling a desired task, such an optimization is called optimal control. Although
optimization has been studied for millennia, optimal control received a fillip in
the 1950s with the publication of two treatises, one by Richard Bellman [13], and
the other by Pontryagin [68]. Bellman essentially introduced the terminology we
find in usage today and proposed the principle that a trajectory which minimizes
an objective for a part of it continues to minimize the same objective through the
remainder of the trajectory. Pontryagin’s minimum principle on the other hand states
that any trajectory which fulfills the necessary (but not sufficient) conditions for the
minimization of an objective function will always have a value of the objective
function greater than that on the optimal trajectory. Both of these observations
may seem trivial, but are the basic foundations of the modern optimal control
theory. For example, a small but arbitrary variation of the control input from its
optimal history can be used to derive the variations in the state variables that are
required for satisfying the conditions of optimality. This chapter briefly presents
such a variational method for obtaining the necessary and sufficient conditions for
minimizing a specific objective function along a trajectory which is the solution to
a dynamic constraint equation.
The dynamic optimization problem involves determining a solution to the
governing equations of a dynamic system, which satisfy the necessary conditions
for minimizing a prescribed objective function, subject to boundary conditions and
interior constraints. Such a trajectory (and the corresponding control history) is said
to be extremal, and all quantities evaluated on the extremal trajectory are indicated
by the superscript ∗ in this book. While an extremal trajectory satisfies the necessary
conditions of optimality, it may or may not be the optimal one. In order to be the
optimal trajectory, an extremal trajectory must satisfy additional conditions, called

https://doi.org/10.1007/978-3-030-03789-5_2
8 2 Analytical Optimal Control
the sufficient conditions for optimality. We will briefly consider in this chapter
how an extremal trajectory can be determined, and then tested for optimality.
The process of analytically deriving the necessary and sufficient conditions for
optimization involves the calculus of variations. An alternative method can also be
employed where computational techniques directly search for an optimal trajectory
without first deriving the conditions for optimality. Such direct search techniques are
computationally intensive and fraught with convergence issues, hence their study
is a separate topic in itself. Our focus in this book is therefore on the analytical
methods of dynamic optimization, whose solution can be derived and implemented
much more easily.
Notation is kept simple in this chapter due to the many variables involved.
Whereas in the later chapters, we will distinguish between scalars and vectors with
(or without) the use of bold symbols, in the present chapter, all the variables are
in normal font but their dimensions are specified. For example, a vector spanning
(n × 1) real space is denoted by x ∈ Rn×1 .
2.2 Optimization of Static Systems
In order to motivate the reader, it is necessary to introduce the concepts of

optimization of a system that does not depend upon time. Such an optimization
consists of time-invariant objective functions and constraints, hence it is considered
to be a static problem. To generalize the theory to dynamic systems, time is
merely added to the set of independent variables. Optimization refers to either the
minimization or maximization of a given function. In the book, we shall pose the
optimization problem as a minimization problem without any loss of generality,
since changing the sign of the given objective function can easily transform it into
a maximization problem. Consider the minimization of a scalar objective function,
L(x, u), called the performance index, with respect to a set of mutually independent,
unconstrained scalar variables, u = (u1 , u2 , . . . , um )T . The set u ∈ Rm×1 is
referred to as control vector, while the set of additional variables, x ∈ Rn×1 ,
necessary to describe the static system is called the state vector. Clearly, control
variables are those with respect to which the performance index is minimized. It is
assumed that L(x, u) is continuous and possesses continuous partial derivatives with
respect to both x and u up to an order which is sufficient to validate all the arguments
presented here. This holds for all the performance indexes contained in this book.
In order that a given candidate point (x, u) minimizes a continuously differentiable
performance index L(x, u) with respect to u, it must satisfy the following necessary
condition:
2.2 Optimization of Static Systems 9
⎧ ∂L ⎫
⎪
⎪ ∂u1 ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
T ⎪
⎪
∂L ⎪
⎪
∂L ⎨ ∂u2 ⎬
= =0 (2.1)
∂u ⎪
⎪ .. ⎪
⎪
⎪
⎪ . ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎩ ∂L ⎪⎭
∂um
where the extreme right-hand side is a null vector of dimension m × 1 and all the
derivatives are evaluated at the candidate point. Any point (x ∗ , u∗ ) that satisfies the
necessary condition for a minimum is called an extremal point (or a stationary point)
of L. However, an extremal point is not guaranteed to be a minimum point, because
the necessary condition of Eq. (2.1) is satisfied also by maximum points and saddle
points. In order to unambiguously identify a minimum point, a separate condition
must be satisfied. This condition requires an increase of L whenever departing from
the stationary point in any direction, hence it must be determined from the sign of
the second-order partial derivatives of L with respect to u.
A stationary point is necessarily a minimum of L(x, u) with respect to u, if the
following Hessian is positive definite (that is, has all real, positive eigenvalues) at
the given point:
⎛ ∂2L ∂2L ∂2L
⎞
∂u1 2 ∂u1 ∂u2 ··· ∂u1 ∂um
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ∂2L ∂2L
··· ∂2L ⎟
∂ L ⎜
2
⎜
∂u2 ∂u1 ∂u2 2 ∂u2 ∂um ⎟
⎟
=⎜ ⎟ (2.2)
∂u2 ⎜ ⎟
⎜ .. .. .. .. ⎟
⎜ . . . . ⎟
⎜ ⎟
⎝ ⎠
∂2L ∂2L ∂2L
∂um ∂u1 ∂um ∂u2 ··· ∂um ∂um
The positive definiteness of the Hessian at an extremal point is therefore a sufficient

condition for the existence of a minimum of L with respect to u and is denoted by:
∂ 2L
>0 (2.3)
∂u2
On the other hand, if the stationary point satisfies
∂ 2L
<0 (2.4)
∂u2
then it is a maximum point of L(x, u) with respect to u.

If the Hessian has both positive and negative real eigenvalues at a stationary
point, then that point is said to be a saddle point of L(x, u) with respect to u. If any
of the eigenvalues of the Hessian vanishes at a stationary point, then it is called a
singular point. Since a singular point does not strictly satisfy the sufficient condition
inequality, it is not an optimal point. Additional condition/s must be imposed to
convert a singular point into an optimal point.
Any point u satisfying both the necessary and sufficient conditions for a
minimum is called a minimum point (or optimal point) of L with respect to u.
If L(x, û) ≤ L(x, u) for all u = û, then the optimal point û is said to be a global
minimum of L(x, u) with respect to u. If a minimum point is not a global minimum,
then it is a local minimum.
Example 2.1 Determine the minimum points of the following function of real
variables:

1 −1 u1
L = {u1 , u2 }
−1 4 u2
Since the performance index is a quadratic function of u = (u1 , u2 )T , its only

stationary point is the origin:

∂L ∗ 1 −1
Lu = = u1 , u∗2 = {0 , 0}
∂u −1 4
or, u∗1 = u∗2 = 0. The Hessian is the following:

∂ 2L 1 −1
Luu = = ,
∂u2 −1 4
whose both eigenvalues are real and positive, (0.7, 4.3). As the sufficient condition
is satisfied by the Hessian, the origin is the only minimum point of L, hence L = 0
is its global minimum value.
Example 2.2 Does the following function of real variables have a minimum point:
L = −u21 + 2u1 u2 + 3u22 ?
Since the performance index is a quadratic function of u = (u1 , u2 )T , its only

stationary point is the origin:

∂L ∗ −1 1
Lu = = u1 , u∗2 = {0 , 0}
∂u 1 3
or, u∗1 = u∗2 = 0. The Hessian is the following:


∂ 2L −1 1
Luu = =
∂u2 1 3
√
whose eigenvalues are 1 ± 5. Since an eigenvalue of the Hessian is negative while
the other is positive, the only stationary point (the origin) is a saddle point of L.
Therefore, there are no minimum points of L.
Example 2.3 Determine whether the origin is a minimum point of the following
function:
L = (u1 − u22 )(u1 − 3u22 )
It is easily seen that u∗ = (0, 0)T satisfies the stationary condition, Lu = 0:

T
∂L 2u∗1 − 4(u∗2 )2 0
= =
∂u −8u∗1 u∗2 + 12(u∗2 )3 0
and is hence an extremal point of L. The Hessian at the extremal point is the
following:
T
∂ 2L 2 − 8u∗2 −8u∗2 2 0
Luu = = =
∂u2 −8u2 −8u1 + 36(u∗2 )2
∗ ∗ 0 0
whose eigenvalues are (0, 2). As the sufficient condition is not satisfied by the
Hessian, the origin is not a minimum point of L. Since one of the eigenvalues of
the Hessian vanishes at the origin, the latter is a singular point of L.
2.2.1 Static Equality Constraints
Consider the minimization of L(x, u) with respect to u and subject to the vector
constraint:
f (x, u) = 0 (2.5)
with f : Rn×1 ×Rm×1 → Rn×1 having continuous partial derivatives with respect to
both the state variables, x, and control variables, u. The equality constraint must be
such that it uniquely determines x from u. This fact translates into the requirement
that a small and arbitrary variation in the control, du, must cause a corresponding
variation in the state, dx, in order to satisfy the equality constraint. The variations
being small allow one to use the following first-order approximation:
df = fx dx + fu du = 0 (2.6)
or
dx = −fx−1 fu du (2.7)
with fx = ∂f/∂x ∈ Rn×n and fu = ∂f/∂u ∈ Rn×m . Equation (2.7) requires the
square matrix, fx (called the Jacobian) must be non-singular. Since a stationary
point must also have a stationary L in the presence of small and arbitrary variation
in u, we must have to a first-order approximation:
dL = Lx dx + Lu du = 0 (2.8)
while holding df = 0, with Lu = ∂L/∂u ∈ R1×m and Lx = ∂L/∂x ∈ R1×n . This

implies that the following must be true:

dL = Lu − Lx fx−1 fu du = 0 (2.9)
Since du is arbitrary, we have the following m scalar equations relating the partial
derivatives of L at the stationary point:
Lu − Lx fx−1 fu = 0 (2.10)
An alternative necessary condition for minimization in the presence of the equal-

ity constraints can be derived by adjoining the constraints with the function being
minimized. Define the following Hamiltonian function in which the function to be
minimized, L(x, u) (called the Lagrangian), is adjoined to the vector constraint
f (x, u) = 0 through a set of additional variables λ ∈ Rn×1 called the Lagrange
multipliers or the adjoint variables:
H (x, u, λ) = L(x, u) + λT f (x, u) (2.11)
The stationary condition in the presence of a small but arbitrary variation in u can
now be expressed as follows:
dH = Hx dx + Hu du = 0 (2.12)
with Hu = ∂H /∂u ∈ R1×m and with Hx = ∂H /∂x ∈ R1×n . For the convenience
of determining the Lagrange multipliers, λ (which are completely arbitrary), the
following condition is imposed at the stationary point:
Hx = Lx + λT fx = 0 (2.13)
This implies that
(λ∗ )T = −Lx fx−1 (2.14)

Upon substituting Hx = 0 in Eq. (2.12), the following stationary condition is

derived:
Hu = Lu + (λ∗ )T fu = 0 (2.15)
Equations (2.13)–(2.15) are the necessary conditions for the minimization of

L(x, u) with respect to u, subject to f (x, u) = 0.
The sufficient conditions for the minimization of L(x, u) with respect to u,
subject to f (x, u) = 0, are derived by taking the variations in L and f up to
terms of the second order. Consider the following Taylor series expansions about
the stationary point:
L
dx 1 T xx Lxu dx
dL = {Lx , Lu } + dx , duT (2.16)
du 2 Lux Luu du
f
dx 1 T xx fxu dx
df = {fx , fu } + dx , duT (2.17)
du 2 fux fuu du
where
T
∂ 2L ∂ ∂L
Lxu = =
∂x∂u ∂u ∂x
∈ Rn×m , etc., are the second-order partial derivatives. The Hamiltonian variation
can similarly be expanded into a second-order Taylor series, and substituting the
necessary conditions for optimality, Hx = 0 and Hu = 0, we have
1 T H
dx
xx Hxu
dH = dx , duT (2.18)
2 Hux Huu du
The substitution of Eqs. (2.17)–(2.18) into
dL = dH − λT df , (2.19)
along with the following Taylor series variation in x,
dx = −fx−1 fu du + terms involving higher powers of (dx, du) (2.20)
results in the following variation in L up to second-order terms:

⎛ ⎞⎧ −1
⎫
1 T T T −1 Hxx Hxu ⎨ −fx fu ⎬
dL = du −fu (fx ) , I ⎝ ⎠ du (2.21)
2 ⎩ ⎭
Hux Huu I
The sufficient condition for the minimization of L(x, u) with respect to u, subject
to f (x, u) = 0, is thus the following:
⎛ ⎞⎧ ⎫

−1
Hxx Hxu ⎨ −fx fu ⎬
∂ 2L
= −fuT (fxT )−1 , I ⎝ ⎠ >0 (2.22)
∂u2 f =0 ⎩ ⎭
Hux Huu I
or
Huu − Hux fx−1 fu − fuT (fxT )−1 Hxu + fuT (fxT )−1 Hxx fx−1 fu > 0 (2.23)
This form of the sufficient condition for optimality in the presence of equality
constraints is referred to as the Legendre–Clebsch sufficient condition.
Example 2.4 Determine the extremal and minimum points for

1 y2 u2
L= +
2 a2 b2
with respect to u, and subject to the following constraint:
y + mu − c = 0
where a, b, c, m are real constants.

We begin with defining the following Hamiltonian function in which the function
to be minimized, L, and the equality constraint are combined:

1 y2 u2
H = 2
+ 2 + λ (y + mu − c)
2 a b
The necessary conditions for minimization are the following:

∂H y∗
= 2 +λ=0
∂y a
∂H u∗
= 2 + mλ = 0
∂u b
solving which along with the equality constraint, y ∗ + mu∗ − c = 0, results in the
following extremal point:
a2c
y∗ =
a 2 + m2 b 2
b2 mc
u∗ =
+ m2 b 2
a2
c
λ∗ = − 2
a + m2 b 2
The sufficient condition for minimization requires that (Luu )f =0 > 0, which is
seen to be satisfied at the extremal point:
1 m2
(Luu )f =0 = + >0
b2 a2
Thus, the extremal point is a minimum point, with the following minimal value of L:
c2
Lmin =
2(a 2 + m2 b 2 )
Example 2.5 Determine the extremal and minimum points for
L = 3z2 + u21 ez + u22 y 2 + (u3 − 2)2
with respect to u = (u1 , u2 , u3 )T ∈ R3×1 , subject to the following constraints:
z − u1 y 2 = 0
z + 2u3 = 0
where x T = (z, y) ∈ R1×3 .

We begin with defining the following Hamiltonian function:
H = 3z2 + u21 ez + u22 y 2 + (u3 − 2)2 + λ1 (z − u1 y 2 ) + λ2 (y + 2u3 )
Here, (λ1 , λ2 ) ∈ R1×2 are the Lagrange multipliers corresponding to the two
constraints, respectively. The problem of minimizing L relative to u while satisfying
the two equality constraints is thus translated into the problem of minimizing the
Hamiltonian function, H , relative to u.
The extremal control, u∗ , satisfying the stationary condition is derived as follows:
⎧ ∗
⎫ ⎧ ⎫
⎪
⎪ 2u∗1 ez − λ∗1 (y ∗ )2 ⎪
⎪ ⎪
⎪ 0⎪⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
T ⎪
⎨ ⎪
⎬ ⎪
⎨ ⎪
⎬
∂H ∗ ∗
= 2u2 (y ) 2 = 0
∂u ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎩ ∗ ∗
⎪
⎭ ⎪
⎩ ⎪
⎭
2(u3 − 2) + 2λ2 0
which is identically satisfied by:
λ∗1 (y ∗ )2
u∗1 = ∗
2ez
u∗2 = 0
u∗3 = 2 − λ∗2
In order to determine the Lagrange multipliers, the following condition is imposed:

⎧ ⎫ ⎧ ∗ ⎫ ⎧ ⎫
⎨ ∂z ⎪
⎪ ∂H
⎬ ⎨ 6z∗ + (u∗1 )2 ez + λ∗1 ⎬ ⎨0⎬
= =
⎩ ∂H ⎪
⎪ ⎭ ⎩ ⎭ ⎩ ⎭
∂y 2(u∗2 )2 y ∗ − 2λ∗1 u∗1 y ∗ + λ∗2 0
which are satisfied by:

∗
λ∗1 = −6x ∗ − (u∗1 )2 ez
λ∗2 = 2u∗1 λ∗1 y ∗
When the extremal control values are substituted into the Lagrange multipliers, we
have a quadratic equation for u∗1 which is solved to yield the following stationary
value of u∗1 :

1 1 6z∗
u∗1 =− ∗ 2 ± ∗ 4
− z∗ , (y ∗ = 0)
(y ) (y ) e
For the extremal control variable u∗1 to have a real value, it is necessary to impose
the restriction:
∗
ez ≥ 6z∗ (y ∗ )4 , (y ∗ = 0)
The stationary control variable u∗1 can now be substituted into the expressions
derived above for λ∗1 , λ∗2 , and u∗3 in order to determine their values. Hence, the
stationary points are obtained.
In order to investigate the minimum points, we derive the Hessian for the
sufficient condition in the following steps:
∂H
Hu = = 2u1 ez − λ1 y 2 , 2u2 y 2 , 2(u3 − 2) + 2λ2
∂u
⎛ ⎞
2ez 0 0
Huu = ⎝ 0 2y 2 0 ⎠
0 0 2
z
2e u1 0 0
T
Hux =
−2yλ1 4yu2 0

Hx = 6z + ez u21 + λ1 , 2yu22 − 2λ1 u1 y + λ2

6 + u21 ez 0
Hxx =
0 2u22 − 2λ1 u1
⎛ ⎞
2u21 ez −2yλ1
T
Hxu =⎝ 0 4yu2 ⎠
0 0

1 −2yu1
fx =
0 1

1 2yu1
fx−1 =
0 1

−y 2 0 0
fu =
0 0 2

−y 2 0 4u1 y
fx−1 fu =
0 0 2
⎛ ⎞
−2y 2 u1 ez 0 8ez yu21 − 4λ1 y
Hux fx−1 fu = ⎝ 0 0 8yu2 ⎠
0 0 0
⎛ ⎞
−2y 2 u1 ez 0 0
(fx−1 fu )T Hxu =⎝ 0 0 0⎠
8e yu1 − 4λ1 y 8yu2 0
z 2
fuT (fxT )−1 Hxx fx−1 fu =

⎛ ⎞
6y 4 + u21 y 4 ez 0 −24u1 y 3 − 4u31 y 3 ez
⎝ 0 0 0 ⎠
−24u1 y − 4y u1 e 0 96u1 y + 16y u1 e + 8u2 − 8u1 λ1
3 3 3 z 2 2 2 4 z 2
∗
Substitution of λ∗1 = −6x ∗ − (u∗1 )2 ez and u∗2 = 0 for the stationary points results
in the following expression for the required Hessian:
(Luu )f =0 = Huu − (fx−1 fu )T Hxu − Hux fx−1 fu + fuT (fxT )−1 Hxx fx−1 fu
⎛ 4 ⎞∗
6y + u21 y 4 ez 0 − 24u1 y 3 + 4u31 y 3 ez
⎜ +2ez + 4y 2 u ez +24xy + 12ez yu21
⎟
⎜ 1 ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
=⎜ 0 2y 2 −8yu2 ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎝ − 24u1 y 3 + 4u31 y 3 ez −8yu2 2 + 96u21 y 2 + 16y 2 u41 ez ⎠

+24xy + 12ez yu21 +8u22 − 8u31 ez − 48xu1
where the stationary expression for u1 , calculated above, must be substituted. The
minimum points of L are the ones for which (Luu )f =0 has real and positive
eigenvalues in the (z, y) space. Of course, this has to be determined by a numerical
search.
As the previous example illustrates, often the solution for the minimum points
cannot be obtained in a closed form, and instead requires a numerical solution.
The search for extremal points that satisfy the sufficient condition of minimization
in the presence of implicit, nonlinear constraints in terms of the control and state
variables must be carried out by a nonlinear programming algorithm. Such iterative
computational schemes form the crux of the current research on multivariable
optimization methods.
2.2.2 Inequality Constraints
Consider general inequality constraints of the form:
f (x, u) ≤ 0 , (2.24)
while minimizing L(x, u) with respect to u. Extending the concept of Lagrange

multipliers, λ, we adjoin L with the inequality constraint function into the following
augmented objective function:
J (x, u) = L(x, u) + λT f (x, u) . (2.25)
Noting the two possibilities for an optimal point (if it exists), i.e., either
f (x̂, û) < 0 , (2.26)
or
f (x̂, û) = 0 , (2.27)
we say that in the first case (Eq. (2.26)), the constraint is inactive and thus it can
be ignored by putting λ = 0 in Eq. (2.25), which is treated as unconstrained
minimization we discussed above. In the second case (Eq. (2.27)), the minimum
point (if it exists) lies on the constraining boundary, and a small control variation,
δu, about the minimum point, (x̂, û), would result in either an increase, or no change
of L:
∂L
δL = δu ≥ 0 , (2.28)
∂u
where δu must satisfy the constraint, Eq. (2.24), with λ > 0:
∂f
λT δf = λT δu ≤ 0 (2.29)
∂u
Since δu is arbitrary, there are only the following two possibilities of reconciling
Eqs. (2.28) and (2.29), i.e., either we should have (with λ > 0)
∂L ∂f
= −λT (2.30)
∂u ∂u
or
∂L ∂f
=0; =0. (2.31)
∂u ∂u
The three possibilities represented by Eqs. (2.26), (2.30), and (2.31) can be
expressed in the following compact form:
∂L ∂f
+ λT =0, (2.32)
∂u ∂u
where

= 0, f (x, u) < 0
λ (2.33)
> 0, f (x, u) = 0
Equation (2.32) is the necessary condition for optimality and must be solved in an
iterative manner to yield an optimal point (if it exists).
Whenever inequality constraints are specified, they denote a bounded (or feasi-
ble) region in which the search for any optimal points is to be performed. If a point
falls at the boundary of the feasible region, then it may not satisfy the necessary
and sufficient conditions for minimization (or maximization) of the performance
index, yet it could be an optimal point simply because it has the smallest (or the
largest) numerical value of the performance index in the feasible region. Thus, we
can categorize the search for a minimum in the presence of inequality constraints as
follows:
1. The minimum point lies inside the feasible region, where both the necessary and
sufficient conditions of optimality are satisfied.
2. The minimum point lies on a boundary of the feasible region and satisfies the
necessary (but not sufficient) conditions of optimality. An example of this case
is a singular control problem in which an inequality constraint supplies the
additional information for determining the minimality of the given point.
3. The minimum point lies on a boundary of the feasible region, but does not satisfy
the necessary conditions of optimality.
2.3 Dynamic Equality Constraints and Unbounded Inputs
We begin the treatment of the optimal control problem with the case of continuously
applied, smooth control inputs which are unbounded by any physical constraints.
Consider the problem of finding an admissible set of control inputs u∗ (t) ∈ Rm×1
to follow an extremal trajectory x ∗ (t) ∈ Rn×1 , which is the solution to the following
state equation:
ẋ = f (x, u, t) (2.34)
subject to the initial condition:
x(t0 ) = x0 (2.35)
such that the following performance index is minimized with respect to the control
inputs:
tf

J = φ[x(tf ), tf ] + L(x, u, t)dt (2.36)

t0
where f : Rn×1 × Rm×1 × R → Rn×1 , φ : Rn×1 × R → R, and L : Rn×1 ×

Rm×1 × R → R are continuous and possess continuous partial derivatives with
respect to their arguments, up to an indefinite order. While φ is the terminal cost to
be minimized, the integral term containing the Lagrangian, L(x, u), represents the
transient cost to be minimized over the control interval, t ∈ [t0 , tf ]. The general
form of the performance index given in Eq. (2.36) is termed the Bolza problem. A
special optimal control problem in which φ[x(tf ), tf ] = 0 is called the Lagrange
problem, whereas the one with L(x, u, t) = 0 is termed the Mayer problem.
We begin by writing the terminal cost as follows:
tf
φ[x(tf ), tf ] = φ̇(x, t)dt + φ[x(t0 ), t0 ] (2.37)
t0
which substituted into Eq. (2.36) results in

tf
J
= φ[x(t0 ), t0 ] + [L(x, u, t) + φ̇(x, t)]dt (2.38)
t0
Since x(t0 ) and t0 are both fixed, they do not affect the minimization, and only the
following performance index needs to be minimized:
tf
J = [L(x, u, t) + φ̇(x, t)]dt (2.39)
t0
2.3 Dynamic Equality Constraints and Unbounded Inputs 21
which is also expressed as follows:

tf
∂φ ∂φ
J = L(x, u, t) + (x, t) + (x, t)ẋ dt (2.40)
t0 ∂t ∂x
Let x ∗ (t) be an extremal trajectory satisfying the necessary conditions for the
minimization of J subject to Eq. (2.34), and let u∗ (t) be the corresponding control
history. We augment J by adjoining the state equation as the dynamic equality
constraint through Lagrange multipliers, λ(t) ∈ Rn×1 (also called the adjoint, or
costate variables):
tf ∂φ ∂φ
Ja = [L + + ẋ + λT (f − ẋ)]dt (2.41)
t0 ∂t ∂x
Introducing the following augmented Lagrangian:

∂φ ∂φ
La = L + + ẋ + λT (f − ẋ) (2.42)
∂t ∂x
we have the following augmented performance index to be minimized:
tf
Ja = La (x, u, λ, ẋ, t)dt (2.43)
t0
Next, consider infinitesimal and arbitrary variations in x(t), u(t), λ(t), ẋ(t) about
the extremal trajectory represented by δx(t), δu(t), δλ(t), δ ẋ(t), respectively.1 The
net change in the augmented performance index over the extremal trajectory is then
given by:
tf
dJa = [La (x ∗ + δx, u∗ + δu, λ∗ + δλ, ẋ ∗ + δ ẋ, t) − La (x ∗ , u∗ , λ∗ , ẋ ∗ , t)]dt
t0
tf +δtf
+ La (x, u, λ, ẋ, t)dt (2.44)
tf
The first integrand on the right-hand side of Eq. (2.44) is expanded in a first-order
Taylor series to yield
tf
∂La ∂La ∂La ∂La
dJa = δx + δu + δ ẋ + δλ dt
t0 ∂x ∂u ∂ ẋ ∂λ
tf +δtf
+ La (x, u, λ, ẋ, t)dt (2.45)
tf
1 Since the dynamic equality constraint of Eq. (2.34) must be enforced separately, ẋ is taken to be
an independent variable in the minimization of Ja .

where all coefficients are evaluated on the admissible trajectory. Now, the second
integral on the right-hand side of Eq. (2.44) is expressed as follows:
tf +δtf
La (x, u, λ, ẋ, t)dt = La (x(tf ), u(tf ), λ(tf ), ẋ(tf ), tf )δtf
tf
+ higher order terms involving δtf (2.46)
Neglecting the second- (and higher-) order terms in δtf in Eq. (2.46), and expanding
the augmented Lagrangian in a first-order Taylor series about the extremal trajectory
at the final time, we have
La (x(tf ), u(tf ), λ(tf ), ẋ(tf ), tf ) = La [x ∗ (tf ), u∗ (tf ), λ∗ (tf ), ẋ ∗ (tf ), tf ]

∂La ∂La
+ δx(tf ) + δu(tf )
∂x t=tf ∂u t=tf

∂La ∂La
+ δλ(tf ) + δ ẋ(tf )
∂λ t=tf ∂ ẋ t=tf
(2.47)
When Eq. (2.47) is substituted into Eq. (2.46), we have the following first-order
expansion of the second integral on the right-hand side of Eq. (2.44):
tf +δtf
La (x, u, λ, ẋ, t)dt = La [x ∗ (tf ), u∗ (tf ), λ∗ (tf ), ẋ ∗ (tf ), tf ]δtf (2.48)
tf
In the first integral on the right-hand side of Eq. (2.45), the term involving δ ẋ is
expanded by integration by parts as follows:
tf tf tf
∂La ∂La d ∂La
δ ẋdt = δx − δxdt
t0 ∂ ẋ ∂ ẋ t0 t0 dt ∂ ẋ

∂La
= δx(tf )
∂ ẋ t=tf
tf
∂La d ∂La
− δx(t0 ) − δxdt (2.49)
∂ ẋ t=t0 t0 dt ∂ ẋ
However, since δx(t0 ) = 0 (the initial condition being fixed), we have

tf tf
∂La ∂La d ∂La
δ ẋdt = δx(tf ) − δxdt (2.50)
t0 ∂ ẋ ∂ ẋ t=tf t0 dt ∂ ẋ
The substitution of Eqs. (2.48) and (2.49) into Eq. (2.45) leads to the following
necessary condition for the minimization of Ja :

∂La
dJa = 0 = δx(tf ) + La [x ∗ (tf ), u∗ (tf ), λ∗ (tf ), ẋ ∗ (tf ), tf ]δtf
∂ ẋ t=tf
tf
∂La d ∂La ∂La ∂La
+ − δx + δu + δλ dt (2.51)
t0 ∂x dt ∂ ẋ ∂u ∂λ
Before proceeding further, it is necessary to express the variation in the terminal

state by the following first-order approximation:
∂x(tf ) = δxf − ẋ ∗ (tf )δtf (2.52)
This is tantamount to a linear extrapolation around the extremal trajectory at the

terminal time, as indicated by the schematic diagram of Fig. 2.1. Clearly, when the
terminal condition is not fixed, the variation of the terminal state depends linearly
upon the variation in the terminal time.
When Eq. (2.52) is substituted into Eq. (2.51), we have the following necessary
condition for minimization:

∂La ∂La ∗
dJa = 0 = δxf + La − δ ẋ δtf
∂ ẋ t=tf ∂ ẋ t=tf
tf
∂La d ∂La ∂La ∂La
+ − δx + δu + δλ dt (2.53)
t0 ∂x dt ∂ ẋ ∂u ∂λ
Fig. 2.1 Extrapolation of the .*

(tf)δtf
terminal state in the
neighborhood of the extremal x
trajectory for a small,
arbitrary variation in the δ xf
terminal time δx(tf)
xf
x*(t)
x(t)
xi t
ti tf tf +δ tf
It is now possible to derive the individual necessary conditions for optimality

from Eq. (2.53). The terms involving the terminal cost, φ, in the augmented
Lagrangian (see Eq. (2.42)) must satisfy the following condition inside the integral
term of Eq. (2.53):

∂ ∂φ ∂φ d ∂ ∂φ
δ ẋ ∗ + − ẋ ∗ (2.54)
∂x ∂x ∂t x=x ∗ dt ∂ ẋ ∂x x=x ∗
2
∂ φ ∗ ∂ 2φ d ∂φ
= ẋ + −
∂x 2 ∂t∂x dt ∂x
2
∂ φ ∗ ∂ 2φ ∂ 2φ ∂ 2φ
= 2
ẋ + − 2
ẋ ∗ −
∂x ∂t∂x ∂x ∂x∂t
Since the partial derivatives of φ must necessarily be continuous, we must have

∂ 2 φ/∂t∂x = ∂ 2 φ/∂x∂t, hence the term on the right-hand side of Eq. (2.54) vanishes
identically. Thus, all the terms involving φ inside the integral term of Eq. (2.53)
cancel out, thereby yielding the following necessary condition for optimality arising
out of that particular term:
tf
∂L ∂f d ∂L ∂f
0= + (λ∗ )T − (−λ∗ )T δx + + (λ∗ )T δu
t0 ∂x ∂x dt ∂u ∂u
T
+ f − ẋ ∗ δλ dt (2.55)
Equation (2.55) must hold irrespective of any boundary conditions on x and λ.

Furthermore, the equation of dynamic constraint (Eq. (2.34)) must be satisfied along
the extremal trajectory:
ẋ ∗ = f (x ∗ , u∗ , t) (2.56)
which implies that the coefficient of δλ must vanish in Eq. (2.55), thereby leaving
the following necessary condition to be satisfied for optimality:
tf
∂L ∗ T ∂f d ∗ T ∂L ∗ T ∂f
+ (λ ) + (λ ) δx + + (λ ) δu dt = 0
t0 ∂x ∂x dt ∂u ∂u
(2.57)
Because Eq. (2.57) must hold irrespective of variations on x and u, the coefficients
of both δx and δu must independently vanish along the extremal trajectory. Thus,
we have
∂L ∂f
(λ̇∗ )T = − − (λ∗ )T (2.58)
∂x ∂x
and
∂L ∂f
+ (λ∗ )T =0 (2.59)
∂u ∂u
These last two equations, along with the state equation, Eq. (2.34), are the necessary
conditions which must be satisfied by the extremal trajectory, x ∗ (t) and the extremal
control history u∗ (t), in the control interval t ∈ [t0 , tf ]. Of these, Eq. (2.58) is
termed the costate equation (or the adjoint equation), while Eq. (2.59) is called the
stationary condition. Furthermore, the term containing δxf and δtf in Eq. (2.53),
representing the terminal boundary variation, must also vanish, resulting in the
following transversality condition:

∂La ∂La ∗
δxf + La − ẋ δtf = 0 (2.60)
∂ ẋ t=tf ∂ ẋ t=tf
In order to combine the requirements of optimality with the dynamic constraint,

a Hamiltonian function is defined as follows:
H (x, u, λ, t) = L(x, u, t) + λT f (x, u, t) (2.61)
Then, the necessary conditions for optimality are expressed as follows in terms of
the Hamiltonian function:

∗ ∗ ∗ ∂H T ∗ ∗ ∗
ẋ = f (x , u , t) = (x , u , λ , t) (2.62)
∂λ
T T T
∗ ∂L ∂f ∗ ∂H
λ̇ = − − λ = (x ∗ , u∗ , λ∗ , t) (2.63)
∂x ∂x ∂x
and
T T
∂L ∂f ∂H ∗ ∗ ∗
+ λ∗ = 0 = (x , u , λ , t) (2.64)
∂u ∂x ∂u
The performance index of optimization can be expressed as follows in terms of the
Hamiltonian:
tf
J = φ[x(tf ), tf ] + H (x, u, λ, t) − λT ẋ dt (2.65)
t0
It is to be noted that the Hamiltonian is related to the augmented Lagrangian,

Eq. (2.42), as follows:
∂φ ∂φ
La = H + + ẋ − λT ẋ (2.66)
∂t ∂x
Then, the transversality condition can be derived from Eq. (2.60) to be the following:
! T "T
∂φ ∗ ∂φ
− λ (tf ) δxf + H + δtf = 0 (2.67)
∂x t=tf ∂t t=tf
Equations (2.62)–(2.64), along with Eq. (2.67), are collectively termed the Euler–
Lagrange equations, which must be satisfied on any trajectory (the extremal)
minimizing the performance index. In addition to the scalar boundary condition
of Eq. (2.67) imposed on the terminal time tf , boundary conditions on x ∗ (t), λ∗ (t)
must be specified in order to satisfy the Euler–Lagrange equations. For a system of
order n, the total number of boundary conditions necessary to pose the optimal
control problem is thus (2n + 1). Of these, n boundary conditions are already
specified by the initial condition, Eq. (2.35). Hence, only (n+1) additional boundary
conditions are necessary to completely pose the optimal control problem.
2.4 Special Boundary Conditions
Optimal control formulations can be classified according to the boundary conditions

imposed on the feasible space and control interval. These are now taken up in detail
depending upon how the terminal time and state are specified.
2.4.1 Fixed Terminal Time
When the terminal time is fixed, we cannot have any variation in tf , therefore
δtf = 0, and the total number of boundary conditions to be imposed on the system
is reduced to 2n. A further classification occurs due to the boundary conditions
imposed upon the terminal state, x(tf ), as follows.
Fixed Terminal State
Here, the terminal state is specified to be a fixed point in the state space:
x(tf ) = xf (2.68)
which implies that there can be no variation on the terminal state at all (δxf = 0).
Consequently, Eq. (2.67) is identically satisfied, and the initial condition, Eq. (2.35),
along with the final boundary condition, Eq. (2.68), furnishes the required 2n
boundary conditions for the problem.
Free Terminal State
When no boundary condition is imposed at all on the terminal state, the transver-
sality condition, Eq. (2.67), results in the following n boundary conditions on the
adjoint (or costate) variables:
2.4 Special Boundary Conditions 27
T
∂φ
λ∗ (tf ) = (2.69)
∂x t=tf
Terminal State on a Hypersurface
Sometimes instead of fixing the terminal state x(tf ) to be at a point, it is given to lie
anywhere in a region of the state space specified by:
F [x(tf )] = 0 (2.70)
where F : Rn×1 → Rk×1 , 1 ≤ k ≤ (n − 1). Note that the requirement n < 1 is

necessary to preclude the case n = 1, which is tantamount to fixing the final state
(δxf = 0). The scalar components of Eq. (2.70) can be expressed as follows:
Fi [x(tf )] = 0 , i = 1, . . . , k (2.71)
The hypersurface is thus represented by contours, Fi = 0, i = 1, . . . , k, and the

condition that x(tf ) must lie on the hypersurface is enforced by ensuring that the
variation of x(tf ) is tangential to each contour. Thus, we have
#
k
∂Fi
ci δx(tf ) = 0 , i = 1, . . . , k (2.72)
∂x t=tf
i=1
However, since the final state x(tf ) should also be on an extremal trajectory
minimizing J with respect to u, the variation in the terminal state, δxf , is exactly
specified by the normal δx(tf ) to each of the hypersurfaces, Fi [x(tf )] = 0,
1 ≤ k ≤ (n − 1), bounding the terminal region at the final extremal point, x ∗ (tf ),
represented by Eq. (2.67) as follows:
! T "T
∂φ ∗
− λ (tf ) δx(tf ) = 0 (2.73)
∂x t=tf
where x(tf ) satisfies Eq. (2.72). Hence, the requirements that x(tf ) should be on
the hypersurface and that it should also fall on the extremal trajectory are combined
into the following n terminal boundary conditions for the state variables:
! T "
∂φ ∗ ∂F1 ∂F2 ∂Fk
−λ = c1 + c2 + · · · + ck
∂x ∂x t=tf ∂x t=tf ∂x t=tf
t=tf
(2.74)
where c1 , c2 , . . . , ck are the constant Lagrange multipliers to be determined. This

implies that apart from the initial condition, Eq. (2.35), there are a total of (n + k)
terminal boundary conditions of which n are specified by Eq. (2.74), and the
remaining k conditions are given by:
F [x ∗ (tf )] = 0 (2.75)
Equation (2.74) is equivalent to replacing the terminal cost φ in Eq. (2.69) with
the following:
Φ = φ + cT F [x(tf )] (2.76)
where c = (c1 , c2 , . . . , ck )T . Thus, the boundary condition for free terminal state
is used to represent the case of a terminal state constrained to be on a hypersurface
using the modified terminal cost, which adjoins the terminal constraint with the
original terminal cost using the additional adjoint parameters, c1 , c2 , . . . , ck , which
are constants due to the static nature of the equality constraint, Eq. (2.75).
2.4.2 Free Terminal Time
When the terminal time tf is not specified while stating the optimal control problem,
the variations on the terminal time are allowed, (δtf = 0), and Eq. (2.67) provides
the boundary condition to determine tf .
Fixed Terminal State
When the terminal state is fixed,
x(tf ) = xf (2.77)
there cannot be any variation on the terminal state (δxf = 0). Consequently,
Eq. (2.67) leads to the following scalar boundary condition:

∂φ
H+ =0 (2.78)
∂t t=tf
Free Terminal State
Since both δxf and δtf are arbitrary and nonzero, their respective coefficients
in Eq. (2.67) must independently vanish. Thus, we have the following boundary
conditions:
2.4 Special Boundary Conditions 29
T
∂φ
λ∗ (tf ) = (2.79)
∂x t=tf
and

∂φ
H+ =0 (2.80)
∂t t=tf
Time-Varying Terminal State
When the terminal boundary condition is a moving point in the state space,
described by a time-varying function h(t) : R → Rn×1 , the variation in the terminal
state is expressed by a first-order time difference as follows:

dh
δxf = δtf (2.81)
dt t=tf
Substituted into Eq. (2.67), this results in the following scalar boundary condition:
! T "T
∂φ ∂φ ∗ dh
H+ + −λ =0 (2.82)
∂t t=tf ∂x dt t=tf
t=tf
Terminal State on a Moving Hypersurface
The terminal state lies anywhere in a region of the state space specified by:
F [x(tf ), tf ] = 0 (2.83)
where F : R(n+1)×1 → Rk×1 , 1 ≤ k ≤ (n − 1). On such a surface, δxf and

δtf are independent of each other, and both, being arbitrary variations, must have
independently vanishing coefficients in Eq. (2.67). Thus, we have the following
boundary conditions:

∂φ
H+ =0 (2.84)
∂t t=tf
and
! T "
∂φ ∗ ∂F1 ∂F2 ∂Fk
−λ = c1 + c2 + · · · + ck
∂x ∂x t=tf ∂x t=tf ∂x t=tf
t=tf
(2.85)
where c1 , c2 , . . . , ck are time-varying coefficients to be treated as additional vari-

ables for adjoining the dynamic terminal constraint, Eq. (2.83), to the performance
index. Hence, there are in total (2n+k+1) boundary conditions for an equal number
of variables, x ∗ , λ∗ , tf , c1 , c2 , . . . , ck .
Terminal State Partially Fixed
The case where some state variables are either fixed or constrained to lie on
a specified hypersurface, while other state variables are free at the unspecified
terminal time, has to be specially treated. Let the partially specified state variables
be referred by the following boundary condition:
xi (tf ) = xf , i = 1, 2, . . . , q (2.86)
where q ≤ n. This can be regarded as a special case of final state lying on a

hypersurface, where
F [x(tf ), tf ] = Gx(tf ) − xf (2.87)
with
$ %
G = Iq×q , 0q×(n−q) (2.88)
with Iq×q and 0q×(n−q) being the identity and null matrices, respectively, of the
assigned dimensions. When substituted into Eq. (2.85), the resulting boundary
condition can be stated as follows:
T T
∂φ ∂F ∂φ
λ∗ (tf ) = + c = + GT c (2.89)
∂x ∂x t=tf ∂x t=tf
where c ∈ Rq×1 is a vector of constant coefficients (Lagrange multipliers), and

⎧ ⎫T
⎪
⎪
0q×1 ⎪
⎪

∂φ
T ⎨ ∂φ/∂xq+1 T
⎪ ⎪
⎬
= .. (2.90)
∂x ⎪
⎪ ⎪
⎪
t=tf ⎪
⎩
. ⎪
⎭
(∂φ/∂xn )T t=tf
Therefore, the terminal boundary conditions for the costate variables can be
expressed as follows:
&
∗
c
λ (tf ) = T , j = (q + 1), . . . , n (2.91)
∂φ/∂xj t=t
f
2.6 Pontryagin’s Minimum Principle 31
2.5 Sufficient Condition for Optimality
Let [x ∗ (t), u∗ (t), λ∗ (t)] be an extremal trajectory and associated control history
satisfying Hu = 0 as well as the other Euler–Lagrange necessary conditions for
optimality. The weak form of the Legendre–Clebsch condition given by:
Huu ≥ 0 (2.92)
is a necessary condition for minimization of H relative to u. However, an extremal

trajectory satisfying all the necessary conditions of optimality may not be the
optimal trajectory.
In addition to satisfying the necessary conditions, if the given extremal trajectory
satisfies the following Legendre–Clebsch condition of a positive definite Hessian of
the Hamiltonian with respect to the control variables:
∂ 2H
Huu = >0 (2.93)
∂u2
then the extremal trajectory is the optimal one. Hence, the condition given by the
strict inequality of Eq. (2.93) (called the strong Legendre–Clebsch condition) is the
sufficient condition for optimality on an extremal trajectory.
In cases where the Hamiltonian is a linear function of the control variables, the
Legendre–Clebsch sufficient condition is not satisfied, and instead we have Huu = 0
on an extremal trajectory. Such a problem is said to be singular in control, and
requires additional conditions to determine the optimal trajectory.
2.6 Pontryagin’s Minimum Principle
Pontryagin’s minimum principle states that if û(t) minimizes the performance index
J given by Eq. (2.40), subject to the dynamics state equality constraint, Eq. (2.34),
then any variation δ û in the control applied on the corresponding optimal trajectory,
[x̂(t), λ̂(t)], produces either an increase, or an invariance of the Hamiltonian, H .
In other words, an optimal control û(t) is the one that minimizes the Hamiltonian
among all admissible controls, u ∈ U ⊂ Rm×1 , at any point on the optimal
trajectory. The minimum principle can thus be expressed as the following inequality:
H (x̂, û + δ û, λ̂, t) ≥ H (x̂, û, λ̂, t) , ∀u ∈ U , t ∈ [t0 , tf ] (2.94)
Since the minimum principle does not require a Hamiltonian to be a smooth

function of the control variables, it provides a more general necessary condition of
optimality than Euler–Lagrange equations. Hence, optimal control problems with a
non-smooth Hamiltonian can be solved using the minimum principle.
It is a trivial matter to show that Euler–Lagrange equations lead to the minimum

principle, if the Hamiltonian function possesses continuous partial derivatives up to
the second order with respect to u. The necessary conditions for the minimization
of a smooth performance index are the following:
Hu (x̂, û, λ̂, t) = 0

Huu (x̂, û, λ̂, t) ≥ 0 (2.95)
A variation in the control about the given optimal trajectory while holding (x̂, λ̂)
unchanged produces the following second-order expansion of the Hamiltonian:

∂H 1 ∂ 2H
H (x̂, û + δ û, λ̂, t) = H (x̂, û, λ̂, t) + δu + δ ûT δ û (2.96)
∂u 2 ∂u2
which, after substitution of Eq. (2.95), produces Eq. (2.94).
2.7 Hamilton–Jacobi–Bellman Formulation
An alternative formulation for the sufficient conditions of optimality is possible

through the Hamilton–Jacobi–Bellman approach. For an optimal trajectory, x̂(t),
minimizing the performance index:
tf
J = φ[x(tf ), tf ] + L[x(t), u(t), t]dt , (2.97)
t0
subject to:
ẋ = f (x, u, t) , x(t0 ) = x0 , (2.98)
and corresponding to the optimal control history, û(t), one can define the following
optimal return function for t0 ≤ t ≤ tf :
tf
V̂ [x̂(t), t] = φ[x̂(tf ), tf ] + L[x̂(τ ), û(τ ), τ ]dτ , (2.99)
t
or
t
V̂ [x̂(t), t] = φ[x̂(tf ), tf ] − L[x̂(τ ), û(τ ), τ ]dτ . (2.100)
tf
Differentiating V̂ with time, we have

2.7 Hamilton–Jacobi–Bellman Formulation 33
dV̂
= −L[x̂(t), û(t), t] , (2.101)
dt
and also
dV̂ ∂ V̂ ∂ V̂
= + ẋ
dt ∂t ∂x
∂ V̂ ∂ V̂
= + f . (2.102)
∂t ∂x
By equating Eqs. (2.101) and (2.102), we have the following partial differential
equation, called the Hamilton–Jacobi–Bellman (HJB) equation, which must be
satisfied by the optimal return function:
∂ V̂ ∂ V̂
− = L[x̂(t), û(t), t] + f . (2.103)
∂t ∂x
The boundary condition for the optimal return function is the following:
V̂ [x̂(tf ), tf ] = φ[x̂(tf ), tf ] . (2.104)
A solution for the HJB equation in the optimal trajectory space, x̂, and time, t,
for a set of initial states, x0 , results in a field of optimal solutions. A finite set
of small initial perturbations about a given optimal path thus produces a set of
neighboring optimal solutions by the solution of HJB equation. However, a solution
of the HJB equation is a formidable task requiring simultaneous integration in
space and time for neighboring extremal trajectories with finite-difference (or finite-
volume) methods. A more practical use of the HJB equation is to provide a sufficient
condition for testing the optimality of a given return function.
On comparing Eqs. (2.103) and (2.61), we note that
∂ V̂
− = H [x̂(t), λ̂(t), û(t), t] (2.105)
∂t
where
∂ V̂
λ̂T = . (2.106)
∂x
With these identities, the HJB equation can be used to provide optimality conditions,
but its main application lies in deriving nonlinear feedback control laws of the form:
u(t) = û [t, x(t)] , (2.107)
from the arguments of the optimal return function, V̂ [x̂(t), t].

For a system described by:
ẋ = f (x, t) ; x(t0 ) = x0 , (2.108)
whose equilibrium point is xe = 0, if the optimal return function, V̂ [x̂(t), t],

satisfies Lyapunov’s theorem [74] for global asymptotic stability for all optimal
trajectories, x̂(t), then the given system is globally asymptotically stable at the
origin, x = 0. In this regard, the optimal return function is regarded as a Lyapunov
function. Hence, the task of a designer is to find a return function, V (x, t), that
is continuously differentiable with respect to the time and the state variables of
the system, and satisfies the following sufficient conditions for global asymptotic
stability about the equilibrium point at origin:
dV
V (0, t) = 0 , V (x, t) > 0 ; (x, t) < 0 ; for all x = 0
dt
(2.109)
| x |→ ∞ implies V (x, t) → ∞ . (2.110)
2.8 Time-Invariant Systems
Time-invariant systems are those whose properties do not change with time. They
are also referred to as autonomous systems in control parlance. The state equation
of such a system does not depend explicitly on time, and hence can be expressed as
follows:
ẋ = f (x, u) (2.111)
with x(t) : R → Rn×1 and u(t) : R → Rm×1 , and the initial condition:
x(t0 ) = x0 (2.112)
If the Lagrangian to be optimized is also time invariant, then the performance index
is given by:
tf
J = φ[x(tf ), tf ] + L(x, u)dt (2.113)
t0
resulting in the following Hamiltonian:
H = L(x, u) + λT f (x, u) (2.114)
To show that the Hamiltonian of a time-invariant system is constant along an

extremal, [x ∗ (t), u∗ (t)], let us consider the following necessary conditions of
optimality derived above:
2.9 Linear Systems with Quadratic Cost Functions 35
T
∂H ∗
λ̇∗ = − (2.115)
∂x
∗ T
∂H
ẋ ∗ = f (x ∗ , u∗ ) = (2.116)
∂λ
and
∗
∂H
=0 (2.117)
∂u
Substituting Eqs. (2.115)–(2.117) into the following time derivative of the Hamilto-
nian evaluated along the extremal trajectory:
∗ ∗ ∗
∂H ∂H ∂H
Ḣ = ẋ ∗ + λ̇∗ + u̇∗ (2.118)
∂x ∂λ ∂u
we have
∗ ∗ T ∗ ∗ T
∂H ∂H ∂H ∂H
Ḣ = − =0 (2.119)
∂x ∂λ ∂λ ∂x
This result is useful in deriving relationships among the variables along an extremal
trajectory.
2.9 Linear Systems with Quadratic Cost Functions
The special case of linear systems with a cost function to be optimized which
contains only quadratic terms of state and input variables can be solved analytically.
Let the state equation in such a case be given by:
ẋ = A(t)x(t) + B(t)u(t) (2.120)
with the initial condition:
x(t0 ) = x0 (2.121)
where A(t) : R →R n × n and B(t) : R →R n × m are the state-space coefficient

matrices.
The objective function for minimization is then expressed as follows:

tf
J = φ[x(tf ), tf ] + L(x, u, t)dt (2.122)
t0
where the terminal cost is given by:
1 T
φ[x(tf ), tf ] = x (tf )Qf (tf )x(tf ) (2.123)
2
and the Lagrangian is given by:
1 T 1
L(x, u, t) = x (t)Q(t)x(t) + x T (t)S(t)u(t) + uT (t)R(t)u(t) (2.124)
2 2
where Q(t) : R →R n × n, Qf (tf ) : R →R n × n, S(t) : R →R m × n, and

R(t) : R →R m × m are the specified cost coefficient matrices. Equations (2.123)
and (2.124) substituted into Eq. (2.122) can also be expressed as follows:
tf
1 Q(t) S(t) x(t)
J = x T (tf )Qf (tf )x(tf ) + {x (t), u (t)}
T T
dt
2 t0 S T (t) R(t) u(t)
(2.125)
From the foregoing discussion, it is evident that the matrices Q, Qf , and R must
be chosen to be symmetric. Since J ≥ 0, it is also requires that Qf (tf ), Q(t), and
R(t) be at least positive semi-definite. Other requirements will be imposed on the
cost coefficients from the sufficient conditions of optimality.
While the focus of the present section is on linear systems, it is important to
realize that an actual physical system is often nonlinear which can be linearized in
the neighborhood of an extremal trajectory. A fuller insight into linearized systems’
optimization with a quadratic cost function can be obtained from the expansion of
the actual nonlinear system about an extremal trajectory. Consider a dynamic plant
with state vector, ξ(t) : R →R n × 1, and control input vector, η(t) : R →R m × 1,
governed by the following state equation:
ξ̇ = f (ξ, η, t) . (2.126)
Let an extremal trajectory, ξ ∗ (t), and the corresponding extremal control history,
η∗ (t), both satisfying Eq. (2.126), be available from the necessary conditions for the
solution of the optimal control problem minimizing an objective function:
tf
J (ξ, η) = φ[ξ(tf ), tf ] + L[ξ(t), η(t), t]dt , (2.127)
t0
subject to certain specific constraints. The extremal control and trajectory are the
nominal functions satisfying Eq. (2.126):
ξ̇ ∗ = f (ξ ∗ , η∗ , t) , (2.128)
about which small control and state deviations:
u(t) = η(t) − η∗ (t) ; t 0 ≤ t ≤ tf (2.129)
x(t) = ξ(t) − ξ ∗ (t) ; t 0 ≤ t ≤ tf (2.130)
are to be minimized. A first-order Taylor series expansion about the extremal tra-
jectory leads to the following linearized state equation governing small deviations,
x(t), u(t):
ẋ(t) = A(t)x(t) + B(t)u(t) ; x(t0 ) = x0 , (2.131)
where
' '
∂f '' ∂f ''
A(t) = ; B(t) = (2.132)
∂ξ 'ξ ∗ ,η∗ ∂η 'ξ ∗ ,η∗
are the Jacobian matrices of the expansion. Similarly, the objective function can be
expanded about the extremal solution, (ξ ∗ , η∗ ), up to the second-order terms:
J (ξ ∗ + x, η∗ + u) J (ξ ∗ , η∗ ) + Δ2 J (x, u) , (2.133)
noting that the first variation of J about the extremal trajectory is identically zero:
' '
∂J '' ∂J ''
dJ (x, u) = x+ u=0. (2.134)
∂ξ 'ξ ∗ ,η∗ ∂η 'ξ ∗ ,η∗
The second variation of J about the extremal trajectory is given by:
1 T
d2 J (x, u) = x (tf )Qf x(tf )
2

1 tf T Q(t) S(t) x(t)
+ {x (t), uT (t)} T , (2.135)
2 t0 S (t) R(t) u(t)
which is a quadratic form with the following cost coefficient matrices:

' '
∂ 2 φ '' ∂ 2 L ''
Qf = ; Q(t) = (2.136)
∂ξ 2 'ξ ∗ (tf ) ∂ξ 2 'ξ ∗ ,η∗
and
' '
∂ 2 L '' ∂ 2 L ''
S(t) = ; R(t) = . (2.137)
∂ξ ∂η 'ξ ∗ ,η∗ ∂η2 'ξ ∗ ,η∗
Thus, the second variation of the objective function, d2 J , about the extremal
trajectory is a quadratic cost function that must be minimized subject to a linearized
dynamic equation for a neighboring extremal trajectory. This forms the basis of the
linear, quadratic, optimal control problem for neighboring extremal trajectories.
The Hamiltonian corresponding to the quadratic cost function, d2 J , and subject
to linear dynamic constraint of Eq. (2.131) is the following:
1 T 1
H = x (t)Q(t)x(t) + x T (t)S(t)u(t) + uT (t)R(t)u(t)
2 2
+ λT (t)[A(t)x(t) + B(t)u(t)] . (2.138)
The necessary conditions for optimality with a fixed terminal time, tf , are then given
by Euler–Lagrange equations as follows:
T
∂H
λ̇ = − = −Q(t)x(t) − S(t)u(t) − AT (t)λ(t) , (2.139)
∂x
T
∂φ
λ(tf ) = = Qf x(tf ) , (2.140)
∂x t=tf
∂H
= 0 = S T (t)x(t) + R(t)u(t) + B T (t)λ(t) , (2.141)
∂u
the last of which is solved for the optimal control as follows:

u(t) = −R −1 (t) S T (t)x(t) + B T (t)λ(t) . (2.142)
This is the well-known linear, quadratic regulator (LQR) control law. The sufficient
condition for optimality requires a positive definite Hessian:
∂ 2H
= R(t) > 0 , (2.143)
∂u2
which is satisfied by selecting a positive definite control cost matrix, R(t), and
ensures the existence of its inverse in Eq. (2.142).
Substitution of Eq. (2.142) into Eqs. (2.131) and (2.139) results in the following
set of linear state and costate equations for the LQR:

ẋ = A(t) − B(t)R −1 (t)S T (t) x(t) − B(t)R −1 (t)B T (t)λ(t) , (2.144)

λ̇ = − AT (t) − S(t)R −1 (t)B T (t) λ(t) + S(t)R −1 (t)S T (t) − Q(t) x(t) ,
(2.145)
that must be solved subject to the boundary conditions:
x(t0 ) = x0 ; λ(tf ) = Qf x(tf ) . (2.146)
The linear, two-point boundary value problem comprising Eqs. (2.144)–(2.146)

must be integrated in time by an iterative, numerical method such that the boundary
conditions are satisfied. However, since the state and costate vectors of the linearized
system are related by Eq. (2.140) at the final time, we look for a solution to
Eqs. (2.144)–(2.146) that satisfies a state transition matrix, thereby ensuring linear
independence of solutions, x(t) and λ(t). Thus, we write the combined state and
costate equations as follows:

ẋ A − BR −1 S T −BR −1 B T x
= (2.147)
λ̇ SR −1 S T − Q −AT + SR −1 B T λ
which must satisfy the boundary conditions of Eq. (2.146). To begin the solution,
we integrate backward in time from t = tf , for which we have

x(t) X(t) 0 x(tf )
= , (2.148)
λ(t) 0 Λ(t) λ(tf )
where X(t) and Λ(t) are the state transition matrices corresponding to the backward
evolution in time of x(t) and λ(t), respectively. Clearly, we must have
λ(t) = Λ(t)λ(tf ) = Λ(t)Qf x(tf ) , (2.149)
x(t) = X(t)x(tf ) , (2.150)
and
x(t0 ) = X(t0 )x(t) . (2.151)
Inversion of Eq. (2.150) and substitution into Eq. (2.149) yields
λ(t) = Λ(t)Qf X−1 (t)x(t) . (2.152)
Since both the state transition matrices must satisfy
X(tf ) = I ; Λ(tf ) = I. (2.153)
Equation (2.152) represents the adjoint relationship between the solutions, x(t) and
λ(t), written as follows:
λ(t) = P (t)x(t) , (2.154)

where P (t) = Λ(t)Qf X−1 (t). Substituting Eq. (2.154) into Eq. (2.142), we have
the following linear, optimal feedback control law:
u(t) = −R −1 (t)[B T (t)P (t) + S T (t)]x(t) . (2.155)
Taking the time derivative of Eq. (2.154) and substituting into Eqs. (2.144),
(2.145), and (2.155), we have the following matrix Riccati equation (MRE) to be
satisfied by the matrix, P (t):
− Ṗ = Q + (A − BR −1 S T )T P + P (A − BR −1 S T )
− P BR −1 B T P − SR −1 S T , (2.156)
which must be solved subject to the boundary condition:
P (tf ) = Qf . (2.157)
An alternative derivation of the matrix Riccati equation is via the HJB formu-
lation, where the following quadratic value function is proposed with a positive
definite, symmetric matrix, P (t):
V [x(t), t] = x T (t)P (t)x(t) , (2.158)
whose optimal value must satisfy the HJB equation with the terminal boundary
condition:
V̂ [x̂(tf ), tf ] = x T (tf )Qf x(tf ) . (2.159)
Thus, we have
∂V
λT = = 2x T (t)P (t) , (2.160)
∂x
and the following Hamiltonian:
H = x T (t)Q(t)x(t) + 2x T (t)S(t)u(t) + uT (t)R(t)u(t)

+ 2x T (t)P (t)[A(t)x(t) + B(t)u(t)] . (2.161)
For deriving the optimal control, we differentiate H with respect to u and equate
the result to zero Hu = 0:
x̂ T (t)S(t) + ûT (t)R(t) + x̂ T (t)P̂ (t)B(t) = 0 , (2.162)
or
û(t) = −R −1 (t)[B T (t)P̂ (t) + S T (t)]x̂(t) . (2.163)

which is the same as Eq. (2.155). Since the matrix P (t) is chosen to be positive
definite, we have V [x(t), t] > 0 for all x(t), and by substituting the optimal
feedback control law into
= −x̂ T P̂˙ x̂ ,
∂ V̂
Ĥ = − (2.164)
∂t
we have the Legendre–Clebsch sufficient condition, Huu > 0, implying a negative
definite matrix, P̂˙ (t) (i.e., V̇ [x(t), t] < 0 for all x(t)). Therefore, the optimal
return function given by Eq. (2.158) with a positive definite P̂ (t) ensures a globally
asymptotically stable system by Lyapunov’s theorem [74]. We again note that the
Legendre–Clebsch sufficient condition implies that R(t) is positive definite.
By substituting Eq. (2.163) into Eq. (2.161), we have the optimal Hamiltonian
which satisfies the following HJB equation, Eq. (2.103) for the LQR problem:
x̂ T P̂˙ x̂ = −x̂ T [(A − BR −1 S T )T P̂ + P̂ (A − BR −1 S T )

− P̂ BR −1 B T P̂ + Q − SR −1 S T ]x̂ , (2.165)
which yields the following matrix Riccati equation to be satisfied by the optimal
matrix, P̂ :
− P̂˙ = Q + (A − BR −1 S T )T P̂ + P̂ (A − BR −1 S T )
− P̂ BR −1 B T P̂ − SR −1 S T , (2.166)
which must be solved subject to the boundary condition:
P̂ (tf ) = Qf . (2.167)
Since the matrix Riccati equation (MRE) is derived both from the necessary
conditions (Euler–Lagrange equations) and the sufficient condition (Hamilton–
Jacobi–Bellman equation) of optimality, it must reflect both necessary and sufficient
conditions for the existence of an optimal control law for linear systems with
a quadratic performance index. As discussed earlier, a sufficient condition for
optimality is the existence of a positive definite solution to MRE, P (t), for all times
in the control interval, t0 ≤ t ≤ tf .
The MRE is at the heart of modern control theory, and can be regarded as
a simplified form of the nonlinear, two-point boundary value problem (2PBVP)
in time represented by the Euler–Lagrange equations, or the nonlinear partial
differential equations in space and time derived through Hamilton–Jacobi–Bellman
(HJB) formulation. However, being nonlinear in character, the solution procedure to
a general MRE is only slightly simpler, often requiring iterative numerical methods
similar to the nonlinear 2PBVP. Although simple iterative schemes based upon
repeated linear, algebraic solutions are usually applied when the coefficient matrices
are either slowly or periodically varying, the convergence to a positive definite (or
even positive semi-definite) solution is not always guaranteed. Given the complexity
of an MRE solution, it is much easier to directly solve the linear state and costate
equations, Eqs. (2.144)–(2.146), by an iterative 2PBVP procedure like the shooting,
or collocation methods. Linearity of the adjoint system of equations, Eq. (2.147),
assures the existence of a state transition matrix, Φ(t, t0 ), such that

x(t) x(t0 )
= Φ(t, t0 ) , (2.168)
λ(t) λ(t0 )
with the boundary conditions:
x(t0 ) = x0 ; λ(tf ) = Qf x(tf ) . (2.169)
The state transition matrix has the usual properties of
Φ(t0 , t0 ) = I ; Φ(t0 , t) = Φ −1 (t, t0 ) , (2.170)
and

A − BR −1 S T −BR −1 B T
Φ̇(t, t0 ) = Φ(t, t0 ) , (2.171)
SR −1 S T − Q −AT + SR −1 B T
as well as the special property of being a symplectic matrix2 , that is:

0 I 0 I
T
Φ (t, t0 ) Φ(t, t0 ) = . (2.172)
−I 0 −I 0
On partitioning Φ(t, t0 ) as follows:

x(t) Φxx (t, t0 ) Φxλ (t, t0 ) x(t0 )
= , (2.173)
λ(t) Φλx (t, t0 ) Φλλ (t, t0 ) λ(t0 )
2A symplectic matrix, A, satisfies AT J A = J , where

0 I
J =
−I 0
√
is analogous to the imaginary number, j = −1, in complex algebra since J 2 = −I . A symplectic
matrix, A, has the following properties:
1. Its determinant is unity, i.e., det(A) = 1.
2. If λ is an eigenvalue of A, then 1/λ is also an eigenvalue of A.
2.10 Illustrative Examples 43
the symplectic nature of the state transition matrix implies that

⎡ ⎤
T (t, t )
Φλλ 0 −Φxλ
T (t, t )
0
Φ −1 (t, t0 ) = ⎣ ⎦ , (2.174)
−Φλx
T (t, t )
0
T (t, t )
Φxx 0
which is very useful in carrying out the matrix operations required for the solution
of the 2PBVP.
In order to prove that Φ(t, t0 ) is indeed symplectic, we note that Φ(t0 , t0 ) = I
satisfies Φ T J Φ = J . Therefore, if it can be shown that
d
(Φ T J Φ) = 0 , (2.175)
dt
then the identity Φ T J Φ = J is satisfied for all times. Making use of Eq. (2.171),
which we re-write as follows:

E F
Φ̇ = Φ, (2.176)
G −E T
where
E = A − BR −1 S T
F = −BR −1 B T (2.177)
G = SR −1 S T − Q ,
we have
d
(Φ T J Φ) = Φ̇ T J Φ + Φ T J Φ̇
dt

G − GT E T − E T
=Φ T
Φ=0, (2.178)
E−E F − FT
because F, G are symmetric matrices. Thus, we have proved that Φ(t, t0 ) is

symplectic.
2.10 Illustrative Examples
Example 2.6 Consider a second-order, linear time-invariant system with the follow-
ing state equations:
ẋ1 = x2
ẋ2 = x1 + u (2.179)
where the scalar input u(t) is unconstrained. The control task is to move the system
from initial state x1 (0) = x2 (0) = 0 to x1 (5) = 1 while minimizing
5
1 1
J = x22 (5) + u2 (t)dt (2.180)
2 2 0
Note that the terminal time is specified to be tf = 5, whereas the terminal cost
function is given as φ = 12 x22 (5). The Lagrangian, L = 12 u2 , represents the total
control effort (or energy) spent during the control interval, 0 ≤ t ≤ 5.
We begin by writing the Hamiltonian of the system as follows:
1 2
H = L + λT f = u + λ1 x2 + λ2 (x1 + u) (2.181)
2
The costate equations of the system are next derived as follows:
T
∗ ∂H ∗
λ̇ = − = Λλ∗ (2.182)
∂x
where

0 −1
Λ= (2.183)
−1 0
The solution to the costate equations subject to the initial condition, λ∗ (0) =
(c1 , c2 )T , is expressed by:
⎛ ⎞⎧ ⎫
et + e−t −et + e−t
⎨ c1 ⎬
1
λ∗ (t) = eΛt λ∗ (0) = L−1 (sI − Λ)−1 λ∗ (0) = ⎝ ⎠
2 ⎩ ⎭
−et + e−t et + e−t c2
(2.184)
Note that the constants c1 , c2 (thus the extremal costate trajectory) are to be
determined from the boundary conditions.
The necessary conditions for optimality include the following stationary condi-
tion of the Hamiltonian with respect to the control history, u(t):
∗ T
∂H
Hu = = u∗ (t) + λ∗2 (t) = 0 (2.185)
∂u
which results in the following extremal control history:
1 $ −t %
u∗ (t) = −λ∗2 (t) = − c1 e − et + c2 et + e−t
2
1$ %
= (c1 − c2 ) et − (c1 + c2 ) e−t , (0 ≤ t ≤ 5) (2.186)
2
In order to determine the constants c1 , c2 , the state equations must be solved, subject
to the boundary conditions. The extremal trajectory resulting from the extremal
control history (Eq. (2.186)) and zero initial state is expressed as follows:
t
x ∗ (t) = eA(t−τ ) Bu∗ (τ )dτ , (0 ≤ t ≤ 5) (2.187)
0
where

0 1 0
A= , B= (2.188)
1 0 1
The state transition matrix is determined to be the following:

⎛ ⎞
et + e−t et − e−t
1⎝ ⎠
eAt = L−1 (sI − A)−1 = (2.189)
2
et − e−t et + e−t
After Eqs. (2.186) and (2.189) are substituted into Eq. (2.187), and the integration is
performed, we have the following extremal trajectory:
⎧ −t $ %⎫
1 ⎨ c1 e − et + t (c1 − c2 ) et + (c1 + c2 ) e−t ⎬
x ∗ (t) = , (0 ≤ t ≤ 5)
4 ⎩ −t $ %⎭
c2 e − et + t (c1 − c2 ) et − (c1 + c2 ) e−t
(2.190)
The boundary conditions are given by:
x1∗ (5) = 1 (2.191)
and
∗
∂φ
λ∗2 (5) = = x2∗ (5) (2.192)
∂x2 t=5
Substituting the boundary conditions into Eqs. (2.184) and (2.190) results in the
following set of linear algebraic equations:
1 5 5 −5
2e + 3e−5 c1 + e − e5 c2 = 1
2 4

7 5 −5 3 −5
e −e c1 − e + 2e c2 = 0
5
(2.193)
4 2
whose solution is easily derived to be the following:
c1 = −0.072 , c2 = −0.063 (2.194)

1 0.8
0.6
x∗1
x2∗
0.5 0.4
0.2
0 0
0 1 2 3 4 5 0 1 2 3 4 5
t t
0 1
−0.2
0.5
λ ∗1
−0.4 λ ∗2
0
−0.6
−0.8 −0.5
0 1 2 3 4 5 0 1 2 3 4 5
t t
Fig. 2.2 The extremal trajectory and the corresponding costate trajectory for the optimal control
problem of Example 2.6
Thus, the optimal control problem is solved, and the extremal trajectory and the
corresponding control history are obtained. The results are plotted in Figs. 2.2
and 2.3. It can be verified that x2∗ (5) = λ∗2 (5) = 0.6689 and x1∗ (5) = 1.
It is quite easily verified that the extremal trajectory is also the optimal one,
because
∗ T
∂ 2H
Huu = =1>0 (2.195)
∂u2
hence the sufficient condition for optimality is satisfied on the extremal trajectory.
Example 2.7 Let the optimizing conditions for the system in Example 2.6 be
changed such that the terminal time, tf , is unspecified, and the system is to be moved
from a zero initial state, x1 (0) = x2 (0) = 0 to x1 (tf ) = 1, while minimizing
tf
1 2 1
J = x2 (tf ) + u2 (t)dt (2.196)
2 2 0
While the state and costate solutions are still given by Eqs. (2.190) and (2.184),
respectively, and the stationary condition:
−0.1
−0.2
−0.3
u∗
−0.4
−0.5
−0.6
−0.7
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t
Fig. 2.3 The extremal control history for the optimal control problem of Example 2.6
Hu = 0 = u∗ (t) + λ∗2 (t) (2.197)
is unchanged, the boundary conditions are now changed to the following:
x1∗ (tf ) = 1 (2.198)

∗
∂φ
λ∗2 (tf ) = = x2∗ (tf ) (2.199)
∂x2 t=tf
and

∂φ ∗ 1$ ∗ %2
H+ = u (tf ) + λ∗1 (tf )x2∗ (tf ) + λ∗2 (tf )x1∗ (tf ) = 0 (2.200)
∂t t=tf 2
Substituting Eqs. (2.197)–(2.199) into Eq. (2.200), we have
1$ ∗ %2
− λ (tf ) + λ∗1 (tf )λ∗2 (tf ) + λ∗2 (tf ) = 0 (2.201)
2 2
or
$ %
λ∗2 (tf ) = 2 λ∗1 (tf ) + 1 (2.202)
The algebraic equations in terms of the unknowns, c1 , c2 , tf , resulting from the

boundary conditions are the following:
tf $ % c1 t
(c1 − c2 ) etf + (c1 + c2 ) e−tf − e f − e−tf = 1 (2.203)
4 4
t
− c1 3e f + e−tf + c2 3etf − e−tf = 4 (2.204)
and

c1 etf − e−tf tf + 2
c2 = t (2.205)
3e f + e−tf + tf etf + e−tf
When Eqs. (2.204) and (2.205) are substituted into Eq. (2.203), the following
necessary condition is derived for the terminal time, tf :
9e3tf + 45etf + 11e−tf − e−3tf + 66tf etf + 26tf e−tf + 6tf e3tf − 2tf e−3tf
+ 21tf2 etf + 11tf2 e−tf + tf2 e3tf − tf2 e−3tf = 0 (2.206)
Newton’s iterative solution method [77] yields the only real solution to Eq. (2.206)
as tf = −0.7167, which is invalid because the terminal time must always be
positive. Hence, the stated optimal control problem does not have a feasible solution.
This example illustrates the fact that not every set of objective function and
boundary conditions has a feasible solution.
Example 2.8 Consider another variation for the system in Example 2.6 where the
terminal time, tf , is unspecified, and the system is to be moved from a zero initial
state, x1 (0) = x2 (0) = 0, to final state, x1 (tf ) = x2 (tf ) = 1, while minimizing
tf
1 1
J = k 2 tf2 + u2 (t)dt (2.207)
2 2 0
Since both the states are fixed at the terminal time, we have the following boundary
conditions for determining the coefficients, c1 , c2 , in Eqs. (2.184) and (2.190):
$ %
tf (c1 − c2 ) etf + (c1 + c2 ) e−tf − c1 etf − e−tf = 4 (2.208)
and
$ %
tf (c1 − c2 ) etf − (c1 + c2 ) e−tf − c2 etf − e−tf = 4 (2.209)
These are solved to yield

$ %
4 + c2 etf − e−tf + tf etf + e−tf
c1 = (2.210)
tf etf − e−tf
Table 2.1 The results of k tf c1 c2

iterative solution in
Example 2.8 0.1 2.9026 −0.2495 −0.2409
1 1.4528 −2.3939 −1.7046
10 0.6490 −33.0767 −11.3930
100 0.2723 −522.0565 −73.7958
1.4
k = 0.1
k = 1
1.2 k = 10
k = 100
1
0.8
x1
0.6
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3
t
Fig. 2.4 The extremal trajectory of the first state variable for different values of the terminal cost
parameter, k, in Example 2.8
where

4 e−tf + etf + 2tf e−tf
c2 = 2 (2.211)
etf − e−tf − 4tf2
The additional boundary condition required to solve for the terminal time, tf , is the
following:

∂φ ∗ 1$ %2
H+ = − λ∗2 (tf ) + λ∗1 (tf ) + λ∗2 (tf ) + k 2 tf = 0 (2.212)
∂t t=tf 2
wherein the terminal boundary condition for the state variables has been substituted,
and the costate variables, λ∗1 , λ∗2 , are given by Eq. (2.184). Equation (2.212) requires
an iterative numerical solution, which is performed by Newton’s method [77].
The results when the cost coefficient k is varied from 0.1 to 100 are tabulated in
Table 2.1 and Figs. 2.4, 2.5, and 2.6. The terminal time, tf , is seen to decrease
almost exponentially with the increase of k whereas the coefficients, c1 , c2 , are
6
k = 0.1
k = 1
5 k = 10
k = 100
4
x2
0
0 0.5 1 1.5 2 2.5 3
t
Fig. 2.5 The extremal trajectory of the second state variable for different values of the terminal
cost parameter, k, in Example 2.8
almost linearly related to k. As k is increased, the state variable, x1∗ , reaches its
final value in a smaller time in a continuously ascending curve (Fig. 2.4), whereas
x2∗ encounters increasing overshoots in its trajectory for k > 1 (Fig. 2.5). The largest
control input magnitude, | u∗ |, increases in proportion with k, its peak magnitude
being roughly the same as k (Fig. 2.6). Hence, the cost parameter, k, offers a direct
trade-off between the terminal time, tf , and the required peak input.
Example 2.9 In order to illustrate the effect of the boundary conditions, consider
a lunar landing problem in which a continuous vertical acceleration input, u(t), is
applied to control the altitude, z(t), of the spacecraft’s center of mass. The lander
is assumed to have approached sufficiently near the lunar surface in order for the
acceleration due to gravity, g, to be considered a constant. The differential equation
governing the spacecraft’s translational motion is thus simply given by:
z̈ = −g + u (2.213)
Let us choose the state variables to be the following:
1
x1 = z + gt 2
2
x2 = ẋ1 = ż + gt (2.214)
0.3 2
0.2 k = 0.1 k=1

1
0.1
u
u
0
0
−0.1 −1
0 1 2 3 0 0.5 1 1.5
t t
20 100
10
k = 10 50 k = 100
0
u
0
−50
−10 −100
0 0.2 0.4 0.6 0.8 0 0.1 0.2 0.3 0.4
t t
Fig. 2.6 The extremal control history for different values of the terminal cost parameter, k, in
Example 2.8
The state equations can then be expressed as follows:
ẋ1 = x2
ẋ2 = u (2.215)
This would appear to be no different than the simple linear acceleration problem
in a zero gravity situation, but this is deceptive because although the acceleration
due to gravity does not appear in the state equations, it is very much a part of the
boundary conditions. For a smooth landing to take place, both vertical displacement
and vertical speed must vanish at some terminal time, tf , implying
1 2
x1 (tf ) = gt
2 f
x2 (tf ) = gtf (2.216)
Therefore, a solution to the optimal landing problem must be sought where
Eq. (2.216) is imposed as a terminal constraint. This must be ensured irrespective
of the initial condition:
x1 (0) = z(0)
x2 (0) = ż(0) (2.217)
A possible choice of the objective function for the smooth vertical landing is the
total control energy given by:
tf
1
J = u2 (t)dt (2.218)
2 0
1 2
H = u + λ1 x2 + λ2 u (2.219)
2
The costate equations of the system are next derived as follows:
T
λ∗1 ∂H ∗ 0
λ̇∗ = =− = (2.220)
λ∗2 ∂x −λ∗1
whose solution subject to the initial condition, λ∗ (0) = (c1 , c2 )T , is simply the
following:
⎧ ⎫
⎨ c1 ⎬
λ∗ (t) = (2.221)
⎩ ⎭
c2 − c1 t
The stationary condition of the Hamiltonian:

∗ T
∂H
Hu = = u∗ (t) + λ∗2 (t) = 0 (2.222)
∂u
results in the following expression for the extremal control history:
u∗ (t) = −λ∗2 (t) = c2 − c1 t, (0 ≤ t ≤ tf ) (2.223)
This is substituted into the state equations, Eq. (2.215), to yield the extremal
trajectory as follows:
⎧ ⎫
⎨ c1 t 2 /2 − c2 t + c3 ⎬
x1∗
x ∗ (t) = = , (0 ≤ t ≤ tf ) (2.224)
x2∗ ⎩ ⎭
c1 t 3 /6 − c2 t 2 /2 + c3 t + c4
where x ∗ (0) = (c3 , c4 )T . The initial condition of the system, Eq. (2.217), requires
that c3 = ż(0) and c4 = z(0). Finally, the terminal condition, Eq. (2.216), is applied
to determine the coefficients, c1 , c2 , as follows:
800
600
x∗1(m)
400
200
0
0 5 10 15 20 25 30
t(s)
60
40
x∗2 (m/s)
20
0
0 5 10 15 20 25 30
t(s)
Fig. 2.7 The optimal state trajectory for the lunar landing problem of Example 2.9
⎛ ⎞⎧ ⎫
6/tf2 −12/tf3 ⎪
⎨ −ż(0) + gtf ⎪
⎬
c1 ⎜ ⎟
=⎝ ⎠ (2.225)
c2 ⎪ ⎪
2/tf −6/tf2 ⎩ −z(0) − ż(0)tf + gtf2 /2 ⎭
Since Huu = 3/2 > 0 on the extremal trajectory, the latter is also the optimal
trajectory.
Figures 2.7, 2.8, and 2.9 are the plots of the optimal state trajectory, control
history, and the vertical displacement and velocity, respectively, for g = 1.6 m/s2 ,
z(0) = 100 m, ż(0) = 10 m/s, and tf = 30 s. For this condition, the values
of the coefficients are c1 = 1/9 and c2 = 2/5. The optimal control increases
linearly from −0.4 m/s2 to 2.93 m/s2 (Fig. 2.8), whereas the vertical displacement
(Fig. 2.9) has an overshoot of 28 m at t = 6 s before smoothly decreasing to
zero at the terminal time. This overshoot is due to the fact that the spacecraft is
initially traveling upward, thereby requiring a negative vertical acceleration input
to be applied initially in order to make it land in the given time. By increasing the
value of tf , one can decrease the magnitude of the control input as well as the
vertical overshoot before landing. By keeping the terminal time open and adding a
terminal cost to the objective function, one can optimally combine the requirements
of minimizing both the control energy and the time. This is to be explored in one of
the exercises at the end of the chapter.
Caution must be exercised when choosing a too small value of tf , which would
not only greatly increase the control input magnitudes, but more importantly, can
2.5
1.5
u∗(m/s2)
0.5
−0.5
0 5 10 15 20 25 30
t(s)
Fig. 2.8 The optimal control history for the lunar landing problem of Example 2.9
150
100
z∗ (m)
50
0
0 5 10 15 20 25 30
t(s)
10
5
ż ∗ (m/s)
−5
−10
0 5 10 15 20 25 30
t(s)
Fig. 2.9 The optimal vertical displacement and velocity for the lunar landing problem of
Example 2.9
2.11 Singular Optimal Control 55
also lead to an undershoot in the displacement near the terminal time. This implies a
crash into the ground at a nonzero vertical velocity, and is automatically avoided by
adding an inequality constraint to the optimization problem, such as x1 (t) ≥ 12 gt 2 ,
0 ≤ t ≤ tf . However, the resulting problem would be a nonlinear one as opposed
to the linear one solved in the present example.
2.11 Singular Optimal Control
A singular control problem is the one in which the necessary conditions of

optimality do not determine the optimal control profile. In such a case, additional
conditions must be specified in order to define the optimal control. A large majority
of optimal control problems in space navigation have a Hamiltonian which is a linear
function of the control variables, and therefore the stationary condition, Hu = 0,
cannot be used to solve for u(t). Furthermore, such a Hamiltonian does not satisfy
the Legendre–Clebsch sufficient condition for optimality, Huu > 0, because the
Hessian, Huu , is singular. Examples of singular control include the minimization
of a performance index which is a linear function of the control variables, such
as the time minimization and fuel minimization maneuvers. Other singular optimal
control problems are the ones in which the state equations are linearized, the control
cost in the performance index is quadratic, and the control variables are bounded
by inequality constraints. Since the sufficient condition for optimality is unusable in
singular control problems, additional conditions must be available to determine the
minimizing input and trajectory.
As an illustrative example of singular control, consider a control-affine system
described by the following state equation:
ẋ = f (x, t) + B(x, t)u(t) (2.226)
and the following Lagrangian which is linear in control:
L(x, u, t) = 1 (x, t) + 2 (t)u(t) (2.227)
H (x, u, λ, t) = 1 + 2 u + λT (f + Bu) (2.228)
Unless other conditions are specified, there is no solution to the optimal control
problem in which H is minimized relative to u, because Huu = 0 on an extremal
trajectory for which Hu = 0. Even though the non-convexity of H does not
guarantee optimality, a practical minimum solution is obtained for the case of
bounded control inputs, such as that given by the following inequality constraint:
| u(t) |≤ um (2.229)
In such a case, the Pontryagin’s minimum principle given by Eq. (2.94) yields the
following condition for minimality on an optimal trajectory, (x̂, û, λ̂):
1 (x̂, t) + 2 (t)û(t) + λ̂T [f (x̂, t) + B(x̂, t)û(t)]

≤ 1 (x̂, t) + 2 (t)u(t) + λ̂T [f (x̂, t) + B(x̂, t)u(t)] (2.230)
or
[2 (t) + λ̂T B(x̂, t)]û(t)] ≤ [2 (t) + λ̂T B(x̂, t)]u(t)] (2.231)
This results in the following switching condition for the optimal control, û(t):
⎧
⎪
⎪ −um , (2 + λ̂T B)i > 0
⎪
⎪
⎪
⎨
ûi (t) = 0 , (2 + λ̂T B)i = 0 , (i = 1, . . . , m) (2.232)
⎪
⎪
⎪
⎪
⎪
⎩
um , (2 + λ̂T B)i < 0
where the index i refers to the i th element of the particular vector. The magnitude of
each element of û thus depends upon the sign of the switching function, (2 + λ̂T B)i ,
and can also be expressed as follows:
ûi (t) = −um sgn(2 + λ̂T B)i , (i = 1, . . . , m) (2.233)
Clearly, the optimality requires that the norm of the control input vector must be on
the boundary, that is either the maximum or the minimum value. Such an optimal
control law switching between the two extreme input values is typical of the singular
problems and is referred to as bang–bang control. Note that the bound on the norm
of the input vector, Eq. (2.229), implies that if the i th element is either ûi = −um or
ûi = um , then ûj = 0 for all i = j .
Example 2.10 Let us now take up an example of time-optimal control with bounded
control magnitude as an example of singular control problems. For illustration, the
following linear system is chosen:
ẋ1 = x2
ẋ2 = u (2.234)
which can represent a rigid spacecraft undergoing a single axis rotation with u(t)
being the angular acceleration input and x1 (t) the rotation angle. Alternatively, these
state equations also model the rectilinear displacement, x1 (t), of a mass moving
in a frictionless groove when forced by a linear acceleration input, u(t). Consider
moving the system from x(0) = (0, 0)T to x(tf ) = (1, 0)T in the minimum time,
tf , with the following restriction on the input magnitude:
|u(t)| ≤ 1 (2.235)
Since the Lagrangian of a time-optimal control problem is L = 1, the Hamiltonian

of the system is given by:
H = 1 + λT f = 1 + λ1 x2 + λ2 u (2.236)
while Pontryagin’s minimum principle applied to this singular problem results in

the following switching condition:

−1 , λ̂2 (t) > 0
û(t) = (2.237)
1 , λ̂2 (t) < 0
where
λ̂1 = c1
λ̂2 = −c1 t + c2 (2.238)
The constants c1 , c2 must be determined in order to meet the boundary conditions

in the minimum possible time, tf , after the switching is applied at t = tˆ. Since λ̂2 (t)
is a linear function of time, it can change sign only once. Therefore, a change in the
system’s state from x(0) = (0, 0)T to x(tf ) = (1, 0)T requires that a positive
control, û(t) = 1, must be applied before switching, t < tˆ, which implies λ̂2 (t) < 0
for 0 ≤ t < tˆ. Clearly, we must have c2 < 0 and tˆ = c2 /c1 > 0, thereby implying
c1 < 0. Thus, we have
t 2
1 t −τ 0 t /2
x̂(t) = dτ = (0 ≤ t < tˆ) (2.239)
0 0 1 1 t
After switching at t = tˆ, the optimal control changes to û(t) = −1, resulting in the
following state response:
t
1 t − tˆ tˆ2 /2 1 t −τ 0
x̂(t) = − dτ
0 1 ˆt tˆ 0 1 1

2t tˆ − tˆ2 − t 2 /2
= (tˆ < t ≤ tf ) (2.240)
2tˆ − t
In order to meet the terminal condition, we must have x̂(tf ) = (1, 0)T , which
requires tˆ = tf /2 and tf = 2. The resulting bang–bang optimal trajectory is plotted
in Fig. 2.10 with a time step of 1/100. Since there can be only one switching due
to the time-linear profile of the costate variable, λ̂2 (t), the resulting terminal time
0.5
x
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
1
0.5
ẋ
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
1
0
u
−1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t
Fig. 2.10 The optimal bang–bang trajectory and control history of the time-optimal control
problem of Example 2.10
tf is the absolute minimum, with the switching performed at the midpoint of the
control interval, tf /2. The velocity, ẋ, reaches a maximum at the switching time
and decreases to zero at the final time.
2.11.1 Generalized Legendre–Clebsch Necessary Condition
Since a singular control history, u(t), satisfies the Legendre–Clebsch necessary

condition, Huu ≥ 0, with an equality, it may (or may not) be the optimal control.
In order to determine the optimality of a singular control solution, a generalized
Legendre–Clebsch necessary condition must be derived. Most of the singular
control problems in space navigation involve a scalar control variable, hence we
first consider the optimal control problem of minimization of
tf
J = φ[x(tf ), tf ] + L(x, t)dt (2.241)
t0
with respect to a scalar control variable, u(t) ∈ R, which satisfies the control-affine
dynamic constraint (state equation):
ẋ = f (x, t) + b(x, t)u (2.242)

with a specified initial state:
x(t0 ) = x0 (2.243)
and the terminal state satisfying:
F [x(tf )] = 0 (2.244)
where tf is specified in advance. The Hamiltonian:
H = L(x, t) + λT [f (x, t) + b(x, t)u] (2.245)
is linear in control, u, thus the optimal control problem is singular.

Theorem 2.11 If u is the optimal control solution of the problem given by
Eqs. (2.241)–(2.244), then
∂ d2 ∂ d2 T
2
Hu = λ b≤0 (2.246)
∂u dt ∂u dt 2
for all t ∈ [t0 , tf ].
The proof of this theorem is derived by taking the second-order variation of the
augmented performance index along the singular trajectory, and requiring it to be
non-negative. See, for example, Kelley et al. [46].
Equation (2.246) is the modified necessary condition for optimality of singular
control problem, that is when Hu = 0. If satisfied with a strict inequality, this
condition gives the information to determine the optimal singular control.
In case Eq. (2.246) is satisfied with an equality, it does not give any information
regarding the optimal control, u. In that case, the following condition must be
checked:
∂ d4 ∂ d4 T
Hu = λ b≤0 (2.247)
∂u dt 4 ∂u dt 2
for all t ∈ [t0 , tf ].
Once again, if Eq. (2.247) is satisfied with an equality, it does not give any
information regarding the optimal control, u. In that case, the next higher derivative
must be considered. Kelley et al. [46] derived the following generalized necessary
condition for optimality:
∂ d2q
(−1)q Hu ≥ 0 (2.248)
∂u dt 2q
for all t ∈ [t0 , tf ], where q = 1, 2, . . .. Thus, an odd time derivative of Hu cannot
appear in the generalized necessary condition for the scalar input.
Example 2.12 Consider the minimization of

π
J = x22 − x12 dt
0
relative to u ∈ R, subject to:
ẋ1 = x2 , x1 (0) = 0
ẋ2 = u , x2 (0) = 1
The Hamiltonian of the problem is the following:

H = x22 − x12 + λ1 x2 + λ2 u
which results in the following costate equations:
λ̇1 = 2x1 , λ1 (π ) = 0
λ̇2 = −2x2 − λ1 , λ2 (π ) = 0
The extremal solution of this singular problem is derived as follows:
Hu = 0 = λ∗2
or
λ∗1 = −2x2∗
which implies that the second state equation is the following along the extremal arc:
ẋ2∗ = u∗ = −x1∗
The extremal trajectory is thus represented by:

ẋ1∗ 0 1 x1∗ x1∗ (0) 0
= , =
ẋ2∗ −1 0 x2∗ x2∗ (0) 1
whose solution is the following:

x1∗ (t) sin t
=
x2∗ (t) cos t
Thus, u∗ (t) = − sin t, 0 ≤ t ≤ π is the singular control history.

To investigate the optimality of the singular control profile, we apply Theo-

rem 2.11 on the extremal arc as follows:
d2
Hu = λ̈∗2 = −2ẋ2∗ − λ̇∗1 = −2u∗ − 2x1∗
dt 2
∂ d2
Hu = −2 < 0
∂u dt 2
Thus, Theorem 2.11 is satisfied for all t, implying that u∗ (t) = − sin t, 0 ≤ t ≤ π
could be an optimal solution of the problem. However, since it is not a sufficient
condition for optimality, additional information is necessary to determine whether
the singular solution is optimal.
The generalized Legendre–Clebsch necessary condition for multi-input singular
optimal control can be derived using a transformational approach [34], wherein
the singular control problem is converted into an equivalent non-singular one by
transforming the control variables. For the control-affine system with u ∈ Rm×1
and x ∈ Rn×1 , represented by:
ẋ = f (x, t) + g(x, t)u (2.249)
and with a specified initial state:
x(t0 ) = x0 (2.250)
the minimization of
tf
J = φ[x(tf ), tf ] + L(x, t)dt (2.251)
t0
subject to the dynamic equality constraint, Eq. (2.249), and a terminal state con-
straint:
F [x(tf ), tf ] = 0 (2.252)
results in the Hamiltonian:
H = L(x, t) + λT [f (x, t) + g(x, t)u] (2.253)
which is linear in control, u. The augmented cost function in terms of the

Hamiltonian is expressed as follows:
tf
Ja = φ[x(tf ), tf ] + c F [x(tf ), tf ] +
T
H (x, u, λ, t) − λT ẋdt (2.254)
t0
The Euler–Lagrange necessary conditions for optimality along an extremal arc,

[x ∗ (t), u∗ (t)], are based on the following first-order variation of Ja along the
extremal trajectory. The first-order variations in the state equation and the boundary
conditions due to δx, δu, δt are expressed as follows:
δ ẋ = Aδx + Bδu (2.255)
δx(t0 ) = 0 (2.256)
and
[Ft δt + Fx (δx + ẋδt)]t=tf = 0 (2.257)
The first-order variation in Ja is the following:

dJa = (φt + cT Ft + H − λT ẋ)dt + (φx + cT Fx )dx
t=tf
tf
+ Hx δx + Hu δu − λT δ ẋ dt (2.258)
t0
As before, we integrate the term involving δ ẋ by parts, express the total derivative
dx = ẋdt + δx, and note that δx(t0 ) = 0. This produces the following first-order
variation:

dJa = (φt + cT Ft + H )δt + (φx + cT Fx − λT ) (ẋδt + δx)
t=tf
tf
+ Hx + λ̇T δx + Hu δu dt (2.259)
t0
Since dJa = 0 in the presence of arbitrary variations δx, δu, δt, we have the
following necessary conditions for minimization of Ja :
∗
∂H
λ̇T = − = −Hx∗ (2.260)
∂x
∗
λT (tf ) = φx + cT Fx (2.261)
t=tf
∗
φt + c T F t + H =0 (2.262)
t=tf
and Hu∗ = 0. Here, A = (∂f/∂x)∗ is the Jacobian matrix, and B = g(x ∗ , t), both
being evaluated on the extremal arc.
The second variation of Ja is now derived by taking the total derivative of
Eq. (2.258) as follows:

d2 Ja = (φt + cT Ft + H − λT ẋ)d2 t + (φx + cT Fx )d2 x
+ (φtt + cT Ftt + Ḣ − λT ẍ)dt 2 + 2(φtx + cT Ftx )dtdx
+ dx T (φxx + cT Fxx )dx + 2Hx δxdt + 2Hu δudt − 2λT δ ẋdt (2.263)
t=tf
tf
+ Hx δ 2 x + Hu δ 2 u − λT δ 2 ẋ+ δx T Hxx δx+ 2δuT Hux δx+ δuT Huu δu dt
t0
where the time derivative of the Hamiltonian is the following:
Ḣ = Ht + Hx ẋ + Hλ λ̇ + Hu u̇ (2.264)
which becomes
Ḣ = Ht + Hx ẋ ∗ − ẋ ∗ HxT + Hu u̇∗ = Ht (2.265)
on the extremal trajectory. Furthermore, integrating the term involving δ ẋ by parts,

expressing the total derivatives of x by:
dx = ẋdt + δx
d2 x = ẋd2 t + ẍdt 2 + 2δ ẋdt + δ 2 x (2.266)
the terminal state variation by:
dx(tf ) = δx(tf ) + ẋ(tf )δtf (2.267)
and substituting Eqs. (2.260)–(2.262) into Eq. (2.264), we have the following
expression for the second variation:

d2 Ja = Φtt δt 2 + Φtx (δx + ẋδt)δt + (δx + ẋδt)T Φxx (δx + ẋδt)
+ (Ht − Hx ẋ)δt 2 + Hx (δx + ẋδt)δt (2.268)

t=tf
tf
+ δx T Hxx δx + 2δuT Hux δx + δuT Huu δu dt
t0
where
Φ[x(tf ), tf ] = φ[x(tf ), tf ] + cT F [x(tf ), tf ] (2.269)
When the terminal time is specified, we have δtf = 0, hence the terminal
constraint variation becomes the following:
(Fx δx)t=tf = 0 (2.270)

As before, a necessary condition for minimization of Ja is that of a positive semi-

definite second variation, d2 Ja , which (as discussed earlier) requires that for the
non-singular case (Huu > 0) the following matrix Riccati equation must be satisfied
with a non-negative, finite square matrix, P (t), along the extremal arc for a specific
terminal time, tf :
−1
− Ṗ = Hxx + P A + AT P − (Hxu + P B)Huu (Hux + B T P ) (2.271)
with the terminal boundary condition:
P (tf ) = Φxx [x(tf ), tf ] (2.272)
As before, we introduce the following notation for convenience:
Hxx = Q , Hux = S , Huu = R
(Φxx )t=tf = Qf , (Fx )t=tf = D
The requirement of a non-negative second variation of Ja with a specific terminal

time, tf , results in the following accessory problem of minimization of
tf
J2 = x(tf ) Qf x(tf ) +
T
x T Qx + 2uT Sx + uT Ru dt (2.273)
t0
subject to:
ẋ = A(t)x + B(t)u , x(t0 ) = 0 (2.274)
and
Dx(tf ) = 0 (2.275)
For the non-singular case, R = Huu > 0, this is the standard linear, quadratic
regulator (LQR) problem discussed above, and has the solution:
u(t) = −R −1 (t)[B T (t)P (t) + S T (t)]x(t) . (2.276)
where P (t) is a non-negative matrix satisfying
− Ṗ = Q + P A + AT P − (S T + P B)R −1 (S + B T P ) (2.277)
with the terminal boundary condition:
P (tf ) = Qf (2.278)
However, for the present singular case we have R = 0, hence there is no solution to
the Riccati equation, Eq. (2.277).
In order to transform the singular problem to a non-singular one (for which the
Riccati solution is available), consider the following definition:
v̇ = u (2.279)
with initial condition:
v(t0 ) = 0 (2.280)
Furthermore, a new state vector is defined by:
z = x − Bv (2.281)
Thus, the first-order variational state equation and boundary conditions, Eqs. (2.274)
and (2.275), become the following:
ż = Az + (AB − Ḃ)v , z(t0 ) = 0 (2.282)
and
D[z(tf ) + Bv(tf )] = 0 (2.283)
The objective function to be minimized associated with the second-order variation

of the original performance index is now the following:
J2 = [z(tf ) + B(tf )v(tf )]T Qf [z(tf ) + B(tf )v(tf )] (2.284)

tf
+ zT Qz + 2v T B T Qz + v T B T QBv + 2v̇ T S(z + Bv) dt
t0
H = zT Qz + 2v T B T Qz + v T B T QBv + 2v̇ T S(z + Bv) + λT [Az + (AB − Ḃ)v]

(2.285)
and the following costate equation:
λ̇ = −HzT = −Qz − 2QBv − 2S T v̇ − AT λ (2.286)
On integrating the term v̇ T Sz in Eq. (2.284) by parts, we have the following

result:
J2 = zT (tf )Qf z(tf ) + 2v T (tf )[S(tf ) + B T (tf )Qf ]z(tf )

tf
+ v T (tf )B T (tf )Qf B(tf )v(tf ) + zT Qz (2.287)
t0

+ 2v T (B T Q − SA − Ṡ)z + v T [B T QB − 2S(AB − Ḃ)]v + 2v̇SBv dt
However, from Eq. (2.285) it can be shown that:
d ∂
B T QB − 2S(AB − Ḃ) − (SB) = − Ḧu (2.288)
dt ∂u
A substitution of Eq. (2.288) into Eq. (2.287) and integration of the last term inside
the integral by parts results in the following:
J2 = zT (tf )Qf z(tf ) + 2v T (tf )[S(tf ) + B T (tf )Qf ]z(tf )

tf
+ v T (tf )[S(tf )B(tf ) + B T (tf )Qf B(tf )]v(tf ) + zT Qz
t0

∂
+ 2v (B Q − SA − Ṡ)z + 2v − Ḧu v
T T T
(2.289)
∂u

+ v̇ T (SB − B T S T )v dt
The term (SB − B T S T ) in Eq. (2.289) is a skew-symmetric matrix, which can make
J2 either positive or negative. But, this cannot be allowed if J2 has a minimum on
the extremal arc. Hence, the following necessary condition for optimality must be
enforced:
SB − B T S T = 0 for all t ∈ [t0 , tf ] (2.290)
which makes the v̇ term vanish inside the integral of Eq. (2.289). It is to be noted
that
∂
SB − B T S T = Ḣu (2.291)
∂u
which corresponds to the generalized necessary Legendre–Clebsch condition given
by Theorem 2.11 (Eq. (2.248)) with q = 1 for the scalar case, and is trivially
satisfied even for the vector case when the condition of Eq. (2.290) is imposed.
Therefore, another transformation is necessary to derive the nontrivial generalized
optimality condition by taking:
ẇ = v , w(t0 ) = 0 (2.292)
for which the second variation becomes the following:
J2 = ψ[ẏ(tf ), y(tf )]
tf
+ ẏ T N ẏ + 2ẏ T Uy + y T My dt
t0
and must be minimized subject to the first-order state variation:
B̄ ẏ + Āy = 0 , y(t0 ) = 0 (2.293)
where y = (zT , w T )T ,
Ā = (−A, 0) , B̄ = [In , (Ḃ − AB)]

⎛ ⎞ ⎛ ⎞
0 0 0 0
N =⎝ ⎠ , U =⎝ ⎠
0 − ∂u
∂
Ḧu (B T Q − SA − Ṡ) 0
and
⎛ ⎞
Q 0
M=⎝ ⎠
0 0
The classical Legendre–Clebsch condition of a non-negative Hessian applied to this

accessory problem yields the following necessary condition to be satisfied along an
optimal trajectory:
r T Nr ≥ 0 (2.294)
for all r ∈ R(n+m)×1 satisfying
B̄r = 0 (2.295)
This results in the following necessary condition:
∂
− Ḧu ≥ 0 for all t ∈ [t0 , tf ] (2.296)
∂u
If this condition is satisfied with an equality for all t ∈ [t0 , tf ], then the problem
is totally singular in v, and therefore another transformation must be made to
determine optimality. This process is repeated until a nontrivial condition is
satisfied.
2.11.2 Jacobson’s Necessary Condition
Another necessary condition for the optimality of singular problems is given by the
following theorem [42]:
Theorem 2.13 A necessary condition for the second variation, J2 , to be non-

negative relative to u for the singular optimal control problem posed by
Eqs. (2.273)–(2.275) is
SB + B T W B ≥ 0 (2.297)
for all t ∈ [t0 , tf ], where
− Ẇ = Q + AT W + W A , W (tf ) = Qf (2.298)
Proof Adjoin the dynamic state constraint Eq. (2.274) to the performance index
given by Eq. (2.273) for the singular case (R = 0) via the costate vector, λ(t) =
W T (t)x(t), with W (t) ∈ Rn×n being an arbitrary, symmetric matrix which is
continuously differentiable with time. This results in the following augmented
performance index (i.e., the second variation on the performance index of the
accessory problem):
J2 = x(tf )T Qf x(tf )
tf
+ x T Qx + 2uT Sx + x T W (Ax + Bu − ẋ) dt (2.299)
t0
On integrating by parts the term involving ẋ, we have

tf
J2 = x(tf )T Qf x(tf ) + x T (Ẇ + Q + AT W + W A)x (2.300)
t0
+ 2uT (S + B T W )x dt + x(tf )T [Qf − W (tf )]x(tf )
The substitution of Eq. (2.298) into Eq. (2.300) yields

tf
J2 = 2 uT (S + B T W )xdt (2.301)
t0
Now, consider an arbitrary control variation, u, which vanishes over the entire
control interval, t ∈ [t0 , tf ], except in a narrow interval, t ∈ [t1 , t1 + ], with
t1 ∈ [t0 , tf ], where it has a constant and arbitrary value, k ∈ Rm×1 . Thus, the
second variation becomes the following:
t1 +
J2 = 2 k T (S + B T W )xdt (2.302)
t1
whose dominant term in the limit, → 0, is

k T (S + B T W )Bk 2 (2.303)
t=t1
2.12 Numerical Solution Procedures 69
The necessary condition for this term to be non-negative for any arbitrary t1 ∈
[t0 , tf ] is therefore SB + B T W B ≥ 0, which completes the proof.
Example 2.14 Let us revisit Example 2.12 in the light of Jacobson’s condition.
Here, we have

−2 0
Q = Hxx = , S = Hux = (0, 0) , B = fu = (0, 1)T
0 2
from which it follows that

w11 w12
S + B T W = (0, 0) + (0, 1) = (w12 , w22 )
w12 w22
SB + B T W B = (w12 , w22 )(0, 1)T = w22
and Eq. (2.298) produces

ẇ11 ẇ12 −2 w11 w11 (π ) w12 (π ) 0 0
− = , =
ẇ12 ẇ22 w11 2 + 2w12 w12 (π ) w22 (π ) 0 0
whose solution is easily derived to be the following:
2
w11 = −2τ , w12 = −τ 2 , w22 = 2τ − τ 3
3
√
where τ = π − t is the time-to-go. Since w22 < 0 in the interval τ > 3,
Jacobson’s necessary condition
√ given by Theorem 2.13 fails to be satisfied for a
part of the control interval, 3 < τ ≤ π . Thus, the singular solution is seen to
be non-optimal. We note that while the generalized Legendre–Clebsch necessary
condition was seen to be satisfied in Example 2.12, it did not guarantee the solution
to be optimal because it was not a sufficient condition. Jacobson’s condition is the
additional test which reveals the same solution to be actually a non-optimal one.
2.12 Numerical Solution Procedures
As has been discussed in the previous sections of this chapter, any optimal
control formulation without specific interior point constraints results in a two-
point boundary value problem (2PBVP), which can be expressed as the following
set of first-order, ordinary differential equations comprising the state and costate
equations:
ẏ = F (y, t) , (2.304)
where

x(t)
y(t) = ∈ R2n×1
λ(t)
and F (.) : R(2n+1)×1 → R2n×1 is a mapping vector functional. The equality

boundary conditions applied at the ends of the control interval, t ∈ [t0 , tf ], are
generally expressed as follows:
g0 [y(t0 )] = 0 ; gf [y(tf )] = 0 . (2.305)
As discussed previously, the initial conditions are specified only for the state
vector, x(t), whereas the terminal boundary conditions are to be satisfied by both
the state and the costate (λ(t)) vectors. A solution procedure for Eq. (2.304) is
thus necessary, which can march both forward and backward in time until all the
boundary conditions, Eq. (2.305), are satisfied. If the boundary conditions were to
be only specified at one end, the resulting problem is called an initial value problem
(IVP) whose integration can be carried out by a variable-step time-marching scheme
such as the Runge–Kutta methods [77] based upon a Taylor series approximation of
Eq. (2.304). However, a 2PBVP is inherently more difficult to solve than an IVP
since boundary conditions must be satisfied at both the ends.
It is not necessary that there should exist a solution to a particular 2PBVP,
or that the solution be unique. The existence and uniqueness of the solution to
a 2PBVP depend upon the nature of the functional, F (y, t), at the interior and
boundary points. The matters become complicated if F (y, t) has either singularities,
or discontinuous partial derivatives. Difficulties of computing a solution to a 2PBVP
are further compounded by errors in the numerical approximation employed for the
purpose. In some cases, there may be spurious numerical solutions when none really
exist, and in others a particular numerical scheme may be unable to find a solution
when one does exist.
Even when a unique solution does exist to a 2PBVP, iterative numerical schemes
are often required and are divided into two broad categories: (a) shooting method,
and (b) collocation method.
2.12.1 Shooting Method
A shooting method is based upon posing the boundary value problem as if it were
an initial value problem (IVP), and then changing the initial conditions by trial
and error until the boundary conditions are satisfied with a given accuracy. Thus,
a shooting method is based upon a repeated integration of the governing differential
equations for various initial conditions, using a standard numerical scheme such
as the Runge–Kutta method [77]. Simple shooting methods solve the IVP over the
entire control interval, t0 ≤ t ≤ tf , and can face convergence problems even if the
governing system of equations is stable and well-conditioned. In order to remove
such difficulties, multiple shooting methods have been developed [77] that partition
the control interval into several grid points, and shooting is carried out in several
stages.
2.12.2 Collocation Method
In contrast with the shooting methods, a collocation method approximates the

solution by a linear combination of a number of piecewise continuous polynomials
that are usually defined on a mesh of collocation points [59, 72]. The approximate
solution is then substituted into the system of ODEs such that the system is exactly
satisfied at each collocation point. The number of collocation points plus the
number of boundary conditions must equal the number of unknown coefficients in
the approximate solution. The most common choice of approximation is a linear
combination of spline functions. The collocation solution strategy is thus based
upon dividing the control interval into a number of sub-intervals, and approximating
the solution in each sub-interval by an appropriate spline interpolated at the mid-
point. In order to achieve a given accuracy, the collocation points must be carefully
selected.
Commercially available software such as Maple and MATLAB contain inbuilt
2PBVP solvers based upon either the shooting or the collocation scheme. For a
further description of the numerical solution procedures, a reader is referred to a
textbook devoted to such methods [59, 72, 77].
Exercises
2.1 A scalar system with the state equation, ẋ = x + u, is to be moved from the
initial state, x(0) = 1, in exactly five units of time, while minimizing the following
performance index:
5
1 1
J = [x(5)]2 + u2 (t)dt
2 2 0
Determine the extremal trajectory and the corresponding control history and find
out whether they are optimal.
2.2 A block sliding in a straight, frictionless groove on a horizontal table has its
equation of motion given as follows:
ÿ = u
where y(t) is the displacement of the block measured from one end of the table,
and u(t) is the applied acceleration input. The system is to be controlled such that
beginning from rest at t = 0, the block should come to rest at a final displacement
of 10 m after 1 s, and the total control effort, given by:
1
1
J = u2 (t)dt
2 0
is a minimum.
Solve for the extremal trajectory and the corresponding control history and
determine whether the extremal control history is optimal.
2.3 Repeat Exercise 2.2 if the final displacement and velocity after 1 s are not
specified, and the performance index is changed to the following:
1
1
J = k12 [y(1) − 10]2 + k22 [ẏ(1)]2 + u2 (t)dt
2 0
where k1 ∈ R, k2 ∈ R. Determine the value of the cost coefficients, k12 , k22 , such that
the final displacement and the velocity of the block exactly after 1 s are within ±1%
of 10 m and 0, respectively. Plot the optimal trajectory and control history for this
set of the cost coefficients.
2.4 Repeat Exercise 2.2 if the terminal time tf is not specified, and the performance
index to be minimized while moving the block from rest to rest by 10 m is the
following:
tf
1
J = 900tf2 + u2 (t)dt
2 0
2.5 Does the optimal control problem for the system in Example 2.6 have any
solution if the terminal time, tf , is unspecified, and the system is to be moved from
a zero initial state, x1 (0) = x2 (0) = 0 to x1 (tf ) = 1, x2 (tf ) = 0, while minimizing
tf
1
J = u2 (t)dt ?
2 0
Explain your answer.

2.6 Consider a system with the following state equations:
ẋ1 = x2 − x1
ẋ2 = −2x1 − 3x2 + u
Determine the extremal trajectory and control history for a fixed terminal time,
tf = 1, if the system is to be taken from an initial state of x(0) = (0, 0)T , to a
free final state such that the following performance index is minimized:
1
1 1 1
J = [x1 (tf ) − 1]2 + [x2 (tf ) − 1]2 + u2 (t)dt
2 2 2 0
2.7 Solve the lunar landing problem of Example 2.9 with the revised initial
conditions, z(0) = 100 m, ż(0) = −10 m/s, and for tf = 30 s. What is the
value of the vertical acceleration input at the final time and why?
2.8 For the lunar landing problem of Example 2.9, revise the objective function to
the following:

1 1 1 tf 2
J = k 2 tf2 + [x2 (tf ) − 1]2 + u (t)dt
2 2 2 0
and find the optimal trajectory and control history for a suitable value of the cost
parameter k.
2.9 For a system with the following state equations:
ẋ1 = x2
ẋ2 = −x2 + u
Determine the extremal trajectory and control history for a fixed terminal time, tf =
2, if x(0) = (0, 0)T , x(2) is not specified, and the following performance index is
minimized:
2
1 1 1
J = [x1 (2) − 5]2 + [x2 (2) − 2]2 + u2 (t)dt
2 2 2 0
2.10 Repeat Exercise 2.9 for x(0) = (0, 0)T , x(2) = (5, 2)T , and
2
1
J = u2 (t)dt
2 0
2.11 Repeat Exercise 2.9 if the system is to be transferred from x(0) = (0, 0)T to
the following straight line at tf = 2:
x1 (t) + 5x2 (t) = 15
while minimizing
2
1
J = u2 (t)dt
2 0
2.12 Consider a particle moving in a plane acted upon by a known acceleration,

a(t), which is produced by a force applied at an angle, β(t), measured from
the direction of motion. The equations of motion with (x, y) as the displacement
coordinates, and (u, v) as the velocity components are the following:
ẋ = u
ẏ = v
u̇ = a cos β
v̇ = a sin β
It is required to move the particle in a given time T such that y(T ) = h and v(T ) =
0. Considering β(t) to be the control input, formulate the two-point boundary value
problem to be solved for the extremal trajectory.
2.13 Consider a bead sliding on a frictionless wire stretched between two points
with the following state equations:
ẋ = v(y) cos θ
ẏ = v(y) sin θ
where x, y are the horizontal and vertical coordinates of the bead, θ (t) is the local
angle made
, by the wire with the horizontal, the speed of the bead is given by
v(y) = v02 + 2gy, with initial speed v0 , and g being the constant acceleration
due to gravity.
State the necessary conditions for the minimum-time path taken by the bead
between two given points, if θ (t) is the control input. (This problem is called the
brachistochrone problem.)
2.14 The state equations of a ship with the displacement coordinates (x, y), moving
at a constant speed V with heading angle, β(t), between strong ocean currents of
known velocity field, [u(x, y), v(x, y)] are the following:
ẋ = V cos β + u(x, y)
ẏ = V sin β + v(x, y)
It is required to steer the ship in the minimum time between two specified positions,
(xi , yi ), and (xf , yf ). (This is called Zermelo’s navigation problem.)
(a) Considering β(t) to be the control input, formulate the two-point boundary
value problem to be solved for the extremal trajectory.
(b) From (a), derive the differential equation to be solved for β(t).
(c) What is the solution to (b), if the ocean currents have constant velocity
components, (u, v)?
2.15 Suppose the scalar system with the state equation, ẋ = u, has its input
bounded by:
|u(t)| ≤ 1
Derive the switching condition on u(t) for the time-optimal transfer from x(0) =
1 to x(tf ) = 0 and solve for the resulting optimal trajectory.
2.16 Does the minimum-time problem for the block sliding in a frictionless, hori-
zontal, straight groove (Exercise 2.2) have any solution if the input is unconstrained?
Explain your answer.
2.17 State the necessary conditions for the minimum-time transfer of the particle
in Exercise 2.11.
2.18 Reformulate the lunar landing problem of Example 2.9 as a time-optimal
control problem with the vertical acceleration input bounded by | u |≤ a, and the
initial conditions, z(0) = 100 m, ż(0) = −10 m/s.
(a) Can the upper bound, a, be smaller than g for a smooth landing (ż(tf ) = 0)?
(b) For a = 1.5g, solve the minimum-time problem and determine the trajectory
and control history.
Chapter 3
Orbital Mechanics and Impulsive
Transfer
3.1 Introduction
Orbital mechanics refers to the study of the translational motion of bodies in mutual
gravitational attraction, and the resulting spatial paths of the centers of mass of the
respective bodies are called the orbits. The relative motion of two spherical bodies
in mutual gravitational attraction is the fundamental problem of translational space
dynamics, called the two-body problem (or the Keplerian motion), and possesses
an analytical solution. The motion of a spacecraft under the influence of a celestial
body is usually approximated as a two-body problem by neglecting the gravitational
effects caused by the other bodies, the actual, aspherical shapes of the two bodies,
and non-gravitational disturbances due to the atmosphere, continuous thrust, and
solar radiation pressure. The gravity field in such a case is exactly spherical,
governed by the inverse-squared law of Newton. However, when a spacecraft
comes under the influence of two (or more) bodies whose gravitational pull is of
comparable magnitudes in certain regions of the flight, then the approximation of
Keplerian motion becomes invalid, and an n-body gravity model with n > 2 is
required. Lunar and interplanetary flights are the examples of such problems. For
a flight within the solar system, it is rarely necessary to take into account more
than three bodies in mutual gravitational attraction. Thus the three-body problem
(n = 3) is the limiting case of analytical space flight navigation problems. For
n > 3, a numerical solution is invariably required, hence a theoretical orbital model
is unavailable. Furthermore, the mass of the spacecraft is negligible in comparison
with those of the bodies attracting it, resulting in what is referred to as a restricted
two- (or three-) body problem.
The present chapter introduces the reader to the two-body orbits and their
impulsive changes (called impulsive orbital transfers). The following chapters will
cover the more general cases of continuous transfers and non-spherical gravity
fields. For a greater detail, a textbook on orbital mechanics can be consulted, such
as Chobotov [23] or Battin [8].

https://doi.org/10.1007/978-3-030-03789-5_3
78 3 Orbital Mechanics and Impulsive Transfer
3.2 Keplerian Motion
Consider spherical masses m1 , m2 in mutual attraction given by Newton’s law of

gravitation. The equation of motion of each mass is given by Newton’s second law
of motion referred to an inertial frame. Subtracting the two equations of motion
from one another, we have
μ
r̈ + r=0 (3.1)
r3
where r(t) is the position of center of mass m2 relative to that of mass m1 , μ =
G(m1 + m2 ), with G = 6.6726 × 10−11 m3 /kg/s2 is the universal gravitational
constant, and overdot represents the time derivative, d/dt. Since the spacecraft’s
mass, m2 , is negligible in comparison with that of the celestial body, m1 , we can
approximate μ ≈ Gm1 . Take the right vector product of both the sides of Eq. (3.1)
with r to show that the specific angular momentum of m2 relative to m1 , defined by
h = r × ṙ = r × v (3.2)
is conserved in Keplerian motion, where v = ṙ is the orbital velocity. Since h is a

constant vector, it follows that
(a) The direction of h is a constant, implying that the vectors r and v are always in
the same plane—called orbital plane—and h is normal to that plane.
(b) The magnitude of h is constant. Writing h in polar coordinates, (r, θ ),
h =| h |=| r × v |= r 2 θ̇ (3.3)
which implies that the radius vector, r, sweeps out area at a constant rate,
thereby confirming Kepler’s second law of planetary motion.
The trajectories in the orbital plane—called orbits—are classified according to
the magnitude and direction of a constant h. The case of h = 0 represents rectilinear
motion along the line joining the two bodies, while h = 0 represents the more
common orbits involving the mass m2 describing a curve about m1 .
A scalar product of Eq. (3.1) with v = ṙ reveals that the specific energy of the
two-body relative motion is also constant, given by
v2 μ
− = (3.4)
2 r
The right vector product of both sides of Eq. (3.1) with h leads to another constant
vector, e, called the eccentricity vector, and defined by
μr
μe = v × h − (3.5)
r
3.2 Keplerian Motion 79
Since the eccentricity vector is normal to h, it always lies in the orbital plane, and
can be used as a reference to specify the direction of the relative position vector,
r(t). The angle, θ (t), made by r(t) with e (measured along the flight direction) is
called the true anomaly, θ (t).
Taking the scalar product of μe with r(t) produces the following orbit equation:
p
r(θ ) = (3.6)
1 + e cos θ
where p = h2 /μ is the parameter of the orbit and e is termed eccentricity of the

orbit. The orbit equation defines the shape of the orbit in polar coordinates, (r, θ ),
and indicates that the orbit is symmetrical about e. Furthermore, it shows that the
orbit is a conic section, with the minimum separation of the two bodies (called
periapsis) occurring for θ = 0. Hence, e points toward the periapsis. For e < 1, the
orbit is an ellipse, whose special case is the circular orbit (e = 0). For e = 1, the
orbit is a parabola, whereas for e > 1, it is a hyperbola. A time derivative of the
orbit equation produces the radial velocity,
-
eh μ
vr = ṙ = sin θ = sin θ (3.7)
p p
while the circumferential velocity is derived from Eqs. (3.3) and (3.6) to be the
following:
-
h μ
vθ = r θ̇ = = (1 + e cos θ ) (3.8)
r p
The ratio of the two velocity components gives the flight-path angle, φ, as follows:
vr ṙ sin θ
tan φ = = = (3.9)
vθ r θ̇ 1 + e cos θ
while their contribution to the velocity magnitude (orbital speed), v, is given by

, -
μ
v= vr2 + vθ2 = (1 + 2e cos θ + e2 ) (3.10)
p
A substitution of Eq. (3.10) into Eq. (3.4) yields
v2 μ μ(1 − e2 ) μ
= − =− =− (3.11)
2 r 2p 2a
where a = p/(1 − e2 ) is defined to be the semi-major axis of the orbit. For an

elliptical orbit, a > 0 has a geometric significance, whereas a = ∞ for a parabolic
orbit, and a < 0 for a hyperbolic orbit. In all cases, the focus of the orbit is at
the center of the celestial body, and the smallest distance between the two bodies
(periapsis radius) is rp = p/(1 + e).
3.2.1 Reference Frames of Keplerian Motion
The velocity vector can be resolved in the local-horizon frame, (ir , iθ , ih ), with ir =
r/r, and iθ being normal to both ir and ih , such that ih = ir × iθ :
-
μ
v = vr ir + vθ iθ = [sin θ ir + (1 + e cos θ )iθ ] (3.12)
p
The local-horizon frame—being attached to the spacecraft—is a moving frame, and

the plane (iθ , ih ), which is normal to the radius vector, r, is called the local horizon.
A right-handed coordinate frame fixed to the orbital plane, with unit vectors ih =
h/ h, ie = e/e, and ip = ih × ie , is used to specify the relative position and velocity
vectors of the spacecraft in the orbital plane. This frame, (ie , ip , ih ), is called the
perifocal frame. The position and velocity of the spacecraft in the perifocal frame
are the following:
r = r cos θ ie + r sin θ ip
v = v(sin φ cos θ − cos φ sin θ )ie + v(sin φ sin θ + cos φ cos θ )ip (3.13)
where φ is the flight-path angle (defined earlier) measured above the local horizon.
On eliminating the flight path angle from the velocity expression, we have
μ μ
v=− sin θ ie + (e + cos θ )ip (3.14)
h h
The spacecraft’s position and velocity can also be resolved in an inertial celestial
frame, (I, J, K), which is fixed with respect to distant stars, with its origin, o, at the
center of the celestial body (Fig. 3.1). The position vector, r, can be expressed in
terms of spherical coordinates, (r, δ, λ), where r is the radius, δ is the declination
(or latitude) measured above the (I, J) plane, and λ the right ascension (or celestial
longitude) measured from the I axis to the projection of r in the (I, J) plane
along the right-hand rotation about K (Fig. 3.1). Another orientation of the local-
horizon frame is useful in representing the velocity vector in terms of the spherical
coordinates. It consists of the unit vectors, (ir , iλ , iδ ), where iλ points toward the
direction of increasing λ, and iδ toward that of increasing δ, such that ir × iλ = iδ
(Fig. 3.1). The velocity vector, v, is then resolved in spherical coordinates, (v, φ, ψ),
relative to the local-horizon frame, where v is the speed, φ the flight-path angle made
by v with the local-horizon plane, (iλ , iδ ), measured positive toward ir , and ψ the
azimuth angle between the projection of v on the local-horizon plane, (iλ , iδ ), and
the axis iδ , measured positive toward iλ , as shown in Fig. 3.1.
Fig. 3.1 Position and v

velocity vectors resolved in il
the celestial, (I, J, K), and id f
local horizon, (ir , iλ , iδ ), ir
frames, respectively K y
d
o
l J
Local
Horizon
K v
f
ih
ip
i
r
ie Periapsis
q
o w J
I W
Vernal Equinox Orbit
n
Ascending Node
Fig. 3.2 Orientation of the perifocal frame relative to celestial plane given by Euler angles, Ω, i, ω
Examples of Earth-based celestial frames are with the (I, J) plane being either
the equatorial plane or the ecliptic plane, with the axis I pointing toward the vernal
equinox. This is the point of intersection between the equatorial and ecliptic planes,
where the Sun appears to cross the equator from the south to the north at mid-day
(which usually happens around March 21). The intersection of the orbital plane
with the reference plane, (I, J), yields the line of nodes, as shown in Fig. 3.2. The
ascending node is the point on the line of nodes where the orbit crosses the plane
(I, J) from the south to the north. A unit vector, n, pointing toward the ascending
node makes an angle Ω with the axis I. The angle Ω is measured in the plane
(I, J) in an anti-clockwise direction (right-handed rotation about K) (Fig. 3.2), and
is termed the right ascension of the ascending node. The orbital inclination, i, is
the angle between the orbital plane and the (I, J) plane, and is the positive (right-
handed) rotation about n required to produce ih from the axis K. The angle ω
represents a positive rotation of n about ih to produce ie in the orbital plane, and
is called the argument of periapsis.
A coordinate transformation between the perifocal and celestial frames is
necessary. The orientation of the perifocal frame relative to the celestial frame is
classically described by the Ω, i, ω Euler angles, sequentially applied in the right-
hand sense about the third, first, and third axes, respectively, of the successively
rotated reference frame, as shown in Fig. 3.2. This results in the following rotation
matrix [80]:
C = C3 (ω)C1 (i)C3 (Ω) (3.15)
where
⎛ ⎞ ⎛ ⎞
1 0 0 cos(.) − sin(.) 0
C1 (.) = ⎝ 0 cos(.) − sin(.) ⎠ , C3 (.) = ⎝ sin(.) cos(.) 0 ⎠ (3.16)
0 sin(.) cos(.) 0 0 1
Hence the transformation between the perifocal and the celestial frames is given as
follows:
⎧ ⎫ ⎧ ⎫
⎨I⎬ ⎨ ie ⎬
J = C T
i (3.17)
⎩ ⎭ ⎩ p⎭
K ih
Similarly, a coordinate transformation between the local-horizon and celestial

frames is described by the Euler angles, λ, −δ, sequentially applied in the right-
hand sense about the third and the second axes, respectively, of the successively
rotated reference frame, as shown in Fig. 3.1:
⎧ ⎫ ⎧ ⎫
⎨I⎬ ⎨ ir ⎬
J = CLH T iλ (3.18)
⎩ ⎭ ⎩ ⎭
K iδ
where
CLH = C2 (−δ)C3 (λ) (3.19)
where
⎛ ⎞
cos(.) 0 sin(.)
C2 (.) = ⎝ 0 1 0 ⎠ (3.20)
− sin(.) 0 cos(.)
3.2.2 Time Equation
While the vectors h and e completely determine the shape and orientation of a two-
body orbit, they do not provide any information about the location of the spacecraft
at a given time along the orbit. This missing information is expressed as a differential
equation which relates the variation of true anomaly with time, θ (t), beginning
with the time of periapsis, τ , for which θ = 0. On substituting the orbit equation,
Eq. (3.6), into Eq. (3.8), we have
-
μ
θ̇ = (1 + e cos θ )2 (3.21)
p3
The integration of Eq. (3.21) provides τ , thereby determining the function θ (t), and
completing the solution to the two-body problem. However, such an integration is
usually carried out by a numerical procedure, depending upon whether the orbit is
elliptical, parabolic, or hyperbolic.
Elliptical Orbit (0 ≤ e < 1)
For an elliptical orbit, Eq. (3.21) becomes the following:

3
(1 − e2 ) 2 dθ
= ndt (3.22)
(1 + e cos θ )2
where
-
μ
n= (3.23)
a3
is the orbital mean motion. Now introduce an eccentric anomaly, E defined by
e + cos θ
cos E = (3.24)
1 + e cos θ
which substituted into Eq. (3.22), yields the following Kepler’s equation:
E − e sin E = M (3.25)
where M = n(t − τ ) is called the mean anomaly. The unambiguous relationship

between the true and eccentric anomalies is given by
-
θ 1+e E
tan = tan (3.26)
2 1−e 2
Note that E2 and θ2 are always in the same quadrant. Solution to Kepler’s equation
is simply obtained by Newton’s method. Consider a first-order, finite-difference
approximation of the function, f (E) = E − e sin E − M, as follows:
f (E + ΔE) f (E) + f
(E)ΔE (3.27)
where f
= df (E)/dE = 1 − e cos E and ΔE is the step size. Newton’s method
is then based upon starting with an initial guess of the solution (usually E = M +
e sin M), updating it to E + ΔE where ΔE = −f (E)/f
(E) until f (E + ΔE) < δ,
a given tolerance.
The position and velocity vectors in an elliptical orbit can be expressed directly
in terms of the eccentric anomaly as follows:
√
r = a(cos E − e)ie + ap sin Eip
√ √
μa μp
v=− sin Eie + cos Eip (3.28)
r r
Example 3.1 A low-Earth satellite is to be de-orbited from a circular orbit of period
90 min. by firing a retro-rocket which applies tangential velocity impulse of Δv =
−0.5iθ km/s. Calculate the radius, the speed, and the flight-path angle 5 min. after
firing the rocket. (For Earth: μ = 398,600.4 km3 /s2 , mean radius, R0 = 6378.14
km.)
We begin by calculating the spacecraft’s orbit following the velocity impulse.
The initial orbital radius is calculated from the time period of the orbit as follows:
2π
n= = 0.0013884 rad/s r1 = (n2 μ)1/3 = 6652.6 km
60 × 90
√
The initial speed in the circular orbit is therefore v1
= μ/r1 = 7.7406, from
which the horizontal impulse is subtracted to yield the initial speed of the new orbit,
v1 = v1
− 0.5 = 7.2406 km/s. The semi-major axis of the new orbit is calculated
using Eq. (3.11) as follows:
μ
a= = 5913.3 km
2μ/r1 − v12
Since the semi-major axis of this elliptical orbit is smaller than Earth’s radius, a <
R0 , the satellite will finally impact Earth. Furthermore, since a < r1 , the initial
tangential impulse point is the apogee (the point of the largest radius) of the new
orbit. Therefore, we put r1 = r(π ) in the orbit equation, Eq. (3.6), and obtain r1 =
a(1 + e), from which the eccentricity is derived to be e = r1 /a − 1 = 0.125.
(The same result can be obtained by calculating the angular momentum, h, of the
new √ orbit, h = r1 v1 = 48,168.49 km2 s, giving p = h2 /μ = 5820.9 km, and
e = 1 − p/a = 0.125.)
Next, Kepler’s equation is solved by taking the initial anomalies for the apogee
point, θ1 = E1 = M1 = π , and determining the final eccentric anomaly, E2 , after
5 min. The mean anomaly at the final position is M2 = π + n × 5 × 60 = 3.5581221
rad., from which the initial guess of E (0) = M2 is made. The steps of Newton’s
iteration are then the following:
(i) f (E (0) ) = E (0) − e sin E (0) − M2 = 0.05058, f
(E (0) ) = 1 − e cos E (0) =
1.1143, ΔE = −f (E (0) )/f
(E (0) ) = −0.045391 rad.
(ii) E (1) = E (0) + ΔE = 3.5127312 rad., f (E (1) ) = E (1) − e sin E (1) −
M2 = −5.03 × 10−5 , f
(E (1) ) = 1 − e cos E (1) = 1.1165, ΔE =
−f (E (1) )/f
(E (1) ) = 4.5 × 10−5 rad.
(iii) E (2) = E (1) + ΔE = 3.5127763 rad., f (E (2) ) = E (2) − e sin E (2) − M2 =
−4.6 × 10−11 rad.
Hence, the solution has converged to within a tolerance of 10−10 rad. in two
iterations, resulting in E2 = E (2) = 3.5127763 rad. The radius at this point is
r2 = a(1 − e cos E2 ) = 6602.21 km, the speed,

2μ μ
v2 = − = 7.3034 km/s
r2 a
and the flight-path angle,

h
φ2 = cos−1 = −2.617◦
r2 v2
where the correct quadrant (π < θ < 2π ) is employed.
Hyperbolic Orbit (e > 1)
The equivalent form of Kepler’s equation can be obtained for hyperbolic orbits by
introducing the hyperbolic anomaly, H. The parametric equation of a hyperbola
centered at x = 0, y = 0, with semi-major axis a, can be written as, x = a cosh H ,
or r cos θ = a(cosh H − e). Hence the relationship between the hyperbolic anomaly
and the true anomaly is given by
cosh H − e
cos θ = (3.29)
1 − e cosh H
whose substitution into Eq. (3.21) and integration from H = 0, t = τ yields
e sinh H − H = n(t − τ ) (3.30)
with the hyperbolic mean motion, n given by

-
μ
n= − 3 (3.31)
a
Equation (3.30) can be solved numerically using Newton’s method in a manner

similar to Kepler’s equation. However, since a hyperbolic orbit is an open trajectory,
the hyperbolic anomaly, H , keeps on increasing with time, and hence a larger
number of iterations are required for an increased difference between the initial
and final times.
Parabolic Escape Trajectory (e = 1)
A parabolic orbit is the minimum energy escape trajectory ( = 0), thus it has zero
speed at r → ∞. It is often a valuable mathematical aid for quickly determining
the minimum fuel requirements for a given mission, as well as for beginning the
iterations of Kepler’s equation in the case e 1 or e > 1. Equation (3.21) for a
parabolic orbit is integrated to yield the following Barker’s equation:
-
θ θ μ
tan3 + 3 tan = 6 (t − τ ) (3.32)
2 2 p3
whose real, unique, and closed-form solution is obtained by substituting
θ 1
tan =α− (3.33)
2 α
and solving the resulting quadratic equation for α 3 , resulting in
θ . 1 . 1
tan = (C + 1 + C 2 ) 3 − (C + 1 + C 2 )− 3 (3.34)
2
where
-
. μ
C=3 (t − τ ) (3.35)
p3
is an equivalent mean anomaly.
3.2.3 Lagrange’s Coefficients
A two-body trajectory expressed in the perifocal frame (ie , ip , and ih ) is the

following (Eq. (3.13)):
r = r cos θ ie + r sin θ ip (3.36)
μ μ
v=− sin θ ie + (e + cos θ )ip (3.37)
h h
Given the position and velocity at a time, t0 , one would like to determine the
position and velocity at some other time, t. In order to do so, the known position and
velocity are expressed as follows:
& / ! "& /
r0 r0 cos θ0 r0 sin θ0 ie
= (3.38)
v0 − μh sin θ0 μ
h (e + cos θ0 ) ip
Since the square matrix in Eq. (3.38) is non-singular (its determinant is equal to h),
it can be inverted to obtain ie and ip as follows:
& / ⎡ ⎤& /
ie 1
p (e + cos θ0 ) − rh0 sin θ0 r0
=⎣ r0
⎦ (3.39)
ip 1
p sin θ0 h (e + cos θ0 ) v0
On substituting Eq. (3.39) into Eqs. (3.36) and (3.37), we have

& / 0 1& /
r f g r0
= (3.40)
v f˙ ġ v0
where
r
f = 1+ [cos(θ − θ0 ) − 1]
p
rr0
g= sin(θ − θ0 )
h
df h
f˙ = = − 2 [sin(θ − θ0 ) + e(sin θ − sin θ0 )]
dt p
dg r0
ġ = = 1 + [cos(θ − θ0 ) − 1] (3.41)
dt p
The functions f and g, called Lagrange’s coefficients, are very useful in deter-
mining a two-body orbit from a known position and velocity. The matrix,
0 1
f g
(t, t0 ) = (3.42)
f˙ ġ
called state transition matrix, has a special significance, because it uniquely

determines the current state, (r, v), from the initial state, (r0 , v0 ). Such a relationship
between the initial and final states is rarely possible in the solution to a nonlinear
differential equation, and is thus a valuable property of the two-body problem. The
state transition matrix is seen to have the following properties:
1. From the conservation of angular momentum,
h = r × v = (f ġ − g f˙)r0 × v0 = r0 × v0 (3.43)
it follows that
| |= f ġ − g f˙ = 1 (3.44)
A consequence of the unity determinant is that the inverse of the state transition
matrix is given by
0 1
−1 . ġ −g
= (3.45)
−f˙ f
Such a matrix is said to be symplectic. (It can be shown [60] that the state
transition matrix of any conservative Hamiltonian system is symplectic.)
2. Given any three points (t0 , t1 , t2 ) along the trajectory, it is true that
(t2 , t0 ) = (t2 , t1 )(t1 , t0 ) (3.46)
3.3 Impulsive Orbital Transfer
A spacecraft’s nominal trajectory relative to the central body is generally the solu-
tion to the two-body problem, unless significant orbital perturbations are present.
Even if orbital perturbations are present, the two-body orbit can be taken to be a
nominal flight path, which should be either followed despite small perturbations by
the application of control inputs or used as an initial guess for obtaining the closed-
loop solution to the perturbed problem by iterative techniques [80]. Therefore, the
problem of guiding the spacecraft from an initial to a desired final position in space
can be generally treated as the problem of determining an optimal trajectory that
passes through the two given positions.
Consider the impulsive orbital transfer from an initial position vector, r1 (t1 ), to a
final position, r2 (t2 ) (see Fig. 3.3). The transfer angle, Δθ = θ2 −θ1 , in Fig. 3.3 is the
difference of the respective true anomalies, (θ1 , θ2 ), along the transfer orbit. If the
time of flight, tf = t2 −t1 , is known a priori, the problem has a unique solution, and is
referred to as Lambert’s problem. However—as is commonly the case—if we have
a choice of varying the time of flight according to the requirement of minimizing an
objective function (such as the time of flight, or the propellant mass) for the transfer,
one arrives at the problem of optimal impulsive transfer that must be solved by
optimal control techniques (Chap. 2). In either case, one has to solve a two-point,
boundary value problem yielding the required orbit, r(t), taking the spacecraft from
r1 to r2 , which can be used as a nominal flight path to be followed by a closed-loop
guidance system despite perturbations.
3.3 Impulsive Orbital Transfer 89
Fig. 3.3 Geometry of Transfer Orbit

impulsive transfer between
two given positions, r1 , r2
c
r1
r2
Dq
b1 b2
Dv1 Dv2
v1¢ v2
a1-f1 a2-f2¢
v1 v2¢
Df1 Df2
f1 a1 f 2¢ a2
(a) (b)
Fig. 3.4 Geometry of the two impulse applications in Example 3.2
Example 3.2 Calculate the magnitudes and the directions of the two velocity
impulses required for transferring a spacecraft around Earth from r1 = 7000 km,
v1 = 7 km/s, φ1 = 25◦ to r2 = 7500 km, v2 = 8 km/s, φ2 = 30◦ via a transfer
ellipse of a = 7000 km and e = 0.5.
Let the first impulse of magnitude Δv1 be applied at an angle α1 from the local
horizon on the given orbit. Figure 3.4a shows the geometry of impulse application,
which yields the following relationships:
φ1
+ β1 − α1 = 0
v1
v1 Δv1
= =
sin(π + φ1 − α1 ) sin β1 sin Δφ1
,
Δv1 = v12 + (v1
)2 − 2v1 v1
cos Δφ1
Here, Δ indicates the change caused by a velocity impulse, and the prime denotes
the velocities at the two extremities of the transfer orbit, (v1
, φ1
) and (v2
, φ2
). Thus
we have φ1
= φ1 + Δφ1 and φ2
= φ2 − Δφ2 . The transfer orbit yields the following
at r = r1 : v1
= 7.54605 km/s, φ1
= 30◦ , implying Δφ1 = 5◦ , from which the
relationships derived from Fig. 3.4a produce the first velocity impulse as Δv1 =
0.83677 km/s and α1 = 76.81◦ .
Similarly, the second impulse produces the geometry shown in Fig. 3.4b, result-
ing in the following relationships:
φ2
+ β2 − α2 = 0
v2 v2
Δv2
= =
sin(π + φ2 − α2 ) sin β2 sin Δφ2
,
Δv2 = v22 + (v2
)2 − 2v2 v2
cos Δφ2
The transfer orbit yields the following at r = r2 : v2

= 7.025 km/s, φ2
= ±29.745◦ ,
implying either Δφ2 = −0.2551◦ or Δφ2 = −59.745◦ . The first option produces
Δv2 = 0.97558 km/s and α2 = 212.1◦ , while the second option (φ2
= −29.745◦ )
yields Δv2 = 7.5311 km/s and α2 = 276.57◦ . Clearly, the second option is
uneconomical in comparison with the first, because it requires a larger velocity
impulse magnitude. It can be easily calculated that the second option also requires
a larger transfer angle, θ2 − θ1 = 113.13◦ , compared to that of the first option,
θ2 − θ1 = 6.87◦ , which translates into a much larger flight time. Therefore, the first
option is selected, Δv2 = 0.97558 km/s and α2 = 212.1◦ .
For a two-body orbital transfer with an arbitrary transfer time, t2 − t1 , we express
the final position vector, r2 , in terms of the initial position, r1 , and initial velocity,
v1 , through Lagrange’s coefficients, f, g, as follows:
r2 = f r1 + gv1 , (3.47)
where
r2
f = 1+ (cos Δθ − 1)
p
r1 r2
g= sin Δθ
h
df h
f˙ = = − 2 [sin Δθ + e(sin θ2 − sin θ1 )]
dt p
dg r1
ġ = = 1 + (cos Δθ − 1) , (3.48)
dt p
and Δθ = θ2 − θ1 is the given transfer angle (Fig. 3.3). Furthermore, the initial
position vector in terms of the final position and velocity is given by
r1 = f
r2 + g
v2 , (3.49)
where
r1
f
= 1 + (cos Δθ − 1) = ġ
p
r1 r2
g
= − sin Δθ = −g . (3.50)
h
Thus, we have
r1 = ġr2 − gv2 . (3.51)
The initial and final velocity vectors can be expressed as components along the
chord, vc , and the local radius, vr . Note from Fig. 3.3 that the chord and local radius
are not necessarily perpendicular to each other (vr = ṙ). Thus we have
1
v1 = [(r2 − r1 ) + (1 − f )r1 ] = vc ic + vr ir1
g
1
v2 = [(r2 − r1 ) − (1 − ġ)r2 ] = vc ic − vr ir2 , (3.52)
g
where
r2 − r1
ic =
c
r1
ir1 =
r1
r2
ir2 = (3.53)
r2
c ch
vc = =
g r1 r2 sin Δθ
h(1 − cos Δθ )
vr = .
p sin Δθ
The last two equations of Eq. (3.53) imply that
r1 r2 (1 − cos Δθ ) vc
p= . (3.54)
c vr
Furthermore, we have
cμ sec2 Δθ
vc vr = 2
, (3.55)
2r1 r2
which is an important result showing that the product of the two velocity compo-
nents depends only upon the geometry of transfer specified by (r1 , r2 , Δθ ) (and not
on the shape or size of the transfer orbit). Thus, we can select various transfer orbits
to meet other requirements (such as the time or fuel for transfer).
3.3.1 Minimum Energy Transfer
A practical solution for the impulsive two-point, boundary value problem is the case
of the minimum energy transfer, which corresponds to the smallest value of orbital
energy = −μ/2a for the transfer orbit. The minimum energy orbit involves the
smallest positive value of the semi-major axis, a, of the orbit joining the two given
radii, (r1 , r2 ), and would normally result in the smallest propellant expenditure. Let
the minimum energy, elliptical orbit have the semi-major axis amin , as shown in
Fig. 3.5, and the respective distances of the initial and final positions from the vacant
focus, oV , be (r1∗ , r2∗ ). For an ellipse of major axis a, we have
r1 + r1∗ = 2a ; r2 + r2∗ = 2a (3.56)
or
r1 + r2 + r1∗ + r2∗ = 4a (3.57)
Minimum Energy Transfer Orbit
r1*
c
r1
r2* oV
Dq
r2 o
2amin
Fig. 3.5 Geometry of minimum energy orbital transfer for spacecraft guidance
Then it follows that the smallest possible value for the distance, r1∗ +r2∗ , is the chord,
c =| r2 − r1 |, thereby implying that the vacant focus, oV , must lie on the chord
line, as shown in Fig. 3.5. Thus we have
1
2amin = (r1 + r2 + c) (3.58)
2
The semi-perimeter, s, of the transfer triangle with sides r1 , c, r2 is given by
r1 + r2 + c
s= . (3.59)
2
Hence, the major axis of the minimum energy transfer ellipse equals the semi-
perimeter,
2amin = s . (3.60)
Note that the minimum energy transfer has the same orbit, am , if the focus, o, is
shifted to the vacant focus, oV , the transfer takes place from r1∗ to r2∗ (i.e., Δθ =
π and c = r1 + r2 ), which implies from Eq. (3.52) that s = c and vc /vr = 1.
Since the ratio vc /vr is unity for the minimum energy transfer, the following is the
corresponding parameter from Eq. (3.54):
r1 r2
pm = (1 − cos Δθ ) . (3.61)
c
The following method can be employed for determining the minimum energy
transfer orbit (and the corresponding time of flight) between two given positions,
(r1 , r2 ):
(a) Determine the chord, the initial and final radii, and the semi-major axis of the
transfer orbit as follows:
c =| r2 − r1 | ; r1 =| r1 | ; r2 =| r2 |
1
a= (r1 + r2 + c)
4
(b) Find the unit vectors along the chord and the two radii, and the transfer angle,
Δθ , as follows:
r2 − r1 r1 r2
ic = ; ir1 = ; ir2 =
c r1 r2
cos Δθ = ir1 · ir2 sin Δθ =| ir1 × ir2 |

(c) Calculate the parameter and the orbital eccentricity as follows:

r1 r2
p= (1 − cos Δθ )
c
-
p
e = 1−
a
(d) Referring to Fig. 3.5, determine the initial and final eccentric anomalies, the
time of periapsis, and the time of flight as follows (see Eqs. (3.6) and (3.24)):

−1 1 r1
E1 = cos −
e ae

1 r2
E2 = 2π − cos−1 −
e ae
-
μ E1 − e sin E1
n= ; τ = t1 −
a3 n
E2 − E1 − e(sin E2 − sin E1 )
t2 − t1 =
n
Example 3.3 Find the minimum energy transfer trajectory and the correspond-
ing time of flight for reaching a 500 km altitude above the Earth (μ =
398,600.4 km3 /s2 , surface radius, r0 = 6378.14 km), latitude 45◦ , and right
ascension 250◦ , beginning from an initial altitude, latitude, and right ascension of
200 km, −10◦ , and 30◦ , respectively.
The necessary computation is performed as follows:
r1 = r0 + 200 = 6578.14 km , r2 = r0 + 500 = 6878.14 km
r1 = 5610.289I + 3239.102J − 1142.282K km
r2 = −1663.442I − 4570.270J + 4863.579K km
c =| r2 − r1 |= 12,245.971 km
1
a= (r1 + r2 + c) = 6425.563 km
4
r2 − r1 r1 r2
ic = ; ir1 = ; ir2 =
c r1 r2
r1 · r2
cos Δθ = = −0.656233760652892
r1 r2
3.4 Lambert’s Transfer 95
.
p = r1 r2 (1 − cos Δθ )/c = 6119.311 km , e= 1 − p/a = 0.218315

−1 1 − r1 /a
E1 = cos = 1.679778 rad.
e

1 − r2 /a
E2 = 2π − cos−1 = 4.383888 rad.
e
,
n= μ/a 3 = 0.00122575 rad/s
E2 − E1 − e(sin E2 − sin E1 )
t2 − t1 = = 2551.719 s
n
Hence the required transfer orbit has elements, a = 6425.563 km, e = 0.218315,
and τ = −(E1 − e sin E1 )/n = −1193.357 s (measured from t1 = 0 at launch),
and the required time of flight is 2551.719 s.
Since the two foci must always lie on the chord for the minimum energy ellipse
(Fig. 3.5), the corresponding time of flight for even a small transfer angle, Δθ , can
be quite large. In fact, it can be shown that the minimum energy orbit yields the
maximum time of flight between a given pair of positions. In any case, a different
transfer strategy becomes necessary when the time of flight is fixed, such as in an
orbital rendezvous between two spacecraft, or interception of a target in space by a
free-flight interceptor.
3.4 Lambert’s Transfer
For an impulsive two-body orbital transfer, Lambert’s theorem states that the
transfer time, t2 − t1 , is a function of only the semi-major axis, a, of the transfer
orbit, the sum of the two radii, r1 + r2 , and the chord, c, joining the two positions.
Thus, we have
t2 − t1 = f (a, r1 + r2 , c) (3.62)
Since there is no dependence of the transfer time on the orbital eccentricity, one can
choose any value of e for the orbit, provided (a, r1 + r2 , c) are unchanged. This
is tantamount to changing the orbital shape for given initial and final positions by
moving the two foci in such a way that both (a, r1 + r2 ) are unaffected. Lambert’s
theorem is readily proved for parabolic, elliptic, and hyperbolic transfer orbits [8].
An outcome of Lambert’s theorem is Lambert’s problem which refers to the two-
point boundary value problem (2PBVP) resulting from an orbital transfer between
two position vectors in a given time. Such a problem is commonly encountered in the
guidance of spacecraft and ballistic missiles, as well as in the orbital determination

of space objects from two observed positions separated by a specific time interval.
For the initial and final positions, (r1 , r2 ), the specified time of flight, t12 = t2 − t1 ,
(or the transfer angle, Δθ , in Fig. 3.3) Lambert’s theorem indicates a unique value
of the semi-major axis, a, which must be determined from the orbital plane, (r1 , r2 ),
and the respective true anomalies, (θ1 , θ2 ).
Since each value of semi-major axis produces two possible transfer times
corresponding to either Δθ ≤ π or Δθ > π , a simple iterative procedure may
not converge to a unique a. Furthermore, the derivative of t12 with respect to a
is infinite for a minimum energy orbit, therefore a gradient-based scheme such
as Newton’s method cannot be used for solving the time equation with respect
to a, if a = am . For practical guidance, a robust algorithm is needed which
converges to a unique, non-singular solution. Examples of Lambert’s algorithms
are the universal variable methods by Battin [8], which can be applied to elliptical,
parabolic, and hyperbolic transfer trajectories. These methods are formulated either
by the evaluation of a power series (Stumpff functions) or via continued fractions
(hypergeometric functions).
3.4.1 Stumpff Function Method
Consider an auxiliary variable, x, given by

√
x = ΔE a , (3.63)
where ΔE = E2 − E1 , and a is the semi-major axis of the transfer orbit. Lagrange’s

coefficients, f, g, are expressed as functions of x as follows [7]:
x2
f = 1− C(z)
r1
√
x μ
f˙ = [zS(z) − 1]
r1 r2
1
g = t12 − √ x 3 S(z) (3.64)
μ
x2
ġ = 1 − C(z)
r2
where
x2
z= , (3.65)
a
and C(z), S(z) are the following Stumpff functions [76]:
1 z z2
C(z) = − + − ···
2! 4! 6!
1 z z2
S(z) = − + − ··· (3.66)
3! 5! 7!
The elliptical transfer is represented by z > 0, and the hyperbolic transfer by z < 0.
The parabolic case, z = 0, results in C(0) = 1/2, S(0) = 1/6, and can be solved
analytically.
Another universal variable, y, is now defined by
r1 r 2
y= (1 − cos Δθ ) (3.67)
p
where p is the parameter of the transfer orbit. The Lagrange’s coefficients are
expressed for the orbital transfer as follows:
r2
f = 1− (1 − cos Δθ )
p
-
˙ μ 1 − cos Δθ 1 − cos Δθ 1 1
f = − −
p Δθ p r1 r2
r1 r2
g= √ sin Δθ (3.68)
μp
r1
ġ = 1 − (1 − cos Δθ )
p
whose comparison with Eq. (3.64) yields the following:

-
y
x= (3.69)
C(z)
where x is the solution to the following cubic equation:

√ .
t12 μ = Ax C(z) + S(z)x 3 , (3.70)
and A is given by
-
r1 r2
A = sin Δθ (3.71)
1 − cos Δθ
Thus we have
y
f = 1−
r1
-
y
g=A (3.72)
μ
y
ġ = 1 − (3.73)
r2
while the remaining Lagrange’s coefficient, f˙, is obtained from the relationship
f ġ − g f˙ = 1. By expressing the transfer time in terms of the universal variables,
(x, y), rather than the classical elements, (a, e) (or (a, p)), the singularity and
convergence issues discussed earlier have been avoided. However, the convergence
of the present formulation is guaranteed only for Δθ < π .
In a terminal control application, it is required to find out the initial and final
velocities of the transfer, which can be finally derived from Lambert’s solution,
f, g, ġ, as follows:
1
v1 = (r2 − f r1 )
g
1
v2 = f˙r1 + ġv1 = (ġr2 − r1 ) (3.74)
g
The relationship among the transfer geometry, time, and Lagrange’s coefficients
is a transcendental one, which requires an iterative solution for the universal
variables, (x, y). Such an algorithm using Newton’s method is coded by Tewari [81].
For carrying out Newton’s iteration, we write
. √
F (x) = Ax C(z) + S(z)x 3 − t12 μ (3.75)
and take the derivative with respect to x,
.
C
(z) dz
F
(x) = A C(z) + 3S(z)x 2 + Ax √ + S
(z)x 3 (3.76)
2 C(z) dx
where
dz 2x
= (3.77)
dx a
and
1 2z 3z2
C
(z) = − + − − ···
4! 6! 8!
1 2z 3z2
S
(z) = − + − − ··· (3.78)
5! 7! 9!
The correction in each step is then calculated by
F (x)
Δx = − (3.79)
F
(x)
3.4.2 Hypergeometric Function Method
Battin [8] presents an alternative procedure for the general Lambert’s problem
through the use of hypergeometric functions evaluated by continued fractions,
instead of the Stumpff functions. Such a formulation improves convergence, even
for transfer angles greater than 180◦ , and renders the method insensitive to the initial
choice of z. Consider a universal variable x such that the semi-major axis, a, of the
transfer orbit is given by
am
a= , (3.80)
1 − x2
where am is the semi-major axis of the minimum energy ellipse connecting the
initial and final radii. Clearly, for an elliptical transfer we have −1 < x < 1
(x = 0 for minimum energy transfer), while for a parabola, x = 1. A hyperbolic
transfer requires x > 1. Furthermore, for a given transfer, (r1 , r2 , Δθ ), we define
the following parameter, λ:
√
r1 r2 Δθ
λ= cos (3.81)
s 2
where s is the semi-perimeter,
r1 + r2 + c
s = 2am = (3.82)
2
Note that −1 < λ < 1. Additional variables, y, η, and z, are defined by
.
y= 1 − λ2 (1 − x 2 ) (3.83)
η = y − xλ (3.84)
and
1
z= (1 − λ − xη) (3.85)
2
which transform the time equation to the following:

-
μ 4
t12 3
= η3 F (z) + 4λη (3.86)
am 3
Here F (z) is a hypergeometric function [8] that can be evaluated by a continued

fraction as follows:
1
F (z) = γ1 z (3.87)
1− γ2 z
1− γ z
1− 3
..
1− .
and
⎧
⎨ (n+2)(n+5)
n odd
(2n+1)(2n+3)
γn = (3.88)
⎩ n(n−3)
n even
(2n+1)(2n+3)
When adopting Newton’s method to solve the time equation for x, we write [81]
-
4 3 μ
f (x) = η F (z) + 4λη − t12 3
(3.89)
3 am
and
dη 4 3
dz
f
(x) = 4 η2 F (z) + λ + η F (z) (3.90)
dx 3 dx
where prime denotes differentiation with respect to the argument of a function. From
the definitions given above, we have
dη λη
=− (3.91)
dx y
dz η2
=− (3.92)
dx 2y
and
6 − 3γ1 G(z)
F
(z) = (3.93)
2(1 − z)[1 − zγ1 G(z)]
3.5 Optimal Impulsive Transfer 101
where
1
G(z) = γ2 z (3.94)
1− γ3 z
1− γ z
1− 4
..
1− .
3.5 Optimal Impulsive Transfer
The problem of transferring a spacecraft from one set of position and velocity,
[r(t0 ), r(t0 )], to another, [r(tf ), r(tf )], can be optimized either by minimizing the
total velocity change magnitude, Δv or the total time of transfer, tf − t0 . The
transfer is made by applying a number N of velocity impulses, Δvk , at different time
instants, t0 ≤ tk ≤ tf , such that immediately after the application of the k th impulse,
there is an instantaneous change in the velocity without affecting the position:
v̄k = v(tk− ) + Δvk , r̄k = r(tk− ) (1 ≤ k ≤ N ) (3.95)
where r(tk− ), v(tk− ) denote the position and velocity, respectively, of the two-body
orbit evolved from the initial state, [r(tk−1 ), v̄(tk−1 )], to the time t = tk , after the
application of the previous velocity impulse at t = tk−1 , as follows:
ṙ(t) = v(t) r( tk−1 ) = r̄k−1

r(t)
v̇(t) = −μ , v( tk−1 ) = v̄k−1 , (tk−1 ≤ t ≤ tk ) (3.96)
r 3 (t)
Clearly, the motion evolves from the first impulse applied at time t = t0 , with
v(t0− ) = v(t0 ), to the final impulse at t = tf = tN , and the terminal boundary
condition requires the final impulse to be the following:
ΔvN = v(tf ) − v(tN − ) (3.97)
A fuel-optimal trajectory would minimize the following Hamiltonian:
#
N
r(tk )
H = | Δvk | +λTrk [v(tk− ) + Δvk ] − μλTvk (3.98)
r 3 (tk )
k=1
relative to Δvk , k = 1, . . . , N , where λrk and λvk are the costate vectors corre-
sponding to r(tk ) and v(tk ), respectively. A spacecraft equipped with a chemical
rocket engine can adjust both the magnitude and the direction of each velocity
impulse, Δvk , their time instants, tk , as well as the total number, N , of the impulses.
In general, due to the large number of variables, such an optimization problem has
many possible solutions, and is therefore indeterminate. However, solutions can be
derived analytically for specific problems.
3.5.1 Coasting Arc
Since an impulsive transfer consists of a zero-thrust (also called null-thrust or

coasting) arc between any two velocity impulses, it is possible to use the optimal
control method to evolve the costate vectors such that the optimal magnitude and
direction of the next impulse is obtained. Consider the following Hamiltonian along
a coasting arc:
H = λTr v + λTv g (3.99)
where g = −μr/r 3 . The costate equations for the arc to be an extremal trajectory
are the following:
T T
∂H ∂g(r)
λ̇r = − =− λv (3.100)
∂r ∂r
T
∂H
λ̇v = − = −λr (3.101)
∂v
where the gravity-gradient matrix, ∂g/∂r, is the following:
∂g(r) μ 3μ
= − 3 I3×3 + 5 rrT (3.102)
∂r r r
A substitution of Eq. (3.101) into Eq. (3.100) yields the following second-order
differential equation for λv (t):
T
∂g(r)
λ̈v = λv (3.103)
∂r
Resolving λv into components along the local-horizon axes, (ir , iθ , ih ), we have
λv = λvr ir + λvθ iθ + λvn ih (3.104)
and substituting into Eqs. (3.102) and (3.103), the latter becomes the following:
μ
λ̈v = (2λvr ir − λvθ iθ − λvn ih ) (3.105)
r3
Now, take the time derivatives of λv as follows:
∂λv
λ̇v = + θ̇ ih × λv (3.106)
∂t
∂ 2 λv
λ̈v = + θ̈ ih × λv
∂t 2
∂λv
+ 2θ̇ ih × + θ̇ 2 ih × (ih × λv ) (3.107)
∂t
where ∂t∂ denotes taking the time derivative of a vector’s components. Introduce
for convenience, λv = rq, with q = (q1 , q2 , q3 )T , and substitute the Keplerian
equations of motion expressed as follows:
μ
r̈ − r θ̇ 2 = −
r2
r θ̈ + 2ṙ θ̇ = 0 (3.108)
Then Eq. (3.107) becomes the following:
∂ 2q ∂q
λ̈v = r + 2r θ̇ih ×
∂t 2 ∂t
μ
+ r θ̇ 2 ih × (ih × q) − 2 q (3.109)
r
In order to integrate the equation, we change the independent variable from the time
to the true anomaly, θ , which results in the following differential equation with the
prime denoting the derivative with respect to θ :

μp

r
λ̈v = q + 2i h × q + i h × (i h × q) − q (3.110)
r3 p
On substituting Eq. (3.110) into Eq. (3.105) with λv = rq, we have the following
scalar differential equations for the elements of q:
3r
q1
− 2q2
= q1
p
q2
+ 2q1
= 0 (3.111)
q3
+ q3 = 0
Of these, the last is readily integrated to yield
q3 = A3 cos θ + B3 sin θ (3.112)
where A3 , B3 are the constants of integration. Furthermore, since the gravity field is
time invariant, the Hamiltonian is constant on the coasting arc:
T
H = −λ̇v v + λTv g
= eq1
sin θ (1 + e cos θ ) + q2
(1 + e cos θ )2 + q1 (2 + e2 + 3e cos θ )
= C = const. (3.113)
This integral is identically satisfied by the first of Eq. (3.111), which is therefore
discarded from future consideration. The second of Eq. (3.111) is now integrated to
q2
= Ae − 2q1 , where A is a constant, and this result is used to eliminate q2
from
Eq. (3.113), which is then integrated to yield the following result [52]:
q1 = (1 + e cos θ )(A cos θ + Be sin θ + CI1 ) (3.114)

D − A sin θ
q2 = (1 + e cos θ ) −A sin θ + B(1 + e cos θ ) + + CI2
1 + e cos θ
where B, D are constants, and
cot θ 1 + e cos θ
I2 = + I1
e(1 + e cos θ ) e sin θ

dθ
I1 = sin θ (3.115)
sin2 θ (1 + e cos θ )2
Analytical expressions for evaluating the integral, I1 , are provided by Lawden [50].
Finally, the substitution of the orbit equation, Eq. (3.6), gives the costate vector
for velocity, λv , as follows:
λvr = A cos θ + Be sin θ + CI2

D − A sin θ
λvθ = −A sin θ + B(1 + e cos θ ) + + CI2 (3.116)
1 + e cos θ
A3 cos θ + B3 sin θ
λvn =
1 + e cos θ
The vector λ̇v = −(λrr , λrθ , λrn )T (the negative of the costate vector for radius
λr ) can now be obtained by the first equation of Eq. (3.106), thereby completing the
optimal solution for the coasting arc:
√
μp A sin θ − D
λrr = −(λ
vr − λvθ )θ̇ = − B + CI3
r2 1 + e cos θ
-
μ
λrθ = −(λvr + λ
vθ )θ̇ = − [−A(e + cos θ ) + De sin θ + C cos θ ]
p3
-

μ
λrn = −λvr θ̇ = − [−A3 sin θ + B3 (e + cos θ )] (3.117)
p3
where
e sin θ − cos θ I1
I3 = − (3.118)
e sin θ (1 + e cos θ )2 e sin θ
Equations (3.116) and (3.117) cannot be applied for the special case of a circular
orbit (e = 0), for which the required integrations can be carried out separately (and
much more easily) by putting e = 0 in the differential equations. The results, only
for e = 0, are the following:
λvr = A cos θ + B sin θ + 2C

λvθ = −2A sin θ + 2B cos θ + D − 3Cθ (3.119)
λvn = A3 cos θ + B3 sin θ
-
μ
λrr = − 3 (A sin θ − B cos θ + 3Cθ − D)
r
-
μ
λrθ = (A sin θ − B cos θ + 3Cθ − D) (3.120)
r3
-
μ
λrn = (A3 sin θ − B3 cos θ )
r3
The necessary conditions of optimality on a coasting arc require that both λv

and λ̇v must be continuous across a velocity impulse, and that the direction of each
impulse must be opposite to λv . These can be obtained by taking a derivative of
the Hamiltonian, Eq. (3.98), with respect to Δvk , k = 1, . . . , N, and equating it to
zero for the stationary condition. We shall return to the optimal thrust direction in
Chap. 5.
3.5.2 Hohmann Transfer
Hohmann showed that the optimal transfer between two circular orbits consists
of two velocity impulses applied tangentially to the respective orbits. Hohmann
transfer between two circular orbits of radii r1 and r2 , respectively, is depicted
in Fig. 3.6. Since both the velocity impulses, Δv1 and Δv2 , are tangential to the
orbits, the semi-major axis of the transfer ellipse is the average of the two radii,
a = (r1 + r2 )/2. It can be shown that Hohmann transfer satisfies the necessary
conditions for optimality as follows.
Let the two velocity impulses be applied at angles φ1 and φ2 , respectively, from
the local horizon, (ir , iθ ):
Δv1 = Δv1θ iθ1 + Δv1r ir1

= (v1
cos φ1 − v1 )iθ1 + v1
sin φ1 ir1
Δv2 = Δv2θ iθ2 + Δv2r ir2 (3.121)
= (v2 − v2
cos φ2 )iθ2 − v2
sin φ2 ir2
Fig. 3.6 Geometry of Dv2

Hohmann transfer between
two circular orbits
r2
r1
Dv1
√ √
where v1 = μ/r1 and v2 = μ/r2 are the two circular orbital speeds, v1
denotes
the speed immediately after the application of the first impulse, and v2
the speed
immediately before the application of the second impulse. The impulse magnitudes
are thus the following:
,
Δv1 = | Δv1 |= (v1
− v1 )2 + 2v1 v1
(1 − cos φ1 )
,
Δv2 = | Δv2 |= (v2
− v2 )2 + 2v2 v2
(1 − cos φ2 ) (3.122)
The total velocity impulse magnitude, J = Δv1 + Δv2 , is the performance index
to be minimized with respect to the control variables, v1
, v2
, φ1 , φ2 , subject to the
following constraints on the transfer orbit:
(i) The specific energy, , is constant along the transfer ellipse. This implies that
(v1
)2 − (v2
)2 + 2(v22 − v12 ) = 0 (3.123)
(ii) The specific angular momentum magnitude, h, is conserved along the transfer
ellipse, thereby implying that
v22 v1
cos φ1 − v12 v2
cos φ2 = 0 (3.124)
The Hamiltonian of the optimal control problem is thus given by

, ,
H = (v1
− v1 )2 + 2v1 v1
(1 − cos φ1 ) + (v2
− v2 )2 + 2v2 v2
(1 − cos φ2 )
+ λ1 [(v1
)2 − (v2
)2 + 2(v22 − v12 )] + λ2 (v22 v1
cos φ1 − v12 v2
cos φ2 ) (3.125)
where λ1 , λ2 are the Lagrange multipliers associated with the two equality con-
straints.
The stationary conditions yield the following:
∂H v1
− v1 cos φ1
= + 2λ1 v1
+ λ2 v22 cos φ1 = 0
∂v1
Δv1
∂H v
− v2 cos φ2

= 2 − 2λ1 v2
− λ2 v12 cos φ2 = 0
∂v2 Δv2
∂H v1 v1
= sin φ1 − λ2 v22 v1
sin φ1 = 0 (3.126)
∂φ1 Δv1
∂H v2 v2
= sin φ2 + λ2 v12 v2
sin φ2 = 0
∂φ2 Δv2
the last two of which require φ1 = φ2 = 0, which substituted into the constraints,
Eqs. (3.123) and (3.124), yield
v12
v1
= ,
v12 +v22
2
v22
v2
= , (3.127)
v12 +v22
2
This proves that Hohmann transfer ellipse is indeed an extremal trajectory, satisfying
the necessary conditions of optimality.
The range of the ratio, r2 /r1 , for which Hohmann transfer is optimal can be
derived by the continuity of the costate vector, λv , and its time derivative, λ̇v , at
both the points of application of the impulses (see Eq. (3.116)). Using e = 0 for
the initial and final circular orbits, the costate solutions at the impulse application
points can be expressed as follows [52]:
λvr = R sin θ
λvθ = 2R cos θ + D (3.128)
√
where R = A2 + B 2 and θ is the true anomaly measured from the first impulse
point. The magnitude of λv on a circular orbit is thus given by
λ2v = λ2vr + λ2vθ = R 2 + D 2 + 4RD cos θ + 3R 2 cos2 θ (3.129)

For optimality, λ2v must be stationary with respect to θ , which happens either at the
points θ = 0, π or when
2D + 3R cos θ = 0 (3.130)
However, for a real solution of Eq. (3.130), we must have | D |≤ 3R/2, in which
range λ2v is stationary only for θ = 0 and θ = π . Therefore, in all cases (unless
R = 0) the impulses must be applied either at the periapsis (θ = 0) or the apoapsis
(θ = π ) of the transfer ellipse.
The cases θ = 0, π require that λvr = 0, which implies the impulses should
be applied along ±iθ , that is, tangentially to the circular orbit. However, to have a
continuity of the time derivative λ̇v , across each impulse applied on different circular
orbits lying at either end of the coasting arc, one can set A = 0, B = (1 − D)/2e,
resulting in the following:
1
λvr = (1 − D) sin θ
2
λvθ = (1 − D) cos θ ) ± D (3.131)
with 0 ≤ D ≤ 1. Here the positive sign is taken for the impulse applied at θ = 0,
and the negative for that at θ = π . For the impulse applied in the direction of flight,
the λ̇v components are given by
⎧ ,
⎨ 1 μ3 (1 + e)2 (2 − e) (θ = 0)
2 p
λrr = 1, μ
⎩
2 p3 (1 − e) (2 + e) (θ = π )
2
λrθ = 0 (3.132)
Applying Eq. (3.132) to the impulse at the periapsis, we have

-
1 μ 1 μ D−1
3
(1 + e)2 (2 − e) = − −D + (3.133)
2 p 2 r13 2
or
√
(2 − e) 1 + e = 1 + D (3.134)
Similarly, for the impulse at the apoapsis, r = r2 , we have

3/2
r2 √
(1 − e)2 (2 + e) = (2 + e) 1 − e = 1 + D̄ (3.135)
p
1.8
1.6
1.4
1.2
f(z)
0.8
0.6
0.4
0.2
0 z2 1.5
−1 −z1 −0.5 0 0.5 1 2
z
Fig. 3.7 Roots for satisfying the optimality constraints on the eccentricity of a Hohmann transfer
ellipse
where D̄ is the value of D for the larger circle. Since both D and D̄ should have
values between zero and unity, we have the following inequalities to be satisfied for
optimality:
√
1 ≤ (2 − e) 1 + e ≤ 2
√
1 ≤ (2 + e) 1 − e ≤ 2 (3.136)
To solve for the appropriate value of the orbital√eccentricity which satisfies these
inequalities, consider the plot of f (z) = (2 − z) 1 + z for −1 ≤ z ≤ 2, shown in
Fig. 3.7, wherein the range 1 ≤ f (z) ≤ 2 corresponds to −z1 ≤ z ≤ z2 , where z1 =
0.8794 and z2 = 1.3473. Therefore, the optimal value of e satisfying Eq. (3.136)
must fall between zero and the smaller of z1 and z2 , that is, 0 ≤ e ≤ 0.8794.
Now, since e = (r2 −r1 )/(r1 +r2 ) for the transfer ellipse, it follows that x = r2 /r1
must satisfy
x−1
0≤ ≤ 0.8794 (3.137)
x+1
or 1 ≤ x ≤ 1.8794/0.1206 = 15.58. Hence, Hohmann transfer is optimal only for

1 ≤ r2 /r1 ≤ 15.58. For larger values of the radius ratio, r2 /r1 > 15.58, a transfer
with more than two impulses would be optimal (instead of Hohmann transfer).
3.5.3 Outer Bi-elliptical Transfer
A generalization of Hohmann transfer for the three impulse case is the outer bi-
elliptical transfer depicted in Fig. 3.8. Here the transfer consists of two ellipses,
and takes the spacecraft to an intermediate radius, rt > r2 > r1 , with all the three
impulses applied tangentially. The third impulse, Δv3 , is applied opposite to the
direction of the other two impulses, as shown in Fig. 3.8. The time of flight is
determined by rt , which itself is a nonlinear function of r1 and r2 . Thus the sole
design parameter for this transfer is the intermediate radius, rt . It has been indicated
above that the outer bi-elliptical transfer can be more efficient than Hohmann
transfer for r2 /r1 > 15.58. The reader can show this by minimizing the net velocity
change, | Δv1 | + | Δv2 | + | Δv3 |, with respect to rt , and equating the result with
that of Hohmann transfer.
Consider x = r2 /r1 ≥ 1 and y = rt /r1 ≥ 1 as the two non-dimensional
parameters of the outer bi-elliptical (OB) transfer. The total velocity change required
Fig. 3.8 Geometry of the Dv2

outer bi-elliptical transfer
between two circular orbits
rt
r2
r1
Dv1
Dv3
for the Hohmann and OB transfers for the same value of r1 and r2 is then
respectively expressed as follows:
- 0-
μ 1 2x 2
ΔvH = −1 + √ + −
r1 x 1+x x(1 + x)
- 0
μ 1 2y 2x 2 2y
ΔvB = −1 − √ + + − +
r1 x 1+y y(x + y) y(1 + y) x(x + y)
(3.138)
The difference between the non-dimensional velocity change of the two transfers is
then expressed by the following function:
ΔvB − ΔvH
f (x, y) = , (3.139)
μ
r1
which vanishes for x = y (that is, when the two transfers become exactly the same).
The value of x = y for which the OB transfer becomes more efficient than Hohmann
transfer is then obtained from the stationary condition of f (x, y) given by
∂f ∂f
= =0
∂x ∂y
-
3y + 1 x
= − (3.140)
(y + 1)3/2 x+y
3x + 1 1
= −√
(x + 1)3/2 2
or
x 3 − 15x 2 − 9x − 1 = 0 (3.141)
whose only positive real root is x = 15.5817. Since the second derivative of f (x, y)
is always negative, given by
∂ 2f 1
=− √ <0 (3.142)
∂x∂y 2 2xy(x + y)3/2
it follows that f (x, y) = 0 is a maximum point for x = y. Therefore, f (x, y) < 0

for x > 15.5817, thereby implying that OB transfer is more efficient than Hohmann
transfer for r2 /r1 > 15.5817.
Exercises
3.1 A spacecraft is detected to be moving at a speed 7 km/s relative to Earth, when

its altitude is 500 km. What additional minimum velocity must be provided to the
spacecraft by an impulsive rocket thrust in order to escape Earth’s gravity? (For
Earth: μ = 398,600.4 km3 /s2 , mean surface radius, R0 = 6378.14 km).
3.2 A spacecraft is initially in an Earth orbit of perigee altitude 300 km and apogee
altitude 1000 km. It is required to transfer the spacecraft to a new coplanar orbit of a
constant altitude of 500 km. Determine the minimum total velocity change required,
and the corresponding time of transfer in each of the following cases:
(a) Only one impulse is applied.
(b) Two impulses are applied.
3.3 Two astronauts called A and B have landed on Charon (μ = 101.25 km3 /s2 ,
surface radius = 605 km), while a third one called C is on a spacecraft orbiting
Charon in a circular orbit of radius 650 km. The astronauts have decided to play a
game of cricket. Astronaut B bowls to A, who hits the ball such that it reaches C
in a minimum energy orbit when C is 10◦ ahead of the radial position of A. The
atmosphere on Charon is negligible.
(a) Determine the magnitude and direction of the velocity provided by A to the ball.
(b) Determine the ball’s velocity (magnitude and direction) when it is caught by C.
(c) Determine the time of transfer of the ball from A to C.
(d) Calculate the position of C relative to A when A hits the ball.
(e) Use the hypergeometric function method to compute the transfer orbit, such that
the transfer from A to C in the given position should take exactly 5 min. Verify
that your answer is correct.
(f) After catching the ball, astronaut C waits until he is directly overhead of A. At
this time, he throws the ball vertically downwards toward A. But the ball instead
of reaching A, returns to C in 5174 s. What is the velocity provided to the ball
by C?
3.4 Repeat parts (a) and (b) of Exercise 3.3, if the transfer angle is not specified,
but the ball is hit horizontally by A, and reaches C via a minimum energy transfer
orbit.
3.5 Determine the minimum energy transfer orbit and the required velocities at both
the ends, for the positions, r1 , r2 , given in Example 3.2. Is it a practical transfer
trajectory?
3.6 Use the hypergeometric function method for solving Lambert’s problem to
compute the transfer orbit for sending a spacecraft around Earth from a position
in the Earth centered frame, (I, J, K),
r1 = −3000I + 6000K km
to the position
r2 = 500I − 7000J + 1000K km
in exactly 20 min. Give your answer in terms of the classical orbital elements.
3.7 Calculate the velocity impulse vectors for transferring a spacecraft around Earth
from a position
r1 = −3000I + 6000K km
and velocity
v1 = 7I − 2J km/s
to the position
r2 = 7000I − 2000J km
and velocity
v2 = 8I + J − K km/s
in exactly 30 min.
3.8 Estimate the smallest total velocity impulse magnitude required for sending a
spacecraft from Earth (orbital radius = 1 astronomical unit (A.U.), orbital period =
365.256 mean solar days) to Mars (orbital period 687 mean solar days). (For Sun,
μ = 1.3271244 × 1011 km3 /s2 , 1 mean solar day = 24 h, 1 A.U. = 1.495978 × 108
km).
3.9 Estimate the smallest total velocity impulse magnitude required for sending a
spacecraft from Earth to Neptune (orbital radius 30.1 A.U.).
Chapter 4
Two-Body Maneuvers with Unbounded
Continuous Inputs
4.1 Introduction
A majority of spacecraft navigation problems involve transfers between two given

positions and velocities. When continuous inputs are applied for the orbital transfer
of a spacecraft around a central body, the perturbed two-body model can be used as a
plant to derive the optimal trajectory and control history. In such a model, the central
gravity field is regarded to be spherical and the gravitational perturbations caused
by other bodies are neglected. The effects of any atmosphere and solar radiation
pressure are also neglected, unless the nature of the trajectory demands that they
be included. After an optimal trajectory has been determined, an optimal feedback
loop performs automatic course corrections at regular intervals to keep the vehicle
on the nominal trajectory despite small, unmodeled disturbances. In this chapter,
we shall consider how the optimal trajectories can be devised using the perturbed
Keplerian equations and unbounded acceleration inputs, and used to perform the
task of optimal orbital regulation.
4.2 A Motivating Example
Consider the problem of optimally controlling the displacement, y(t), of a block of

unit mass sliding in a horizontal, frictionless track by the application of a force, u(t).
Since the block faces no natural resistance to its motion for u(t) = 0, this system
can be regarded as the simplest possible idealization of a spaceflight problem where
neither the atmospheric drag nor any other dissipative natural forces are present.
The second-order, linear time-invariant system can be represented by the following
state equations:
ẋ1 = x2
ẋ2 = u (4.1)

https://doi.org/10.1007/978-3-030-03789-5_4
116 4 Two-Body Maneuvers with Unbounded Continuous Inputs
where x1 (t) = y(t). The control task is to move the system from an initial state of
rest, x1 (0) = x2 (0) = 0, to a final rest position of x1 (tf ) = 1 and x2 (tf ) = 0 at an
unspecified time tf , while minimizing the total control effort given by
tf
J = u2 (t)dt (4.2)
0
The extremal control history, u∗ (t), and the corresponding trajectory,

x1∗ (t), x2∗ (t),
are readily solved from the Euler–Lagrange necessary conditions
(Chap. 2). The Hamiltonian of the problem is given by
H = L + λT f = u2 + λ1 x2 + λ2 u (4.3)
which results in the following costate equations:

T
∗ λ̇∗1 ∂H ∗ 0
λ̇ = =− = (4.4)
λ̇∗2 ∂x −λ∗1
The solution to the costate equations subject to the initial condition, λ∗ (0) =
(c1 , c2 )T , is expressed by
∗
∗ λ1 c1
λ (t) = = (4.5)
λ∗2 c2 − c1 t
where the constants c1 , c2 are to be determined from the boundary conditions.

The stationary condition of the Hamiltonian with respect to the control history,
u(t),
T
∂H ∗
Hu = = 2u∗ (t) + λ∗2 (t) = 0 (4.6)
∂u
yields the following extremal control history:
1
u∗ (t) = −λ∗2 (t)/2 = − (−c2 + c1 t) , (0 ≤ t ≤ tf ) (4.7)
2
In order to determine the constants c1 , c2 , the state equations are solved, subject to
the boundary conditions. The extremal trajectory resulting from the extremal control
history (Eq. (4.7)) and zero initial state is expressed as follows:

x1∗ (t) c1 t 3 /12 − c2 t 2 /4
x ∗ (t) = = , (0 ≤ t ≤ tf ) (4.8)
x2∗ (t) c1 t 2 /4 − c2 t/2
The terminal boundary conditions are now applied to determine the constants
c1 , c2 , tf :
x2∗ (tf ) = c1 tf2 /4 − c2 tf /2 = 0 (4.9)

4.3 Equations of Motion 117
which yields tf = 2c2 /c1 .
x1∗ (tf ) = c1 tf3 /12 − c2 tf2 /4 = 1 (4.10)
into which the substitution of Eq. (4.9) yields
tf3 = −24/c1
We must have c1 < 0 (and c2 < 0) for tf > 0. Finally, the terminal boundary
condition for extremal control is applied,

∂φ ∗
H+ = (u∗ (tf ))2 + λ∗1 (tf )x2∗ (tf ) + λ∗2 (tf )u∗ (tf ) = 0 (4.11)
∂t t=tf
to produce the result u∗ (tf ) = 0, or tf = c2 /c1 . Clearly, there is a contradiction

between Eqs. (4.9) and (4.11), both of which cannot be simultaneously satisfied.
Hence, there does not exist any extremal trajectory satisfying the boundary condition
x2 (tf ) = 0. This fact can be explained physically as follows. Since there is no
resistance to the unforced motion, the velocity at the terminal time cannot be made
zero by an extremal control, because the optimality condition, Eq. (4.11), demands
the control force—which is initially positive (u∗ (0) = −c2 /2)—must vanish at the
terminal time. In fact, to make the terminal velocity zero, a negative force must
act for some time t < tf , and due to the time linearity of the control history,
should become increasingly negative as the time increases. Thus u(tf ) < 0 is
necessary for satisfying x2 (tf ) = 0, which conflicts with the necessary condition for
optimality. Therefore, the boundary condition x2 (tf ) = 0 causes the optimal control
problem with continuous inputs to be ill posed. Such a characteristic of the energy-
optimal, continuous control history is valid for spaceflight problems wherein a given
unforced terminal state can only be achieved by a suboptimal path. As discussed in
Chap. 3, the application of discontinuous (or impulsive) thrust inputs can solve the
problem of optimal transfer between orbits (which are the unforced states of the
system). Alternatively, a well-posed spaceflight optimal navigation problem with
continuous thrust is the one where the terminal velocity constraint is replaced by a
suitable terminal cost, φ[x(tf ), tf ], in the objective function to be minimized. This
will be further explored in this chapter for controlling spaceflight trajectories with
continuous, unbounded inputs.
4.3 Equations of Motion
The orbital motion of a spacecraft around a spherical body perturbed by a

continuous acceleration input, a(t) ∈ R3×1 , can be represented by the following
vector differential equations:
r
v̇ + μ =a (4.12)
r3
and
ṙ = v (4.13)
where r(t) ∈ R3×1 (with r(t) = |r(t)|) is the position vector measured from the
center of gravity of the two-body system, and μ is the gravitational constant. If the
spacecraft’s mass is assumed to be negligible relative to the central body, the origin
of the coordinate system shifts to the center of the body, which is then assumed
to be at rest. Hence a coordinate frame with origin at the center of the large body
can be regarded as an inertial reference frame, and r̈(t) is the absolute acceleration
of the spacecraft. Equations (4.12) and (4.13) can generally be expressed either in
spherical or Cartesian coordinates. However, an alternative system of coordinates
regards the plane of the unperturbed Keplerian orbit (i.e., when a = 0) as a
reference plane. Here the polar coordinates, r(t), θ (t), can be used to resolve
the spacecraft’s position vector, r(t), in the unperturbed orbital plane, while the
angle, α(t), represents the plane change caused by the out-of-plane component
of a(t). Such a representation is useful especially when the magnitude of the
perturbing acceleration, a(t) = |a(t)|, is always small compared to the gravitational
acceleration (i.e., a << rμ2 ), resulting in a slow departure from the Keplerian
motion. In such a case, at any given instant, t0 , one can imagine an osculating
orbit, which is the solution to Eq. (4.12) for a(t) = 0, t ≥ t0 . The osculating
orbit is thus the imaginary two-body orbit with initial condition, r(t0 ), v(t0 ), and
will be the trajectory of the spacecraft if the perturbing acceleration, a(t), suddenly
vanishes at t = t0 and remains zero for t > t0 . If the applied perturbing acceleration
is small1 at all times, then the osculating orbit is nearly tangential to the actual
trajectory of the spacecraft at any given instant, t0 , thereby explaining the adjective
“osculating” (or “touching”). Whenever continuous acceleration inputs are applied
in practice, their magnitudes are always much smaller than the acceleration due to
gravity, except in certain special situations such as lift-off and landing (which may
be considered separately). Therefore, the approximation, a << rμ2 , is generally
valid in orbital maneuvers involving continuous inputs. The osculating orbit is
thus a useful reference trajectory for space navigation problems involving small,
continuous inputs because it allows a simplification of the governing equations of
motion.
1 An osculating orbit is defined to be the tangential trajectory that will be followed if the perturbing
acceleration were suddenly to be removed. If the perturbation is large and were suddenly to be
removed, there would be an almost impulsive change in the acceleration at the instant of removal,
which translates into a step change of velocity. Therefore, in that case the resulting orbit would
not be tangential to the original (perturbed) orbit. In fact, there will be a “bump” in the trajectory.
Hence the concept of osculating orbit is valid only if the applied perturbation is always small, so
that at any instant, an abrupt vanishing of the acceleration would not cause an abrupt change in
the velocity, thereby approximating the resulting orbit as being tangential (or osculating) to the
original (perturbed) orbit.
4.3 Equations of Motion 119
Utilizing the set of nonlinear ordinary differential equations, Eqs. (4.12) and
(4.13), expressed in Cartesian or spherical coordinates for optimal guidance appli-
cation requires numerical integration, which has both computational and implemen-
tation difficulties. Instead, an approximate model based upon either linearization
about the osculating orbit or a variation of Keplerian orbital parameters is employed
in practical applications. The linearized model is useful when the reference orbit
is fixed, such as in orbital regulation of and maneuvers around a given circular
orbit. On the other hand, the variational model is more useful when apart from
the thrust acceleration, other perturbations are also present, such as due to non-
spherical gravity field, atmospheric drag, or solar radiation pressure. The classical
variational model is via the Lagrange’s (or alternatively, Gauss’) perturbation
equations, wherein the variation of the classical orbital elements of the osculating is
used to represent the perturbed motion. A more general perturbation model utilizes
the universal or canonical variables instead of the classical elements, in order to
avoid the singularity of the latter in certain conditions. In any case, a set of six
state variables is necessary to represent the position and velocity, r(t), ṙ(t), of the
spacecraft in an inertial reference frame. When the osculating orbital model is used
as a reference, then only four state variables are required to represent the motion of
the spacecraft relative to the instantaneous osculating plane at t = t0 . In that case,
the deviation of the spacecraft’s trajectory from the osculating orbit is represented
by the respective changes in the orbital radius, r, the radial velocity, ṙ, the orbital
angular momentum, h = r 2 θ̇ , and the instantaneous plane change angle, α. The
remaining two variables required to complete the description are the two angles
used to resolve the orbital angular momentum vector, h(t0 ) = r(t0 ) × v(t0 ), of
the osculating orbit in the inertial frame. The osculating orbit model is depicted in
Fig. 4.1.
The vector differential equation, Eq. (4.12), results in the following scalar
constituents when the deviation from a reference osculating orbit is considered:
h2 μ
r̈ − + 2 = ar (4.14)
r3 r
Fig. 4.1 The osculating orbit

h(t) = r(t)´v(t) v(t)
used as a reference for the Actual orbit
actual trajectory of a
spacecraft under the influence
of small perturbing
acceleration inputs, a r(t) v(t0)
a
a
h(t0) = r(t0)´v(t0)
r(t0)
Osculating orbit
ḣ = raθ (4.15)
and
r
α̇ = an (4.16)
h
where the input acceleration vector is resolved into the radial, circumferential, and
normal components, ar , aθ , and an , respectively:
a = (ar , aθ , an )T (4.17)
4.4 Optimal Low-Thrust Orbital Transfer
The slow departure from a Keplerian orbit when a small input acceleration is applied
by a low-thrust ion/plasma engine is readily modeled by Eqs. (4.14)–(4.16). Let the
state vector be the following:
x = (r, ṙ, h, α)T (4.18)
which results in the following state equation:
ẋ = f(x, a) (4.19)
where
⎛ ⎞
ṙ
⎜ ⎟
⎜ h2 ⎟
⎜ r 3 − rμ2 + ar ⎟
f=⎜
⎜
⎟
⎟ (4.20)
⎜ raθ ⎟
⎝ ⎠
r
h an
The simplest optimal control problem relates to the low-thrust orbital transfer
from the initial state
x(ti ) = xi (4.21)
to a final state
x(tf ) = xf (4.22)
such that the net control effort,
L = aT Ra (4.23)
4.4 Optimal Low-Thrust Orbital Transfer 121
is minimized for the control interval, t ∈ [ti , tf ], with R ∈ R3×3 being a

symmetric cost coefficient matrix. The performance index to be minimized is thus
the following:
tf
J = a(t)T R(t)a(t)dt (4.24)
ti
and the Hamiltonian for the problem is given by
h2 μ r
H = L + λT f = aT Ra + λ1 ṙ + λ2 ( 3
− 2 + ar ) + λ3 raθ + λ4 an (4.25)
r r h
where λ = (λ1 , λ2 , λ3 , λ4 )T is the costate vector.

The transverse condition for optimality,
T
∗ ∂H ∗
λ̇ = − (4.26)
∂x
yields the following costate equations:

3(h∗ )2 2μ an∗
λ̇∗1 = − ∗ 3 λ∗2 − λ∗3 aθ∗ − λ∗4 (4.27)
(r ∗ )4 (r ) h∗
λ̇∗2 = −λ∗1 (4.28)
2h∗ ∗ r ∗a∗
λ̇∗3 = − ∗ 3
λ2 + ∗ n2 λ∗4 (4.29)
(r ) (h )
λ̇∗4 = 0 (4.30)
Finally, the stationary condition, Ha = 0 on the extremal trajectory, gives the

following equation to be solved for the extremal control, a∗ :
⎛ ⎞
λ∗2
2Ra∗ + ⎝ λ∗3 r ∗ ⎠ = 0 (4.31)
λ∗4 r ∗ / h∗
where the cost coefficient matrix R must be positive definite in order to satisfy the
sufficient condition for optimality on the extremal trajectory:
Haa = 2R > 0 (4.32)
Thus by choosing a positive definite R, we have â(t) = a∗ (t), t ∈ [ti , tf ].

4.4.1 Coplanar Orbital Transfer
The maneuvers which do not require a change of plane are termed coplanar
transfers, and must have an∗ (t) = 0, t ∈ [ti , tf ], which implies λ∗4 (t) = 0 throughout
the control interval. The fourth costate equation, Eq. (4.30), as well as the fourth
state variable, α, is now redundant in solving the optimal control problem. Thus we
have for the special case of coplanar maneuvers, x = (r, ṙ, h)T , a = (ar , aθ )T ,
λ = (λ1 , λ2 , λ3 )T , and
⎛ ⎞
ṙ
⎜ 2 ⎟
f = ⎝ h3 − μ2 + ar ⎠ (4.33)
r r
raθ
which result in the following costate equations:

3(h∗ )2 2μ
λ̇∗1 = − λ∗2 − λ∗3 aθ∗ (4.34)
(r ∗ )4 (r ∗ )3
λ̇∗2 = −λ∗1 (4.35)
2h∗ ∗
λ̇∗3 = − λ (4.36)
(r ∗ )3 2
It is to be noted that the extremal control can be easily derived in the coplanar case
by choosing

R1 0
R= (4.37)
0 R2
with R1 > 0, R2 > 0, resulting in the following:
λ∗2
âr = ar∗ = − (4.38)
2R1
and
λ∗3 r ∗
âθ = aθ∗ = − (4.39)
2R2
4.4.2 Plane Change Maneuver
When changing the orbital plane without affecting the shape of the orbit, the normal
acceleration, an , is the only control input, whereas the angular deviation, α, from the
osculating orbit is the sole state variable. The formulation is greatly simplified owing
to the fact that neither r nor ṙ are changed by the control, and evolve naturally by
4.4 Optimal Low-Thrust Orbital Transfer 123
Keplerian motion. The orbital angular momentum magnitude, h, remains constant

since aθ = 0. Thus we have the following scalar state equation:
r
α̇ = an (4.40)
h
which results in the following Hamiltonian:
ran
H = Ran2 + λ (4.41)
h
The costate equation is now the following:
λ̇∗ = 0 (4.42)
which implies that λ∗ is constant. The stationary condition produces the following
optimal control input:
λ∗ r ∗
ân = an∗ = − (4.43)
2h∗ R
where R > 0. A practical choice for R is the following:
r(tf )∗
R= (4.44)
2h∗
where r ∗ (tf ) is known from the Kepler’s equation for a specific value of terminal
time, tf . Thus we have
λ∗ r ∗ (t)
ân = − (4.45)
r(tf )∗
The value of λ∗ can be chosen based upon the maximum available acceleration
magnitude.
4.4.3 General Orbital Transfer
A more general optimal transfer is for the minimization of

tf
J = φ[x(tf ), tf ] + [ L[x(t), a(t)]dt (4.46)
ti
such that the terminal state, x(tf ), lies on a hypersurface defined by
F[x(tf ), tf ] = 0 (4.47)
This results in the change in the terminal boundary conditions to the following
transversality condition on the costate vector (Chap. 2):
T
∂Φ
λ∗ (tf ) = (4.48)
∂x t=tf
where
Φ = φ + cT F (4.49)
with c = (c1 , c2 , . . . , ck )T being arbitrary constants. The terminal time, tf , must

satisfy

∂H
Φ+ =0 (4.50)
∂t t=tf
Since neither the costate vector, λ, nor the coefficients, c, have any physical
significance, these boundary conditions can be satisfied in a variety of ways,
which can cause difficulty in their implementation. The transversality conditions
are common when a transfer is to be made either to a specific final orbit or to a
specific relative position in a constellation of spacecraft (formation flying).
4.5 Variational Model
The problem of Keplerian motion perturbed by a continuous, small acceleration,

u, caused by the thrust inputs and external disturbances, can be modeled as
a continuous, slow deviation of the trajectory from a two-body orbit. Such a
variational model is developed by approximating the orbital elements by first-order
time variations, while the others are averaged over an osculating orbit.
The Gauss’ form of Lagrange’s planetary model subject to a perturbing accelera-
tion, u = (ar , aθ , an )T , is expressed as the following set of variational equations [8]
in terms of the classical orbital elements, x = (a, e, i, Ω, ω, θ )T :
da 2a 2 v
= aθ (4.51)
dt μ
de 1 r
= [2(e + cos θ )aθ + ar sin θ ] (4.52)
dt v a
di r
= cos(θ + ω)an (4.53)
dt h
dΩ r
= sin(θ + ω)an (4.54)
dt h sin i
dω 1 a r
= 2aθ sin θ − 2e + cos θ ar − an sin(θ + ω) cos i (4.55)
dt ev r h sin i
4.5 Variational Model 125
dθ h 1 a
= 2− 2aθ sin θ − 2e + cos θ ar (4.56)
dt r ev r
where θ is the true anomaly obtained from solving the Kepler’s equation for the
osculating orbit from the time of periapsis, τ , to the present time, t, and (r, v, h)
are the instantaneous radius, speed, and angular momentum of the osculating orbit.
Equations (4.51)–(4.56) can be alternatively expressed in terms of the variation with
. anomaly, E, instead of time, by dividing them by dE/dt =
respect to the eccentric
na/r, where n = μ/a 3 is the orbital mean motion. An averaging procedure is
utilized to determine each element’s mean rate of change by calculating its variation
over a single revolution, and then dividing by the orbital period, T = 2π/n. The
averaged state vector is thus computed as follows:
E2 r dx
1
x̄ = dE (4.57)
T E1 na dt
where dx/dt is calculated from Eqs. (4.51)–(4.56) with the terms on the right-
hand side fixed at their mean values over a complete revolution. The integration
limits of the averaging integral, E1 , E2 = E1 + 2π , are traditionally selected based
upon astronomical observation issues, such as the points of entering and exiting the
Earth’s shadow.
Most of the low-thrust transfer problems are posed as time-optimal problems,
because with a nearly constant mass exhaust rate of rocket engines, the minimum
time also results in the minimum fuel consumption. Furthermore, a practical transfer
between two orbits requires a terminal condition, Eq. (4.47), in which the final shape
(a, e) and orbital inclination, i, are specified. For a time-optimal transfer involving
only the changes in (a, e, i), the Hamiltonian of the variational model is given by
da de di
H = 1 + λT ẋ = 1 + λa + λe + λi (4.58)
dt dt dt
Since the Hamiltonian is linear in control variables, (ar , aθ , an ), the time-optimal

problem is singular in terms of the acceleration input, u. Its solution requires either
imposing constraints on the magnitude of input acceleration, u, and solving the
singular control problem using the method of Chap. 5 or by introducing new control
variables which make the Hamiltonian non-singular. In order to do the latter, we
resolve the thrust acceleration in spherical coordinates using the in-plane pitch
angle, α, and out-of-plane yaw angle, β, as follows:
u = u(sin α cos βir + cos α cos βiθ + sin βih ) (4.59)
where ir = r/r and ih = h/ h = ir × iθ (withh = r × v). This gives

ar = a sin α cos β, aθ = a cos α cos β, and an = a sin β. Substitution of these
components into Eqs. (4.51)–(4.56), and treating the thrust angles, (α, β), as the
new control variables, produces the following stationary conditions:
∂H 2a 2 v
= −uλ∗a sin α ∗ cos β ∗
∂α μ
u r
+ λ∗e [−2(e + cos θ ) sin α ∗ + sin θ cos α ∗ ] cos β ∗ = 0 (4.60)
v a
∂H 2a 2 v u r
= −uλ∗a cos α ∗ sin β ∗ −λ∗e [2(e + cos θ ) cos α ∗ + sin θ sin α ∗ ] sin β ∗
∂β μ v a
r
+ uλ∗i cos(θ +ω) cos β ∗ = 0 (4.61)
h
The quadrant ambiguity in the optimal pitch and yaw angles is avoided by ensuring
Hαα > 0 and Hββ > 0, which yields the following optimal angles:
−λe ar sin θ
sin α̂ = -
2 2 2 2
4 λ∗a a μv + λ∗e (e + cos θ ) + (λ∗e )2 ar 2 sin2 θ
2 2
−2 λ∗a a μv + λ∗e (e + cos θ )
cos α̂ = - (4.62)
2 2 2 2
4 λ∗a a μv + λe (e + cos θ ) + (λ∗e )2 ar 2 2
sin θ
λ∗i rv
h cos(θ + ω)
sin β̂ = −
Δ
2 2 $ %
2λ∗a a μv cos α ∗ + λ∗e 2(e + cos θ ) cos α ∗ + r
a sin θ sin α ∗
cos β̂ = − (4.63)
Δ
where

2
r 2 v2 a2 v2 r
Δ = (λ∗i )2 cos 2 (θ + ω) + 2λ∗
a cos α ∗ + λ∗ 2(e + cos θ) cos α ∗ +
e sin θ sin α ∗
h2 μ a
The costate equations to be solved for the optimal steering laws of Eqs. (4.62) and
(4.63) are the following:
∂H av r
λ̇∗a = − = −4uλ∗a cos α ∗ cos β ∗ + λ∗e 2 sin θ sin α ∗ cos β ∗ (4.64)
∂a μ va
∂H 2
λ̇∗e = − = −λ∗e cos α ∗ cos β ∗ (4.65)
∂e v
∂H
λ̇∗i = − =0 (4.66)
∂i
4.6 Optimal Regulation of Circular Orbits 127
the last of which yields λ∗i (t) = λi (tf ) = const. The other two costate
variables must be determined from the transversality conditions given by Eqs. (4.48)
and (4.50). The resulting 2PBVP is to be solved numerically.
Edelbaum [28] derived the low-thrust optimal profiles for transfers between
circular orbits of different inclinations with Gauss perturbation equations, using
ar = 0, and a constant thrust acceleration magnitude, u. The optimal paths √ with
a low-thrust acceleration are nearly circular, i.e., e 0, r a, and v = μ/r.
Putting these conditions into Eqs. (4.62) and (4.63), along with Edelbaum’s arbitrary
selection of λa = 1/a, reduces them to the following time-optimal steering laws for
a nearly circular transfer between inclined circular orbits [28]:
λ∗e sin θ
tan α ∗ = (4.67)
2(1 + λ∗e cos θ )
λ∗i cos(θ + ω)
tan β ∗ = (4.68)
2
Chobotov [23] presents a Mayer formulation (see Chap. 2) using the equinoctial
elements, which avoid the singularities of the classical elements for e = 0, i =
0, π/2 (but are singular for i = π instead).
4.6 Optimal Regulation of Circular Orbits
Most of the maneuvering tasks in the life of a satellite involve the regulation
of radial, in-track, and out-of-the plane position around a desired circular orbit.
Here we consider the optimal station keeping maneuvers where the initial errors
in position and velocity relative to a point in the circular orbit must be driven
to zero with the minimum expenditure of the propellant. Let the constant radius
and frequency of the desired orbit be r0 and n, respectively. The state variables
of this problem can be chosen to be the deviation from the radius, δr = r − r0 ,
the radial velocity, ṙ, the in-track angular deviation, δθ , and out-of-plane angular
displacement, α (both measured from a given point in the orbit), and the orbital
angular momentum, h. The state equations can then be expressed as follows:
δ ṙ = ṙ
h2 μ
r̈ = − + ar
(r0 + δr)3 (r0 + δr)2
h
δ θ̇ = (4.69)
(r0 + δr)2
ḣ = (r0 + δr)aθ
(r0 + δr)
α̇ = an
h
These nonlinear state equations couple the coplanar and out-of-the plane motion,
whose solution requires numerical solution. However, a decoupling of the two types
of maneuvers is possible by applying either coplanar or normal acceleration inputs,
which simplifies their control. Furthermore, nearly all practical maneuvers with low
thrust require only small displacements from a reference point in the given circular
orbit, which allows a linearization of the governing equations.
4.6.1 Coplanar Regulation with Radial Thrust
Regulating the orbital position is possible by radial acceleration input, ar , alone,

which allows the angular momentum, h, to remain unchanged and simplifies the
state equations, derived from Eq. (4.69) to be the following:
δ ṙ = ṙ
h2 μ
r̈ = − + ar (4.70)
(r0 + δr)3 (r0 + δr)2
h
δ θ̇ =
(r0 + δr)2
√
where the angular momentum of the desired orbit, h = μr0 , is constant. Since
the positional errors are usually small in comparison with the orbital radius, it is
reasonable to linearize the state equations around a reference point in the circular
orbit through the binomial approximations:

kδr
(r0 + δr) −k
r0−k 1− , (k = 2, 3) (4.71)
r0
Thus the following linearized state equations are derived by neglecting the second
and higher-order terms involving δr and r0 δθ :
δ ṙ = ṙ
0 1
3h2 2μ
r̈ = − − 3 δr + ar (4.72)
r04 r0
2h
δ θ̇ = −
r03
By re-defining the state vector to be the following:
x = (δr, ṙ, r0 δθ )T (4.73)

,
and noting that h/r02 = μ/r03 = n and
3h2 2μ μ
4
− 3 = 3 = n2 (4.74)
r0 r0 r0
the linearized state equations are expressed as follows:
ẋ = Ax + Bar (4.75)
where
⎛ ⎞ ⎛ ⎞
0 1 0 0
A = ⎝ −n2 0 0 ⎠ , B = ⎝1⎠ (4.76)
−2n 0 0 0
Example 4.1 An orbital regulation system is to be designed for a spacecraft in a

circular orbit of radius r0 and frequency n around a planet, for maintaining the
circular shape of the orbit with a continuous, radial acceleration control input,
ar (t), in the presence of an initial perturbation, δ ṙ(0). The initial radial and in-track
deviations, δr(0), δθ (0), are zeros. Use the linearized state equations to determine
the extremal trajectory, δr ∗ (t), and the extremal control history for minimizing the
following performance index:

1 1 tf 2
J = [δ ṙ(tf )]2 + a (t)dt .
2 2 0 r
where tf is specified to be the orbital period. The final deviations, δr(tf ), δ ṙ(tf ),
and δθ (tf ), are unspecified.
The Hamiltonian of this problem is given by
H = L + λT (Ax + Bar ) (4.77)
where L = 12 ar2 , x and (A, B) are given by Eqs. (4.73) and (4.76), respectively,
and λ = (λr , λṙ , λθ )T is the costate vector. The terminal cost function is φ =
2 [δ ṙ(tf )] . The costate equations are thus the following:
1 2
∗T
∗ ∂H
λ̇ = − = −AT λ∗ (4.78)
∂x
Since the costate equations are linear and homogeneous, their solution to the initial
condition, λ(0) = (c1 , c2 , c3 )T , can be expressed as follows:
⎛ ⎞⎧ ⎫
cos nt n sin nt 2 sin nt ⎨ c1 ⎬
λ∗ (t) = e −AT t
λ(0) = ⎝ − n1 sin nt cos nt 2
(cos nt − 1) ⎠ c2 (4.79)
n ⎩ ⎭
0 0 1 c3
The transverse boundary condition for this case is the following:

⎧ ⎫
∗T ⎨
0 ⎬
∂φ
λ∗ (tf ) = = δ ṙ(tf ) (4.80)
∂x ⎩ ⎭
0
Since we have tf = 2π/n, this condition results in λ∗r (tf ) = c1 = 0 and λ∗θ (tf ) =
c3 = 0, therefore, the solution for the costate vector on the extremal trajectory is
given by
⎧ ⎫
⎨ n sin nt ⎬
λ∗ (t) = c2 cos nt (4.81)
⎩ ⎭
0
The stationary condition, Har = 0, produces ar∗ (t) = −λ∗ṙ (t) = −c2 cos nt, and
the state solution is derived to be the following:
t
x∗ (t) = eAt x(0) + eA(t−τ ) Bar∗ (τ )dτ
0
⎧ $ % ⎫
⎪
⎨
1
n δ ṙ(0) −c2 2t sin nt ⎪
⎬
c2
= δ ṙ(0) cos nt − sin nt
+ t cos nt (4.82)
⎪
⎩
2 n ⎪
⎭
0
Finally, the application of the terminal boundary condition gives

c2 sin ntf
λ∗ṙ (tf ) = c2 = δ ṙ(tf ) = δ ṙ(0) cos ntf − + tf cos ntf
2 n
π
= δ ṙ(0) − c2 (4.83)
n
or c2 = δ ṙ(0)/(1 + πn ). Hence, the extremal control history is
δ ṙ(0)
ar∗ (t) = − cos nt , (0 ≤ t ≤ 2π/n) (4.84)
1 + πn
and the extremal trajectory is the following:

δ ṙ(0) t
δr ∗ (t) = 1+ sin nt , (0 ≤ t ≤ 2π/n) (4.85)
n 1 + πn
The extremal control, ar∗ (t), is also optimal because Har ar = 1 > 0.
4.6.2 Coplanar Regulation with Tangential Thrust
When only a tangential thrust acceleration, aθ , is applied to the spacecraft the radial
acceleration is applied only indirectly through a continuous change of the orbital
angular momentum, h. The state equations are then expressed as follows:
δ ṙ = ṙ
h2 μ
r̈ = −
(r0 + δr)3 (r0 + δr)2
h
δ θ̇ = (4.86)
(r0 + δr)2
ḣ = (r0 + δr)aθ
When a linearization is carried out based upon the assumption of small accel-
eration input and small displacements, the angular momentum can be expressed as
follows:
h = r 2 θ̇ = (r0 + δr)2 (n + δ θ̇ ) r02 n + 2nr0 δr + r02 δ θ̇ (4.87)
whose substitution into Eq. (4.86) results in the following approximate model:
δ ṙ = ṙ
r̈ = −n2 δr (4.88)
r0 δ θ̈ = −2nr0 δ ṙ + r0 aθ
By re-defining the state vector to be the following:
x = (δr, ṙ, r0 δθ, r0 δ θ̇ )T (4.89)
we express the linearized state equations as follows:
ẋ = Ax + Bu (4.90)
where
⎛ ⎞ ⎛ ⎞
0 1 0 0 0
⎜ −n2 0 0 0⎟ ⎜0⎟
A=⎜
⎝ 0
⎟ , B=⎜ ⎟ (4.91)
0 0 1⎠ ⎝0⎠
0 −2nr0 0 r0
4.7 General Orbital Tracking
Many space missions require a spacecraft to track a predetermined (nominal)

trajectory in the presence of disturbing forces. This could either be a Keplerian orbit
in an inertial frame (Chap. 3) or a multi-body trajectory in a rotating frame (Chap. 6).
With the use of the appropriate motion coordinates, it is possible to linearize the
governing equations close to the given trajectory, and then apply control inputs to
track the latter in the presence of nonlinear perturbations. Here we shall consider a
Keplerian nominal orbit for tracking, which can be extended to a more general case,
such as a halo orbit around the collinear Lagrangian point of a three-body problem
(Chap. 6).
The equation governing perturbed two-body motion can be expressed as follows:
ṙ = v
μ
v̇ + 3 r = u + ad (4.92)
r
where ad (t) represents a disturbing acceleration, and u(t) is the required corrective
control input. Unless modeled by appropriate physical processes, ad (t) is treated
as a stochastic disturbance called the process noise, ν(t). Furthermore, the mea-
surement of the position and velocity vectors is subject to a random disturbance
called the measurement noise, w(t). The nominal orbit being described by Keplerian
motion is modeled by
ṙn = vn
μ
v̇n + 3 rn = 0 (4.93)
rn
This results in a first-order Taylor series expansion of the position and velocity
vectors about their nominal values as follows:
r(t) rn (t) + δr(t)

v(t) vn (t) + δv(t) (4.94)
which is a valid approximation as long as

μ μ
| ad | ; | u |
r2 r2
Writing the spherical gravitational acceleration as follows:

μ μ ∂g μ
g = − 3 r = − 3 rn + δr = − rn + δg (4.95)
r rn ∂r r=rn rn 3
or,
δg = Gn δr (4.96)
4.7 General Orbital Tracking 133
where Gn is the gravity-gradient matrix evaluated on the nominal trajectory:

∂g
Gn =
∂r r=rn
Similarly, the perturbing acceleration can be linearized about the nominal trajectory
as follows:
ad = Pn δr + Qn δv (4.97)
where

∂ad
Pn =
∂r r=rn
and

∂ad
Qn =
∂v v=vn
The design of a tracking system is thus based upon the following linearized plant
derived above with the small perturbation approximation:
ṙ = δv
v̇ = (Gn + Pn ) δr + Qn δv + u (4.98)
which can be expressed in the following state-space form:
ẋ(t) = A(t)x(t) + B(t)u(t) x(0) = x0 (4.99)
where

δr
x=
δv

0 I 0
A= ; B=
(Gn + Pn ) Qn I
If the disturbance coefficient matrices Pn , Qn cannot be determined by a physical

model, one has to include their effects by the stochastic process and measurement
noise vectors, ν(t) and w(t), respectively. In such a case, the state vector is not
determined directly, but instead estimated from the measurement of (noisy) output
variables using an optimal observer (Kalman filter) [79]. Hence, the stochastic
tracking plant is expressed by
ẋ(t) = A(t)x(t) + B(t)u(t) + F(t)ν(t) x(0) = x0 (4.100)

where

0 I 0
A= ; B=
Gn 0 I
along with the output equation,
y(t) = C(t)x(t) + D(t)u(t) + w(t) (4.101)
We note that the state transition matrix, (t, t0 ), corresponding to the homoge-
neous part of the state equation, Eq. (4.100), is a symplectic matrix (see Chap. 2),
because the gravity-gradient matrix, Gn (t), is always symmetric. Thus the state
transition matrix of the nominal plant [ν(t) = 0, w(t) = 0] is expressed as follows:

R̃(t) R(t)
(t, t0 ) = (4.102)
Ṽ(t) V(t)
with the property

0 1
−1 VT (t) −RT (t)
(t, t0 ) = (t0 , t) = T T (4.103)
−Ṽ (t) R̃ (t)
Here the partition matrices, R, R̃, V, Ṽ, must obey Eq. (4.103), which implies that
the following matrices are symmetric:
T T T
RR̃ , VṼ , R̃ Ṽ, RT V
−1
VR−1 , R−1 R̃, V−1 Ṽ, ṼR̃
−1 −1 −1
RV−1 , R̃ R, Ṽ V, R̃Ṽ
Furthermore, the following identity must be satisfied:
T
VR̃ − ṼRT = I (4.104)
and the partition matrices should also satisfy the governing, homogeneous equa-
tions, and the initial conditions, that is,
˙ = Ṽ
R̃
˙ = G R̃
Ṽ n
R̃(t0 ) = I (4.105)
Ṽ(t0 ) = 0
and
Ṙ = V
V̇ = Gn R
R(t0 ) = 0 (4.106)
V(t0 ) = I
Battin [8] presents an algorithm for computing the state transition matrix based upon
Lagrange’s coefficients and universal functions which require series (or continued
fraction) evaluations. For practical guidance problems, it appears much simpler
to solve the nonlinear 2PBVP by a Lambert algorithm, rather than determine a
linearized solution in this manner.
The general planar orbit tracking requires the solution of a linear, time varying
regulator problem. Since the nominal trajectory is unenforced (that is, un = 0), there
is no need for a feedforward control [81]. Linear optimal control (Chap. 2) is best
suited to this problem. Such a controller can be later modified to include an observer
to account for stochastic disturbances, ν(t), w(t), by a robust, multivariable control
strategy, such as either the linear, quadratic Gaussian (LQG) or H∞ method [79].
Since the disturbance acceleration is unknown, it is not possible to make the
trajectory converge to the nominal at any given time. Instead, the best that can
be achieved is to drive the state vector, x(t), asymptotically to a zero steady state
(t → ∞) by an infinite-time, linear, quadratic regulator (LQR) with the following
cost function:

1 ∞ T
J∞ (t) = x (τ )Q(τ )x(τ ) + uT (τ )R(τ )u(τ ) dτ (4.107)
2 t
subject to the state constraint equation
ẋ(t) = A(t)x(t) + B(t)u(t) x(0) = x0 (4.108)
and the linear, full-state feedback control law (see Chap. 2):
u(t) = −K∞ (t)x(t) (4.109)
Since the plant is slowly time varying, the stabilizing steady-state solution for the
feedback gain matrix at each time instant, K∞ (t), is given by
K∞ (t) = R−1 (t)BT (t)P(t) (4.110)
where P∞ (t) is the symmetric, positive semi-definite, quasi-steady solution of the

following algebraic Riccati equation:
P∞ (t)A(t) + AT (t)P∞ (t) − P∞ (t)B(t)R−1 (t)BT (t)P∞ (t) + Q(t) = 0 (4.111)

The quasi-steady approximation is tantamount to finding a steady-state solution to

Eq. (4.111) at each time instant, provided the variation of A, B, Q, R with time is
much slower than the settling time of all the transients of the closed-loop dynamics
matrix, A − BK.
For an observer-based application, a suitable Kalman filter can be designed for
a practical LQG implementation, with loop transfer recovery (LTR) at either the
plant’s input or its output [79].
Example 4.2 Consider, for example, a planar tracking problem, in which the
perturbations always lie in the plane of the orbit. The orbital tracking plant can
be expressed in a state-space form as follows:
ẋ(t) = A(t)x(t) + B(t)u(t) x(0) = x0 (4.112)
where
⎧ ⎫
⎨ δr(t) ⎬
x(t) = δ ṙ(t)
⎩ ⎭
δh(t)

ar
u(t) =
aθ
⎛ ⎞
0 1 0
⎜ 3h2n 2μ ⎟
A(t) = ⎝ − rn4 (t)
− rn3 (t)
0 2hn
rn3 (t) ⎠
0 0 0
and
⎛ ⎞
0 0
B(t) = ⎝ 1 0 ⎠
0 rn (t)
The stability of the linear, planar orbital plant is analyzed by finding the
eigenvalues of A as follows:

3h2n 2μ
|λI − A| = λ λ + 4
2
− (4.113)
rn (t) rn3 (t)
Clearly, for stability we must have
3h2n 2μ
4
> 3 (4.114)
rn (t) rn (t)
or,
2
h2n > μrn (t)
3
Since there will be points on any given orbit where the stability condition will
be unsatisfied, we need a feedback controller for stability augmentation at such
points. Even for those points where the stability condition is satisfied, the plant is
not asymptotically stable, thereby requiring a feedback controller to provide active
damping for bringing the orbital deviations to zero in the steady state.
Since the plant is unconditionally controllable, i.e., the controllability test
matrix [79],
⎡ 2hn
⎤
0 0 1 rn3 (t)
0 0
⎢ ⎥
⎢1 0 0 2μ
−
3h2n 4μhn
−
6h3n ⎥
⎣ 0 rn3 (t) rn4 (t) rn6 (t) rn7 (t) ⎦
0 rn (t) 0 0 0 0
is of rank three (the order of plant) at all times, we can satisfy the sufficient
conditions for the existence of a unique, positive semi-definite solution to the
algebraic Riccati equation by selecting suitable state and control cost coefficient
matrices, Q(t), R(t). A possible choice is the following constant pair:
⎛ ⎞
Q1 0 0
R1 0
Q=⎝ 0 Q2 0 ⎠ ; R=
0 R2
0 0 Q3
where Qi ≥ 0, Ri > 0 are selected based upon the relative importance of the state
and control variables.
The algebraic Riccati equation, Eq. (4.111), must be solved in real time with a
changing set of state-space coefficient matrices. Because such a controller is rarely
implementable, a constant feedback gain matrix based on a specific time instant
is selected, and applied throughout the control interval. Another approach would
be to use the averaged orbital parameters of a perturbation model over a complete
orbit. A linear regulator based upon either a time-frozen or time-averaged plant
dynamics, Ā, Ā, can be designed to be stabilizing because of the slowly varying
plant dynamics. In a real application, a control interval of a few hundred seconds is
quite small compared to the orbital period of a satellite. Therefore, we modify the
control law by taking a constant gain matrix,
−1 T
K̄∞ = R̄ B̄ P̄∞ (4.115)
where P̄∞ is the constant solution to

T −1 T
P̄∞ Ā + Ā P̄∞ − P̄∞ B̄R̄ B̄ P̄∞ + Q̄ = 0 (4.116)
Referring to the orbital plant dynamics, the constant gain approximation consists of
neglecting the variation in the nominal orbital radius with time, rn (t) ≈ rn (0).
4.8 Basic Guidance with Continuous Inputs
Either positioning the spacecraft relative to a given point in an orbit or changing

its position and velocity to achieve a new orbit, both require an ability to move the
vehicle relative to a desired terminal state. Hence the term guidance refers to the
process of changing the flight path in order to meet desired terminal conditions.
Guiding a space vehicle involves modifying the translational motion of its center
of mass (i.e., orbital mechanics) by the application of acceleration inputs. These
inputs can be applied either impulsively (as seen in Chap. 3) or continuously.
In a spacecraft—as in any other vehicle—there are two broad ways of applying
continuous acceleration inputs for guidance purposes: (a) varying the flight direction
via acceleration inputs applied normal to the flight path, and (b) changing the kinetic
energy via tangential inputs. The adjectives “tangential” and “normal” refer to the
instantaneous flight direction, which could be changing continuously. Method (a)
is referred to as maneuvering (or steering) whereas method (b) is called energy
control. Each of the two methods require a guidance logic whose derivation can be
carried out by optimal control theory. While it is possible to devise multi-variable
control laws for performing both maneuvering and energy control simultaneously,
it is much more intuitive to develop guidance techniques for them separately and
analytically. Being much simpler than a numerical 2PBVP solution, basic guidance
methods have been practically used in even the most challenging missions, such as
guiding the landers to the surface of the various planets and the moons, beginning
with the Apollo program, as well as in making orbital rendezvous between two (or
more) spacecraft.
Guidance for changing the flight direction is referred to as the maneuvering
guidance. Here the application of normal acceleration control inputs causes a
variation in the flight path, which in turn results in a new orbit. The most common
example of maneuvering guidance is for achieving a desired final position from
a given current position and velocity, such as in a terminal interception problem.
Since the terminal velocity is an unspecified target, the flight energy is not a part of
maneuvering, and can be controlled separately as discussed in the previous section.
The dependence of the optimal control input on the ever changing current
position makes the maneuvering guidance a time-dependent (or non-autonomous)
system. Maneuvering guidance is aimed at intercepting a specific target, which can
either be another spacecraft or a specified set of position and velocity coordinates
defining a moving point in space. Such an interception requires a continuous
reduction in the distance between the target and the maneuvering spacecraft. The
acceleration control input is applied normal to the relative velocity vector such
that the spacecraft closes in on the target at any given instant, often maintaining
a constant relative speed. Let us begin the treatment of guidance laws with a simple
4.9 Line-of-Sight Guidance 139
motivating method. Although the line-of-sight navigation was originally developed

for air-to-air missiles, it can be applied to spacecraft maneuvers around an intended
target.
4.9 Line-of-Sight Guidance
Consider a target moving with a position rt (t) and velocity vt (t). A maneuvering
spacecraft is required to intercept the target by applying an acceleration input
u(t). Both the target and the interceptor are assumed to be sufficiently close to
disregard the differences in the acceleration due to gravity between them. This is
possible when the two points are roughly in the same orbit but separated by a small
distance, and have a small relative velocity. The instantaneous position, rm (t), and
velocity, vm (t), of the spacecraft are thus changed such that the relative position,
r(t) = rm (t) − rt (t), decreases in magnitude r(t) =| r(t) | at every instant. Without
knowing the instantaneous velocity and acceleration of the target, the simplest way
in which an interception can be ensured is by applying the control u(t) normal to
the relative velocity vector, v(t) = vm (t) − vt (t), such that component of v(t)
in the direction of r(t) is always negative. Since r(t) is along the line of sight
from the maneuvering spacecraft to the target, such a scheme is referred to as line-
of-sight interception. This situation is depicted in Fig. 4.2 where the right-handed
coordinate frame attached to the target, iθ , ir , in , with in = iθ × ir , is used to resolve
the relative position and velocity, with angles θ (t) and φ(t) denoting the inertial
directions of r(t) and v(t), respectively. The line-of-sight direction, given by ir (t),
can be determined either by an optical or radar sensor on board the maneuvering
spacecraft or from an inertial navigation computer if the target is not a physical
object.
Fig. 4.2 The geometry of vm

line-of-sight interception
v
f
Inertial
reference
q
u
r
ir
iq
vt
The angle (φ + θ ) is subtended by the relative velocity v with the line of sight,
ir (t), in either the second or the third quadrant, otherwise the closure rate ṙ would
be non-negative. Therefore, the relative velocity can be expressed as follows:
v = v [ir cos(π − φ − θ ) + iθ sin(π − φ − θ )]

= v [−ir cos(φ + θ ) + iθ sin(φ + θ )] (4.117)
The kinematic relationship between r and v is the following:

dr
v= = ṙir + (−θ̇ in ) × r
dt
= ṙir + r θ̇iθ (4.118)
On comparing Eqs. (4.117) and (4.118), we have the following kinematic scalar
equations of motion:
ṙ = −v cos(φ + θ ) (4.119)
and
v
θ̇ = sin(φ + θ ) (4.120)
r
For the kinetic equation of motion, we recall that the target and the interceptor are
roughly in the same orbit and separated by a small distance compared to the orbital
radius, which implies that the difference in their accelerations due to gravity can be
neglected. We will consider a more accurate model later in the chapter where the
difference in gravity is also accounted for. Thus the relative acceleration is supplied
by the control acceleration, u, applied to the interceptor in a direction normal to v.
Hence we write
dv
u= = v̇ [−ir cos(φ + θ ) + iθ sin(φ + θ )]
dt
+ v(φ̇ + θ̇) [ir sin(φ + θ ) + iθ cos(φ + θ )]
$ %
+ v − −θ̇ in × ir cos(φ + θ ) + −θ̇ in × iθ sin(φ + θ ) (4.121)
or, noting that in × ir = −iθ and in × iθ = ir , we have
u = v̇ [−ir cos(φ + θ ) + iθ sin(φ + θ )]

+ v φ̇ [ir sin(φ + θ ) + iθ cos(φ + θ )] (4.122)
Since u is normal to v, we also have the following:

v
u = u in ×
v
= u [ir sin(φ + θ ) + iθ cos(φ + θ )] (4.123)
Therefore, on comparing Eqs. (4.122) and (4.123), it follows that

u
φ̇ = (4.124)
v
and
v̇ = 0 (4.125)
This implies that the relative speed v always remains a constant, which is not
surprising since the input acceleration is normal to the relative velocity vector. A
substitution of Eq. (4.120) into Eq. (4.124) gives the following rate of change of the
net angle between ir and v:
u v
φ̇ + θ̇ = + sin(φ + θ ) (4.126)
v r
The equations governing the relative motion are thus Eqs. (4.119) and (4.126) where
the relative speed, v, is constant. The acceleration normal to the projected line of
sight is given by Eq. (4.124) to be the following:
v2
z̈ = v(φ̇ + θ̇ ) = u + sin(φ + θ ) (4.127)
r
Equation (4.127) can be integrated twice to yield the projected miss distance, z(t),
defined to be the lateral distance of closest approach between the target and the
interceptor if the acceleration input, u, remains unchanged. Referring to Fig. 4.2,
the projected miss distance is given by
z = r sin(φ + θ ) (4.128)
whose time derivative is the following:
ż = ṙ sin(φ + θ ) + r(φ̇ + θ̇ ) cos(φ + θ ) (4.129)
or, by Eqs. (4.119) and (4.126), we have

ur
ż = −v sin(φ + θ ) cos(φ + θ ) + cos(φ + θ ) + v sin(φ + θ ) cos(φ + θ )
v
ur
= cos(φ + θ ) (4.130)
v
For a successful interception, it must be ensured that z(t) → 0 as t → tf . A
simple navigation law for accomplishing this is derived by treating the net angle,
(φ + θ ), to be small at all times, resulting in the following approximations:
z r(φ + θ ) (4.131)
v zv
θ̇ (φ + θ ) = 2 (4.132)
r r
ur
ż (4.133)
v
and
ṙ −v (4.134)
It follows from Eq. (4.134) that
r = rf − vt (4.135)
or,
r
f
ż = u −t (4.136)
v
If we regard the time, tf = rf /v, as the projected terminal time at which the
interception takes place, that is z(tf ) = 0, then the control u(t) is applied in the
remaining time, i.e., the time to go, (tf − t), before the projected interception. When
Eq. (4.135) is substituted into Eqs. (4.132) and (4.133), the following linearized state
equations are obtained for the line-of-sight navigation problem:
z
θ̇ = (4.137)
v(tf − t)2
and
ż = u(tf − t) (4.138)
These can serve as the basis of designing an optimal line-of-sight navigation system.
Equation (4.138) shows that the normal velocity, ż, vanishes at the terminal time,
unless u(tf ) → ∞. Furthermore, if u(t) were to remain constant throughout the
time to go, the integration of Eq. (4.138) results in the following:

t
z = ut tf − (4.139)
2
which implies that the actual miss distance, z(tf ) = utf2 /2 = 0, could be large.
Clearly, the acceleration input, u(t), must not remain constant, but should evolve in
time in order to make z(tf ) = 0.
Example 4.3 Consider a spacecraft approaching a space station for docking in a
circular orbit. When the two spacecraft are in the same orbit and separated by a small
lateral distance, x(t), the spacecraft is guided by line-of-sight navigation, resulting

in the following equation of relative motion:

ẋ = tf − t u , (4.140)
where u(t) is the normal acceleration input applied to the maneuvering spacecraft.
It is assumed that the two spacecraft are traveling in approximately parallel straight
lines with a constant relative speed called closure rate, which will bring them exactly
abreast in time tf . A successful docking thus requires the final lateral separation,
x(tf ), must vanish while requiring only small transient input magnitudes, u(t), t ≤
tf , both of which are satisfied by minimizing the following objective function:
tf
J = φ[x(tf ), tf ] + L(x, u)dτ (4.141)
t
where the terminal penalty function
φ[x(tf ), tf ] = k 2 x 2 (tf ) , (4.142)
ensures that x(tf ) is driven to small values by appropriate selection of the cost
parameter, k, and the minimization of the transient cost with Lagrangian,
L = u2 (t) , (4.143)
begins at the current time, t ≤ tf . The Hamiltonian function for this scalar state
variable, x(t), is thus the following:
H = L(x, u) + λT f (x, u) = u2 + (tf − t)λu , (4.144)
with the costate equation

T
∗ ∂H ∗
λ̇ (t) = − =0 (4.145)
∂x
with boundary condition
λ∗ (tf ) = 2k 2 x ∗ (tf ) (4.146)
The stationary condition for optimization,

∗ T
∂H
Hu = = 2u∗ (t) + (tf − t)λ∗ (t) = 0 (4.147)
∂u
gives the following expression for the extremal control:
1
u∗ (t) = − (tf − t)λ∗ (t) = −k 2 (tf − t)x ∗ (tf ) (4.148)
2
Thus we have derived a simple boundary-value problem for determining the current
extremal state from the projected terminal state, x(tf ). An extremal solution to
this problem is easily obtained by substituting Eq. (4.148) into Eq. (4.140) and
integrating back in time from tf to t:
t t
x ∗ (t) − x ∗ (tf ) = (tf − τ )u∗ (τ )dτ = −k 2 x ∗ (tf ) (tf − τ )2 dτ (4.149)
tf tf
or,
x ∗ (t)
x ∗ (tf ) = (4.150)
k2
1+ 3 (tf − t)3
which substituted into Eq. (4.148) results in the following feedback control law:
k 2 (tf − t)
u∗ (t) = − x ∗ (t) , (t ≤ tf ) (4.151)
k2
1+ 3 (tf − t)3
This control law is optimal for any real value of k, hence û(t) = u∗ (t). The
requirement for an exactly zero terminal separation for any given value of tf is
satisfied in the limit k → ∞, for which the optimal control law becomes the
following:
3
lim u∗ (t) = − x ∗ (t) , (t ≤ tf ) (4.152)
k→∞ (tf − t)2
The simple guidance problem presented in Example 4.2 could be solved in

a closed form. However, a general three-dimensional navigation problem is not
amenable to a closed-form solution. Even so, certain insights can be derived from
this simple example, which cannot be obtained from a numerical solution. These are
listed as follows:
(a) The optimal guidance problem for interception of a target in a given terminal
time, tf , with quadratic cost functions results in a linear feedback control law,
which relates the current optimal input, û(t), to the current state, x̂(t).
(b) The optimal control law depends upon the remaining time in the control interval,
(tf − t), called the time to go.
(c) The requirement of precise interception, i.e., for x̂(tf ) = 0, translates into a
control input which increases in magnitude as the time to-go becomes smaller,
reaching an infinite magnitude at t = tf .
4.10 Cross-Product Steering 145
4.10 Cross-Product Steering
The Apollo program was perhaps the most complex space mission conducted with
rather primitive computational resources. It could be successfully carried out only
due to the application of simple guidance techniques. One such method, called
cross-product steering, was used to align the velocity vector with an optimal
direction at various stages of a mission. This yielded both maneuvering and energy
control in a practical way. Its discussion here is along the lines of Battin [8]. The
method is based upon the ability to measure the craft’s velocity vector, v(t), and the
thrust acceleration vector, a(t), with some precision. Furthermore, a reference (or
desired) velocity vector, vR (t), is available from the navigation computer at every
instant. Consider vg (t) = vR (t) − v(t) as the velocity-to-be-gained at any given
time, t, when the spacecraft’s position vector measured in an inertial reference
frame is r(t). For an optimal interception of a final position, rf , in a future time
tf , the velocity-to-be-gained must vanish at t = tf , otherwise the fuel is wasted.
Hence, vg (tf ) is the objective function to be minimized with respect to a(t), which
is applied in the remaining time, tf − t. Normally, such an optimization would
require a sophisticated feedback scheme based upon the measurement of position
and velocity as well as the integration of the equations of motion. We demonstrate
that the same task can be achieved with a simple guidance law.
The required acceleration at any time t is given by
dvR ∂vR ∂vR dr

= +
dt ∂t ∂r dt
∂vR ∂vR
= + v (4.153)
∂t ∂r
∂vR ∂vR
= + (vR − v)
∂t ∂r
The total acceleration is the sum of the thrust acceleration and that due to gravity:
dv
= g(r) + a (4.154)
dt
where g(r) denotes the acceleration due to gravity produced by an axisymmetric

body. This yields the following time derivative of velocity to be gained:
dvg dvR
= − g(r) − a (4.155)
dt dt
whose substitution into Eq. (4.153) produces the guidance law,
dvR ∂vR
= g(r) − vg (4.156)
dt ∂r
or
dvg ∂vR
=− vg − a (4.157)
dt ∂r
The square matrix, ∂vR /∂r, is pre-computed and stored on board. Equa-
tion (4.157)provides the future thrust direction, a(t), which continues to be applied
until vg (tf ) = 0 at some final time, tf > t. The velocity-to-be-gained, vg (t), can
be driven to zero irrespective of the gravity field under consideration, by selecting
a(t) such that the following condition is satisfied:
dvg
× vg = 0 (4.158)
dt
which requires from Eq. (4.157) that
∂vR
a × vg = − vg × vg (4.159)
∂r
Hence, the velocity-to-be-gained is driven to zero without requiring any integration
of the governing differential equations. A post-cross product of Eq. (4.159) with
vg (t) results in the following:
(a · vg )vg − vg2 a = (p · vg )vg − vg2 p (4.160)
where
∂vR
p=− vg (4.161)
∂r
and vg =| vg |. From Eq. (4.160) it follows that

vg vg
a=p+ − ·p (4.162)
vg vg
where
2
vg vg
=a· = a 2 − p2 + ·p (4.163)
vg vg
Equation (4.162) provides the optimal thrust program as a simple algebraic guidance
logic, given the measurement of the thrust acceleration magnitude, a, by on-
board accelerometers, and the determination of the velocity-to-be-gained from
Eq. (4.158).
4.10 Cross-Product Steering 147
Example 4.4 To illustrate the cross-product steering, consider a problem with a

constant gravity field, g = const. This model—called the flat-trajectory approxi-
mation—is based upon a negligible variation in the radius vector from the vertical
direction, and is valid when the craft is very close to the surface of a spherical body
(such as during a lunar landing), or is traveling in a nearly circular orbit.
The equations of motion for a flat-trajectory model are the following:
dr
=v (4.164)
dt
dv
= g+a
dt
The solution for the unforced case (a = 0) with the initial conditions, r(0) = r0 ,
v(0) = v0 , is given by
1
r(t) = r0 + v0 t + gt 2 (4.165)
2
v(t) = v0 + gt
The velocity-to-be-gained at any time t is derived as follows by regarding the

current position, r(t), and the reference velocity, vR (t), as the initial conditions,
such that the desired final position, rf , is reached at a future time, tf :

1 1
vR (t) = rf − r(t) − g(tf − t) 2
(4.166)
tf − t 2
Thus we have
∂vR 1
=− I (4.167)
∂r tf − t
where I is the identity matrix. The substitution of Eqs. (4.166) and (4.167) into
Eq. (4.162) yields the following:
dvg 1
= vg − a (4.168)
dt tf − t
A guidance law based upon Eq. (4.168) is necessary for correcting the thrust
acceleration, a(t), in the time to go, tf − t, such that vg (tf ) = 0. Since the reference
velocity, vR (t), is derived for the unforced case, a successful application of thrust
should make vg = 0 at a time either before or at tf . For an optimal thrust program,
the engine must cut off at the time tc ≤ tf , exactly when the velocity-to-be-gained
becomes zero, otherwise the propellant is wasted. The vehicle travels on a coasting
trajectory (i.e., a = 0) in the remaining flight duration, tf − tc .
To derive the optimal thrust profile, take the scalar product of Eq. (4.168) with
vg , and obtain
1 d 2 1
(vg ) = v 2 − a · vg (4.169)
2 dt tf − t g
which is integrated by parts from the present time, t, to the engine cut-off time, tc ,
with vg (tc ) = 0, as follows:
tc
(tf − t)vg2 (t) = [2(tf − τ )a(τ ) · vg (τ ) − vg2 (τ )]dτ (4.170)
t
It is clear from Eq. (4.170) that the powered flight duration (called the remaining
burn time), tc − t, which governs the fuel expenditure, is minimized by maximizing
the scalar product, a · vg . This is achieved by aligning a(t) with the velocity-
to-be-gained, vg (t) at every instant, t ≤ tc , thereby making dvg /dt × vg = 0
(cross-product steering). This can be seen by putting a = c(t)vg , with c(t) being
an arbitrary function of time, into Eq. (4.168). Hence, the optimal solution is
achieved without numerically integrating Eq. (4.168). The velocity-to-be-gained
vector can be simply chosen as vg = (1, c1 , c2 )vgx (t), where c1 , c2 are constants to
be determined from the initial conditions.
4.11 Energy-Optimal Guidance
A practical but simple guidance law for driving the craft from the present initial
position, r(t), and velocity, v(t), to a specified final position, rf , and velocity, vf , at
a given time, tf , while minimizing the total control power in a known gravity field,
g(r, t), can be derived as an energy-optimal problem of minimizing the following
objective function:
tf
1
J (t) = uT (τ )u(τ )]dτ (4.171)
2 t
where u(t) = g(r, t)+a(t) is the net acceleration. The state equations of the problem
are given by
ṙ = v (4.172)
v̇ = u
with the following Hamiltonian:
1 T
H = u u + λTr v + λTv u (4.173)
2
4.11 Energy-Optimal Guidance 149
and the following costate equations:

∂H T
λ̇r = − =0 (4.174)
∂r

∂H T
λ̇v = − = −λr
∂v
whose solution can be expressed as
λr (t) = c1 (4.175)
λv (t) = −c1 t + c2
where c1 , c2 are constant vectors to be determined from the boundary conditions.

The stationary condition for optimality is given by
∂H
= u + λv = 0 (4.176)
∂u
or
û = −λv = c1 t − c2 (4.177)
which also satisfies the sufficient condition for optimality, Huu = I > 0. Thus,
the energy-optimal acceleration profile is linear, from which the thrust acceleration,
a(t), can be obtained by subtracting the known function, g(r, t).
Integration of the state equations beginning from the current position and
velocity, r(t), v(t), respectively, for τ = tf − t, yields the following:
1
v̂(τ ) = c1 τ 2 − c2 τ + v(t) (4.178)
2
and
1 1
r̂(τ ) = c1 τ 3 − c2 τ 2 + v(t)τ + r(t) (4.179)
6 2
from which the constants, c1 , c2 , are derived after applying the terminal boundary
condition as follows:
τ vf − v(t)
c2 = c1 − (4.180)
2 τ
1 τ
rf − r(t) = − c1 τ 3 + [vf − v(t)] + v(t)τ (4.181)
12 2

12 [vf + v(t)]
c1 = 3 r(t) − rf + τ (4.182)
τ 2

6 [vf + v(t)] vf − v(t)
c2 = 2 r(t) − rf + τ − (4.183)
τ 2 τ
The substitution of the constants, c1 , c2 , yields the following acceleration profile:
4$ % 6 $ %
û(τ ) = vf − v(t) + 2 r(t) − rf + τ v(t) (4.184)
τ τ
But we also have û(t) = c1 t − c2 , and û(t) = [vf − v(t)]/(tf − t), from which it
follows that
vf − v(t)
û(0) = −c2 = − (4.185)
τ
Substitution of Eq. (4.185) into Eq. (4.183) produces
τ
rf − r(t) = [vf + v(t)] (4.186)
2
or
rf − r(t) − vf τ = r(t) − rf + τ v(t) (4.187)
whose substitution into Eq. (4.184) yields the following expression for the optimal
acceleration profile:
4 $ % 6 $ %
û(t) = vf − v(t) + rf − r(t) − vf (tf − t) (4.188)
tf − t (tf − t)2
Example 4.5 To consider an example of the energy-optimal guidance, let us return

to the flat-trajectory approximation of constant gravity field. In such a case, we have
the following optimal thrust acceleration:
4 $ % 6 $ %
û = vf − v(t) + rf − r(t) + vf (tf − t) − g (4.189)
tf − t (tf − t) 2
In order to keep u(t) finite in the limit t → tf , the time to go, τ = tf − t, is

approximated by a constant, τ tgo , resulting in the following approximation:
& $ % $ %
4
tgo vf − v(t) + 6
2 rf − r(t) + vf tgo − g (t < tf − tgo )
û (tgo
0 (t ≥ tf − tgo )
(4.190)
4.12 Hill–Clohessy–Wiltshire Model 151
4.12 Hill–Clohessy–Wiltshire Model
A guidance method which combines both maneuvering and energy control can be
devised by using the approximate equations of relative motion between a spacecraft
and a target point moving in a fixed circular orbit. Referred to as the Hill–Clohessy–
Wiltshire (or HCW) model [24], these equations of relative motion are valid for
small relative separation. Being a linearized system of equations, they have an
analytical solution, which is easily used to derive an optimal control strategy.
Consider a right-handed reference frame called the HCW frame, (ir , iθ , ih ),
attached to a target moving in a circular orbit of radius, C, around a spherical planet,
such that ir is pointed radially outward, iθ is tangential to the orbit and pointed in the
direction of target’s motion, and ih is normal to the orbital plane, along the direction
of the angular momentum. Now consider a spacecraft which can be maneuvered by
the application of thrust, located at radius r from the planetary center, and having
a relative position ρ from the target. These two position vectors are resolved in the
HCW frame as follows:
r = Cir + ρ (4.191)
ρ = δrir + Cδθ iθ + δzih
where the coordinates (δr, Cδθ, δz) denote the small radial, in-track, and out-of-
plane displacements, respectively, of the maneuvering spacecraft from the . target.
Since the HCW frame is rotating with an angular velocity, nih , where n = μ/C 3 ,
the velocity of the maneuvering spacecraft is given by the following time derivative:
ṙ = δ ṙir + Cδ θ̇ iθ + δ żih + nih × r (4.192)

= (δ ṙ − nCδθ )ir + (Cδ θ̇ + nC + δr)iθ + δ żih
Similarly, the acceleration of the maneuvering spacecraft is resolved in the HCW

frame as follows:
r̈ = (δ r̈ − nCδ θ̇ )ir + (Cδ θ̈ + δ ṙ)iθ + δ z̈ih + nih × ṙ (4.193)

= (δ r̈ − 2nCδθ + n2 (C + δr) ir + (Cδ θ̈ + 2nδ ṙ − n2 Cδθ )iθ + δ z̈ih
The acceleration due to gravity of the spherical planet on the spacecraft is the
following:
r (C + δr)ir + Cδθ iθ + δzih

g = −μ = −μ 2
r 3 (C + 2Cδr + δr 2 + C 2 δθ 2 + δz2 )3/2
μ μ
−C 3 ir − 3 (−2δrir + Cδθ iθ + δzih ) (4.194)
C C
= n2 (2δr − C)ir − n2 Cδθ iθ − n2 δzih
where the binomial theorem has been applied in the light of the approximation,
δr C, δθ 1, and δz C.
Equations (4.193) and (4.194) substituted into Newton’s second law for the
maneuvering spacecraft, r̈ = g + a, produce the following scalar equations of
approximate relative motion, commonly referred to as the Hill–Clohessy–Wiltshire
(or HCW) equations:
δ r̈ − 2nCδθ − 3n2 δr = ar
Cδ θ̈ + 2nδ ṙ = aθ (4.195)
z̈ + n2 z = an
where a = ar ir + aθ iθ + an ih is the applied thrust acceleration. It is traditional to

abbreviate the deviations such that the in-track displacement is denoted x = Cδθ
and the radial displacement as y = δr, with the corresponding thrust accelerations
given by ax = aθ , ay = ar , and az = an . Then Eq. (4.195) is expressed as follows:
ẍ + 2nẏ = ax
ÿ − 2nẋ − 3ω2 y = ay (4.196)
z̈ + n2 z = az
Equation (4.196) can be represented by the following state-space model:
ṙ = v
v̇ = Ar + Bv + a (4.197)
where r = (x, y, z)T , v = (ẋ, ẏ, ż)T , a = (ax , ay , az )T , and

⎛ ⎞ ⎛ ⎞
0 0 0 0 −2n 0
A = ⎝ 0 3n2 0 ⎠ , B = ⎝ 2n 0 0 ⎠
0 0 −n2 0 0 0
The initial state is specified by r(0) = r0 , and v(0) = v0 .

The Hamiltonian for an energy-optimal transfer between two given points is thus
the following:
1 T
H = a a + λTr v + λTv (Ar + Bv + a) (4.198)
2
with the costate equations given by
T
∗ ∂H
λ̇r = − = −AT λ∗v (4.199)
∂r
T
∗ ∂H
λ̇v = − = −λ∗r − BT λ∗v
∂v
and the boundary conditions on the state and the costate variables being dependent
upon the type of transfer. For example, an interception problem is posed as a
transfer from the initial state to a final null relative position, r(tf ) = 0 in a given
time, tf . The terminal boundary condition, λv (tf ) = 0, is then specified to close
the problem. On the other hand, a rendezvous problem would have the terminal
boundary conditions as r(tf ) = 0 and v(tf ) = 0 for a given time, tf . For multiple
spacecraft involved in a formation flying constellation, the terminal condition on
the k th spacecraft is specified by a given separation from the (k − 1) remaining
spacecraft, which is specified as a boundary constraint, F[r(tf )] = 0.
The stationary condition, Ha = 0, provides the following extremal input profile:
a∗ (t) = −λ∗v (t) , (0 ≤ t ≤ tf ) (4.200)
The optimality of the extremal history is proved by seeing that it satisfies the
sufficient condition, Haa = I > 0.
Example 4.6 Consider a rendezvous between a space station in a circular orbit
around a spherical Earth of 90 min. period and a spacecraft initially located at
x(0) = −100 km, y(0) = 20 km, z(0) = −5 km, relative to the station, having
the initial relative velocity components, ẋ(0) = −1.0 m/s, ẏ(0) = 2.0 m/s, and
ż(0) = −0.5 m/s. If the spacecraft is equipped with a low-thrust engine which
can provide a continuous acceleration of magnitude 1 m/s2 , select the smallest
possible time of transfer, tf , for an energy-optimal rendezvous. Determine the
optimal trajectory and input profile.
For a rendezvous, the relative separation and velocity should both vanish
simultaneously at t = tf , that is r(tf ) = 0, and v(tf ) = 0. The Hamiltonian for an
energy-optimal transfer is given by Eq. (4.198), the costate equations by Eq. (4.199),
and the extremal input by Eq. (4.200). The associated 2PBVP is numerically solved
by a collocation scheme [81], and the results are plotted in Figs. 4.3, 4.4, 4.5, and
4.6. However, the upper limit of input acceleration is not imposed in the solution
process, but instead the final time is selected by trial and error such that the
maximum input magnitude does not exceed 1 m/s2 . In the next chapter, we will
consider how an input acceleration bound can be imposed.
Figure 4.4 shows the plots of the trajectory, (x, y, z), as well as the input
acceleration magnitude, a, against the traverse angle (converted to degree), nt, for
three values of the terminal time of rendezvous, ntf = 55.26◦ , 90◦ and 180◦ . As
seen in Fig. 4.4, the smallest possible value of terminal time is for ntf = 55.26◦
(corresponding to tf = 828.9 s), for which the largest required thrust acceleration,
a, is just under the maximum allowable value of 1 m/s2 . As the value of tf
is increased, the peak value of optimal a decreases. The maximum overshoot of
(x, y, z) also become smaller as tf is increased.
50 40
0 20
x (km)
y (km)
−50 0
−100 −20
−150 −40
0 60 120 180 0 60 120 180
ntf = 55.26° ntf = 90° ntf = 180°
0 1
0.8
−2
a (m/s2 )
z (km)
0.6
0.4
−4
0.2
−6 0
0 60 120 180 0 60 120 180
nt (deg.) nt (deg.)
Fig. 4.3 The optimal trajectory for rendezvous, and the corresponding thrust acceleration magni-
tude plotted for three values of the terminal time, tf
Figure 4.4 shows the velocity components, (ẋ, ẏ, ż), for the three values of
ntf , whereas Fig. 4.5 plots the thrust acceleration components, (ax , ay , az ). The
three-dimensional path taken by the spacecraft in the HCW frame until rendezvous
(marked by an “x”) is shown in Fig. 4.6 for the three terminal times, illustrating
the case ntf = 55.26◦ results in the shortest possible distance between the initial
and final points. Such a shortening of the optimal path (hence the time of transfer)
with an increased input magnitude forms the basis of bang–bang profiles, which
will be studied in the following chapter when the constraint on the maximum thrust
acceleration is enforced.
Apart from the HCW model, more accurate models of relative motion can be
derived, for which state transition matrices can be analytically evolved, such as the
one for elliptical orbit of the target by Yamanaka and Ankersen [86]. These can be
applied with greater fidelity to problems requiring accurate relative positioning of
spacecraft, as in formation flying applications. Further enhancements are possible
by including the atmospheric drag and oblateness effects [44] in a relative motion
model, which are crucial in close-proximity formation flying missions in low Earth
orbits.
ntf = 55.26° ntf = 90° ntf = 180°

200
ẋ (m/s)
100
−100
0 20 40 60 80 100 120 140 160 180
50
ẏ (m/s)
−50
−100
0 20 40 60 80 100 120 140 160 180
10
ż (m/s)
−5
0 20 40 60 80 100 120 140 160 180
nt (deg.)
Fig. 4.4 The optimal relative velocity components for rendezvous plotted for three values of the
terminal time, tf
Exercises
4.1 For the orbital regulation system of a satellite in a circular orbit of radius C
around a planet, use the linearized state equations with low-thrust radial acceleration
control input, ar (t), to derive the two-point boundary value problem for reducing
all initial perturbations, δr(0), δ ṙ(0), δθ (0) to zero in a fixed final time, tf , while
minimizing the total control power given by
tf
J = ar2 (t)dt .
0
4.2 Repeat Exercise 4.1 if the final angular error, δθ (tf ), is unspecified, but both
radial displacement and radial velocity errors must finally be made zero at an
unspecified time tf , while minimizing the total control power.
4.3 Derive the optimal trajectory for a low-thrust, coplanar orbital transfer using
only the input, aθ , between two given positions, (r1 , θ1 ) and (r2 , θ2 ), with a free
final time, tf , and an energy-optimal profile. (Hint: Resolve the flight path along
radial and tangential directions, and assume that due to the small thrust acceleration,
the change in flight path angle is negligible.)
ntf = 55.26° ntf = 90° ntf = 180°

1
ax (m/s2 )
−1
0 20 40 60 80 100 120 140 160 180
0
ay (m/s2 )
−0.5
−1
0 20 40 60 80 100 120 140 160 180
0.05
az (m/s2 )
−0.05
0 20 40 60 80 100 120 140 160 180
nt (deg.)
Fig. 4.5 The optimal thrust acceleration components for rendezvous plotted for three values of the
terminal time, tf
t = tf
50
0
−2 0
z (km)
t=0
−4
−50
−6
30 −100
20
10
0 x (km)
−10
−20 −150
y (km) −30
Fig. 4.6 The optimal trajectory for rendezvous plotted in three dimensions for three values of the
terminal time, tf , showing the paths taken in the HCW frame
4.4 Consider the low-thrust, energy-optimal rendezvous to a target in a circular

orbit of frequency n and radius C, using the approximate equations of relative
motion, when only the input aθ is applied. Derive the expression for the energy-
optimal trajectory for a transfer with tf = π/n with δr(0) = C/1000, δ ṙ(0) =
δθ (0) = δ θ̇ (0) = δz(0) = δ ż(0) = 0.
4.5 Repeat Exercise 4.4 if only the normal acceleration input, an , is applied, and
the initial condition is changed to δz(0) = C/1000, δr(0) = δ ṙ(0) = δθ (0) =
δ θ̇ (0) = δ ż(0) = 0.
4.6 Derive the terminal boundary conditions to be satisfied by the coefficients of
the extremal trajectory in Example 4.1, if the final time, tf , is unspecified.
4.7 A spacecraft is to be launched from the Moon’s surface by applying a known
thrust acceleration profile, a(t), at an angle β(t) from the horizontal direction. This
must be done such that the following objective is maximized before the engine cuts
out in a fixed time, tf :
1
J = gr(tf ) + v 2 (tf )
2
where g is the constant acceleration due to gravity, assumed to be always vertically
downward. Derive a closed-form expression for the optimum thrust direction profile,
β(t).
4.8 A lunar lander must make a smooth touchdown on the Moon at a point
located (1680, −300, 300)T km in an inertial reference frame. Its final velocity
at touchdown should be zero. Assuming that the position and velocity of the
spacecraft 500 s before the touchdown are (1750, 300, −300)T km and (0, 1, 1)T
km/s, respectively, and the Moon’s acceleration due to gravity is constant at
(−1.6, 0, 0) m/s2 , calculate the optimal thrust acceleration profile, and the resulting
optimal trajectory.
4.9 Consider two spacecraft traveling in nearly circular orbits of approximately
the same radius, C, and in the same plane. It is desired to apply the line-of-sight
guidance to one of the spacecraft, such that a constant closure rate v is maintained
while the line of sight rotates in the orbital plane at a rate θ̇ , thereby producing
a relative lateral acceleration, v θ̇ . A corrective normal acceleration input, u(t), is
continuously applied to the maneuvering spacecraft such that an interception takes
place at a given terminal time, tf . The projected miss distance is given by
2
z(t) = tf − t v θ̇
Using the HCW model of approximate relative motion, determine an energy-optimal

input profile, u(t).
4.10 Derive Edelbaum’s [28] solution for the pitch and yaw angles for a time-
optimal transfer between inclined circular orbits, (a0 , i0 ), and (af , if ), using Gauss’
variational equations, and using only the acceleration components,
, aθ , an , with
a constant, low-thrust acceleration magnitude, u = aθ2 + an2 , such that the
approximately circular shape (e 0, r a) is maintained throughout the transfer.
Chapter 5
Optimal Maneuvers with Bounded Inputs
5.1 Introduction
Spacecraft engines can generate only limited thrust magnitudes. This implies that
the acceleration inputs in space navigation must necessarily be bounded. The
trajectory optimization must therefore be performed taking bounded inputs into
account. As discussed in the earlier chapters, the nature of optimal trajectories
with bounded acceleration inputs can be classified as either impulsive thrust or
continuous thrust maneuvers. A further classification is possible according to the
gravity field in which such maneuvers are performed. This chapter presents a general
optimization method which can be applied to the various kinds of space navigation
problems. While the approach follows the formal development of optimal control
presented earlier, the focus is on the so-called primer vector method of Lawden [52]
and its further applications.
5.2 Optimal Thrust Direction
The equations of motion of a spacecraft in a general gravity field with thrust

acceleration input, a(t) = a(t)n(t), where n is a unit vector in the thrust direction,
can be expressed as follows:
v̇ = g(r, t) + an (5.1)
and
ṙ = v (5.2)

https://doi.org/10.1007/978-3-030-03789-5_5
160 5 Optimal Maneuvers with Bounded Inputs
with the initial condition, r(t0 ) = r0 , where g(r, t) is the acceleration due to gravity.
The thrust acceleration input is bounded in magnitude by the engine’s capacity. The
thrust magnitude, T (t), of an ideal rocket engine with an exhaust speed, ve (t), is
given by
T (t) = −ve (t)ṁ (5.3)
where ṁ is the rate of change of the vehicle’s mass (equal in magnitude and opposite
in sign to the mass exhaust rate). A more traditional expression for thrust magnitude
is in terms of the specific impulse, Is (t), which is the ratio of the exhaust speed and
acceleration due to gravity on Earth’s surface, Is = ve /gs , given by
T (t) = −gs Is (t)ṁ (5.4)
where gs = 9.81 m/s2 . Note that Eq. (5.4) gives a constant value of Is for a chemical
rocket engine, because the exhaust speed ve of such engines is nearly constant. The
thrust magnitude of such an engine can still be varied by changing the mass flow
rate (by a process called throttling) up to a certain extent:
βm ≤ ṁ ≤ 0 (5.5)
thereby yielding
0 ≤ T (t) ≤ Tm (5.6)
where Tm = −gs Is βm is the maximum-thrust magnitude (constant) corresponding

to the maximum mass exhaust rate, −βm (also constant). The bounds on the
acceleration magnitude provided by the constant specific impulse engine are
therefore the following:
Tm
0≤a≤ (5.7)
m(t)
This implies that the upper bound of the acceleration magnitude provided by a
constant specific impulse engine is not constant, but increases as the vehicle’s mass
is consumed. For a constant maximum throttle setting, βm , the mass variation with
time is given by
m(t) = m0 + βm (t − t0 ) (5.8)
with m0 = m(t0 ), which yields the maximum acceleration limit, 0 ≤ a ≤ am (t), as

follows:
Tm βm
am (t) = = −gs Is (5.9)
m0 + βm (t − t0 ) m0 + βm (t − t0 )
5.2 Optimal Thrust Direction 161
An electric (or ion) propulsion engine has a constant power output, P =

T (t)ve (t) = −ṁve2 (t), instead of a constant exhaust speed, thereby producing
a time-variable specific impulse, Is (t). Such an engine also has a variable mass
exhaust rate, −ṁ, inversely proportional to the square of the exhaust speed. The
bounds on the mass flow rate are however constant:
βm ≤ ṁ ≤ 0 (5.10)
The acceleration magnitude bounds of a constant power engine are thus represented
by
0 ≤ a ≤ am (t) (5.11)
where
√
Tm (t) −Pβm
am (t) = = (5.12)
m(t) m(t)
The upper bound, am , can be approximated as being constant, if the propellant

consumed, m0 − m(t), is minimized to be a small fraction of the initial mass, m0 , at
any time t.
It is often useful to define a characteristic speed, c(t), as the integral of the
acceleration input magnitude:
t
c(t) = a(τ )dτ (5.13)
t0
This is the maximum possible change in the vehicle’s velocity if the thrust is applied
tangentially, and the acceleration due to gravity is negligible in comparison, which
is the case for a large (or impulsive) acceleration input. If the vehicle’s current mass,
m(t), is to be determined, the state space is enhanced by an additional variable, c(t),
where
ċ = a (5.14)
with the initial condition c(t0 ) = 0. This implies that the characteristic speed, c, is a
measure of the total elapsed time, and influences the current mass, m, as well as the
upper limit of acceleration, am . The final characteristic speed, c(tf ), gives a measure
of the total propellant expenditure, m(t0 ) − m(tf ), and therefore must be minimized
for an optimal maneuver. Hence, the performance index to be minimized for the
smallest possible propellant requirement, subject to the dynamic equality constraints
given by Eqs. (5.1)–(5.2) and (5.14), is the following Mayer type function:
tf
J = c(tf ) = a(t)dt (5.15)
t0
For the special case of a constant specific impulse, the characteristic speed allows
an exact integration of Eq. (5.3) for the current mass, m(t), in terms of the initial
mass, m0 , and the constant exhaust speed, ve , expressed as follows:
c(t)
m(t) = m0 e− ve (5.16)
and
βm gc(t)I
am (t) = −gs Is e ss (5.17)
m0
Irrespective of the type of the engine employed, a minimization of the terminal

characteristic speed, J = φ = c(tf ), results in the following Hamiltonian for the
optimal control problem:
H = λTr v + λTv [g(r, t) + an] + λc a (5.18)
where λr and λv are the costate vectors corresponding to r and v, respectively, and
λc corresponds to the characteristic speed. The costate equations for λr and λv are
derived from the necessary conditions of optimality as follows:
T T
∂H ∂g(r, t)
λ̇r = − =− λv (5.19)
∂r ∂r
T
∂H
λ̇v = − = −λr (5.20)
∂v
Here the gravity-gradient matrix, ∂g(r, t)/∂r, depends upon the nature of the gravity
field. For example, for a spherical gravity field of the restricted two-body problem,
we have a time-invariant acceleration due to gravity:
μ
g(r) = − r (5.21)
r3
∂g(r) μ 3μ
= − 3 I3×3 + 5 rrT (5.22)
∂r r r
The general optimal maneuvers with the final time, tf , unspecified can be posed
as optimal boundary value problems with the given initial position, r(t0 ) = r0 , and
a terminal state lying on a moving hypersurface (see Chap. 2) expressed as follows:
F[r(tf ), v(tf ), tf ] = 0 (5.23)
where F : R7×1 → Rk×1 , 1 ≤ k ≤ 5. An example of such maneuvers is the

orbital rendezvous with a target moving in a specified orbit, but with the rendezvous
point being unspecified. As covered in Chap. 2, the associated terminal boundary

condition from the Euler–Lagrange formulation is the following:
T
∂F
λr (tf ) = d (5.24)
∂r t=tf
T
∂F
λv (tf ) = d (5.25)
∂v t=tf
where d = (d1 , d2 , . . . , dk )T are the additional adjoint variables used to adjoin the
terminal constraint equation, Eq. (5.23), to the performance index
J = φ[x(tf ), tf ] + dT F = c(tf ) + dT F (5.26)
The control variables of this problem are u = (a, nT )T , with n being the unit
thrust direction vector. The optimal control problem is singular for a, because of the
linearity of the Hamiltonian with respect to a. However, it is possible to minimize
H relative to the thrust direction, n, such that the optimal thrust direction, n̂, is
obtained. Representing the thrust direction unit vector in terms of its direction
cosines, n = (1 , 2 , 3 )T , such that
21 + 22 + 23 = 1 (5.27)
we express the Hamiltonian as follows:

H = λTr v + λTv [g(r, t) + an] + λc a + λn 21 + 22 + 23 − 1 (5.28)
where the static constraint, Eq. (5.27), is adjoined with the help of the additional
costate parameter, λn . The costate equation corresponding to the static constraint is
given by
T
∂H
0= = aλv + 2λn n̂ (5.29)
∂n
which is a necessary condition for the minimization of H relative to n.

On examining Eq. (5.27), it is found that unless both a and λn vanish, the optimal
thrust direction, n̂, must be aligned opposite to the direction of the costate vector
corresponding to velocity, λv , that is
λv
n̂ = − (5.30)
λv
with λn = 2λa v . This optimal thrust direction is referred to by Lawden [52] as the
primer vector, p(t) = −λv (t), and is an important result of the general optimal
space navigation problem, irrespective of the bound on the input magnitude, a.
Hence we write
p
n̂ = (5.31)
p
If a = 0, then the optimal thrust direction, n̂, becomes indeterminate.

The substitution of Eq. (5.31) into Eqs. (5.19) and (5.20) yields the following
differential equations for the costate variables with the optimal thrust direction:
T
∂g
λ̇r = p (5.32)
∂r
and
ṗ = λr (5.33)
combining both of which we have the following second-order ordinary differential

equation for the primer vector:
T
∂g
p̈ = p (5.34)
∂r
The boundary conditions for the primer vector and its time derivative are obtained
from Eqs. (5.24) and (5.25) as follows:

∂F T
ṗ(tf ) = d (5.35)
∂r t=tf
T
∂F
p(tf ) = − d (5.36)
∂v t=tf
Hence, we have the complete set of boundary conditions for integrating Eq. (5.34)
backward in time in order to determine the primer vector, p(t).
The costate equation corresponding to the characteristic speed is the following:
T T
∂H ∂H
λ̇c = − = −a (5.37)
∂c ∂t
where the Hamiltonian varies with time due to the dependence of the upper limit,
am (t), on time. Furthermore, the terminal boundary condition on λc yields the
following:

∂φ
λc (tf ) = =1 (5.38)
∂c t=tf
The Hamiltonian can be expressed in terms of the primer vector as follows:
H = a(λc − p) + ṗT v − pT g (5.39)
Due to the linearity of H relative to a, the optimal control problem is singular in

acceleration magnitude. However, the bounds on a given by Eq. (5.7) allow the
application of the Pontryagin’s minimum principle (Chap. 2) as follows:
â(λc − p) + ṗT v − pT g ≤ a(λc − p) + ṗT v − pT g (5.40)
or
â(λc − p) ≤ a(λc − p) (5.41)
where â refers to the optimal control, and
0 ≤ a ≤ am (t) (5.42)
The last two inequalities result in the following switching condition for the optimal
control magnitude:
&
am (t) , (p − λc ) > 0
â(t) = (5.43)
0, (p − λc ) < 0
The sign of the switching function, (p − λc ), is determined by whether the

magnitude, p, of the primer vector is either greater than or less than λc . While
λc (t) depends solely upon the engine characteristics (Eq. (5.37)), it is the primer
vector which gives a control over the optimal control magnitude. The bang–bang
profile switches between the “full on” and zero (or “off”) acceleration magnitudes,
and thus the optimal trajectory consists of the maximum-thrust (MT) and null-thrust
(NT) arcs. Such a feature is common to all singular space navigation problems, such
as the fuel minimization mission considered here, as well as the minimum-time
orbital transfer missions.
It is also possible to have an “intermediate-thrust” (IT) arc satisfying the
switching condition, for which the switching function vanishes (p = λc ), hence
the control magnitude can fall between the two bounds:
0 < â < am (t) , p = λc (5.44)
Such arcs are typically extremal arcs which satisfy the first-order necessary
conditions of optimality. However, due to their being singular arcs, they do not
satisfy the Legendre–Clebsch sufficient optimality condition, therefore it is difficult
to prove their optimality based upon second-order Hamiltonian variation.
5.2.1 Constant Acceleration Bound
The problem in which the acceleration upper bound, am (t), varies with time is
difficult to analyze and solve. However, if the assumption of a small, minimum
propellant mass is made, we write
Tm Tm
am = (5.45)
m(t) m0
Furthermore, if the gravity field is time invariant, it can be shown that the
Hamiltonian becomes constant on the optimal trajectory, resulting in the following:
λ̇c = 0 (5.46)
This implies that λc = const., which of course results in λc = λc (tf ) = 1, and

yields the following Hamiltonian:
H = a(1 − p) + ṗT v − pT g (5.47)
The switching condition for the constant acceleration input bound, am , is therefore
the following:
⎧
⎪
⎪ am , (p − 1) > 0
⎨
â(t) = 0, (p − 1) < 0 (5.48)
⎪
⎪
⎩
0 < â < am , p=1
5.2.2 Bounded Exhaust Rate
Often it is more convenient to express the acceleration magnitude, a, as a function

of the mass exhaust rate, β = −ṁ, which is always physically bounded by constant
limits:
0 ≤ β ≤ βm (5.49)
These inequalities can alternatively be expressed as the following equation:
β(β − βm ) − α 2 = 0 (5.50)
where α is an additional (fictitious) control variable without any physical signifi-

cance. The equations of motion are recast in terms of β as the input variable instead
of a as follows:
ṙ = v
ve
v̇ = g + β n (5.51)
m
ṁ = −β
where (β, n) are the input variables. For the constant specific-impulse case (that
is, ve = const.), the objective function for minimization, subject to the equality
constraints, Eqs. (5.50) and (5.51), is now expressed as follows:
tf tf
β(t)ve ṁve m0
J = c(tf ) = dt = − dt = ve ln (5.52)
t0 m(t) t0 m m(tf )
with the initial condition given by r0 = r(t0 ), v0 = v(t0 ), and m0 = m(t0 ), and
the terminal state satisfying Eq. (5.23). The minimization of J is thus equivalent to
maximizing the final mass, m(tf ), subject to the given constraints, thereby resulting
in the minimum propellant consumption. With the optimum thrust direction, n =
p/p, the Hamiltonian is then given by
βve
H = ṗT v − pT g − p − λm β + λβ [β(β − βm ) − α 2 ] (5.53)
m
The costate equations, apart from the primer vector differential equation,
Eq. (5.34), are the following:
T
∂H βve
λ̇m = − =− p
∂m m2
∂H ve
0=− = p + λm − λβ (2β − βm ) (5.54)
∂β m
∂H
0=− = 2λβ α
∂α
where the last two equations arise from the static equality constraint, Eq. (5.50).
Beginning with the last equation of Eq. (5.54), it is seen to be satisfied either by
λβ = 0 implying an intermediate-thrust (IT) arc (0 < β < βm ) or by α = 0,
which stands for either of the two extremes, that is the maximum-thrust (MT) arc
(β = βm ), or the null-thrust (NT) arc (β = 0). Hence, all three possibilities are
encompassed by this equation.
The terminal boundary condition on λm yields the following:

∂φ ve
λm (tf ) = =− (5.55)
∂m t=tf m(tf )
This boundary condition and the first equation of Eq. (5.54) imply that λm (t) < 0
for all t0 ≤ t ≤ tf .
The switching condition derived from Pontryagin’s minimum principle is the

following:
v v
e e
p + λm β ≤ p + λm β̂ (5.56)
m m
which yields λm + vme p > 0 for the NT arc, λm + ve

mp < 0 for the MT arc, and
λm + vme p = 0 for the IT arc.
5.3 Time-Invariant Gravity Field
Integration of the primer vector differential equation, Eq. (5.34), is readily made
for the null-thrust (NT) and maximum-thrust (MT) arcs in the gravity field of a
spherical central body, such as by Lawden [52] and Pines [67]. More generally, if
the gravity field is invariant in time, we have g = g(r), and the costate (adjoint)
equations expressed in terms of the primer vector are the following:
ṗ = λr (5.57)
and
T
∂g(r)
p̈ = λ̇r = p (5.58)
∂r
Furthermore, if the acceleration input has a constant upper bound, am , the time
derivative of the Hamiltonian
H = a(λc − p) + ṗT v − pT g (5.59)
vanishes along an optimal arc satisfying Eq. (5.34). Thus, we have
H = a(λc − p) + ṗT v − pT g = const. (5.60)
on an optimal trajectory. For a null-thrust (NT) arc, (a = 0), we have
C = ṗT v − pT g = const. (5.61)
Equation (5.61) also holds along an intermediate-thrust (IT) arc (p = λc ). This

result is also valid for the case of a bounded exhaust rate, because it also yields a
constant Hamiltonian along the optimal trajectory for a time-invariant gravity field.
A central gravity field, g(r) = g(r)r/r, has the following gravity gradient:

∂g(r) g r ∂g g rT
= I+ − (5.62)
∂r r r ∂r r r
5.4 Null-Thrust Arc in Central Gravity Field 169
which substituted into Eq. (5.58), yields the following:

g r ∂g g rT
p̈ = p + − p (5.63)
r r ∂r r r
For the optimal thrust direction, we have
ṙ = v (5.64)
and
r p
v̇ = g(r) + a (5.65)
r p
These, along with Eqs. (5.57) and (5.63), are used to demonstrate the following
constant throughout the optimal trajectory:
d
(p × v − λr × r) = ṗ × v + p × v̇ − λ̇r × r − λr × ṙ
dt
=0 (5.66)
or
D = p × v − λr × r = const. (5.67)
Thus the optimal trajectory in a time-invariant gravity field (such as that of a

two-body problem) has important constants of integration, irrespective of the type
of input acceleration profile employed.
5.4 Null-Thrust Arc in Central Gravity Field
While the primer vector is unnecessary for determining the thrust direction in a null-
thrust (NT) arc, its relationship with the radius and velocity vectors can be used to
derive important information about the motion. Furthermore, in a trajectory which
consists of both NT and powered (IT or MT) arcs, the evolution of primer vector
along an NT arc is necessary to determine its value at the beginning of a subsequent
powered arc.
Consider the flight in a general, central gravity field, g(r) = g(r)r/r. Along an
NT arc (a = 0), Eqs. (5.1) and (5.2) yield the following constant angular momentum
vector:
h = r × v = const. (5.68)
which implies that the flight is confined to a plane normal to h. Other constants
for a null-thrust arc in a central gravity field are given by Eqs. (5.61) and (5.67).
Yet other integration constants can be derived based upon the specific nature of the
central field, such as the inverse-square law gravity field of a spherical body covered
in Chap. 3.
We collect the various governing equations for the NT arc as follows:
ṙ = v
r
v̇ = g(r)
r
ṗ = λr (5.69)
T
∂g(r)
λ̇r = p̈ = p
∂r
with constants of the motion given by Eqs. (5.61), (5.67), and (5.68).
To resolve the coasting flight in a central gravity field, g(r) = g(r)r/r, consider
a rotating, right-handed coordinate frame, (i, j, k), with origin at the center of the
field, and oriented such that (i, j) is the plane of the motion containing both r and
v (i.e., the plane normal to h = r × v) at all times. Thus k is along the constant
angular momentum vector, (k = h/ h). The orientation of the i axis is selected to be
always along the radius vector, (i = r/r), and j completes the triad with i × j = k.
The various vectors of motion are thus resolved along (i, j, k) as follows:
r = ri
v = vx i + vy j
λr = λx i + λy j + λz k (5.70)
p = p x i + py j + pz k
By substituting Eq. (5.70) into Eq. (5.47) we have the following Hamiltonian along
a null-thrust arc:
H = ṗT v − pT g = λx vx + λy vy − px g = const. (5.71)
which validates Eq. (5.61).

The velocity vector can also be expressed in terms of polar coordinates (r, θ ),
where θ is the angle made by r with an inertial reference line, I, in the plane of
motion:
v = vx i + vy j = ṙi + r θ̇j (5.72)
Hence, the constant angular momentum becomes the following:
h = r × v = rvy k = r 2 θ̇ k (5.73)
5.4 Null-Thrust Arc in Central Gravity Field 171
To solve Eq. (5.67) for the primer vector, it is convenient to resolve the vectors
of motion in an inertial frame, (I, J, k), where J is normal to both I and k such that
I × J = k. This requires the coordinate transformation
⎧ ⎫ ⎛ ⎞⎧ ⎫
⎨i⎬ cos θ sin θ 0 ⎨ I ⎬
j = ⎝ − sin θ cos θ 0 ⎠ J (5.74)
⎩ ⎭ ⎩ ⎭
k 0 0 1 k
Consequently, Eq. (5.67) produces the following scalar component equations:
− pz vy cos θ − (pz vx − λz r) sin θ = hc1

−pz vy sin θ + (pz vx − λz r) cos θ = hc2 (5.75)
px vy − py vx + λy r = hC3
where the constant vector on the right-hand side of Eq. (5.67) is taken to be
D = h(c1 I + c2 J + c3 k), with h = r 2 θ̇. By resolving the velocity components
in polar coordinates according to Eq. (5.72), the first two equations of Eq. (5.75) are
algebraically solved to yield
pz = −r (c1 cos θ + c2 sin θ ) (5.76)
Furthermore, scaling the constant Hamiltonian (Eq. (5.71)) by h2 , we write
h2 H = λx vx + λy vy − px g (5.77)
and using the third equation of Eq. (5.75), we solve for λx and λy as follows:
1 2
λx = h H − r 2 θ̇ 2 C3 + px (g + r θ̇ 2 ) − py θ̇
ṙ
λy = r θ̇C3 − px θ̇ + py ṙ/r (5.78)
The time-derivative of the primer vector resolved along the rotating axes can be
δp
ṗ = +ω×p (5.79)
δt
where
δp
= ṗx i + ṗy j + ṗz k (5.80)
δt
is the partial time derivative derived by treating (i, j, k) to be fixed and
ω = θ̇k (5.81)
is the angular velocity of the orbital motion. A substitution of Eqs. (5.80) and (5.81)
into Eq. (5.79) yields
ṗ = (ṗx − py θ̇ )i + (ṗy + px θ̇ )j + ṗz k (5.82)
Similarly, the second-order derivative is obtained to be the following:

δ ṗ
p̈ = + ω × ṗ
δt

= p̈x − py θ̈ − 2ṗy θ̇ − px θ̇ 2 i (5.83)

+ p̈y + px θ̈ + 2ṗx θ̇ − py θ̇ 2 j + p̈z k
5.4.1 Inverse-Square Gravity Field
A null-thrust (NT) arc in a gravity field of a spherical body is directly determined

from the solution to the two-body problem of Keplerian motion (Chap. 3). As
discussed in Chap. 3 for coasting arcs between impulses, the primer vector can
be integrated along an NT arc using either spherical coordinates or the classical
orbital elements [52]. The result of the integration can also be expressed in
terms of Lagrange’s coefficients (Chap. 3), or by an appropriate universal variable
(generalized coordinates) [8] to preclude the singularity of the classical orbital
elements in near equatorial and circular orbits. A generalized canonical variable
approach can also be used to integrate the primer vector along MT and NT arcs
in a spherical gravity field. Da Silva [26] extends a canonical variable method via
Sundman transformation for analytically determining NT arcs in a closed form.
Integrating the primer vector along an NT arc around the inverse-square field
of a spherical body, g(r) = −μr/r 3 , requires the expression of the primer-vector
differential equation as follows:

∂g(r) T μ 3μ
p̈ = p = − 3 p + 5 (rT p)r (5.84)
∂r r r
The right-hand side of Eq. (5.84) is then expressed as follows:
μ 3μ μ
− 3
p + 5 (rT p)r = 3 2px i − py j − pz k (5.85)
r r r
and the following scalar differential equations are derived by the substitution of
Eqs. (5.83) and (5.85) into Eq. (5.84):
2μ
p̈x − py θ̈ − 2ṗy θ̇ − px θ̇ 2 = px
r3
5.5 Intermediate-Thrust Arc 173
μ
p̈y + px θ̈ + 2ṗx θ̇ − py θ̇ 2 = − py (5.86)
r3
μ
p̈z = − 3 pz
r
Eliminating θ from these equations by using θ̇ = h/r 2 and θ̈ = −2hṙ/r 3 yields the
following:
h h h2 2μ
p̈x + 2py ṙ 3
− 2ṗy 2 − px 4 = 3 px
r r r r
h h h2 μ
p̈y − 2px ṙ + 2ṗ x − p y = − 3 py (5.87)
r3 r2 r4 r
μ
p̈z = − 3 pz
r
For the further development of the integrals of motion, refer to the method presented
in Chap. 3.
5.5 Intermediate-Thrust Arc
Intermediate-thrust arcs have been of much interest in the literature on optimal

space navigation with bounded inputs. Before a rigorous analysis of singular control
problems could be undertaken, there was a doubt whether the intermediate-thrust
arcs can be regarded as optimal control solutions. In early works [31, 49, 50],
it was unclear whether the intermediate-thrust arcs could be solutions of the
minimum-fuel, orbital transfer problems. Leitmann [53] first suggested that the
intermediate-thrust arcs may be candidate optimal arcs. This was followed by
further developments by Lawden [51, 52], wherein it was found that for the
minimum-fuel problems in a spherical gravity field with the final time unspecified,
the intermediate-thrust arc is a spiral centered on the attracting body. Such an arc
became famous in literature as Lawden’s spiral. Many researchers [10, 11, 16, 17,
32, 34, 39, 42, 45, 47, 71, 75, 83], and [3] investigated the optimality of intermediate-
thrust arcs such as Lawden’s spiral. The primary focus was on the derivation of
a generalized form of the Legendre–Clebsch condition (see Chap. 2) involving
second- and higher-order variations on control applied on the singular arcs. A survey
of these developments can be found in Bell [9] and Bell and Jacobson [12]. Such
investigations into the singular optimal control problems led to the maturing of the
optimal control theory, which was formalized into texts such as Bryson and Ho [19],
and Athans and Falb [4].
We are now ready to determine the optimality of singular, intermediate-thrust
arcs of the space navigation problem, such as Lawden’s spiral.
5.6 Lawden’s Spiral
Consider the flight of a spacecraft in an inverse-square (spherical) gravity field with

a constant specific impulse engine having an exhaust speed, ve , and a variable mass
exhaust rate, ṁ = −β. For simplicity, the thrust is assumed to always lie in the plane
of flight, thereby producing a two-dimensional motion. The position of the vehicle’s
center of mass, o, from the center of the gravity field, O, is given by the radius
vector resolved in polar coordinates, (r, θ ). Now consider a moving coordinate
frame, (i, k), such that i is along the instantaneous thrust direction, which makes
an angle ψ from an inertial reference line, and an angle φ from the normal to the
radius vector. The instantaneous velocity vector is resolved along and normal to the
thrust direction as follows:
v = ui + wk , (5.88)
with u denoting the vehicle’s velocity component along the thrust direction, i, and
w that along the direction, k, perpendicular to i. The rotational kinematics of the
thrust vector gives the following:
i̇ = ψ̇k , k̇ = −ψ̇i , (5.89)
which results in the following expression for acceleration resolved along (i, j):
v̇ = (u̇ − w ψ̇)i + (ẇ + uψ̇)k (5.90)
The equations of motion are then expressed as follows:
ṙ = u sin φ − w cos φ
r θ̇ = u cos φ + w sin φ
1 μ
u̇ = w ψ̇ + βve − 2 sin φ (5.91)
m r
μ
ẇ = −uψ̇ + 2 cos φ
r
ṁ = −β
where
π
θ =φ+ψ − (5.92)
2
The control variables of this fifth-order system are u = (β, φ), while the initial and
final boundary conditions are the following:
r(t0 ) = r0
5.6 Lawden’s Spiral 175
θ (t0 ) = θ0
m(t0 ) = m0
r(tf ) = rf
θ (tf ) = θf
u(t0 ) = ṙ(t0 ) sin φ(t0 ) + r0 θ̇ (t0 ) cos φ(t0 ) (5.93)
w(t0 ) = −ṙ(t0 ) cos φ(t0 ) + r0 θ̇ (t0 ) sin φ(t0 )
π
ψ(t0 ) + φ(t0 ) = θ0 +
2
u(t0 ) = ṙ(tf ) sin φ(tf ) + rf θ̇ (tf ) cos φ(tf ) (5.94)
w(tf ) = −ṙ(tf ) cos φ(tf ) + rf θ̇ (tf ) sin φ(tf )
π
ψ(tf ) + φ(tf ) = θf +
2
The objective function for minimization is the final characteristic speed given by
tf β(t)ve m0
J = c(tf ) = dt = ve ln (5.95)
t0 m(t) m(tf )
which, of course, means a maximization of the final mass, m(tf ).

It is clear from the state equations, Eq. (5.92), that the Hamiltonian is linear in β,
and hence the optimization problem is singular relative to β:
u w
H = λr (u sin φ−w cos φ)+λθ cos φ+ sin φ (5.96)
r r

βve μ μ π
+ λu w ψ̇+ − 2 sin φ + λw −uψ̇+ 2 cos φ −λm β+ν θ −φ−ψ+
m r r 2
where the mass exhaust rate, β, is physically bounded as follows:
0 ≤ β ≤ βm (5.97)
with βm being constant. The solution for an extremal intermediate-thrust arc with β
falling between the two limits, 0 < β < βm , is obtained from the Euler–Lagrange
necessary conditions for optimality. To this end, we switch to the primer vector
method introduced earlier in the chapter. Since the mass exhaust rate, β (rather
than the acceleration magnitude, a) is now a control variable affecting the thrust
magnitude, the application of Pontryagin’s minimum principle results in a switching
condition with p = const. along an intermediate-thrust arc (which is equivalent to
p = 1 for the constant acceleration bound am ). Therefore, we have p = pi for
the optimal thrust direction. The time derivatives of the primer vector are thus the
following:
ṗ = pi̇ = pψ̇k
p̈ = pψ̈k − pψ̇ 2 i (5.98)
The last equation is compared with Eq. (5.58),

T
∂g(r)
p̈ = p (5.99)
∂r
and noting that for a spherical gravity field, g(r) = −(μ/r 3 )r, we have
T
∂g(r) μ 3μ
p=− p + 5 rT pr (5.100)
∂r r3 r
Furthermore, we have
r = r sin φi − r cos φk (5.101)
and
rT p = rp sin φ (5.102)
Hence
T
∂g(r) μ 3μ
p=p 3
(3 sin2 φ − 1)i − p 3 sinφ cos φk (5.103)
∂r r r
Substituting Eq. (5.103) into Eq. (5.99), and comparing its components along i, j
with those of the second equation of Eq. (5.98), we have the following scalar
differential equations:
μ
ψ̇ 2 = (1 − 3 sin2 φ) (5.104)
r3
3μ
ψ̈ = − sinφ cos φ (5.105)
r3
By substituting
μ
g=− (sin φi − cos φk) (5.106)
r2
p = pi (5.107)
and
ṗ = pi̇ = pψ̇k (5.108)

into Eq. (5.61), we have the following constant of the IT arc:

μ
ṗT v − pT g = pw ψ̇ + p sin φ = const. (5.109)
r2
Since p = const. along an optimally directed IT arc, we write
μ
w ψ̇ + sin φ = C = const. (5.110)
r2
Another constant of the motion is obtained by using Eqs. (5.104) and (5.105) to
yield the following:
μ
ψ̇ 2 cos φ − ψ̈ sin φ = cos φ (5.111)
r3
which substituted into the fourth equation of Eq. (5.91), produces
ẇ + uψ̇ − r ψ̇ 2 cos φr ψ̈ sin φ = 0 (5.112)
Finally, the radial and transverse velocity components, given by the first two
equations of Eq. (5.91), are resolved along i, to yield
u = ṙ sin φ + r θ̇ cos φ = ṙ sin φ + r(φ̇ + ψ̇) cos φ (5.113)
which substituted into Eq. (5.112) gives the following result:

d
w + r ψ̇ sin φ = 0 (5.114)
dt
or,
w + r ψ̇ sin φ = A = const. (5.115)
Equations (5.110) and (5.115) are integrated to obtain Lawden’s spiral. The time
can be eliminated from these two differential equations, and instead φ can be used
as the independent variable. From Eqs. (5.91) and (5.92), we have
d
w=− (r cos φ) + r ψ̇ sin φ (5.116)
dt
which, substituted into Eq. (5.115), results in
d
A − 2w = (r cos φ) (5.117)
dt
This equation can be rewritten as follows, where the prime denotes the derivative
relative to φ:
d
(A − 2w)t
= (r cos φ) (5.118)
dφ
which can be integrated to yield φ as a function of time. Similarly, Eq. (5.115) yields
A−w
ψ
= t (5.119)
r sin φ
which is integrated to give the shape of the IT arc, with A being a constant shape
parameter.
Another useful information about the trajectory is derived by eliminating ψ
from Eqs. (5.110) and (5.115), while using Eq. (5.104), which yields the following
expressions for w and r in terms of the independent variable φ:
μ
w(A − w) = Cr − sin φ sin φ (5.120)
r
and
r μ 2
w 2 (1 − 3 sin2 φ) = Cr − sin φ (5.121)
μ r
An integration of the parametric differential equations, Eqs. (5.118) and (5.119),

is readily performed for the case where the terminal time, tf , is unspecified. In such
a case of open control interval, tf − t0 , we are free to select C = 0 in Eq. (5.110),
which results in the following relationships from Eqs. (5.120) and (5.121) [52]:
as 6
r= (5.122)
1 − 3s 2
and
-
1 μ
w=± 2 (5.123)
s a
where s = sin φ and a > 0 is a characteristic length. From Eq. (5.118), it follows
that

a 3 s 7 (5s 2 − 3)
t =± (5.124)
μ (1 − 3s 2 )2
which is integrated to obtain

-
μ 1 31 5 1 c
±t = − c5 + c3 − c −
a3 9 81 9 81 (3c2 − 2)
- 0 √ √ 1
2 2 c 3− 2
− ln √ √ + const. (5.125)
81 3 c 3+ 2
.
where c = cos φ. Equations (5.119) and (5.123) yield
3
ψ
= −5 (5.126)
s2
which is integrated to obtain
ψ = −3 cot φ − 5φ + const. (5.127)
and from Eq. (5.92) it follows that
θ = θ0 − 3 cot φ − 4φ (5.128)
where θ0 is a constant. Equations (5.122)–(5.128) represent a spiral with a mono-

tonically increasing
√ radius from 0 to ∞ as the angle φ is increased from 0 to
α = sin−1 (1/ 3). In the same range of φ, the angle θ increases from −∞ to
θ0 − 4α − 3/2. The sign ambiguity in Eqs. (5.123) and (5.124) is resolved by always
having a positive value of the angular rate, θ̇ , which requires a negative sign when
sin φ > 0, and a positive sign when sin φ < 0.
To investigate the optimality of Lawden’s spiral, we substitute Eq. (5.110) into
the expression for the Hamiltonian:
u w
H = λr (u sin φ − w cos φ) + λθ cos φ + sin φ
r r

2μ βve μ
+ λu C − 2 sin φ + + λw −uψ̇ + 2 cos φ (5.129)
r m r
π
− λm β + ν θ − φ − ψ +
2
The derivative relative to exhaust rate is then the following:
∂H ve
Hβ = = λu − λm (5.130)
∂β m
where the costate equations are the following:
∂H 2μ
λ̇r = − =− (λu sin φ − λw cos φ)
∂r r3
∂H
λ̇θ = − = −ν
∂θ
∂H λθ
λ̇u = − = λw ψ̇ − λr sin φ − cos φ (5.131)
∂u r
∂H λθ
λ̇w = − = −λu ψ̇ + λr cos φ − sin φ
∂w r
∂H βve
λ̇m = − = λu 2
∂m m
The second-order time derivative of Hβ is now derived as follows:
dHβ ve βve ve
Ḣβ = = λ̇u + λu 2 − λ̇m = λ̇u (5.132)
dt m m m
and
d2 Hβ ve βve
Ḧβ = = λ̈u + λ̇u 2 (5.133)
dt 2 m m
The necessary condition of optimality is given by Theorem 2.11 to be the following:
∂ ve
Ḧβ = λ̇u 2 ≤ 0 (5.134)
∂β m
for all t ∈ [t0 , tf ]. This condition is violated along Lawden’s spiral, because
ve > 0, m > 0 for all t ∈ [t0 , tf ], and λ̇u > 0 can be chosen to satisfy the boundary
conditions given by Eq. (5.93), thereby making the inequality of Eq. (5.134) strictly
positive. Hence, an intermediate-thrust arc like Lawden’s spiral does not minimize
the propellant consumption, and is therefore not an optimal trajectory. Bell and
Jacobson [12] arrive at the same result via a second-order variation of J along
Lawden’s spiral.
5.7 Powered Arcs
While null-thrust arcs can be derived in a closed form, the maximum-thrust (MT)
and intermediate-thrust (IT) arcs require numerical integration of the two-point
boundary value problem (2PBVP).
Example 5.1 Consider an escape trajectory consisting of null-thrust (NT) and
maximum-thrust (MT) arcs, with a constant specific-impulse engine of maximum
acceleration am = 1.0 m/s2 . The transfer is made beginning with the geocentric
position 8000I km and inertial velocity 1.8269J + 6.8182K km/s, measured in the
geocentric, right-handed inertial reference frame, (I, J, K), with K being the polar
axis, and I pointed at the initial launch point. Since there is no terminal boundary
constraint in this case, the engine is switched on and off at arbitrary intervals, while
the primer vector evolves from a given initial condition to maintain a constant thrust
direction. The simulation is carried out by a Runge–Kutta solver for 15,000 s of
flight, and the resulting trajectory is plotted in Fig. 5.1, with the velocity shown in
Fig. 5.2. The abrupt changes in the acceleration caused by switching between NT
and MT arcs are evident in the velocity plot.
5.7 Powered Arcs 181
4
x 10
1
Null−thrust arc
0.5
Max−thrust arc
0
Z (km)
−0.5
−1
−1.5
2000
3
0
2
−2000 1
4
0 x 10
Y (km) −4000 −1
X (km)
Fig. 5.1 The escape trajectory consisting of null-thrust and maximum-thrust arcs of Example 5.1
Null−thrust arc
10 Max−thrust arc
5
Ż (km/s)
−5
−10
2
10
0
5
−2 0
−5
−4 −10
Ẏ (km/s) Ẋ (km/s)
Fig. 5.2 The velocity plot for the escape trajectory of Example 5.1
Example 5.2 Consider a variation of the escape trajectory of Example 5.1 wherein
an initial primer vector of p(0) = −10−6 (3, 5, 8)T and ṗ(0) = 0 are specified,
signifying a fixed terminal velocity constraint. The optimal trajectory with the same
initial state and engine as in Example 5.1 is simulated for 12,000 s, and plotted in
Fig. 5.3, with the velocity plot shown in Fig. 5.4. The circular orbit is seen to be
preserved during the null-thrust arc, and changes when the thrust is applied. The
4
x 10
0.5
0
Z (km)
−0.5
−1
−1.5
4000
2000
1
0 0.5
−2000 0
4
−0.5 x 10
Y (km) −4000 −1
X (km)
Fig. 5.3 The optimal escape trajectory consisting of a null-thrust, an intermediate-thrust, and a
maximum-thrust arc for Example 5.2
10
5
Ż (km/s)
−5
−10
2
10
0
5
−2 0
−5
−4
Ẏ (km/s) −10
Ẋ (km/s)
Fig. 5.4 The velocity plot for the optimal escape trajectory of Example 5.2
evolution of the primer vector magnitude, p, is seen to be monotonically increasing

in Fig. 5.5, thereby yielding a single switching point from NT to MT arc at around
9600 s into the flight. The switching involves an intermediate-thrust (IT) arc with
a = am /2, resulting in a much smoother velocity plot when compared to that of
Example 5.1. The optimal thrust direction is seen to be constant throughout the
1.5
p
0.5
0
8600 8800 9000 9200 9400 9600 9800 10000
t (s)
Fig. 5.5 The monotonically increasing p of the optimal escape trajectory of Example 5.2,
indicating the switch from null thrust to powered flight regime at p = 1
40
30
φ = sin −1(px /p)

20 ψ = tan−1(py /pz )
(deg.)
10
−10
−20
0 2000 4000 6000 8000 10000 12000
t (s)
Fig. 5.6 The optimal thrust direction indicated by the angles φ and ψ for the optimal escape
trajectory of Example 5.2
trajectory in Fig. 5.6, but is used only in the powered portion of the flight. The
radius and speed are plotted in Fig. 5.7, and show the escape nature of the powered
trajectory, evident in a steeply increasing radius and a burst in the speed as the engine
is switched on at 9600 s point.
12000
r (km) 11000
10000
9000
8000
7000
0 2000 4000 6000 8000 10000 12000
7.8
v (km/s)
7.6
7.4
7.2
7
0 2000 4000 6000 8000 10000 12000
t (s)
Fig. 5.7 The radius and speed profiles for the optimal escape trajectory of Example 5.2
Example 5.3 In the previous two examples, the terminal constraints were not
explicitly given, and instead an initial primer vector was specified. Now consider
the planar transfer of a spacecraft around a spherical Earth from an initial position,
r(0) = (6000, 4000)T km, and velocity v(0) = (6, 2)T km/s, to a final position of
r(tf ) = (7000, 2000)T km by the application of constant specific-impulse engine
with the constant acceleration bounds, 0 ≤ a ≤ am . The results of this 2PBVP—
computed by a collocation technique [81]—are plotted for am = 0.1, 1.0, and
10 m/s2 , respectively, in Figs. 5.8 and 5.9. Figure 5.8 plots the inertial planar
trajectory, r = (x, y)T , for the three chosen values of am . The increasing terminal
positional error with the decreasing upper bound of the acceleration magnitude
can be clearly observed. Figure 5.9 shows the optimal thrust direction, n = p/p,
as the inertial angle made by the primer vector, p = (px , py )T , with the x-axis,
tan−1 (py /px ), as a function of time. The trajectory in each case is revealed in this
figure to be more than 99% the maximum-thrust (MT) arc, with a = am , with the
remaining less than (final) 1% being the null-thrust (NT) arc (p < 1). The value
of the terminal time, tf , is seen to increase as the thrust acceleration magnitude is
reduced. For the highly underactuated case of am = 0.1 m/s2 , the largest terminal
positional error as well as the largest terminal time is observed. The optimal speed
profile is plotted in Fig. 5.10.
Example 5.4 Let us now consider a three-dimensional transfer for escape from a
circular Earth orbit of radius, 42,000 km, inclination 75◦ , when passing the equator
with initial velocity, v(0) = 0.7973J + 2.9757K km/s, to the point rf = 51,336I +
5000
4500
4000
3500
y (km)
3000
2500
am = 0.1 m/s2
2000 am = 1.0 m/s2
am = 10 m/s2
1500
6000 6500 7000 7500 8000 8500 9000 9500 10000 10500
x (km)
Fig. 5.8 The optimal trajectory of the spacecraft for the interception problem of Example 5.3
60
40
20
0
tan−1(py /px )
−20 am = 0.1 m/s2

am = 1.0 m/s2
−40
am = 10 m/s2
−60
−80
−100
0 500 1000 1500 2000 2500 3000
t (s)
Fig. 5.9 The optimal thrust direction of the spacecraft for the interception problem of Example 5.3
8038J + 30,000K km, measured in the geocentric, right-handed inertial reference

frame, (I, J, K), with K being the polar axis, and I pointed at the initial launch
point. The transfer is made using a constant specific-impulse engine of maximum
acceleration am = 1.0 m/s2 . The time of transfer is selected to be 6051 s such
8
am = 0.1 m/s2
7 am = 1.0 m/s2
am = 10 m/s2
6
5
v (km/s)
0
0 500 1000 1500 2000 2500 3000
t (s)
Fig. 5.10 The optimal speed of the spacecraft for the interception problem of Example 5.3
that only powered flight is required. The trajectory includes an intermediate-thrust

(IT) arc, for which a = am /2 is used. As shown earlier, the IT arc is a suboptimal
solution, but it extends the flight time for reducing the terminal positional error,
when compared to a null-thrust arc. The performance index for minimization is the
following:
1
J = c(tf ) + φ[x(tf ), tf ] , φ= [r(tf ) − rf ]2
2
where
1
φ[x(tf ), tf ] = [r(tf ) − rf ]T [r(tf ) − rf ]
2
The boundary conditions are thus the following:
r(0) = (42,000, 0, 0)T (km) , v(0) = (0, 0.7973, 2.9757)T km/s

T
∂φ
λr (tf ) = = r(tf ) − rf , p(tf ) = 0
∂r t=tf
The solution to the 2PBVP is obtained using a collocation method [81], and
the results are plotted in Figs. 5.11, 5.12, 5.13, and 5.14, showing an almost linear
4
x 10
2
Z (km)
0
10000
8000 5.2
6000 5
4.8
4000 4
x 10
4.6
2000 4.4
Y (km) 0 4.2 X (km)
Fig. 5.11 The trajectory of the spacecraft for the three-dimensional transfer problem of Exam-
ple 5.4
45
φ = sin −1(py /p)
40
ψ = tan−1(pz /p)
35
30
25
(deg.)
20
15
10
0
0 1000 2000 3000 4000 5000 6000 7000
t (s)
Fig. 5.12 The optimal thrust direction of the spacecraft for the three-dimensional transfer problem
of Example 5.4
increase in the flight speed as the optimal thrust direction (indicated by the angles
φ = 45◦ and ψ = 15◦ ) is achieved and maintained throughout the flight. The
trajectory culminates with an IT arc covering the final 60 s of flight until the desired
position is reached with a hyperbolic escape velocity of 7.802 km/s.
60
58
56
54
r(×1000 km)
52
50
48
46
44
42
0 1000 2000 3000 4000 5000 6000 7000
t (s)
Fig. 5.13 The optimal radius of the spacecraft for the interception problem of Example 5.4
7.5
6.5
6
v (km/s)
5.5
4.5
3.5
3
0 1000 2000 3000 4000 5000 6000 7000
t (s)
Fig. 5.14 The optimal speed of the spacecraft for the interception problem of Example 5.4
The last two examples demonstrate the determination of a fuel-optimal trajectory

subject to the specific terminal boundary conditions (also referred to as the
transversality conditions). These conditions often cause convergence problems in
any numerical scheme used to solve the associated 2PBVP, because of the arbitrary
5.8 Linearization Relative to a Circular Orbit 189
nature of the costate boundary conditions. Since costate variables have no physical
meaning, they are generally chosen by trial and error to achieve convergence for
a given final time, tf . Resolving such convergence issues in either a shooting or
collocation technique is a topic of active research, such as the paper by Minter and
Fuller-Rowell [61] who propose setting the initial conditions of the adjoint variables
from an analytic solution of a simplified set of differential equations, and that by Pan
et al. [65], who propose a list of reduced transversality conditions based upon the
classical orbital elements.
5.8 Linearization Relative to a Circular Orbit
The optimal transfer relative to a circular orbit in an inverse-square gravity field

can be addressed using the linearized Hill–Clohessy–Wiltshire (HCW) model of
approximate relative motion (see Chap. 4). Consider the HCW equations of a
constant specific impulse powered spacecraft relative .
to a point (target) moving in a
circular orbit of radius R with orbital frequency, ω = μ/R 3 , expressed as follows:
ẍ + 2ωẏ = anx
ÿ − 2ωẋ − 3ω2 y = any (5.135)
z̈ + ω2 z = anz
Here x = Rδθ is the small in-track displacement, y = δr the small radial

displacement, and z = Rδα the small out-of-plane displacement, all measured from
the target as the origin of the HCW frame, and n = (nx , ny , ny )T is the thrust
direction unit vector,
a = an (5.136)
with the thrust acceleration magnitude, a =| a |, is bounded by the maximum

possible acceleration magnitude, am = Tm /m:
0 ≤ a(t) ≤ am (5.137)
Assuming a negligible change in the vehicle’s mass during the low-thrust maneuver,
the upper bound on the acceleration magnitude, am , can be taken to be constant,
thereby simplifying the control problem.
In a state-space form, Eq. (5.135) can be expressed as follows:
ṙ = v
v̇ = Ar + Bv + an (5.138)
where r = (x, y, z)T and v = (ẋ, ẏ, ż)T , and

⎛ ⎞ ⎛ ⎞
0 0 0 0 −2ω 0
A = ⎝ 0 3ω2 0 ⎠ , B = ⎝ 2ω 0 0 ⎠
0 0 −ω2 0 0 0
The problem of a fuel-optimal transfer from an initial state, [r(0), v(0)], to a final
state [r(tf ), v(tf )], in a specified time, tf , can be stated as the minimization of
tf
J = c(tf ) = a(t)dt (5.139)
0
subject to the constraints of Eqs. (5.137) and (5.138), resulting in the following
Hamiltonian:
H = a + λTr v + λTv (Ar + Bv + an) (5.140)
The application of the primer vector method yields the optimal thrust direction to
be the following:
p
n̂ = , (a = 0) (5.141)
p
where p = −λv is the primer vector. If a = 0, then the optimal thrust direction, n̂,
is indeterminate. The Hamiltonian is now expressed in terms of the primer vector as
follows:
H = a(1 − p) + ṗT v − pT (Ar + Bv) (5.142)

acceleration magnitude. However, the constant bounds on a given by Eq. (5.137)
allow the application of the Pontryagin’s minimum principle, resulting in the
following switching condition for the optimal control magnitude:
&
am , (p − 1) > 0
â(t) = (5.143)
0 , (p − 1) < 0
The costate equations produce the primer vector differential equation as follows:
λ̇r = AT p (5.144)
ṗ = λr − BT p (5.145)
which yields
p̈ + BT ṗ − AT p = 0 (5.146)
or, considering the symmetric and skew-symmetric nature of A and B, respectively,

we have
p̈ = Ap + Bṗ (5.147)
which is the same as the second equation of Eq. (5.138) with a = 0. Hence both the
primer vector and the HCW model have the same state transition matrix.
The boundary conditions for the primer vector and its time derivative for a
transfer with both the end conditions fixed and the terminal time specified are
completely arbitrary. They could be expressed as follows:
p(0) = p0 , ṗ(0) = ṗ0 (5.148)
and must be chosen for integrating Eq. (5.146) in order to determine the primer
vector, p(t), 0 ≤ t ≤ tf , thus the optimal thrust direction and magnitude.
The arbitrary nature of the primer vector indicates that infinitely many possible
fuel-optimal solutions exist for transferring the spacecraft from an initial state,
[r(0), v(0)], to a final state [r(tf ), v(tf )], in a specified time, tf . Such a degeneracy
of the singular problem with hard end-point constraints has been explored by
Carter [21] using the linearized HCW model, and by Marec [57] using the linearized
orbital elements formulation. The degeneracy of the optimal solutions is confined
to the linearized problem with specific position and velocity boundary conditions,
but the original (nonlinear) transfer problem discussed earlier in this chapter has a
unique optimal solution.
In order to derive a unique solution, the orbital transfer problem relative to a
circular orbit can be posed as either an interception or a rendezvous problem with
soft end-point constraints. For this purpose, a terminal cost function is added to the
performance index as follows:
J = c(tf ) + φ[x(tf ), tf ] (5.149)
where
1 1
φ[x(tf ), tf ] = Qr [r(tf ) − rf ]T [r(tf ) − rf ] + Qv [v(tf ) − vf ]T [v(tf ) − vf ]
2 2
(5.150)
where rf and vf are the target point’s position and velocity, respectively, at a
specific terminal time tf , and the weights Qr > 0, Qv > 0 can be chosen as desired
in comparison with the fuel cost, c(tf ). For example, an interception problem can
have Qv = 0, while a rendezvous would require both the weights to be selected
appropriately. The terminal conditions for the costate vectors can then be obtained
as follows:
T
∂φ
λr (tf ) = = Qr [r(tf ) − rf ]
∂r t=tf
T
∂φ
p(tf ) = − = −Qv [v(tf ) − vf ] (5.151)
∂v t=tf
Since the primer-vector differential equation, Eq. (5.146), is linear, it has a

closed-form solution, and can be expressed in terms of the individual elements of
the vector, p = (px , py , pz )T . We begin by writing Eq. (5.146) in terms of its scalar
components as follows:
p̈x + 2ωṗy = 0
p̈y − 2ωṗx − 3ω2 py = 0 (5.152)
p̈z + ω2 pz = 0
which have the same form as the unforced HCW equations, Eq. (5.146), and whose
general solution is the following:
px (t) = 2A1 sin(ωt + ϕ) + 3c1 t + c2

py (t) = A1 cos(ωt + ϕ) + 2c1 /ω (5.153)
pz (t) = A2 sin(ωt + ϕ) + A3 cos(ωt + ϕ)
The coefficients, A1, A2 , A3 , c1 , c2 , as well as the phase angle, ϕ, are determined

from the boundary conditions, Eq. (5.151), which also require the state solution at
the terminal time, r(tf ), v(tf ). Thus the iterative nature of the 2PBVP solution is
revealed.
Another way of expressing the primer vector solution is as follows:
p(t) = rr (t)p(0) + rv (t)ṗ(0) (5.154)
where rr , rv are the following transition matrices of the unforced HCW problem:
⎡ ⎤
1 6(sin ωt − ωt) 0
rr (t) = ⎣ 0 4 − 3 cos ωt 0 ⎦ (5.155)
0 0 cos ωt
⎡ ⎤
ω sin ωt − 3t ω (cos ωt − 1)
4 2
0
rv (t) = ⎣ − 2 (cos ωt − 1) 1 ⎦ (5.156)
ω ω sin ωt 0
1
0 0 ω sin ωt
The initial vectors, p(0), ṗ(0) are to be determined from the terminal boundary
conditions, Eq. (5.151), as well as the state solution at t = tf .
Recall that both the primer vector dynamics and the HCW model have the same
state transition matrix, expressed by

rr (t) rv (t)
(t) = ˙ rv (t)
˙ rr (t) (5.157)

where
⎡ ⎤
0 ω(6 cos ωt − 1) 0
˙ rr (t) = ⎣ 0
3ω sin ωt 0 ⎦ (5.158)
0 0 −ω sin ωt
⎡ ⎤
4 cos ωt − 3 −2 sin ωt 0
˙ rv (t) = ⎢
⎣ 2 sin ωt cos ωt 0 ⎥
⎦ (5.159)
0 0 cos ωt
the respective solutions can be written as follows:

p(t) p(tf )
= (t − tf ) , (0 ≤ t ≤ tf ) (5.160)
ṗ(t) ṗ(tf )

t & /
r(t) r(0) 0
= (t) + â(τ )(t − τ ) p(τ ) dτ , (0 ≤ t ≤ tf )
v(t) v(0) 0 p(τ )
(5.161)
or
t
r(t) r(0) (t − τ ) p(τ )
= (t) + â(τ ) ˙ rv dτ , (0 ≤ t ≤ tf )
v(t) v(0) 0 rv (t − τ ) p(τ )
(5.162)
Equations (5.160) and (5.162) are to be solved simultaneously, given the bound-
ary conditions, r(0), v(0), p(tf ), and ṗ(tf ), as well as the switching condition,
Eq. (5.143).
Since the HCW model has decoupled out-of-plane (z) and coplanar (x, y)
dynamics, it is possible to solve the two problems independently of each other. How-
ever, a three-dimensional rendezvous would require that both the out-of-plane and
the coplanar displacement and velocity components should vanish simultaneously
(that is, for the same value of tf ).
5.8.1 Out-of-Plane Rendezvous
In order to illustrate the procedure of solving the two-point boundary value problem
(2PBVP) associated with the linearized HCW model, consider a rendezvous
problem in which the spacecraft has the same circular orbit as the target, but
is displaced in plane by an initial displacement, z(0), and velocity, ż(0). The
rendezvous must be made with the minimum fuel in time tf , using a constant
specific-impulse engine of maximum (constant) acceleration, am .
The linearized scalar equation of out-of-plane motion is simply the following:
z̈ + ω2 z = anz (5.163)
with n̂z = 1 being the optimal thrust direction. The switching condition, Eq. (5.143),
is governed by the primer vector magnitude, p(t), obeying the following differential
equation:
p̈ + ω2 p = 0 (5.164)
whose solution is given by
p(t) = c1 cos ωt + c2 sin ωt (5.165)
where the coefficients c1 , c2 are determined from the terminal boundary condition:
ṗ(tf ) = Qr z(tf )
p(tf ) = −Qv ż(tf ) (5.166)
Since p(t) is oscillatory, it can produce infinitely many switching points, p(ts ) =
1. Hence, the rendezvous problem (i.e., z(tf ) = ż(tf ) = 0) can have infinitely many
solutions, depending upon the acceleration bound, am . To see this fact, the solution
for the out-of-plane displacement is expressed as follows:
1

z(t) cos ωt sin ωt z(0)
= ω
ż(t) −ω sin ωt cos ωt ż(0)
t
cos ω(t − τ ) ω1 sin ω(t − τ ) 0
+ â(τ ) dτ (5.167)
0 −ω sin ω(t − τ ) cos ω(t − τ ) 1
where 0 ≤ t ≤ tf . To apply the terminal boundary condition, we must have

p(tf ) 0 −Qv z(tf )
= (5.168)
ṗ(tf ) Qr 0 ż(tf )
Suppose it were possible to make the rendezvous with only one switching. Then
the trajectory must begin with an MT arc, â = am , for which p(0) = c1 > 1, and
the switching point, t = ts ≤ tf , should satisfy
c1 cos ωts + c2 sin ωts = 1 (5.169)

The displacement and velocity at the switching point are then obtained from
Eq. (5.167) as follows:
ts 1
ω sin ω(t − τ ) dτ
1
z(ts ) cos ωts sin ωts z(0)
= ω + am
ż(ts ) −ω sin ωts cos ωts ż(0) 0 cos ω(t − τ )
(5.170)
or
& /
z(ts ) z(0) cos ωts + ż(0) am
ω sin ωts + ω2 (1 − cos ωts )
= , (0 ≤ ts ≤ tf )
ż(ts ) −ωz(0) sin ωts + ż(0) cos ωts + aωm sin ωts
(5.171)
This is the initial condition for the NT arc, for which it should be true that
1

z(t) cos ωt sin ωt z(ts )
= ω , (ts ≤ t ≤ tf ) (5.172)
ż(t) −ω sin ωt cos ωt ż(ts )
By putting t = tf in Eq. (5.172), it may be verified that a perfect rendezvous

(z(tf ) = ż(tf ) = 0) requires that both z(ts ) and ż(ts ) should be made to vanish
for an arbitrary initial condition, z(0), ż(0), which implies the rendezvous could
take place at both t = ts and at t = tf . If this condition were to be satisfied, it would
mean that an NT arc is unnecessary, and an MT arc is itself sufficient to make the
rendezvous. Since the relative displacement and velocity both vanish at t = ts as
well as at t = tf (that is, at either end of the NT arc), it follows that tf = ts + 2π ω .
Therefore, we can replace ts with tf in Eqs. (5.169) and (5.171), resulting in the
following:
c1 cos ωtf + c2 sin ωtf = 1 (5.173)

& /
z(tf ) z(0) cos ωtf + ż(0) am
ω sin ωtf + ω2 (1 − cos ωtf )
= (5.174)
ż(tf ) −ωz(0) sin ωtf + ż(0) cos ωtf + aωm sin ωtf
or,
ωż(0)(1 − c1 )
sin ωtf = (5.175)
c1 ω2 z(0) + c2 ωż(0)
Given z(0), ż(0), the coefficients c2 and c1 > 1, and the terminal time, tf , should
satisfy Eq. (5.175), as well as the requirement

z(0) cos ωtf + ż(0)
ω sin ωtf
am = ω 2
>0 (5.176)
cos ωtf − 1
The last equation implies that the terminal time for exact rendezvous in one MT arc
is dependent upon the upper bound of the acceleration input, and can have many
solutions. A similar result can be obtained for multiple switching points.
For deriving a unique solution for the terminal time, tf , for rendezvous, the soft
terminal condition of Eq. (5.168) should be applied, which requires that

−c1 ω sin ωtf +c2 ω cos ωtf ωQv sin ωtf −Qv cos ωtf
= Qr (5.177)
c1 cos ωtf +c2 sin ωtf Qr cos ωtf ω sin ωtf
& /
z(0) cos ωts + ż(0) sin ωt s + am
(1 − cos ωt s )
× ω ω 2
−ωz(0) sin ωts + ż(0) cos ωts + aωm sin ωts
Given tf , Qr , and Qv , Eqs. (5.169) and (5.177) are to be solved for c1 , c2 , and ts .
However, such a rendezvous is not exact because the displacement and velocity do
not vanish simultaneously for any value of tf .
5.8.2 Coplanar Rendezvous
The approach considered above for out-of-plane rendezvous can also be applied to
the coplanar problem. The state transition matrix for the coplanar HCW model is
the following:
⎛ ⎞
1 6(sin ωt − ωt) ω sin ωt − 3t
4
ω (cos ωt − 1)
2
⎜ 0 4 − 3 cos ωt − 2 (cos ωt − 1) 1 ⎟
(t) = ⎜ ω ω sin ωt ⎟ (5.178)
⎝ 0 ω(6 cos ωt − 1) 4 cos ωt − 3 −2 sin ωt ⎠
0 3ω sin ωt 2 sin ωt cos ωt
Let the optimal trajectory begin with an MT arc which, by definition, extends until
the first switching time, t = ts ≤ tf , given by
p2 (ts ) = px2 (ts ) + py2 (ts ) = 1 (5.179)
The displacement and velocity at the switching point are then obtained to be the
following:

r(ts ) r(0)
= (ts ) (5.180)
v(ts ) v(0)
⎡4 ⎤
sin ω(t−τ ) − 3(t − τ ) ω [cos ω(t−τ ) − 1] & p (τ ) /
2
ts ω 2
⎢ − [cos ω(t − τ ) − 1] 1 ⎥ p(τ
x
+ am ⎢ ω ω sin ω(t−τ ) ⎥ p (τ)) dτ
⎣ 4 cos ω(t − τ ) − 3 −2 sin ω(t−τ ) ⎦ p(τy
0 )
2 sin ω(t − τ ) cos ω(t−τ )
with the primer vector elements evolving as follows:

p(t) p(tf )
= (t − tf ) , (0 ≤ t ≤ tf ) (5.181)
ṗ(t) ṗ(tf )
from the terminal boundary condition,
ṗ(tf ) = Qr r(tf )
p(tf ) = −Qv v(tf ) (5.182)
This completes the formulation needed to solve the optimal coplanar rendezvous
problem with the linearized HCW model.
Exercises
5.1 Formulate a two-point boundary value problem using only the radial thrust
acceleration bounded by | ar |≤ c for an open time transfer from an initial circular
orbit of radius, r1 , around a spherical body, to a coplanar circular orbit of radius, r2 .
5.2 Find the time-optimal input profile for a plane change maneuver by angle α0 in
a circular orbit of radius C, if | an |≤ 1 m/s2 .
5.3 A solar-sail spacecraft is to be transferred from an initial circular orbit of radius,
r(0) = r1 , around the Sun to a coplanar circular orbit of radius r, by applying
a solar-radiation acceleration of small magnitude, a = c/r 2 , where c μ is a
constant. The direction of the input acceleration from the radial direction is given
by the angle β(t), and can be adjusted by rotating the sail in the limits, −bπ/2 ≤
β(t) ≤ bπ/2, where 0 < b ≤ 1 is constant. Formulate the boundary value problem
to be solved such that r 2 (tf ) is maximized in a fixed time, tf , relative to β(t).
5.4 Consider an orbital regulation system of a satellite in a circular orbit of radius
C and frequency n around a spherical planet, with low-thrust radial acceleration
control input, |ar (t)| ≤ Cn2 , and initial perturbations, δr(0) = 0.01C, δ ṙ(0) =
0, δθ (0) = 0. Derive the time-optimal control history and trajectory for driving the
final radial response, δr(tf ), δ ṙ(tf ), to zero at an unspecified time tf .
5.5 A satellite in a circular orbit of radius C and frequency n around a planet is hit
by space debris, resulting in an initial out-of-plane velocity of Δż(0) = 0.01nC.
For restoring the original orbital plane, a normal acceleration input is applied by the
rocket engine, which is limited by
|an (t)| ≤ Cn2
Calculate the time-optimal trajectory until the first switching time.

5.6 Solve the optimal coplanar transfer problem using only the input, uθ , between
two given positions, (r1 , θ1 ) and (r2 , θ2 ), with a free final time, tf , and an energy-
optimal profile, if the magnitude of the input is limited by | uθ |≤ 1 m/s2 . (Hint:
Resolve the flight path along radial and tangential directions, and assume that due
to the small thrust acceleration, the change in flight path angle is negligible.)
5.7 Formulate the optimal control problem for the constant upper limit of acceler-
ation input magnitude, am , for the minimization of
tf
J = a(t)dt
0
subject to the dynamic constraints of Eqs. (5.1)–(5.2), but without taking c(t) as
a state variable. Show that the same switching condition is obtained for the thrust
direction as given by Eq. (5.43).
Chapter 6
Flight in Non-spherical Gravity Fields
6.1 Introduction
Spaceflight involving orbital transfers around irregularly shaped bodies or in the

gravity field of several large bodies is fundamentally different from the flight in
the gravity field of a single spherical body, which was covered in the previous
chapters. The primary reason for this difference is that the spacecraft is no longer
in a time-invariant gravity field of the two-body problem, but instead encounters a
time-dependent field due to the relative motion of the multiple large bodies with
respect to one another, or due to the changing position of the spacecraft relative
to a rotating, non-spherical body. A general description of such a flight is by the
following governing equation:
r̈ = g[r(t), t] + a (6.1)
where r is the instantaneous position vector of the spacecraft in an inertial frame, g

is the time-varying acceleration due to gravity caused either by an irregularly shaped
body or by multiple large bodies in mutual relative motion, and a is the acceleration
input applied by the engines. The determination of an optimal acceleration input
profile, a(t), in a specified control interval while minimizing a given objective forms
the crux of the spaceflight navigation problem. For example, an orbital transfer
between two given positions minimizing a cost function,
tf
J = φ[r(tf ), v(tf ), tf ] + aT adt , (6.2)
0
subject to the dynamics of Eq. (6.1), requires the solution of the unbounded accel-
eration input profile, a, from the following Euler–Lagrange equations (Chap. 2):
∂H
= λTv g + aT = 0 (6.3)
∂a

https://doi.org/10.1007/978-3-030-03789-5_6
200 6 Flight in Non-spherical Gravity Fields
T T
∂H ∂g
λ̇r = − =− λv (6.4)
∂r ∂r
T
∂H
λ̇v = − = −λr (6.5)
∂v

∂φ
H+ =0 (6.6)
∂t t=tf
where the Hamiltonian is given by
H = λTr v + λTv (g + a) + aT a (6.7)
and appropriate boundary conditions are specified for r, v, λr , and λv . Modeling a

non-spherical gravity field thus requires deriving expressions for the acceleration
due to gravity, g, and the gravity-gradient matrix, ∂g/r, both of which are time
varying at any given location, r, in the inertial space. Such a time-varying gravity
model would be cumbersome to employ in the solution of an optimal control
problem. However, a great simplification is possible if the gravity field is rendered
time invariant by transforming the equations of orbital motion, Eq. (6.1), to an
appropriate frame rotating at a constant angular velocity. The essence of a non-
spherical gravity model is to find such a coordinate frame where both an orbital
stability analysis and an optimal control solution can be conveniently carried out.
A prime example of a non-spherical gravity field is the one created by multiple
bodies influencing the flight of a spacecraft. In such a case, the gravitational
acceleration, g, experienced by the spacecraft—regarded as a test mass—at any
instant depends upon the orbital dynamics of the primary bodies (e.g., the Sun and
a planet, or a planet and its Moon). The simplest multi-body gravity model is that of
spaceflight in the vicinity of two spherical bodies orbiting each other with a constant
angular velocity, and is referred to as the restricted three-body problem (R3B) where
the spacecraft’s mass is neglected in comparison to the masses of the two primaries.
A special case of the R3B model is that of the circular orbit of the two primaries
about a common barycenter, and is called the circular restricted three-body problem
(CR3BP). The derivation of a time-invariant CR3BP gravity model is carried out in
a rotating frame attached to the two primary bodies, which give rise to equilibrium
points of the CR3BP model in the rotating frame. The equilibrium points can be
utilized for stability analysis of the restricted three-body orbits. An orbital transfer
problem can then be posed as an optimal transfer trajectory between two (or more)
equilibrium points of the transformed system. Similarly, an optimal control strategy
can be devised for stationing the spacecraft close to an equilibrium point in the
rotating coordinate frame.
The gravity field of an irregularly shaped body (such as an asteroid) rotating
uniformly about a body axis is carried out in a body-fixed frame, which renders the
gravity field time invariant. A series approximation for the gravitational potential
6.2 Gravity Field of a Non-spherical Body 201
is then carried out using the transformed spatial coordinates, and coefficients are
determined by fitting with the observational data obtained by an orbiting spacecraft
equipped with accelerometers. Either a spherical harmonics expansion or a finite-
element model (such as a polyhedron model using cubic elements) can be employed
for deriving the gravitational potential in the rotating frame, whose gradients yield
time-invariant g and ∂g/r. As in the case of the CR3BP model, the equilibrium
points in the rotating frame are used for stability analysis of an orbit in the vicinity
of the body. An orbital transfer around the irregularly shaped body can be posed
as an optimal transfer trajectory between two (or more) equilibrium points in the
rotating frame. Another optimal control strategy can be devised for maintaining
desired orbit around the body.
This chapter introduces the essential methodology necessary to model the non-
spherical gravity fields for application to optimal spaceflight trajectories. Due to the
advanced nature of these topics, it would be beyond the present scope to go deeply
into their modeling details. Instead, the reader is referred to specialized textbooks
and research articles which cover those topics in detail, and focus is maintained on
the optimal spaceflight applications.
6.2 Gravity Field of a Non-spherical Body
The gravity field produced by any heavenly body is non-spherical due to its irregular
shape and rotation. The larger bodies such as the planets and large moons, due to
their gravitational mass accretion, display a nearly axisymmetric mass distribution
with only a slight longitudinal dependence. On the other hand, smaller bodies such
as the asteroids and comets depart from any regularity in their mass distribution,
thereby producing a highly uncertain gravity field at any given point. Modeling the
gravity of such bodies is thus a challenging task.
6.2.1 Gravity of an Axisymmetric Planet
A large heavenly body (referred to here by the generic term planet) can be regarded
as rotating about an axis of symmetry, but departs from the perfect symmetry of
spherical shape due to various reasons. One such departure—called oblateness—is
due to the centrifugal mass displacement caused by rotation, causing a bulge at the
equator, and flattening near the poles. The net departure from the spherical shape
can be approximated as a linear superposition of all such deforming effects, called
spherical harmonics. Being a conservative force, gravity is expressed as the gradient
of a scalar potential function, U (r), at a point, r, from the planetary center:
∂U
gT = (6.8)
∂r
Considering the central body’s mass to be distributed over a large number, (N −

1), of elemental masses, mi (i = 1, . . . , N − 1), the net gravitational acceleration
acting on the N th body (spacecraft) of mass m is the vector sum of the gravitational
accelerations caused by the individual masses,
#
N −1
∂Ui
gT = (6.9)
∂ri
i=1
where ri is the position of the test mass, m, from the i th point mass, mi . This
approach also yields the N-body problem where the spacecraft of mass m is under
the gravitational influence of (N − 1) large bodies.
For the purpose of modeling the motion of the spacecraft in the vicinity of a large
planet, the test mass, m, is considered to be negligible in comparison with the sum of
remaining (N − 1) masses, which are densely clustered together, away from the test
mass, the partial derivative can be moved outside the summation sign in Eq. (6.9),
resulting in
N −1
∂ # Ui
g =
T
(6.10)
∂ri ri
i=1
where Ui is the gravitational potential due to the point mass, mi , given by
Gmi
Ui = (6.11)
ri
with G being the universal gravitational constant. 4 −1

For all the (N − 1) particles constituting the planetary mass, M = N
i=1 mi , we
take the limit of an infinitesimal elemental mass, mi → dM and N → ∞, whereby
the summation in Eq. (6.10) is replaced by the following integral:

∂ G
gT = dM (6.12)
∂r s
where s is the distance of the test mass, m, from the elemental mass, dM, and can
be expressed as follows:
,
s= r 2 + ρ 2 − 2rρ cos γ (6.13)
with r and ρ being the position vectors of the test mass, m, and elemental mass,
dM, respectively, from the center of mass of the planet, and γ , is the angle between
r and ρ Hence,
r=s+ρ (6.14)
and ρ is a constant, because the planet is regarded to be a rigid body. Therefore, the
gravitational potential of the planetary mass distribution is given by

GdM
U= . (6.15)
r 2 + ρ 2 − 2rρ cos γ
In order to carry out the integration in Eq. (6.15), it is convenient to expand the
integrand in a series (with r > ρ) as follows:

1 1 ρ 1 ρ
. = 1 + cos γ + ( )2 (3 cos2 γ − 1)
r 2 + ρ 2 − 2rρ cos γ r r 2 r

1 ρ
+ ( )3 cos γ (5 cos2 γ − 3) + · · · (6.16)
2 r
or,
1 1 ρ ρ
. = P0 (cos γ ) + P1 (cos γ ) + ( )2 P2 (cos γ )
r 2 + ρ 2 − 2rρ cos γ r r r
ρ
+ · · · + ( )n Pn (cos γ ) + . . . (6.17)
r
where Pn (ν) is the Legendre polynomial of degree n. The first few Legendre
polynomials are the following:
P0 (ν) = 1
P1 (ν) = ν
1
P2 (ν) = (3ν 2 − 1) (6.18)
2
1
P3 (ν) = (5ν 3 − 3ν)
2
1
P4 (ν) = (35ν 4 − 30ν 2 + 3)
8
1
P5 (ν) = (63ν 5 − 70ν 3 + 15ν)
8
Some important properties of Legendre polynomials are as follows:
Pn (1) = 1
Pn (−1) = (−1)n (6.19)
Pn (−ν) = (−1)n Pn (ν)
A Legendre polynomial can be generated from those of lower degree with the help
of recurrence formulae, such as
(2n − 1)νPn−1 (ν) − (n − 1)Pn−2 (ν)

Pn (ν) = (6.20)
n
Other recursive formulae, called generating functions, such as the following [1]:
1 dn 2
Pn (ν) = (ν − 1) (6.21)
2n n! dν n
can also be useful in determining the Legendre polynomials. From these properties,
it is clear that the Legendre polynomials satisfy the condition | Pn (cos γ ) |< 1,
which implies that the series in Eq. (6.17) is convergent. Therefore, one can
approximate the integrand of Eq. (6.15) by retaining only a finite number of terms
in the series. The exact gravitational potential is expressed as follows in a series of
spherical harmonics:
∞
G# ρ
U= ( )n Pn (cos γ )dM (6.22)
r r
n=0
It is possible to further simplify the gravitational potential, before carrying out

the complete integration. The integral arising out of P0 (cos γ ) in Eq. (6.22) yields
the mass, M, of the planet; hence, we have

G GM
P0 (cos γ )dM = (6.23)
r r
Furthermore, the integral containing P1 (cos γ ) vanishes, because the axis ρ = 0
is the axis of symmetry of the planet. Therefore, the gravitational potential can be
expressed in terms of the Legendre polynomials of the second and higher degrees as
! ∞
"
GM # ρ n dM
U= 1+ ( ) Pn (cos γ ) (6.24)
r r M
n=2
A further simplification of the gravitational potential requires the planetary mass

distribution.
Consider a right-handed triad, (i, j, k), representing an inertial frame with origin
at the center of an oblate spheroid of equatorial radius, r0 , such that (i, j) is the
equatorial plane, and k locates the north pole. Due to the planetary symmetry about
k, the gravitational potential depends only upon the radius, r, from the planetary
center and the co-latitude, φ = π/2 − δ, measured from k, as shown in Fig. 6.1.
Let the planet’s mass density be given by D(ρ, β, λ) in terms of the spherical
coordinates of Fig. 6.1. Hence, the elemental mass can be expressed as
dM = D(ρ, β, λ)ρ 2 sin βdρdβdλ (6.25)

f
b •dM
Equatorial
r0
r
r •
m ir
Plane j
l
if
i q
Fig. 6.1 Geometry for the non-spherical gravity model
where β and λ are the co-latitude and longitude, respectively, of the elemental mass.
The assumption of symmetry about the polar axis translates into neglecting any
longitudinal dependence of the mass distribution; thus, we have
dM = D(ρ, β)ρ 2 sin βdρdβdλ (6.26)
Using the spherical trigonometry [85] of Fig. 6.1, we can write
cos γ = cos β cos φ + sin β sin φ cos(θ − λ) (6.27)
where φ and θ are the co-latitude and longitude, respectively, of the test mass,
m, as shown in Fig. 6.1. Upon neglecting the longitudinal variations in the mass
distribution, we have from Eq. (6.27)
Pn (cos γ ) = Pn (cos β)Pn (cos φ) (6.28)
resulting in
∞
μ # An
U (r, φ) = + Pn (cos φ) (6.29)
r r n+1
n=2
where μ = GM and

An = G D(ρ, β)ρ n+2 Pn (cos β) sin βdρdβdλ (6.30)
A more useful expression for the gravitational potential can be obtained as follows
in terms of the non-dimensional distance, rr0 (where r0 is the equatorial radius of the
planet):
& ∞
/
μ # r0 n
U (r, φ) = 1− Jn Pn (cos φ) (6.31)
r r
n=2
where
An
Jn = − (6.32)
μr0 n
are called Jeffery’s constants, and are unique for a planet. Jeffery’s constants
represent the spherical harmonics of the planetary mass distribution, and diminish
in magnitude as the order, n, increases. The largest of these constants, J2 , denotes a
non-dimensional difference between the moments of inertia about the polar axis,
k, and an axis in the equatorial plane (i or j in Fig. 6.1), and is a measure of
oblateness of the planet. The higher-order term, J3 , indicates the pear-shaped or
triangular harmonic, whereas J4 and J5 are the measures of square and pentagonal
shaped harmonics, respectively. For a reasonably large body, it is seldom necessary
to include more than the first four Jeffery’s constants. For example, the Earth’s
spherical harmonics are given by J2 = 0.00108263, J3 = −0.00000254, and
J4 = −0.00000161.
The acceleration due to gravity of a non-spherical, axisymmetric planet can be
obtained according to Eq. (6.12) by taking the gradient of the gravitational potential,
Eq. (6.31), with respect to the position vector, r = rir + rφiφ , as follows:
∂U T ∂U ∂U
g = −( ) =− ir − iφ (6.33)
∂r ∂r r∂φ
or,
g = gr ir + g φ iφ (6.34)
where
μ r0 r0
gr = − 1 − 3J2 ( )2 P2 (cos φ) − 4J3 ( )3 P3 (cos φ)
r2 r r
r0 4
− 5J4 ( ) P4 (cos φ) (6.35)
r
and

3μ r0 2 1 r0
gφ = 2 ( ) sin φ cos φ J2 + J3 ( ) sec φ(5 cos2 φ − 1)
r r 2 r

5 r0
+ J4 ( )2 (7 cos2 φ − 3) (6.36)
6 r
The unit vectors ir and iφ denote the radial and southward directions in the local
horizon frame attached to the test mass (Fig. 6.1). Due to a nonzero transverse
gravity component, gφ , the direction of g differs from the radial direction, while
its radial component, gr , is smaller in magnitude compared to that predicted by a
spherical gravity model.
6.2.2 Gravity Field of an Oblate Planet
In many flight applications close to a planet, only the oblateness effect need to be
modeled. This greatly simplifies the non-spherical model, as well as the resulting
optimal control application. From the series expansion of the gravitational potential,
Eq. (6.31), truncated to the largest Jeffery’s constant, J2 , we have
r 2
μ 0
U (r, φ) = 1− J2 P2 (cos φ) (6.37)
r r
This yields the following acceleration due to gravity at the given point, (r, φ):
T
∂U μ 2 3z 3 15z2
g= =− r − μJ r
2 0 k + − r (6.38)
∂r r3 r5 2r 5 2r 7
with z = r cos φ. The gravity-gradient matrix required in the primer vector

integration (see Chap. 5) is then evaluated as follows:
! " ! "
∂g μ 3μJ2 r02 5z2 3μ r02 7z2 ∂r
=− 3+ 5
1− 2 I+ 4
−15μJ2 6 2
−1 r
∂r r 2r r r 2r r ∂r

z ∂r r2 ∂z
+ 15μJ2 r02 6 k − 3μJ2 05 k (6.39)
r ∂r r ∂r

z ∂z
+ 15μJ2 r02 7 r
r ∂r
where
∂r
= rT /r (6.40)
∂r
∂z
= kT (6.41)
∂r
A substitution of Eqs. (6.40) and (6.41) into Eq. (6.39) yields the following
expression for the gravity-gradient matrix:
! "
∂g μ 3μJ2 r02 5z2
=− 3 + 1− 2 I
∂r r 2r 5 r
! "
3μ r02 7z2 z
+ − 15μJ2 7 − 1 rrT + 15μJ2 r02 7 krT (6.42)
r5 2r r2 r

r02 5z T
+ 3μJ2 5 rk − kk T
r r2
6.2.3 Gravity of an Irregular Body
Irregular bodies such as asteroids and comets do not have an axisymmetric shape,
and thus the longitudinal dependence of the gravity potential in Eq. (6.24) cannot be
neglected. The rotation of the body about a principal axis therefore results in a time-
varying gravity field at any given point, (r, φ, θ ). Deriving an analytical expression,
such as the one given by Eq. (6.33), is generally not feasible by the spherical
harmonics expansion method presented above for the axisymmetric planet, because
a detailed density distribution, D(ρ, β, λ), is unavailable. Furthermore, the spherical
harmonics model, which converges at a point external to the body, may not converge
inside a circumscribing sphere centered on the body [62]. The third disadvantage
of the spherical harmonics model is that it cannot detect whether a point is located
outside or inside the irregularly shaped body. Thus a problem in which a soft landing
is to be carried out cannot be studied with a spherical harmonics model.
Instead of the spherical harmonics model, a more useful method of deriving
a gravity model of an irregularly shaped body is by the polyhedral approxima-
tion [84]. This method derives an analytical gravity field by approximating the
body as a constant-density polyhedron, which can have peaks, overhangs, cavities,
caves, and through holes. Hence there is no penalty associated with detailed surface
features, which would require a high-order harmonics model. Furthermore, the
model is applicable right up to the surface of the body, and can detect whether a point
is located either inside or outside the body. Therefore, the polyhedral gravity model
is especially useful for deriving optimal landing strategies on asteroids and comets.
The successful orbiting and landing of the comet 67P/Churyumov-Gerasimenko by
the ESA Rosetta probe and Philae lander mission in 2014 demonstrated the practical
application of a polyhedral gravity model. Yang et al. [87] employ the polyhedral
model to devise time-optimal powered descent and landing strategies on an asteroid.
6.3 Circular Restricted Three-Body Problem
The three-body problem refers to the motion of three masses under mutual grav-
itational attraction. It arises when either the two-body orbital motion is perturbed
by a distant third body or a much smaller body (like a spacecraft) moves in the
6.3 Circular Restricted Three-Body Problem 209
gravitational field formed by two larger bodies. Examples of the former kind are
the orbital motion of a Moon around a planet perturbed by the Sun, and that of a
satellite around the Earth perturbed by the Sun or Moon. The latter model is applied
to a comet under the combined gravity of the Sun and Jupiter, a spacecraft under
mutual gravity of the Earth–Moon system, and an interplanetary probe approaching
(or departing) a planet. The three-body problem is a much better approximation
of the actual motion of a spacecraft at any given time in the solar system when
compared to the two-body problem considered earlier. An important application
of the three-body problem is the design of optimal circumlunar and interplanetary
trajectories based upon the solutions that take the spacecraft around multiple large
bodies on largely natural, fuel-free paths. These will be explored later in this chapter.
The three-body problem has attracted attention of mathematicians and physicists
over the past three centuries, primarily in studying the motion of the Moon. The
first systematic study of the three-body problem was by Lagrange, who derived
the particular solution of the coplanar problem, and investigated the stability of
the equilibrium solutions. The equations of motion for the three-body problem are
d2 Ri # mj3
2
=G rij ; (i = 1, 2, 3) (6.43)
dt rij3
j =i
where G is the universal gravitational constant, Ri denotes the position of the center
.
of mass of the i th body, and rij = Rj − Ri denotes the relative position of the centers
of mass of bodies, i,j (i = j ).
Lagrange showed that certain particular solutions of the three-body problem exist
when the motion of the three bodies is always confined to a single plane. Each mass
in the coplanar three-body problem traces a conic section about the common center
of mass (called the barycenter). From the coplanar motion derived by Lagrange,
certain stationary solutions can be obtained in terms of fixed locations of the masses
in the rotating frame. These solutions represent equilibrium points of the three-
body problem, and describe the three bodies moving in concentric, coplanar circles.
Such a motion has a constant separation of the three bodies, and the line joining
any two bodies rotates at a constant angular speed, n = θ̇, by angular momentum
conservation. The equations of motion of the coplanar three-body problem are the
following:
0 1
n2 m2 m3 m2 m3
− 3 − 3 R1 + 3 R2 + 3
R3 = 0
G r12 r13 r12 r13
0 1
m1 n2 m1 m3 m3
3
R1 + − 3 − 3 R2 + 3
R3 = 0 (6.44)
r12 G r12 r23 r23
m1 R1 + m2 R2 + m3 R3 = 0
Particular equilibrium solutions of Eq. (6.44) include the equilateral triangle

configurations of the three masses, wherein the masses are at the same constant
distance from the common center of mass (r12 = r13 = r23 = R), hence the angular
speed is given by
-
G(m1 + m2 + m3 )
n = θ̇ = (6.45)
R3
Another set of equilibrium points are the collinear solutions, where the three
masses share a straight line. Let the axial location of the three masses be given by
x1 , x2 , x3 , where we assume x1 < x2 < x3 . Then the equations of motion, written
in terms of a distance ratio, α = r23 /r12 , yield the following [8]:
m2 + (1 + α)m3
x1 = −r12
m1 + m2 + m3
G(m1 + m2 + m3 ) m2 (1 + α)2 + m3
ω2 = (6.46)
3 (1 + α)2
r12 m2 + (1 + α)m3
(m1 + m2 )α 5 + (3m1 + 2m2 )α 4 + (3m1 + m2 )α 3

− (m2 + 3m3 )α 2 − (2m2 + 3m3 )α − (m2 + m3 ) = 0
The last of Eq. (6.46), called the quintic equation of Lagrange, has only one positive
root, α, which can be seen by examining the signs of the coefficients (they change
sign only once). However, for each specific configuration of the three collinear
masses, the positive value of α is different.
For example, the values of the distance ratio, α, can be numerically determined
for the collinear Earth, Moon, spacecraft system by taking the ratio of the Moon’s
mass to the Earth’s mass is 1/81.3 and neglecting the spacecraft’s mass in compari-
son with the masses of heavenly bodies. These are α = 0.1678, 0.1778, or 1.0071,
depending upon whether the Moon, the spacecraft, or the Earth, respectively, falls
between the other two bodies.
When the mass of one of the three bodies, say m3 , is negligible in comparison
with that of the other two bodies (called the primaries) which latter are in a circular
orbit about the barycenter, a major simplification occurs in the coplanar three-
body equations (Eq. (6.44)). In that case, the gravitational pull of m3 on both m1
and m2 is negligible, and the natural motion of m3 relative to the primaries is
referred to as the circular restricted three-body problem (or CR3BP). The barycenter
is then approximated by the common center of mass of the two primaries, i.e.,
m1 R1 m2 R2 . For convenience, the CR3BP equations of motion are rendered
non-dimensional by dividing each of the masses by the total mass of the primaries,
m1 + m2 , and dividing all the distances by the constant separation between the
primaries, R12 .
. m2
μ=
m1 + m2
. r13
r1 = (6.47)
r12
. r23
r2 =
r12
Conventionally, we take m1 ≥ m2 , which makes μ ∈ [0, 1/2]. The normalized

masses of the primaries, m1 and m2 , are thus (1 − μ) and μ, respectively. The
non-dimensional distances of the two primaries from the barycenter are R1 = μ
and R2 = (1 − μ), respectively. The value of gravitational constant, G, is non-
dimensionalized such that the angular velocity, n, of the primaries is unity
, .
n = θ̇ = G(m1 + m2 )/r12
3 = μ + (1 − μ) = 1 (6.48)
which produces a non-dimensional time, t, such that the orbital period of the
primaries is t = 2π . Let the coordinates of the primaries m1 and m2 in an inertial
frame with origin at the barycenter be (ξ1 , η1 , 0) and (ξ2 , η2 , 0), respectively.
If the position of the negligible mass, m3 , is described by the non-dimensional
coordinates, (ξ, η, ζ ), they are governed by the following differential equations:
(1 − μ)(ξ − ξ1 ) μ(ξ − ξ2 )
ξ̈ = − −
r13 r23
(1 − μ)(η − η1 ) μ(η − η2 )
η̈ = − − (6.49)
r13 r23
(1 − μ)ζ μζ
ζ̈ = − −
r13 r23
where the overdot represents derivative in non-dimensional time, t, and

,
r1 = (ξ − ξ1 )2 + (η − η1 )2 + ζ 2
,
r2 = (ξ − ξ2 )2 + (η − η2 )2 + ζ 2 (6.50)
Now, consider a frame (i, j, k), with origin at the common center of mass
(barycenter) (Fig. 6.2), and rotating with the non-dimensional velocity, k. Such a
rotating frame is termed the synodic frame. The position of the mass, m3 , in the
rotating frame is given by r = xi + yj + zk. The coordinate transformation between
the inertial and the rotating frames is the following:
⎧ ⎫ ⎛ ⎞⎧ ⎫
⎨ξ ⎬ cos t − sin t 0 ⎨ x ⎬
η = ⎝ sin t cos t 0 ⎠ y (6.51)
⎩ ⎭ ⎩ ⎭
ζ 0 0 1 z
Orbital plane (i,j) of primaries
i
2
1
Fig. 6.2 Geometry of the circular restricted three-body problem
The substitution of Eq. (6.51) into Eq. (6.49) and elimination of sin t and cos t yields
the following:
(1 − μ)(x − x1 ) μ(x − x2 )
ẍ − 2ẏ − x = − −
r13 r23
(1 − μ)(y − y1 ) μ(y − y2 )
ÿ + 2ẋ − y = − − (6.52)
r13 r23
(1 − μ)z μz
z̈ = − −
r13 r23
Finally, the direction of the x-axis (i) is chosen to point from the barycenter toward
the center of the smaller primary, m2 , and the y-axis (j) being normal to both i
and k completes the right-handed triad, i × j = k. Thus we have x1 = −μ, x2 =
1 − μ, y1 = y2 = 0, and the following equations of CR3BP motion in terms of the
coordinates, (x, y, z):
(1 − μ)(x + μ) μ(x − 1 + μ)
ẍ − 2ẏ − x = − −
r13 r23
(1 − μ)y μy
ÿ + 2ẋ − y = − − (6.53)
r13 r23
(1 − μ)z μz
z̈ = − −
r13 r23
where
,
r1 = (x + μ)2 + y 2 + z2
,
r2 = (x − 1 + μ)2 + y 2 + z2 (6.54)
The set of coupled, nonlinear, ordinary differential equations, Eq. (6.53), repre-
senting the CR3BP cannot be solved in a closed form. However, we can obtain
certain important analytical insights into the problem without actually solving it.
An important aspect of the CR3BP equations is the symmetrical nature of its
solutions. If x(t), y(t), z(t) is a solution of the equations, then it is easily verified
that x(t), y(t), −z(t) is also a solution, thereby implying a reflection across the
(x, y) plane. Similarly, x(−t), −y(−t), z(−t) is also a solution, which implies a
reflection across the (x, z) plane with time reversal. The third symmetry is related
to the parameter, μ. If x(t), y(t), z(t) satisfies the equations with parameter μ, then
−x(t), −y(t), z(t) also satisfies the equations with parameter 1 − μ. This means
that one needs to study the CR3BP only for μ ∈ [0, 1/2].
6.3.1 Lagrangian Points and Stability
As noted earlier, certain equilibrium solutions are possible to the coplanar three-
body problem represented by Eq. (6.44). These equilibrium points are called the
Lagrangian points (or libration points), and can be derived for the CR3BP model
by equating the time derivatives in Eq. (6.53) to zero, resulting in the following
algebraic equations:
(1 − μ)(x + μ) μ(x − 1 + μ)
x= +
r13 r23
(1 − μ)y μy
y= + (6.55)
r13 r23
(1 − μ)z μz
0= +
r13 r23
(6.55), we have the result z.= 0 for all the Lagrangian points,
From the last of Eq. .
which implies r1 = (x + μ)2 + y 2 and r2 = (x + μ − 1)2 + y 2 . The coplanar
equilibrium solutions consist of the three possible collinear positions of m3 , called

the collinear Lagrangian points, L1 , L2 , L3 , as well as the two equilateral triangle
positions, called triangular Lagrangian points, L4 , L5 , with the two primaries
(Fig. 6.2). The points L1 , L2 , L3 are obtained from the solution of the quintic
equation of Lagrange (the last of Eq. (6.46)), which for m1 = 1 − μ, m2 = μ,
m3 = 0, and α = r23 /r12 is written as follows:
α 5 + (3 − μ)α 4 + (3 − 2μ)α 3 − μα 2 − 2μα − μ = 0 (6.56)
As discussed earlier, there is only one real root of Eq. (6.56). For example, with
μ = 1/82.3 for the Earth–Moon system, Eq. (6.56) gives α = 0.167833, which
yields x = 1.15568 for Lagrangian point L2 . With m1 = μ (Moon), m2 = 1 − μ
(Earth), and m3 = 0 (spacecraft), the quintic equation becomes the following:
α 5 + (2 + μ)α 4 + (1 + 2μ)α 3 − (1 − μ)α 2 − 2(1 − μ)α − (1 − μ) = 0 (6.57)
whose real root for the Earth–Moon system is α = 0.992912, thereby yielding
x = −α − μ = −1.00506 for the Lagrangian point L3 .
Finally, the Lagrangian point L1 is obtained by putting m2 = 0 (spacecraft),
m1 = μ (Moon), and m3 = 1 − μ (Earth). The quintic equation then becomes
μα 5 + 3μα 4 + 3μα 3 − 3(1 − μ)α 2 − 3(1 − μ)α − (1 − μ) = 0 (6.58)
whose real root for the Earth–Moon system is α = r23 /r12 = 5.62538553 with
r12 + r23 = 1, thereby yielding
1
x = r23 − μ = − μ = 0.836914719
1 + 1/α
√ L4 , L5 , correspond to r1 = r2 = 1, and are given

The triangular Lagrangian points,
by the coordinates ( 12 − μ, ± 3/2).
In order to investigate the stability of a Lagrangian point, consider infinitesimal
displacements, δx, δy, δz, from the equilibrium position, (x0 , y0 , 0). If the displace-
ments remain small, the equilibrium point is said to be stable, otherwise, unstable.
Such an analysis easily reveals that the out-of-plane motion represented by δz is
unconditionally stable about all the equilibrium points. However, the stability of the
coplanar motion described by (δx, δy) crucially depends upon the location (x0 , y0 )
of the given equilibrium point.
For the small displacement, coplanar motion about a triangular Lagrangian point,
say L4 , the following linearized model can be applied:
√
3 3 3(μ − 12 )
δ ẍ − 2δ ẏ − δx − δy = 0
4 2
√
3 3(μ − 12 ) 9
δ ÿ + 2δ ẋ − δx − δy = 0 (6.59)
2 4
whose general solution to initial displacement δx0 , δy0 , can be written as
δr = δr0 eλt (6.60)
where δr = (δx, δy)T , δr0 = (δx0 , δy0 )T , and λ is the eigenvalue of the following
matrix:
⎛ ⎞
0 0 1 0
⎜ 0 1⎟
⎜ 0 √
0 ⎟
A=⎜ ⎜
1 ⎟ (6.61)
0 2⎟
3 3 3(μ− 2 )
⎝ √ 4 1 2 ⎠
3 3(μ− 2 )
2
9
4 −2 0
The resulting characteristic equation for λ is thus the following:

27
λ4 + λ2 + μ(1 − μ) = 0 (6.62)
4
whose roots are
1 .
λ2 = [± 1 − 27μ(1 − μ) − 1] (6.63)
2
For stability, the values of λ must be purely imaginary, representing a simple
harmonic oscillation about the equilibrium point. If the quantity in the square-root
is negative, there is at least one value of λ with a positive real part and represents
an unstable system. Therefore, the critical values of μ representing the boundary
between stable and unstable behavior of L4 are those that correspond to
1 − 27μ(1 − μ) = 0 (6.64)
Hence, for the stability of the triangular Lagrangian points, we require either
μ ≤ 0.0385209 or μ ≥ 0.9614791, which is always satisfied by any two CR3BP
primaries in the solar system. Therefore, we expect that the triangular Lagrangian
points would provide stable locations for smaller bodies in the solar system. The
existence of such bodies for the Sun–Jupiter system has been verified in the form of
Trojan asteroids. For the Earth–Moon system, μ = 0.01215 is well within the stable
region, hence one could expect the triangular points to be populated. However, the
Earth–Moon system is not a good example of the restricted three-body problem,
because of an appreciable influence of Sun’s gravity (a fourth body), which renders
the triangular points of Earth–Moon system unstable. Hence, L4 , L5 for Earth–
Moon system are empty regions in space, where one cannot park a spacecraft.
For a given CR3BP system, one can determine the actual small displacement
dynamics about the stable triangular Lagrangian points, in terms of the eigenvalues,
λ. For example, the Earth–Moon system has non-dimensional natural frequencies
0.2982 and 0.9545, which indicate long-period and short-period modes of the small
displacement motion.
For the stability of the collinear Lagrangian points, we employ y0 = 0, and the
resulting equations of small displacement can be written as follows:

2(1 − μ) 2μ
δ ẍ − 2δ ẏ − + + 1 δx
(x0 − μ)3 (x0 + 1 − μ)3

1−μ μ
δ ÿ + 2δ ẋ + + − 1 δy (6.65)
(x0 − μ)3 (x0 + 1 − μ)3
The eigenvalues of this system are the eigenvalues of the following square matrix:
⎛ ⎞
0 0 1 0
⎜0 0 0 1⎟
A=⎜
⎝α 0 0
⎟ (6.66)
2⎠
0 −β −2 0
where
2(1 − μ) 2μ
α= + +1
(x0 − μ)3 (x0 + 1 − μ)3
1−μ μ
β= + −1 (6.67)
(x0 − μ) 3 (x0 + 1 − μ)3
The resulting characteristic equation for the eigenvalues, λ, is the following:
λ4 + (4 + β − α)λ2 − αβ = 0 (6.68)
For the collinear Lagrangian points, α > 0, and β < 0, which implies that the
constant term in the characteristic equation is always positive. Therefore, there
is always at least one eigenvalue with a positive real part, hence the system is
unconditionally unstable. Thus, the collinear Lagrangian points of any restricted
three-body system are always unstable. However, as discussed later, a spacecraft
can successfully orbit a collinear point with small energy expenditure. Such an orbit
around a Lagrangian point is termed a halo orbit , and is useful in practical space
missions, such as stationing a probe around the Sun–Earth L1 and L2 points for
monitoring the solar and interplanetary zone ahead of and behind the Earth.
6.3.2 Hamiltonian Formulation and Jacobi’s Integral
The CR3BP equations of motion can be alternatively derived by the Lagrangian

approach using the canonical coordinates, q = (x, y, z)T , as follows:
q̈ − 2Kq̇ = ∇Ω (6.69)
where ∇ = (∂/∂x, ∂/∂y, ∂/∂z)T is the gradient operator

⎛ ⎞
0 1 0
K = ⎝ −1 0 0 ⎠ (6.70)
0 0 0
and
1 μ (1 − μ) 1 (1 − μ) μ
Ω = − qT K2 q + + = (x 2 + y 2 ) + +
2 | q − (1 − μ)i | | q + μi | 2 r1 r2
(6.71)
is the pseudo-potential function with i = (1, 0, 0)T and r1 , r2 given by Eq. (6.54).
The scalar components of Eq. (6.69) are the following:
∂Ω
ẍ − 2ẏ =
∂x
∂Ω
ÿ + 2ẋ = (6.72)
∂y
∂Ω
z̈ =
∂z
The Hamiltonian associated with Eq. (6.69) is the following:
1
H (q, q̇) = − q̇T (q̇ − Kq) − L(q, q̇) (6.73)
2
with the Lagrangian given by
1
L(q, q̇) = (q̇ − Kq)T (q̇ − Kq) − U (q) (6.74)
2
and the gravitational potential by
1 (1 − μ) μ
U (q) = −Ω − qT K2 q = − − (6.75)
2 r1 r2
It is to be noted that the terms “Hamiltonian” and “Lagrangian” used here in the
dynamical context are different from the same terms occurring in the optimal context
(Chap. 2). By defining the generalized momentum vector as follows:
∂L
p= = q̇ − Kq (6.76)
∂ q̇
the Hamiltonian is expressed as
1 T
H (q, p) = p p + pT Kq + U (q) (6.77)
2
This yields the equations of motion, Eq. (6.69):
∂H
q̇ =
∂p
∂H
ṗ = − (6.78)
∂q
which are compactly expressed in terms of the state vector, X = (qT , pT )T , as

follows:
Ẋ = J∇X H (X) (6.79)
where ∇X is the gradient operator relative to X, and

03×3 I3×3
J= (6.80)
−I3×3 03×3
An important scalar constant of the restricted three-body motion can be derived

by taking the scalar product of Eq. (6.69) with q̇:
∂Ω T
q̈ · q̇ − 2(Kq̇) · q̇ = · q̇ (6.81)
∂q
or
1 dq̇ 2 dΩ
= (6.82)
2 dt dt
which is an exact differential, and can be integrated to obtain
1 2
Ω= (q̇ + C) (6.83)
2
Here, C is the constant of integration, called Jacobi’s integral:
1−μ μ
C = 2Ω − q̇ 2 = (x 2 + y 2 ) − 2 − 2 − (ẋ 2 + ẏ 2 + ż2 ) (6.84)
r1 r2
The Hamiltonian, H , is related to Jacobi’s integral as follows:
1 T
H = p p + pT Kq + U (q)
2
1 1
= q̇ 2 + qT K2 q + U (q) (6.85)
2 2
1 2 1 2
= q̇ − (x + y 2 ) + U = −C/2
2 2
6.4 Special Three-Body Trajectories 219
Jacobi’s integral can be thought to represent a pseudo-energy of the mass m3 ,

which is a sum of the relative kinetic energy, gravitational potential energy, and
an additional pseudo-potential energy, (x 2 + y 2 ). The value of C at any point is a
measure of its relative energy. The most useful interpretation of Jacobi’s integral
lies in the contours of zero relative speed, v = 0, which represent the boundary
between a motion toward higher potential, or the return trajectory to the lower
potential [78]. On such a boundary, the mass must stop and turn back. Hence, the
zero relative speed contours demarcate the regions accessible to a spacecraft, and
cannot be crossed for the given value of C. The closed zero speed contours around a
primary indicate that a flight from one primary to the other is impossible at the given
energy level. As the value of C increases, the contours expand in size touch one
another for some special values of C, and a further increase of C causes the region
around a collinear point to open up, indicating either a flight at nonzero relative
velocity between the primaries or an escape from the gravitational influence of the
primaries [78]. In such a case, contours enclose the triangular Lagrangian points,
indicating that a flight to reach them is impossible from any of the primaries. As
the value of C becomes very large, the forbidden regions around L4 , L5 , shrink, but
never actually vanish, thereby denoting that the stable triangular Lagrangian points
can be approached (but never reached) only with a very high initial energy.
6.4 Special Three-Body Trajectories
Although the circular restricted problem of three bodies (CR3BP) appears more
amenable to a solution than the general problem due to the availability of Jacobi’s
integral, a closed-form solution is still not possible. This fact was first demonstrated
by Poincaré using phase-space surfaces. The CR3BP problem is mathematically
interesting as it belongs to the group of chaotic systems, wherein a small change in
the initial condition results in an arbitrarily large change in the trajectory. Numerical
solutions to the CR3BP problem for special cases reveal special trajectories that are
possible flight paths for spacecraft in lunar and interplanetary space. Such solutions
rarely have a periodic behavior, which requires special integration schemes, such
as a higher-order Runge–Kutta technique, to prevent the truncation errors from
growing to unacceptable magnitudes within a few orbital time periods of the
primaries. Here some typical trajectories are explored which can be derived within
one period of primary rotation.
6.4.1 Perturbed Orbits About a Primary
When a spacecraft is orbiting one of the primaries, its trajectory is perturbed by

the presence of the more distant primary, causing a precession in the plane of the
orbit. This leads to both an apsidal rotation and a nodal regression of the two-body
orbital plane, thereby producing a quasi-periodic trajectory. Consider, for example,
0.04
0.02
0
m1 t =0
−0.02
−0.04
y
−0.06
−0.08
−0.1
−0.12 t =1
−0.14
−0.04 −0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
x
Fig. 6.3 Trajectory around the Earth perturbed by the Moon simulated for t = 1 (4.8 days)
the apsidal rotation of a spacecraft orbiting the Earth due to the Moon’s gravity,
shown in Fig. 6.3, when the three bodies are in the same plane. The value of the non-
dimensional mass parameter for the Earth–Moon system is μ = 0.01215, which
implies the Earth’s location at (−0.01215, 0) marked by the red dot in Fig. 6.3. The
trajectory is simulated for t = 1 (recall that a time period of the primaries in the
non-dimensional time is t = 2π , or about 30 mean solar days). The initial condition
for the given coplanar trajectory is x(0) = 0.15, y(0) = 0, ẋ = 0, and ẏ = 0.45.
The elliptical shape of the trajectory is evident in this planar problem.
Figure 6.4 shows the apsidal rotation of a spacecraft’s orbital plane around the
Moon caused by the Earth. The initial condition for the given coplanar trajectory
plotted for t = 1 is x(0) = 1.0, y(0) = 0, ẋ = 0, and ẏ = 0.4. The perturbed orbits
shown in Figs. 6.3 and 6.4 are special quasi-periodic solutions of the CR3BP in the
vicinity of a collinear Lagrangian point.
6.4.2 Free-Return Trajectories Between the Primaries
An interesting periodic solution of the CR3BP is the free-return trajectory, which

passes close to both the primaries, and can be used for an essentially fuel-free
transfer from one primary to the other. The free-return concept formed the basis of
Jules Verne’s famous novel From the Earth to the Moon (1866), and was ultimately
realized about a century later by the manned lunar exploration through the Apollo
lunar missions, wherein the initial insertion was performed at a given location with
the desired velocity, both on the outward and the return journey. Being a natural
6.4 Special Three-Body Trajectories 221
−3
x 10
4
t
2
0
m2
−2
y
−4
−6
−8
0.984 0.986 0.988 0.99 0.992 0.994 0.996 0.998 1
x
Fig. 6.4 Trajectory around the Moon perturbed by the Earth simulated for t = 1 (4.8 days)
0.8
0.6
0.4
0.2
m1 L1 m2 L2
0
y
−0.2
−0.4
−0.6
−0.8
−1
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2
x
Fig. 6.5 Free-return trajectory passing close to the Earth and the Moon in t = 3.5 (16.7 days)
solution, it is a fuel-optimal path of the planar CR3BP. An example of the free-

return Earth–Moon trajectory is plotted in Fig. 6.5 for t = 3.57, which extends from
45,000 km radius near the Earth (green circle) to a point close to the Moon (red
circle), and orbits the Lagrangian point, L1 . Another possible free-return trajectory
0.5
0.4
0.3
0.2
0.1
m1 L2
L1 m2
y
−0.1
−0.2
−0.3
−0.4
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
x
Fig. 6.6 Hourglass free-return trajectory passing close to the Earth and the Moon in t = 2.387
(11.4 days)
is the hourglass shaped one (Fig. 6.6) passing between the second primary, m2 ,
and the Lagrangian point, L2 . The total flight time is reduced significantly in this
case to about t = 2.387, because the spacecraft passes around the Moon, between
L1 and L2 , and in the process, receives a kinetic energy boost due to the Moon.
Such an increase in the Keplerian orbital energy derived by passing close to a third
body is called a gravity-assist maneuver (or a swing-by). An extreme case of the
lunar swing-by trajectory is depicted in Fig. 6.7, where the spacecraft escapes the
Earth–Moon system by passing very close to the Moon. The initial launch speed
(from the same initial location) for the escape trajectory of Fig. 6.7 is smaller
(v = 6.935, C = 2.3982) than that of the free-return trajectory of Fig. 6.6, which
has v = 6.9493 and C = 2.1995. Lunar swing-by trajectories have been employed
in cheaply boosting several spacecraft to the Sun–Earth Lagrangian points, such as
the ISEE-3, MAP, and ACE missions of NASA [27]. A similar approach of multiple
planetary swing-bys has been found useful in reducing the mission cost of spacecraft
bound for distant planets. Figure 6.8 shows a trajectory beginning near the Moon,
and orbiting all five Lagrangian points of the Earth–Moon system in t = 28.5.
6.5 Optimal Low-Thrust Three-Body Transfer
Navigating a spacecraft by continuous thrust in the CR3BP gravity field is a practical

problem of space navigation with applications in both lunar and interplanetary
flight. Consider a spacecraft with a continuously operating, low-thrust engine
6.5 Optimal Low-Thrust Three-Body Transfer 223
1
y
0 L3 L1 L2
−1
−2
−3
−3 −2 −1 0 1 2 3 4
x
Fig. 6.7 Lunar swing-by escape trajectory simulated for t = 10 (47.76 days)
1.5
1
L4
0.5
0 L3 L1 L2
y
−0.5
−1 L5
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x
Fig. 6.8 Free trajectory orbiting all five Lagrangian points of the Earth–Moon system in t = 28.5
(136 days)
generating an acceleration input, a = a(t)n(t). The non-dimensional state equations

representing the spacecraft’s position, r = xi + yj + zk and velocity v = vx i +
vy j + vz k in the rotating frame, (i, j, k), with origin at the common center of mass
(barycenter) (Fig. 6.2) of the circular, restricted three-body problem (CR3BP) can
be derived from Eq. (6.53) to be the following:
⎧ ⎫ ⎧ ⎫
⎨ ẋ ⎬ ⎨ vx ⎬
ṙ = ẏ = vy = v
⎩ ⎭ ⎩ ⎭
ż vz
⎧ ⎫ ⎧ ⎫
⎪ v̇ ⎪ ⎪
⎪ x + 2vy − (1−μ)(x+μ) − μ(x−1+μ) + ax ⎪
⎪
⎨ ⎪
⎪ x
⎬ ⎪ ⎨ r1
3 r2
3
⎪
⎬
(1−μ)y μy
v̇ = v̇y = y − 2vx − 3 − 3 + ay (6.86)
⎪
⎪ ⎪ ⎪ r1 r2 ⎪
⎩ ⎪ ⎭ ⎪ ⎪
⎩
⎪
⎪
⎭
v̇z v̇z − (1−μ)z
3 − μz3 + az
r1 r2
where n = x i + y j + z k is the unit vector in the thrust direction, and

,
r1 = (x + μ)2 + y 2 + z2 (6.87)
,
r2 = (x − 1 + μ)2 + y 2 + z2
The thrust direction cosines, (x , y , z ), must satisfy the equality constraint
2x + 2y + 2z = 1 (6.88)
Both the velocity, v, and the thrust acceleration input, a, can alternatively be
resolved in spherical coordinates as shown in Fig. 6.9 via the angles, (φ, ψ) and
(α, β), respectively, through the following relationships:
Fig. 6.9 Resolving the

velocity and thrust
acceleration in the CR3BP a
frame using spherical
coordinates
a
v
y b
v = v(cos φ cos ψi + cos φ sin ψj + sin φk) (6.89)

a = a[cos(φ + α) cos(ψ + β)i + cos(φ + α) sin(ψ + β)j + sin(φ + α)k]
,
which implies φ = tan−1 (vz / (vx2 + vy2 ) and ψ = tan−1 (vy /vx ).
Irrespective of whether the Cartesian or spherical coordinates are used to write
the state equations, the position, velocity, and acceleration require a transformation
from the rotating frame, (i, j, k), to an inertial (sidereal) frame, (I, J, k). Further-
more, the velocity and acceleration must be corrected as follows to yield the
corresponding inertial values:
v ⇒ v+ω×r
v̇ ⇒ v̇ + ω × v (6.90)
One can use the state equations in the rotating frame to compute the optimal
trajectory and control history in the rotating frame, and then transform them to the
inertial frame. Thus for the purpose of control derivation, Eq. (6.86) can be regarded
as the state equations, and expressed in the form
Ẋ = f(X, a) (6.91)
where X(t) = (rT , vT )T is the state vector, a(t) ∈ R3×1 is the acceleration input
vector, and the functional f(.) : R6×1 × R3×1 → R6×1 is given by Eq. (6.86). A
maneuver can be represented by the initial state,
X(0) = X0 (6.92)
and a terminal condition
F [X(tf ), tf ] = 0 (6.93)
with the terminal time, tf , either specified in advance or not. The optimization
problem is of the Mayer type, and can be stated as finding the control history which
minimizes the following performance index:
J = φ[X(tf ), tf ] (6.94)
subject to the equality constraints given by Eqs. (6.86)–(6.88). The optimal maneu-
ver is the one that minimizes the propellant consumption, which implies the
minimization of the terminal characteristic speed (φ = c(tf )), and is equivalent
to the maximization of the final mass (φ = −m(tf )) (see Chap. 5). In either case,
an additional scalar state equation must be added, such as
ċ = a (6.95)
or
ṁ = −β (6.96)
with bounds being applied either to the acceleration input magnitude
0 ≤ a ≤ am (t) (6.97)
or to the exhaust rate
0 ≤ β ≤ βm (t) (6.98)
depending upon the type of the engine employed. Let us consider the bounded input
magnitude formulation, for which we have φ = c(tf ), and the Hamiltonian of the
optimal control problem is the following:
$ %
H = λTr v + λTv g
+ an + λc a (6.99)
where λr and λv are the costate vectors corresponding to r and v, respectively, and
λc corresponds to the characteristic speed, c(t), obeying Eq. (6.95). Here, the non-
inertial acceleration due to gravity, g
, in the rotating frame is the following:
⎧ ⎫
⎪
⎪ x + 2vy − (1−μ)(x+μ) − μ(x−1+μ) ⎪
⎪
⎨ r13 r23 ⎬

(1−μ)y μy
g = y − 2vx − 3 − 3 (6.100)
⎪
⎪
r1 r2 ⎪
⎪
⎩ v̇ − (1−μ)z − μz ⎭
z r13 r23
The costate equations for λr and λv are derived from the necessary conditions of
optimality as follows:
T T
∂H ∂g
λ̇r = − =− λv (6.101)
∂r ∂r
T
∂H
λ̇v = − = −λr (6.102)
∂v

∂H ∂H
λ̇c = − = −a (6.103)
∂c ∂t
where the non-inertial gravity-gradient matrix, ∂g

/∂r, is given by:
⎛ ⎞
Ax 0 0
∂g
⎜ ⎟
=⎜
⎝ 0 Ay 0 ⎟
⎠ (6.104)
∂r
0 0 Az
with

(1 − μ) (x + μ)2 μ (x − 1 + μ)2
Ax = 1 + −1 + 3 +1
r13 2 r2 2
2 ! "
y (1 − μ) μ
Ay = 1 − +1 + 3 (6.105)
2 r13 r2
2 ! "
z (1 − μ) μ
Az = − +1 + 3
2 r13 r2
The boundary conditions for the costate variables are the following:
T
∂F
λr (tf ) = d (6.106)
∂r t=tf
T
∂F
λv (tf ) = d (6.107)
∂v t=tf

∂φ
λc (tf ) = =1 (6.108)
∂c t=tf
where d = (d1 , d2 , . . . , dk )T are the additional parameters used to adjoin the

terminal constraint equation, Eq. (6.93), to the performance index as follows:
Ja = φ[x(tf ), tf ] + dT F = c(tf ) + dT F (6.109)
The primer vector method (see Chap. 5) produces the following optimal thrust
direction in a CR3BP field:
p
n̂ = (6.110)
p
where p(t) = −λv (t) is the primer vector governed by the following differential
equations:
ṗ = λr (6.111)
T
∂g
p̈ = p (6.112)
∂r
with the boundary conditions

T
∂F
ṗ(tf ) = d (6.113)
∂r t=tf
T
∂F
p(tf ) = − d (6.114)
∂v t=tf
The Hamiltonian expressed in terms of the primer vector is the following:
H = a(λc − p) + ṗT v − pT g
(6.115)

the acceleration magnitude. However, the bounds on a given by Eq. (6.97) allow
the application of the Pontryagin’s minimum principle to yield for the optimal
magnitude, â(t):
â(λc − p) ≤ a(λc − p) (6.116)
The inequalities of Eqs. (6.97) and (6.116) result in the following switching
condition for the optimal control magnitude:

am (t) , (p − λc ) > 0
â(t) = (6.117)
0 , (p − λc ) < 0
If the upper bound am is assumed to be constant, based upon the approximation that
the propellant consumed, m0 − m(t), is a small fraction of the initial mass, m0 , at
any time t, then the Hamiltonian is time invariant in the non-inertial frame, thereby
producing λc = const. = 1. In such a case, the switching condition becomes the
following:

am , p > 1
â(t) = (6.118)
0, p<1
Example 6.1 Consider a low-thrust interstellar mission which requires sending

a spacecraft from a low-Earth orbit to an out-of-ecliptic radial position, while
maximizing the terminal speed, vf , as the propellant is exhausted at t = tf . The
initial conditions in the CR3BP frame are x(0) = −0.1, y(0) = z(0) = 0, ẋ(0) = 4,
ẋ(0) = 3, ż(0) = 0.01, and an ion-propulsion engine is assumed to produce
a maximum acceleration magnitude, am = 0.01 m/s2 , which is transformed to
the non-dimensional units of Earth–Moon CR3BP system by multiplying it with
367.1975 s2 /m. A three-dimensional plot of the trajectory (shown in red), simulated
with a fourth-order Runge–Kutta integration, is compared with the zero-thrust case
(am = 0) (plotted in blue) in Fig. 6.10, showing a more rapidly increasing radius
(clearly seen in the planar plot of Fig. 6.11), as well as a much larger out-of-plane
motion (Fig. 6.12) achieved with the continuous thrust. The effects of thrusting on
the radius and velocity are depicted in Fig. 6.13, clearly showing the variations as
the spacecraft initially orbits the Earth, and then departs on the outward journey.
Figure 6.14 shows the flight-path angle, φ, and the azimuth angle, ψ, showing the
am = 0
am = 0. 01 m / s2
1
−1
z
−2
−3
−4
20
10
−10
−20 20 30
10
y −10 0
−30 −30 −20
−40 x
Fig. 6.10 Optimal low-thrust trajectory (red) for out-of-plane, interstellar mission compared with
the free flight case (blue)
am = 0 am = 0. 01 m / s2
20
15
10
0
y
−5
−10
−15
−20
−25
−40 −30 −20 −10 0 10 20 30
x
Fig. 6.11 Projection of the optimal low-thrust trajectory (red) in the (x, y) plane compared with
am = 0 am = 0. 01 m /s2
0.5
−0.5
−1
−1.5
z
−2
−2.5
−3
−3.5
−4
−40 −30 −20 −10 0 10 20 30
x
Fig. 6.12 Projection of the optimal low-thrust trajectory (red) in the (x, z) plane compared with
am = 0 am = 0. 01 m /s2
40
30
20
r
10
0
0 1 2 3 4 5 6 7 8 9 10
40
30
20
v
10
0
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 6.13 Radius and speed of the optimal low-thrust trajectory (red) compared with the free flight
case (blue)
am = 0 am = 0. 01 m /s2
20
0
φ (deg. )
−20
−40
−60
0 1 2 3 4 5 6 7 8 9 10
90
45
ψ (deg. )
−45
−90
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 6.14 Flight-path and velocity azimuth angles of the optimal low-thrust trajectory (red)
compared with the free flight case (blue)
optimal thrusting asymptotically approaches the flight direction, with the azimuth
angle varying between ±π/2 on the spiral trajectory. Figure 6.15 is a plot of the
optimal thrust direction angles, α and β, resolved in the CR3BP frame (see Fig. 6.9).
The maximum thrust acceleration, a = am , is maintained until tf = 10 (approx. 48
days).
Zhang and Zhao [88] present a shooting method for solving the minimum-fuel,
low-thrust orbital transfer problem in CR3BP field. An interesting application of
the optimal low-thrust transfer method is in designing solar-sail missions around
heliocentric and Earth centered orbits, such as by Otten and McInnes [64] and
Macdonald and McInnes [56]. However, solar-sailing strategies are only locally
optimal, and face numerical problems in addressing the transversality conditions,
especially for out-of-plane maneuvers.
The optimal mission design of a multi-body transfer trajectory involves the
solution of a two-point boundary value problem, in which either the fuel or the time
of flight is minimized, and the control is applied either impulsively or continuously.
The effects of other bodies on either a two-body (Keplerian) or a restricted three-
body model (CR3BP or ER3BP) is modeled by patching the numerical solutions at
the boundaries of the feasible space for each optimization problem. Such approaches
include patching of impulsive two-body trajectories on the spheres of influence [80],
or transferring a vehicle to an invariant manifold of a CR3BP model using a
60
40
α (deg. )
20
−20
0 1 2 3 4 5 6 7 8 9 10
180
135
β (deg. )
90
45
−45
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 6.15 Optimal thrust direction angles of the low-thrust trajectory (red) compared with the free
flight case (blue)
continuous, optimal thrust. The design methods range from the numerical patching
of two-body Lambert solutions [20] to analytical searches for particular manifolds
of periodic and semi-periodic orbits around collinear Lagrangian points [37, 41].
Such approximate solutions are indispensable, and have served well in practical
mission designs, such as the exploration of Jovian and Saturnian systems carried
out by Galileo and Cassini missions, respectively.
6.6 Missions Around Collinear Lagrangian Points
The collinear Lagrangian points of a planet–Sun system are unique vantage points
for continuously observing either the planet or the Sun with a minimum-fuel
expenditure. If these equilibrium points were unavailable, a large radius, Keplerian
orbit would have to be established around the target body, with the associated
high cost of transfer. A similar station near the Lagrangian point, L2 , of the
Earth–Moon system can be used to relay communication signals between the far
side of the Moon and the Earth. Such an idea was proposed by Farquhar [29]
for the Apollo missions, but was first applied to the International Sun–Earth
Explorer Satellite (ISEE-3) mission, wherein a spacecraft was stationed near the
6.6 Missions Around Collinear Lagrangian Points 233
Sun–Earth Lagrangian point, L1 , to observe the effects of the solar wind, and its
interaction with the Earth’s magnetic field. Further practical examples of stationing
a spacecraft near the Sun–Earth L1 or L2 points are the Wind, the Solar and
Heliospheric Observatory (SOHO), the Advanced Composition Explorer (ACE),
the Wilkinson Microwave Anisotropy Probe (WMAP), the GEOTAIL, the Genesis,
ARTEMIS, Herschel, Planck, Chang‘e-5T1, Gaia, DSCOVR, and LISA Pathfinder
missions.
Although the collinear Lagrangian points are equilibrium points of the CR3BP,
and provide excellent vantage points for observational purposes, they are also
inherently unstable points. This implies that even a small, arbitrary perturbation
will cause a spacecraft stationed at such a point to depart from it, and to approach
either one of the two primaries. Such a perturbation occurs naturally due to the
neglected dynamics in the CR3BP model, such as the elliptical (rather than circular)
orbit of the primaries and the gravitational effect of other large bodies. Due to the
inherent instability of a collinear point, a control strategy (called station keeping) is
required to maintain the spacecraft in the vicinity of it by the periodic application of
thrust. Such a position must be changing periodically in a small amplitude motion
around a collinear point (like trying to balance a stick vertically in one’s palm),
and is called a halo orbit. It should be noted that halo orbits are not the natural
solutions of the CR3BP system to arbitrary initial conditions, but are artificially
forced either by choosing special initial conditions that preclude the non-periodic (or
secular) part of the motion or by providing external thrust inputs, which requires fuel
expenditure.
An analytical method of deriving the periodic and semi-periodic, small-
amplitude orbits about the collinear Lagrangian points is based upon the local
linearization of the CR3BP equations. By neglecting the higher-order terms in
a series expansion about a collinear point, and by retaining only those initial
conditions which do not excite the unstable response, an approximate linear model
of the halo orbits can be obtained [78]. Such special (or particular) solutions of
the linearized model, although unrealistic as accurate solutions of the nonlinear,
inherently unstable, unforced dynamics, are however useful in providing nominal
paths for maintaining the spacecraft in the vicinity of a collinear point, by the
application of control. Other small-amplitude orbits of the collinear points are also
possible, such as the quasi-halo [36], Lissajous [22], and rectilinear out-of-plane
trajectories, which are not exactly periodic, but display a semi-periodic nature.
Stable and unstable manifolds around the periodic and semi-periodic orbits can be
constructed as feasible paths of transferring the spacecraft to and from, respectively,
the collinear points in a practical Lagrangian point mission.
The CR3BP equations expanded about the Lagrangian points L2 and L1 in a
third-order series were presented by Farquhar and Kamel [30] and Richardson [69],
respectively, for deriving the approximate periodic motions by neglecting the
non-periodic (secular) part of the response by the classical method of Lindstedt–
Poincaré. We will next consider such a model here for illustrative purposes.
6.6.1 Motion About the Collinear Points
Consider a change in the CR3BP coordinates such that the origin is shifted from
the barycenter to the collinear Lagrangian point, Li , i = 1, 2, 3. Since the motion
around L3 is the same as that around L2 , except for the changed non-dimensional
mass from μ to 1 − μ, we need to consider only L1 and L2 , whose locations in the
original coordinate system are
i = i i, i = 1, 2, where i is given by the root of
the quintic equation for the i th Lagrangian point. The vectors from the Lagrangian
point to the primaries m1 and m2 in new coordinates are thus the following:
d1 = (1 − μ − i )i (6.119)
d2 = −(μ + i )i
and ρ = (ρx , ρy , ρz )T is the location of the third mass, m3 , from the Lagrangian
point, normalized by d1 =| d1 |. The Lagrangian function, normalized by d12 , and
after dropping the constant and linear terms in ρ̇ (as they do not contribute to the
equations of motion) is expressed as follows:
1 ρx
L(ρ, ρ̇) = (ρ̇ − Kρ)T (ρ̇ − Kρ) + i − U (ρ) (6.120)
2 d1
where K is given by Eq. (6.70), and
μ 1 (1 − μ) d2 /d1
U (ρ) = − − (6.121)
d13 |ρ∓i| 2
d1 d2 | ρ + (d2 /d1 )i |
is the gravitational potential with ρ =| ρ |. In the denominator of the first term in

Eq. (6.121), the upper sign is taken for the Lagrangian point L1 and the lower sign
for L2 . A single series expansion of the Lagrangian function with d2 ≥ d1 would
give the orbit about L1 for 0 < μ ≤ 1/2, and about L2 for 0 < μ < 1. When
the following series expansion by Legendre polynomials, Pn (.), is introduced into
Eq. (6.121):
# ∞
1
= ρ n Pn (±ρx /ρ) (6.122)
|ρ∓i|
n=0
# ∞
d2 /d1
= (d1 ρ/d2 )n Pn (−ρx /ρ) (6.123)
| ρ + (d2 /d1 )i |
n=0
we have the following Lagrangian:

1
L(ρ, ρ̇) = (ρ̇ − Kρ)T (ρ̇ − Kρ)
2
∞
#
+ cn ρ n Pn (ρx /ρ) (6.124)
n=2
where
μ (1 − μ)d1n−2
cn = (±1)n + (−1)n (6.125)
d13 d2n+1
where the upper sign is taken for the Lagrangian point L1 , and the lower sign for
L2 . The Lagrangian in the neighborhood of L3 can be derived using the expression
for L2 , normalizing ρ by d2 , and replacing the non-dimensional mass μ by 1 − μ.
This yields the following coefficient for L3 :
1−μ μd2n−2
cn = (−1)n + (−1)n (6.126)
d23 d1n+1
Finally, the Lagrangian results in the following Euler–Lagrange equations of motion

in the neighborhood of a Lagrangian point:
∞
#
ρ̈ − 2Kρ̇ + K2 ρ = ncn ρ n−2 Pn (ρx /ρ)ρ
n=2
∞
#
+ cn ρ n−2 Pn
(ρx /ρ) (ρi − ρx ρ/ρ) (6.127)
n=2
where
dPn (nu) # n−1
Pn
(ν) = = (2n − 4k − 1)Pn−2k−1 (ν) , (0 ≤ k ≤)
dν 2
(6.128)
A linearization of Eq. (6.127) in the vicinity of the collinear point requires
taking only the quadratic terms in the Lagrangian given by Eq. (6.124). Such an
approximation results in the following:
1
L(ρ, ρ̇) L̄(ρ, ρ̇) = (ρ̇ − Kρ)T (ρ̇ − Kρ) + c2 ρ 2 P2 (ρx /ρ) (6.129)
2
where we recall that P2 (ν) = (3ν 2 − 1)/2 and
μ 1−μ
c2 = + (6.130)
d13 d23
Substitution of Eqs. (6.129) and (6.130) into Eq. (6.127) with n = 2 yields the
following linearized equations of motion:
ρ̈ − 2Kρ̇ + (K2 + C)ρ = 0 (6.131)

where
⎛ ⎞
−2c2 0 0
C = ⎝ 0 c2 0 ⎠ (6.132)
0 0 c2
The characteristic equation of the linearized system is given by
det{s 2 I3×3 −2sK+K2 +C} = (s 2 +c2 )[s 4 +s 2 (2−c2 )+1+c2 −2c22 ] = 0 (6.133)
which yields the eigenvalues

5 ,
6
6 c − 2 ± 9c2 − 8c
√ 7 2 2 2
s1,2 = ±j c2 , s3,4,...,6 =±
2
√
with j = −1.
The coefficient c2 determines the nature of the linearized initial response. Its
value can be analyzed as follows:
μ 1−μ
c2 = a + b , a = , b=
d13 d23
From the location of the collinear point, it follows that
μ(1 − μ − i ) (1 − μ)(μ + i )
i + − =0
| 1 − μ − i |3 | μ + i |3
which implies that
i − b(μ + i )
a=
i − 1 + μ
or

i + μ i
c2 = a + b = b 1 − + >1
i − 1 + μ i − 1 + μ
which holds for both L1 and L2 (since | 1 + μ |< 1 and | 2 + μ |> 1). Thus
c2 > 1 always holds.
Because c2 > 1, the coplanar (x, y) motion involves a conjugate pair of
imaginary eigenvalues, ±λj , and a pair of real eigenvalues, ±α, where
5 , 5 ,
6 6
6 2 − c + 9c2 − 8c 6 c − 2 + 9c2 − 8c
7 2 2 2 7 2 2 2
λ= , α=
2 2
and the out-of-plane motion (z) (which is decoupled from the coplanar motion)
√
consists of a conjugate pair of imaginary eigenvalues, ± c2 j . Such a configuration
of eigenvalues is termed “saddle×center×center” type, and indicates naturally
unstable equilibrium. Consequently, the initial response of the unforced linearized
system is expressed as follows:
x(t) = A1 eαt + A2 e−αt + A3 cos λt + A4 sin λt

y(t) = −k1 A1 eαt − A2 e−αt − k2 (A3 sin λt − A4 cos λt) (6.134)
√ √
z(t) = A5 cos c2 t + A6 sin c2 t
where
1 + 2c2 − α 2 1 + 2c2 + λ2
k1 = , k2 =
2α 2λ
and A1 , A2 , . . . , A6 are real coefficients to be determined from the initial conditions.
As pointed out earlier, we are interested only in the periodic and semi-periodic
motions around the collinear points. Therefore, we carefully select only those
initial conditions which make the secular (or non-periodic) terms vanish, that is
A1 = A2 = 0, thereby producing the following initial response of the linear system:
x(t) = −Ax cos(λt + φ)

y(t) = kAx sin(λt + φ) (6.135)
√
z(t) = Az sin( c2 t + ψ)
where k = k2 , and the amplitudes, Ax , Az , as well as the phase angles, φ, ψ, are

to be determined from initial conditions. We note that a purely periodic, three-
√
dimensional response (i.e., a halo orbit) is possible, if and only if λ/ c2 is an
integer. The smallest value of this integer being unity, the smallest energy halo orbit
√
would be produced for the resonant condition (λ = c2 in which the coplanar and
out-of-plane motions sustain each other). Unfortunately, this is not possible, because
√
λ/ c2 is an irrational number. Thus the linear approximation cannot produce a halo
orbit, but only gives rise to semi-periodic solutions, such as Lissajous tori. Hence,
one must look for higher-order approximations to generate a halo orbit solution.
6.6.2 Lindstedt–Poincaré Method for Halo Orbits
In order to devise a higher-order dynamic model for producing periodic orbits—

called halo orbits—around the collinear points, the perturbation method of
Lindstedt–Poincaré [58] is applied to neglect the secular part of the initial response,
while retaining its periodic behavior. The method involves refining a frequency
scale by successively increasing the order of the series approximation. Consider

the periodic motion to have a natural frequency, ω, such as the time scale of
motion is changed from t to τ = t/ω. The series expansion of ω in terms of the
non-dimensional parameter, 0 < 1, is the following:
ω = 1 + ω1 + 2 ω2 + · · · (6.136)
where ω1 , ω2 , . . . are the frequency corrections applied to first, second, and higher-
order terms. The time derivatives in τ are denoted by primes, and are related to the
derivatives in t as follows:
d(.) d(.)
=ω = ω(.)
dt dτ
d2 (.) (
2 d .)
= ω = ω2 (.)
(6.137)
dt 2 dτ 2

= 1 + 2ω1 + 2 (ω12 + 2ω2 ) + · · · (.)
The nonlinear equations of motion in the neighborhood of a collinear point,

Eq. (6.127), are expanded up to the third-degree polynomials as follows:
3
ω2 x
− 2ωy
− (1 + 2c2 )x = c3 (2x 2 − y 2 − z2 ) + 2c4 (2x 2 − 3y 2 − 3z2 )x
2
3
ω2 y
+ 2ωx
+ (c2 − 1)y = −3c3 xy − c4 (4x 2 − y 2 − z2 )y (6.138)
2
3
ω2 z
+ λ2 z = −3c3 xz − c4 (4x 2 − y 2 − z2 )z + (λ2 − c2 )z

2
On expanding the solution in a third-order series of , we have the following result:
x(τ ) = x1 (τ ) + 2 x2 (τ ) + 3 x3 (τ ) + · · ·
y(τ ) = y1 (τ ) + 2 y2 (τ ) + 3 y3 (τ ) + · · · (6.139)
z(τ ) = z1 (τ ) + 2 z2 (τ ) + 3 z3 (τ ) + · · ·
where the assumption of the same order expansion for both the coplanar and out-of-
plane motions is made, based upon the approximation that the natural frequencies
of the two motions are very close, that is
.
Δ = λ2 − c2 = O( 2 ) (6.140)
Such a frequency matching (or resonance) of the coplanar and out-of-plane motions
is crucial in achieving the periodic motion.
After substituting the linearized equations, Eq. (6.131), the secular terms are
made to vanish by taking ω1 = 0, in the second-order response (O()). Next, the
third-order equations are derived (O( 2 )) by including the first- and second-order
terms, and putting ω1 = 0 in the series expansion. The removal of secular terms
from the third-order expansion is not possible simply by choosing an appropriate
value for the frequency correction, ω2 . The secular part of the out-of-plane response
is due to δ4 sin(2τ1 − τ2 ), which can be adjusted by selecting the phase angles such
that
π
φ =ψ +n , (n = 0, 1, 2, 3) (6.141)
2
Then the secular terms in both of the coplanar equations can be made to vanish
by imposing amplitude constraints, as well as selecting a particular third-order
frequency correction, ω2 . The final step in the approximation involves removing
from the equations by substituting the amplitudes Ax and Az by Ax / and Az /,
respectively, thereby implying the two amplitudes to be of O(). The first- to
third-order solutions obtained above are now combined to produce the following
result:
x(τ ) = a21 A2x + a22 A2z − Ax cos τ1 + (a23 A2x + ηa24 A2z ) cos 2τ1
+ (a31 A3x + ηa32 Ax A2z ) cos 3τ1
y(τ ) = (kAx b33 A3x + b34 Ax A2z + ηb35 Ax A2z ) sin τ1
+ (b21 A2x + ηb22 A2z ) sin 2τ1 + (b31 A3x + ηb32 Ax A2z ) sin 3τ1 (6.142)
z(τ ) =
&
(−1)n/2 (Az sin τ1 + d21 Ax Az ) sin 2τ1 + (d31 A3z + d32 A2x Az ) sin 3τ1 ) (n = 0, 2)
z cos τ1 + d21 Ax Az )(cos 2τ1 − 3) + (d32 Ax Az − d31 Az ) cos 3τ1 ) (n = 1, 3)
(−1)(n−1)/2 (A 2 3
where the coefficients a21 − d32 are listed in Thurman and Worfolk [82], and η =
(−1)n , (n = 0, 1, 2, 3). The qualitative change in the periodic behavior due to the
switching of the out-of-plane phase angle, ψ, between odd and even multiples of
π/2 is termed bifurcation. Hence the two distinct classes of periodic solutions are
indicated by n = (0, 2) and n = (1, 3), respectively.
Example 6.2 Consider a halo orbit around the Sun–Earth Lagrangian point, L1 ,
representing the ISEE-3 mission parameters [70] with λ = 2.08645, and Δ =
0.292214, and simulated for Az = 125,000 km in Figs. 6.16, 6.17, 6.18, and 6.19.
The period of the orbit is 177.704 mean solar days, and the (x, y) symmetry as
well as the phase change in the out-of-plane (z) motion from n = 1 (blue) and
n = 3 (red) are clearly seen. The coplanar part of the motion is unaffected by
the change in ψ (Fig. 6.17), whereas both (y, z) (Fig. 6.18) and (x, z) (Fig. 6.19)
projections show a sign change of the z motion from ψ = π/2 (blue) to
ψ = 3π/2 (red).
−3
x 10
1
0.5
0
z
−0.5
−1
0.5
0.15 0.2
y 0.05 0.1
−0.5 −0.05 0
−0.1
−0.15 x
Fig. 6.16 Three-dimensional halo orbit around the Sun–Earth collinear point L1 with Az =
125,000 km and period 177.704 days, showing out-of-plane phase change from ψ = π/2 (blue)
to ψ = 3π/2 (red)
0.5
0.4
0.3
0.2
0.1
0
y
−0.1
−0.2
−0.3
−0.4
−0.5
−0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
x
Fig. 6.17 The (x, y) projection of the halo orbit around the Sun–Earth collinear point L1 with
Az = 125,000 km and period 177.704 days, unchanged by the out-of-plane phase change from
ψ = π/2 (blue) to ψ = 3π/2 (red)
−3
x 10
1
0.8
0.6
0.4
0.2
0
z
−0.2
−0.4
−0.6
−0.8
−1
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
y
Fig. 6.18 The (y, z) projection of the halo orbit around the Sun–Earth collinear point L1 with
Az = 125,000 km and period 177.704 days, showing a sign change in the z response due to the
out-of-plane phase change from ψ = π/2 (blue) to ψ = 3π/2 (red)
−3
x 10
1
0.8
0.6
0.4
0.2
0
z
−0.2
−0.4
−0.6
−0.8
−1
−0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
x
Fig. 6.19 The (x, z) projection of the halo orbit around the Sun–Earth collinear point L1 with
Az = 125,000 km and period 177.704 days, showing a sign change in the z response due to the
out-of-plane phase change from ψ = π/2 (blue) to ψ = 3π/2 (red)
6.7 Numerical Computation of Halo Orbits and Manifolds
The CR3BP equations can be expressed as follows in terms of the canonical

coordinate vector, X = (qT , pT )T , in the Hamiltonian form given by Eq. (6.79),
repeated here for convenience as follows:
Ẋ = J∇X H = f(X) (6.143)
where ∇X is the gradient operator relative to X,

03×3 I3×3
J= (6.144)
−I3×3 03×3
the Hamiltonian is given by
1 T
H = p p + pT Kq + U (q) (6.145)
2
the generalized momentum vector by
p = q̇ − Kq , (6.146)
⎛ ⎞
0 1 0
K = ⎝ −1 0 0 ⎠ (6.147)
0 0 0
the gravitational potential by
1 (1 − μ) μ
U (q) = −Ω − qT K2 q = − − (6.148)
2 r1 r2
with
1 μ (1 − μ) 1 (1 − μ) μ
Ω = − qT K2 q + + = (x 2 + y 2 ) + +
2 | q − (1 − μ)i | | q + μi | 2 r1 r2
(6.149)
and i = (1, 0, 0)T and r1 , r2 given by Eq. (6.54). When expressed in the functional
form, Eq. (6.143) is the following:

Kq + p
Ẋ = f(X) = (6.150)
−∇q (U ) − KT p
6.7 Numerical Computation of Halo Orbits and Manifolds 243
where ∇q (.) is the gradient operator relative to q.1
6.7.1 Monodromy Matrix
Due to its nonlinear nature, a closed-form solution subject to a given initial

condition, X(0) = X0 , is unavailable. However, an approximate numerical solution
can be attempted for a periodic orbit where the motion repeats itself after a fixed
time period, T , that is X(t + T , X0 ) = X(t, X0 ). Consider the change in the solution
due to a small, arbitrary change in the initial condition, (Δt, ΔX0 ), expressed in the
following Taylor series:

∂X
X(t + Δt, X0 ) = X(t, X0 ) + ΔX0
∂X0 (t,X0 )

∂X
+ Δt + O(Δ2 t, | ΔX0 |2 ) (6.151)
∂t (t,X0 )
By truncating the series at the linear terms, the solution is treated as if it were that
of a linear, time-varying system with the following state transition matrix:

∂X
(t, X0 ) = (6.152)
∂X0 (t,X0 )
which satisfies the initial condition
(0, X0 ) = I6×6 (6.153)
Taking the partial time derivative of , we have
∂ ∂ ∂X
=
∂t ∂X0 ∂t
1A different state-space representation is obtained by replacing the generalized momentum vector,

p, with the velocity, q̇:

q̇ q̇
f(X) = =
q̈ KT Kq − ∇q (U ) + 2Kq̇
whose Jacobian is given by

0 I
f
(X) =
Ωqq 2K

∂f(X) ∂X
= (6.154)
∂X X0 ∂X0 (t,X0 )
= f
(X)
where

∂f(X) K I
f
(X) = = (6.155)
∂X X0 −Uqq −KT
is the Jacobian matrix and

∂ 2U
Uqq =
∂q∂q q0
For a periodic orbit, the state transition matrix, , computed over one time period
is termed the monodromy matrix, denoted by M = (T , X0 ). Since it is evaluated
at the initial condition, X(0) = X0 , the monodromy matrix repeats itself after each
time period, T , because X(T ) = X(0). It has been noted in both Chaps. 2 and 3 that
the state transition matrix of a conservative Hamiltonian system is symplectic [60],
hence the same is true for (t, X0 ) as well as the monodromy matrix, that is, they
both obey M−1 = −JMT J, where J is given by Eq. (6.144). The symplectic nature
(see Chap. 2) of the monodromy matrix, as well as the time-reversal symmetry of the
CR3BP solutions across the (x, z) plane (noted above), enables a direct computation
of a halo orbit.
Computation of the state transition matrix for a fixed initial state, X0 , is carried
out by solving the differential equation governing its time evolution, given by
˙ = f
(X) , (0) = I

Such an iteration utilizes the symmetry of the CR3BP solutions discussed earlier.
For example, consider Newton’s method consisting of the following steps:
(i) Assume an initial guess of the solution on the (x, z) plane, X(0) =
(x0 , 0, z0 , 0, py0 , 0)T . Then the first return of the symmetric, periodic orbit to
the (x, z) plane is given by X(T /2, X0 ) = (x̄, 0, z̄, p̄x , p̄y , p̄z )T , where T is
the time period.
(ii) To search for the first half of the orbit, the initial conditions (x0 , z0 , py0 ) are to
be found which produce v̄x = p̄x − x̄ = 0 and v̄z = p̄z = 0 on the (x, z) plane.
For this purpose, a correction, ΔX0 = (Δx0 , 0, Δz0 , 0, Δpy0 , 0)T , is applied
to the initial state such that the state evolves to cancel the nonzero terms at
t = T /2, v̄x , v̄z , by driving them to −v̄x , −v̄z at a time step t = T /2 + δt:
X(T /2 + Δt, X0 + ΔX0 ) X(T /2, X0 ) + (T /2, X0 )ΔX0

+ f[X(T /2, X0 )]Δt (6.156)
which yields
⎛⎧ ⎫ ⎞
⎪
⎪ ··· ⎪
⎪
⎜⎪⎪ ⎪
⎪ ⎟
⎜⎪⎪ 0 ⎪
⎪ ⎟
⎜⎪⎨ ⎪
⎬ ⎟
−1 ⎜ ··· ⎟
ΔX0 = (T /2, X0 ) ⎜ − f[X(T /2, X0 )]Δt ⎟ (6.157)
⎜⎪⎪ −(p̄x − x̄) ⎪
⎪ ⎟
⎜⎪ ⎪ ⎟
⎝⎪⎪
⎪ · · · ⎪
⎪
⎪ ⎠
⎪
⎩ ⎪
⎭
−p̄z
The three initial variable corrections, (Δx0 , Δz0 , Δpy0 ), and the time step,
Δt, are determined from the three equations given by Eq. (6.157). Since there
are four unknowns and only three equations, the z0 variation is set to zero,
(Δz0 = 0), by selecting an amplitude, z0 , of the out-of-plane motion. This can
be done by taking a derivative of Jacobi’s integral at the initial time.
(iii) The state transition matrix computed for half orbit, (T /2, X0 + ΔX0 ), is
corrected for the new initial condition, and step (ii) is repeated after replacing
T /2 → T /2 + Δt. This process is continued until v̄x , v̄z are driven to within a
small tolerance.
The iterative procedure given above converges to an initial condition, X0 , and
a semi-period, T /2, from which the monodromy matrix, M = (T , X0 ), can be
computed by using the symmetry (reflection with time reversal) of the orbit about
the (x, z) plane as follows [40]:
M = (T , X0 ) = A−1 (T /2, X0 )A(T /2, X0 ) (6.158)
where
⎛ ⎞
1 0 0
0
A= , = ⎝ 0 −1 0 ⎠ (6.159)
0 −
0 0 1
In the above given calculations, the symplectic nature of allows its following
inversion:
−1 (T /2, X0 ) = V−1 S−1 V−T T (T /2, X0 )VT SV (6.160)
where

I 0 0 I
V= , S= (6.161)
−K I −I 0
and K is given by Eq. (6.147). The final expression for M is thus the following:

0 −I −2K I
M=A T (T /2, X0 ) A(T /2, X0 ) (6.162)
I −2K −I 0
The eigenvalues of the monodromy matrix for a nearly planar orbit [40] can be
classified as two reciprocal pair of real eigenvalues, (λ1 , λ1 ), (λ2 , λ2 ), respectively,
and a complex conjugate pair, λ3 , λ∗3 . Since monodromy matrix is periodic (with a
unit determinant due to its symplectic character), a stable motion should have all
eigenvalues lying either inside or on the unit circle in the complex plane. However,
due to the instability of the collinear points, there is a real eigenvalue greater than
unity, λ1 > 1, its reciprocal eigenvalue less than unity, while λ2 = 1 and | λ3 |=|
λ∗3 |= 1 are on the unit circle. The two reciprocal eigenvalues, λ1 , λ1 , not falling on
the unit circle are termed the unstable and stable eigenvalues, respectively, as they
represent either growing or decaying periodic motion.
6.7.2 Stable and Unstable Manifolds
A manifold is a set of trajectories in the CR3BP space, x(t), y(t), z(t), which satisfy
a particular condition, F[x(t), y(t), z(t)] = 0. The zero-speed contours of constant
Jacobi’s integral are an example of a CR3BP manifold. Manifolds associated with
a halo orbit around a collinear point are especially useful as a set of transfer
trajectories either to or from the collinear point, and are utilized in practical missions
for stationing spacecraft around such a point. Manifolds structured around periodic
orbits are also termed invariant manifolds, because a trajectory starting on such a
manifold always remains on it. Hence manifolds can be used to design essentially
fuel-free paths in optimal low-thrust transfer missions, as discussed later in the
chapter.
A stable manifold begins from a halo orbit and evolves in forward time, t > 0,
along the eigenvector, vk− , associated with the stable eigenvalues, λk− < 1, of the
monodromy matrix, M = (T , X0 ) [38]. Thus the stable manifold is tangential to
both the halo orbit and a stable eigenvector, and evolves away from the halo orbit. An
unstable manifold terminates at a halo orbit, and evolves in backward time, t < 0,
along the eigenvector, vk+ , associated with an unstable eigenvalue, λk+ > 1, of the
monodromy matrix. Thus an unstable manifold serves as an envelope of possible
trajectories evolving toward the halo orbit in forward time. For a nearly planar halo
orbit, there is only one stable eigenvalue, and also only one unstable eigenvalue.
Thus the stable and unstable manifolds of a planar orbit can be constructed by
evaluating the corresponding stable and unstable eigenvectors at each point on
the orbit, and propagating the point tangentially along the respective eigenvectors
using the state transition matrix, (t, X0 ). With a linearized system dynamics,
the construction of manifolds becomes much easier, which has been proposed by
Thurman and Worfolk [82] and further studied by Bando and Scheeres [6] using the
linear optimal control formulation. Howell et al. [41] carry out a similar construction
of manifolds around Lissajous (semi-periodic) orbits for the Genesis mission.
For the system linearized in the neighborhood of a collinear Lagrangian point,
the Hamiltonian is in the following quadratic form:
1 T 1
H = p p + pT Kq + qT Cq (6.163)
2 2
where
⎛ ⎞
−2c2 0 0
C = ⎝ 0 c2 0 ⎠ (6.164)
0 0 c2
with c2 > 1 being the coefficient associated with the second-degree Legendre
polynomial (Eq. (6.130)). Consequently, the linearized canonical equations are time
invariant, and expressed in a state-space form as follows:

K I
Ẋ = X = AX (6.165)
−C −KT
which is the same system as Eq. (6.131), with the eigenvalues ±λj and ±α for the
√
coplanar motion and ± c2 j , with
5 , 5 ,
6 6
6 2 − c + 9c2 − 8c 6 c − 2 + 9c2 − 8c
7 2 2 2 7 2 2 2
λ= , α=
2 2
The state transformation from canonical state space, X = (qT , pT )T , to the

eigenspace, X̄ = VX, with
˙ = ĀX̄
X̄ (6.166)
such that
⎛ ⎞
α 0 0 0 0 0
⎜0 0 0 λ 0 ⎟
⎜ 0
√ ⎟
⎜ ⎟
⎜0 0 0 0 0 c2 ⎟
Ā = VAV−1 =⎜ ⎟ (6.167)
⎜0 0 0 −α 0 0 ⎟
⎜ ⎟
⎝ 0 −λ 0 0 0 0 ⎠
√
0 0 − c2 0 0 0
is given by the transformation matrix

0 1−1
1 1 1 1 1 1/4
V= v+ , vR , 1/4 e3 , v− , vI , c2 e3 (6.168)
w1 w2 c2 w 1 w 2
where v+ is the eigenvector of A associated with the unstable eigenvalue, α, v−

is that corresponding to the stable eigenvalue, −α, vR ± j vI the eigenvectors
corresponding to the imaginary coplanar eigenvalues, ±j λ, e3 = (0, 0, 1, 0, 0, 0)T ,
e6 = (0, 0, 0, 0, 0, 1)T , and
, ,
w1 = 2α[(4 + 3c2 )α 2 + 4 + 5c2 − 6c22 ] , w2 = λ[(4 + 3c2 )λ2 − 4 − 5c2 + 6c22 ]
are the normalizing factors associated with α and λ, respectively.

Example 6.3 For the halo orbit around the Sun–Earth Lagrangian point, L1 ,
considered in Example 6.2, we have c2 = 4.06107, which produces the following
linearized state dynamics matrix:
⎛ ⎞
0 1 0 1 0 0
⎜ −1 0⎟
⎜ 0 0 0 1 ⎟
⎜ ⎟
⎜ 0 0 0 0 0 1⎟
A=⎜ ⎟
⎜ 8.12214 0 0 0 1 0⎟
⎜ ⎟
⎝ 0 −4.06107 0 −1 0 0⎠
0 0 −4.06107 0 0 0
has the state transformation matrix

⎛ ⎞
−1.7527 −0.2022 0 −0.5714 −0.3055 0
⎜ 0.4441 0 ⎟
⎜ 0 0 0 1.2550 ⎟
⎜ ⎟
⎜ 0 0 1.4196 0 0 0 ⎟
V=⎜ ⎟
⎜ 1.7527 −0.2022 0 −0.5714 0.3055 0 ⎟
⎜ ⎟
⎝ −0.0000 −2.2299 0 −0.3886 0 0 ⎠
0 0 0 0 0 0.7044
and the following transformed dynamics matrix in the eigenspace:

⎛ ⎞
2.5327 0 0 0 0 0
⎜ 0 0 ⎟
⎜ 0 0 0 2.0865 ⎟
⎜ ⎟
⎜ 0 0 0 0 0 2.0152 ⎟
Ā = ⎜ ⎟
⎜ 0 0 0 −2.5327 0 0 ⎟
⎜ ⎟
⎝ 0 −2.0865 0 0 0 0 ⎠
0 0 −2.0152 0 0 0
For transferring a spacecraft to (or from) a collinear point orbit, the optimal
strategy consists of transferring the spacecraft to an invariant manifold associated
with the given collinear point. The remaining journey (or a large part of it) is then
carried out on the manifold without fuel expenditure. Anderson and Lo [2] study the
practicality of invariant manifolds in designing a low-thrust trajectory. Martin and
Conway [25] apply a Gauss–Lobatto collocation method for solving the 2PBVP
associated with an optimal transfer on an invariant manifold, and consider optimal
transfers to Earth–Moon L1 halo orbit. Such nonlinear programming methods are
valuable in determining either the shortest (time optimal) or the most efficient (fuel
optimal) low-thrust transfer paths in CR3BP and ER3BP gravity models.
6.8 Optimal Station Keeping around Collinear Points 249
6.8 Optimal Station Keeping around Collinear Points
A halo orbit either constructed analytically by the method of the previous section or
computed by a numerical scheme can be utilized as the nominal trajectory to station
a spacecraft around a collinear Lagrangian point. The design of an orbital regulation
system can be carried out using a suitable optimal feedback control law (see Chap. 2)
to track the halo orbit in the presence of disturbances. Such disturbances (or noise)
are problematic, because as mentioned earlier, the collinear points are naturally
unstable equilibrium points of the CR3BP system, and therefore necessitate an
automatic regulation system. It is crucial for such a controller to be optimal for a
prolonged mission, because frequent, arbitrary corrections can quickly deplete the
fuel, thereby shortening the life of the spacecraft. We shall briefly consider some of
the optimal station keeping methods that either have been applied in actual missions
or are being explored for future applications.
6.8.1 Impulsive Control
Impulsive control methods are traditionally applied in most missions, as they are
powered by chemical rockets. If N impulses are applied per orbit,
4 the optimization
consists of minimizing the net velocity impulse magnitude, N i | Δvi |, per orbit.
This invariably requires that each impulse be applied in an optimal direction. The
simplest possible optimal impulsive strategy is to determine each velocity impulse,
Δvi = (Δẋ, Δẏ, Δż)T , such that the coefficient of the unstable term, A1 , in the
linearized response, Eq. (6.134), is driven to zero as follows:
∇A1
Δvi = −A1 (6.169)
| ∇A1 |2
where ∇ = (∂/∂Δẋ, ∂/∂Δẏ, ∂/∂Δż)T denotes the velocity gradient. Such a

problem can be solved either by analytically or by dynamic programming. An
alternative approach is to minimize
t0 +T
1
J = A21 (t)dt (6.170)
T t0
where t0 is the initial time and T the time interval between two impulses. Another
method is to determine the velocity impulses for maximizing the time between
corrections. Shirobokov et al. [73] survey such impulsive methods, mostly from
the Russian archive, and suggest improvements in the optimal strategy.
A more sophisticated approach is to apply each impulse opposite to the unstable
manifold direction, determined by the eigenvector of the monodromy matrix
corresponding to the unstable eigenvalue of the linearized system (see the previous
section). This would attempt to cancel the instability in the interval between the
impulses. However, as the errors in the linearized dynamics increase with time, and
the period T between two impulses must be reasonably large for fuel minimization,
it is necessary to compute the state transition matrix, Φ(t0 , t), of the first variational
equations more accurately. Gomez et al. [35] apply the Floquet mode approach
to evolve the state transition matrix by Monte Carlo simulation for stationing a
spacecraft near Earth–Moon L2 point. A nonlinear sliding mode application [74]
for station keeping using the discrete linear, quadratic regulator (LQR) formulation
has also been studied [55], and applied to Lissajous trajectories around Earth–Moon
L2 point. Other impulsive methods applied to the station keeping problem include
the targeting methods [66] and Chebyshev–Picard iteration methods [5].
6.8.2 Continuous Thrust Control
Continuous, low-thrust inputs can be applied optimally to maintain a spacecraft

in the vicinity of a collinear point by tracking a given halo orbit in the synodic
(CR3BP) frame. Similarly, tracking a manifold of a halo orbit can optimally transfer
the spacecraft either toward or away from the collinear point. Consider the following
equations of motion linearized about such a nominal trajectory (either the halo or its
manifold), qn (t) = {xn (t), yn (t), zn (t)}T :
Ẋ(t) = A(t)X(t) + B(t)u(t) X(0) = X0 (6.171)
where X = (δqT , δpT )T is the state vector comprising first-order deviations from
the nominal trajectory, Xn = (qTn , pTn )T :
q(t) qn (t) + δq(t)

p(t) pn (t) + δp(t) (6.172)
u the control acceleration input, and

∂f(X) K I 0
A = f (X) = = ; B= (6.173)
∂X Xn −Uqq −KT I
are the Jacobian matrices, with f(X) the same as in Eq. (6.143) for the equations
of motion referred to the synodic frame with the origin at the barycenter, and the
potential Hessian, Uqq , evaluated on the nominal trajectory as follows:
⎛ ⎞
1−μ 0 0
∂ 2U 1−μ⎝
Uqq = = 3/2 0 1 0⎠
∂q∂q qn r1 0 0 1
6.8 Optimal Station Keeping around Collinear Points 251
⎧ ⎫ ⎛ ⎞
⎨ xn − μ ⎬ μ−1 0 0
3(1 − μ) μ ⎝
− yn {xn − μ, yn , zn } + 3/2 0 1 0⎠
5/2
r1 ⎩ ⎭ r
zn 2 0 0 1
⎧ ⎫
x + μ − 1⎬
3μ ⎨ n
− 5/2 yn {xn + μ − 1, yn , zn } (6.174)
r2 ⎩ zn
⎭
where
,
r1 = (xn − μ)2 + yn2 + zn2
,
r2 = (xn + μ − 1)2 + yn2 + zn2 (6.175)
An optimal control approach can be applied, such as the linear, quadratic

regulator (LQR) (see Chap. 2) for minimizing the following performance index for
driving the trajectory variations to zero:
∞
1
J (t) = XT (τ )Q(τ )X(τ ) + uT (τ )R(τ )u(τ ) dτ (6.176)
2 t
subject to the state constraint equation, Eq. (6.171), resulting in the linear, full-state
feedback control law:
u(t) = −K∞ (t)x(t) (6.177)
Since the plant is slowly time varying, the stabilizing steady-state solution for the
feedback gain matrix at each time instant, K∞ (t), is given by
K∞ (t) = R−1 (t)BT (t)P(t) (6.178)
where P∞ (t) is the symmetric, positive semi-definite, quasi-steady solution of the

following algebraic Riccati equation:
P∞ (t)A(t) + AT (t)P∞ (t) − P∞ (t)B(t)R−1 (t)BT (t)P∞ (t) + Q(t) = 0 (6.179)
For output feedback applications, a Kalman filter can be added in an LQG/LTR

formulation [79] to minimize the effects of the measurement noise inputs. Breakwell
et al. [18] apply an LQR-based approach to the ISEE-3 mission for station keeping
around the Sun–Earth L2 point. Ghorbani and Assadian [33] study the Earth–Moon
Lagrangian point missions by the LQR method, and also include the elliptical
restricted three-body (ER3BP) model, as well as the gravitational perturbations
of Sun and other planets. Jones and Bishop [43] apply the H2 robust control
methodology (a variation of the LQG/LTR method using the LQR at its core) [79]
for tracking Lissajous trajectories around Earth–Moon L2 point, whereas Kulkarni

et al. [48] apply a similar H∞ approach for halo orbit station keeping around Sun–
Earth L1 point.
Other methods applied to collinear point station keeping include the backstep-
ping [63] and model-predictive [54] control. Thus the collinear point station keeping
problem has been explored from nearly all possible angles in terms of control theory,
and continues to be an active research topic.
Answers to Selected Exercises
Exercises of Chap. 2
2.1

∗ e5 −t e5
x (t) = e + 5
t
e − et , u∗ (t) = −2 e−t , (0 ≤ t ≤ 5)
e − e−5 e5 − e−5
∗ = 1 > 0.
Optimal, because Huu
2.2
y ∗ (t) = −20t 3 + 30t 2 , u∗ (t) = −120t + 60, (0 ≤ t ≤ 1)
∗ = 1 > 0.
Optimal, because Huu
2.3

1+k22
10k12 −3t + t
1 3 2
1+2k22
y ∗ (t) = ,
1+k22
6 + 2
5 1
2
1+2k2
! 0 1"
∗ 20k12 1 + k22
u (t) = −t + (0 ≤ t ≤ 1)
1+k22 1 + 2k22
5
+ 1
6 2 1+2k22
k12 = 6.5, k22 = 3000 (Fig. 1).

https://doi.org/10.1007/978-3-030-03789-5
254 Answers to Selected Exercises
10
y (m)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t (s)
75
50
25
u(m / s2 )
0
−25
−50
−75
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t (s)
Fig. 1 The optimal trajectory and the corresponding control history for k12 = 6.5, k22 = 3000
(Exercise 2.3)
2.4 tf = 1, y ∗ (t) = −20t 3 + 30t 2 , u∗ (t) = −120t + 60 (0 ≤ t ≤ tf )

2
2.5 No, because it results in the necessary condition, etf + e−tf = −1, which
does not have a solution for any real value of tf .
2.7 The optimal displacement and velocity for this case are plotted in Fig. 2:
The final value of vertical acceleration input is u(tf ) = g = 1.6 m/s2 , because
ẋ2 (tf ) = 0.
2.8

∂φ ∗ 1$ ∗ %2
H+ = u (tf ) + λ∗1 (tf )x2∗ (tf ) + λ∗2 (tf )u∗ (tf ) = 0
∂t t=tf 2
or
1 2
− c1 tf − c2 + c1 gtf + k 2 tf = 0
2
implying
,
1
2 2 − c2 c2
tf = c c
1 2 + c1 g + k 2
± c c
1 2 + c1 g + k 1 2
c12
Answers to Selected Exercises 255
100
z*(m)
50
0
0 5 10 15 20 25 30
5
ż*(m / s)
−5
−10
0 5 10 15 20 25 30
2.5
u*(m /s2)
1.5
0 5 10 15 20 25 30
t(s)
Fig. 2 The optimal trajectory for lunar landing (Exercise 2.7)
A real and positive solution for tf is derived for k = 1/100, for which tf = 57.0248
s, c1 = 0.0249, and c2 = −0.714. The optimal displacement and velocity for this
case are plotted in Fig. 3.
2.15

−1 , λ̂(t) > 0
û(t) =
1 , λ̂(t) < 0
H = 1 + λu = 0, x̂(0) = 1, û = −1, (λ̂ = 1), x̂(t) = 1 − t, (0 ≤ t ≤ tf ),

x̂(tf ) = 0, tf = tˆ = 1.
2.18
(a) No.
(b) For a = 1.5g = 2.4 m/s2 ,
,
2 $ %
tf = ż(0) + ż2 (0) + 5 6gz(0) + ż2 (0) = 15.8712 s
5g
ż(0) + 12 gtf
tˆ = = 0.5619 s
3g
200
150
z*(m)
100
50
0
0 10 20 30 40 50 60
t(s)
10
5
ż *(m/s)
−5
−10
0 10 20 30 40 50 60
t(s)
Fig. 3 The optimal trajectory for lunar landing (Exercise 2.8)
100
50
z(m)
−50
0 2 4 6 8 10 12 14 16
0
ż(m / s)
−5
−10
−15
0 2 4 6 8 10 12 14 16
5
u(m / s2 )
−5
0 2 4 6 8 10 12 14 16
t(s)
Fig. 4 Time-optimal trajectory for lunar landing (Exercise 2.18)


−1.5g (0 ≤ t ≤ tˆ)
u(t) =
1.5g (tˆ < t ≤ tf )

z(0) + ż(0)t − 54 gt 2 (0 ≤ t ≤ tˆ)
z(t) =
z(0) + ż(0)t + 14 gt 2 + 32 g tˆ2 − 3g tˆt (tˆ < t ≤ tf )
The time-optimal displacement and velocity for this case are plotted in Fig. 4.
3.3
(a) v = 0.1995 km/s, φ = 53.643◦ .
(b) v = 0.1289 km/s, φ = −31.357◦ .
(c) 1018.3 s.
(d) C is 25.4246◦ behind the radial position of A.
(e) z = 0.0024, x = 0.6588, a = 606.591 km, e = 0.3820.
(f) Zero.
3.4
(a) v = 0.4164 km/s, φ = 0.
(b) v = 0.3875 km/s, φ = 0.
3.5 a = 3875.483 km, e = 0.950225, v1
= 3.32175 km/s, φ1
= 58.221◦ , v2
=
1.855 km/s, φ2
= −28.344◦ . No, because the orbit is Earth intersecting.
3.6 a = 8003.0968 km, e = 0.1754826, i = 63.6658◦ , = 98.1301◦ , ω =
59.3285◦ , τ = −373.756 s.
3.7
v1 = −0.193607I − 0.41983J + 3.32604K km/s
v2 = 7.28047I + 0.16851J + 6.2595K km/s
3.8 5.5939 km/s.

3.9
lim (v1 + v2 + v3 ) = 14.586 km/s

rt →∞
For rt = 100 A.U., v1 + v2 + v3 = 15.04 km/s.

4.1
ar∗ (t) = −λ∗ṙ (t) ,

⎧ ∗ ⎫ ⎧ ⎫
⎪ c1 cos nt + nc2 sin nt + 2c3 sin nt ⎪
⎪
⎪ λ (t) ⎪
⎪ ⎪
⎪ ⎪
⎪
⎨ r
⎬ ⎨ c c3 ⎬
∗ 1
λṙ (t) = − sin nt + c2 cos nt + 2 (cos nt − 1) (0 ≤ t ≤ tf )
⎪
⎪ ⎪ ⎪ n n ⎪
⎩ ∗ ⎪ ⎭ ⎪
⎪
⎩
⎪
⎪
⎭
λθ (t) c 3
where c1 , c2 , c3 are to be solved from the terminal boundary conditions, δr(tf ) =

δ ṙ(tf ) = δθ (tf ) = 0
4.2 λ∗θ (tf ) = 0 = c3

∂φ 1
H+ = 0 = − [λ∗ṙ (tf )]2
∂t t=tf 2
which implies λ∗ṙ (tf ) = 0. This along with δr(tf ) = δ ṙ(tf ) = 0 yields
2n2 [nδr(0) cos ntf + δ ṙ(0) sin ntf ] cos ntf

c1 =
ntf − sin ntf cos ntf
2n[nδr(0) cos ntf + δ ṙ(0) sin ntf ] sin ntf

c2 =
ntf − sin ntf cos ntf
and tf satisfies the following equation:
n2 tf δr(0) sin ntf + δ ṙ(0)(sin3 ntf + cos2 ntf sin ntf − ntf cos ntf ) = 0
4.4
aθ∗ (t) = −λ∗θ̇ (t)
⎧ ⎫
⎪ ∗
⎪ λr (t) ⎪
⎪ ⎪
⎨ λ∗ (t) ⎪⎬
θ
⎪ ∗
⎪ λṙ (t) ⎪
⎪ ⎪
⎩ λ∗ (t) ⎪⎭
θ̇
⎧ ⎫
⎪
⎪ c1 (4 − 3 cos nt) + 6c2 (nt − sin nt) + 3nc3 sin nt − 6nc4 (1 − cos nt) ⎪ ⎪
⎪
⎨ ⎪
⎬
c2
= (0 ≤ t ≤ tf )
⎪
⎪ −c1 n − 2c2
sin nt (1−cos nt)
+ c3 cos nt + 2c4 sin nt ⎪
⎪
⎪
⎩ n ⎪
⎭
2c1 (1−cos
n
nt)
+ c2 (3nt−4nsin nt) − 2c3 sin nt + c4 (4 cos nt − 3)
200
δr
100 Cδθ
(km)
−100
0 20 40 60 80 100 120 140 160 180
0.2
δṙ
(km / s)
Cδ θ̇
0
−0.2
0 20 40 60 80 100 120 140 160 180
1
aθ (m / s2 )
0.5
−0.5
0 20 40 60 80 100 120 140 160 180
nt (deg. )
Fig. 5 The optimal trajectory for coplanar rendezvous (Exercise 4.4)
where c1 , c2 , c3 , c4 are to be solved from the following terminal boundary condi-

tions:

2π 2 3π 2 14 8 2π
3
c1 + 2
− c2 − 2 c3 − 2 c4 = 0.07C
n n 2n n n n

2 3π 2 π 3π 2 2π 3π 2
− 3 14 − c1 + 3 16 − c2 − 2 c3 +3 16 − c4 = 0.06π C
n 2 n 2 n 2
8 2π
− c1 + c2 − 2π c3 + 12c4 = 0
n n

2π 3 3π 2 12 π
c1 + 2 16 − c2 − c3 − c4 = 0.12Cn
n2 n 2 n n
The optimal trajectory and control history are plotted in Fig. 5.

4.5

λ∗z (t) c1 cos nt + c2 n sin nt)
an∗ (t) = −λ∗ż (t) , = (0 ≤ t ≤ tf )
λ∗ż (t) −c1 sinnnt + c2 cos nt
where c1 , c2 are solved from the terminal boundary conditions to be the following:
n3
c1 = 0.002 C, c2 = 0
π
which give the following extremal trajectory:

∗ n sin nt
δz (t) = 0.001C cos nt + − t cos nt (0 ≤ t ≤ tf )
π n
n
δ ż∗ (t) = 0.001nC t sin nt − sin nt (0 ≤ t ≤ tf )
π
and control history, u∗ (t) = 0.002n2 C sinπnt , (0 ≤ t ≤ tf ).

4.7
vy∗ (tf ) + g(tf − t)

tan β ∗ =
vx∗ (tf )
5.2

−1 , λn > 0
ân (t) =
1 , λn < 0
√
with λn = const. With α0 > 0, an = 1, α(t) = Ct/ h, and tf = α0 μ/C.
5.3

cos−1 (bπ/2) , λ2 > 0
β̂(t) =
− cos−1 (bπ/2) , λ2 < 0
where
c
λ̈2 = −c2 sin β − 3h 2 4
/r − 2μ/r 3
+ 2c cos β/r 3
λ2
r2
where c2 is constant and r(t), β(t) are on the optimal trajectory.
5.5

Cn2 , λ2 < 0
ân (t) =
−Cn2 , λ2 > 0
where
c2
λ2 = − sin nt + c2 cos nt
n
The response until the first switching time, ts = tan−1 (nc2 /c1 ), with c1 > 0, c2 > 0,
is the following:
δz(t) = C(0.01 sin nt + cos nt − 1) , (0 ≤ t ≤ ts )

References
1. Abramowitz, M., Stegun, I.A.: Handbook of Mathematical Functions. Dover, New York (1974)
2. Anderson, R.L., Lo, M.W.: Role of invariant manifolds in low-thrust trajectory design. J. Guid.
Control. Dyn. 32, 1921–1930 (2009)
3. Archenti, A.R., Vinh, N.X.: Intermediate thrust arcs and their optimality in a central force field.
J. Optim. Theory Appl. 11, 293–304 (1973)
4. Athans, M., Falb, P.L.: Optimal Control. Dover, New York (2007)
5. Bai, X., Junkins, J.L.: Modified Chebyshev-Picard iteration methods for station-keeping of
translunar halo orbits. Math. Probl. Eng. 2012, 1–18 (2012)
6. Bando, M., Scheeres, D.J.: Attractive sets to unstable orbits using optimal feedback control. J.
Guid. Control. Dyn. 39, 2725–2739 (2016)
7. Battin, R.H.: Astronautical Guidance. McGraw-Hill, New York (1964)
8. Battin, R.H.: An Introduction to the Mathematics and Methods of Astrodynamics. AIAA
Education Series, Reston (1999)
9. Bell, D.J.: Optimal space trajectories – a review of published work. Aeronaut. J. R. Aeronaut.
Soc. 72, 141–146 (1968)
10. Bell, D.J.: Second variation and singular space trajectories. Int. J. Control 14, 697–703 (1971)
11. Bell, D.J.: The non-optimality of Lawden’s spiral. Acta Astronaut. 16, 317–324 (1971)
12. Bell, D.J., Jacobson, D.H.: Singular Optimal Control Problems. Academic, New York (1975)
13. Bellman, R.: Dynamic Programming. Princeton University Press, Princeton (1957)
14. Betts, J.T.: Survey of numerical methods for trajectory optimization. J. Guid. Control Dyn. 21,
193–207 (1998)
15. Bliss, G.A.: Lectures on the Calculus of Variations. University of Chicago Press, Chicago
(1946)
16. Bolonkin, A.A.: Special Extrema in Optimal Control Problems. Eng. Cybern. 2, 170–183
(1969)
17. Breakwell, J.V.: A doubly singular problem in optimal interplanetary guidance. SIAM J.
Control 3, 71–77 (1965)
18. Breakwell, J.V., Kamel, A.A., Ratner, M.J.: Station-keeping for a translunar communication
station. Celest. Mech. 10, 357–373 (1974)
19. Bryson, A.E., Ho, Y.: Applied Optimal Control. Wiley, New York (1979)
20. Byrnes, D.V.: Application of the pseudostate theory to the three-body lambert problem. J.
Astronaut. Sci. 37, 221–232 (1989)
21. Carter, T.: Fuel-optimal maneuvers of a spacecraft relative to a point in circular orbit. J. Guid.
Control Dyn. 7, 710–716 (1984)

https://doi.org/10.1007/978-3-030-03789-5
264 References
22. Celletti, A., Pucacco, G., Stella, D.: Lissajous and halo orbits in the restricted three-body
problem. J. Nonlinear Sci. 25, 343–370 (2015)
23. Chobotov, V.A. (ed.): Orbital Mechanics, Chap. 14. AIAA Education Series, Reston (2002)
24. Clohessy, W.H., Wiltshire, R.S.: Terminal guidance system for satellite rendezvous. J.
Aerospace Sci. 27, 653–658 (1960)
25. Conway, B.A. (ed.): Spacecraft Trajectory Optimization. Cambridge University Press, New
York (2010)
26. Da Silva, F.S.: Solution of the coast-arc problem using Sundman transformation. Acta
Astronaut. 50, 1–11 (2001)
27. Dunham, D.W., Davis, S.A.: Optimization of a multiple Lunar-Swingby trajectory sequence.
J. Astronaut. Sci. 33, 275–288 (1985)
28. Edelbaum, T.N.: Optimal space trajectories. Analytical Mechanics Associates Report, No. 69-
4, Jericho, New York (1969)
29. Farquhar, R.W.: The control and use of libration-point satellites. Goddard space flight center,
Tech. Rep. NASA TR R-346 (September 1970)
30. Farquhar, R.W., Kamel, A.A.: Quasi-periodic orbits about the translunar libration point. Celest.
Mech. 7, 458–473 (1973)
31. Forbes, G.F.: The trajectory of a powered rocket in space. J. Br. Interplanet. Soc. 9, 75–79
(1950)
32. Gabasov, R.: Necessary conditions for optimality of singular control. Eng. Cybern. 5, 28–37
(1968)
33. Ghorbani, M., Assadian, N.: Optimal station-keeping near Earth-Moon collinear libration
points using continuous and impulsive maneuvers. Adv. Space Res. 52, 2067–2079 (2013)
34. Goh, B.S.: Compact forms of the generalized Legendre conditions and the derivation of
the singular extremals. In: Proceedings of 6th Hawaii International Conference on System
Sciences, pp. 115–117 (1973)
35. Gomez, G., Howell, K.C., Simo, C., Masdemont, J.: Station-keeping strategies for translunar
libration point orbits. In: Proceedings of AAS/AIAA Spaceflight Mechanics Meeting, AAS
Paper 98-168, Monterey (1998)
36. Gomez, G., Masdemont, J., Simo, C.: Quasihalo orbits associated with libration points. J.
Astronaut. Sci. 46, 135–176 (1998)
37. Gomez, G., Koon, W.S., Lo, M.W., Marsden, J.E., Masdemont, J., Ross, S.D.: Connecting
orbits and invariant manifolds in the spatial restricted three-body problem. Nonlinearity 17,
1571–1606 (2004)
38. Guckenheimer, J., Holmes, P.: Nonlinear Oscillations, Dynamical Systems, and Bifurcations
of Vector Fields. Springer, New York (1983)
39. Hermes, H.: Controllability and the singular problem. SIAM J. Control 2, 241–260 (1964)
40. Howell, K.: Three-dimensional, periodic, ‘Halo’ orbits. Celest. Mech. 32, 53–71 (1984)
41. Howell, K., Barden, B., Lo, M.: Application of dynamical systems theory to trajectory design
for a libration point mission. J. Astronaut. Sci. 45, 161–178 (1997)
42. Jacobson, D.H.: A new necessary condition of optimality for singular control problems. SIAM
J. Control 7, 578–595 (1969)
43. Jones, B.L., Bishop, R.L.: H2 optimal halo orbit guidance. J. Guid. Control. Dyn. 16, 1118–
1124 (1993)
44. Kechichian, J.A.: The algorithm of the two-impulse time-fixed noncoplanar rendezvous
with drag and oblateness effects. In: Proceedings of AAS/AIAA Astrodynamics Specialist
Conference, AAS Paper 97-645, Sun Valley (1997)
45. Kelley, H.J.: Singular extremals in Lawden’s problem of optimal rocket flight. AIAA J. 1,
1578–1580 (1963)
46. Kelley, H.J., Kopp, R.E., Moyer, H.G.: Singular extremals. In: Leitmann, G. (ed.) Topics in
Optimization, pp. 63–101. Academic, New York (1967)
47. Kopp, R.E., Moyer, H.G.: Necessary conditions for singular extremals. AIAA J. 3, 1439–1444
(1965)
References 265
48. Kulkarni, J.E., Campbell, M.E., Dullerud, G.E.: Stabilization of spacecraft flight in halo orbits:
an H∞ approach. IEEE Trans. Control Syst. Technol. 14, 572–578 (2006)
49. Lawden, D.F.: Inter-orbital transfer of a rocket. Annual Report, pp. 321–333. British Interplan-
etary Society (1952)
50. Lawden, D.F.: Fundamentals of space navigation. J. Br. Interplanet. Soc. 13, 87–101 (1954)
51. Lawden, D.F.: Optimal intermediate-thrust arcs in a gravitational field. Acta Astronaut. 8, 106–
123 (1954)
52. Lawden, D.F.: Optimal Trajectories for Space Navigation. Butterworths, London (1963)
53. Leitmann, G.: On a class of variational problems in rocket flight. J. Aerospace Sci. 26, 586–591
(1959)
54. Li, C., Liu, G., Huang, J., Gao, T., Guo, Y.: Stationkeeping control for libration point orbits
using NMPC. In: Proceedings of AAS/AIAA Astrodynamics Specialist Conference, AAS
Paper 15-692, Vail (2015)
55. Lian, Y., Gomez, G., Masdemont, J.J., Tang, G.: Station keeping of real Earth-Moon libration
point orbits using discrete-time sliding mode approach. Commun. Nonlinear Sci. Numer.
Simul. 19, 3792–3807 (2014)
56. Macdonald, M., McInnes, C.R.: Analytical control laws for planet-centred solar sailing. J.
Guid. Control. Dyn. 28, 1038–1048 (2005)
57. Marec, J.P.: Optimal Space Trajectories. Elsevier, New York (1979)
58. Marinca, V., Herisanu, N.: Nonlinear Dynamical Systems in Engineering. Springer, Berlin
(2012)
59. Mattheij, R.M.M., Molenaar, J.: Ordinary Differential Equations in Theory and Practice. SIAM
Classics in Applied Mathematics, vol. 43. SIAM, Philadelphia (2002)
60. Meyer, K.R., Hall, G.R.: Introduction to Hamiltonian Dynamical Systems and the N-body
Problem. Springer, New York (1992)
61. Minter, C.F., Fuller-Rowell, T.J.: A robust algorithm for solving unconstrained two-point
boundary value problems. In: Proceedings of AAS/AIAA Spaceflight Mechanics Meeting,
AAS Paper 05-129, Copper Mountain (2005)
62. Moritz, H.: Advanced Physical Geodesy. Abacus Press, New York (1980)
63. Nazari, M., Anthony, W., Butcher, E.A.: Continuous thrust stationkeeping in Earth-Moon
L1 halo orbits based on LQR control and Floquet theory. In: Proceedings of AAS/AIAA
Astrodynamics Specialist Conference, AIAA Paper 2014-4140, San Diego, CA (2014)
64. Otten, M., McInnes, C.R.: Near minimum-time trajectories for solar sails. J. Guid. Control.
Dyn. 24, 632–634 (2001)
65. Pan, B., Zheng, C., Lu, P., Gao, B.: Reduced transversality conditions in optimal space
trajectories. J. Guid. Control. Dyn. 36, 1289–1300 (2013)
66. Pavlak, T.A., Howell, K.C.: Strategy for optimal, long-term stationkeeping of libration point
orbits in the Earth-Moon system. In: Proceedings of AAS/AIAA Astrodynamics Specialist
Conference, AIAA Paper 2012-4665, Minneapolis (2012)
67. Pines, S.: Constants of the motion for optimum thrust trajectories in a central force field. AIAA
J. 2, 2010–2014 (1964)
68. Pontryagin, L.S., Boltyanskii, V.G., Gamkrelidze, R.V., Mishchenko, E.F.: The Mathematical
Theory of Optimal Processes. Wiley, New York (1962)
69. Richardson, D.L.: Analytic construction of periodic orbits about the collinear points. Celest.
Mech. 22, 241–253 (1980)
70. Richardson, D.L.: Halo orbit formulation for the ISEE-3 mission. J. Guid. Control. Dyn. 3,
543–548 (1980)
71. Robbins, H.M.: Optimality of intermediate-thrust arcs of rocket trajectories. AIAA J. 3, 1094–
1098 (1965)
72. Shampine, L.F., Gladwell, I., Thompson, S.: Solving ODEs with MATLAB. Cambridge
University Press, Cambridge (2003)
73. Shirobokov, M., Trofimov, S., Ovchinnikov, M.: Survey of station-keeping techniques for
libration point orbits. J. Guid. Control. Dyn. 40, 1085–1105 (2017)
74. Slotine, J.E., Li, W.: Applied Nonlinear Control. Prentice-Hall, Englewood Cliffs (1991)
266 References
75. Snow, D.R.: Singular optimal controls for a class of minimum effort problems. SIAM J. Control
2, 203–219 (1964)
76. Stiefel, E.L., Scheifele, G.: Linear and Regular Celestial Mechanics. Springer, New York
(1971)
77. Stoer, J., Bulirsch, R.: Introduction to Numerical Analysis. Springer, New York (2002)
78. Szebehely, V.: Theory of Orbits: The Restricted Problem of Three Bodies. Academic, New
York (1967)
79. Tewari, A.: Modern Control Design with MATLAB and Simulink. Wiley, Chichester (2002)
80. Tewari, A.: Atmospheric and Space Flight Dynamics–Modeling and Simulation with MAT-
LAB and Simulink. Birkhäuser, Boston (2006)
81. Tewari, A.: Advanced Control of Aircraft, Spacecraft, and Rockets. Wiley, Chichester (2011)
82. Thurman, R., Worfolk, P.A.: Geometry of halo orbits in the circular restricted three-body
problem. Technical Report GCG95, University of Minnesota, Minneapolis (1996)
83. Fraeijs de Veubeke, B.: Canonical transformations and the thrust-coast-thrust optimal transfer
problem. Acta Astronaut. 11, 271–282 (1965)
84. Werner, R.A., Scheeres, D.J.: Exterior gravitation of a polyhedron derived and compared with
harmonic and mascon gravitation representations of asteroid 4769 CASTALIA. Celest. Mech.
Dyn. Astron. 65, 313–344 (1996)
85. Wertz, J.R. (ed.): Spacecraft Attitude Determination and Control. Kluwer Academic Publish-
ers, Dordrecht (1978)
86. Yamanaka, K., Ankersen, F.: New state transition matrix for relative motion on an arbitrary
elliptical orbit. J. Guid. Control. Dyn. 25, 60–66 (2002)
87. Yang, H., Bai, X., Baoyin, H.: Rapid generation of time-optimal trajectories for asteroid
landing via convex optimization. J. Guid. Control. Dyn. 40, 628–641 (2017)
88. Zhang, C., Zhao, Y.: Low-thrust minimum-fuel optimization in the circular restricted three-
body problem. J. Guid. Control. Dyn. 38, 1501–1509 (2015)
Index
A Circular restricted three-body problem, 200,

Accessory problem, 64 210
Algebraic Riccati equation, 135, 137 Circumferential thrust, 120
Angular momentum, 78 Classical orbital elements, 124
Angular speed, 209 Closure rate, 143
Angular velocity, 151 Coasting arc, 102
Argument of periapsis, 82 Co-latitude, 205
Ascending node, 81 Collinear Lagrangian points, 210, 214, 232
Attitude control, 3 Collocation method, 71, 186, 248
Attitude kinematics, 4 Constant acceleration bound, 166
Augmented Lagrangian, 21 Continued fraction, 100, 135
Augmented performance index, 21, 68 Control-affine system, 55, 61
Autonomous system, 34 Control interval, 2, 20
Axisymmetric mass distribution, 201, 205 Controllability, 137
Azimuth, 80 Control variables, 2, 8
Coplanar orbital transfer, 122
Coplanar rendezvous, 196
B Coplanar three-body problem, 209
Backstepping, 252 Corrective control, 132
Bang–bang control, 56 Costate equation, 25
Barker’s equation, 86 Costate variables, 21
Barycenter, 209, 211 Cross-product steering, 145
Bifurcation, 239
Body axes, 3
Bolza problem, 20 D
Bounded exhaust rate, 166 Declination, 80
Brachistochrone problem, 74 Disturbing acceleration, 132
Docking, 142
Dynamic optimization, 7
C
Celestial frame, 80
Central gravity field, 168, 169 E
Characteristic speed, 161 Earth–Moon system, 214, 221
Chebyshev–Picard iteration, 250 Eccentric anomaly, 83
Chord, 93 Eccentricity, 79

https://doi.org/10.1007/978-3-030-03789-5
268 Index
Eccentricity vector, 78 Hyperbolic orbit, 85

Ecliptic plane, 81 Hypergeometric function, 99
Electric propulsion, 161
Elliptical orbit, 83
Energy control, 138 I
Energy-optimal guidance, 148, 150 Impulsive control, 249
Equality constraint Impulsive orbital transfer, 77, 88
dynamic, 20 Inclination, 81
static, 11 Inequality constraint, 2, 18, 55
Equatorial plane, 81 Inertial frame, 3
Equatorial radius, 205 Initial condition, 20
Equilateral triangle, 210 Initial value problem, 70
Equinoctial elements, 127 Intermediate-thrust arc, 165, 173
Euler–Lagrange equations, 26 Interstellar mission, 228
Exhaust rate, 160 In-track displacement, 127, 151, 189
Exhaust speed, 160 Invariant manifold, 246, 248
Extremal point, 9 Inverse-square gravity field, 172
Extremal trajectory, 7 Irregularly shaped body, 199, 208
F J
Feasible region, 19 Jacobian, 12, 37, 62, 244
Feedback controller, 5 Jacobi’s integral, 218
Feedforward controller, 5 Jacobson’s necessary condition, 68
Flat-trajectory approximation, 147 Jeffery’s constants, 206
Flight-path angle, 79
Floquet mode, 250
Focus, 79 K
Formation flying, 153 Kalman filter, 133, 251
Free-return trajectory, 220 Kepler’s equation, 83
Frequency matching, 238
L
G Lagrange’s coefficients, 87, 97, 135
Gauss’ perturbation equations, 119, 124 Lagrange’s planetary model, 124
Generalized momentum, 217 Lagrange multipliers, 12
Generating functions, 204 Lagrange problem, 20
Global asymptotic stability, 34, 41 Lagrangian, 2, 12, 20, 217, 234
Global minimum, 10 Lagrangian points, 213
Gradient operator, 218 Lambert’s problem, 88, 95
Gravitational potential, 201, 217 Lambert’s theorem, 95
Gravity-assist maneuver, 222 Lawden’s spiral, 173, 177
Gravity-gradient matrix, 134, 162, 207 Legendre–Clebsch
Guidance, 138 generalized necessary condition, 58, 59
necessary condition, 31
sufficient condition, 14, 31, 55
H Legendre polynomial, 203, 234
Halo orbit, 216, 233, 237, 239 Lindstedt–Poincaré method, 237
Hamiltonian, 12, 25, 217 Line-of-sight interception, 139
Hamilton–Jacobi–Bellman equation, 33 Linear
Hessian, 9 optimal control, 135
Hill–Clohessy–Wiltshire model, 151, 152, 189 quadratic regulator, 38, 64, 135, 250, 251
Hohmann transfer, 105 Line of nodes, 81
Hyperbolic anomaly, 85 Line of sight, 140
Index 269
Lissajous trajectory, 233 Orbit equation, 79

Local-horizon frame, 80 Osculating orbit, 118
Local minimum, 10 Out-of-plane displacement, 127, 151, 189
Longitude, 205 Outer bi-elliptical transfer, 110
Low-thrust orbital transfer, 120
Low-thrust three-body transfer, 222
LQG/LTR formulation, 251 P
Lunar landing problem, 50 Parabolic orbit, 86
Lyapunov function, 34 Parameter, 79
Patched two-body trajectory, 231
Performance index, 2, 8, 20
M Periapsis, 79
Maneuvering, 138 Perifocal frame, 80
Maneuvering spacecraft, 151 Perturbation model, 119
Manifold, 246 Perturbing acceleration, 124
Matrix Riccati equation, 40, 64 Plane-change maneuver, 122
Maximum-thrust arc, 165 Polyhedral model, 208
Mayer problem, 20 Pontryagin’s minimum principle, 7, 31, 56,
Mean anomaly, 83 165, 168, 175
Mean motion, 83 Primaries, 210
Measurement noise, 132 Primer vector, 164
Minimum energy escape, 86 Primer vector method, 159
Minimum energy transfer, 92 Process noise, 132
Minimum point, 10 Projected miss distance, 141
Model-predictive control, 252
Monodromy matrix, 244
Monte Carlo simulation, 250 Q
Quadratic cost, 35
Quasi-halo orbit, 233
N Quasi-steady approximation, 136
Navigation, 3 Quintic equation of Lagrange, 210
Necessary condition, 7, 8, 13, 23, 25
Neighboring optimal solution, 33
Newton’s method, 48, 84, 100, 244 R
Nonlinear feedback control, 33 Radial displacement, 151, 189
Non-spherical body, 201 Radial thrust, 120, 128
Non-spherical gravity field, 199 Recurrence formulae, 204
Normal thrust, 120 Remaining burn time, 148
Null-thrust arc, 102, 165 Rendezvous problem, 153, 191, 194
Restricted three-body problem, 200
Right-ascension of the ascending node, 81
O Right ascension, 80
Oblateness, 201, 206, 207 Robust control, 135
Optimal control, 1, 7
Optimal impulsive transfer, 88, 101
Optimal orbital regulation, 127 S
Optimal orbital transfer, 123, 199 Saddle point, 10
Optimal return function, 32 Semi-major axis, 79
Optimal thrust direction, 163, 175 Shooting method, 70, 231
Orbit, 78 Singular control, 31, 55, 165
Orbital motion, 3, 117 Singular point, 10
Orbital plane, 78 Skew-symmetric matrix, 66
Orbital tracking, 132 Sliding mode, 250
Orbital velocity, 78 Solar-sail mission, 231
270 Index
Specific energy, 78 Thrust pitch angle, 125

Specific impulse, 160 Thrust yaw angle, 125
Sphere of influence, 231 Time-averaged dynamics, 137
Spherical harmonics, 201, 204, 208 Time of flight, 88
Stability augmentation, 137 Time-invariant gravity field, 168
Stable manifold, 246 Time-invariant system, 34
State equation, 2, 20 Time of periapsis, 83
State transition matrix, 42, 87, 134, 193, 243 Time to go, 142
State variables, 2, 8 Tracking system, 133
Static optimization, 8 Transformational approach, 61
Stationary condition, 13, 25 Transversality condition, 25, 188, 231
Station keeping, 233, 249 Triangular Lagrangian points, 214
Stumpff function, 97 True anomaly, 79
Sufficient condition, 8, 9, 13, 34 Two-body problem, 77
Sun–Earth system, 239 Two-point boundary-value problem, 39
Sun–Jupiter system, 215
Swing-by, 222
Switching condition, 56, 165 U
Switching function, 56 Universal variable, 97
Symplectic matrix, 42, 88, 134, 244 Unstable equilibrium, 237
Synodic frame, 211 Unstable manifold, 246, 249
T V
Tangential thrust, 131 Variational model, 119, 124
Target, 151, 189 Velocity-to-be-gained, 145
Terminal cost, 20 Vernal equinox, 81
Terminal time, 2
Three-body problem, 77, 208
Throttling, 160 Z
Thrust acceleration, 117, 120, 159 Zermelo’s navigation problem, 74
Thrust direction, 159, 163, 189, 224 Zero relative speed contour, 219

Optimal Space Flight Navigation

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Optimal Space Flight Navigation

Hochgeladen von

Copyright:

Verfügbare Formate

Control Engineering

Editorial Advisory Board

More information about this series at http://www.springer.com/series/4988

Optimal Space Flight

ISSN 2373-7719 ISSN 2373-7727 (electronic)

Library of Congress Control Number: 2018961705

© Springer Nature Switzerland AG 2019

This book is primarily written to consolidate optimal space flight navigation

resources present on board Apollo 11 could be surpassed by those of today’s pocket

assumed of the reader, which can be supplemented by textbooks on these topics,

Kanpur, India Ashish Tewari

3.2.2 Time Equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.8 Linearization Relative to a Circular Orbit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

Answers to Selected Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

1.1 Optimal Control

© Springer Nature Switzerland AG 2019 1

are termed non-deterministic and must be handled by stochastic (probabilistic)

1.2 Space Vehicle Guidance

attached to the spacecraft. The attitude is controlled by externally applied torque

If the spacecraft is assumed to be rigid with a constant inertia tensor, J ∈ R3×3 ,

Td(t) Spacecraft (t)

Navigational rd(t), vd(t) Navigational a(t) Spacecraft r(t), v(t)

© Springer Nature Switzerland AG 2019 7

2.2 Optimization of Static Systems

In order to motivate the reader, it is necessary to introduce the concepts of

The positive definiteness of the Hessian at an extremal point is therefore a sufficient

then it is a maximum point of L(x, u) with respect to u.

Since the performance index is a quadratic function of u = (u1 , u2 )T , its only

or, u∗1 = u∗2 = 0. The Hessian is the following:

L = −u21 + 2u1 u2 + 3u22 ?

Since the performance index is a quadratic function of u = (u1 , u2 )T , its only

or, u∗1 = u∗2 = 0. The Hessian is the following:

L = (u1 − u22 )(u1 − 3u22 )

It is easily seen that u∗ = (0, 0)T satisfies the stationary condition, Lu = 0:

2.2.1 Static Equality Constraints

while holding df = 0, with Lu = ∂L/∂u ∈ R1×m and Lx = ∂L/∂x ∈ R1×n . This

An alternative necessary condition for minimization in the presence of the equal-

H (x, u, λ) = L(x, u) + λT f (x, u) (2.11)

This implies that

(λ∗ )T = −Lx fx−1 (2.14)

Upon substituting Hx = 0 in Eq. (2.12), the following stationary condition is

Equations (2.13)–(2.15) are the necessary conditions for the minimization of

The substitution of Eqs. (2.17)–(2.18) into

along with the following Taylor series variation in x,

dx = −fx−1 fu du + terms involving higher powers of (dx, du) (2.20)

results in the following variation in L up to second-order terms:

where a, b, c, m are real constants.

The necessary conditions for minimization are the following:

Example 2.5 Determine the extremal and minimum points for

L = 3z2 + u21 ez + u22 y 2 + (u3 − 2)2

with respect to u = (u1 , u2 , u3 )T ∈ R3×1 , subject to the following constraints:

where x T = (z, y) ∈ R1×3 .

H = 3z2 + u21 ez + u22 y 2 + (u3 − 2)2 + λ1 (z − u1 y 2 ) + λ2 (y + 2u3 )

which is identically satisfied by:

In order to determine the Lagrange multipliers, the following condition is imposed:

which are satisfied by:

fuT (fxT )−1 Hxx fx−1 fu =

2.2.2 Inequality Constraints

Consider general inequality constraints of the form:

while minimizing L(x, u) with respect to u. Extending the concept of Lagrange

J (x, u) = L(x, u) + λT f (x, u) . (2.25)

f (x̂, û) < 0 , (2.26)

f (x̂, û) = 0 , (2.27)

where δu must satisfy the constraint, Eq. (2.24), with λ > 0: